[Robust Speech Event] Add guides (#15155)
* up * improve readme * up * up * more info * up * up * Apply suggestions from code review Co-authored-by: Anton Lozhkov <aglozhkov@gmail.com> * add more stuff for eval * update * up * Update README.md * Update examples/research_projects/xls_r/README.md Co-authored-by: Omar Sanseviero <osanseviero@users.noreply.github.com> * apply omar's suggestions Co-authored-by: Anton Lozhkov <aglozhkov@gmail.com> Co-authored-by: Omar Sanseviero <osanseviero@users.noreply.github.com>
This commit is contained in:
committed by
GitHub
parent
1a354d53c4
commit
e118e085ea
722
examples/research_projects/robust-speech-event/README.md
Normal file
722
examples/research_projects/robust-speech-event/README.md
Normal file
@@ -0,0 +1,722 @@
|
|||||||
|
# Robust Speech Challange 🤗
|
||||||
|
|
||||||
|
Welcome to the robust speech recognition challenge 🎙️ !
|
||||||
|
|
||||||
|
The goal of this event is to build **robust**, **real-world** speech recognition (ASR) systems in as many languages as possible 🌏🌍🌎.
|
||||||
|
If necessary and available, free access to a V100 32 GB GPU will kindly be provided by the [OVH could team](https://us.ovhcloud.com/) 🚀.
|
||||||
|
This document summarizes all the relevant information required for the speech community event 📋.
|
||||||
|
|
||||||
|
To sign-up, please see [this forum post](https://discuss.huggingface.co/t/open-to-the-community-robust-speech-recognition-challenge/13614) 🤗. Please make sure to:
|
||||||
|
- Read it in detail
|
||||||
|
- Fill the google form
|
||||||
|
- Join our Discord server in the #join-sprint channel.
|
||||||
|
|
||||||
|
## Table of Contents
|
||||||
|
|
||||||
|
- [TLDR;](#tldr)
|
||||||
|
- [Important dates](#important-dates)
|
||||||
|
- [How to install pytorch, transformers, datasets](#how-to-install-relevant-libraries)
|
||||||
|
- [Data and Preprocessing](#data-and-preprocessing)
|
||||||
|
- [How to fine-tune an acoustic model](#how-to-finetune-an-acoustic-model)
|
||||||
|
- [How to fine-tune with OVH could](#how-to-finetune-with-ovh-cloud)
|
||||||
|
- [How to combine n-gram language models with acoustic model](#how-to-combine-n-gram-with-acoustic-model)
|
||||||
|
- [Evaluation](#evaluation)
|
||||||
|
- [Prizes](#prizes)
|
||||||
|
- [Communication and Problems](#communication-and-problems)
|
||||||
|
- [Talks](#talks)
|
||||||
|
- [General Tips & Tricks](#general-tips-and-tricks)
|
||||||
|
|
||||||
|
## TLDR
|
||||||
|
|
||||||
|
Participants are encouraged to leverage pre-trained speech recognition checkpoints,
|
||||||
|
preferably [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53),
|
||||||
|
to train a speech recognition system in a language of their choice.
|
||||||
|
|
||||||
|
Speech recognition systems should be trained using **PyTorch**, **🤗 Transformers**, and, **🤗 Datasets**.
|
||||||
|
For more information on how to install the above libraries, please read through
|
||||||
|
[How to install pytorch, transformers, datasets](#how-to-install-relevant-libraries).
|
||||||
|
|
||||||
|
Participants can make use of whatever data they think is useful to build a
|
||||||
|
speech recognition system for **real-world** audio data -
|
||||||
|
**except** the Common Voice `"test"` split of their chosen language.
|
||||||
|
The section [Data and preprocessing](#data-and-preprocessing) explains
|
||||||
|
in more detail what audio data can be used, how to find suitable audio data, and
|
||||||
|
how the audio data can be processed.
|
||||||
|
|
||||||
|
For training, it is recommended to use the [official training script](https://github.com/huggingface/transformers/blob/master/examples/pytorch/speech-recognition/run_speech_recognition_ctc.py) or a modification thereof. A step-by-step guide on how to fine-tune
|
||||||
|
an acoustic model for a speech recognition system can be found under [How to fine-tune an acoustic model](#how-to-finetune-an-acoustic-model).
|
||||||
|
If possible it is encouraged to fine-tune the acoustic models on local GPU machines, but
|
||||||
|
if those are not available, the OVH could team kindly provides a limited
|
||||||
|
number of GPUs for the event. Simply fill out [this google form](https://forms.gle/GFZkMkKLiufi75g28) to get access to a GPU.
|
||||||
|
For more information on how to train an acoustic model on one of OVH's GPU - see [How to fine-tune a speech recognition model with OVHcould](#how-to-fine-tune-with-ovh-cloud).
|
||||||
|
|
||||||
|
The performance of speech recognition system can often significantly be improved by adding a
|
||||||
|
language model for decoding. For more information on how to add a language model, please
|
||||||
|
take a look at [How to combine n-gram language models with speech recognition models](#how-to-combine-n-gram-with-model).
|
||||||
|
|
||||||
|
During the event, the speech recognition system will be evaluated on both the Common Voice `"test"` split
|
||||||
|
of the participants' chosen language as well as the *real-world* `"dev"` data provided by
|
||||||
|
the Hugging Face team.
|
||||||
|
At the end of the robust speech recognition challenge, the speech recognition system will also be evaluated on the
|
||||||
|
*real-world* `"test"` data provided by the Hugging Face team. Each participant should add an
|
||||||
|
`eval.py` script to her/his model repository in a specific format that lets one easily
|
||||||
|
evaluate the speech recognition system on both Common Voice's `"test"` data as well as the *real-world* audio
|
||||||
|
data. Please read through the [Evaluation](#evaluation) section to make sure your evaluation script is in the correct format. Speech recognition systems
|
||||||
|
with evaluation scripts in an incorrect format can sadly not be considered for the Challenge.
|
||||||
|
|
||||||
|
At the end of the event, the best performing speech recognition system
|
||||||
|
will receive a prize 🏆 - more information regarding the prizes can be found under [Prizes](#prizes).
|
||||||
|
|
||||||
|
We believe that framing the event as a competition is more fun, but at the core, the event is about
|
||||||
|
creating speech recognition systems in as many languages as possible as a community.
|
||||||
|
This can be achieved by working together, helping each other to solve bugs, share important findings, etc...🤗
|
||||||
|
|
||||||
|
**Note**:
|
||||||
|
Please, read through the section on [Communication & Problems](#communication-and-problems) to make sure you
|
||||||
|
know how to ask for help, etc...
|
||||||
|
All important announcements will be made on discord. Please make sure that
|
||||||
|
you've joined [this discord channel](https://discord.gg/SHr5wC7m)
|
||||||
|
|
||||||
|
Also, please make sure that you have been added to the [Speech Event Organization](https://huggingface.co/speech-recognition-community-v2).
|
||||||
|
You should have received an invite by email. If you didn't receive an invite, please contact the organizers, *e.g.* Anton, Patrick, or Omar directly on discord.
|
||||||
|
|
||||||
|
## Important dates
|
||||||
|
|
||||||
|
- **12.01.** Official announcement of the community event. Make sure to sign-up in [this google form](TODO: )
|
||||||
|
- **12.01. - 19.01.** Participants sign up for the event.
|
||||||
|
- **19.01.** Release of all relevant guides and training scripts.
|
||||||
|
- **24.01.** Start of the community week! OVH & Hugging Face gives access to GPUs.
|
||||||
|
- **24.01. - 07.02.** The OVH & Hugging Face team will be available for any questions, problems the participants might have.
|
||||||
|
- **07.02.** Access to GPU is deactivated and community week officially ends.
|
||||||
|
|
||||||
|
|
||||||
|
## Data and preprocessing
|
||||||
|
|
||||||
|
In this section, we will quickly go over how to find suitable training data and
|
||||||
|
how to preprocess it.
|
||||||
|
|
||||||
|
To begin with, **all data except Common Voice's `"test"` data can be used as training data.**
|
||||||
|
The exception includes all Common Voice versions as the test data split of later Common Voice versions often
|
||||||
|
overlaps with the one of previous versions, *e.g.* the test data of Common Voice 7 in English is
|
||||||
|
to a big part identical to the test data of Common Voice 6 in English:
|
||||||
|
|
||||||
|
```python
|
||||||
|
load_dataset("mozilla-foundation/common_voice_7_0", "en", split="test")
|
||||||
|
```
|
||||||
|
|
||||||
|
includes more or less the same data as
|
||||||
|
|
||||||
|
```python
|
||||||
|
load_dataset("mozilla-foundation/common_voice_6_1", "en", split="test")
|
||||||
|
```
|
||||||
|
|
||||||
|
However, we strongly encourage participants to make use of Common Voice's other splits, *e.g.* `"train"` and `"validation"`.
|
||||||
|
For most languages, the Common Voice dataset offers already a decent amount of training data. It is usually
|
||||||
|
always advantageous to collect additional data. To do so, the participants are in a first step encouraged to search the
|
||||||
|
Hugging Face Hub for additional audio data, for example by selecting the category
|
||||||
|
["speech-processing"](https://huggingface.co/datasets?task_categories=task_categories:speech-processing&sort=downloads).
|
||||||
|
All datasets that are available on the Hub can be downloaded via the 🤗 Datasets library in the same way Common Voice is downloaded.
|
||||||
|
If one wants to combine multiple datasets for training, it might make sense to take a look at
|
||||||
|
the [`interleave_datasets`](https://huggingface.co/docs/datasets/package_reference/main_classes.html?highlight=interleave#datasets.interleave_datasets) function.
|
||||||
|
|
||||||
|
In addition, participants can also make use of their audio data. Here, please make sure that you **are allowed to use the audio data**. E.g., if audio data
|
||||||
|
is taken from media platforms, such as YouTube, it should be verified that the media platform and the owner of the data have given her/his approval to use the audio
|
||||||
|
data in the context of machine learning research. If you are not sure whether the data you want to use has the appropriate licensing, please contact the Hugging Face
|
||||||
|
team on discord.
|
||||||
|
|
||||||
|
Next, let's talk about preprocessing. Audio data and transcriptions have to be brought into the correct format when
|
||||||
|
training the acoustic model (example shown in [How to fine-tune an acoustic model](#how-to-finetune-an-acoustic-model)).
|
||||||
|
It is recommended that this is done by using 🤗 Datasets `.map()` function as shown
|
||||||
|
[here](https://github.com/huggingface/transformers/blob/9a2dabae7002258e41419491c73dd43ad61b5de7/examples/pytorch/speech-recognition/run_speech_recognition_ctc.py#L444). As can be
|
||||||
|
see we can pass some characters that will be removed from the transcriptions, *e.g.*: `--chars_to_ignore , ? . ! - \; \: \" “ % ‘ ” <20> \`
|
||||||
|
on the official ["Single GPU Example"](https://github.com/huggingface/transformers/tree/master/examples/pytorch/speech-recognition#single-gpu-ctc).
|
||||||
|
The participants are free to modify this preprocessing by removing more characters or even replacing characters as
|
||||||
|
it is done in the [official blog post](https://github.com/huggingface/transformers/blob/9a2dabae7002258e41419491c73dd43ad61b5de7/examples/pytorch/speech-recognition/run_speech_recognition_ctc.py#L444).
|
||||||
|
**However**, there are some rules regarding what characters are allowed to be removed/replaced and which are not.
|
||||||
|
These rules are not this straightforward and therefore often have to be evaluated case-by-case.
|
||||||
|
It is allowed (and recommended) to normalize the data to only have lower-case characters. It is also allowed (and recommended) to remove typographical
|
||||||
|
symbols and punctuation marks. A list of such symbols can *e.g.* be found [here](https://en.wikipedia.org/wiki/List_of_typographical_symbols_and_punctuation_marks) - however here we already must be careful. We should **not** remove a symbol that would change the meaning of the words, *e.g.* in English,
|
||||||
|
we should not remove the single quotation mark `'` since it would change the meaning of the word `"it's"` to `"its"` which would then be incorrect.
|
||||||
|
So the golden rule here is to not remove any characters that could change the meaning of a word into another word. This is not always obvious and should
|
||||||
|
be given some consideration. As another example, it is fine to remove the "Hyphen-minus" sign "`-`" since it doesn't change the
|
||||||
|
meaning of a word to another one. *E.g.* "`fine-tuning`" would be changed to "`finetuning`" which has still the same meaning.
|
||||||
|
|
||||||
|
Since those choices are not always obvious when in doubt feel free to ask on Discord or even better post your question on the forum, as was
|
||||||
|
done, *e.g.* [here](https://discuss.huggingface.co/t/spanish-asr-fine-tuning-wav2vec2/4586).
|
||||||
|
|
||||||
|
## How to install relevant libraries
|
||||||
|
|
||||||
|
The following libraries are required to fine-tune a speech model with 🤗 Transformers and 🤗 Datasets in PyTorch.
|
||||||
|
|
||||||
|
- [PyTorch](https://pytorch.org/)
|
||||||
|
- [Transformers](https://github.com/huggingface/transformers)
|
||||||
|
- [Datasets](https://github.com/huggingface/datasets)
|
||||||
|
|
||||||
|
We recommend installing the above libraries in a [virtual environment](https://docs.python.org/3/library/venv.html).
|
||||||
|
If you're unfamiliar with Python virtual environments, check out the [user guide](https://packaging.python.org/guides/installing-using-pip-and-virtual-environments/). Create a virtual environment with the version of Python you're going
|
||||||
|
to use and activate it.
|
||||||
|
|
||||||
|
You should be able to run the command:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
python3 -m venv <your-venv-name>
|
||||||
|
```
|
||||||
|
|
||||||
|
You can activate your venv by running
|
||||||
|
|
||||||
|
```bash
|
||||||
|
source ~/<your-venv-name>/bin/activate
|
||||||
|
```
|
||||||
|
|
||||||
|
To begin with please make sure you have PyTorch and CUDA correctly installed.
