SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Adding the LXMERT pretraining model (MultiModal languageXvision) to HuggingFace's suite of models (#5793)

Browse Source 
* added template files for LXMERT and competed the configuration_lxmert.py

* added modeling, tokization, testing, and finishing touched for lxmert [yet to be tested]

* added model card for lxmert

* cleaning up lxmert code

* Update src/transformers/modeling_lxmert.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Update src/transformers/modeling_tf_lxmert.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Update src/transformers/modeling_tf_lxmert.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Update src/transformers/modeling_lxmert.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* tested torch lxmert, changed documtention, updated outputs, and other small fixes

* Update src/transformers/convert_pytorch_checkpoint_to_tf2.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Update src/transformers/convert_pytorch_checkpoint_to_tf2.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Update src/transformers/convert_pytorch_checkpoint_to_tf2.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* renaming, other small issues, did not change TF code in this commit

* added lxmert question answering model in pytorch

* added capability to edit number of qa labels for lxmert

* made answer optional for lxmert question answering

* add option to return hidden_states for lxmert

* changed default qa labels for lxmert

* changed config archive path

* squshing 3 commits: merged UI + testing improvments + more UI and testing

* changed some variable names for lxmert

* TF LXMERT

* Various fixes to LXMERT

* Final touches to LXMERT

* AutoTokenizer order

* Add LXMERT to index.rst and README.md

* Merge commit test fixes + Style update

* TensorFlow 2.3.0 sequential model changes variable names

Remove inherited test

* Update src/transformers/modeling_tf_pytorch_utils.py

* Update docs/source/model_doc/lxmert.rst

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update docs/source/model_doc/lxmert.rst

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/modeling_tf_lxmert.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* added suggestions

* Fixes

* Final fixes for TF model

* Fix docs

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>
Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
This commit is contained in:
Antonio V Mendoza
2020-09-03 04:02:25 -04:00
committed by GitHub
parent 4ebb52afdb
commit ea2c6f1afc
23 changed files with 4798 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
README.md
View File

@@ -172,8 +172,9 @@ for Open-Domain Question Answering](https://arxiv.org/abs/2004.04906) by Vladimi
Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
23. **[Pegasus](https://github.com/google-research/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777)> by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
24. **[MBart](https://github.com/pytorch/fairseq/tree/master/examples/mbart)** (from Facebook) released with the paper [Multilingual Denoising Pre-training for Neural Machine Translation](https://arxiv.org/abs/2001.08210) by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
25. **[Other community models](https://huggingface.co/models)**, contributed by the [community](https://huggingface.co/users).
26. Want to contribute a new model? We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder of the repository. Be sure to check the [contributing guidelines](./CONTRIBUTING.md) and contact the maintainers or open an issue to collect feedbacks before starting your PR.
25. **[LXMERT](https://github.com/airsplay/lxmert)** (from UNC Chapel Hill) released with the paper [LXMERT: Learning Cross-Modality Encoder Representations from Transformers for Open-Domain Question Answering](https://arxiv.org/abs/1908.07490) by Hao Tan and Mohit Bansal.
26. **[Other community models](https://huggingface.co/models)**, contributed by the [community](https://huggingface.co/users).
27. Want to contribute a new model? We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder of the repository. Be sure to check the [contributing guidelines](./CONTRIBUTING.md) and contact the maintainers or open an issue to collect feedbacks before starting your PR.
These implementations have been tested on several datasets (see the example scripts) and should match the performances of the original implementations (e.g. ~93 F1 on SQuAD for BERT Whole-Word-Masking, ~88 F1 on RocStories for OpenAI GPT, ~18.3 perplexity on WikiText 103 for Transformer-XL, ~0.916 Pearson R coefficient on STS-B for XLNet). You can find more details on the performances in the Examples section of the [documentation](https://huggingface.co/transformers/examples.html).

6
docs/source/index.rst
View File

@@ -128,7 +128,10 @@ conversion utilities for the following models:
<https://arxiv.org/abs/1912.08777>`_ by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
24. `MBart <https://github.com/pytorch/fairseq/tree/master/examples/mbart>`_ (from Facebook) released with the paper `Multilingual Denoising Pre-training for Neural Machine Translation <https://arxiv.org/abs/2001.08210>`_ by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov,
Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
25. `Other community models <https://huggingface.co/models>`_, contributed by the `community
25. `LXMERT <https://github.com/airsplay/lxmert>`_ (from UNC Chapel Hill) released with the paper `LXMERT: Learning
Cross-Modality Encoder Representations from Transformers for Open-Domain Question
Answering <https://arxiv.org/abs/1908.07490>`_ by Hao Tan and Mohit Bansal.
26. `Other community models <https://huggingface.co/models>`_, contributed by the `community
<https://huggingface.co/users>`_.
.. toctree::
@@ -213,6 +216,7 @@ conversion utilities for the following models:
model_doc/dpr
model_doc/pegasus
model_doc/mbart
model_doc/lxmert
internal/modeling_utils
internal/tokenization_utils
internal/pipelines_utils

