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

Add Donut (#18488)

Browse Source 
* First draft

* Improve script

* Update script

* Make conversion work

* Add final_layer_norm attribute to Swin's config

* Add DonutProcessor

* Convert more models

* Improve feature extractor and convert base models

* Fix bug

* Improve integration tests

* Improve integration tests and add model to README

* Add doc test

* Add feature extractor to docs

* Fix integration tests

* Remove register_buffer

* Fix toctree and add missing attribute

* Add DonutSwin

* Make conversion script work

* Improve conversion script

* Address comment

* Fix bug

* Fix another bug

* Remove deprecated method from docs

* Make Swin and Swinv2 untouched

* Fix code examples

* Fix processor

* Update model_type to donut-swin

* Add feature extractor tests, add token2json method, improve feature extractor

* Fix failing tests, remove integration test

* Add do_thumbnail for consistency

* Improve code examples

* Add code example for document parsing

* Add DonutSwin to MODEL_NAMES_MAPPING

* Add model to appropriate place in toctree

* Update namespace to appropriate organization

Co-authored-by: Niels Rogge <nielsrogge@Nielss-MacBook-Pro.local>
This commit is contained in:
NielsRogge
2022-08-12 16:40:58 +02:00
committed by GitHub
parent a5ca56ff15
commit 2ab790e82d
31 changed files with 2924 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
tests/models/donut/__init__.py Normal file
View File

203
tests/models/donut/test_feature_extraction_donut.py Normal file
View File

@@ -0,0 +1,203 @@
# coding=utf-8
# Copyright 2022 HuggingFace Inc.
#
# 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
from transformers.testing_utils import require_torch, require_vision
from transformers.utils import is_torch_available, is_vision_available
from ...test_feature_extraction_common import FeatureExtractionSavingTestMixin, prepare_image_inputs
if is_torch_available():
import torch
if is_vision_available():
from PIL import Image
from transformers import DonutFeatureExtractor
class DonutFeatureExtractionTester(unittest.TestCase):
def __init__(
self,
parent,
batch_size=7,
num_channels=3,
image_size=18,
min_resolution=30,
max_resolution=400,
do_resize=True,
size=[20, 18],
do_thumbnail=True,
do_align_axis=False,
do_pad=True,
do_normalize=True,
image_mean=[0.5, 0.5, 0.5],
image_std=[0.5, 0.5, 0.5],
):
self.parent = parent
self.batch_size = batch_size
self.num_channels = num_channels
self.image_size = image_size
self.min_resolution = min_resolution
self.max_resolution = max_resolution
self.do_resize = do_resize
self.size = size
self.do_thumbnail = do_thumbnail
self.do_align_axis = do_align_axis
self.do_pad = do_pad
self.do_normalize = do_normalize
self.image_mean = image_mean
self.image_std = image_std
def prepare_feat_extract_dict(self):
return {
"do_resize": self.do_resize,
"size": self.size,
"do_thumbnail": self.do_thumbnail,
"do_align_long_axis": self.do_align_axis,
"do_pad": self.do_pad,
"do_normalize": self.do_normalize,
"image_mean": self.image_mean,
"image_std": self.image_std,
}
@require_torch
@require_vision
class DonutFeatureExtractionTest(FeatureExtractionSavingTestMixin, unittest.TestCase):
feature_extraction_class = DonutFeatureExtractor if is_vision_available() else None
def setUp(self):
self.feature_extract_tester = DonutFeatureExtractionTester(self)
@property
def feat_extract_dict(self):
return self.feature_extract_tester.prepare_feat_extract_dict()
def test_feat_extract_properties(self):
feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
self.assertTrue(hasattr(feature_extractor, "do_resize"))
self.assertTrue(hasattr(feature_extractor, "size"))
self.assertTrue(hasattr(feature_extractor, "do_thumbnail"))
self.assertTrue(hasattr(feature_extractor, "do_align_long_axis"))
self.assertTrue(hasattr(feature_extractor, "do_pad"))
self.assertTrue(hasattr(feature_extractor, "do_normalize"))
self.assertTrue(hasattr(feature_extractor, "image_mean"))
self.assertTrue(hasattr(feature_extractor, "image_std"))
def test_batch_feature(self):
pass
def test_call_pil(self):
# Initialize feature_extractor
feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
# create random PIL images
image_inputs = prepare_image_inputs(self.feature_extract_tester, equal_resolution=False)
for image in image_inputs:
self.assertIsInstance(image, Image.Image)
# Test not batched input
encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
self.assertEqual(
encoded_images.shape,
(
1,
self.feature_extract_tester.num_channels,
self.feature_extract_tester.size[1],
self.feature_extract_tester.size[0],
),
)
# Test batched
encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
self.assertEqual(
encoded_images.shape,
(
self.feature_extract_tester.batch_size,
self.feature_extract_tester.num_channels,
self.feature_extract_tester.size[1],
self.feature_extract_tester.size[0],
),
)
def test_call_numpy(self):
# Initialize feature_extractor
feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
# create random numpy tensors
image_inputs = prepare_image_inputs(self.feature_extract_tester, equal_resolution=False, numpify=True)
for image in image_inputs:
self.assertIsInstance(image, np.ndarray)
# Test not batched input
encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
self.assertEqual(
encoded_images.shape,
(
1,
self.feature_extract_tester.num_channels,
self.feature_extract_tester.size[1],
self.feature_extract_tester.size[0],
),
)
# Test batched
encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
self.assertEqual(
encoded_images.shape,
(
self.feature_extract_tester.batch_size,
self.feature_extract_tester.num_channels,
self.feature_extract_tester.size[1],
self.feature_extract_tester.size[0],
),
)
def test_call_pytorch(self):
# Initialize feature_extractor
feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
# create random PyTorch tensors
image_inputs = prepare_image_inputs(self.feature_extract_tester, equal_resolution=False, torchify=True)
for image in image_inputs:
self.assertIsInstance(image, torch.Tensor)
# Test not batched input
encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
self.assertEqual(
encoded_images.shape,
(
1,
self.feature_extract_tester.num_channels,
self.feature_extract_tester.size[1],
self.feature_extract_tester.size[0],
),
)
# Test batched
encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
self.assertEqual(
encoded_images.shape,
(
self.feature_extract_tester.batch_size,
self.feature_extract_tester.num_channels,
self.feature_extract_tester.size[1],
self.feature_extract_tester.size[0],
),
)