|
||||||
|
The following command should return ``True``:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
python -c "import torch; print(torch.cuda.is_available())"
|
||||||
|
```
|
||||||
|
|
||||||
|
If the above command doesn't print ``True``, in the first step, please follow the
|
||||||
|
instructions [here](https://pytorch.org/) to install PyTorch with CUDA.
|
||||||
|
|
||||||
|
We strongly recommend making use of the provided PyTorch examples scripts in [transformers/examples/pytorch/speech-recognition](https://github.com/huggingface/transformers/tree/master/examples/pytorch/speech-recognition) to train your speech recognition
|
||||||
|
system.
|
||||||
|
In all likelihood, you will adjust one of the example scripts, so we recommend forking and cloning the 🤗 Transformers repository as follows.
|
||||||
|
|
||||||
|
1. Fork the [repository](https://github.com/huggingface/transformers) by
|
||||||
|
clicking on the 'Fork' button on the repository's page. This creates a copy of the code
|
||||||
|
under your GitHub user account.
|
||||||
|
|
||||||
|
2. Clone your fork to your local disk, and add the base repository as a remote:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
$ git clone https://github.com/<your Github handle>/transformers.git
|
||||||
|
$ cd transformers
|
||||||
|
$ git remote add upstream https://github.com/huggingface/transformers.git
|
||||||
|
```
|
||||||
|
|
||||||
|
3. Create a new branch to hold your development changes. This is especially useful to share code changes with your team:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
$ git checkout -b a-descriptive-name-for-my-project
|
||||||
|
```
|
||||||
|
|
||||||
|
4. Set up a PyTorch environment by running the following command your virtual environment:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
$ pip install -e ".[torch-speech]"
|
||||||
|
```
|
||||||
|
|
||||||
|
(If transformers was already installed in the virtual environment, remove
|
||||||
|
it with `pip uninstall transformers` before reinstalling it in editable
|
||||||
|
mode with the `-e` flag.)
|
||||||
|
|
||||||
|
If you have already cloned that repo, you might need to `git pull` to get the most recent changes in the `transformers`
|
||||||
|
library.
|
||||||
|
|
||||||
|
Running this command will automatically install `torch` and the most relevant
|
||||||
|
libraries required for fine-tuning a speech recognition system.
|
||||||
|
|
||||||
|
Next, you should also install the 🤗 Datasets library. We strongly recommend installing the
|
||||||
|
library from source to profit from the most current additions during the community week.
|
||||||
|
|
||||||
|
Simply run the following steps:
|
||||||
|
|
||||||
|
```
|
||||||
|
$ cd ~/
|
||||||
|
$ git clone https://github.com/huggingface/datasets.git
|
||||||
|
$ cd datasets
|
||||||
|
$ pip install -e ".[streaming]"
|
||||||
|
```
|
||||||
|
|
||||||
|
If you plan on contributing a specific dataset during
|
||||||
|
the community week, please fork the datasets repository and follow the instructions
|
||||||
|
[here](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-create-a-pull-request).
|
||||||
|
|
||||||
|
To verify that all libraries are correctly installed, you can run the following command in a Python shell.
|
||||||
|
It verifies that both `transformers` and `datasets` have been correclty installed.
|
||||||
|
|
||||||
|
```python
|
||||||
|
from transformers import AutoModelForCTC, AutoProcessor
|
||||||
|
from datasets import load_dataset
|
||||||
|
|
||||||
|
dummy_dataset = load_dataset("common_voice", "ab", split="test")
|
||||||
|
|
||||||
|
model = AutoModelForCTC.from_pretrained("hf-internal-testing/tiny-random-wav2vec2")
|
||||||
|
model.to("cuda")
|
||||||
|
|
||||||
|
processor = AutoProcessor.from_pretrained("hf-internal-testing/tiny-random-wav2vec2")
|
||||||
|
|
||||||
|
input_values = processor(dummy_dataset[0]["audio"]["array"], return_tensors="pt", sampling_rate=16_000).input_values
|
||||||
|
input_values = input_values.to("cuda")
|
||||||
|
|
||||||
|
logits = model(input_values).logits
|
||||||
|
|
||||||
|
assert logits.shape[-1] == 32
|
||||||
|
```
|
||||||
|
|
||||||
|
## How to finetune an acoustic model
|
||||||
|
|
||||||
|
In this section, we show you how to fine-tune a pre-trained [XLS-R Model](https://huggingface.co/docs/transformers/model_doc/xls_r) on the [Common Voice 7 dataset](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0).
|
||||||
|
|
||||||
|
We recommend fine-tuning one of the following pre-trained XLS-R checkpoints:
|
||||||
|
|
||||||
|
- [300M parameters version](https://huggingface.co/facebook/wav2vec2-xls-r-300m)
|
||||||
|
- [1B parameters version](https://huggingface.co/facebook/wav2vec2-xls-r-1b)
|
||||||
|
- [2B parameters version](https://huggingface.co/facebook/wav2vec2-xls-r-2b)
|
||||||
|
|
||||||
|
To begin with, please note that to use the Common Voice dataset, you
|
||||||
|
have to accept that **your email address** and **username** are shared with the
|
||||||
|
mozilla-foundation. To get access to the dataset please click on "*Access repository*" [here](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0).
|
||||||
|
|
||||||
|
Next, we recommended that you get familiar with the XLS-R model and its capabilities.
|
||||||
|
In collaboration with [Fairseq's Wav2Vec2 team](https://github.com/pytorch/fairseq/tree/main/examples/wav2vec),
|
||||||
|
we've written ["Fine-tuning XLS-R for Multi-Lingual ASR with 🤗 Transformers"](https://huggingface.co/blog/fine-tune-xlsr-wav2vec2) which gives an in-detail explanation of how XLS-R functions and how it can be fine-tuned.
|
||||||
|
|
||||||
|
The blog can also be opened and directly fine-tuned in a google colab notebook.
|
||||||
|
In this section, we will explain how to fine-tune the model on a local machine.
|
||||||
|
|
||||||
|
1. **Log in**
|
||||||
|
|
||||||
|
To begin with, you should check that you are correctly logged in and that you have `git-lfs` installed so that your fine-tuned model can automatically be uploaded.
|
||||||
|
|
||||||
|
Run:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
huggingface-cli login
|
||||||
|
```
|
||||||
|
|
||||||
|
to login. It is recommended to login with your access token that can be found under your hugging face profile (icon in the top right corner on [hf.co](http://hf.co/), then Settings -> Access Tokens -> User Access Tokens -> New Token (if haven't generated one already)
|
||||||
|
|
||||||
|
You can then copy-paste this token to log in locally.
|
||||||
|
|
||||||
|
2. **Create your model repository**
|
||||||
|
|
||||||
|
First, let's make sure that `git-lfs` is correctly installed. To so, simply run:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
git-lfs -v
|
||||||
|
```
|
||||||
|
|
||||||
|
The output should show something like `git-lfs/2.13.2 (GitHub; linux amd64; go 1.15.4)`. If your console states that the `git-lfs` command was not found, please make
|
||||||
|
sure to install it [here](https://git-lfs.github.com/) or simply via:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
sudo apt-get install git-lfs
|
||||||
|
```
|
||||||
|
|
||||||
|
Now you can create your model repository which will contain all relevant files to
|
||||||
|
reproduce your training. You can either directly create the model repository on the
|
||||||
|
Hub (Settings -> New Model) or via the CLI. Here we choose to use the CLI instead.
|
||||||
|
|
||||||
|
Assuming that we want to call our model repository *xls-r-ab-test*, we can run the
|
||||||
|
following command:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
huggingface-cli repo create xls-r-ab-test
|
||||||
|
```
|
||||||
|
|
||||||
|
You can now see the model on the Hub, *e.g.* under https://huggingface.co/hf-test/xls-r-ab-test .
|
||||||
|
|
||||||
|
Let's clone the repository so that we can define our training script inside.
|
||||||
|
|
||||||
|
```bash
|
||||||
|
git lfs install
|
||||||
|
git clone https://huggingface.co/hf-test/xls-r-ab-test
|
||||||
|
```
|
||||||
|
|
||||||
|
3. **Add your training script and `run`-command to the repository**
|
||||||
|
|
||||||
|
We encourage participants to add all relevant files for training directly to the
|
||||||
|
directory so that everything is fully reproducible.
|
||||||
|
|
||||||
|
Let's first copy-paste the official training script from our clone
|
||||||
|
of `transformers` to our just created directory:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
cp ~/transformers/examples/pytorch/speech-recognition/run_speech_recognition_ctc.py ./
|
||||||
|
```
|
||||||
|
|
||||||
|
Next, we'll create a bash file to define the hyper-parameters and configurations
|
||||||
|
for training. More detailed information on different settings (single-GPU vs. multi-GPU) can be found [here](https://github.com/huggingface/transformers/tree/master/examples/pytorch/speech-recognition#connectionist-temporal-classification).
|
||||||
|
|
||||||
|
For demonstration purposes, we will use a dummy XLS-R model `model_name_or_path="hf-test/xls-r-dummy"` on the very low-resource language of "Abkhaz" of [Common Voice 7](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0): `dataset_config_name="ab"` for just a single epoch.
|
||||||
|
|
||||||
|
Before starting to train, let's make sure we have installed all the required libraries. You might want to run:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
pip install -r ~/transformers/examples/pytorch/speech-recognition/requirements.txt
|
||||||
|
```
|
||||||
|
|
||||||
|
Alright, finally we can define the training script. We'll simply use some
|
||||||
|
dummy hyper-parameters and configurations for demonstration purposes.
|
||||||
|
|
||||||
|
Note that we add the flag `--use_auth_token` so that datasets requiring access,
|
||||||
|
such as [Common Voice 7](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0) can be downloaded. In addition, we add the `--push_to_hub` flag to make use of the
|
||||||
|
[Trainers `push_to-hub` functionality](https://huggingface.co/docs/transformers/master/en/main_classes/trainer#transformers.Trainer.push_to_hub) so that your model will be automatically uploaded to the Hub.
|
||||||
|
|
||||||
|
Let's copy the following code snippet in a file called `run.sh`
|
||||||
|
|
||||||
|
```bash
|
||||||
|
echo '''python run_speech_recognition_ctc.py \
|
||||||
|
--dataset_name="mozilla-foundation/common_voice_7_0" \
|
||||||
|
--model_name_or_path="hf-test/xls-r-dummy" \
|
||||||
|
--dataset_config_name="ab" \
|
||||||
|
--output_dir="./" \
|
||||||
|
--overwrite_output_dir \
|
||||||
|
--max_steps="10" \
|
||||||
|
--per_device_train_batch_size="2" \
|
||||||
|
--learning_rate="3e-4" \
|
||||||
|
--save_total_limit="1" \
|
||||||
|
--evaluation_strategy="steps" \
|
||||||
|
--text_column_name="sentence" \
|
||||||
|
--length_column_name="input_length" \
|
||||||
|
--save_steps="5" \
|
||||||
|
--layerdrop="0.0" \
|
||||||
|
--freeze_feature_encoder \
|
||||||
|
--gradient_checkpointing \
|
||||||
|
--fp16 \
|
||||||
|
--group_by_length \
|
||||||
|
--push_to_hub \
|
||||||
|
--use_auth_token \
|
||||||
|
--do_train --do_eval''' > run.sh
|
||||||
|
```
|
||||||
|
|
||||||
|
4. **Start training**
|
||||||
|
|
||||||
|
Now all that is left to do is to start training the model by executing the
|
||||||
|
run file.
|
||||||
|
|
||||||
|
```bash
|
||||||
|
bash run.sh
|
||||||
|
```
|
||||||
|
|
||||||
|
The training should not take more than a couple of minutes.
|
||||||
|
During the training intermediate saved checkpoints are automatically uploaded to
|
||||||
|
your model repository as can be seen [on this commit](https://huggingface.co/hf-test/xls-r-ab-test/commit/0eb19a0fca4d7d163997b59663d98cd856022aa6) .
|
||||||
|
|
||||||
|
At the end of the training, the [Trainer](https://huggingface.co/docs/transformers/master/en/main_classes/trainer) automatically creates a nice model card and all
|
||||||
|
relevant files are uploaded.
|
||||||
|
|
||||||
|
5. **Tips for real model training**
|
||||||
|
|
||||||
|
The above steps illustrate how a model can technically be fine-tuned.
|
||||||
|
However as you can see on the model card [hf-test/xls-r-ab-test](https://huggingface.co/hf-test/xls-r-ab-test), our demonstration has a very poor performance which is
|
||||||
|
not surprising given that we trained for just 10 steps on a randomly initialized
|
||||||
|
model.
|
||||||
|
|
||||||
|
For real model training, it is recommended to use one of the actual pre-trained XLS-R models:
|
||||||
|
|
||||||
|
- [300M parameters version](https://huggingface.co/facebook/wav2vec2-xls-r-300m)
|
||||||
|
- [1B parameters version](https://huggingface.co/facebook/wav2vec2-xls-r-1b)
|
||||||
|
- [2B parameters version](https://huggingface.co/facebook/wav2vec2-xls-r-2b)
|
||||||
|
|
||||||
|
Also, the hyper-parameters should be carefully chosen depending on the dataset.
|
||||||
|
As an example, we will fine-tune the 300M parameters model on Swedish on a single
|
||||||
|
TITAN RTX 24GB GPU.
|
||||||
|
|
||||||
|
The model will be called `"xls-r-300m-sv"`.
