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

FlaxBart (#11537)

Browse Source 
* Start working on FlaxBart

* Create modeling_flax_bart.py

* Write FlaxBartAttention

* Add FlaxBartEncoderLayer

* Add FlaxBartDecoderLayer and some typing

* Add helepr function for FlaxBart

* shift_tokens_right

* _make_causal_mask

* _expand_mask

* Add PositionalEmbedding and fix init_std naming

* Add FlaxBartPretrainedModel

* Add FlaxBartEncoder

* Add FlaxBartEncoder

* Add FlaxBartEncoder among modules to be imported

* YET WE CANNOT INITIALIZE THAT!! :(

* Make BartEncoder working

Change BartEncoder to instance of nn.Module so far

* Add FlaxBartDecoder

* Add FlaxBartModel

* TODO to make model run -> Prepapre model inputs

* Resolve padding

* Add FlaxBartModel

* Add FlaxBartModel into importable modules

* Remove FlaxBartEncoder and FlaxBartDecoder from importable modules

* make style; not properly working

* make style; make quality not pass due to some import I left

* Remove TODO for padding_idx in nn.Embed so far

* Add FlaxBartForConditionalGeneration

* Incorporate Flax model output classes, i.e. return_dict

* Add another models and incorporate use_cache arg

* Add FlaxBartForSequenceClassification and FlaxBartForQuestionAnswering

* Incorporate use_cache arg from PyTorch implementation

* Add all necessary Flax output utils

* Add FlaxBartForCausalLM; not working yet'

* Add minor improvements; still lacks some functionality

* Update docs, src and tests

* Add support of FlaxBart to docs/source

* Fix some bugs in FlaxBart souce code

* Add some neccessary tests for FlaxBart models - jit_compilation not passing

* Fix tests and add test_head_masking

* Fix tests for @jax.jit computation

* Add test_head_masking

* Migrate FlaxBart tests from jax.numpy to numpy

* Remove FlaxBartForCausalLM

* Clean repo

* fix bart model weight structure

* Fix FlaxBartForSequenceClassification

Slicing is not possible to use below jit, therefore, selecting sentence
representation from hidden_states must be changed.

* Allow FlaxBartForSequenceClassification for testing pt_flax equivalence

* Allow testing for FlaxBartForQA for pt_flax equivalence

* Add a comment to FlaxBartForSequenceClassification + change noise from 1e-3 to 1e-6

* remove past_key_values

* remove inputs_mebeds and make input_ids required

* add position ids

* re-write attention layer

* fix dataclass

* fix pos embeds and attention output

* fix pos embeds

* expose encode method

* expose decode method

* move docstring to top

* add cache for causal attn layer

* remove head masking for now

* s2s greedy search first pass

* boom boom

* fix typos

* fix greedy generate for bart

* use encoder, decoder layers instead of num_hidden_layers

* handle encoder_outputs

* cleanup

* simplify decoding

* more clean-up

* typos

* Change header + add {decoder_,}position_ids into 2 models

* add BartConfig

* fix existing tests

* add encode, decode methods

* Fix shift_tokens_right for JIT compilation + clarify one condition

* fix decode

* encoder => encode

* simplify generate

* add tests for encode and decode

* style

* add tests for cache

* fix equivalence tests

* sample generate now works with seq2seq

* generation tests

* initialize dense layers

* docstring and cleanup

* quality

* remove get/set input_embeddings

* address Patricks suggestions

* decode for every model, remove encoder_outputs from call

* update tests accordingly

* decode returns only decoder outputs and logits

* fix arguments

* doc encode, decode methods

* correct base_model_prefix

* fix test for seq classif model

* fix docs

Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Suraj Patil <surajp815@gmail.com>
This commit is contained in:
Daniel Stancl
2021-06-14 11:46:08 +02:00
committed by GitHub
parent d36fce8237
commit 4a51b1dd9b
11 changed files with 2726 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