464
tests/models/donut/test_modeling_donut_swin.py Normal file
View File

@@ -0,0 +1,464 @@
# coding=utf-8
# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" Testing suite for the PyTorch Donut Swin model. """
import collections
import inspect
import os
import pickle
import tempfile
import unittest
from transformers import DonutSwinConfig
from transformers.testing_utils import require_torch, slow, torch_device
from transformers.utils import is_torch_available, is_torch_fx_available
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor, ids_tensor
if is_torch_available():
import torch
from torch import nn
from transformers import DonutSwinModel
from transformers.models.donut.modeling_donut_swin import DONUT_SWIN_PRETRAINED_MODEL_ARCHIVE_LIST
if is_torch_fx_available():
from transformers.utils.fx import symbolic_trace
class DonutSwinModelTester:
def __init__(
self,
parent,
batch_size=13,
image_size=32,
patch_size=2,
num_channels=3,
embed_dim=16,
depths=[1, 2, 1],
num_heads=[2, 2, 4],
window_size=2,
mlp_ratio=2.0,
qkv_bias=True,
hidden_dropout_prob=0.0,
attention_probs_dropout_prob=0.0,
drop_path_rate=0.1,
hidden_act="gelu",
use_absolute_embeddings=False,
patch_norm=True,
initializer_range=0.02,
layer_norm_eps=1e-5,
is_training=True,
scope=None,
use_labels=True,
type_sequence_label_size=10,
encoder_stride=8,
):
self.parent = parent
self.batch_size = batch_size
self.image_size = image_size
self.patch_size = patch_size
self.num_channels = num_channels
self.embed_dim = embed_dim
self.depths = depths
self.num_heads = num_heads
self.window_size = window_size
self.mlp_ratio = mlp_ratio
self.qkv_bias = qkv_bias
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.drop_path_rate = drop_path_rate
self.hidden_act = hidden_act
self.use_absolute_embeddings = use_absolute_embeddings
self.patch_norm = patch_norm
self.layer_norm_eps = layer_norm_eps
self.initializer_range = initializer_range
self.is_training = is_training
self.scope = scope
self.use_labels = use_labels
self.type_sequence_label_size = type_sequence_label_size
self.encoder_stride = encoder_stride
def prepare_config_and_inputs(self):
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
labels = None
if self.use_labels:
labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
config = self.get_config()
return config, pixel_values, labels
def get_config(self):
return DonutSwinConfig(
image_size=self.image_size,
patch_size=self.patch_size,
num_channels=self.num_channels,
embed_dim=self.embed_dim,
depths=self.depths,
num_heads=self.num_heads,
window_size=self.window_size,
mlp_ratio=self.mlp_ratio,
qkv_bias=self.qkv_bias,
hidden_dropout_prob=self.hidden_dropout_prob,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
drop_path_rate=self.drop_path_rate,
hidden_act=self.hidden_act,
use_absolute_embeddings=self.use_absolute_embeddings,
path_norm=self.patch_norm,
layer_norm_eps=self.layer_norm_eps,
initializer_range=self.initializer_range,
encoder_stride=self.encoder_stride,
)
def create_and_check_model(self, config, pixel_values, labels):
model = DonutSwinModel(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
expected_seq_len = ((config.image_size // config.patch_size) ** 2) // (4 ** (len(config.depths) - 1))
expected_dim = int(config.embed_dim * 2 ** (len(config.depths) - 1))
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, expected_seq_len, expected_dim))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
config,
pixel_values,
labels,
) = config_and_inputs
inputs_dict = {"pixel_values": pixel_values}
return config, inputs_dict
@require_torch
class DonutSwinModelTest(ModelTesterMixin, unittest.TestCase):
all_model_classes = (DonutSwinModel,) if is_torch_available() else ()
fx_compatible = True
test_pruning = False
test_resize_embeddings = False
test_head_masking = False
def setUp(self):
self.model_tester = DonutSwinModelTester(self)
self.config_tester = ConfigTester(self, config_class=DonutSwinConfig, embed_dim=37)
def test_config(self):
self.create_and_test_config_common_properties()
self.config_tester.create_and_test_config_to_json_string()
self.config_tester.create_and_test_config_to_json_file()
self.config_tester.create_and_test_config_from_and_save_pretrained()
