From 87b4a52b3c22446a2265cdf90500b6655497bfd7 Mon Sep 17 00:00:00 2001 From: "jiaqi.sjq" Date: Wed, 19 Oct 2022 23:29:58 +0800 Subject: [PATCH 01/10] Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/10462681 --- modelscope/hub/git.py | 7 +++++-- tests/hub/test_hub_upload.py | 39 +++++++++--------------------------- 2 files changed, 14 insertions(+), 32 deletions(-) diff --git a/modelscope/hub/git.py b/modelscope/hub/git.py index db76506e..1369fddf 100644 --- a/modelscope/hub/git.py +++ b/modelscope/hub/git.py @@ -184,8 +184,11 @@ class GitCommandWrapper(metaclass=Singleton): info = [ line.strip() for line in rsp.stdout.decode('utf8').strip().split(os.linesep) - ][1:] - return ['/'.join(line.split('/')[1:]) for line in info] + ] + if len(info) == 1: + return ['/'.join(info[0].split('/')[1:])] + else: + return ['/'.join(line.split('/')[1:]) for line in info[1:]] def pull(self, repo_dir: str): cmds = ['-C', repo_dir, 'pull'] diff --git a/tests/hub/test_hub_upload.py b/tests/hub/test_hub_upload.py index 2250164b..a10f7c81 100644 --- a/tests/hub/test_hub_upload.py +++ b/tests/hub/test_hub_upload.py @@ -7,12 +7,13 @@ import uuid from modelscope.hub.api import HubApi from modelscope.hub.constants import Licenses, ModelVisibility +from modelscope.hub.errors import HTTPError, NotLoginException from modelscope.hub.repository import Repository from modelscope.hub.upload import upload_folder from modelscope.utils.constant import ModelFile from modelscope.utils.logger import get_logger from modelscope.utils.test_utils import test_level -from .test_utils import TEST_ACCESS_TOKEN1, delete_credential +from .test_utils import TEST_ACCESS_TOKEN1, TEST_MODEL_ORG, delete_credential logger = get_logger() @@ -22,7 +23,7 @@ class HubUploadTest(unittest.TestCase): def setUp(self): logger.info('SetUp') self.api = HubApi() - self.user = os.environ.get('TEST_MODEL_ORG', 'citest') + self.user = TEST_MODEL_ORG logger.info(self.user) self.create_model_name = '%s/%s_%s' % (self.user, 'test_model_upload', uuid.uuid4().hex) @@ -39,7 +40,10 @@ class HubUploadTest(unittest.TestCase): def tearDown(self): logger.info('TearDown') shutil.rmtree(self.model_dir, ignore_errors=True) - self.api.delete_model(model_id=self.create_model_name) + try: + self.api.delete_model(model_id=self.create_model_name) + except Exception: + pass def test_upload_exits_repo_master(self): logger.info('basic test for upload!') @@ -119,48 +123,23 @@ class HubUploadTest(unittest.TestCase): logger.info('test upload without login!') self.api.login(TEST_ACCESS_TOKEN1) delete_credential() - try: - upload_folder( - model_id=self.create_model_name, - model_dir=self.finetune_path, - visibility=ModelVisibility.PUBLIC, - license=Licenses.APACHE_V2) - except Exception as e: - logger.info(e) - self.api.login(TEST_ACCESS_TOKEN1) + with self.assertRaises(NotLoginException): upload_folder( model_id=self.create_model_name, model_dir=self.finetune_path, visibility=ModelVisibility.PUBLIC, license=Licenses.APACHE_V2) - Repository( - model_dir=self.repo_path, clone_from=self.create_model_name) - assert os.path.exists( - os.path.join(self.repo_path, 'configuration.json')) - shutil.rmtree(self.repo_path, ignore_errors=True) @unittest.skipUnless(test_level() >= 0, 'skip test in current test level') def test_upload_invalid_repo(self): logger.info('test upload to invalid repo!') self.api.login(TEST_ACCESS_TOKEN1) - try: + with self.assertRaises(HTTPError): upload_folder( model_id='%s/%s' % ('speech_tts', 'invalid_model_test'), model_dir=self.finetune_path, visibility=ModelVisibility.PUBLIC, license=Licenses.APACHE_V2) - except Exception as e: - logger.info(e) - upload_folder( - model_id=self.create_model_name, - model_dir=self.finetune_path, - visibility=ModelVisibility.PUBLIC, - license=Licenses.APACHE_V2) - Repository( - model_dir=self.repo_path, clone_from=self.create_model_name) - assert os.path.exists( - os.path.join(self.repo_path, 'configuration.json')) - shutil.rmtree(self.repo_path, ignore_errors=True) if __name__ == '__main__': From 089cadab4b0f0e6d17a76ef0ef88ae0c1c9d4fe5 Mon Sep 17 00:00:00 2001 From: Yingda Chen Date: Thu, 20 Oct 2022 08:51:25 +0800 Subject: [PATCH 02/10] [to #42322933] disable unstable trainer test --- tests/trainers/test_image_denoise_trainer.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tests/trainers/test_image_denoise_trainer.py b/tests/trainers/test_image_denoise_trainer.py index 68ddf616..c4abca6a 100644 --- a/tests/trainers/test_image_denoise_trainer.py +++ b/tests/trainers/test_image_denoise_trainer.py @@ -51,7 +51,7 @@ class ImageDenoiseTrainerTest(unittest.TestCase): shutil.rmtree(self.tmp_dir, ignore_errors=True) super().tearDown() - @unittest.skipUnless(test_level() >= 0, 'skip test in current test level') + @unittest.skipUnless(test_level() >= 2, 'skip test in current test level') def test_trainer(self): kwargs = dict( model=self.model_id, @@ -65,7 +65,7 @@ class ImageDenoiseTrainerTest(unittest.TestCase): for i in range(2): self.assertIn(f'epoch_{i+1}.pth', results_files) - @unittest.skipUnless(test_level() >= 1, 'skip test in current test level') + @unittest.skipUnless(test_level() >= 2, 'skip test in current test level') def test_trainer_with_model_and_args(self): model = NAFNetForImageDenoise.from_pretrained(self.cache_path) kwargs = dict( From 3cd5f73da076cc550b747473199585a80bc90917 Mon Sep 17 00:00:00 2001 From: "yingda.chen" Date: Thu, 20 Oct 2022 10:28:15 +0800 Subject: [PATCH 03/10] [to #42322933] refactor push_model Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/10463992 --- modelscope/hub/api.py | 128 +++++++++++++++++++++++++--- modelscope/hub/file_download.py | 3 +- modelscope/hub/git.py | 4 +- modelscope/hub/repository.py | 4 +- modelscope/hub/snapshot_download.py | 2 +- modelscope/hub/upload.py | 117 ------------------------- tests/hub/test_hub_operation.py | 2 +- tests/hub/test_hub_upload.py | 15 ++-- 8 files changed, 130 insertions(+), 145 deletions(-) delete mode 100644 modelscope/hub/upload.py diff --git a/modelscope/hub/api.py b/modelscope/hub/api.py index 2aab142e..f8ca683a 100644 --- a/modelscope/hub/api.py +++ b/modelscope/hub/api.py @@ -1,8 +1,11 @@ # Copyright (c) Alibaba, Inc. and its affiliates. +# yapf: disable +import datetime import os import pickle import shutil +import tempfile from collections import defaultdict from http import HTTPStatus from http.cookiejar import CookieJar @@ -16,17 +19,25 @@ from modelscope.hub.constants import (API_RESPONSE_FIELD_DATA, API_RESPONSE_FIELD_GIT_ACCESS_TOKEN, API_RESPONSE_FIELD_MESSAGE, API_RESPONSE_FIELD_USERNAME, - DEFAULT_CREDENTIALS_PATH) + DEFAULT_CREDENTIALS_PATH, Licenses, + ModelVisibility) +from modelscope.hub.errors import (InvalidParameter, NotExistError, + NotLoginException, RequestError, + datahub_raise_on_error, + handle_http_post_error, + handle_http_response, is_ok, raise_on_error) +from modelscope.hub.git import GitCommandWrapper +from modelscope.hub.repository import Repository +from modelscope.hub.utils.utils import (get_endpoint, + model_id_to_group_owner_name) from modelscope.utils.config_ds import DOWNLOADED_DATASETS_PATH from modelscope.utils.constant import (DEFAULT_DATASET_REVISION, DEFAULT_MODEL_REVISION, DatasetFormations, DatasetMetaFormats, - DownloadMode) + DownloadMode, ModelFile) from modelscope.utils.logger import get_logger -from .errors import (InvalidParameter, NotExistError, RequestError, - datahub_raise_on_error, handle_http_post_error, - handle_http_response, is_ok, raise_on_error) -from .utils.utils import get_endpoint, model_id_to_group_owner_name + +# yapf: enable logger = get_logger() @@ -169,11 +180,106 @@ class HubApi: else: r.raise_for_status() - def list_model(self, - owner_or_group: str, - page_number=1, - page_size=10) -> dict: - """List model in owner or group. + def push_model(self, + model_id: str, + model_dir: str, + visibility: int = ModelVisibility.PUBLIC, + license: str = Licenses.APACHE_V2, + chinese_name: Optional[str] = None, + commit_message: Optional[str] = 'upload model', + revision: Optional[str] = DEFAULT_MODEL_REVISION): + """ + Upload model from a given directory to given repository. A valid model directory + must contain a configuration.json file. + + This function upload the files in given directory to given repository. If the + given repository is not exists in remote, it will automatically create it with + given visibility, license and chinese_name parameters. If the revision is also + not exists in remote repository, it will create a new branch for it. + + This function must be called before calling HubApi's login with a valid token + which can be obtained from ModelScope's website. + + Args: + model_id (`str`): + The model id to be uploaded, caller must have write permission for it. + model_dir(`str`): + The Absolute Path of the finetune result. + visibility(`int`, defaults to `0`): + Visibility of the new created model(1-private, 5-public). If the model is + not exists in ModelScope, this function will create a new model with this + visibility and this parameter is required. You can ignore this parameter + if you make sure the model's existence. + license(`str`, defaults to `None`): + License of the new created model(see License). If the model is not exists + in ModelScope, this function will create a new model with this license + and this parameter is required. You can ignore this parameter if you + make sure the model's existence. + chinese_name(`str`, *optional*, defaults to `None`): + chinese name of the new created model. + commit_message(`str`, *optional*, defaults to `None`): + commit message of the push request. + revision (`str`, *optional*, default to DEFAULT_MODEL_REVISION): + which branch to push. If the branch is not exists, It will create a new + branch and push to it. + """ + if model_id is None: + raise InvalidParameter('model_id cannot be empty!') + if model_dir is None: + raise InvalidParameter('model_dir cannot be empty!') + if not os.path.exists(model_dir) or os.path.isfile(model_dir): + raise InvalidParameter('model_dir must be a valid directory.') + cfg_file = os.path.join(model_dir, ModelFile.CONFIGURATION) + if not os.path.exists(cfg_file): + raise ValueError(f'{model_dir} must contain a configuration.json.') + cookies = ModelScopeConfig.get_cookies() + if cookies is None: + raise NotLoginException('Must login before upload!') + files_to_save = os.listdir(model_dir) + try: + self.get_model(model_id=model_id) + except Exception: + if visibility is None or license is None: + raise InvalidParameter( + 'visibility and license cannot be empty if want to create new repo' + ) + logger.info('Create new model %s' % model_id) + self.create_model( + model_id=model_id, + visibility=visibility, + license=license, + chinese_name=chinese_name) + tmp_dir = tempfile.mkdtemp() + git_wrapper = GitCommandWrapper() + try: + repo = Repository(model_dir=tmp_dir, clone_from=model_id) + branches = git_wrapper.get_remote_branches(tmp_dir) + if revision not in branches: + logger.info('Create new branch %s' % revision) + git_wrapper.new_branch(tmp_dir, revision) + git_wrapper.checkout(tmp_dir, revision) + for f in files_to_save: + if f[0] != '.': + src = os.path.join(model_dir, f) + if os.path.isdir(src): + shutil.copytree(src, os.path.join(tmp_dir, f)) + else: + shutil.copy(src, tmp_dir) + if not commit_message: + date = datetime.datetime.now().strftime('%Y_%m_%d_%H_%M_%S') + commit_message = '[automsg] push model %s to hub at %s' % ( + model_id, date) + repo.push(commit_message=commit_message, branch=revision) + except Exception: + raise + finally: + shutil.rmtree(tmp_dir, ignore_errors=True) + + def list_models(self, + owner_or_group: str, + page_number=1, + page_size=10) -> dict: + """List models in owner or group. Args: owner_or_group(`str`): owner or group. diff --git a/modelscope/hub/file_download.py b/modelscope/hub/file_download.py index 1cc5645b..8ffc60bc 100644 --- a/modelscope/hub/file_download.py +++ b/modelscope/hub/file_download.py @@ -11,13 +11,12 @@ from typing import Dict, Optional, Union from uuid import uuid4 import requests -from filelock import FileLock from tqdm import tqdm from modelscope import __version__ +from modelscope.hub.api import HubApi, ModelScopeConfig from modelscope.utils.constant import DEFAULT_MODEL_REVISION from modelscope.utils.logger import get_logger -from .api import HubApi, ModelScopeConfig from .constants import FILE_HASH from .errors import FileDownloadError, NotExistError from .utils.caching import ModelFileSystemCache diff --git a/modelscope/hub/git.py b/modelscope/hub/git.py index 1369fddf..fe1d1554 100644 --- a/modelscope/hub/git.py +++ b/modelscope/hub/git.py @@ -1,13 +1,10 @@ # Copyright (c) Alibaba, Inc. and its affiliates. import os -import re import subprocess from typing import List -from xmlrpc.client import Boolean from modelscope.utils.logger import get_logger -from .api import ModelScopeConfig from .errors import GitError logger = get_logger() @@ -132,6 +129,7 @@ class GitCommandWrapper(metaclass=Singleton): return response def add_user_info(self, repo_base_dir, repo_name): + from modelscope.hub.api import ModelScopeConfig user_name, user_email = ModelScopeConfig.get_user_info() if user_name and user_email: # config user.name and user.email if exist diff --git a/modelscope/hub/repository.py b/modelscope/hub/repository.py index d92089ed..35c831a9 100644 --- a/modelscope/hub/repository.py +++ b/modelscope/hub/repository.py @@ -7,7 +7,6 @@ from modelscope.hub.errors import GitError, InvalidParameter, NotLoginException from modelscope.utils.constant import (DEFAULT_DATASET_REVISION, DEFAULT_MODEL_REVISION) from modelscope.utils.logger import get_logger -from .api import ModelScopeConfig from .git import GitCommandWrapper from .utils.utils import get_endpoint @@ -47,6 +46,7 @@ class Repository: err_msg = 'a non-default value of revision cannot be empty.' raise InvalidParameter(err_msg) + from modelscope.hub.api import ModelScopeConfig if auth_token: self.auth_token = auth_token else: @@ -166,7 +166,7 @@ class DatasetRepository: err_msg = 'a non-default value of revision cannot be empty.' raise InvalidParameter(err_msg) self.revision = revision - + from modelscope.hub.api import ModelScopeConfig if auth_token: self.auth_token = auth_token else: diff --git a/modelscope/hub/snapshot_download.py b/modelscope/hub/snapshot_download.py index cde6ad34..ac57d1b1 100644 --- a/modelscope/hub/snapshot_download.py +++ b/modelscope/hub/snapshot_download.py @@ -5,9 +5,9 @@ import tempfile from pathlib import Path from typing import Dict, Optional, Union +from modelscope.hub.api import HubApi, ModelScopeConfig from modelscope.utils.constant import DEFAULT_MODEL_REVISION from modelscope.utils.logger import get_logger -from .api import HubApi, ModelScopeConfig from .constants import FILE_HASH from .errors import NotExistError from .file_download import (get_file_download_url, http_get_file, diff --git a/modelscope/hub/upload.py b/modelscope/hub/upload.py deleted file mode 100644 index 9dffc60e..00000000 --- a/modelscope/hub/upload.py +++ /dev/null @@ -1,117 +0,0 @@ -# Copyright (c) Alibaba, Inc. and its affiliates. - -import datetime -import os -import shutil -import tempfile -import uuid -from typing import Dict, Optional -from uuid import uuid4 - -from filelock import FileLock - -from modelscope import __version__ -from modelscope.hub.api import HubApi, ModelScopeConfig -from modelscope.hub.errors import InvalidParameter, NotLoginException -from modelscope.hub.git import GitCommandWrapper -from modelscope.hub.repository import Repository -from modelscope.utils.constant import DEFAULT_MODEL_REVISION, ModelFile -from modelscope.utils.logger import get_logger - -logger = get_logger() - - -def upload_folder(model_id: str, - model_dir: str, - visibility: int = 0, - license: str = None, - chinese_name: Optional[str] = None, - commit_message: Optional[str] = None, - revision: Optional[str] = DEFAULT_MODEL_REVISION): - """ - Upload model from a given directory to given repository. A valid model directory - must contain a configuration.json file. - - This function upload the files in given directory to given repository. If the - given repository is not exists in remote, it will automatically create it with - given visibility, license and chinese_name parameters. If the revision is also - not exists in remote repository, it will create a new branch for it. - - This function must be called before calling HubApi's login with a valid token - which can be obtained from ModelScope's website. - - Args: - model_id (`str`): - The model id to be uploaded, caller must have write permission for it. - model_dir(`str`): - The Absolute Path of the finetune result. - visibility(`int`, defaults to `0`): - Visibility of the new created model(1-private, 5-public). If the model is - not exists in ModelScope, this function will create a new model with this - visibility and this parameter is required. You can ignore this parameter - if you make sure the model's existence. - license(`str`, defaults to `None`): - License of the new created model(see License). If the model is not exists - in ModelScope, this function will create a new model with this license - and this parameter is required. You can ignore this parameter if you - make sure the model's existence. - chinese_name(`str`, *optional*, defaults to `None`): - chinese name of the new created model. - commit_message(`str`, *optional*, defaults to `None`): - commit message of the push request. - revision (`str`, *optional*, default to DEFAULT_MODEL_REVISION): - which branch to push. If the branch is not exists, It will create a new - branch and push to it. - """ - if model_id is None: - raise InvalidParameter('model_id cannot be empty!') - if model_dir is None: - raise InvalidParameter('model_dir cannot be empty!') - if not os.path.exists(model_dir) or os.path.isfile(model_dir): - raise InvalidParameter('model_dir must be a valid directory.') - cfg_file = os.path.join(model_dir, ModelFile.CONFIGURATION) - if not os.path.exists(cfg_file): - raise ValueError(f'{model_dir} must contain a configuration.json.') - cookies = ModelScopeConfig.get_cookies() - if cookies is None: - raise NotLoginException('Must login before upload!') - files_to_save = os.listdir(model_dir) - api = HubApi() - try: - api.get_model(model_id=model_id) - except Exception: - if visibility is None or license is None: - raise InvalidParameter( - 'visibility and license cannot be empty if want to create new repo' - ) - logger.info('Create new model %s' % model_id) - api.create_model( - model_id=model_id, - visibility=visibility, - license=license, - chinese_name=chinese_name) - tmp_dir = tempfile.mkdtemp() - git_wrapper = GitCommandWrapper() - try: - repo = Repository(model_dir=tmp_dir, clone_from=model_id) - branches = git_wrapper.get_remote_branches(tmp_dir) - if revision not in branches: - logger.info('Create new branch %s' % revision) - git_wrapper.new_branch(tmp_dir, revision) - git_wrapper.checkout(tmp_dir, revision) - for f in files_to_save: - if f[0] != '.': - src = os.path.join(model_dir, f) - if os.path.isdir(src): - shutil.copytree(src, os.path.join(tmp_dir, f)) - else: - shutil.copy(src, tmp_dir) - if not commit_message: - date = datetime.datetime.now().strftime('%Y_%m_%d_%H_%M_%S') - commit_message = '[automsg] push model %s to hub at %s' % ( - model_id, date) - repo.push(commit_message=commit_message, branch=revision) - except Exception: - raise - finally: - shutil.rmtree(tmp_dir, ignore_errors=True) diff --git a/tests/hub/test_hub_operation.py b/tests/hub/test_hub_operation.py index c96db986..f2bdb2d3 100644 --- a/tests/hub/test_hub_operation.py +++ b/tests/hub/test_hub_operation.py @@ -127,7 +127,7 @@ class HubOperationTest(unittest.TestCase): return None def test_list_model(self): - data = self.api.list_model(TEST_MODEL_ORG) + data = self.api.list_models(TEST_MODEL_ORG) assert len(data['Models']) >= 1 diff --git a/tests/hub/test_hub_upload.py b/tests/hub/test_hub_upload.py index a10f7c81..e1f61467 100644 --- a/tests/hub/test_hub_upload.py +++ b/tests/hub/test_hub_upload.py @@ -9,7 +9,6 @@ from modelscope.hub.api import HubApi from modelscope.hub.constants import Licenses, ModelVisibility from modelscope.hub.errors import HTTPError, NotLoginException from modelscope.hub.repository import Repository -from modelscope.hub.upload import upload_folder from modelscope.utils.constant import ModelFile from modelscope.utils.logger import get_logger from modelscope.utils.test_utils import test_level @@ -54,14 +53,14 @@ class HubUploadTest(unittest.TestCase): license=Licenses.APACHE_V2) os.system("echo '111'>%s" % os.path.join(self.finetune_path, 'add1.py')) - upload_folder( + self.api.push_model( model_id=self.create_model_name, model_dir=self.finetune_path) Repository(model_dir=self.repo_path, clone_from=self.create_model_name) assert os.path.exists(os.path.join(self.repo_path, 'add1.py')) shutil.rmtree(self.repo_path, ignore_errors=True) os.system("echo '222'>%s" % os.path.join(self.finetune_path, 'add2.py')) - upload_folder( + self.api.push_model( model_id=self.create_model_name, model_dir=self.finetune_path, revision='new_revision/version1') @@ -73,7 +72,7 @@ class HubUploadTest(unittest.TestCase): shutil.rmtree(self.repo_path, ignore_errors=True) os.system("echo '333'>%s" % os.path.join(self.finetune_path, 'add3.py')) - upload_folder( + self.api.push_model( model_id=self.create_model_name, model_dir=self.finetune_path, revision='new_revision/version2', @@ -88,7 +87,7 @@ class HubUploadTest(unittest.TestCase): add4_path = os.path.join(self.finetune_path, 'temp') os.mkdir(add4_path) os.system("echo '444'>%s" % os.path.join(add4_path, 'add4.py')) - upload_folder( + self.api.push_model( model_id=self.create_model_name, model_dir=self.finetune_path, revision='new_revision/version1') @@ -105,7 +104,7 @@ class HubUploadTest(unittest.TestCase): self.api.login(TEST_ACCESS_TOKEN1) os.system("echo '111'>%s" % os.path.join(self.finetune_path, 'add1.py')) - upload_folder( + self.api.push_model( model_id=self.create_model_name, model_dir=self.finetune_path, revision='new_model_new_revision', @@ -124,7 +123,7 @@ class HubUploadTest(unittest.TestCase): self.api.login(TEST_ACCESS_TOKEN1) delete_credential() with self.assertRaises(NotLoginException): - upload_folder( + self.api.push_model( model_id=self.create_model_name, model_dir=self.finetune_path, visibility=ModelVisibility.PUBLIC, @@ -135,7 +134,7 @@ class HubUploadTest(unittest.TestCase): logger.info('test upload to invalid repo!') self.api.login(TEST_ACCESS_TOKEN1) with self.assertRaises(HTTPError): - upload_folder( + self.api.push_model( model_id='%s/%s' % ('speech_tts', 'invalid_model_test'), model_dir=self.finetune_path, visibility=ModelVisibility.PUBLIC, From eb7bd5b8256d474e4382ee73f19860bed3d21689 Mon Sep 17 00:00:00 2001 From: "jiaqi.sjq" Date: Thu, 20 Oct 2022 11:46:50 +0800 Subject: [PATCH 04/10] [to #42322933]Add lock to tts's AM and model loaded since its AR part not lockfree Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/10464620 Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/10464620 --- modelscope/models/audio/tts/voice.py | 78 +++++++++++++++------------- 1 file changed, 41 insertions(+), 37 deletions(-) diff --git a/modelscope/models/audio/tts/voice.py b/modelscope/models/audio/tts/voice.py index dc830db5..b7240088 100644 --- a/modelscope/models/audio/tts/voice.py +++ b/modelscope/models/audio/tts/voice.py @@ -2,6 +2,7 @@ import os import pickle as pkl +from threading import Lock import json import numpy as np @@ -27,6 +28,7 @@ class Voice: self.__am_config = AttrDict(**am_config) self.__voc_config = AttrDict(**voc_config) self.__model_loaded = False + self.__lock = Lock() if 'am' not in self.__am_config: raise TtsModelConfigurationException( 'modelscope error: am configuration invalid') @@ -71,34 +73,35 @@ class Voice: self.__generator.remove_weight_norm() def __am_forward(self, symbol_seq): - with torch.no_grad(): - inputs_feat_lst = self.__ling_unit.encode_symbol_sequence( - symbol_seq) - inputs_sy = torch.from_numpy(inputs_feat_lst[0]).long().to( - self.__device) - inputs_tone = torch.from_numpy(inputs_feat_lst[1]).long().to( - self.__device) - inputs_syllable = torch.from_numpy(inputs_feat_lst[2]).long().to( - self.__device) - inputs_ws = torch.from_numpy(inputs_feat_lst[3]).long().to( - self.__device) - inputs_ling = torch.stack( - [inputs_sy, inputs_tone, inputs_syllable, inputs_ws], - dim=-1).unsqueeze(0) - inputs_emo = torch.from_numpy(inputs_feat_lst[4]).long().to( - self.__device).unsqueeze(0) - inputs_spk = torch.from_numpy(inputs_feat_lst[5]).long().to( - self.__device).unsqueeze(0) - inputs_len = torch.zeros(1).to(self.__device).long( - ) + inputs_emo.size(1) - 1 # minus 1 for "~" - res = self.__am_net(inputs_ling[:, :-1, :], inputs_emo[:, :-1], - inputs_spk[:, :-1], inputs_len) - postnet_outputs = res['postnet_outputs'] - LR_length_rounded = res['LR_length_rounded'] - valid_length = int(LR_length_rounded[0].item()) - postnet_outputs = postnet_outputs[ - 0, :valid_length, :].cpu().numpy() - return postnet_outputs + with self.