417
tests/test_modeling_flax_bart.py Normal file
View File

@@ -0,0 +1,417 @@
# Copyright 2021 The HuggingFace 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
# limitations under the License.
import unittest
import numpy as np
import timeout_decorator # noqa
from transformers import BartConfig, is_flax_available
from transformers.testing_utils import require_flax, slow
from .test_generation_flax_utils import FlaxGenerationTesterMixin
from .test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor
if is_flax_available():
import os
# The slow tests are often failing with OOM error on GPU
# This makes JAX allocate exactly what is needed on demand, and deallocate memory that is no longer needed
# but will be slower as stated here https://jax.readthedocs.io/en/latest/gpu_memory_allocation.html
os.environ["XLA_PYTHON_CLIENT_ALLOCATOR"] = "platform"
import jax
import jax.numpy as jnp
from transformers.models.bart.modeling_flax_bart import (
FlaxBartForConditionalGeneration,
FlaxBartForQuestionAnswering,
FlaxBartForSequenceClassification,
FlaxBartModel,
shift_tokens_right,
)
def prepare_bart_inputs_dict(
config,
input_ids,
decoder_input_ids=None,
attention_mask=None,
decoder_attention_mask=None,
head_mask=None,
decoder_head_mask=None,
cross_attn_head_mask=None,
):
if attention_mask is None:
attention_mask = np.where(input_ids != config.pad_token_id, 1, 0)
if decoder_attention_mask is None:
decoder_attention_mask = np.where(decoder_input_ids != config.pad_token_id, 1, 0)
if head_mask is None:
head_mask = np.ones((config.encoder_layers, config.encoder_attention_heads))
if decoder_head_mask is None:
decoder_head_mask = np.ones((config.decoder_layers, config.decoder_attention_heads))
if cross_attn_head_mask is None:
cross_attn_head_mask = np.ones((config.decoder_layers, config.decoder_attention_heads))
return {
"input_ids": input_ids,
"decoder_input_ids": decoder_input_ids,
"attention_mask": attention_mask,
"decoder_attention_mask": attention_mask,
}
class FlaxBartModelTester(unittest.TestCase):
def __init__(
self,
parent,
batch_size=13,
seq_length=7,
is_training=True,
use_labels=False,
vocab_size=99,
hidden_size=16,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=4,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=32,
eos_token_id=2,
pad_token_id=1,
bos_token_id=0,
initializer_range=0.02,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
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.eos_token_id = eos_token_id
self.pad_token_id = pad_token_id
self.bos_token_id = bos_token_id
self.initializer_range = initializer_range
def prepare_config_and_inputs(self):
input_ids = np.clip(ids_tensor([self.batch_size, self.seq_length - 1], self.vocab_size), 3, self.vocab_size)
input_ids = np.concatenate((input_ids, 2 * np.ones((self.batch_size, 1), dtype=np.int64)), -1)
decoder_input_ids = shift_tokens_right(input_ids, 1, 2)
config = BartConfig(
vocab_size=self.vocab_size,
d_model=self.hidden_size,
encoder_layers=self.num_hidden_layers,
decoder_layers=self.num_hidden_layers,
encoder_attention_heads=self.num_attention_heads,
decoder_attention_heads=self.num_attention_heads,
encoder_ffn_dim=self.intermediate_size,
decoder_ffn_dim=self.intermediate_size,
dropout=self.hidden_dropout_prob,
attention_dropout=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
eos_token_id=self.eos_token_id,
bos_token_id=self.bos_token_id,
pad_token_id=self.pad_token_id,
initializer_range=self.initializer_range,
use_cache=False,
)
inputs_dict = prepare_bart_inputs_dict(config, input_ids, decoder_input_ids)
return config, inputs_dict
def prepare_config_and_inputs_for_common(self):
config, inputs_dict = self.prepare_config_and_inputs()
return config, inputs_dict
def check_use_cache_forward(self, model_class_name, config, inputs_dict):
max_decoder_length = 20
model = model_class_name(config)
encoder_outputs = model.encode(inputs_dict["input_ids"])
decoder_input_ids, decoder_attention_mask = (
inputs_dict["decoder_input_ids"],