self.config_tester.create_and_test_config_with_num_labels()
self.config_tester.check_config_can_be_init_without_params()
self.config_tester.check_config_arguments_init()
def create_and_test_config_common_properties(self):
return
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_inputs_embeds(self):
# DonutSwin does not use inputs_embeds
pass
def test_model_common_attributes(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
x = model.get_output_embeddings()
self.assertTrue(x is None or isinstance(x, nn.Linear))
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.forward)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = ["pixel_values"]
self.assertListEqual(arg_names[:1], expected_arg_names)
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
for model_class in self.all_model_classes:
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = False
config.return_dict = 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))
attentions = outputs.attentions
expected_num_attentions = len(self.model_tester.depths)
self.assertEqual(len(attentions), expected_num_attentions)
# check that output_attentions also work using config
del inputs_dict["output_attentions"]
config.output_attentions = True
window_size_squared = config.window_size**2
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
attentions = outputs.attentions
self.assertEqual(len(attentions), expected_num_attentions)
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_heads[0], window_size_squared, window_size_squared],
)
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))
if hasattr(self.model_tester, "num_hidden_states_types"):
added_hidden_states = self.model_tester.num_hidden_states_types
else:
# also another +1 for reshaped_hidden_states
added_hidden_states = 2
self.assertEqual(out_len + added_hidden_states, len(outputs))
self_attentions = outputs.attentions
self.assertEqual(len(self_attentions), expected_num_attentions)
self.assertListEqual(
list(self_attentions[0].shape[-3:]),
[self.model_tester.num_heads[0], window_size_squared, window_size_squared],
)
def check_hidden_states_output(self, inputs_dict, config, model_class, image_size):
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
hidden_states = outputs.hidden_states
expected_num_layers = getattr(
self.model_tester, "expected_num_hidden_layers", len(self.model_tester.depths) + 1
)
self.assertEqual(len(hidden_states), expected_num_layers)
# DonutSwin has a different seq_length
patch_size = (
config.patch_size
if isinstance(config.patch_size, collections.abc.Iterable)
else (config.patch_size, config.patch_size)
)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
self.assertListEqual(
list(hidden_states[0].shape[-2:]),
[num_patches, self.model_tester.embed_dim],
)
reshaped_hidden_states = outputs.reshaped_hidden_states
self.assertEqual(len(reshaped_hidden_states), expected_num_layers)
batch_size, num_channels, height, width = reshaped_hidden_states[0].shape
reshaped_hidden_states = (
reshaped_hidden_states[0].view(batch_size, num_channels, height * width).permute(0, 2, 1)
)
self.assertListEqual(
list(reshaped_hidden_states.shape[-2:]),
[num_patches, self.model_tester.embed_dim],
)
def test_hidden_states_output(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
image_size = (
self.model_tester.image_size
if isinstance(self.model_tester.image_size, collections.abc.Iterable)
else (self.model_tester.image_size, self.model_tester.image_size)
)
for model_class in self.all_model_classes:
inputs_dict["output_hidden_states"] = True
self.check_hidden_states_output(inputs_dict, config, model_class, image_size)
# check that output_hidden_states also work using config
del inputs_dict["output_hidden_states"]
config.output_hidden_states = True
self.check_hidden_states_output(inputs_dict, config, model_class, image_size)
def test_hidden_states_output_with_padding(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.patch_size = 3
image_size = (
self.model_tester.image_size
if isinstance(self.model_tester.image_size, collections.abc.Iterable)
else (self.model_tester.image_size, self.model_tester.image_size)
)
patch_size = (
config.patch_size
if isinstance(config.patch_size, collections.abc.Iterable)