109
docs/source/model_doc/lxmert.rst Normal file
View File

@@ -0,0 +1,109 @@
LXMERT
----------------------------------------------------
Overview
~~~~~~~~~~~~~~~~~~~~~
The LXMERT model was proposed in `LXMERT: Learning Cross-Modality Encoder Representations from Transformers <https://arxiv.org/abs/1908.07490>`__
by Hao Tan & Mohit Bansal. It is a series of bidirectional transformer encoders (one for the vision modality, one for the language modality, and then one to fuse both modalities)
pre-trained using a combination of masked language modeling, visual-language text alignment, ROI-feature regression, masked visual-attribute modeling, masked visual-object modeling, and visual-question answering objectives.
The pretraining consists of multiple multi-modal datasets: MSCOCO, Visual-Genome + Visual-Genome Question Answering, VQA 2.0, and GQA.
The abstract from the paper is the following:
*Vision-and-language reasoning requires an understanding of visual concepts, language semantics, and, most importantly, the alignment and relationships between these two
modalities. We thus propose the LXMERT
(Learning Cross-Modality Encoder Representations from Transformers) framework to learn
these vision-and-language connections. In
LXMERT, we build a large-scale Transformer
model that consists of three encoders: an object relationship encoder, a language encoder,
and a cross-modality encoder. Next, to endow our model with the capability of connecting vision and language semantics, we
pre-train the model with large amounts of
image-and-sentence pairs, via five diverse representative pre-training tasks: masked language modeling, masked object prediction
(feature regression and label classification),
cross-modality matching, and image question answering. These tasks help in learning both intra-modality and cross-modality relationships. After fine-tuning from our pretrained parameters, our model achieves the
state-of-the-art results on two visual question answering datasets (i.e., VQA and GQA).
We also show the generalizability of our pretrained cross-modality model by adapting it to
a challenging visual-reasoning task, NLVR
,
and improve the previous best result by 22%
absolute (54% to 76%). Lastly, we demonstrate detailed ablation studies to prove that
both our novel model components and pretraining strategies significantly contribute to
our strong results; and also present several
attention visualizations for the different encoders*
Tips:
- Bounding boxes are not necessary to be used in the visual feature embeddings, any kind of visual-spacial features will work.
- Both the language hidden states and the visual hidden states that LXMERT outputs are passed through the cross-modality layer, so they
contain information from both modalities. To access a modality that only attends to itself, select the vision/language hidden states from the first input in the tuple.
- The bi-directional cross-modality encoder attention only returns attention values when the language modality is used as the input and the vision modality is used as the context vector. Further,
while the cross-modality encoder contains self-attention for each respective modality and cross-attention, only the cross attention is returned and both self attention outputs are disregarded.
The code can be found `here <https://github.com/airsplay/lxmert>`__
LxmertConfig
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.LxmertConfig
:members:
LxmertTokenizer
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.LxmertTokenizer
:members: build_inputs_with_special_tokens, get_special_tokens_mask,
create_token_type_ids_from_sequences, save_vocabulary
Lxmert specific outputs
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.modeling_lxmert.LxmertModelOutput
:members:
.. autoclass:: transformers.modeling_lxmert.LxmertForPreTrainingOutput
:members:
.. autoclass:: transformers.modeling_lxmert.LxmertForQuestionAnsweringOutput
:members:
.. autoclass:: transformers.modeling_tf_lxmert.TFLxmertModelOutput
:members:
.. autoclass:: transformers.modeling_tf_lxmert.TFLxmertForPreTrainingOutput
:members:
LxmertModel
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.LxmertModel
:members:
LxmertForPreTraining
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.LxmertForPreTraining
:members:
LxmertForQuestionAnswering
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.LxmertForQuestionAnswering
:members:
TFLxmertModel
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.TFLxmertModel
:members:
TFLxmertForPreTraining
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.TFLxmertForPreTraining
:members:

4
docs/source/pretrained_models.rst
View File

@@ -364,3 +364,7 @@ For a list that includes community-uploaded models, refer to `https://huggingfac
| | ``facebook/mbart-large-en-ro`` | | 24-layer, 1024-hidden, 16-heads, 610M parameters |
| | | | mbart-large-cc25 model finetuned on WMT english romanian translation. |
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| Lxmert | ``lxmert-base-uncased`` | | 9-language layers, 9-relationship layers, and 12-cross-modality layers |
| | | | 768-hidden, 12-heads (for each layer) ~ 228M parameters |
| | | | Starting from lxmert-base checkpoint, trained on over 9 million image-text couplets from COCO, VisualGenome, GQA, VQA |
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+

21
model_cards/uncnlp/lxmert-base-uncased/LICENSE Normal file
View File

@@ -0,0 +1,21 @@
MIT License
Copyright (c) 2019 Hao Tan
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

34
model_cards/uncnlp/lxmert-base-uncased/README.md Normal file
View File

@@ -0,0 +1,34 @@
# LXMERT
## Model Description
[LXMERT](https://arxiv.org/abs/1908.07490) is a pre-trained multimodal transformer. The model takes an image and a sentence as input and compute cross-modal representions. The model is converted from [LXMERT github](https://github.com/airsplay/lxmert) by [Antonio Mendoza](https://avmendoza.info/) and is authored by [Hao Tan](https://www.cs.unc.edu/~airsplay/).
![](./lxmert_model-1.jpg?raw=True)
## Usage
## Training Data and Prodcedure
The model is jointly trained on multiple vision-and-language datasets.
We included two image captioning datsets (i.e., [MS COCO](http://cocodataset.org/#home), [Visual Genome](https://visualgenome.org/)) and three image-question answering datasets (i.e., [VQA](https://visualqa.org/), [GQA](https://cs.stanford.edu/people/dorarad/gqa/), [VG QA](https://github.com/yukezhu/visual7w-toolkit)). The model is pre-trained on the above datasets for 20 epochs (roughly 670K iterations with batch size 256), which takes around 8 days on 4 Titan V cards. The details of training could be found in the [LXMERT paper](https://arxiv.org/pdf/1908.07490.pdf).
## Eval Results
| Split | [VQA](https://visualqa.org/) | [GQA](https://cs.stanford.edu/people/dorarad/gqa/) | [NLVR2](http://lil.nlp.cornell.edu/nlvr/) |
|----------- |:----: |:---: |:------:|
| Local Validation | 69.90% | 59.80% | 74.95% |
| Test-Dev | 72.42% | 60.00% | 74.45% (Test-P) |
| Test-Standard | 72.54% | 60.33% | 76.18% (Test-U) |
## Reference
```bibtex
@inproceedings{tan2019lxmert,
title={LXMERT: Learning Cross-Modality Encoder Representations from Transformers},
author={Tan, Hao and Bansal, Mohit},
booktitle={Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing},
year={2019}
}
```

BIN
model_cards/uncnlp/lxmert-base-uncased/lxmert_model-1.jpg Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 275 KiB

19
src/transformers/__init__.py
View File

@@ -31,6 +31,7 @@ from .configuration_encoder_decoder import EncoderDecoderConfig
from .configuration_flaubert import FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, FlaubertConfig
from .configuration_gpt2 import GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP, GPT2Config
from .configuration_longformer import LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, LongformerConfig
from .configuration_lxmert import LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP, LxmertConfig
from .configuration_marian import MarianConfig
from .configuration_mbart import MBartConfig
from .configuration_mmbt import MMBTConfig
@@ -156,6 +157,7 @@ from .tokenization_electra import ElectraTokenizer, ElectraTokenizerFast
from .tokenization_flaubert import FlaubertTokenizer
from .tokenization_gpt2 import GPT2Tokenizer, GPT2TokenizerFast
from .tokenization_longformer import LongformerTokenizer, LongformerTokenizerFast
from .tokenization_lxmert import LxmertTokenizer, LxmertTokenizerFast
from .tokenization_mbart import MBartTokenizer
from .tokenization_mobilebert import MobileBertTokenizer, MobileBertTokenizerFast
from .tokenization_openai import OpenAIGPTTokenizer, OpenAIGPTTokenizerFast
@@ -343,6 +345,15 @@ if is_torch_available():
LongformerModel,
LongformerSelfAttention,
)
from .modeling_lxmert import (
LxmertEncoder,
LxmertForPreTraining,
LxmertForQuestionAnswering,
LxmertModel,
LxmertPreTrainedModel,
LxmertVisualFeatureEncoder,
LxmertXLayer,
)
from .modeling_marian import MarianMTModel
from .modeling_mbart import MBartForConditionalGeneration
from .modeling_mmbt import MMBTForClassification, MMBTModel, ModalEmbeddings
@@ -573,6 +584,14 @@ if is_tf_available():
TFLongformerModel,
TFLongformerSelfAttention,
)
from .modeling_tf_lxmert import (
TF_LXMERT_PRETRAINED_MODEL_ARCHIVE_LIST,
TFLxmertForPreTraining,
TFLxmertMainLayer,
TFLxmertModel,
TFLxmertPreTrainedModel,
TFLxmertVisualFeatureEncoder,
)
from .modeling_tf_mobilebert import (
TF_MOBILEBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
TFMobileBertForMaskedLM,

10
src/transformers/commands/convert.py
View File

@@ -155,5 +155,13 @@ class ConvertCommand(BaseTransformersCLICommand):
)