|
||||||
|
Following the above steps we first create the model:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
huggingface-cli repo create xls-r-300m-sv
|
||||||
|
```
|
||||||
|
|
||||||
|
, clone it locally (assuming the `<username>` is `hf-test`)
|
||||||
|
|
||||||
|
```bash
|
||||||
|
git clone hf-test/xls-r-300m-sv
|
||||||
|
```
|
||||||
|
|
||||||
|
, and, define the following hyperparameters for training
|
||||||
|
|
||||||
|
```bash
|
||||||
|
echo '''python run_speech_recognition_ctc.py \
|
||||||
|
--dataset_name="mozilla-foundation/common_voice_7_0" \
|
||||||
|
--model_name_or_path="facebook/wav2vec2-xls-r-300m" \
|
||||||
|
--dataset_config_name="sv-SE" \
|
||||||
|
--output_dir="./" \
|
||||||
|
--overwrite_output_dir \
|
||||||
|
--num_train_epochs="50" \
|
||||||
|
--per_device_train_batch_size="8" \
|
||||||
|
--per_device_eval_batch_size="8" \
|
||||||
|
--gradient_accumulation_steps="4" \
|
||||||
|
--learning_rate="7.5e-5" \
|
||||||
|
--warmup_steps="2000" \
|
||||||
|
--length_column_name="input_length" \
|
||||||
|
--evaluation_strategy="steps" \
|
||||||
|
--text_column_name="sentence" \
|
||||||
|
--chars_to_ignore , ? . ! \- \; \: \" “ % ‘ ” <20> — ’ … – \
|
||||||
|
--save_steps="500" \
|
||||||
|
--eval_steps="500" \
|
||||||
|
--logging_steps="100" \
|
||||||
|
--layerdrop="0.0" \
|
||||||
|
--activation_dropout="0.1" \
|
||||||
|
--save_total_limit="3" \
|
||||||
|
--freeze_feature_encoder \
|
||||||
|
--feat_proj_dropout="0.0" \
|
||||||
|
--mask_time_prob="0.75" \
|
||||||
|
--mask_time_length="10" \
|
||||||
|
--mask_feature_prob="0.25" \
|
||||||
|
--mask_feature_length="64" \
|
||||||
|
--gradient_checkpointing \
|
||||||
|
--use_auth_token \
|
||||||
|
--fp16 \
|
||||||
|
--group_by_length \
|
||||||
|
--do_train --do_eval \
|
||||||
|
--push_to_hub''' > run.sh
|
||||||
|
```
|
||||||
|
|
||||||
|
The training takes *ca.* 7 hours and yields a reasonable test word
|
||||||
|
error rate of 27% as can be seen on the automatically generated [model card](https://huggingface.co/hf-test/xls-r-300m-sv).
|
||||||
|
|
||||||
|
The above-chosen hyperparameters probably work quite well on a range of different
|
||||||
|
datasets and languages but are by no means optimal. It is up to you to find a good set of
|
||||||
|
hyperparameters.
|
||||||
|
|
||||||
|
|
||||||
|
## How to finetune with OVH cloud
|
||||||
|
|
||||||
|
For a more detailed guide on setting up OVHcloud please watch this video: TODO
|
||||||
|
|
||||||
|
### Creating an OVHCloud account
|
||||||
|
*TIP*: If you haven't created a project on OVHcloud yet, make sure you've received your GPU voucher code *beforehand*,
|
||||||
|
so that you can skip entering the credit card information.
|
||||||
|
1. If you're a US citizen, create an account via [OVHcloud.CA](https://ovhcloud.ca/).
|
||||||
|
If you're from anywhere else in the world, create an account via [OVHcloud.COM](https://ovhcloud.com/).
|
||||||
|
2. Once logged in, click `Public Cloud` from the top menu and then click `Create your first OVH Public Cloud project`.
|
||||||
|
Then enter a project name (e.g. "huggingface"), enter your voucher code, and click `Continue` -> `Create my project`.
|
||||||
|
*Note: if you see a request for credit card details during the last step, and you can't skip it, then your voucher code
|
||||||
|
is invalid. Please report it to the [#ovh-support](https://discord.gg/p4qqDV3M) channel on Discord.*
|
||||||
|
|
||||||
|
### Setting up an AI notebook
|
||||||
|
1. Go to the `Public Cloud` page and select `Project Management` -> `Users & Roles` from the menu on the left.
|
||||||
|
2. Click `+ Add user`. Write a user description (e.g. `AI Trainer`), and select an `AI Training Operator` user role.
|
||||||
|
Click `Confirm`.
|
||||||
|
3. Write down the *username* and *password* (at the top of the screen) somewhere. They will be needed during step 7.
|
||||||
|
4. Select `AI & Machine Learning` -> `AI Training` from the menu on the left.
|
||||||
|
Click `+ Launch a new job` on the AI Training page.
|
||||||
|
5. On the `Launch a new job` page:
|
||||||
|
* In `1. Choose a region` select a region closest to you.
|
||||||
|
* In `2. Enter the Docker image` select `Custom image` -> `baaastijn/ovh_huggingface`.
|
||||||
|
* You can skip steps `3.` and `4.` if you will be using the Hugging Face Hub to store the models after training.
|
||||||
|
* In `5. Configure your job` select **1** `GPU`.
|
||||||
|
* Validate the info and Create the job.
|
||||||
|
6. On the `AI Training Jobs` screen wait until the job's status changes from `Pending` to `Running`.
|
||||||
|
7. Click `HTTP Access` and log in with the AI training user you've created earlier.
|
||||||
|
Once logged in, you can close the page and click `HTTP Access` to launch a JupyterLab notebook.
|
||||||
|
8. Awesome, now you have a free GPU-enabled Jupyter instance!
|
||||||
|
|
||||||
|
**Note**: If you're an experienced Docker user, feel free to create a custom docker image with all of the needed packages
|
||||||
|
like the one in step 5. The Dockerfile for it is available here:
|
||||||
|
[baaastijn/Dockerimages](https://github.com/baaastijn/Dockerimages/tree/main/Hugginface_challenge_speech).
|
||||||
|
Once you've built your image, push it to https://hub.docker.com/ and select it during the OVHcloud job creation.
|
||||||
|
|
||||||
|
## How to combine n-gram with acoustic model
|
||||||
|
|
||||||
|
Having trained a speech recognition model with CTC as shown in the section above,
|
||||||
|
one can further improve the model's performance by adding an **n-gram language model**
|
||||||
|
to the decoding process of the model. By doing so, we are replacing the naive greedy decoding
|
||||||
|
with **n-gram-boosted** beam search decoding.
|
||||||
|
|
||||||
|
N-gram language models can be built on CPU in just a few minutes. *N-gram-boosted* beam search decoding noticeably slows down the
|
||||||
|
inference time, but also yields significant word error rates improvements - usually between 10-40 %.
|
||||||
|
|
||||||
|
You can find an in-detail blog post on how to build an *n-gram* [here](https://huggingface.co/blog/wav2vec2-with-ngram).
|
||||||
|
The blog post can be opened in a google colab and by adapting three lines of the example for your use case, one can directly
|
||||||
|
create an *n-gram* in the google colab.
|
||||||
|
The blog post gives in-detail instructions on how to build an n-gram and how to add it to your trained speech recognition model.
|
||||||
|
|
||||||
|
- why one should add an *n-gram* to her/his speech recognition system,
|
||||||
|
- how to build an *n-gram*, and,
|
||||||
|
- how to add the built *n-gram* the speech recognition system for seamless decoding
|
||||||
|
|
||||||
|
Our previously trained model - [xls-r-300m-sv](https://huggingface.co/hf-test/xls-r-300m-sv) - enjoys a 30% word error rate reduction after
|
||||||
|
having added an n-gram. As shown in the example of the blog post, we strongly advise participants to upload all files required for combining
|
||||||
|
the *n-gram* with a trained speech recognition model directly into the same model repository.
|
||||||
|
|
||||||
|
## Evaluation
|
||||||
|
|
||||||
|
Finally, we have arrived at the most fun part of the challenge - sitting back and
|
||||||
|
watching the model transcribe audio. If possible, every participant should evaluate
|
||||||
|
the speech recognition system on the test set of Common Voice 7 and
|
||||||
|
ideally also on the real-world audio data (if available).
|
||||||
|
For languages that have neither a Common Voice evaluation dataset nor a real world
|
||||||
|
evaluation dataset, please contact the organizers on Discord so that we can work
|
||||||
|
together to find some evaluation data.
|
||||||
|
|
||||||
|
As a first step, one should copy the official `eval.py` script to her/his model
|
||||||
|
repository. Let's use our previously trained [xls-r-300m-sv](https://huggingface.co/hf-test/xls-r-300m-sv) again as an example.
|
||||||
|
|
||||||
|
Assuming that we have a clone of the model's repo under `~/xls-r-300m-sv`, we can
|
||||||
|
copy the `eval.py` script to the repo.
|
||||||
|
|
||||||
|
```bash
|
||||||
|
cp ~/transformers/examples/research_projects/robust-speech-event/eval.py ~/xls-r-300m-sv
|
||||||
|
```
|
||||||
|
|
||||||
|
Next, we should adapt `eval.py` so that it fits our evaluation data. Here it is
|
||||||
|
important to keep the `eval.py` file in the following format:
|
||||||
|
|
||||||
|
- 1. The following input arguments should not be changed and keep their original functionality/meaning (being to load the model and dataset): `"--model_id"`, `"--dataset"`, `"--config"`, `"--split"`. We recommend to not change any of the code written under `if __name__ == "__main__":`.
|
||||||
|
- 2. The function `def log_results(result: Dataset, args: Dict[str, str])` should also not be changed. The function expects the above names attached to the `args` object as well as a `datasets.Dataset` object, called `result` which includes all predictions and target transcriptions under the names `"predictions"` and `"targets"` respectively.
|
||||||
|
- 3. All other code can be changed and adapted. Participants are especially invited to change the `def normalize_text(text: str) -> str:` function as this might be a very language and model-training specific function.
|
||||||
|
- 4. **Important**: It is not allowed to "cheat" in any way when in comes to pre-and postprocessing. In short, "cheating" refers to any of the following:
|
||||||
|
- a. Somehow giving the model access to the target transcriptions to improve performance. The model is not allowed to use the target transcriptions to generate its predictions.
|
||||||
|
- b. Pre-processing the target transcriptions in a way that makes the target transcriptions lose their original meaning. This corresponds to what has already been said in [Data and Preprocessing](#data-and-preprocessing) and is somewhat of a grey zone. It means that one should not remove characters that would make a word to lose its meaning. E.g., it is not allowed to replace all `e` in English with `i` and simply make the model learn that `e` and `i` are the same letter for a better word error rate. This would destroy the meaning of words such as `fell -> fill`. However, it is totally fine to normalize (*e.g.* lowercase) all letters, remove punctuation. There can be a lot of language-specific exceptions and in case you are not sure whether your target transcription pre-processing is allowed, please ask on the Discord channel.
|
||||||
|
|
||||||
|
Uff, that was a lot of text describing how to make sure your `eval.py` script
|
||||||
|
is in the correct format. If you have any questions, please ask openly in Discord.
|
||||||
|
|
||||||
|
Great, now that we have adapted the `eval.py` script, we can lean back and run the
|
||||||
|
evaluation.
|
||||||
|
First, one should evaluate the model on Common Voice 7's test data. This might
|
||||||
|
already have been done for your acoustic model during training but in case you
|
||||||
|
added an *n-gram* language model after having fine-tuned the acoustic model, you
|
||||||
|
should now see a nice improvement.
|
||||||
|
|
||||||
|
The command to evaluate our test model [xls-r-300m-sv](https://huggingface.co/hf-test/xls-r-300m-sv) on Common Voice 7's test data is the following:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
cd xls-r-300m-sv
|
||||||
|
./eval.py --model_id ./ --dataset mozilla-foundation/common_voice_7_0 --config sv-SE --split test --log_outputs
|
||||||
|
```
|
||||||
|
|
||||||
|
To log each of the model's predictions with the target transcriptions, you can just
|
||||||
|
add the `--log_outputs` flag.
|
||||||
|
|
||||||
|
Running this command should automatically create the file:
|
||||||
|
`mozilla-foundation_common_voice_7_0_sv-SE_test_eval_results.txt` that contains
|
||||||
|
both the word- and character error rate.
|
||||||
|
|
||||||
|
In a few days, we will give everybody access to some real-world audio data for as many languages as possible.
|
||||||
|
If your language has real-world audio data, it will most likely have audio input
|
||||||
|
of multiple minutes. 🤗Transformer's [ASR pipeline](https://huggingface.co/docs/transformers/master/en/main_classes/pipelines#transformers.AutomaticSpeechRecognitionPipeline) supports audio chunking out-of-the-box. You only need to specify
|
||||||
|
how song each audio chunk should be (`chunk_length_s`) and how much audio stride
|
||||||
|
(`stride_length_s`) each chunk should use.
|
||||||
|
For more information on the chunking works, please have a look at [this nice blog post](TODO: ).
|
||||||
|
|
||||||
|
In the case of `xls-r-300m-sv`, the following command can be run:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
cd xls-r-300m-sv
|
||||||
|
./eval.py --model_id hf-test/xls-r-300m-sv --dataset <to-be-announced> --config sv --split validation --chunk_length_s 5.0 --stride_length_s 1.0 --log_outputs
|
||||||
|
```
|
||||||
|
|
||||||
|
Great, now you should have successfully evaluated your model. Finally, there is one
|
||||||
|
**important** thing you should do so that your model is taken into account
|
||||||
|
for the final evaluation. You should add two tags to your model, one being `robust-speech-event`, one being the ISO code of your chosen language, *e.g.* `"sv"` for the
|
||||||
|
exemplary model we used above. You can find a list of all available languages and
|
||||||
|
their ISO code [here](https://huggingface.co/languages).
|
||||||
|
|
||||||
|
To add the tags, simply edit the README.md of your model repository and add
|
||||||
|
|
||||||
|
```
|
||||||
|
- "sv"
|
||||||
|
- "robust-speech-event"
|
||||||
|
```
|
||||||
|
|
||||||
|
under `tags:` as done [here](https://huggingface.co/hf-test/xls-r-300m-sv/commit/a495fd70c96bb7d019729be9273a265c2557345e).
|
||||||
|
|
||||||
|
To verify that you've added the tags correctly make sure that your model
|
||||||
|
appears when clicking on [this link](https://huggingface.co/models?other=robust-speech-event).