__lock: + with torch.no_grad(): + inputs_feat_lst = self.__ling_unit.encode_symbol_sequence( + symbol_seq) + inputs_sy = torch.from_numpy(inputs_feat_lst[0]).long().to( + self.__device) + inputs_tone = torch.from_numpy(inputs_feat_lst[1]).long().to( + self.__device) + inputs_syllable = torch.from_numpy( + inputs_feat_lst[2]).long().to(self.__device) + inputs_ws = torch.from_numpy(inputs_feat_lst[3]).long().to( + self.__device) + inputs_ling = torch.stack( + [inputs_sy, inputs_tone, inputs_syllable, inputs_ws], + dim=-1).unsqueeze(0) + inputs_emo = torch.from_numpy(inputs_feat_lst[4]).long().to( + self.__device).unsqueeze(0) + inputs_spk = torch.from_numpy(inputs_feat_lst[5]).long().to( + self.__device).unsqueeze(0) + inputs_len = torch.zeros(1).to(self.__device).long( + ) + inputs_emo.size(1) - 1 # minus 1 for "~" + res = self.__am_net(inputs_ling[:, :-1, :], inputs_emo[:, :-1], + inputs_spk[:, :-1], inputs_len) + postnet_outputs = res['postnet_outputs'] + LR_length_rounded = res['LR_length_rounded'] + valid_length = int(LR_length_rounded[0].item()) + postnet_outputs = postnet_outputs[ + 0, :valid_length, :].cpu().numpy() + return postnet_outputs def __vocoder_forward(self, melspec): dim0 = list(melspec.shape)[-1] @@ -118,14 +121,15 @@ class Voice: return audio def forward(self, symbol_seq): - if not self.__model_loaded: - torch.manual_seed(self.__am_config.seed) - if torch.cuda.is_available(): + with self.__lock: + if not self.__model_loaded: torch.manual_seed(self.__am_config.seed) - self.__device = torch.device('cuda') - else: - self.__device = torch.device('cpu') - self.__load_am() - self.__load_vocoder() - self.__model_loaded = True + if torch.cuda.is_available(): + torch.manual_seed(self.__am_config.seed) + self.__device = torch.device('cuda') + else: + self.__device = torch.device('cpu') + self.__load_am() + self.__load_vocoder() + self.__model_loaded = True return self.__vocoder_forward(self.__am_forward(symbol_seq)) From 01d521dd7851ee273de3793d7ad123ad99c0de9c Mon Sep 17 00:00:00 2001 From: "shouzhou.bx" Date: Thu, 20 Oct 2022 11:51:52 +0800 Subject: [PATCH 05/10] [to #42322933]add face 2d keypoints finetune test case Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/10421808 * add face 2d keypoints & human wholebody keypoint finrtune test case --- modelscope/metainfo.py | 4 +- modelscope/msdatasets/cv/easycv_base.py | 13 ++-- requirements/cv.txt | 2 +- .../test_easycv_trainer_face_2d_keypoints.py | 71 +++++++++++++++++++ 4 files changed, 83 insertions(+), 7 deletions(-) create mode 100644 tests/trainers/easycv/test_easycv_trainer_face_2d_keypoints.py diff --git a/modelscope/metainfo.py b/modelscope/metainfo.py index 6c8d91fa..31bef3b8 100644 --- a/modelscope/metainfo.py +++ b/modelscope/metainfo.py @@ -452,9 +452,9 @@ class Datasets(object): """ Names for different datasets. """ ClsDataset = 'ClsDataset' - Face2dKeypointsDataset = 'Face2dKeypointsDataset' + Face2dKeypointsDataset = 'FaceKeypointDataset' HandCocoWholeBodyDataset = 'HandCocoWholeBodyDataset' - HumanWholeBodyKeypointDataset = 'HumanWholeBodyKeypointDataset' + HumanWholeBodyKeypointDataset = 'WholeBodyCocoTopDownDataset' SegDataset = 'SegDataset' DetDataset = 'DetDataset' DetImagesMixDataset = 'DetImagesMixDataset' diff --git a/modelscope/msdatasets/cv/easycv_base.py b/modelscope/msdatasets/cv/easycv_base.py index a45827a3..7b6df6e0 100644 --- a/modelscope/msdatasets/cv/easycv_base.py +++ b/modelscope/msdatasets/cv/easycv_base.py @@ -26,11 +26,16 @@ class EasyCVBaseDataset(object): if self.split_config is not None: self._update_data_source(kwargs['data_source']) + def _update_data_root(self, input_dict, data_root): + for k, v in input_dict.items(): + if isinstance(v, str) and self.DATA_ROOT_PATTERN in v: + input_dict.update( + {k: v.replace(self.DATA_ROOT_PATTERN, data_root)}) + elif isinstance(v, dict): + self._update_data_root(v, data_root) + def _update_data_source(self, data_source): data_root = next(iter(self.split_config.values())) data_root = data_root.rstrip(osp.sep) - for k, v in data_source.items(): - if isinstance(v, str) and self.DATA_ROOT_PATTERN in v: - data_source.update( - {k: v.replace(self.DATA_ROOT_PATTERN, data_root)}) + self._update_data_root(data_source, data_root) diff --git a/requirements/cv.txt b/requirements/cv.txt index d23fab3a..f29b296b 100644 --- a/requirements/cv.txt +++ b/requirements/cv.txt @@ -19,7 +19,7 @@ moviepy>=1.0.3 networkx>=2.5 numba onnxruntime>=1.10 -pai-easycv>=0.6.3.7 +pai-easycv>=0.6.3.9 pandas psutil regex diff --git a/tests/trainers/easycv/test_easycv_trainer_face_2d_keypoints.py b/tests/trainers/easycv/test_easycv_trainer_face_2d_keypoints.py new file mode 100644 index 00000000..4dffa998 --- /dev/null +++ b/tests/trainers/easycv/test_easycv_trainer_face_2d_keypoints.py @@ -0,0 +1,71 @@ +# Copyright (c) Alibaba, Inc. and its affiliates. +import glob +import os +import shutil +import tempfile +import unittest + +import torch + +from modelscope.metainfo import Trainers +from modelscope.msdatasets import MsDataset +from modelscope.trainers import build_trainer +from modelscope.utils.constant import DownloadMode, LogKeys, Tasks +from modelscope.utils.logger import get_logger +from modelscope.utils.test_utils import test_level + + +@unittest.skipIf(not torch.cuda.is_available(), 'cuda unittest') +class EasyCVTrainerTestFace2DKeypoints(unittest.TestCase): + model_id = 'damo/cv_mobilenet_face-2d-keypoints_alignment' + + def setUp(self): + self.logger = get_logger() + self.logger.info(('Testing %s.%s' % + (type(self).__name__, self._testMethodName))) + + def _train(self, tmp_dir): + cfg_options = {'train.max_epochs': 2} + + trainer_name = Trainers.easycv + + train_dataset = MsDataset.load( + dataset_name='face_2d_keypoints_dataset', + namespace='modelscope', + split='train', + download_mode=DownloadMode.REUSE_DATASET_IF_EXISTS) + eval_dataset = MsDataset.load( + dataset_name='face_2d_keypoints_dataset', + namespace='modelscope', + split='train', + download_mode=DownloadMode.REUSE_DATASET_IF_EXISTS) + + kwargs = dict( + model=self.model_id, + train_dataset=train_dataset, + eval_dataset=eval_dataset, + work_dir=tmp_dir, + cfg_options=cfg_options) + + trainer = build_trainer(trainer_name, kwargs) + trainer.train() + + @unittest.skipUnless(test_level() >= 0, 'skip test in current test level') + def test_trainer_single_gpu(self): + temp_file_dir = tempfile.TemporaryDirectory() + tmp_dir = temp_file_dir.name + if not os.path.exists(tmp_dir): + os.makedirs(tmp_dir) + + self._train(tmp_dir) + + results_files = os.listdir(tmp_dir) + json_files = glob.glob(os.path.join(tmp_dir, '*.log.json')) + self.assertEqual(len(json_files), 1) + self.assertIn(f'{LogKeys.EPOCH}_2.pth', results_files) + + temp_file_dir.cleanup() + + +if __name__ == '__main__': + unittest.main() From 535acaef5bc01f8d6059e6acf364eaebb1147f7b Mon Sep 17 00:00:00 2001 From: "wenmeng.zwm" Date: Thu, 20 Oct 2022 12:13:19 +0800 Subject: [PATCH 06/10] [to #42322933]add test case to check xtcocotools availbility Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/10462622 * add test case to check xtcocotools availbility --- tests/utils/test_compatibility.py | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+) create mode 100644 tests/utils/test_compatibility.py diff --git a/tests/utils/test_compatibility.py b/tests/utils/test_compatibility.py new file mode 100644 index 00000000..f5222261 --- /dev/null +++ b/tests/utils/test_compatibility.py @@ -0,0 +1,19 @@ +# Copyright (c) Alibaba, Inc. and its affiliates. + +import unittest + + +class CompatibilityTest(unittest.TestCase): + + def setUp(self): + print(('Testing %s.%s' % (type(self).__name__, self._testMethodName))) + + def tearDown(self): + super().tearDown() + + def test_xtcocotools(self): + from xtcocotools.coco import COCO + + +if __name__ == '__main__': + unittest.main() From e7c7be6aae33537e25cf4187336a0c46479a75f3 Mon Sep 17 00:00:00 2001 From: "xingguang.zxg" Date: Thu, 20 Oct 2022 12:51:48 +0800 Subject: [PATCH 07/10] fix cpu error Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/10432300 * fix cpu error --- modelscope/models/cv/text_driven_segmentation/lseg_model.