convert_xlm_checkpoint_to_pytorch(self._tf_checkpoint, self._pytorch_dump_output)
elif self._model_type == "lxmert":
from transformers.convert_lxmert_original_pytorch_checkpoint_to_pytorch import (
convert_lxmert_checkpoint_to_pytorch,
)
convert_lxmert_checkpoint_to_pytorch(self._tf_checkpoint, self._pytorch_dump_output)
else:
raise ValueError("--model_type should be selected in the list [bert, gpt, gpt2, transfo_xl, xlnet, xlm]")
raise ValueError(
"--model_type should be selected in the list [bert, gpt, gpt2, transfo_xl, xlnet, xlm, lxmert]"
)

6
src/transformers/configuration_auto.py
View File

@@ -28,6 +28,7 @@ from .configuration_encoder_decoder import EncoderDecoderConfig
from .configuration_flaubert import FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, FlaubertConfig
from .configuration_gpt2 import GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP, GPT2Config
from .configuration_longformer import LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, LongformerConfig
from .configuration_lxmert import LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP, LxmertConfig
from .configuration_marian import MarianConfig
from .configuration_mbart import MBART_PRETRAINED_CONFIG_ARCHIVE_MAP, MBartConfig
from .configuration_mobilebert import MobileBertConfig
@@ -66,6 +67,7 @@ ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = dict(
ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP,
LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP,
RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
]
for key, value, in pretrained_map.items()
)
@@ -166,6 +168,10 @@ CONFIG_MAPPING = OrderedDict(
"encoder-decoder",
EncoderDecoderConfig,
),
(
"lxmert",
LxmertConfig,
),
]
)

179
src/transformers/configuration_lxmert.py Normal file
View File

@@ -0,0 +1,179 @@
# coding=utf-8
# Copyright 2018, Hao Tan, Mohit Bansal
#
# 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
# limitations under the License.
""" LXMERT model configuration """
import logging
from .configuration_utils import PretrainedConfig
logger = logging.getLogger(__name__)
LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
"unc-nlp/lxmert-base-uncased": "",
}
class LxmertConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a :class:`~transformers.BertModel`.
It is used to instantiate an Lxmert model according to the specified arguments, defining the model
architecture.
Args:
vocab_size (:obj:`int`, optional, defaults to 30522):
Vocabulary size of the BERT model. Defines the different tokens that
can be represented by the `inputs_ids` passed to the forward method of :class:`~transformers.BertModel`.
hidden_size (:obj:`int`, optional, defaults to 768):
Dimensionality of the encoder layers and the pooler layer.
r_layers (:obj:`int`, optional, defaults to 5):
Number of hidden layers in the Transformer visual encoder.
l_layers (:obj:`int`, optional, defaults to 9):
Number of hidden layers in the Transformer language encoder.
x_layers (:obj:`int`, optional, defaults to 5):
Number of hidden layers in the Transformer cross modality encoder.
num_attention_heads (:obj:`int`, optional, defaults to 5):
Number of attention heads for each attention layer in the Transformer encoder.
intermediate_size (:obj:`int`, optional, defaults to 3072):
Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
hidden_act (:obj:`str` or :obj:`function`, optional, defaults to "gelu"):
The non-linear activation function (function or string) in the encoder and pooler.
If string, "gelu", "relu", "swish" and "gelu_new" are supported.
hidden_dropout_prob (:obj:`float`, optional, defaults to 0.1):
The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
attention_probs_dropout_prob (:obj:`float`, optional, defaults to 0.1):
The dropout ratio for the attention probabilities.
max_position_embeddings (:obj:`int`, optional, defaults to 512):
The maximum sequence length that this model might ever be used with.
Typically set this to something large just in case (e.g., 512 or 1024 or 2048).
type_vocab_size (:obj:`int`, optional, defaults to 2):
The vocabulary size of the `token_type_ids` passed into :class:`~transformers.BertModel`.
initializer_range (:obj:`float`, optional, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
layer_norm_eps (:obj:`float`, optional, defaults to 1e-12):
The epsilon used by the layer normalization layers.
visual_feat_dim (:obj:`int`, optional, defaults to 2048):
This represents the last dimension of the pooled-object features used as input for the model,
representing the size of each object feature itself.
visual_pos_dim (:obj:`int`, optional, defaults to 4):
This represents the number of spacial features that are mixed into the visual features.
The default is set to 4 because most commonly this will represent the location of a bounding box.
i.e. (x, y, width, height)
visual_loss_normalizer (:obj:`float`, optional, defaults to 1/15):
This represents the scaling factor in which each visual loss is multiplied by if during pretraining,
one decided to train with multiple vision-based loss objectives.
num_qa_labels (:obj:`int`, optional, defaults to 9500):
This represents the total number of different question answering (QA) labels there are. If using more than one dataset with QA,
the user will need to account for the total number of labels that all of the datasets have in total.
num_object_labels (:obj:`int`, optional, defaults to 1600):
This represents the total number of semantically unique objects that lxmert will be able to classify a pooled-object feature
as belonging too.
num_attr_labels (:obj:`int`, optional, defaults to 400):
This represents the total number of semantically unique attributes that lxmert will be able to classify a pooled-object feature
as possessing.
task_matched (:obj:`bool`, optional, defaults to True):
This task is used for sentence-image matching. If the sentence correctly describes the image the label will be 1.
If the sentence does not correctly describe the image, the label will be 0.
task_mask_lm (:obj:`bool`, optional, defaults to True):
This task is the defacto masked langauge modeling used in pretraining models such as BERT.
task_obj_predict (:obj:`bool`, optional, defaults to True):
This task is set to true if the user would like to perform one of the following loss objectives:
object predicition, atrribute predicition, feature regression
task_qa (:obj:`bool`, optional, defaults to True):
This task specifies whether or not Lxmert will calculate the question-asnwering loss objective
visual_obj_loss (:obj:`bool`, optional, defaults to True):
This task specifies whether or not Lxmert will calculate the object-prediction loss objective
visual_attr_loss (:obj:`bool`, optional, defaults to True):
This task specifies whether or not Lxmert will calculate the attribute-prediction loss objective
visual_feat_loss (:obj:`bool`, optional, defaults to True):
This task specifies whether or not Lxmert will calculate the feature-regression loss objective
output_attentions (:obj:`bool`, optional, defaults to False):
if True, the vision, langauge, and cross-modality layers will be returned
output_hidden_states (:obj:`bool`, optional, defaults to False):
if True, final cross-modality hidden states for language and vision features will be returned
"""
model_type = "lxmert"
def __init__(
self,
vocab_size=30522,
hidden_size=768,
num_attention_heads=12,
num_labels=2,
num_qa_labels=9500,
num_object_labels=1600,
num_attr_labels=400,
intermediate_size=3072,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=512,
type_vocab_size=2,
initializer_range=0.02,
layer_norm_eps=1e-12,
pad_token_id=0,
l_layers=9,
x_layers=5,
r_layers=5,
visual_feat_dim=2048,
visual_pos_dim=4,
visual_loss_normalizer=6.67,
task_matched=True,
task_mask_lm=True,
task_obj_predict=True,
task_qa=True,
visual_obj_loss=True,
visual_attr_loss=True,
visual_feat_loss=True,
output_attentions=False,
output_hidden_states=False,
**kwargs,
):
super().__init__(**kwargs)
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_attention_heads = num_attention_heads
self.num_labels = num_labels
self.hidden_act = hidden_act
self.intermediate_size = intermediate_size
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.type_vocab_size = type_vocab_size
self.initializer_range = initializer_range
self.layer_norm_eps = layer_norm_eps
self.num_qa_labels = num_qa_labels
self.num_object_labels = num_object_labels
self.num_attr_labels = num_attr_labels
self.l_layers = l_layers
self.x_layers = x_layers
self.r_layers = r_layers
self.visual_feat_dim = visual_feat_dim
self.visual_pos_dim = visual_pos_dim
self.visual_loss_normalizer = visual_loss_normalizer
self.task_matched = task_matched
self.task_mask_lm = task_mask_lm
self.task_obj_predict = task_obj_predict
self.task_qa = task_qa
self.visual_obj_loss = visual_obj_loss
self.visual_attr_loss = visual_attr_loss
self.visual_feat_loss = visual_feat_loss
self.output_hidden_states = output_hidden_states
self.output_attentions = self.output_attentions
self.num_hidden_layers = {"vision": r_layers, "cross_encoder": x_layers, "language": l_layers}

61
src/transformers/convert_lxmert_original_tf_checkpoint_to_pytorch.py Executable file
View File

@@ -0,0 +1,61 @@
# coding=utf-8
# Copyright 2018 The HuggingFace Inc. team.
#
# 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
# limitations under the License.
"""Convert LXMERT checkpoint."""
import argparse
import logging
import torch
from transformers import LxmertConfig, LxmertForPreTraining, load_tf_weights_in_lxmert
logging.basicConfig(level=logging.INFO)
def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, config_file, pytorch_dump_path):
# Initialise PyTorch model
config = LxmertConfig.from_json_file(config_file)