inputs_dict["decoder_attention_mask"],
)
past_key_values = model.init_cache(decoder_input_ids.shape[0], max_decoder_length, encoder_outputs)
decoder_attention_mask = jnp.ones((decoder_input_ids.shape[0], max_decoder_length), dtype="i4")
decoder_position_ids = jnp.broadcast_to(
jnp.arange(decoder_input_ids.shape[-1] - 1)[None, :],
(decoder_input_ids.shape[0], decoder_input_ids.shape[-1] - 1),
)
outputs_cache = model.decode(
decoder_input_ids[:, :-1],
encoder_outputs,
decoder_attention_mask=decoder_attention_mask,
past_key_values=past_key_values,
decoder_position_ids=decoder_position_ids,
)
decoder_position_ids = jnp.array(decoder_input_ids.shape[0] * [[decoder_input_ids.shape[-1] - 1]], dtype="i4")
outputs_cache_next = model.decode(
decoder_input_ids[:, -1:],
encoder_outputs,
decoder_attention_mask=decoder_attention_mask,
past_key_values=outputs_cache.past_key_values,
decoder_position_ids=decoder_position_ids,
)
outputs = model.decode(decoder_input_ids, encoder_outputs)
diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5])))
self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
def check_use_cache_forward_with_attn_mask(self, model_class_name, config, inputs_dict):
max_decoder_length = 20
model = model_class_name(config)
encoder_outputs = model.encode(inputs_dict["input_ids"])
decoder_input_ids, decoder_attention_mask = (
inputs_dict["decoder_input_ids"],
inputs_dict["decoder_attention_mask"],
)
decoder_attention_mask_cache = jnp.concatenate(
[
decoder_attention_mask,
jnp.zeros((decoder_attention_mask.shape[0], max_decoder_length - decoder_attention_mask.shape[1])),
],
axis=-1,
)
past_key_values = model.init_cache(decoder_input_ids.shape[0], max_decoder_length, encoder_outputs)
decoder_position_ids = jnp.broadcast_to(
jnp.arange(decoder_input_ids.shape[-1] - 1)[None, :],
(decoder_input_ids.shape[0], decoder_input_ids.shape[-1] - 1),
)
outputs_cache = model.decode(
decoder_input_ids[:, :-1],
encoder_outputs,
decoder_attention_mask=decoder_attention_mask_cache,
past_key_values=past_key_values,
decoder_position_ids=decoder_position_ids,
)
decoder_position_ids = jnp.array(decoder_input_ids.shape[0] * [[decoder_input_ids.shape[-1] - 1]], dtype="i4")
outputs_cache_next = model.decode(
decoder_input_ids[:, -1:],
encoder_outputs,
past_key_values=outputs_cache.past_key_values,
decoder_attention_mask=decoder_attention_mask_cache,
decoder_position_ids=decoder_position_ids,
)
outputs = model.decode(decoder_input_ids, encoder_outputs, decoder_attention_mask=decoder_attention_mask)
diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5])))
self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
@require_flax
class BartHeadTests(unittest.TestCase):
vocab_size = 99
def _get_config_and_data(self):
input_ids = np.array(
[
[71, 82, 18, 33, 46, 91, 2],
[68, 34, 26, 58, 30, 82, 2],
[5, 97, 17, 39, 94, 40, 2],
[76, 83, 94, 25, 70, 78, 2],
[87, 59, 41, 35, 48, 66, 2],
[55, 13, 16, 58, 5, 2, 1], # note padding
[64, 27, 31, 51, 12, 75, 2],
[52, 64, 86, 17, 83, 39, 2],
[48, 61, 9, 24, 71, 82, 2],
[26, 1, 60, 48, 22, 13, 2],
[21, 5, 62, 28, 14, 76, 2],
[45, 98, 37, 86, 59, 48, 2],
[70, 70, 50, 9, 28, 0, 2],
],
dtype=np.int64,
)
batch_size = input_ids.shape[0]
config = BartConfig(
vocab_size=self.vocab_size,
d_model=24,
encoder_layers=2,
decoder_layers=2,
encoder_attention_heads=2,
decoder_attention_heads=2,
encoder_ffn_dim=32,
decoder_ffn_dim=32,
max_position_embeddings=48,
eos_token_id=2,
pad_token_id=1,
bos_token_id=0,
)
return config, input_ids, batch_size
def test_sequence_classification_forward(self):
config, input_ids, batch_size = self._get_config_and_data()
model = FlaxBartForSequenceClassification(config)
outputs = model(input_ids=input_ids, decoder_input_ids=input_ids)
expected_shape = (batch_size, config.num_labels)
self.assertEqual(outputs["logits"].shape, expected_shape)
def test_question_answering_forward(self):
config, input_ids, batch_size = self._get_config_and_data()
model = FlaxBartForQuestionAnswering(config)