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modelscope/models/cv/text_driven_segmentation/lseg_model.py b/modelscope/models/cv/text_driven_segmentation/lseg_model.py index 9a5754c6..ec381356 100644 --- a/modelscope/models/cv/text_driven_segmentation/lseg_model.py +++ b/modelscope/models/cv/text_driven_segmentation/lseg_model.py @@ -93,7 +93,7 @@ class TextDrivenSeg(TorchModel): """ with torch.no_grad(): if self.device_id == -1: - output = self.model(image) + output = self.model(image, [text]) else: device = torch.device('cuda', self.device_id) output = self.model(image.to(device), [text]) From 1483c64638a002279443752adc195df6ef2e7494 Mon Sep 17 00:00:00 2001 From: "shichen.fsc" Date: Thu, 20 Oct 2022 12:54:37 +0800 Subject: [PATCH 08/10] [to #42322933] Fix ASR error when resample failed, and add all asr models UT, add apply-cmvn for pytorch models Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/10465241 --- data/test/audios/asr_example_8K.wav | 3 + data/test/audios/asr_example_cn_dialect.wav | 3 + data/test/audios/asr_example_cn_en.wav | 3 + data/test/audios/asr_example_en.wav | 3 + data/test/audios/asr_example_es.wav | 3 + data/test/audios/asr_example_id.wav | 3 + data/test/audios/asr_example_ja.wav | 3 + data/test/audios/asr_example_ko.wav | 3 + data/test/audios/asr_example_ru.wav | 3 + .../pipelines/audio/asr_inference_pipeline.py | 19 +- .../pipelines/audio/kws_kwsbp_pipeline.py | 4 +- modelscope/preprocessors/asr.py | 6 + modelscope/utils/audio/audio_utils.py | 13 +- .../test_automatic_speech_recognition.py | 267 ++++++++++++++---- 14 files changed, 270 insertions(+), 66 deletions(-) create mode 100644 data/test/audios/asr_example_8K.wav create mode 100644 data/test/audios/asr_example_cn_dialect.wav create mode 100644 data/test/audios/asr_example_cn_en.wav create mode 100644 data/test/audios/asr_example_en.wav create mode 100644 data/test/audios/asr_example_es.wav create mode 100644 data/test/audios/asr_example_id.wav create mode 100644 data/test/audios/asr_example_ja.wav create mode 100644 data/test/audios/asr_example_ko.wav create mode 100644 data/test/audios/asr_example_ru.wav diff --git a/data/test/audios/asr_example_8K.wav b/data/test/audios/asr_example_8K.wav new file mode 100644 index 00000000..956aad27 --- /dev/null +++ b/data/test/audios/asr_example_8K.wav @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e999c247bfebb03d556a31722f0ce7145cac20a67fac9da813ad336e1f549f9f +size 38954 diff --git a/data/test/audios/asr_example_cn_dialect.wav b/data/test/audios/asr_example_cn_dialect.wav new file mode 100644 index 00000000..e18fb05d --- /dev/null +++ b/data/test/audios/asr_example_cn_dialect.wav @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32eb8d4d537941bf0edea69cd6723e8ba489fa3df64e13e29f96e4fae0b856f4 +size 93676 diff --git a/data/test/audios/asr_example_cn_en.wav b/data/test/audios/asr_example_cn_en.wav new file mode 100644 index 00000000..8baf3193 --- /dev/null +++ b/data/test/audios/asr_example_cn_en.wav @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f57aee13ade70be6b2c6e4f5e5c7404bdb03057b63828baefbaadcf23855a4cb +size 472012 diff --git a/data/test/audios/asr_example_en.wav b/data/test/audios/asr_example_en.wav new file mode 100644 index 00000000..fa996eec --- /dev/null +++ b/data/test/audios/asr_example_en.wav @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fee8e0460ca707f108782be0d93c555bf34fb6b1cb297e5fceed70192cc65f9b +size 71244 diff --git a/data/test/audios/asr_example_es.wav b/data/test/audios/asr_example_es.wav new file mode 100644 index 00000000..95b22dc3 --- /dev/null +++ b/data/test/audios/asr_example_es.wav @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:450e31f9df8c5b48c617900625f01cb64c484f079a9843179fe9feaa7d163e61 +size 181964 diff --git a/data/test/audios/asr_example_id.wav b/data/test/audios/asr_example_id.wav new file mode 100644 index 00000000..54c30614 --- /dev/null +++ b/data/test/audios/asr_example_id.wav @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:255494c41bc1dfb0c954d827ec6ce775900e4f7a55fb0a7881bdf9d66a03b425 +size 112078 diff --git a/data/test/audios/asr_example_ja.wav b/data/test/audios/asr_example_ja.wav new file mode 100644 index 00000000..e953fee2 --- /dev/null +++ b/data/test/audios/asr_example_ja.wav @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22a55277908bbc3ef60a0cf56b230eb507b9e837574e8f493e93644b1d21c281 +size 200556 diff --git a/data/test/audios/asr_example_ko.wav b/data/test/audios/asr_example_ko.wav new file mode 100644 index 00000000..0dad1be3 --- /dev/null +++ b/data/test/audios/asr_example_ko.wav @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee92191836c76412463d8b282a7ab4e1aa57386ba699ec011a3e2c4d64f32f4b +size 162636 diff --git a/data/test/audios/asr_example_ru.wav b/data/test/audios/asr_example_ru.wav new file mode 100644 index 00000000..b0cb8f2f --- /dev/null +++ b/data/test/audios/asr_example_ru.wav @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77d1537fc584c1505d8aa10ec8c86af57ab661199e4f28fd7ffee3c22d1e4e61 +size 160204 diff --git a/modelscope/pipelines/audio/asr_inference_pipeline.py b/modelscope/pipelines/audio/asr_inference_pipeline.py index 4e8b658d..6a4864bf 100644 --- a/modelscope/pipelines/audio/asr_inference_pipeline.py +++ b/modelscope/pipelines/audio/asr_inference_pipeline.py @@ -47,22 +47,28 @@ class AutomaticSpeechRecognitionPipeline(Pipeline): if isinstance(audio_in, str): # load pcm data from url if audio_in is url str - self.audio_in = load_bytes_from_url(audio_in) + self.audio_in, checking_audio_fs = load_bytes_from_url(audio_in) elif isinstance(audio_in, bytes): # load pcm data from wav data if audio_in is wave format - self.audio_in = extract_pcm_from_wav(audio_in) + self.audio_in, checking_audio_fs = extract_pcm_from_wav(audio_in) else: self.audio_in = audio_in + # set the sample_rate of audio_in if checking_audio_fs is valid + if checking_audio_fs is not None: + self.audio_fs = checking_audio_fs + if recog_type is None or audio_format is None: self.recog_type, self.audio_format, self.audio_in = asr_utils.type_checking( audio_in=self.audio_in, recog_type=recog_type, audio_format=audio_format) - if hasattr(asr_utils, 'sample_rate_checking') and audio_fs is None: - self.audio_fs = asr_utils.sample_rate_checking( + if hasattr(asr_utils, 'sample_rate_checking'): + checking_audio_fs = asr_utils.sample_rate_checking( self.audio_in, self.audio_format) + if checking_audio_fs is not None: + self.audio_fs = checking_audio_fs if self.preprocessor is None: self.preprocessor = WavToScp() @@ -80,7 +86,7 @@ class AutomaticSpeechRecognitionPipeline(Pipeline): logger.info(f"Decoding with {inputs['audio_format']} files ...") - data_cmd: Sequence[Tuple[str, str]] + data_cmd: Sequence[Tuple[str, str, str]] if inputs['audio_format'] == 'wav' or inputs['audio_format'] == 'pcm': data_cmd = ['speech', 'sound'] elif inputs['audio_format'] == 'kaldi_ark': @@ -88,6 +94,9 @@ class AutomaticSpeechRecognitionPipeline(Pipeline): elif inputs['audio_format'] == 'tfrecord': data_cmd = ['speech', 'tfrecord'] + if inputs.__contains__('mvn_file'): + data_cmd.append(inputs['mvn_file']) + # generate asr inference command cmd = { 'model_type': inputs['model_type'], diff --git a/modelscope/pipelines/audio/kws_kwsbp_pipeline.py b/modelscope/pipelines/audio/kws_kwsbp_pipeline.py index 5555c9e6..db6fc65d 100644 --- a/modelscope/pipelines/audio/kws_kwsbp_pipeline.py +++ b/modelscope/pipelines/audio/kws_kwsbp_pipeline.py @@ -51,10 +51,10 @@ class KeyWordSpottingKwsbpPipeline(Pipeline): if isinstance(audio_in, str): # load pcm data from url if audio_in is url str - audio_in = load_bytes_from_url(audio_in) + audio_in, audio_fs = load_bytes_from_url(audio_in) elif isinstance(audio_in, bytes): # load pcm data from wav data if audio_in is wave format - audio_in = extract_pcm_from_wav(audio_in) + audio_in, audio_fs = extract_pcm_from_wav(audio_in) output = self.preprocessor.forward(self.model.forward(), audio_in) output = self.forward(output) diff --git a/modelscope/preprocessors/asr.py b/modelscope/preprocessors/asr.py index facaa132..91bf5860 100644 --- a/modelscope/preprocessors/asr.py +++ b/modelscope/preprocessors/asr.py @@ -133,6 +133,12 @@ class WavToScp(Preprocessor): else: inputs['asr_model_config'] = asr_model_config + if inputs['model_config'].__contains__('mvn_file'): + mvn_file = os.path.join(inputs['model_workspace'], + inputs['model_config']['mvn_file']) + assert os.path.exists(mvn_file), 'mvn_file does not exist' + inputs['mvn_file'] = mvn_file + elif inputs['model_type'] == Frameworks.tf: assert inputs['model_config'].