|
||||||
|
|
||||||
|
Great that's it! This should give you all the necessary information to evaluate
|
||||||
|
your model. For the final evaluation, we will verify each evaluation result to
|
||||||
|
determine the final score and thereby the winning models for each language.
|
||||||
|
|
||||||
|
The final score is calculated as follows:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
FINAL_SCORE = 1/3 * WER_Common_Voice_7_test + 1/3 * WER_REAL_AUDIO_DEV + 1/3 * WER_REAL_AUDIO_TEST
|
||||||
|
```
|
||||||
|
|
||||||
|
The dataset `WER_REAL_AUDIO_TEST` is hidden and will only be published
|
||||||
|
at the end of the robust speech challenge.
|
||||||
|
|
||||||
|
If there is no real audio data for your language the final score will be
|
||||||
|
computed solely based on the Common Voice 7 test dataset. If there is also
|
||||||
|
no Common Voice 7 test dataset for your language, we will see together how to
|
||||||
|
score your model - if this is the case, please don't be discouraged. We are
|
||||||
|
especially excited about speech recognition systems of such low-resource
|
||||||
|
languages and will make sure that we'll decide on a good approach to evaluating
|
||||||
|
your model.
|
||||||
|
|
||||||
|
## Prizes
|
||||||
|
|
||||||
|
TODO(Patrick, Omar, ...)
|
||||||
|
|
||||||
|
## Communication and Problems
|
||||||
|
|
||||||
|
If you encounter any problems or have any questions, you should use one of the following platforms
|
||||||
|
depending on your type of problem. Hugging Face is an "open-source-first" organization meaning
|
||||||
|
that we'll try to solve all problems in the most public and most transparent way possible so that everybody
|
||||||
|
in the community profits.
|
||||||
|
|
||||||
|
The following table summarizes what platform to use for which problem.
|
||||||
|
|
||||||
|
- Problem/question/bug with the 🤗 Datasets library that you think is a general problem that also impacts other people, please open an [Issues on Datasets](https://github.com/huggingface/datasets/issues/new?assignees=&labels=bug&template=bug-report.md&title=) and ping @anton-l and @patrickvonplaten.
|
||||||
|
- Problem/question/bug with the 🤗 Transformers library that you think is a general problem that also impacts other people, please open an [Issues on Transformers](https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title=) and ping @anton-l and @patrickvonplaten.
|
||||||
|
- Problem/question with a modified, customized training script that is less likely to impact other people, please post your problem/question [on the forum](https://discuss.huggingface.co/) and ping @anton-l and @patrickvonplaten.
|
||||||
|
- Questions regarding access to the OVHcloud GPU, please ask in the Discord channel **#ovh-support**.
|
||||||
|
- Other questions regarding the event, rules of the event, or if you are not sure where to post your question, please ask in the Discord channel **#sprint-discussions**.
|
||||||
|
|
||||||
|
## Talks
|
||||||
|
|
||||||
|
We are very excited to be hosting 2 days of talks from Kensho-Technologies, Mozilla's Common Voice, Kensho-Technologies, Hugging Face.
|
||||||
|
|
||||||
|
### Thursday, January 20th
|
||||||
|
|
||||||
|
- [Watch the talks on YouTube](TODO)
|
||||||
|
- [Chat history](TODO)
|
||||||
|
|
||||||
|
Speaker | Topic | Time | Video |
|
||||||
|
|-------------|---------------------------------|------------------------|------------------------|
|
||||||
|
| Patrick von Platen, Hugging Face | TODO | ??? UTC | [](TODO)
|
||||||
|
| Raymond Grossman and Jeremy Lopez, Kensho-Technologies | Pyctcdecode & Speech2text decoding | 5h30pm - 6h00pm UTC | [](TODO)
|
||||||
|
|
||||||
|
### Friday, January 21th
|
||||||
|
|
||||||
|
- [Watch the talks on YouTube](TODO)
|
||||||
|
- [Chat history](TODO)
|
||||||
|
|
||||||
|
Speaker | Topic | Time | Video |
|
||||||
|
|-------------|---------------------------------|------------------------|------------------------|
|
||||||
|
| Gabriel Habayeb, Mozilla Common Voice | TODO | 4h30pm - 5h00pm UTC | [](TODO)
|
||||||
|
| Changhan Wang, Meta AI Research | XLS-R: Large-Scale Cross-lingual Speech Representation Learning on 128 Languages | 5h30pm - 6h00pm UTC | [](TODO)
|
||||||
|
|
||||||
|
### Talks & Speakers
|
||||||
|
|
||||||
|
#### Patrick von Platen, Research Engineer, Hugging Face
|
||||||
|
- Talk: Introduction to Robust Speech Challenge
|
||||||
|
- Abstract: In this talk, Patrick outlines the Robust Speech Challenge and gives tips and tricks on how to train and evaluate speech recognition systems with 🤗 Transformers and 🤗 Datasets, and PyTorch.
|
||||||
|
- Speaker info: Patrick von Platen is a research engineer at Hugging Face and one of the core maintainers of the popular Transformers library. He specializes in speech recognition, encoder-decoder models, and long-range sequence modeling. Before joining Hugging Face, Patrick researched speech recognition at Uber AI, Cambridge University, and RWTH Aachen University.
|
||||||
|
|
||||||
|
#### Raymond Grossman, Jeremy Lopez, Machine Learning Engineer, Kensho Technologies
|
||||||
|
- Talk: PyCTCDecode & Speech2text decoding
|
||||||
|
- Abstract: PyCTCDecode is a fast and feature-rich CTC beam search decoder for speech recognition written in Python, providing n-gram (kenlm) language model support similar to PaddlePaddle's decoder, but incorporating many new features such as byte pair encoding and real-time decoding to support models like Nvidia's Conformer-CTC or Facebook's Wav2Vec2.
|
||||||
|
- Speaker info :
|
||||||
|
- Raymond works as a machine learning engineer at Kensho Technologies, specializing in speech and natural language domains. Before coming to Kensho, he studied mathematics at Princeton and was an avid Kaggler under the moniker @ToTrainThemIsMyCause.
|
||||||
|
- Jeremy is a machine learning engineer at Kensho Technologies and has worked on a variety of different topics including search and speech recognition. Before working at Kensho, he earned a PhD in experimental particle physics at MIT and continued doing physics research as a postdoc at the University of Colorado Boulder.
|
||||||
|
|
||||||
|
#### Gabriel Habayeb, Data Engineer, Common Voice @ Mozilla
|
||||||
|
- Talk: Common Voice
|
||||||
|
- Abstract:
|
||||||
|
- Speaker info:
|
||||||
|
|
||||||
|
#### Changhan Wang, Main author of XLS-R and Research Engineer, Meta AI Research
|
||||||
|
- Talk: XLS-R: Large-Scale Cross-lingual Speech Representation Learning on 128 Languages
|
||||||
|
- Abstract: In this talk, Changhan will present XLS-R, a large-scale model for cross-lingual speech representation learning based on wav2vec 2.0. XLS-R has up to 2B parameters and was trained on nearly half a million hours of publicly available speech audio in 128 languages, an order of magnitude more public data than the largest known prior work. On the CoVoST-2 speech translation benchmark, XLS-R improves the previous state of the art by an average of 7.4 BLEU over 21 translation directions into English. For speech recognition, XLS-R improves over the best known prior work on BABEL, MLS, CommonVoice as well as VoxPopuli, lowering error rates by 14-34% relative on average. XLS-R also sets a new state of the art on VoxLingua107 language identification. The XLS-R team hopes to work together with the open-source community to improve speech processing tasks for many more languages of the world.
|
||||||
|
|
||||||
|
## General Tips and Tricks
|
||||||
|
|
||||||
|
- Memory efficient training:
|
||||||
|
|
||||||
|
In case, you are getting out-of-memory errors on your GPU, we recommend to use
|
||||||
|
[bitsandbytes](https://github.com/facebookresearch/bitsandbytes) to replace the
|
||||||
|
native memory-intensive Adam optimizer with the one of `bitsandbytes`. You
|
||||||
|
can simply run the script `./run_speech_recognition_ctc_bnb.py` provided in this
|
||||||
|
folder that makes use of `bitsandbytes` instead of the official one.
|
||||||
|
|
||||||
|
- Dataset streaming
|
||||||
|
|
||||||
|
TODO(Patrick)
|
||||||
128
examples/research_projects/robust-speech-event/eval.py
Executable file
128
examples/research_projects/robust-speech-event/eval.py
Executable file
@@ -0,0 +1,128 @@
|
|||||||
|
#!/usr/bin/env python3
|
||||||
|
import argparse
|
||||||
|
import re
|
||||||
|
from typing import Dict
|
||||||
|
|
||||||
|
from datasets import Audio, Dataset, load_dataset, load_metric
|
||||||
|
|
||||||
|
from transformers import AutoFeatureExtractor, pipeline
|
||||||
|
|
||||||
|
|
||||||
|
def log_results(result: Dataset, args: Dict[str, str]):
|
||||||
|
"""DO NOT CHANGE. This function computes and logs the result metrics."""
|
||||||
|
|
||||||
|
log_outputs = args.log_outputs
|
||||||
|
dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
|
||||||
|
|
||||||
|
# load metric
|
||||||
|
wer = load_metric("wer")
|
||||||
|
cer = load_metric("cer")
|
||||||
|
|
||||||
|
# compute metrics
|
||||||
|
wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
|
||||||
|
cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
|
||||||
|
|
||||||
|
# print & log results
|
||||||
|
result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
|
||||||
|
print(result_str)
|
||||||
|
|
||||||
|
with open(f"{dataset_id}_eval_results.txt", "w") as f:
|
||||||
|
f.write(result_str)
|
||||||
|
|
||||||
|
# log all results in text file. Possibly interesting for analysis
|
||||||
|
if log_outputs is not None:
|
||||||
|
pred_file = f"log_{dataset_id}_predictions.txt"
|
||||||
|
target_file = f"log_{dataset_id}_targets.txt"
|
||||||
|
|
||||||
|
with open(pred_file, "w") as p, open(target_file, "w") as t:
|
||||||
|
|
||||||
|
# mapping function to write output
|
||||||
|
def write_to_file(batch, i):
|
||||||
|
p.write(f"{i}" + "\n")
|
||||||
|
p.write(batch["prediction"] + "\n")
|
||||||
|
t.write(f"{i}" + "\n")
|
||||||
|
t.write(batch["target"] + "\n")
|
||||||
|
|
||||||
|
result.map(write_to_file, with_indices=True)
|
||||||
|
|
||||||
|
|
||||||
|
def normalize_text(text: str) -> str:
|
||||||
|
"""DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
|
||||||
|
|
||||||
|
chars_to_ignore_regex = '[,?.!\-\;\:"“%‘”<EFBFBD>—’…–]' # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
|
||||||
|
|
||||||
|
text = re.sub(chars_to_ignore_regex, "", text.lower())
|
||||||
|
|
||||||
|
# In addition, we can normalize the target text, e.g. removing new lines characters etc...
|
||||||
|
# note that order is important here!
|
||||||
|
token_sequences_to_ignore = ["\n\n", "\n", " ", " "]
|
||||||
|
|
||||||
|
for t in token_sequences_to_ignore:
|
||||||
|
text = " ".join(text.split(t))
|
||||||
|
|
||||||
|
return text
|
||||||
|
|
||||||
|
|
||||||
|
def main(args):
|
||||||
|
# load dataset
|
||||||
|
dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
|
||||||
|
|
||||||
|
# for testing: only process the first two examples as a test
|
||||||
|
dataset = dataset.select(range(10))
|
||||||
|
|
||||||
|
# load processor
|
||||||
|
feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
|
||||||
|
sampling_rate = feature_extractor.sampling_rate
|
||||||
|
|
||||||
|
# resample audio
|
||||||
|
dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
|
||||||
|
|
||||||
|
# load eval pipeline
|
||||||
|
asr = pipeline("automatic-speech-recognition", model=args.model_id)
|
||||||
|
|
||||||
|
# map function to decode audio
|
||||||
|
def map_to_pred(batch):
|
||||||
|
prediction = asr(
|
||||||
|
batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
|
||||||
|
)
|
||||||
|
|
||||||
|
batch["prediction"] = prediction["text"]
|
||||||
|
batch["target"] = normalize_text(batch["sentence"])
|
||||||
|
return batch
|
||||||
|
|
||||||
|
# run inference on all examples
|
||||||
|
result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
|
||||||
|
|
||||||
|
# compute and log_results
|
||||||
|
# do not change function below
|
||||||
|
log_results(result, args)
|
||||||
|
|
||||||
|
|
||||||
|
if __name__ == "__main__":
|
||||||
|
parser = argparse.ArgumentParser()
|
||||||
|
|
||||||
|
parser.add_argument(
|
||||||
|
"--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
|
||||||
|
)
|
||||||
|
parser.add_argument(
|
||||||
|
"--dataset",
|
||||||
|
type=str,
|
||||||
|
required=True,
|
||||||
|
help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
|
||||||
|
)
|
||||||
|
parser.add_argument(
|
||||||
|
"--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
|
||||||
|
)
|
||||||
|
parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
|
||||||
|
parser.add_argument(
|
||||||
|
"--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
|
||||||
|
)
|
||||||
|
parser.add_argument(
|
||||||
|
"--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
|
||||||
|
)
|
||||||
|
parser.add_argument(
|
||||||
|
"--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
|
||||||
|
)
|
||||||
|
args = parser.parse_args()
|
||||||
|
|
||||||
|
main(args)