outputs = model(input_ids=input_ids)
self.assertEqual(outputs["start_logits"].shape, input_ids.shape)
self.assertEqual(outputs["end_logits"].shape, input_ids.shape)
# @timeout_decorator.timeout(1) # not working with the decorator so far
def test_lm_forward(self):
config, input_ids, batch_size = self._get_config_and_data()
lm_model = FlaxBartForConditionalGeneration(config)
outputs = lm_model(input_ids=input_ids)
expected_shape = (batch_size, input_ids.shape[1], config.vocab_size)
self.assertEqual(outputs["logits"].shape, expected_shape)
def test_lm_uneven_forward(self):
config = BartConfig(
vocab_size=self.vocab_size,
d_model=14,
encoder_layers=2,
decoder_layers=2,
encoder_attention_heads=2,
decoder_attention_heads=2,
encoder_ffn_dim=8,
decoder_ffn_dim=8,
max_position_embeddings=48,
)
lm_model = FlaxBartForConditionalGeneration(config)
context = np.array([[71, 82, 18, 33, 46, 91, 2], [68, 34, 26, 58, 30, 2, 1]], dtype=np.int64)
summary = np.array([[82, 71, 82, 18, 2], [58, 68, 2, 1, 1]], dtype=np.int64)
outputs = lm_model(input_ids=context, decoder_input_ids=summary)
expected_shape = (*summary.shape, config.vocab_size)
self.assertEqual(outputs["logits"].shape, expected_shape)
def test_shift_tokens_right(self):
input_ids = np.array([[71, 82, 18, 33, 2, 1, 1], [68, 34, 26, 58, 30, 82, 2]], dtype=np.int64)
shifted = shift_tokens_right(input_ids, 1, 2)
n_pad_before = np.equal(input_ids, 1).astype(np.float32).sum()
n_pad_after = np.equal(shifted, 1).astype(np.float32).sum()
self.assertEqual(shifted.shape, input_ids.shape)
self.assertEqual(n_pad_after, n_pad_before - 1)
self.assertTrue(np.equal(shifted[:, 0], 2).all())
@require_flax
class FlaxBartModelTest(FlaxModelTesterMixin, unittest.TestCase, FlaxGenerationTesterMixin):
is_encoder_decoder = True
all_model_classes = (
(
FlaxBartModel,
FlaxBartForConditionalGeneration,
FlaxBartForSequenceClassification,
FlaxBartForQuestionAnswering,
)
if is_flax_available()
else ()
)
all_generative_model_classes = (FlaxBartForConditionalGeneration,) if is_flax_available() else ()
def setUp(self):
self.model_tester = FlaxBartModelTester(self)
def test_use_cache_forward(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
for model_class in self.all_model_classes:
self.model_tester.check_use_cache_forward(model_class, config, inputs_dict)
def test_use_cache_forward_with_attn_mask(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
for model_class in self.all_model_classes:
self.model_tester.check_use_cache_forward_with_attn_mask(model_class, config, inputs_dict)
def test_encode(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
model = model_class(config)
@jax.jit
def encode_jitted(input_ids, attention_mask=None, **kwargs):
return model.encode(input_ids=input_ids, attention_mask=attention_mask)
with self.subTest("JIT Enabled"):
jitted_outputs = encode_jitted(**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = encode_jitted(**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
def test_decode(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
model = model_class(config)
encoder_outputs = model.encode(inputs_dict["input_ids"], inputs_dict["attention_mask"])
prepared_inputs_dict = {
"decoder_input_ids": inputs_dict["decoder_input_ids"],
"decoder_attention_mask": inputs_dict["decoder_attention_mask"],
"encoder_outputs": encoder_outputs,
}
@jax.jit
def decode_jitted(decoder_input_ids, decoder_attention_mask, encoder_outputs):
return model.decode(
decoder_input_ids=decoder_input_ids,
decoder_attention_mask=decoder_attention_mask,
encoder_outputs=encoder_outputs,
)
with self.subTest("JIT Enabled"):
jitted_outputs = decode_jitted(**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = decode_jitted(**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained("facebook/bart-base", from_pt=True)
# FlaxBartForSequenceClassification expects eos token in input_ids
input_ids = np.ones((1, 1)) * model.config.eos_token_id
outputs = model(input_ids)
self.assertIsNotNone(outputs)