__contains__( 'vocab_file'), 'vocab_file does not exist' diff --git a/modelscope/utils/audio/audio_utils.py b/modelscope/utils/audio/audio_utils.py index 647d9521..32e2fa54 100644 --- a/modelscope/utils/audio/audio_utils.py +++ b/modelscope/utils/audio/audio_utils.py @@ -57,6 +57,7 @@ def update_conf(origin_config_file, new_config_file, conf_item: [str, str]): def extract_pcm_from_wav(wav: bytes) -> bytes: data = wav + sample_rate = None if len(data) > 44: frame_len = 44 file_len = len(data) @@ -70,29 +71,33 @@ def extract_pcm_from_wav(wav: bytes) -> bytes: 'Subchunk1ID'] == 'fmt ': header_fields['SubChunk1Size'] = struct.unpack( ' Union[bytes, str]: + sample_rate = None result = urlparse(url) if result.scheme is not None and len(result.scheme) > 0: storage = HTTPStorage() data = storage.read(url) - data = extract_pcm_from_wav(data) + data, sample_rate = extract_pcm_from_wav(data) else: data = url - return data + return data, sample_rate diff --git a/tests/pipelines/test_automatic_speech_recognition.py b/tests/pipelines/test_automatic_speech_recognition.py index 303fb6b9..c37a6a3f 100644 --- a/tests/pipelines/test_automatic_speech_recognition.py +++ b/tests/pipelines/test_automatic_speech_recognition.py @@ -45,6 +45,10 @@ class AutomaticSpeechRecognitionTest(unittest.TestCase, 'checking_item': OutputKeys.TEXT, 'example': 'wav_example' }, + 'test_run_with_url_pytorch': { + 'checking_item': OutputKeys.TEXT, + 'example': 'wav_example' + }, 'test_run_with_url_tf': { 'checking_item': OutputKeys.TEXT, 'example': 'wav_example' @@ -74,6 +78,170 @@ class AutomaticSpeechRecognitionTest(unittest.TestCase, } } + all_models_info = [ + { + 'model_group': 'damo', + 'model_id': + 'speech_paraformer_asr_nat-zh-cn-16k-common-vocab8358-tensorflow1', + 'wav_path': 'data/test/audios/asr_example.wav' + }, + { + 'model_group': 'damo', + 'model_id': 'speech_paraformer_asr_nat-aishell1-pytorch', + 'wav_path': 'data/test/audios/asr_example.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8358-tensorflow1', + 'wav_path': 'data/test/audios/asr_example.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_paraformer_asr_nat-zh-cn-8k-common-vocab8358-tensorflow1', + 'wav_path': 'data/test/audios/asr_example_8K.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-zh-cn-16k-common-vocab8358-tensorflow1-online', + 'wav_path': 'data/test/audios/asr_example.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-zh-cn-16k-common-vocab8358-tensorflow1-offline', + 'wav_path': 'data/test/audios/asr_example.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-zh-cn-8k-common-vocab8358-tensorflow1-online', + 'wav_path': 'data/test/audios/asr_example_8K.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-zh-cn-8k-common-vocab8358-tensorflow1-offline', + 'wav_path': 'data/test/audios/asr_example_8K.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR-large_asr_2pass-zh-cn-16k-common-vocab8358-tensorflow1-offline', + 'wav_path': 'data/test/audios/asr_example.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-cn-en-moe-16k-vocab8358-tensorflow1-online', + 'wav_path': 'data/test/audios/asr_example_cn_en.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-cn-en-moe-16k-vocab8358-tensorflow1-offline', + 'wav_path': 'data/test/audios/asr_example_cn_en.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-cn-dialect-16k-vocab8358-tensorflow1-online', + 'wav_path': 'data/test/audios/asr_example_cn_dialect.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-cn-dialect-16k-vocab8358-tensorflow1-offline', + 'wav_path': 'data/test/audios/asr_example_cn_dialect.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_paraformer_asr_nat-zh-cn-16k-common-vocab3444-tensorflow1-online', + 'wav_path': 'data/test/audios/asr_example.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_paraformer_asr_nat-zh-cn-8k-common-vocab3444-tensorflow1-online', + 'wav_path': 'data/test/audios/asr_example_8K.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-en-16k-common-vocab1080-tensorflow1-offline', + 'wav_path': 'data/test/audios/asr_example_en.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-en-16k-common-vocab1080-tensorflow1-online', + 'wav_path': 'data/test/audios/asr_example_en.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-ru-16k-common-vocab1664-tensorflow1-offline', + 'wav_path': 'data/test/audios/asr_example_ru.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-ru-16k-common-vocab1664-tensorflow1-online', + 'wav_path': 'data/test/audios/asr_example_ru.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-es-16k-common-vocab3445-tensorflow1-offline', + 'wav_path': 'data/test/audios/asr_example_es.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-es-16k-common-vocab3445-tensorflow1-online', + 'wav_path': 'data/test/audios/asr_example_es.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-ko-16k-common-vocab6400-tensorflow1-offline', + 'wav_path': 'data/test/audios/asr_example_ko.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-ko-16k-common-vocab6400-tensorflow1-online', + 'wav_path': 'data/test/audios/asr_example_ko.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-ja-16k-common-vocab93-tensorflow1-online', + 'wav_path': 'data/test/audios/asr_example_ja.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-ja-16k-common-vocab93-tensorflow1-offline', + 'wav_path': 'data/test/audios/asr_example_ja.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-id-16k-common-vocab1067-tensorflow1-online', + 'wav_path': 'data/test/audios/asr_example_id.wav' + }, + { + 'model_group': 'damo', + 'model_id': + 'speech_UniASR_asr_2pass-id-16k-common-vocab1067-tensorflow1-offline', + 'wav_path': 'data/test/audios/asr_example_id.wav' + }, + ] + def setUp(self) -> None: self.am_pytorch_model_id = 'damo/speech_paraformer_asr_nat-aishell1-pytorch' self.am_tf_model_id = 'damo/speech_paraformer_asr_nat-zh-cn-16k-common-vocab8358-tensorflow1' @@ -90,7 +258,7 @@ class AutomaticSpeechRecognitionTest(unittest.TestCase, def run_pipeline(self, model_id: str, audio_in: Union[str, bytes], - sr: int = 16000) -> Dict[str, Any]: + sr: int = None) -> Dict[str, Any]: inference_16k_pipline = pipeline( task=Tasks.auto_speech_recognition, model=model_id) @@ -136,33 +304,26 @@ class AutomaticSpeechRecognitionTest(unittest.TestCase, return audio, fs @unittest.skipUnless(test_level() >= 0, 'skip test in current test level') - def test_run_with_wav_pytorch(self): - """run with single waveform file + def test_run_with_pcm(self): + """run with wav data """ - logger.info('Run ASR test with waveform file (pytorch)...') + logger.info('Run ASR test with wav data (tensorflow)...') - wav_file_path = os.path.join(os.getcwd(), WAV_FILE) + audio, sr = self.wav2bytes(os.path.join(os.getcwd(), WAV_FILE)) rec_result = self.run_pipeline( - model_id=self.am_pytorch_model_id, audio_in=wav_file_path) - self.check_result('test_run_with_wav_pytorch', rec_result) - - @unittest.skipUnless(test_level() >= 0, 'skip test in current test level') - def test_run_with_pcm_pytorch(self): - """run with wav data - """ + model_id=self.am_tf_model_id, audio_in=audio, sr=sr) + self.check_result('test_run_with_pcm_tf', rec_result) logger.info('Run ASR test with wav data (pytorch)...') - audio, sr = self.wav2bytes(os.path.join(os.getcwd(), WAV_FILE)) - rec_result = self.run_pipeline( model_id=self.am_pytorch_model_id, audio_in=audio, sr=sr) self.check_result('test_run_with_pcm_pytorch', rec_result) @unittest.skipUnless(test_level() >= 0, 'skip test in current test level') - def test_run_with_wav_tf(self): + def test_run_with_wav(self): """run with single waveform file """ @@ -174,21 +335,14 @@ class AutomaticSpeechRecognitionTest(unittest.TestCase, model_id=self.am_tf_model_id, audio_in=wav_file_path) self.check_result('test_run_with_wav_tf', rec_result) - @unittest.skipUnless(test_level() >= 0, 'skip test in current test level') - def test_run_with_pcm_tf(self): - """run with wav data - """ - - logger.info('Run ASR test with wav data (tensorflow)...') - - audio, sr = self.wav2bytes(os.path.join(os.getcwd(), WAV_FILE)) + logger.info('Run ASR test with waveform file (pytorch)...') rec_result = self.run_pipeline( - model_id=self.am_tf_model_id, audio_in=audio, sr=sr) - self.check_result('test_run_with_pcm_tf', rec_result) + model_id=self.am_pytorch_model_id, audio_in=wav_file_path) + self.check_result('test_run_with_wav_pytorch', rec_result) @unittest.skipUnless(test_level() >= 