Skip to content

Commit 64e92ed

Browse files
authored
Update transformers metadata (huggingface#14724)
* Wip on metadata update * Most of the script * Add a job to auto-update the transformers metadata * Style
1 parent c3cd88a commit 64e92ed

File tree

2 files changed

+264
-0
lines changed

2 files changed

+264
-0
lines changed

.github/workflows/update_metdata.yml

Lines changed: 36 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,36 @@
1+
name: Build documentation
2+
3+
on:
4+
push:
5+
branches:
6+
- master
7+
- update_transformers_metadata
8+
9+
jobs:
10+
build_and_package:
11+
runs-on: ubuntu-latest
12+
defaults:
13+
run:
14+
shell: bash -l {0}
15+
16+
steps:
17+
- uses: actions/checkout@v2
18+
19+
- name: Loading cache.
20+
uses: actions/cache@v2
21+
id: cache
22+
with:
23+
path: ~/.cache/pip
24+
key: v1-metadata
25+
restore-keys: |
26+
v1-metadata-${{ hashFiles('setup.py') }}
27+
v1-metadata
28+
29+
- name: Setup environment
30+
run: |
31+
pip install git+https://github.com/huggingface/transformers#egg=transformers[dev]
32+
33+
- name: Update metadata
34+
run: |
35+
python utils/update_metadata.py --token ${{ secrets.SYLVAIN_HF_TOKEN }} --commit_sha ${{ github.sha }}
36+