|
||||||
760
examples/research_projects/robust-speech-event/run_speech_recognition_ctc_bnb.py
Executable file
760
examples/research_projects/robust-speech-event/run_speech_recognition_ctc_bnb.py
Executable file
@@ -0,0 +1,760 @@
|
|||||||
|
#!/usr/bin/env python
|
||||||
|
# coding=utf-8
|
||||||
|
# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
|
||||||
|
#
|
||||||
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||||
|
# you may not use this file except in compliance with the License.
|
||||||
|
# You may obtain a copy of the License at
|
||||||
|
#
|
||||||
|
# http://www.apache.org/licenses/LICENSE-2.0
|
||||||
|
#
|
||||||
|
# Unless required by applicable law or agreed to in writing, software
|
||||||
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||||
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||||
|
# See the License for the specific language governing permissions and
|
||||||
|
|
||||||
|
""" Fine-tuning a 🤗 Transformers CTC model for automatic speech recognition"""
|
||||||
|
|
||||||
|
import functools
|
||||||
|
import json
|
||||||
|
import logging
|
||||||
|
import os
|
||||||
|
import re
|
||||||
|
import sys
|
||||||
|
import warnings
|
||||||
|
from dataclasses import dataclass, field
|
||||||
|
from typing import Dict, List, Optional, Union
|
||||||
|
|
||||||
|
import datasets
|
||||||
|
import numpy as np
|
||||||
|
import torch
|
||||||
|
from datasets import DatasetDict, load_dataset, load_metric
|
||||||
|
|
||||||
|
import bitsandbytes as bnb
|
||||||
|
import transformers
|
||||||
|
from transformers import (
|
||||||
|
AutoConfig,
|
||||||
|
AutoFeatureExtractor,
|
||||||
|
AutoModelForCTC,
|
||||||
|
AutoProcessor,
|
||||||
|
AutoTokenizer,
|
||||||
|
HfArgumentParser,
|
||||||
|
Trainer,
|
||||||
|
TrainingArguments,
|
||||||
|
Wav2Vec2Processor,
|
||||||
|
set_seed,
|
||||||
|
)
|
||||||
|
from transformers.trainer_pt_utils import get_parameter_names
|
||||||
|
from transformers.trainer_utils import get_last_checkpoint, is_main_process
|
||||||
|
from transformers.utils import check_min_version
|
||||||
|
from transformers.utils.versions import require_version
|
||||||
|
|
||||||
|
|
||||||
|
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
|
||||||
|
check_min_version("4.16.0.dev0")
|
||||||
|
|
||||||
|
require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
|
||||||
|
|
||||||
|
|
||||||
|
logger = logging.getLogger(__name__)
|
||||||
|
|
||||||
|
|
||||||
|
def list_field(default=None, metadata=None):
|
||||||
|
return field(default_factory=lambda: default, metadata=metadata)
|
||||||
|
|
||||||
|
|
||||||
|
@dataclass
|
||||||
|
class ModelArguments:
|
||||||
|
"""
|
||||||
|
Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
|
||||||
|
"""
|
||||||
|
|
||||||
|
model_name_or_path: str = field(
|
||||||
|
metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
|
||||||
|
)
|
||||||
|
tokenizer_name_or_path: Optional[str] = field(
|
||||||
|
default=None,
|
||||||
|
metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
|
||||||
|
)
|
||||||
|
cache_dir: Optional[str] = field(
|
||||||
|
default=None,
|
||||||
|
metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
|
||||||
|
)
|
||||||
|
freeze_feature_encoder: bool = field(
|
||||||
|
default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
|
||||||
|
)
|
||||||
|
attention_dropout: float = field(
|
||||||
|
default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
|
||||||
|
)
|
||||||
|
activation_dropout: float = field(
|
||||||
|
default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
|
||||||
|
)
|
||||||
|
feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
|
||||||
|
hidden_dropout: float = field(
|
||||||
|
default=0.0,
|
||||||
|
metadata={
|
||||||
|
"help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
|
||||||
|
},
|
||||||
|
)
|
||||||
|
final_dropout: float = field(
|
||||||
|
default=0.0,
|
||||||
|
metadata={"help": "The dropout probability for the final projection layer."},
|
||||||
|
)
|
||||||
|
mask_time_prob: float = field(
|
||||||
|
default=0.05,
|
||||||
|
metadata={
|
||||||
|
"help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
|
||||||
|
"span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
|
||||||
|
"vectors will be masked along the time axis."
|
||||||
|
},
|
||||||
|
)
|
||||||
|
mask_time_length: int = field(
|
||||||
|
default=10,
|
||||||
|
metadata={"help": "Length of vector span to mask along the time axis."},
|
||||||
|
)
|
||||||
|
mask_feature_prob: float = field(
|
||||||
|
default=0.0,
|
||||||
|
metadata={
|
||||||
|
"help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
|
||||||
|
"span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
|
||||||
|
},
|
||||||
|
)
|
||||||
|
mask_feature_length: int = field(
|
||||||
|
default=10,
|
||||||
|
metadata={"help": "Length of vector span to mask along the feature axis."},
|
||||||
|
)
|
||||||
|
layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
|
||||||
|
ctc_loss_reduction: Optional[str] = field(
|
||||||
|
default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
|
||||||
|
)
|
||||||
|
|
||||||
|
|
||||||
|
@dataclass
|
||||||
|
class DataTrainingArguments:
|
||||||
|
"""
|
||||||
|
Arguments pertaining to what data we are going to input our model for training and eval.
|
||||||
|
|
||||||
|
Using `HfArgumentParser` we can turn this class
|
||||||
|
into argparse arguments to be able to specify them on
|
||||||
|
the command line.
|
||||||
|
"""
|
||||||
|
|
||||||
|
dataset_name: str = field(
|
||||||
|
metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
|
||||||
|
)
|
||||||
|
dataset_config_name: str = field(
|
||||||
|
default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
|
||||||
|
)
|
||||||
|
train_split_name: str = field(
|
||||||
|
default="train+validation",
|
||||||
|
metadata={
|
||||||
|
"help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
|
||||||
|
},
|
||||||
|
)
|
||||||
|
eval_split_name: str = field(
|
||||||
|
default="test",
|
||||||
|
metadata={
|
||||||
|
"help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
|
||||||
|
},
|
||||||
|
)
|
||||||
|
audio_column_name: str = field(
|
||||||
|
default="audio",
|
||||||
|
metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
|
||||||
|
)
|
||||||
|
text_column_name: str = field(
|
||||||
|
default="text",
|
||||||
|
metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
|
||||||
|
)
|
||||||
|
overwrite_cache: bool = field(
|
||||||
|
default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
|
||||||
|
)
|
||||||
|
preprocessing_num_workers: Optional[int] = field(
|
||||||
|
default=None,
|
||||||
|
metadata={"help": "The number of processes to use for the preprocessing."},
|
||||||
|
)
|
||||||
|
max_train_samples: Optional[int] = field(
|
||||||
|
default=None,
|
||||||
|
metadata={
|
||||||
|
"help": "For debugging purposes or quicker training, truncate the number of training examples to this "
|
||||||
|
"value if set."
|
||||||
|
},
|
||||||
|
)
|
||||||
|
max_eval_samples: Optional[int] = field(
|
||||||
|
default=None,
|
||||||
|
metadata={
|
||||||
|
"help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
|
||||||
|
"value if set."
|
||||||
|
},
|
||||||
|
)
|
||||||
|
chars_to_ignore: Optional[List[str]] = list_field(
|
||||||
|
default=None,
|
||||||
|
metadata={"help": "A list of characters to remove from the transcripts."},
|
||||||
|
)
|
||||||
|
eval_metrics: List[str] = list_field(
|
||||||
|
default=["wer"],
|
||||||
|
metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
|
||||||
|
)
|
||||||
|
max_duration_in_seconds: float = field(
|
||||||
|
default=20.0,
|
||||||
|
metadata={
|
||||||
|
"help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
|
||||||
|
},
|
||||||
|
)
|
||||||
|
min_duration_in_seconds: float = field(
|
||||||
|
default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
|
||||||
|
)
|
||||||
|
preprocessing_only: bool = field(
|
||||||
|
default=False,
|
||||||
|
metadata={
|
||||||
|
"help": "Whether to only do data preprocessing and skip training. "
|
||||||
|
"This is especially useful when data preprocessing errors out in distributed training due to timeout. "
|
||||||
|
"In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
|
||||||
|
"so that the cached datasets can consequently be loaded in distributed training"
|
||||||
|
},
|
||||||
|
)
|
||||||
|
use_auth_token: bool = field(
|
||||||
|
default=False,
|
||||||
|
metadata={
|
||||||
|
"help": "If :obj:`True`, will use the token generated when running"
|
||||||
|
":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
|
||||||
|
},
|
||||||
|
)
|
||||||
|
unk_token: str = field(
|
||||||
|
default="[UNK]",
|
||||||
|
metadata={"help": "The unk token for the tokenizer"},
|
||||||
|
)
|
||||||
|
pad_token: str = field(
|
||||||
|
default="[PAD]",
|
||||||
|
metadata={"help": "The padding token for the tokenizer"},
|
||||||
|
)
|
||||||
|
word_delimiter_token: str = field(
|
||||||
|
default="|",
|
||||||
|
metadata={"help": "The word delimiter token for the tokenizer"},
|
||||||
|
)
|
||||||
|
phoneme_language: Optional[str] = field(
|
||||||
|
default=None,
|
||||||
|
metadata={
|
||||||
|
"help": "The target language that should be used be"
|
||||||
|
" passed to the tokenizer for tokenization. Note that"
|
||||||
|
" this is only relevant if the model classifies the"
|
||||||
|
" input audio to a sequence of phoneme sequences."
|
||||||
|
},
|
||||||
|
)
|
||||||
|
|
||||||
|
|
||||||
|
@dataclass
|
||||||
|
class DataCollatorCTCWithPadding:
|
||||||
|
"""
|
||||||
|
Data collator that will dynamically pad the inputs received.
|
||||||
|
Args:
|
||||||
|
processor (:class:`~transformers.AutoProcessor`)
|
||||||
|
The processor used for proccessing the data.
|
||||||
|
padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
|
||||||
|
Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
|
||||||
|
among:
|
||||||
|
* :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
|
||||||
|
sequence if provided).
|
||||||
|
* :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
|
||||||
|
maximum acceptable input length for the model if that argument is not provided.
|
||||||
|
* :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
|
||||||
|
different lengths).
|
||||||
|
max_length (:obj:`int`, `optional`):
|
||||||
|
Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
|
||||||
|
max_length_labels (:obj:`int`, `optional`):
|
||||||
|
Maximum length of the ``labels`` returned list and optionally padding length (see above).
|
||||||
|
pad_to_multiple_of (:obj:`int`, `optional`):
|
||||||
|
If set will pad the sequence to a multiple of the provided value.
|
||||||
|
This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
|
||||||
|
7.5 (Volta).
|
||||||
|
"""
|
||||||
|
|
||||||
|
processor: AutoProcessor
|
||||||
|
padding: Union[bool, str] = "longest"
|
||||||
|
pad_to_multiple_of: Optional[int] = None
|
||||||
|
pad_to_multiple_of_labels: Optional[int] = None
|
||||||
|
|
||||||
|
def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
|
||||||
|
# split inputs and labels since they have to be of different lenghts and need
|
||||||
|
# different padding methods
|
||||||
|
input_features = [{"input_values": feature["input_values"]} for feature in features]
|
||||||
|
label_features = [{"input_ids": feature["labels"]} for feature in features]
|
||||||
|
|
||||||
|
batch = self.processor.pad(
|
||||||
|
input_features,
|
||||||
|
padding=self.padding,
|
||||||
|
pad_to_multiple_of=self.pad_to_multiple_of,
|
||||||
|
return_tensors="pt",
|
||||||
|
)
|
||||||
|
|
||||||
|
with self.processor.as_target_processor():
|
||||||
|
labels_batch = self.processor.pad(
|
||||||
|
label_features,
|
||||||
|
padding=self.padding,
|
||||||
|
pad_to_multiple_of=self.pad_to_multiple_of_labels,
|
||||||
|
return_tensors="pt",
|
||||||
|
)
|
||||||
|
|
||||||
|
# replace padding with -100 to ignore loss correctly
|
||||||
|
labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
|
||||||
|
|
||||||
|
batch["labels"] = labels
|
||||||
|
|
||||||
|
return batch
|
||||||
|
|
||||||
|
|
||||||
|
def create_vocabulary_from_data(
|
||||||
|
datasets: DatasetDict,
|
||||||
|
word_delimiter_token: Optional[str] = None,
|
||||||
|
unk_token: Optional[str] = None,
|
||||||
|
pad_token: Optional[str] = None,
|
||||||
|
):
|
||||||
|
# Given training and test labels create vocabulary
|
||||||
|
def extract_all_chars(batch):
|
||||||
|
all_text = " ".join(batch["target_text"])
|
||||||
|
vocab = list(set(all_text))
|
||||||
|
return {"vocab": [vocab], "all_text": [all_text]}
|
||||||
|
|
||||||
|
vocabs = datasets.map(
|
||||||
|
extract_all_chars,
|
||||||
|
batched=True,
|
||||||
|
batch_size=-1,
|
||||||
|
keep_in_memory=True,
|
||||||
|
remove_columns=datasets["train"].column_names,
|
||||||
|
)
|
||||||
|
|
||||||
|
# take union of all unique characters in each dataset
|
||||||
|
vocab_set = functools.reduce(
|
||||||
|
lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
|
||||||
|
)
|
||||||
|
|
||||||
|
vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
|
||||||
|
|
||||||
|
# replace white space with delimiter token
|
||||||
|
if word_delimiter_token is not None:
|
||||||
|
vocab_dict[word_delimiter_token] = vocab_dict[" "]
|
||||||
|
del vocab_dict[" "]
|
||||||
|
|
||||||
|
# add unk and pad token
|
||||||
|
if unk_token is not None:
|
||||||
|
vocab_dict[unk_token] = len(vocab_dict)
|
||||||
|
|
||||||
|
if pad_token is not None:
|
||||||
|
vocab_dict[pad_token] = len(vocab_dict)
|
||||||
|
|
||||||
|
return vocab_dict
|
||||||
|
|
||||||
|
|
||||||
|
def main():
|
||||||
|
# See all possible arguments in src/transformers/training_args.py
|
||||||
|
# or by passing the --help flag to this script.
|
||||||
|
# We now keep distinct sets of args, for a cleaner separation of concerns.
|
||||||
|
|
||||||
|
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
|
||||||
|
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
|
||||||
|
# If we pass only one argument to the script and it's the path to a json file,
|
||||||
|
# let's parse it to get our arguments.
|
||||||
|
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
|
||||||
|
else:
|
||||||
|
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
|
||||||
|
|
||||||
|
# Detecting last checkpoint.
|
||||||
|
last_checkpoint = None
|
||||||
|
if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
|
||||||
|
last_checkpoint = get_last_checkpoint(training_args.output_dir)
|
||||||
|
if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
|
||||||
|
raise ValueError(
|
||||||
|
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
|
||||||
|
"Use --overwrite_output_dir to overcome."