print("Building PyTorch model from configuration: {}".format(str(config)))
model = LxmertForPreTraining(config)
# Load weights from tf checkpoint
load_tf_weights_in_lxmert(model, config, tf_checkpoint_path)
# Save pytorch-model
print("Save PyTorch model to {}".format(pytorch_dump_path))
torch.save(model.state_dict(), pytorch_dump_path)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--tf_checkpoint_path", default=None, type=str, required=True, help="Path to the TensorFlow checkpoint path."
)
parser.add_argument(
"--config_file",
default=None,
type=str,
required=True,
help="The config json file corresponding to the pre-trained model. \n"
"This specifies the model architecture.",
)
parser.add_argument(
"--pytorch_dump_path", default=None, type=str, required=True, help="Path to the output PyTorch model."
)
args = parser.parse_args()
convert_tf_checkpoint_to_pytorch(args.tf_checkpoint_path, args.config_file, args.pytorch_dump_path)

18
src/transformers/convert_pytorch_checkpoint_to_tf2.py
View File

@@ -27,6 +27,7 @@ from transformers import (
ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP,
FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP,
LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP,
ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
T5_PRETRAINED_CONFIG_ARCHIVE_MAP,
@@ -43,6 +44,7 @@ from transformers import (
ElectraConfig,
FlaubertConfig,
GPT2Config,
LxmertConfig,
OpenAIGPTConfig,
RobertaConfig,
T5Config,
@@ -57,6 +59,8 @@ from transformers import (
TFElectraForPreTraining,
TFFlaubertWithLMHeadModel,
TFGPT2LMHeadModel,
TFLxmertForPreTraining,
TFLxmertVisualFeatureEncoder,
TFOpenAIGPTLMHeadModel,
TFRobertaForMaskedLM,
TFRobertaForSequenceClassification,
@@ -94,6 +98,8 @@ if is_torch_available():
ElectraForPreTraining,
FlaubertWithLMHeadModel,
GPT2LMHeadModel,
LxmertForPreTraining,
LxmertVisualFeatureEncoder,
OpenAIGPTLMHeadModel,
RobertaForMaskedLM,
RobertaForSequenceClassification,
@@ -204,6 +210,18 @@ MODEL_CLASSES = {
DistilBertForQuestionAnswering,
DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
),
"lxmert": (
LxmertConfig,
TFLxmertForPreTraining,
LxmertForPreTraining,
LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
),
"lxmert-visual-feature-encoder": (
LxmertConfig,
TFLxmertVisualFeatureEncoder,
LxmertVisualFeatureEncoder,
LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
),
"ctrl": (
CTRLConfig,
TFCTRLLMHeadModel,

4
src/transformers/modeling_auto.py
View File

@@ -31,6 +31,7 @@ from .configuration_auto import (
FlaubertConfig,
GPT2Config,
LongformerConfig,
LxmertConfig,
MBartConfig,
MobileBertConfig,
OpenAIGPTConfig,
@@ -116,6 +117,7 @@ from .modeling_longformer import (
LongformerForTokenClassification,
LongformerModel,
)
from .modeling_lxmert import LxmertForPreTraining, LxmertModel
from .modeling_marian import MarianMTModel
from .modeling_mbart import MBartForConditionalGeneration
from .modeling_mobilebert import (
@@ -200,6 +202,7 @@ MODEL_MAPPING = OrderedDict(
(CTRLConfig, CTRLModel),
(ElectraConfig, ElectraModel),
(ReformerConfig, ReformerModel),
(LxmertConfig, LxmertModel),
]
)
@@ -224,6 +227,7 @@ MODEL_FOR_PRETRAINING_MAPPING = OrderedDict(
(XLMConfig, XLMWithLMHeadModel),
(CTRLConfig, CTRLLMHeadModel),
(ElectraConfig, ElectraForPreTraining),
(LxmertConfig, LxmertForPreTraining),
]
)

1426
src/transformers/modeling_lxmert.py Normal file
View File

File diff suppressed because it is too large Load Diff

1378
src/transformers/modeling_tf_lxmert.py Normal file
View File

File diff suppressed because it is too large Load Diff

12
src/transformers/modeling_tf_mobilebert.py
View File

@@ -883,7 +883,7 @@ MOBILEBERT_START_DOCSTRING = r"""
MOBILEBERT_INPUTS_DOCSTRING = r"""
Args:
input_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`):
input_ids (:obj:`np.ndarray` or :obj:`tf.Tensor` of shape :obj:`{0}`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using :class:`transformers.MobileBertTokenizer`.
@@ -891,28 +891,28 @@ MOBILEBERT_INPUTS_DOCSTRING = r"""
:func:`transformers.PreTrainedTokenizer.__call__` for details.
`What are input IDs? <../glossary.html#input-ids>`__
attention_mask (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
attention_mask (:obj:`np.ndarray` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
Mask to avoid performing attention on padding token indices.
Mask values selected in ``[0, 1]``:
``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
`What are attention masks? <../glossary.html#attention-mask>`__
token_type_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
token_type_ids (:obj:`np.ndarray` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
Segment token indices to indicate first and second portions of the inputs.
Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
corresponds to a `sentence B` token
`What are token type IDs? <../glossary.html#token-type-ids>`__
position_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
position_ids (:obj:`np.ndarray` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
Indices of positions of each input sequence tokens in the position embeddings.
Selected in the range ``[0, config.max_position_embeddings - 1]``.
`What are position IDs? <../glossary.html#position-ids>`__
head_mask (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`, defaults to :obj:`None`):
head_mask (:obj:`np.ndarray` or :obj:`tf.Tensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`, defaults to :obj:`None`):
Mask to nullify selected heads of the self-attention modules.
Mask values selected in ``[0, 1]``:
:obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
inputs_embeds (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, embedding_dim)`, `optional`, defaults to :obj:`None`):
inputs_embeds (:obj:`np.ndarray` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, embedding_dim)`, `optional`, defaults to :obj:`None`):
Optionally, instead of passing :obj:`input_ids` you can to directly pass an embedded representation.
This is useful if you want more control over how to convert `input_ids` indices into associated vectors
than the model's internal embedding lookup matrix.

2
src/transformers/modeling_tf_utils.py
View File

@@ -191,7 +191,7 @@ class TFSequenceClassificationLoss:
"""
def compute_loss(self, labels, logits):
if shape_list(logits)[1] == 1:
if len(shape_list(logits)) == 1 or shape_list(logits)[1] == 1:
loss_fn = tf.keras.losses.MeanSquaredError(reduction=tf.keras.losses.Reduction.NONE)
else:
loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(

5
src/transformers/tokenization_auto.py
View File

@@ -29,6 +29,7 @@ from .configuration_auto import (
FlaubertConfig,
GPT2Config,
LongformerConfig,
LxmertConfig,
MarianConfig,
MBartConfig,
MobileBertConfig,
@@ -55,6 +56,7 @@ from .tokenization_electra import ElectraTokenizer, ElectraTokenizerFast
from .tokenization_flaubert import FlaubertTokenizer
from .tokenization_gpt2 import GPT2Tokenizer, GPT2TokenizerFast
from .tokenization_longformer import LongformerTokenizer, LongformerTokenizerFast
from .tokenization_lxmert import LxmertTokenizer, LxmertTokenizerFast
from .tokenization_marian import MarianTokenizer
from .tokenization_mbart import MBartTokenizer
from .tokenization_mobilebert import MobileBertTokenizer, MobileBertTokenizerFast
@@ -91,6 +93,7 @@ TOKENIZER_MAPPING = OrderedDict(
(RobertaConfig, (RobertaTokenizer, RobertaTokenizerFast)),
(ReformerConfig, (ReformerTokenizer, None)),
(ElectraConfig, (ElectraTokenizer, ElectraTokenizerFast)),
(LxmertConfig, (LxmertTokenizer, LxmertTokenizerFast)),
(BertConfig, (BertTokenizer, BertTokenizerFast)),
(OpenAIGPTConfig, (OpenAIGPTTokenizer, OpenAIGPTTokenizerFast)),
(GPT2Config, (GPT2Tokenizer, GPT2TokenizerFast)),
@@ -128,6 +131,7 @@ class AutoTokenizer:
- `xlm`: XLMTokenizer (XLM model)
- `ctrl`: CTRLTokenizer (Salesforce CTRL model)
- `electra`: ElectraTokenizer (Google ELECTRA model)
- `lxmert`: LxmertTokenizer (Lxmert model)
This class cannot be instantiated using `__init__()` (throw an error).
"""
@@ -163,6 +167,7 @@ class AutoTokenizer:
- `xlm`: XLMTokenizer (XLM model)
- `ctrl`: CTRLTokenizer (Salesforce CTRL model)
- `electra`: ElectraTokenizer (Google ELECTRA model)
- `lxmert`: LxmertTokenizer (Lxmert model)
Params:
pretrained_model_name_or_path: either:

80
src/transformers/tokenization_lxmert.py Normal file
View File

@@ -0,0 +1,80 @@
# coding=utf-8
# Copyright 2020 The Google AI Team, Stanford University and The HuggingFace Inc. team.
#
# 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
# limitations under the License.
from .tokenization_bert import BertTokenizer, BertTokenizerFast
####################################################
# Mapping from the keyword arguments names of Tokenizer `__init__`
# to file names for serializing Tokenizer instances
####################################################
VOCAB_FILES_NAMES = {"vocab_file": "vocab.txt"}
####################################################
# Mapping from the keyword arguments names of Tokenizer `__init__`
# to pretrained vocabulary URL for all the model shortcut names.
####################################################
PRETRAINED_VOCAB_FILES_MAP = {
"vocab_file": {
"unc-nlp/lxmert-base-uncased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased-vocab.txt",
}
}
####################################################
# Mapping from model shortcut names to max length of inputs
####################################################
PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
"unc-nlp/lxmert-base-uncased": 512,
}
####################################################
# Mapping from model shortcut names to a dictionary of additional
# keyword arguments for Tokenizer `__init__`.
# To be used for checkpoint specific configurations.
####################################################