utils/update_metadata.py

Lines changed: 228 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,228 @@
1+
# coding=utf-8
2+
# Copyright 2021 The HuggingFace Inc. team.
3+
#
4+
# Licensed under the Apache License, Version 2.0 (the "License");
5+
# you may not use this file except in compliance with the License.
6+
# You may obtain a copy of the License at
7+
#
8+
# http://www.apache.org/licenses/LICENSE-2.0
9+
#
10+
# Unless required by applicable law or agreed to in writing, software
11+
# distributed under the License is distributed on an "AS IS" BASIS,
12+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13+
# See the License for the specific language governing permissions and
14+
# limitations under the License.
15+
16+
import argparse
17+
import collections
18+
import importlib.util
19+
import os
20+
import re
21+
import tempfile
22+
23+
import pandas as pd
24+
from datasets import Dataset
25+
26+
from huggingface_hub import Repository
27+
28+
29+
# All paths are set with the intent you should run this script from the root of the repo with the command
30+
# python utils/update_metadata.py
31+
TRANSFORMERS_PATH = "src/transformers"
32+
33+
34+
# This is to make sure the transformers module imported is the one in the repo.
35+
spec = importlib.util.spec_from_file_location(
36+
"transformers",
37+
os.path.join(TRANSFORMERS_PATH, "__init__.py"),
38+
submodule_search_locations=[TRANSFORMERS_PATH],
39+
)
40+
transformers_module = spec.loader.load_module()
41+
42+
43+
# Regexes that match TF/Flax/PT model names.
44+
_re_tf_models = re.compile(r"TF(.*)(?:Model|Encoder|Decoder|ForConditionalGeneration)")
45+
_re_flax_models = re.compile(r"Flax(.*)(?:Model|Encoder|Decoder|ForConditionalGeneration)")
46+
# Will match any TF or Flax model too so need to be in an else branch afterthe two previous regexes.
47+
_re_pt_models = re.compile(r"(.*)(?:Model|Encoder|Decoder|ForConditionalGeneration)")
48+
49+
50+
# Fill this with tuples (pipeline_tag, model_mapping, auto_model)
51+
PIPELINE_TAGS_AND_AUTO_MODELS = [
52+
("pretraining", "MODEL_FOR_PRETRAINING_MAPPING_NAMES", "AutoModelForPreTraining"),
53+
("feature-extraction", "MODEL_MAPPING_NAMES", "AutoModel"),
54+
("audio-classification", "MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES", "AutoModelForAudioClassification"),
55+
("text-generation", "MODEL_FOR_CAUSAL_LM_MAPPING_NAMES", "AutoModelForCausalLM"),
56+
("automatic-speech-recognition", "MODEL_FOR_CTC_MAPPING_NAMES", "AutoModelForCTC"),
57+
("image-classification", "MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES", "AutoModelForImageClassification"),
58+
("image-segmentation", "MODEL_FOR_IMAGE_SEGMENTATION_MAPPING_NAMES", "AutoModelForImageSegmentation"),
59+
("fill-mask", "MODEL_FOR_MASKED_LM_MAPPING_NAMES", "AutoModelForMaskedLM"),
60+
("object-detection", "MODEL_FOR_OBJECT_DETECTION_MAPPING_NAMES", "AutoModelForObjectDetection"),
61+
("question-answering", "MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES", "AutoModelForQuestionAnswering"),
62+
("text2text-generation", "MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES", "AutoModelForSeq2SeqLM"),
63+
("text-classification", "MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES", "AutoModelForSequenceClassification"),
64+
("automatic-speech-recognition", "MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES", "AutoModelForSpeechSeq2Seq"),
65+
(
66+
"table-question-answering",
67+
"MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING_NAMES",
68+
"AutoModelForTableQuestionAnswering",
69+
),
70+
("token-classification", "MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES", "AutoModelForTokenClassification"),
71+
("multiple-choice", "MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES", "AutoModelForMultipleChoice"),
72+
(
73+
"next-sentence-prediction",
74+
"MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING_NAMES",
75+
"AutoModelForNextSentencePrediction",
76+
),
77+
]
78+
79+
80+
# Thanks to https://stackoverflow.com/questions/29916065/how-to-do-camelcase-split-in-python
81+
def camel_case_split(identifier):
82+
"Split a camelcased `identifier` into words."
83+
matches = re.finditer(".+?(?:(?<=[a-z])(?=[A-Z])|(?<=[A-Z])(?=[A-Z][a-z])|$)", identifier)
84+
return [m.group(0) for m in matches]
85+
86+
87+
def get_frameworks_table():
88+
"""
89+
Generates a dataframe containing the supported auto classes for each model type, using the content of the auto
90+
modules.
91+
"""
92+
# Dictionary model names to config.
93+
config_maping_names = transformers_module.models.auto.configuration_auto.CONFIG_MAPPING_NAMES
94+
model_prefix_to_model_type = {
95+
config.replace("Config", ""): model_type for model_type, config in config_maping_names.items()
96+
}
97+
98+
# Dictionaries flagging if each model prefix has a backend in PT/TF/Flax.
99+
pt_models = collections.defaultdict(bool)
100+
tf_models = collections.defaultdict(bool)
101+
flax_models = collections.defaultdict(bool)
102+
103+
# Let's lookup through all transformers object (once) and find if models are supported by a given backend.
104+