|
||||||
|
)
|
||||||
|
elif last_checkpoint is not None:
|
||||||
|
logger.info(
|
||||||
|
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
|
||||||
|
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
|
||||||
|
)
|
||||||
|
|
||||||
|
# Setup logging
|
||||||
|
logging.basicConfig(
|
||||||
|
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
|
||||||
|
datefmt="%m/%d/%Y %H:%M:%S",
|
||||||
|
handlers=[logging.StreamHandler(sys.stdout)],
|
||||||
|
)
|
||||||
|
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
|
||||||
|
|
||||||
|
# Log on each process the small summary:
|
||||||
|
logger.warning(
|
||||||
|
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
|
||||||
|
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
|
||||||
|
)
|
||||||
|
# Set the verbosity to info of the Transformers logger (on main process only):
|
||||||
|
if is_main_process(training_args.local_rank):
|
||||||
|
transformers.utils.logging.set_verbosity_info()
|
||||||
|
logger.info("Training/evaluation parameters %s", training_args)
|
||||||
|
|
||||||
|
# Set seed before initializing model.
|
||||||
|
set_seed(training_args.seed)
|
||||||
|
|
||||||
|
# 1. First, let's load the dataset
|
||||||
|
raw_datasets = DatasetDict()
|
||||||
|
|
||||||
|
if training_args.do_train:
|
||||||
|
raw_datasets["train"] = load_dataset(
|
||||||
|
data_args.dataset_name,
|
||||||
|
data_args.dataset_config_name,
|
||||||
|
split=data_args.train_split_name,
|
||||||
|
use_auth_token=data_args.use_auth_token,
|
||||||
|
)
|
||||||
|
|
||||||
|
if data_args.audio_column_name not in raw_datasets["train"].column_names:
|
||||||
|
raise ValueError(
|
||||||
|
f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
|
||||||
|
"Make sure to set `--audio_column_name` to the correct audio column - one of "
|
||||||
|
f"{', '.join(raw_datasets['train'].column_names)}."
|
||||||
|
)
|
||||||
|
|
||||||
|
if data_args.text_column_name not in raw_datasets["train"].column_names:
|
||||||
|
raise ValueError(
|
||||||
|
f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
|
||||||
|
"Make sure to set `--text_column_name` to the correct text column - one of "
|
||||||
|
f"{', '.join(raw_datasets['train'].column_names)}."
|
||||||
|
)
|
||||||
|
|
||||||
|
if data_args.max_train_samples is not None:
|
||||||
|
raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
|
||||||
|
|
||||||
|
if training_args.do_eval:
|
||||||
|
raw_datasets["eval"] = load_dataset(
|
||||||
|
data_args.dataset_name,
|
||||||
|
data_args.dataset_config_name,
|
||||||
|
split=data_args.eval_split_name,
|
||||||
|
use_auth_token=data_args.use_auth_token,
|
||||||
|
)
|
||||||
|
|
||||||
|
if data_args.max_eval_samples is not None:
|
||||||
|
raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
|
||||||
|
|
||||||
|
# 2. We remove some special characters from the datasets
|
||||||
|
# that make training complicated and do not help in transcribing the speech
|
||||||
|
# E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
|
||||||
|
# that could be easily picked up by the model
|
||||||
|
chars_to_ignore_regex = (
|
||||||
|
f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
|
||||||
|
)
|
||||||
|
text_column_name = data_args.text_column_name
|
||||||
|
|
||||||
|
def remove_special_characters(batch):
|
||||||
|
if chars_to_ignore_regex is not None:
|
||||||
|
batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
|
||||||
|
else:
|
||||||
|
batch["target_text"] = batch[text_column_name].lower() + " "
|
||||||
|
return batch
|
||||||
|
|
||||||
|
with training_args.main_process_first(desc="dataset map special characters removal"):
|
||||||
|
raw_datasets = raw_datasets.map(
|
||||||
|
remove_special_characters,
|
||||||
|
remove_columns=[text_column_name],
|
||||||
|
desc="remove special characters from datasets",
|
||||||
|
)
|
||||||
|
|
||||||
|
# save special tokens for tokenizer
|
||||||
|
word_delimiter_token = data_args.word_delimiter_token
|
||||||
|
unk_token = data_args.unk_token
|
||||||
|
pad_token = data_args.pad_token
|
||||||
|
|
||||||
|
# 3. Next, let's load the config as we might need it to create
|
||||||
|
# the tokenizer
|
||||||
|
# load config
|
||||||
|
config = AutoConfig.from_pretrained(
|
||||||
|
model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
|
||||||
|
)
|
||||||
|
|
||||||
|
# 4. Next, if no tokenizer file is defined,
|
||||||
|
# we create the vocabulary of the model by extracting all unique characters from
|
||||||
|
# the training and evaluation datasets
|
||||||
|
# We need to make sure that only first rank saves vocabulary
|
||||||
|
# make sure all processes wait until vocab is created
|
||||||
|
tokenizer_name_or_path = model_args.tokenizer_name_or_path
|
||||||
|
tokenizer_kwargs = {}
|
||||||
|
if tokenizer_name_or_path is None:
|
||||||
|
# save vocab in training output dir
|
||||||
|
tokenizer_name_or_path = training_args.output_dir
|
||||||
|
|
||||||
|
vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
|
||||||
|
|
||||||
|
with training_args.main_process_first():
|
||||||
|
if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
|
||||||
|
os.remove(vocab_file)
|
||||||
|
|
||||||
|
with training_args.main_process_first(desc="dataset map vocabulary creation"):
|
||||||
|
if not os.path.isfile(vocab_file):
|
||||||
|
os.makedirs(tokenizer_name_or_path, exist_ok=True)
|
||||||
|
vocab_dict = create_vocabulary_from_data(
|
||||||
|
raw_datasets,
|
||||||
|
word_delimiter_token=word_delimiter_token,
|
||||||
|
unk_token=unk_token,
|
||||||
|
pad_token=pad_token,
|
||||||
|
)
|
||||||
|
|
||||||
|
# save vocab dict to be loaded into tokenizer
|
||||||
|
with open(vocab_file, "w") as file:
|
||||||
|
json.dump(vocab_dict, file)
|
||||||
|
|
||||||
|
# if tokenizer has just been created
|
||||||
|
# it is defined by `tokenizer_class` if present in config else by `model_type`
|
||||||
|
tokenizer_kwargs = {
|
||||||
|
"config": config if config.tokenizer_class is not None else None,
|
||||||
|
"tokenizer_type": config.model_type if config.tokenizer_class is None else None,
|
||||||
|
"unk_token": unk_token,
|
||||||
|
"pad_token": pad_token,
|
||||||
|
"word_delimiter_token": word_delimiter_token,
|
||||||
|
}
|
||||||
|
|
||||||
|
# 5. Now we can instantiate the feature extractor, tokenizer and model
|
||||||
|
# Note for distributed training, the .from_pretrained methods guarantee that only
|
||||||
|
# one local process can concurrently download model & vocab.
|
||||||
|
|
||||||
|
# load feature_extractor and tokenizer
|
||||||
|
tokenizer = AutoTokenizer.from_pretrained(
|
||||||
|
tokenizer_name_or_path,
|
||||||
|
use_auth_token=data_args.use_auth_token,
|
||||||
|
**tokenizer_kwargs,
|
||||||
|
)
|
||||||
|
feature_extractor = AutoFeatureExtractor.from_pretrained(
|
||||||
|
model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
|
||||||
|
)
|
||||||
|
|
||||||
|
# adapt config
|
||||||
|
config.update(
|
||||||
|
{
|
||||||
|
"feat_proj_dropout": model_args.feat_proj_dropout,
|
||||||
|
"attention_dropout": model_args.attention_dropout,
|
||||||
|
"hidden_dropout": model_args.hidden_dropout,
|
||||||
|
"final_dropout": model_args.final_dropout,
|
||||||
|
"mask_time_prob": model_args.mask_time_prob,
|
||||||
|
"mask_time_length": model_args.mask_time_length,
|
||||||
|
"mask_feature_prob": model_args.mask_feature_prob,
|
||||||
|
"mask_feature_length": model_args.mask_feature_length,
|
||||||
|
"gradient_checkpointing": training_args.gradient_checkpointing,
|
||||||
|
"layerdrop": model_args.layerdrop,
|
||||||
|
"ctc_loss_reduction": model_args.ctc_loss_reduction,
|
||||||
|
"pad_token_id": tokenizer.pad_token_id,
|
||||||
|
"vocab_size": len(tokenizer),
|
||||||
|
"activation_dropout": model_args.activation_dropout,
|
||||||
|
}
|
||||||
|
)
|
||||||
|
|
||||||
|
# create model
|
||||||
|
model = AutoModelForCTC.from_pretrained(
|
||||||
|
model_args.model_name_or_path,
|
||||||
|
cache_dir=model_args.cache_dir,
|
||||||
|
config=config,
|
||||||
|
use_auth_token=data_args.use_auth_token,
|
||||||
|
)
|
||||||
|
|
||||||
|
# freeze encoder
|
||||||
|
if model_args.freeze_feature_encoder:
|
||||||
|
model.freeze_feature_encoder()
|
||||||
|
|
||||||
|
# 6. Now we preprocess the datasets including loading the audio, resampling and normalization
|
||||||
|
# Thankfully, `datasets` takes care of automatically loading and resampling the audio,
|
||||||
|
# so that we just need to set the correct target sampling rate and normalize the input
|
||||||
|
# via the `feature_extractor`
|
||||||
|
|
||||||
|
# make sure that dataset decodes audio with correct sampling rate
|
||||||
|
dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
|
||||||
|
if dataset_sampling_rate != feature_extractor.sampling_rate:
|
||||||
|
raw_datasets = raw_datasets.cast_column(
|
||||||
|
data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
|
||||||
|
)
|
||||||
|
|
||||||
|
# derive max & min input length for sample rate & max duration
|
||||||
|
max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
|
||||||
|
min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
|
||||||
|
audio_column_name = data_args.audio_column_name
|
||||||
|
num_workers = data_args.preprocessing_num_workers
|
||||||
|
|
||||||
|
# `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
|
||||||
|
phoneme_language = data_args.phoneme_language
|
||||||
|
|
||||||
|
# Preprocessing the datasets.
|
||||||
|
# We need to read the audio files as arrays and tokenize the targets.
|
||||||
|
def prepare_dataset(batch):
|
||||||
|
# load audio
|
||||||
|
sample = batch[audio_column_name]
|
||||||
|
|
||||||
|
inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
|
||||||
|
batch["input_values"] = inputs.input_values[0]
|
||||||
|
batch["input_length"] = len(batch["input_values"])
|
||||||
|
|
||||||
|
# encode targets
|
||||||
|
additional_kwargs = {}
|
||||||
|
if phoneme_language is not None:
|
||||||
|
additional_kwargs["phonemizer_lang"] = phoneme_language
|
||||||
|
|
||||||
|
batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
|
||||||
|
return batch
|
||||||
|
|
||||||
|
with training_args.main_process_first(desc="dataset map preprocessing"):
|
||||||
|
vectorized_datasets = raw_datasets.map(
|
||||||
|
prepare_dataset,
|
||||||
|
remove_columns=next(iter(raw_datasets.values())).column_names,
|
||||||
|
num_proc=num_workers,
|
||||||
|
desc="preprocess datasets",
|
||||||
|
)
|
||||||
|
|
||||||
|
def is_audio_in_length_range(length):
|
||||||
|
return length > min_input_length and length < max_input_length
|
||||||
|
|
||||||
|
# filter data that is shorter than min_input_length
|
||||||
|
vectorized_datasets = vectorized_datasets.filter(
|
||||||
|
is_audio_in_length_range,
|
||||||
|
num_proc=num_workers,
|
||||||
|
input_columns=["input_length"],
|
||||||
|
)