PRETRAINED_INIT_CONFIGURATION = {
"unc-nlp/lxmert-base-uncased": {"do_lower_case": True},
}
class LxmertTokenizer(BertTokenizer):
r"""
Constructs an Lxmert tokenizer.
:class:`~transformers.LxmertTokenizer` is identical to :class:`~transformers.BertTokenizer` and runs end-to-end
tokenization: punctuation splitting + wordpiece.
Refer to superclass :class:`~transformers.BertTokenizer` for usage examples and documentation concerning
parameters.
"""
vocab_files_names = VOCAB_FILES_NAMES
pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
pretrained_init_configuration = PRETRAINED_INIT_CONFIGURATION
class LxmertTokenizerFast(BertTokenizerFast):
r"""
Constructs a "Fast" Lxmert Fast tokenizer (backed by HuggingFace's `tokenizers` library).
:class:`~transformers.LxmertTokenizerFast` is identical to :class:`~transformers.BertTokenizerFast` and runs end-to-end
tokenization: punctuation splitting + wordpiece.
Refer to superclass :class:`~transformers.BertTokenizerFast` for usage examples and documentation concerning
parameters.
"""
vocab_files_names = VOCAB_FILES_NAMES
pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
pretrained_init_configuration = PRETRAINED_INIT_CONFIGURATION

684
tests/test_modeling_lxmert.py Normal file
View File

@@ -0,0 +1,684 @@
# coding=utf-8
# Copyright 2018 LXMERT Authors.
#
# 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
# limitations under the License.
import unittest
from transformers import is_torch_available
from transformers.testing_utils import require_torch, slow, torch_device
from .test_configuration_common import ConfigTester
from .test_modeling_common import ModelTesterMixin, ids_tensor
if is_torch_available():
import torch
from transformers import LxmertConfig, LxmertForPreTraining, LxmertForQuestionAnswering, LxmertModel
from transformers.modeling_lxmert import LXMERT_PRETRAINED_MODEL_ARCHIVE_LIST
class LxmertModelTester:
"""You can also import this e.g from .test_modeling_bart import BartModelTester """
def __init__(
self,
parent,
vocab_size=300,
hidden_size=28,
num_attention_heads=2,
num_labels=2,
intermediate_size=64,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=512,
type_vocab_size=2,
initializer_range=0.02,
layer_norm_eps=1e-12,
pad_token_id=0,
num_qa_labels=30,
num_object_labels=16,
num_attr_labels=4,
num_visual_features=10,
l_layers=2,
x_layers=1,
r_layers=1,
visual_feat_dim=128,
visual_pos_dim=4,
visual_loss_normalizer=6.67,
seq_length=20,
batch_size=4,
is_training=True,
task_matched=True,
task_mask_lm=True,
task_obj_predict=True,
task_qa=True,
visual_obj_loss=True,
visual_attr_loss=True,
visual_feat_loss=True,
use_token_type_ids=True,
use_lang_mask=True,
output_attentions=False,
output_hidden_states=False,
scope=None,
):
self.parent = parent
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_attention_heads = num_attention_heads
self.num_labels = num_labels
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.type_vocab_size = type_vocab_size
self.initializer_range = initializer_range
self.layer_norm_eps = layer_norm_eps
self.pad_token_id = pad_token_id
self.num_qa_labels = num_qa_labels
self.num_object_labels = num_object_labels
self.num_attr_labels = num_attr_labels
self.l_layers = l_layers
self.x_layers = x_layers
self.r_layers = r_layers
self.visual_feat_dim = visual_feat_dim
self.visual_pos_dim = visual_pos_dim
self.visual_loss_normalizer = visual_loss_normalizer
self.seq_length = seq_length
self.batch_size = batch_size
self.is_training = is_training
self.use_lang_mask = use_lang_mask
self.task_matched = task_matched
self.task_mask_lm = task_mask_lm
self.task_obj_predict = task_obj_predict
self.task_qa = task_qa
self.visual_obj_loss = visual_obj_loss
self.visual_attr_loss = visual_attr_loss
self.visual_feat_loss = visual_feat_loss
self.num_visual_features = num_visual_features
self.use_token_type_ids = use_token_type_ids
self.output_attentions = output_attentions
self.output_hidden_states = output_hidden_states
self.scope = scope
self.num_hidden_layers = {"vision": r_layers, "cross_encoder": x_layers, "language": l_layers}
def prepare_config_and_inputs(self):
output_attentions = self.output_attentions
input_ids = ids_tensor([self.batch_size, self.seq_length], vocab_size=self.vocab_size)
visual_feats = torch.rand(self.batch_size, self.num_visual_features, self.visual_feat_dim)
bounding_boxes = torch.rand(self.batch_size, self.num_visual_features, 4)
input_mask = None
if self.use_lang_mask:
input_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
token_type_ids = None
if self.use_token_type_ids:
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
obj_labels = None
if self.task_obj_predict:
obj_labels = {}
if self.visual_attr_loss and self.task_obj_predict:
obj_labels["attr"] = (
ids_tensor([self.batch_size, self.num_visual_features], self.num_attr_labels),
ids_tensor([self.batch_size, self.num_visual_features], self.num_attr_labels),
)
if self.visual_feat_loss and self.task_obj_predict:
obj_labels["feat"] = (
ids_tensor(
[self.batch_size, self.num_visual_features, self.visual_feat_dim], self.num_visual_features
),
ids_tensor([self.batch_size, self.num_visual_features], self.num_visual_features),
)
if self.visual_obj_loss and self.task_obj_predict:
obj_labels["obj"] = (
ids_tensor([self.batch_size, self.num_visual_features], self.num_object_labels),
ids_tensor([self.batch_size, self.num_visual_features], self.num_object_labels),
)
ans = None
if self.task_qa:
ans = ids_tensor([self.batch_size], self.num_qa_labels)
masked_lm_labels = None
if self.task_mask_lm:
masked_lm_labels = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
matched_label = None
if self.task_matched:
matched_label = ids_tensor([self.batch_size], self.num_labels)
config = LxmertConfig(
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
num_attention_heads=self.num_attention_heads,
num_labels=self.num_labels,
intermediate_size=self.intermediate_size,
hidden_act=self.hidden_act,
hidden_dropout_prob=self.hidden_dropout_prob,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
type_vocab_size=self.type_vocab_size,
initializer_range=self.initializer_range,
layer_norm_eps=self.layer_norm_eps,
pad_token_id=self.pad_token_id,
num_qa_labels=self.num_qa_labels,
num_object_labels=self.num_object_labels,
num_attr_labels=self.num_attr_labels,
l_layers=self.l_layers,
x_layers=self.x_layers,
r_layers=self.r_layers,
visual_feat_dim=self.visual_feat_dim,
visual_pos_dim=self.visual_pos_dim,
visual_loss_normalizer=self.visual_loss_normalizer,
task_matched=self.task_matched,
task_mask_lm=self.task_mask_lm,
task_obj_predict=self.task_obj_predict,
task_qa=self.task_qa,
visual_obj_loss=self.visual_obj_loss,
visual_attr_loss=self.visual_attr_loss,
visual_feat_loss=self.visual_feat_loss,
output_attentions=self.output_attentions,
output_hidden_states=self.output_hidden_states,
)
return (
config,
input_ids,
visual_feats,
bounding_boxes,
token_type_ids,
input_mask,
obj_labels,
masked_lm_labels,
matched_label,
ans,
output_attentions,
)
def create_and_check_lxmert_model(
self,
config,
input_ids,
visual_feats,
bounding_boxes,
token_type_ids,
input_mask,
obj_labels,
masked_lm_labels,
matched_label,
ans,
output_attentions,
):
model = LxmertModel(config=config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
output_attentions=output_attentions,
)
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
output_attentions=not output_attentions,
)
result = model(input_ids, visual_feats, bounding_boxes, return_dict=False)
result = model(input_ids, visual_feats, bounding_boxes, return_dict=True)
self.parent.assertEqual(result.language_output.shape, (self.batch_size, self.seq_length, self.hidden_size))
self.parent.assertEqual(
result.vision_output.shape, (self.batch_size, self.num_visual_features, self.hidden_size)
)
self.parent.assertEqual(result.pooled_output.shape, (self.batch_size, self.hidden_size))
def create_and_check_lxmert_for_question_answering(
self,
config,
input_ids,
visual_feats,
bounding_boxes,
token_type_ids,
input_mask,
obj_labels,
masked_lm_labels,
matched_label,
ans,
output_attentions,
):
model = LxmertForQuestionAnswering(config=config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
labels=ans,
output_attentions=output_attentions,
return_dict=True,
)
result = model(input_ids, visual_feats, bounding_boxes, labels=ans)
result = model(
input_ids,
visual_feats,
bounding_boxes,
labels=ans,
token_type_ids=token_type_ids,
attention_mask=input_mask,
output_attentions=output_attentions,
)
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
labels=ans,
output_attentions=not output_attentions,
return_dict=True,
)
self.parent.assertEqual(result.question_answering_score.shape, (self.batch_size, self.num_qa_labels))
def create_and_check_lxmert_for_pretraining(
self,
config,
input_ids,
visual_feats,
bounding_boxes,
token_type_ids,
input_mask,
obj_labels,
masked_lm_labels,
matched_label,
ans,
output_attentions,
):
model = LxmertForPreTraining(config=config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
masked_lm_labels=masked_lm_labels,
obj_labels=obj_labels,
matched_label=matched_label,
ans=ans,
output_attentions=output_attentions,
return_dict=True,
)
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
masked_lm_labels=masked_lm_labels,
output_attentions=not output_attentions,
return_dict=False,
)
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
masked_lm_labels=masked_lm_labels,
)
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
obj_labels=obj_labels,
)
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
matched_label=matched_label,
)
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
ans=ans,
)