for attr_name in dir(transformers_module):
105+
lookup_dict = None
106+
if _re_tf_models.match(attr_name) is not None:
107+
lookup_dict = tf_models
108+
attr_name = _re_tf_models.match(attr_name).groups()[0]
109+
elif _re_flax_models.match(attr_name) is not None:
110+
lookup_dict = flax_models
111+
attr_name = _re_flax_models.match(attr_name).groups()[0]
112+
elif _re_pt_models.match(attr_name) is not None:
113+
lookup_dict = pt_models
114+
attr_name = _re_pt_models.match(attr_name).groups()[0]
115+
116+
if lookup_dict is not None:
117+
while len(attr_name) > 0:
118+
if attr_name in model_prefix_to_model_type:
119+
lookup_dict[model_prefix_to_model_type[attr_name]] = True
120+
break
121+
# Try again after removing the last word in the name
122+
attr_name = "".join(camel_case_split(attr_name)[:-1])
123+
124+
all_models = set(list(pt_models.keys()) + list(tf_models.keys()) + list(flax_models.keys()))
125+
all_models = list(all_models)
126+
all_models.sort()
127+
128+
data = {"model_type": all_models}
129+
data["pytorch"] = [pt_models[t] for t in all_models]
130+
data["tensorflow"] = [tf_models[t] for t in all_models]
131+
data["flax"] = [flax_models[t] for t in all_models]
132+
133+
# Now let's use the auto-mapping names to make sure
134+
processors = {}
135+
for t in all_models:
136+
if t in transformers_module.models.auto.processing_auto.PROCESSOR_MAPPING_NAMES:
137+
processors[t] = "AutoProcessor"
138+
elif t in transformers_module.models.auto.tokenization_auto.TOKENIZER_MAPPING_NAMES:
139+
processors[t] = "AutoTokenizer"
140+
elif t in transformers_module.models.auto.feature_extraction_auto.FEATURE_EXTRACTOR_MAPPING_NAMES:
141+
processors[t] = "AutoFeatureExtractor"
142+
else:
143+
# Default to AutoTokenizer if a model has nothing, for backward compatibility.
144+
processors[t] = "AutoTokenizer"
145+
146+
data["processor"] = [processors[t] for t in all_models]
147+
148+
return pd.DataFrame(data)
149+
150+
151+
def update_pipeline_and_auto_class_table(table):
152+
"""
153+
Update the table of model class to (pipeline_tag, auto_class) without removing old keys if they don't exist
154+
anymore.
155+
"""
156+
auto_modules = [
157+
transformers_module.models.auto.modeling_auto,
158+
transformers_module.models.auto.modeling_tf_auto,
159+
transformers_module.models.auto.modeling_flax_auto,
160+
]
161+
for pipeline_tag, model_mapping, auto_class in PIPELINE_TAGS_AND_AUTO_MODELS:
162+
model_mappings = [model_mapping, f"TF_{model_mapping}", f"FLAX_{model_mapping}"]
163+
auto_classes = [auto_class, f"TF_{auto_class}", f"Flax_{auto_class}"]
164+
# Loop through all three frameworks
165+
for module, cls, mapping in zip(auto_modules, auto_classes, model_mappings):
166+
# The type of pipeline may not exist in this framework
167+
if not hasattr(module, mapping):
168+
continue
169+
# First extract all model_names
170+
model_names = []
171+
for name in getattr(module, mapping).values():
172+
if isinstance(name, str):
173+
model_names.append(name)
174+
else:
175+
model_names.extend(list(name))
176+
177+
# Add pipeline tag and auto model class for those models
178+
table.update({model_name: (pipeline_tag, cls) for model_name in model_names})
179+
180+
return table
181+
182+
183+
def update_metadata(token, commit_sha):
184+
"""
185+
Update the metada for the Transformers repo.
186+
"""
187+
with tempfile.TemporaryDirectory() as tmp_dir:
188+
repo = Repository(
189+
tmp_dir, clone_from="huggingface/transformers-metadata", repo_type="dataset", use_auth_token=token
190+
)
191+
192+
frameworks_table = get_frameworks_table()
193+
frameworks_dataset = Dataset.from_pandas(frameworks_table)
194+
frameworks_dataset.to_json(os.path.join(tmp_dir, "frameworks.json"))
195+
196+
tags_dataset = Dataset.from_json(os.path.join(tmp_dir, "pipeline_tags.json"))
197+
table = {
198+
tags_dataset[i]["model_class"]: (tags_dataset[i]["pipeline_tag"], tags_dataset[i]["auto_class"])
199+
for i in range(len(tags_dataset))
200+
}
201+
table = update_pipeline_and_auto_class_table(table)
202+
203+
# Sort the model classes to avoid some nondeterministic updates to create false update commits.
204+
model_classes = sorted(list(table.keys()))
205+
tags_table = pd.DataFrame(
206+
{
207+
"model_class": model_classes,
208+
"pipeline_tag": [table[m][0] for m in model_classes],
209+
"auto_class": [table[m][1] for m in model_classes],
210+
}
211+
)
212+
tags_dataset = Dataset.from_pandas(tags_table)
213+
tags_dataset.to_json(os.path.join(tmp_dir, "pipeline_tags.json"))
214+
215+
if repo.is_repo_clean():
216+
print("Nothing to commit!")
217+
else:
218+
commit_message = f"Update with commit {commit_sha}" if commit_sha is not None else "Update"
219+
repo.push_to_hub(commit_message)
220+
221+
222+
if __name__ == "__main__":
223+
parser = argparse.ArgumentParser()
224+
parser.add_argument("--token", type=str, help="The token to use to push to the transformers-metadata dataset.")
225+
parser.add_argument("--commit_sha", type=str, help="The sha of the commit going with this update.")
226+
args = parser.parse_args()
227+
228+
update_metadata(args.token, args.commit_sha)

0 commit comments

Comments
 (0)