|
||||||
|
|
||||||
|
# 7. Next, we can prepare the training.
|
||||||
|
# Let's use word error rate (WER) as our evaluation metric,
|
||||||
|
# instantiate a data collator and the trainer
|
||||||
|
|
||||||
|
# Define evaluation metrics during training, *i.e.* word error rate, character error rate
|
||||||
|
eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
|
||||||
|
|
||||||
|
# for large datasets it is advised to run the preprocessing on a
|
||||||
|
# single machine first with ``args.preprocessing_only`` since there will mostly likely
|
||||||
|
# be a timeout when running the script in distributed mode.
|
||||||
|
# In a second step ``args.preprocessing_only`` can then be set to `False` to load the
|
||||||
|
# cached dataset
|
||||||
|
if data_args.preprocessing_only:
|
||||||
|
logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
|
||||||
|
return
|
||||||
|
|
||||||
|
def compute_metrics(pred):
|
||||||
|
pred_logits = pred.predictions
|
||||||
|
pred_ids = np.argmax(pred_logits, axis=-1)
|
||||||
|
|
||||||
|
pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
|
||||||
|
|
||||||
|
pred_str = tokenizer.batch_decode(pred_ids)
|
||||||
|
# we do not want to group tokens when computing the metrics
|
||||||
|
label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
|
||||||
|
|
||||||
|
metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
|
||||||
|
|
||||||
|
return metrics
|
||||||
|
|
||||||
|
# Now save everything to be able to create a single processor later
|
||||||
|
if is_main_process(training_args.local_rank):
|
||||||
|
# save feature extractor, tokenizer and config
|
||||||
|
feature_extractor.save_pretrained(training_args.output_dir)
|
||||||
|
tokenizer.save_pretrained(training_args.output_dir)
|
||||||
|
config.save_pretrained(training_args.output_dir)
|
||||||
|
|
||||||
|
try:
|
||||||
|
processor = AutoProcessor.from_pretrained(training_args.output_dir)
|
||||||
|
except (OSError, KeyError):
|
||||||
|
warnings.warn(
|
||||||
|
"Loading a processor from a feature extractor config that does not"
|
||||||
|
" include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
|
||||||
|
" attribute to your `preprocessor_config.json` file to suppress this warning: "
|
||||||
|
" `'processor_class': 'Wav2Vec2Processor'`",
|
||||||
|
FutureWarning,
|
||||||
|
)
|
||||||
|
processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
|
||||||
|
|
||||||
|
# Instantiate custom data collator
|
||||||
|
data_collator = DataCollatorCTCWithPadding(processor=processor)
|
||||||
|
|
||||||
|
decay_parameters = get_parameter_names(model, [torch.nn.LayerNorm])
|
||||||
|
decay_parameters = [name for name in decay_parameters if "bias" not in name]
|
||||||
|
optimizer_grouped_parameters = [
|
||||||
|
{
|
||||||
|
"params": [p for n, p in model.named_parameters() if n in decay_parameters],
|
||||||
|
"weight_decay": training_args.weight_decay,
|
||||||
|
},
|
||||||
|
{
|
||||||
|
"params": [p for n, p in model.named_parameters() if n not in decay_parameters],
|
||||||
|
"weight_decay": 0.0,
|
||||||
|
},
|
||||||
|
]
|
||||||
|
optimizer = bnb.optim.Adam8bit(
|
||||||
|
params=optimizer_grouped_parameters,
|
||||||
|
betas=(training_args.adam_beta1, training_args.adam_beta2),
|
||||||
|
eps=training_args.adam_epsilon,
|
||||||
|
)
|
||||||
|
|
||||||
|
optimizers = (optimizer, None)
|
||||||
|
|
||||||
|
# Initialize Trainer
|
||||||
|
trainer = Trainer(
|
||||||
|
model=model,
|
||||||
|
data_collator=data_collator,
|
||||||
|
args=training_args,
|
||||||
|
compute_metrics=compute_metrics,
|
||||||
|
train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
|
||||||
|
eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
|
||||||
|
tokenizer=feature_extractor,
|
||||||
|
optimizers=optimizers,
|
||||||
|
)
|
||||||
|
|
||||||
|
# 8. Finally, we can start training
|
||||||
|
|
||||||
|
# Training
|
||||||
|
if training_args.do_train:
|
||||||
|
|
||||||
|
# use last checkpoint if exist
|
||||||
|
if last_checkpoint is not None:
|
||||||
|
checkpoint = last_checkpoint
|
||||||
|
elif os.path.isdir(model_args.model_name_or_path):
|
||||||
|
checkpoint = model_args.model_name_or_path
|
||||||
|
else:
|
||||||
|
checkpoint = None
|
||||||
|
|
||||||
|
train_result = trainer.train(resume_from_checkpoint=checkpoint)
|
||||||
|
trainer.save_model()
|
||||||
|
|
||||||
|
metrics = train_result.metrics
|
||||||
|
max_train_samples = (
|
||||||
|
data_args.max_train_samples
|
||||||
|
if data_args.max_train_samples is not None
|
||||||
|
else len(vectorized_datasets["train"])
|
||||||
|
)
|
||||||
|
metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
|
||||||
|
|
||||||
|
trainer.log_metrics("train", metrics)
|
||||||
|
trainer.save_metrics("train", metrics)
|
||||||
|
trainer.save_state()
|
||||||
|
|
||||||
|
# Evaluation
|
||||||
|
results = {}
|
||||||
|
if training_args.do_eval:
|
||||||
|
logger.info("*** Evaluate ***")
|
||||||
|
metrics = trainer.evaluate()
|
||||||
|
max_eval_samples = (
|
||||||
|
data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
|
||||||
|
)
|
||||||
|
metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
|
||||||
|
|
||||||
|
trainer.log_metrics("eval", metrics)
|
||||||
|
trainer.save_metrics("eval", metrics)
|
||||||
|
|
||||||
|
# Write model card and (optionally) push to hub
|
||||||
|
config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
|
||||||
|
kwargs = {
|
||||||
|
"finetuned_from": model_args.model_name_or_path,
|
||||||
|
"tasks": "speech-recognition",
|
||||||
|
"tags": ["automatic-speech-recognition", data_args.dataset_name],
|
||||||
|
"dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
|
||||||
|
"dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
|
||||||
|
}
|
||||||
|
if "common_voice" in data_args.dataset_name:
|
||||||
|
kwargs["language"] = config_name
|
||||||
|
|
||||||
|
if training_args.push_to_hub:
|
||||||
|
trainer.push_to_hub(**kwargs)
|
||||||
|
else:
|
||||||
|
trainer.create_model_card(**kwargs)
|
||||||
|
|
||||||
|
return results
|
||||||
|
|
||||||
|
|
||||||
|
if __name__ == "__main__":
|
||||||
|
main()
|
||||||
@@ -1,395 +0,0 @@
|
|||||||
# Speech recognition community week - version 2 🤗
|
|
||||||
|
|
||||||
Welcome to the 2nd version of the speech recognition community event🎙️ !
|
|
||||||
The goal of this event is to build **robust**, **real-world** speech recognition (ASR) models in as many languages as possible🌏🌍🌎.
|
|
||||||
|
|
||||||
If necessary and available, free access to a V100 32 GB GPU will kindly be provided by the [OVH team](https://us.ovhcloud.com/) 🚀.
|
|
||||||
|
|
||||||
This document summarizes all the relevant information required for the speech community event📋.
|
|
||||||
|
|
||||||
Don't forget to sign up [here](TODO: Create google from)🤗.
|
|
||||||
|
|
||||||
## Table of Contents
|
|
||||||
|
|
||||||
- [Organization](#organization)
|
|
||||||
- [Important dates](#important-dates)
|
|
||||||
- [How to install pytorch, transformers, datasets](#how-to-install-relevant-libraries)
|
|
||||||
- [How to fine-tune a speech recognition model](#how-to-finetune-a-model)
|
|
||||||
- [Talks](#talks)
|
|
||||||
- [Project evaluation](#project-evaluation)
|
|
||||||
- [General Tips & Tricks](#general-tips-and-tricks)
|
|
||||||
- [FAQ](#faq)
|
|
||||||
|
|
||||||
## Organization
|
|
||||||
|
|
||||||
Participants are encouraged to leverage pre-trained speech recognition checkpoints,
|
|
||||||
preferably [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53), to train a speech recognition system in a language of their
|
|
||||||
choice.
|
|
||||||
|
|
||||||
Participants can make use of whatever data they think is useful to build a
|
|
||||||
**robust** speech recognition system for **real-world** audio data. We strongly
|
|
||||||
recommend making use of [Mozilla's diverse Common Voice dataset](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0) when training the model.
|
|
||||||
Please do **not** use the `"test"` split of the Common Voice datasets for training
|
|
||||||
as we will likely use this split for the final evaluation of your model.
|
|
||||||
We kindly ask you to make sure that the dataset that you are using for training
|
|
||||||
has the appropriate licensing - see [here](TODO: ) for more information.
|
|
||||||
|
|
||||||
During the event, the fine-tuned models will regularly be tested on a **development
|
|
||||||
dataset** provided by the Hugging Face team and at the end of the event, all models
|
|
||||||
will be tested on a **test dataset**. For each language,
|
|
||||||
the best performing model will receive a prize 🏆 - more information regarding
|
|
||||||
the testing [here](TODO: ) and prizes [here](TODO: ). We believe that framing the
|
|
||||||
event as a competition is more fun, but at the core, we strongly encourage
|
|
||||||
participants to work together by helping each other to solve bugs, share important findings, etc...🤗
|
|
||||||
|
|
||||||
If possible it is encouraged to fine-tune the models on local GPU machines, but
|
|
||||||
if those are not available, the OVH cloud team kindly provides a limited
|
|
||||||
number of GPUs for the event. For more information on how to get access to the GPU - see [here](TODO: ).
|
|
||||||
|
|
||||||
|
|
||||||
**Please note**:
|
|
||||||
All important announcements will be made on discord. Please make sure that
|
|
||||||
you've joined the following discord server: TODO: fill out.
|
|
||||||
Please make sure that you have been added to the [Speech Event Organization](https://huggingface.co/speech-recognition-community-v2). You should have received an
|
|
||||||
invite by email. If you didn't receive an invite, please contact the organizers, *e.g.* Anton, Patrick, or Omar on discord.
|
|
||||||
|
|
||||||
|
|
||||||
## Important dates
|
|
||||||
|
|
||||||
- **12.01.** Official announcement of the community event. Make sure to sign-up in [this google form](TODO: )
|
|
||||||
- **12.01. - 19.01.** Participants sign up for the event.
|
|
||||||
- **19.01.** Release of all relevant guides and training scripts.
|
|
||||||
- **24.01.** Start of the community week! OVH & Hugging Face gives access to GPUs.
|
|
||||||
- **24.01. - 07.02.** The OVH & Hugging Face team will be available for any questions, problems the participants might have.
|
|
||||||
- **07.02.** Access to GPU is deactivated and community week officially ends.
|
|
||||||
|
|
||||||
## How to install relevant libraries
|
|
||||||
|
|
||||||
The following libraries are required to fine-tune a speech model with 🤗 Transformers and 🤗 Datasets in PyTorch.
|
|
||||||
|
|
||||||
- [PyTorch](https://pytorch.org/)
|
|
||||||
- [Transformers](https://github.com/huggingface/transformers)
|
|
||||||
- [Datasets](https://github.com/huggingface/datasets)
|
|
||||||
|
|
||||||
We recommend installing the above libraries in a [virtual environment](https://docs.python.org/3/library/venv.html).
|
|
||||||
If you're unfamiliar with Python virtual environments, check out the [user guide](https://packaging.python.org/guides/installing-using-pip-and-virtual-environments/). Create a virtual environment with the version of Python you're going
|
|
||||||
to use and activate it.
|
|
||||||
|
|
||||||
You should be able to run the command:
|
|
||||||
|
|
||||||
```bash
|
|
||||||
python3 -m venv <your-venv-name>
|
|
||||||
```
|
|
||||||
|
|
||||||
You can activate your venv by running
|
|
||||||
|
|
||||||
```bash
|
|
||||||
source ~/<your-venv-name>/bin/activate
|
|
||||||
```
|
|
||||||
|
|
||||||
To begin with please make sure you have PyTorch and CUDA correctly installed.
|
|
||||||
The following command should return ``True``:
|
|
||||||
|
|
||||||
```bash
|
|
||||||
python -c "import torch; print(torch.cuda.is_available())"
|
|
||||||
```
|
|
||||||
|
|
||||||
If the above command doesn't print ``True``, in a first step, please follow the
|
|
||||||
instructions [here](https://pytorch.org/) to install PyTorch with CUDA.
|
|
||||||
|
|
||||||
We strongly recommend making use of the provided PyTorch examples scripts in [transformers/examples/pytorch/speech-recognition](https://github.com/huggingface/transformers/tree/master/examples/pytorch/speech-recognition) to train your speech recognition
|
|
||||||
system.
|
|
||||||
In all likelihood, you will adjust one of the example scripts, so we recommend forking and cloning the 🤗 Transformers repository as follows.
|
|
||||||
|
|
||||||
1. Fork the [repository](https://github.com/huggingface/transformers) by
|
|
||||||
clicking on the 'Fork' button on the repository's page. This creates a copy of the code
|
|
||||||
under your GitHub user account.
|
|
||||||
|
|
||||||
2. Clone your fork to your local disk, and add the base repository as a remote:
|
|
||||||
|
|
||||||
```bash
|
|
||||||
$ git clone https://github.com/<your Github handle>/transformers.git
|
|
||||||
$ cd transformers
|
|
||||||
$ git remote add upstream https://github.com/huggingface/transformers.git
|
|
||||||
```
|
|
||||||
|
|
||||||
3. Create a new branch to hold your development changes. This is especially useful to share code changes with your team:
|
|
||||||
|
|
||||||
```bash
|
|
||||||
$ git checkout -b a-descriptive-name-for-my-project
|
|
||||||
```
|
|
||||||
|
|
||||||
4. Set up a PyTorch environment by running the following command your virtual environment:
|
|
||||||
|
|
||||||
```bash
|
|
||||||
$ pip install -e ".[torch-speech]"
|
|
||||||
```
|
|
||||||
|
|
||||||
(If transformers was already installed in the virtual environment, remove
|
|
||||||
it with `pip uninstall transformers` before reinstalling it in editable
|
|
||||||
mode with the `-e` flag.)
|
|
||||||
|
|
||||||
If you have already cloned that repo, you might need to `git pull` to get the most recent changes in the `transformers`
|
|
||||||
library.
|
|
||||||
|
|
||||||
Running this command will automatically install `pytorch` and the most relevant
|
|
||||||
libraries required for fine-tuning a speech recognition system.
|
|
||||||
|
|
||||||
Next, you should also install the 🤗 Datasets library. We strongly recommend installing the
|
|
||||||
library from source to profit from the most current additions during the community week.
|
|
||||||
|
|
||||||
Simply run the following steps:
|
|
||||||
|
|
||||||
```
|
|
||||||
$ cd ~/
|
|
||||||
$ git clone https://github.com/huggingface/datasets.git
|
|
||||||
$ cd datasets
|
|
||||||
$ pip install -e ".[streaming]"
|
|
||||||
```
|
|
||||||
|
|
||||||
If you plan on contributing a specific dataset during
|
|
||||||
the community week, please fork the datasets repository and follow the instructions
|
|
||||||
[here](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-create-a-pull-request).