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
masked_lm_labels=masked_lm_labels,
obj_labels=obj_labels,
matched_label=matched_label,
ans=ans,
output_attentions=not output_attentions,
return_dict=True,
)
self.parent.assertEqual(result.prediction_logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def resize_lxmert_num_qa_labels(
self,
config,
input_ids,
visual_feats,
bounding_boxes,
token_type_ids,
input_mask,
obj_labels,
masked_lm_labels,
matched_label,
ans,
output_attentions,
):
start_labels = config.num_qa_labels
num_large_labels = config.num_qa_labels * 2
num_small_labels = int(config.num_qa_labels * 2)
less_labels_ans = ids_tensor([self.batch_size], num_small_labels)
more_labels_ans = ids_tensor([self.batch_size], num_large_labels)
model_pretrain = LxmertForPreTraining(config=config)
model_qa = LxmertForQuestionAnswering(config=config)
config.num_labels = num_small_labels
end_labels = config.num_labels
result_pretrain = model_pretrain(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
ans=ans,
return_dict=True,
)
result_qa = model_qa(
input_ids,
visual_feats,
bounding_boxes,
labels=ans,
token_type_ids=token_type_ids,
attention_mask=input_mask,
return_dict=True,
)
model_pretrain.resize_num_qa_labels(num_small_labels)
model_qa.resize_num_qa_labels(num_small_labels)
result_pretrain_less = model_pretrain(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
ans=less_labels_ans,
return_dict=True,
)
result_qa_less = model_qa(
input_ids,
visual_feats,
bounding_boxes,
labels=less_labels_ans,
token_type_ids=token_type_ids,
attention_mask=input_mask,
return_dict=True,
)
model_pretrain.resize_num_qa_labels(num_large_labels)
model_qa.resize_num_qa_labels(num_large_labels)
result_pretrain_more = model_pretrain(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
ans=more_labels_ans,
return_dict=True,
)
result_qa_more = model_qa(
input_ids,
visual_feats,
bounding_boxes,
labels=more_labels_ans,
token_type_ids=token_type_ids,
attention_mask=input_mask,
return_dict=True,
)
model_qa_labels = model_qa.num_qa_labels
self.parent.assertNotEqual(start_labels, end_labels)
self.parent.assertNotEqual(model_qa_labels, start_labels)
self.parent.assertEqual(result_qa.question_answering_score.shape, (self.batch_size, start_labels))
self.parent.assertEqual(result_pretrain.question_answering_score.shape, (self.batch_size, start_labels))
self.parent.assertEqual(result_qa_less.question_answering_score.shape, (self.batch_size, num_small_labels))
self.parent.assertEqual(
result_pretrain_less.question_answering_score.shape, (self.batch_size, num_small_labels)
)
self.parent.assertEqual(result_qa_more.question_answering_score.shape, (self.batch_size, num_large_labels))
self.parent.assertEqual(
result_pretrain_more.question_answering_score.shape, (self.batch_size, num_large_labels)
)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
config,
input_ids,
visual_feats,
bounding_boxes,
token_type_ids,
input_mask,
obj_labels,
masked_lm_labels,
matched_label,
ans,
output_attentions,
) = config_and_inputs
inputs_dict = {
"input_ids": input_ids,
"visual_feats": visual_feats,
"visual_pos": bounding_boxes,
"token_type_ids": token_type_ids,
"attention_mask": input_mask,
}
return config, inputs_dict
@require_torch
class LxmertModelTest(ModelTesterMixin, unittest.TestCase):
all_model_classes = (LxmertModel, LxmertForPreTraining, LxmertForQuestionAnswering) if is_torch_available() else ()
test_head_masking = False
test_pruning = False
test_torchscript = False
test_head_masking = False
test_pruning = False
test_torchscript = False
def setUp(self):
self.model_tester = LxmertModelTester(self)
self.config_tester = ConfigTester(self, config_class=LxmertConfig, hidden_size=37)
def test_config(self):
self.config_tester.run_common_tests()
def test_lxmert_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_lxmert_model(*config_and_inputs)
def test_lxmert_question_answering(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_lxmert_for_question_answering(*config_and_inputs)
def test_lxmert_pretraining(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_lxmert_for_pretraining(*config_and_inputs)
def test_lxmert_question_answering_labels_resize(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.resize_lxmert_num_qa_labels(*config_and_inputs)
@slow
def test_model_from_pretrained(self):
for model_name in LXMERT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
model = LxmertModel.from_pretrained(model_name)
self.assertIsNotNone(model)
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
seq_len = getattr(self.model_tester, "seq_length", None)
encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
chunk_length = getattr(self.model_tester, "chunk_length", None)
if chunk_length is not None and hasattr(self.model_tester, "num_hashes"):
encoder_seq_length = encoder_seq_length * self.model_tester.num_hashes
for model_class in self.all_model_classes:
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = False
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
language_attentions, vision_attentions, cross_encoder_attentions = (outputs[-3], outputs[-2], outputs[-1])
self.assertEqual(len(language_attentions), self.model_tester.num_hidden_layers["language"])
self.assertEqual(len(vision_attentions), self.model_tester.num_hidden_layers["vision"])
self.assertEqual(len(cross_encoder_attentions), self.model_tester.num_hidden_layers["cross_encoder"])
# check that output_attentions also work using config
del inputs_dict["output_attentions"]
config.output_attentions = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
language_attentions, vision_attentions, cross_encoder_attentions = (outputs[-3], outputs[-2], outputs[-1])
self.assertEqual(len(language_attentions), self.model_tester.num_hidden_layers["language"])
self.assertEqual(len(vision_attentions), self.model_tester.num_hidden_layers["vision"])
self.assertEqual(len(cross_encoder_attentions), self.model_tester.num_hidden_layers["cross_encoder"])
attentions = [language_attentions, vision_attentions, cross_encoder_attentions]
attention_shapes = [
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
[
self.model_tester.num_attention_heads,
self.model_tester.num_visual_features,
self.model_tester.num_visual_features,
],
[self.model_tester.num_attention_heads, encoder_key_length, self.model_tester.num_visual_features],
]
for attention, attention_shape in zip(attentions, attention_shapes):
self.assertListEqual(list(attention[0].shape[-3:]), attention_shape)
out_len = len(outputs)
# Check attention is always last and order is fine
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
# 2 hidden states were added
self.assertEqual(out_len + 2, len(outputs))
language_attentions, vision_attentions, cross_encoder_attentions = (outputs[-3], outputs[-2], outputs[-1])
self.assertEqual(len(language_attentions), self.model_tester.num_hidden_layers["language"])
self.assertEqual(len(vision_attentions), self.model_tester.num_hidden_layers["vision"])
self.assertEqual(len(cross_encoder_attentions), self.model_tester.num_hidden_layers["cross_encoder"])
attentions = [language_attentions, vision_attentions, cross_encoder_attentions]
attention_shapes = [
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
[
self.model_tester.num_attention_heads,
self.model_tester.num_visual_features,
self.model_tester.num_visual_features,
],
[self.model_tester.num_attention_heads, encoder_key_length, self.model_tester.num_visual_features],
]
for attention, attention_shape in zip(attentions, attention_shapes):
self.assertListEqual(list(attention[0].shape[-3:]), attention_shape)
def test_hidden_states_output(self):
def check_hidden_states_output(inputs_dict, config, model_class):
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
language_hidden_states, vision_hidden_states = outputs[-2], outputs[-1]
self.assertEqual(len(language_hidden_states), self.model_tester.num_hidden_layers["language"] + 1)
self.assertEqual(len(vision_hidden_states), self.model_tester.num_hidden_layers["vision"] + 1)
seq_length = self.model_tester.seq_length
num_visual_features = self.model_tester.num_visual_features
self.assertListEqual(
list(language_hidden_states[0].shape[-2:]),
[seq_length, self.model_tester.hidden_size],
)
self.assertListEqual(
list(vision_hidden_states[0].shape[-2:]),
[num_visual_features, self.model_tester.hidden_size],
)
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
inputs_dict["output_hidden_states"] = True
check_hidden_states_output(inputs_dict, config, model_class)
# check that output_hidden_states also work using config
del inputs_dict["output_hidden_states"]
config.output_hidden_states = True
check_hidden_states_output(inputs_dict, config, model_class)

680
tests/test_modeling_tf_lxmert.py Normal file
View File

@@ -0,0 +1,680 @@
# coding=utf-8
# Copyright 2018 XXX Authors.
#
# 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
# limitations under the License.
import tempfile
import unittest
from transformers import LxmertConfig, is_tf_available
from transformers.testing_utils import require_tf, slow
from .test_configuration_common import ConfigTester
from .test_modeling_tf_common import TFModelTesterMixin, ids_tensor
if is_tf_available():
import tensorflow as tf
from transformers.modeling_tf_lxmert import TFLxmertForPreTraining, TFLxmertModel
class TFLxmertModelTester(object):
def __init__(
self,
parent,
vocab_size=300,
hidden_size=28,
num_attention_heads=2,
num_labels=2,
intermediate_size=64,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=512,
type_vocab_size=2,
initializer_range=0.02,
layer_norm_eps=1e-12,
pad_token_id=0,
num_qa_labels=30,
num_object_labels=16,
num_attr_labels=4,
num_visual_features=10,
l_layers=2,
x_layers=1,
r_layers=1,
visual_feat_dim=128,