else (config.patch_size, config.patch_size)
)
padded_height = image_size[0] + patch_size[0] - (image_size[0] % patch_size[0])
padded_width = image_size[1] + patch_size[1] - (image_size[1] % patch_size[1])
for model_class in self.all_model_classes:
inputs_dict["output_hidden_states"] = True
self.check_hidden_states_output(inputs_dict, config, model_class, (padded_height, padded_width))
# check that output_hidden_states also work using config
del inputs_dict["output_hidden_states"]
config.output_hidden_states = True
self.check_hidden_states_output(inputs_dict, config, model_class, (padded_height, padded_width))
@slow
def test_model_from_pretrained(self):
for model_name in DONUT_SWIN_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
model = DonutSwinModel.from_pretrained(model_name)
self.assertIsNotNone(model)
def test_initialization(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
configs_no_init = _config_zero_init(config)
for model_class in self.all_model_classes:
model = model_class(config=configs_no_init)
for name, param in model.named_parameters():
if "embeddings" not in name and param.requires_grad:
self.assertIn(
((param.data.mean() * 1e9).round() / 1e9).item(),
[0.0, 1.0],
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
)
def _create_and_check_torch_fx_tracing(self, config, inputs_dict, output_loss=False):
if not is_torch_fx_available() or not self.fx_compatible:
return
configs_no_init = _config_zero_init(config) # To be sure we have no Nan
configs_no_init.return_dict = False
for model_class in self.all_model_classes:
model = model_class(config=configs_no_init)
model.to(torch_device)
model.eval()
inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=output_loss)
try:
if model.config.is_encoder_decoder:
model.config.use_cache = False # FSTM still requires this hack -> FSTM should probably be refactored similar to BART afterward
labels = inputs.get("labels", None)
input_names = ["input_ids", "attention_mask", "decoder_input_ids", "decoder_attention_mask"]
if labels is not None:
input_names.append("labels")
filtered_inputs = {k: v for (k, v) in inputs.items() if k in input_names}
input_names = list(filtered_inputs.keys())
model_output = model(**filtered_inputs)
traced_model = symbolic_trace(model, input_names)
traced_output = traced_model(**filtered_inputs)
else:
input_names = ["input_ids", "attention_mask", "token_type_ids", "pixel_values"]
labels = inputs.get("labels", None)
start_positions = inputs.get("start_positions", None)
end_positions = inputs.get("end_positions", None)
if labels is not None:
input_names.append("labels")
if start_positions is not None:
input_names.append("start_positions")
if end_positions is not None:
input_names.append("end_positions")
filtered_inputs = {k: v for (k, v) in inputs.items() if k in input_names}
input_names = list(filtered_inputs.keys())
model_output = model(**filtered_inputs)
traced_model = symbolic_trace(model, input_names)
traced_output = traced_model(**filtered_inputs)
except RuntimeError as e:
self.fail(f"Couldn't trace module: {e}")
def flatten_output(output):
flatten = []
for x in output:
if isinstance(x, (tuple, list)):
flatten += flatten_output(x)
elif not isinstance(x, torch.Tensor):
continue
else:
flatten.append(x)
return flatten
model_output = flatten_output(model_output)
traced_output = flatten_output(traced_output)
num_outputs = len(model_output)
for i in range(num_outputs):
self.assertTrue(
torch.allclose(model_output[i], traced_output[i]),
f"traced {i}th output doesn't match model {i}th output for {model_class}",
)
# Test that the model can be serialized and restored properly
with tempfile.TemporaryDirectory() as tmp_dir_name:
pkl_file_name = os.path.join(tmp_dir_name, "model.pkl")
try:
with open(pkl_file_name, "wb") as f:
pickle.dump(traced_model, f)
with open(pkl_file_name, "rb") as f:
loaded = pickle.load(f)
except Exception as e:
self.fail(f"Couldn't serialize / deserialize the traced model: {e}")
loaded_output = loaded(**filtered_inputs)
loaded_output = flatten_output(loaded_output)
for i in range(num_outputs):
self.assertTrue(
torch.allclose(model_output[i], loaded_output[i]),
f"serialized model {i}th output doesn't match model {i}th output for {model_class}",
)