|
|
||||||
|
|
||||||
To verify that all libraries are correctly installed, you can run the following command in a Python shell.
|
|
||||||
It verifies that both `transformers` and `datasets` have been correclty installed.
|
|
||||||
|
|
||||||
```python
|
|
||||||
from transformers import AutoModelForCTC, AutoProcessor
|
|
||||||
from datasets import load_dataset
|
|
||||||
|
|
||||||
dummy_dataset = load_dataset("common_voice", "ab", split="test")
|
|
||||||
|
|
||||||
model = AutoModelForCTC.from_pretrained("hf-internal-testing/tiny-random-wav2vec2")
|
|
||||||
model.to("cuda")
|
|
||||||
|
|
||||||
processor = AutoProcessor.from_pretrained("hf-internal-testing/tiny-random-wav2vec2")
|
|
||||||
|
|
||||||
input_values = processor(dummy_dataset[0]["audio"]["array"], return_tensors="pt", sampling_rate=16_000).input_values
|
|
||||||
input_values = input_values.to("cuda")
|
|
||||||
|
|
||||||
logits = model(input_values).logits
|
|
||||||
|
|
||||||
assert logits.shape[-1] == 32
|
|
||||||
```
|
|
||||||
|
|
||||||
## How to finetune a model
|
|
||||||
|
|
||||||
In this section, we show you how to fine-tune a pre-trained [XLS-R Model](https://huggingface.co/docs/transformers/model_doc/xls_r) on the [Common Voice 7 dataset](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0).
|
|
||||||
|
|
||||||
We recommend fine-tuning one of the following pre-trained XLS-R checkpoints:
|
|
||||||
|
|
||||||
- [300M parameters version](https://huggingface.co/facebook/wav2vec2-xls-r-300m)
|
|
||||||
- [1B parameters version](https://huggingface.co/facebook/wav2vec2-xls-r-1b)
|
|
||||||
- [2B parameters version](https://huggingface.co/facebook/wav2vec2-xls-r-2b)
|
|
||||||
|
|
||||||
To begin with, please note that to use the Common Voice dataset, you
|
|
||||||
have to accept that **your email address** and **username** are shared with the
|
|
||||||
mozilla-foundation. To get access to the dataset please click on "*Access repository*" [here](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0).
|
|
||||||
|
|
||||||
Next, we recommended that you get familiar with the XLS-R model and its capabilities.
|
|
||||||
In collaboration with [Fairseq's Wav2Vec2 team](https://github.com/pytorch/fairseq/tree/main/examples/wav2vec),
|
|
||||||
we've written ["Fine-tuning XLS-R for Multi-Lingual ASR with 🤗 Transformers"](https://huggingface.co/blog/fine-tune-xlsr-wav2vec2) which gives an in-detail explanation of how XLS-R functions and how it can be fine-tuned.
|
|
||||||
|
|
||||||
The blog can also be opened and directly fine-tuned in a google colab notebook.
|
|
||||||
In this section, we will explain how to fine-tune the model on a local machine.
|
|
||||||
|
|
||||||
1. **Log in**
|
|
||||||
|
|
||||||
To begin with you should check that you are correctly logged in and that you have `git-lfs` installed so that your fine-tuned model can automatically be uploaded.
|
|
||||||
|
|
||||||
Run:
|
|
||||||
|
|
||||||
```bash
|
|
||||||
huggingface-cli login
|
|
||||||
```
|
|
||||||
|
|
||||||
to login. It is recommend to login with your personal access token that can be found under your hugging face profile (icon in the top right corner on [hf.co](http://hf.co/), then Settings -> Access Tokens -> User Access Tokens -> New Token (if haven't generated one already)
|
|
||||||
|
|
||||||
You can then copy-paste this token to log in locally.
|
|
||||||
|
|
||||||
2. **Create your model repository**
|
|
||||||
|
|
||||||
First, let's make sure that `git-lfs` is correctly installed. To so, simply run:
|
|
||||||
|
|
||||||
```bash
|
|
||||||
git-lfs -v
|
|
||||||
```
|
|
||||||
|
|
||||||
The output should show something like `git-lfs/2.13.2 (GitHub; linux amd64; go 1.15.4)`. If your console states that the `git-lfs` command was not found, please make
|
|
||||||
sure to install it [here](https://git-lfs.github.com/) or simply via:
|
|
||||||
|
|
||||||
```bash
|
|
||||||
sudo apt-get install git-lfs
|
|
||||||
```
|
|
||||||
|
|
||||||
Now you can create your model repository which will contain all relevant files to
|
|
||||||
reproduce your training. You can either directly create the model repository on the
|
|
||||||
Hub (Settings -> New Model) or via the CLI. Here we choose to use the CLI instead.
|
|
||||||
|
|
||||||
Assuming that we want to call our model repository *xls-r-ab-test*, we can run the
|
|
||||||
following command:
|
|
||||||
|
|
||||||
```bash
|
|
||||||
huggingface-cli repo create xls-r-ab-test
|
|
||||||
```
|
|
||||||
|
|
||||||
You can now see the model on the Hub, *e.g.* under https://huggingface.co/hf-test/xls-r-ab-test .
|
|
||||||
|
|
||||||
Let's clone the repository so that we can define our training script inside.
|
|
||||||
|
|
||||||
```bash
|
|
||||||
git lfs install
|
|
||||||
git clone https://huggingface.co/hf-test/xls-r-ab-test
|
|
||||||
```
|
|
||||||
|
|
||||||
3. **Add your training script and `run`-command to the repository**
|
|
||||||
|
|
||||||
We encourage participants to add all relevant files for training directly to the
|
|
||||||
directory so that everything is fully reproducible.
|
|
||||||
|
|
||||||
Let's first copy-paste the official training script from our clone
|
|
||||||
of `transformers` to our just created directory:
|
|
||||||
|
|
||||||
```bash
|
|
||||||
cp ~/transformers/examples/pytorch/speech-recognition/run_speech_recognition_ctc.py ./
|
|
||||||
```
|
|
||||||
|
|
||||||
Next, we'll create a bash file to define the hyper-parameters and configurations
|
|
||||||
for training. More detailed information on different settings (single-GPU vs. multi-GPU) can be found [here](https://github.com/huggingface/transformers/tree/master/examples/pytorch/speech-recognition#connectionist-temporal-classification).
|
|
||||||
|
|
||||||
For demonstration purposes, we will use a dummy XLS-R model `model_name_or_path="hf-test/xls-r-dummy"` on the very low-resource language of "Abkhaz" of [Common Voice 7](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0): `dataset_config_name="ab"` for just a single epoch.
|
|
||||||
|
|
||||||
Before starting to train, let's make sure we have installed all the required libraries. You might want to run:
|
|
||||||
|
|
||||||
```bash
|
|
||||||
pip install -r ~/transformers/examples/pytorch/speech-recognition/requirements.txt
|
|
||||||
```
|
|
||||||
|
|
||||||
Alright, finally we can define the training script. We'll simply use some
|
|
||||||
dummy hyper-parameters and configurations for demonstration purposes.
|
|
||||||
|
|
||||||
Note that we add the flag `--use_auth_token` so that datasets requiring access,
|
|
||||||
such as [Common Voice 7](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0) can be downloaded. In addition, we add the `--push_to_hub` flag to make use of the
|
|
||||||
[Trainers `push_to-hub` functionality](https://huggingface.co/docs/transformers/master/en/main_classes/trainer#transformers.Trainer.push_to_hub) so that your model will be automatically uploaded to the Hub.
|
|
||||||
|
|
||||||
Let's copy the following code snippet in a file called `run.sh`
|
|
||||||
|
|
||||||
```bash
|
|
||||||
echo '''python run_speech_recognition_ctc.py \
|
|
||||||
--dataset_name="mozilla-foundation/common_voice_7_0" \
|
|
||||||
--model_name_or_path="hf-test/xls-r-dummy" \
|
|
||||||
--dataset_config_name="ab" \
|
|
||||||
--output_dir="./" \
|
|
||||||
--overwrite_output_dir \
|
|
||||||
--max_steps="10" \
|
|
||||||
--per_device_train_batch_size="2" \
|
|
||||||
--learning_rate="3e-4" \
|
|
||||||
--save_total_limit="1" \
|
|
||||||
--evaluation_strategy="steps" \
|
|
||||||
--text_column_name="sentence" \
|
|
||||||
--length_column_name="input_length" \
|
|
||||||
--save_steps="5" \
|
|
||||||
--layerdrop="0.0" \
|
|
||||||
--freeze_feature_encoder \
|
|
||||||
--gradient_checkpointing \
|
|
||||||
--fp16 \
|
|
||||||
--group_by_length \
|
|
||||||
--push_to_hub \
|
|
||||||
--use_auth_token \
|
|
||||||
--do_train --do_eval''' > run.sh
|
|
||||||
```
|
|
||||||
|
|
||||||
4. **Start training**
|
|
||||||
|
|
||||||
Now all that is left to do is to start training the model by executing the
|
|
||||||
run file.
|
|
||||||
|
|
||||||
```bash
|
|
||||||
bash run.sh
|
|
||||||
```
|
|
||||||
|
|
||||||
The training should not take more than a couple of minutes.
|
|
||||||
During the training intermediate saved checkpoints are automatically uploaded to
|
|
||||||
your model repository as can be seen [on this commit](https://huggingface.co/hf-test/xls-r-ab-test/commit/0eb19a0fca4d7d163997b59663d98cd856022aa6) .
|
|
||||||
|
|
||||||
At the end of the training, the [Trainer](https://huggingface.co/docs/transformers/master/en/main_classes/trainer) automatically creates a nice model card and all
|
|
||||||
relevant files are uploaded.
|
|
||||||
|
|
||||||
5. **Tips for real model training**
|
|
||||||
|
|
||||||
The above steps illustrate how a model can technically be fine-tuned.
|
|
||||||
However as you can see on the model card [hf-test/xls-r-ab-test](https://huggingface.co/hf-test/xls-r-ab-test), our demonstration has a very poor performance which is
|
|
||||||
not surprising given that we trained for just 10 steps on a randomly initialized
|
|
||||||
model.
|
|
||||||
|
|
||||||
For a real model training, one of the actual pre-trained XLS-R models should be used:
|
|
||||||
|
|
||||||
- [300M parameters version](https://huggingface.co/facebook/wav2vec2-xls-r-300m)
|
|
||||||
- [1B parameters version](https://huggingface.co/facebook/wav2vec2-xls-r-1b)
|
|
||||||
- [2B parameters version](https://huggingface.co/facebook/wav2vec2-xls-r-2b)
|
|
||||||
|
|
||||||
Also, the hyper-parameters should be carefully chosen depending on the dataset.
|
|
||||||
As an example, we will fine-tune the 300M parameters model on Swedish on a single
|
|
||||||
TITAN RTX 24GB GPU.
|
|
||||||
|
|
||||||
The model will be called `"xls-r-300m-sv"`.
|
|
||||||
Following the above steps we first create the model:
|
|
||||||
|
|
||||||
```bash
|
|
||||||
huggingface-cli repo create xls-r-300m-sv
|
|
||||||
```
|
|
||||||
|
|
||||||
and then clone it locally:
|
|
||||||
|
|
||||||
```bash
|
|
||||||
|
|
||||||
|
|
||||||
and we define the following
|
|
||||||
|
|
||||||
hyperparameters for training
|
|
||||||
|
|
||||||
```bash
|
|
||||||
|
|
||||||
echo '''python run_speech_recognition_ctc.py \
|
|
||||||
--dataset_name="mozilla-foundation/common_voice_7_0" \
|
|
||||||
--model_name_or_path="facebook/wav2vec2-xls-r-300m" \
|
|
||||||
--dataset_config_name="sv-SE" \
|
|
||||||
--output_dir="./" \
|
|
||||||
--overwrite_output_dir \
|
|
||||||
--num_train_epochs="50" \
|
|
||||||
--per_device_train_batch_size="8" \
|
|
||||||
--per_device_eval_batch_size="8" \
|
|
||||||
--gradient_accumulation_steps="4" \
|
|
||||||
--learning_rate="7.5e-5" \
|
|
||||||
--warmup_steps="2000" \
|
|
||||||
--length_column_name="input_length" \
|
|
||||||
--evaluation_strategy="steps" \
|
|
||||||
--text_column_name="sentence" \
|
|
||||||
--chars_to_ignore , ? . ! \- \; \: \" “ % ‘ ” <20> — ’ … – \
|
|
||||||
--save_steps="500" \
|
|
||||||
--eval_steps="500" \
|
|
||||||
--logging_steps="100" \
|
|
||||||
--layerdrop="0.0" \
|
|
||||||
--activation_dropout="0.1" \
|
|
||||||
--save_total_limit="3" \
|
|
||||||
--freeze_feature_encoder \
|
|
||||||
--feat_proj_dropout="0.0" \
|
|
||||||
--mask_time_prob="0.75" \
|
|
||||||
--mask_time_length="10" \
|
|
||||||
--mask_feature_prob="0.25" \
|
|
||||||
--mask_feature_length="64" \
|
|
||||||
--gradient_checkpointing \
|
|
||||||
--use_auth_token \
|
|
||||||
--fp16 \
|
|
||||||
--group_by_length \
|
|
||||||
--do_train --do_eval \
|
|
||||||
--push_to_hub''' > run.sh
|
|
||||||
```
|
|
||||||
|
|
||||||
The training takes *ca.* 7 hours and yields a reasonable test word
|
|
||||||
error rate of 27% as can be seen on the automatically generated [model card](https://huggingface.co/hf-test/xls-r-300m-sv).
|
|
||||||
|
|
||||||
The above-chosen hyperparameters probably work quite well on a range of different
|
|
||||||
datasets and languages, but are by no means optimal. It is up to you to find a good set of
|
|
||||||
hyperparameters.
|
|
||||||
Reference in New Issue
Block a user