visual_pos_dim=4,
visual_loss_normalizer=6.67,
seq_length=20,
batch_size=8,
is_training=True,
task_matched=True,
task_mask_lm=True,
task_obj_predict=True,
task_qa=True,
visual_obj_loss=True,
visual_attr_loss=True,
visual_feat_loss=True,
use_token_type_ids=True,
use_lang_mask=True,
output_attentions=False,
output_hidden_states=False,
scope=None,
):
self.parent = parent
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_attention_heads = num_attention_heads
self.num_labels = num_labels
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.type_vocab_size = type_vocab_size
self.initializer_range = initializer_range
self.layer_norm_eps = layer_norm_eps
self.pad_token_id = pad_token_id
self.num_qa_labels = num_qa_labels
self.num_object_labels = num_object_labels
self.num_attr_labels = num_attr_labels
self.l_layers = l_layers
self.x_layers = x_layers
self.r_layers = r_layers
self.visual_feat_dim = visual_feat_dim
self.visual_pos_dim = visual_pos_dim
self.visual_loss_normalizer = visual_loss_normalizer
self.seq_length = seq_length
self.batch_size = batch_size
self.is_training = is_training
self.use_lang_mask = use_lang_mask
self.task_matched = task_matched
self.task_mask_lm = task_mask_lm
self.task_obj_predict = task_obj_predict
self.task_qa = task_qa
self.visual_obj_loss = visual_obj_loss
self.visual_attr_loss = visual_attr_loss
self.visual_feat_loss = visual_feat_loss
self.num_visual_features = num_visual_features
self.use_token_type_ids = use_token_type_ids
self.output_attentions = output_attentions
self.output_hidden_states = output_hidden_states
self.scope = scope
self.num_hidden_layers = {"vision": r_layers, "cross_encoder": x_layers, "language": l_layers}
def prepare_config_and_inputs(self):
output_attentions = self.output_attentions
input_ids = ids_tensor([self.batch_size, self.seq_length], vocab_size=self.vocab_size)
visual_feats = tf.random.uniform((self.batch_size, self.num_visual_features, self.visual_feat_dim))
bounding_boxes = tf.random.uniform((self.batch_size, self.num_visual_features, 4))
input_mask = None
if self.use_lang_mask:
input_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
token_type_ids = None
if self.use_token_type_ids:
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
obj_labels = None
if self.task_obj_predict:
obj_labels = {}
if self.visual_attr_loss and self.task_obj_predict:
obj_labels["attr"] = (
ids_tensor([self.batch_size, self.num_visual_features], self.num_attr_labels),
ids_tensor([self.batch_size, self.num_visual_features], self.num_attr_labels),
)
if self.visual_feat_loss and self.task_obj_predict:
obj_labels["feat"] = (
ids_tensor(
[self.batch_size, self.num_visual_features, self.visual_feat_dim], self.num_visual_features
),
ids_tensor([self.batch_size, self.num_visual_features], self.num_visual_features),
)
if self.visual_obj_loss and self.task_obj_predict:
obj_labels["obj"] = (
ids_tensor([self.batch_size, self.num_visual_features], self.num_object_labels),
ids_tensor([self.batch_size, self.num_visual_features], self.num_object_labels),
)
ans = None
if self.task_qa:
ans = ids_tensor([self.batch_size], self.num_qa_labels)
masked_lm_labels = None
if self.task_mask_lm:
masked_lm_labels = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
matched_label = None
if self.task_matched:
matched_label = ids_tensor([self.batch_size], self.num_labels)
config = LxmertConfig(
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
num_attention_heads=self.num_attention_heads,
num_labels=self.num_labels,
intermediate_size=self.intermediate_size,
hidden_act=self.hidden_act,
hidden_dropout_prob=self.hidden_dropout_prob,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
type_vocab_size=self.type_vocab_size,
initializer_range=self.initializer_range,
layer_norm_eps=self.layer_norm_eps,
pad_token_id=self.pad_token_id,
num_qa_labels=self.num_qa_labels,
num_object_labels=self.num_object_labels,
num_attr_labels=self.num_attr_labels,
l_layers=self.l_layers,
x_layers=self.x_layers,
r_layers=self.r_layers,
visual_feat_dim=self.visual_feat_dim,
visual_pos_dim=self.visual_pos_dim,
visual_loss_normalizer=self.visual_loss_normalizer,
task_matched=self.task_matched,
task_mask_lm=self.task_mask_lm,
task_obj_predict=self.task_obj_predict,
task_qa=self.task_qa,
visual_obj_loss=self.visual_obj_loss,
visual_attr_loss=self.visual_attr_loss,
visual_feat_loss=self.visual_feat_loss,
output_attentions=self.output_attentions,
output_hidden_states=self.output_hidden_states,
)
return (
config,
input_ids,
visual_feats,
bounding_boxes,
token_type_ids,
input_mask,
obj_labels,
masked_lm_labels,
matched_label,
ans,
output_attentions,
)
def create_and_check_lxmert_model(
self,
config,
input_ids,
visual_feats,
bounding_boxes,
token_type_ids,
input_mask,
obj_labels,
masked_lm_labels,
matched_label,
ans,
output_attentions,
):
model = TFLxmertModel(config=config)
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
output_attentions=output_attentions,
)
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
output_attentions=not output_attentions,
)
result = model(input_ids, visual_feats, bounding_boxes, return_dict=False)
result = model(input_ids, visual_feats, bounding_boxes, return_dict=True)
self.parent.assertEqual(result.language_output.shape, (self.batch_size, self.seq_length, self.hidden_size))
self.parent.assertEqual(
result.vision_output.shape, (self.batch_size, self.num_visual_features, self.hidden_size)
)
self.parent.assertEqual(result.pooled_output.shape, (self.batch_size, self.hidden_size))
def prepare_config_and_inputs_for_common(self, return_obj_labels=False):
config_and_inputs = self.prepare_config_and_inputs()
(
config,
input_ids,
visual_feats,
bounding_boxes,
token_type_ids,
input_mask,
obj_labels,
masked_lm_labels,
matched_label,
ans,
output_attentions,
) = config_and_inputs
inputs_dict = {
"input_ids": input_ids,
"visual_feats": visual_feats,
"visual_pos": bounding_boxes,
"token_type_ids": token_type_ids,
"attention_mask": input_mask,
}
if return_obj_labels:
inputs_dict["obj_labels"] = obj_labels
return config, inputs_dict
def create_and_check_lxmert_for_pretraining(
self,
config,
input_ids,
visual_feats,
bounding_boxes,
token_type_ids,
input_mask,
obj_labels,
masked_lm_labels,
matched_label,
ans,
output_attentions,
):
model = TFLxmertForPreTraining(config=config)
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
masked_lm_labels=masked_lm_labels,
obj_labels=obj_labels,
matched_label=matched_label,
ans=ans,
output_attentions=output_attentions,
return_dict=True,
)
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
masked_lm_labels=masked_lm_labels,
output_attentions=not output_attentions,
return_dict=False,
)
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
masked_lm_labels=masked_lm_labels,
)
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
obj_labels=obj_labels,
)
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
matched_label=matched_label,
)
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
ans=ans,
)
result = model(
input_ids,
visual_feats,
bounding_boxes,
token_type_ids=token_type_ids,
attention_mask=input_mask,
masked_lm_labels=masked_lm_labels,
obj_labels=obj_labels,
matched_label=matched_label,
ans=ans,
output_attentions=not output_attentions,
return_dict=True,
)
self.parent.assertEqual(result.prediction_logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
@require_tf
class TFLxmertModelTest(TFModelTesterMixin, unittest.TestCase):
all_model_classes = (TFLxmertModel, TFLxmertForPreTraining) if is_tf_available() else ()
def setUp(self):
self.model_tester = TFLxmertModelTester(self)
self.config_tester = ConfigTester(self, config_class=LxmertConfig, hidden_size=37)
def test_config(self):
self.config_tester.run_common_tests()
def test_lxmert_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_lxmert_model(*config_and_inputs)
def test_lxmert_for_pretraining(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_lxmert_for_pretraining(*config_and_inputs)
@slow
def test_model_from_pretrained(self):
for model_name in ["unc-nlp/lxmert-base-uncased"]:
model = TFLxmertModel.from_pretrained(model_name)
self.assertIsNotNone(model)
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
encoder_seq_length = (
self.model_tester.encoder_seq_length
if hasattr(self.model_tester, "encoder_seq_length")
else self.model_tester.seq_length
)
encoder_key_length = (
self.model_tester.key_length if hasattr(self.model_tester, "key_length") else encoder_seq_length
)
for model_class in self.all_model_classes:
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = False
model = model_class(config)
outputs = model(self._prepare_for_class(inputs_dict, model_class))
language_attentions, vision_attentions, cross_encoder_attentions = (outputs[-3], outputs[-2], outputs[-1])
self.assertEqual(model.config.output_hidden_states, False)
self.assertEqual(len(language_attentions), self.model_tester.num_hidden_layers["language"])
self.assertEqual(len(vision_attentions), self.model_tester.num_hidden_layers["vision"])
self.assertEqual(len(cross_encoder_attentions), self.model_tester.num_hidden_layers["cross_encoder"])
attentions = [language_attentions, vision_attentions, cross_encoder_attentions]
attention_shapes = [
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
[
self.model_tester.num_attention_heads,
self.model_tester.num_visual_features,
self.model_tester.num_visual_features,
],
[self.model_tester.num_attention_heads, encoder_key_length, self.model_tester.num_visual_features],
]
for attention, attention_shape in zip(attentions, attention_shapes):
self.assertListEqual(list(attention[0].shape[-3:]), attention_shape)
out_len = len(outputs)
# Check attention is always last and order is fine
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = True
model = model_class(config)
outputs = model(self._prepare_for_class(inputs_dict, model_class))
# 2 hidden states were added
self.assertEqual(out_len + 2, len(outputs))
language_attentions, vision_attentions, cross_encoder_attentions = (outputs[-3], outputs[-2], outputs[-1])
self.assertEqual(len(language_attentions), self.model_tester.num_hidden_layers["language"])
self.assertEqual(len(vision_attentions), self.model_tester.num_hidden_layers["vision"])
self.assertEqual(len(cross_encoder_attentions), self.model_tester.num_hidden_layers["cross_encoder"])
attentions = [language_attentions, vision_attentions, cross_encoder_attentions]
attention_shapes = [
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
[
self.model_tester.num_attention_heads,
self.model_tester.num_visual_features,
self.model_tester.num_visual_features,
],
[self.model_tester.num_attention_heads, encoder_key_length, self.model_tester.num_visual_features],
]
for attention, attention_shape in zip(attentions, attention_shapes):
self.assertListEqual(list(attention[0].shape[-3:]), attention_shape)
def test_hidden_states_output(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
def check_hidden_states_output(config, inputs_dict, model_class):
model = model_class(config)
outputs = model(self._prepare_for_class(inputs_dict, model_class))
language_hidden_states, vision_hidden_states = outputs[-2], outputs[-1]
self.assertEqual(len(language_hidden_states), self.model_tester.num_hidden_layers["language"] + 1)
self.assertEqual(len(vision_hidden_states), self.model_tester.num_hidden_layers["vision"] + 1)
seq_length = self.model_tester.seq_length
num_visual_features = self.model_tester.num_visual_features
self.assertListEqual(
list(language_hidden_states[0].shape[-2:]),
[seq_length, self.model_tester.hidden_size],
)
self.assertListEqual(
list(vision_hidden_states[0].shape[-2:]),
[num_visual_features, self.model_tester.hidden_size],
)
for model_class in self.all_model_classes:
inputs_dict["output_hidden_states"] = True
check_hidden_states_output(config, inputs_dict, model_class)
del inputs_dict["output_hidden_states"]
config.output_hidden_states = True
check_hidden_states_output(config, inputs_dict, model_class)
def test_pt_tf_model_equivalence(self):
from transformers import is_torch_available
if not is_torch_available():
return
import torch
import transformers
for model_class in self.all_model_classes:
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
return_obj_labels="PreTraining" in model_class.__name__
)
pt_model_class_name = model_class.__name__[2:] # Skip the "TF" at the beggining
pt_model_class = getattr(transformers, pt_model_class_name)
config.output_hidden_states = True
config.task_obj_predict = False
tf_model = model_class(config)
pt_model = pt_model_class(config)
# Check we can load pt model in tf and vice-versa with model => model functions
tf_model = transformers.load_pytorch_model_in_tf2_model(
tf_model, pt_model, tf_inputs=self._prepare_for_class(inputs_dict, model_class)
)
pt_model = transformers.load_tf2_model_in_pytorch_model(pt_model, tf_model)
# Check predictions on first output (logits/hidden-states) are close enought given low-level computational differences
pt_model.eval()
# Delete obj labels as we want to compute the hidden states and not the loss
if "obj_labels" in inputs_dict:
del inputs_dict["obj_labels"]
def torch_type(key):
if key in ("visual_feats", "visual_pos"):
return torch.float32
else:
return torch.long
def recursive_numpy_convert(iterable):
return_dict = {}
for key, value in iterable.items():
if isinstance(value, dict):
return_dict[key] = recursive_numpy_convert(value)
else:
if isinstance(value, (list, tuple)):
return_dict[key] = (
torch.from_numpy(iter_value.numpy()).to(torch_type(key)) for iter_value in value
)
else:
return_dict[key] = torch.from_numpy(value.numpy()).to(torch_type(key))
return return_dict
pt_inputs_dict = recursive_numpy_convert(self._prepare_for_class(inputs_dict, model_class))
# need to rename encoder-decoder "inputs" for PyTorch
if "inputs" in pt_inputs_dict and self.is_encoder_decoder:
pt_inputs_dict["input_ids"] = pt_inputs_dict.pop("inputs")
with torch.no_grad():
pto = pt_model(**pt_inputs_dict)
tfo = tf_model(self._prepare_for_class(inputs_dict, model_class), training=False)
tf_hidden_states = tfo[0].numpy()
pt_hidden_states = pto[0].numpy()
import numpy as np
tf_nans = np.copy(np.isnan(tf_hidden_states))
pt_nans = np.copy(np.isnan(pt_hidden_states))
pt_hidden_states[tf_nans] = 0
tf_hidden_states[tf_nans] = 0
pt_hidden_states[pt_nans] = 0
tf_hidden_states[pt_nans] = 0
max_diff = np.amax(np.abs(tf_hidden_states - pt_hidden_states))
# Debug info (remove when fixed)
if max_diff >= 2e-2:
print("===")
print(model_class)
print(config)
print(inputs_dict)
print(pt_inputs_dict)
self.assertLessEqual(max_diff, 6e-2)
# Check we can load pt model in tf and vice-versa with checkpoint => model functions
with tempfile.TemporaryDirectory() as tmpdirname:
import os
pt_checkpoint_path = os.path.join(tmpdirname, "pt_model.bin")
torch.save(pt_model.state_dict(), pt_checkpoint_path)
tf_model = transformers.load_pytorch_checkpoint_in_tf2_model(tf_model, pt_checkpoint_path)
tf_checkpoint_path = os.path.join(tmpdirname, "tf_model.h5")
tf_model.save_weights(tf_checkpoint_path)
pt_model = transformers.load_tf2_checkpoint_in_pytorch_model(pt_model, tf_checkpoint_path)
# Check predictions on first output (logits/hidden-states) are close enought given low-level computational differences
pt_model.eval()
pt_inputs_dict = dict(
(name, torch.from_numpy(key.numpy()).to(torch.long))
for name, key in self._prepare_for_class(inputs_dict, model_class).items()
)
for key, value in pt_inputs_dict.items():
if key in ("visual_feats", "visual_pos"):
pt_inputs_dict[key] = value.to(torch.float32)
else:
pt_inputs_dict[key] = value.to(torch.long)
with torch.no_grad():
pto = pt_model(**pt_inputs_dict)
tfo = tf_model(self._prepare_for_class(inputs_dict, model_class))
tfo = tfo[0].numpy()
pto = pto[0].numpy()
tf_nans = np.copy(np.isnan(tfo))
pt_nans = np.copy(np.isnan(pto))
pto[tf_nans] = 0
tfo[tf_nans] = 0
pto[pt_nans] = 0
tfo[pt_nans] = 0
max_diff = np.amax(np.abs(tfo - pto))
self.assertLessEqual(max_diff, 6e-2)
def test_save_load(self):
for model_class in self.all_model_classes:
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
return_obj_labels="PreTraining" in model_class.__name__
)
model = model_class(config)
outputs = model(self._prepare_for_class(inputs_dict, model_class))
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
model = model_class.from_pretrained(tmpdirname)
after_outputs = model(self._prepare_for_class(inputs_dict, model_class))
self.assert_outputs_same(after_outputs, outputs)
def test_compile_tf_model(self):
optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
metric = tf.keras.metrics.SparseCategoricalAccuracy("accuracy")
for model_class in self.all_model_classes:
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common(
return_obj_labels="PreTraining" in model_class.__name__
)
input_ids = tf.keras.Input(
batch_shape=(self.model_tester.batch_size, self.model_tester.seq_length),
name="input_ids",
dtype="int32",
)
visual_feats = tf.keras.Input(
batch_shape=(
self.model_tester.batch_size,
self.model_tester.num_visual_features,
self.model_tester.visual_feat_dim,
),
name="visual_feats",
dtype="int32",
)
visual_pos = tf.keras.Input(
batch_shape=(self.model_tester.batch_size, self.model_tester.num_visual_features, 4),
name="visual_pos",
dtype="int32",
)
# Prepare our model
model = model_class(config)
# Let's load it from the disk to be sure we can use pretrained weights
with tempfile.TemporaryDirectory() as tmpdirname:
outputs = model(self._prepare_for_class(inputs_dict, model_class)) # build the model
model.save_pretrained(tmpdirname)
model = model_class.from_pretrained(tmpdirname)
outputs_dict = model(input_ids, visual_feats, visual_pos)
hidden_states = outputs_dict[0]
# Add a dense layer on top to test integration with other keras modules
outputs = tf.keras.layers.Dense(2, activation="softmax", name="outputs")(hidden_states)
# Compile extended model
extended_model = tf.keras.Model(inputs=[input_ids, visual_feats, visual_pos], outputs=[outputs])
extended_model.compile(optimizer=optimizer, loss=loss, metrics=[metric])

65
tests/test_tokenization_lxmert.py Normal file
View File

@@ -0,0 +1,65 @@
# coding=utf-8
# Copyright 2018 LXMERT Authors.
#
# 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
# limitations under the License.
import os
import unittest
from transformers.tokenization_bert import VOCAB_FILES_NAMES
from transformers.tokenization_lxmert import LxmertTokenizer
from .test_tokenization_common import TokenizerTesterMixin
class LxmertTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
tokenizer_class = LxmertTokenizer
def setUp(self):
super().setUp()
vocab_tokens = [
"[UNK]",
"[CLS]",
"[SEP]",
"want",
"##want",
"##ed",
"wa",
"un",
"runn",
"##ing",
",",
"low",
"lowest",
]
self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
with open(self.vocab_file, "w", encoding="utf-8") as vocab_writer:
vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
def get_tokenizer(self, **kwargs):
return LxmertTokenizer.from_pretrained(self.tmpdirname, **kwargs)
def get_input_output_texts(self, tokenizer):
input_text = "UNwant\u00E9d,running"
output_text = "unwanted, running"
return input_text, output_text
def test_full_tokenizer(self):
tokenizer = self.tokenizer_class(self.vocab_file)
tokens = tokenizer.tokenize("UNwant\u00E9d,running")
self.assertListEqual(tokens, ["un", "##want", "##ed", ",", "runn", "##ing"])
self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [7, 4, 5, 10, 8, 9])