0, 'skip test in current test level') - def test_run_with_url_tf(self): + def test_run_with_url(self): """run with single url file """ @@ -198,6 +352,12 @@ class AutomaticSpeechRecognitionTest(unittest.TestCase, model_id=self.am_tf_model_id, audio_in=URL_FILE) self.check_result('test_run_with_url_tf', rec_result) + logger.info('Run ASR test with url file (pytorch)...') + + rec_result = self.run_pipeline( + model_id=self.am_pytorch_model_id, audio_in=URL_FILE) + self.check_result('test_run_with_url_pytorch', rec_result) + @unittest.skipUnless(test_level() >= 1, 'skip test in current test level') def test_run_with_wav_dataset_pytorch(self): """run with datasets, and audio format is waveform @@ -217,7 +377,6 @@ class AutomaticSpeechRecognitionTest(unittest.TestCase, data.text # hypothesis text """ - logger.info('Run ASR test with waveform dataset (pytorch)...') logger.info('Downloading waveform testsets file ...') dataset_path = download_and_untar( @@ -225,40 +384,38 @@ class AutomaticSpeechRecognitionTest(unittest.TestCase, LITTLE_TESTSETS_URL, self.workspace) dataset_path = os.path.join(dataset_path, 'wav', 'test') + logger.info('Run ASR test with waveform dataset (tensorflow)...') + + rec_result = self.run_pipeline( + model_id=self.am_tf_model_id, audio_in=dataset_path) + self.check_result('test_run_with_wav_dataset_tf', rec_result) + + logger.info('Run ASR test with waveform dataset (pytorch)...') + rec_result = self.run_pipeline( model_id=self.am_pytorch_model_id, audio_in=dataset_path) self.check_result('test_run_with_wav_dataset_pytorch', rec_result) - @unittest.skipUnless(test_level() >= 1, 'skip test in current test level') - def test_run_with_wav_dataset_tf(self): - """run with datasets, and audio format is waveform - datasets directory: - - wav - test # testsets - xx.wav - ... - dev # devsets - yy.wav - ... - train # trainsets - zz.wav - ... - transcript - data.text # hypothesis text + @unittest.skipUnless(test_level() >= 2, 'skip test in current test level') + def test_run_with_all_models(self): + """run with all models """ - logger.info('Run ASR test with waveform dataset (tensorflow)...') - logger.info('Downloading waveform testsets file ...') - - dataset_path = download_and_untar( - os.path.join(self.workspace, LITTLE_TESTSETS_FILE), - LITTLE_TESTSETS_URL, self.workspace) - dataset_path = os.path.join(dataset_path, 'wav', 'test') - - rec_result = self.run_pipeline( - model_id=self.am_tf_model_id, audio_in=dataset_path) - self.check_result('test_run_with_wav_dataset_tf', rec_result) + logger.info('Run ASR test with all models') + + for item in self.all_models_info: + model_id = item['model_group'] + '/' + item['model_id'] + wav_path = item['wav_path'] + rec_result = self.run_pipeline( + model_id=model_id, audio_in=wav_path) + if rec_result.__contains__(OutputKeys.TEXT): + logger.info(ColorCodes.MAGENTA + str(item['model_id']) + ' ' + + ColorCodes.YELLOW + + str(rec_result[OutputKeys.TEXT]) + + ColorCodes.END) + else: + logger.info(ColorCodes.MAGENTA + str(rec_result) + + ColorCodes.END) @unittest.skip('demo compatibility test is only enabled on a needed-basis') def test_demo_compatibility(self): From acba1786b0ac263140e2743b51fc7fc7c36d0980 Mon Sep 17 00:00:00 2001 From: "yuze.zyz" Date: Thu, 20 Oct 2022 15:29:34 +0800 Subject: [PATCH 09/10] [to #42322933] Fix bug in UT daily 1. Fix bugs in daily test 2. Fix a bug that the updating of lr is before the first time of updating of optimizer TODO this will still cause warnings when GA is above 1 3. Remove the judgement of mode in text-classification's preprocessor to fit the base trainer(Bug) Update some regression bins to fit the preprocessor 4. Update the regression tool to let outer code modify atol and rtol 5. Add the default metric for text-classification task 6. Remove the useless ckpt conversion method in bert to avoid the requirement of tf when loading modeling_bert Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/10430764 --- data/test/regression/sbert-base-tnews.bin | 3 + data/test/regression/sbert_nli.bin | 4 +- data/test/regression/sbert_sen_sim.bin | 4 +- ...rt_for_sequence_classification_exporter.py | 4 +- modelscope/metrics/builder.py | 1 + modelscope/models/nlp/bert/modeling_bert.py | 78 --------------- modelscope/preprocessors/nlp/nlp_base.py | 7 +- .../trainers/hooks/lr_scheduler_hook.py | 2 +- modelscope/trainers/trainer.py | 2 +- modelscope/utils/regress_test_utils.py | 94 ++++++++++++------- tests/msdatasets/test_ms_dataset.py | 3 +- .../test_finetune_sequence_classification.py | 33 ++++++- tests/trainers/test_trainer_with_nlp.py | 24 +++-- 13 files changed, 124 insertions(+), 135 deletions(-) create mode 100644 data/test/regression/sbert-base-tnews.bin diff --git a/data/test/regression/sbert-base-tnews.bin b/data/test/regression/sbert-base-tnews.bin new file mode 100644 index 00000000..1546860f --- /dev/null +++ b/data/test/regression/sbert-base-tnews.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2bce1341f4b55d536771dad6e2b280458579f46c3216474ceb8a926022ab53d0 +size 151572 diff --git a/data/test/regression/sbert_nli.bin b/data/test/regression/sbert_nli.bin index a5f680bb..68efb778 100644 --- a/data/test/regression/sbert_nli.bin +++ b/data/test/regression/sbert_nli.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:44e3925c15d86d8596baeb6bd1d153d86f57b7489798b2cf988a1248e110fd62 -size 62231 +oid sha256:6af5024a26337a440c7ea2935fce84af558dd982ee97a2f027bb922cc874292b +size 61741 diff --git a/data/test/regression/sbert_sen_sim.bin b/data/test/regression/sbert_sen_sim.bin index a59cbe0b..362f762c 100644 --- a/data/test/regression/sbert_sen_sim.bin +++ b/data/test/regression/sbert_sen_sim.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1ff17a0272752de4c88d4254b2e881f97f8ef022f03609d03ee1de0ae964368a -size 62235 +oid sha256:bbce084781342ca7274c2e4d02ed5c5de43ba213a3b76328d5994404d6544c41 +size 61745 diff --git a/modelscope/exporters/nlp/sbert_for_sequence_classification_exporter.py b/modelscope/exporters/nlp/sbert_for_sequence_classification_exporter.py index dc1e2b92..52dab4bc 100644 --- a/modelscope/exporters/nlp/sbert_for_sequence_classification_exporter.py +++ b/modelscope/exporters/nlp/sbert_for_sequence_classification_exporter.py @@ -23,12 +23,14 @@ class SbertForSequenceClassificationExporter(TorchModelExporter): def generate_dummy_inputs(self, shape: Tuple = None, + pair: bool = False, **kwargs) -> Dict[str, Any]: """Generate dummy inputs for model exportation to onnx or other formats by tracing. @param shape: A tuple of input shape which should have at most two dimensions. shape = (1, ) batch_size=1, sequence_length will be taken from the preprocessor. shape = (8, 128) batch_size=1, sequence_length=128, which will cover the config of the preprocessor. + @param pair: Generate sentence pairs or single sentences for dummy inputs. @return: Dummy inputs. """ @@ -55,7 +57,7 @@ class SbertForSequenceClassificationExporter(TorchModelExporter): **sequence_length }) preprocessor: Preprocessor = build_preprocessor(cfg, field_name) - if preprocessor.pair: + if pair: first_sequence = preprocessor.tokenizer.unk_token second_sequence = preprocessor.tokenizer.unk_token else: diff --git a/modelscope/metrics/builder.py b/modelscope/metrics/builder.py index ee4d2840..1c8e16d7 100644 --- a/modelscope/metrics/builder.py +++ b/modelscope/metrics/builder.py @@ -32,6 +32,7 @@ task_default_metrics = { Tasks.sentiment_classification: [Metrics.seq_cls_metric], Tasks.token_classification: [Metrics.token_cls_metric], Tasks.text_generation: [Metrics.text_gen_metric], + Tasks.text_classification: [Metrics.seq_cls_metric], Tasks.image_denoising: [Metrics.image_denoise_metric], Tasks.image_color_enhancement: [Metrics.image_color_enhance_metric], Tasks.image_portrait_enhancement: diff --git a/modelscope/models/nlp/bert/modeling_bert.py b/modelscope/models/nlp/bert/modeling_bert.py index e91a6433..7c1dfcf5 100755 --- a/modelscope/models/nlp/bert/modeling_bert.py +++ b/modelscope/models/nlp/bert/modeling_bert.py @@ -15,7 +15,6 @@ """PyTorch BERT model. """ import math -import os import warnings from dataclasses import dataclass from typing import Optional, Tuple @@ -41,7 +40,6 @@ from transformers.modeling_utils import (PreTrainedModel, find_pruneable_heads_and_indices, prune_linear_layer) -from modelscope.models.base import TorchModel from modelscope.utils.logger import get_logger from .configuration_bert import BertConfig @@ -50,81 +48,6 @@ logger = get_logger(__name__) _CONFIG_FOR_DOC = 'BertConfig' -def load_tf_weights_in_bert(model, config, tf_checkpoint_path): - """Load tf checkpoints in a pytorch model.""" - try: - import re - - import numpy as np - import tensorflow as tf - except ImportError: - logger.error( - 'Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see ' - 'https://www.tensorflow.org/install/ for installation instructions.' - ) - raise - tf_path = os.path.abspath(tf_checkpoint_path) - logger.info(f'Converting TensorFlow checkpoint from {tf_path}') - # Load weights from TF model - init_vars = tf.train.list_variables(tf_path) - names = [] - arrays = [] - for name, shape in init_vars: - logger.info(f'Loading TF weight {name} with shape {shape}') - array = tf.train.load_variable(tf_path, name) - names.append(name) - arrays.append(array) - - for name, array in zip(names, arrays): - name = name.split('/') - # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v - # which are not required for using pretrained model - if any(n in [ - 'adam_v', 'adam_m', 'AdamWeightDecayOptimizer', - 'AdamWeightDecayOptimizer_1', 'global_step' - ] for n in name): - logger.info(f"Skipping {'/'.join(name)}") - continue - pointer = model - for m_name in name: - if re.fullmatch(r'[A-Za-z]+_\d+', m_name): - scope_names = re.split(r'_(\d+)', m_name) - else: - scope_names = [m_name] - if scope_names[0] == 'kernel' or scope_names[0] == 'gamma': - pointer = getattr(pointer, 'weight') - elif scope_names[0] == 'output_bias' or scope_names[0] == 'beta': - pointer = getattr(pointer, 'bias') - elif scope_names[0] == 'output_weights': - pointer = getattr(pointer, 'weight') - elif scope_names[0] == 'squad': - pointer = getattr(pointer, 'classifier') - else: - try: - pointer = getattr(pointer, scope_names[0]) - except AttributeError: - logger.info(f"Skipping {'/'.join(name)}") - continue - if len(scope_names) >= 2: - num = int(scope_names[1]) - pointer = pointer[num] - if m_name[-11:] == '_embeddings': - pointer = getattr(pointer, 'weight') - elif m_name == 'kernel': - array = np.transpose(array) - try: - if pointer.shape != array.shape: - raise ValueError( - f'Pointer shape {pointer.shape} and array shape {array.shape} mismatched' - ) - except AssertionError as e: - e.args += (pointer.shape, array.shape) - raise - logger.info(f'Initialize PyTorch weight {name}') - pointer.data = torch.from_numpy(array) - return model - - class BertEmbeddings(nn.Module): """Construct the embeddings from word, position and token_type embeddings.""" @@ -750,7 +673,6 @@ class BertPreTrainedModel(PreTrainedModel): """ config_class = BertConfig - load_tf_weights = load_tf_weights_in_bert base_model_prefix = 'bert' supports_gradient_checkpointing = True _keys_to_ignore_on_load_missing = [r'position_ids'] diff --git a/modelscope/preprocessors/nlp/nlp_base.py b/modelscope/preprocessors/nlp/nlp_base.py index 267dbb8c..bc96f569 100644 --- a/modelscope/preprocessors/nlp/nlp_base.py +++ b/modelscope/preprocessors/nlp/nlp_base.py @@ -2,7 +2,7 @@ import os.path as osp import re -from typing import Any, Dict, Iterable, Optional, Tuple, Union +from typing import Any, Dict, Iterable, List, Optional, Tuple, Union import numpy as np import sentencepiece as spm @@ -217,7 +217,7 @@ class NLPTokenizerPreprocessorBase(Preprocessor): return isinstance(label, str) or isinstance(label, int) if labels is not None: - if isinstance(labels, Iterable) and all([label_can_be_mapped(label) for label in labels]) \ + if isinstance(labels, (tuple, list)) and all([label_can_be_mapped(label) for label in labels]) \ and self.label2id is not None: output[OutputKeys.LABELS] = [ self.label2id[str(label)] for label in labels @@ -314,8 +314,7 @@ class SequenceClassificationPreprocessor(NLPTokenizerPreprocessorBase): def __init__(self, model_dir: str, mode=ModeKeys.INFERENCE, **kwargs): kwargs['truncation'] = kwargs.get('truncation', True) - kwargs['padding'] = kwargs.get( - 'padding', False if mode == ModeKeys.INFERENCE else 'max_length') + kwargs['padding'] = kwargs.get('padding', 'max_length') kwargs['max_length'] = kwargs.pop('sequence_length', 128) super().__init__(model_dir, mode=mode, **kwargs) diff --git a/modelscope/trainers/hooks/lr_scheduler_hook.py b/modelscope/trainers/hooks/lr_scheduler_hook.py index ca0ec01b..32fb0250 100644 --- a/modelscope/trainers/hooks/lr_scheduler_hook.py +++ b/modelscope/trainers/hooks/lr_scheduler_hook.py @@ -47,7 +47,7 @@ class LrSchedulerHook(Hook): return lr def before_train_iter(self, trainer): - if not self.by_epoch: + if not self.by_epoch and trainer.iter > 0: if self.warmup_lr_scheduler is not None: self.warmup_lr_scheduler.step() else: diff --git a/modelscope/trainers/trainer.py b/modelscope/trainers/trainer.py index 9eaff762..61d11aa6 100644 --- a/modelscope/trainers/trainer.py +++ b/modelscope/trainers/trainer.py @@ -651,7 +651,7 @@ class EpochBasedTrainer(BaseTrainer): # TODO: support MsDataset load for cv if hasattr(data_cfg, 'name'): dataset = MsDataset.load( - dataset_name=data_cfg.name, + dataset_name=data_cfg.pop('name'), **data_cfg, ) cfg = ConfigDict(type=self.cfg.model.type, mode=mode) diff --git a/modelscope/utils/regress_test_utils.py b/modelscope/utils/regress_test_utils.py index 47bbadfe..3c1e5c1c 100644 --- a/modelscope/utils/regress_test_utils.py +++ b/modelscope/utils/regress_test_utils.py @@ -65,7 +65,8 @@ class RegressTool: def monitor_module_single_forward(self, module: nn.Module, file_name: str, - compare_fn=None): + compare_fn=None, + **kwargs): """Monitor a pytorch module in a single forward. @param module: A torch module @@ -107,7 +108,7 @@ class RegressTool: baseline = os.path.join(tempfile.gettempdir(), name) self.load(baseline, name) with open(baseline, 'rb') as f: - baseline_json = pickle.load(f) + base = pickle.load(f) class NumpyEncoder(json.JSONEncoder): """Special json encoder for numpy types @@ -122,9 +123,9 @@ class RegressTool: return obj.tolist() return json.JSONEncoder.default(self, obj) - print(f'baseline: {json.dumps(baseline_json, cls=NumpyEncoder)}') + print(f'baseline: {json.dumps(base, cls=NumpyEncoder)}') print(f'latest : {json.dumps(io_json, cls=NumpyEncoder)}') - if not compare_io_and_print(baseline_json, io_json, compare_fn): + if not compare_io_and_print(base, io_json, compare_fn, **kwargs): raise ValueError('Result not match!') @contextlib.contextmanager @@ -136,7 +137,8 @@ class RegressTool: ignore_keys=None, compare_random=True, reset_dropout=True, - lazy_stop_callback=None): + lazy_stop_callback=None, + **kwargs): """Monitor a pytorch module's backward data and cfg data within a step of the optimizer. This is usually useful when you try to change some dangerous code @@ -265,14 +267,15 @@ class RegressTool: baseline_json = pickle.load(f) if level == 'strict' and not compare_io_and_print( - baseline_json['forward'], io_json, compare_fn): + baseline_json['forward'], io_json, compare_fn, **kwargs): raise RuntimeError('Forward not match!') if not compare_backward_and_print( baseline_json['backward'], bw_json, compare_fn=compare_fn, ignore_keys=ignore_keys, - level=level): + level=level, + **kwargs): raise RuntimeError('Backward not match!') cfg_opt1 = { 'optimizer': baseline_json['optimizer'], @@ -286,7 +289,8 @@ class RegressTool: 'cfg': summary['cfg'], 'state': None if not compare_random else summary['state'] } - if not compare_cfg_and_optimizers(cfg_opt1, cfg_opt2, compare_fn): + if not compare_cfg_and_optimizers(cfg_opt1, cfg_opt2, compare_fn, + **kwargs): raise RuntimeError('Cfg or optimizers not match!') @@ -303,7 +307,8 @@ class MsRegressTool(RegressTool): compare_fn=None, ignore_keys=None, compare_random=True, - lazy_stop_callback=None): + lazy_stop_callback=None, + **kwargs): if lazy_stop_callback is None: @@ -319,7 +324,7 @@ class MsRegressTool(RegressTool): trainer.register_hook(EarlyStopHook()) - def _train_loop(trainer, *args, **kwargs): + def _train_loop(trainer, *args_train, **kwargs_train): with self.monitor_module_train( trainer, file_name, @@ -327,9 +332,11 @@ class MsRegressTool(RegressTool): compare_fn=compare_fn, ignore_keys=ignore_keys, compare_random=compare_random, - lazy_stop_callback=lazy_stop_callback): + lazy_stop_callback=lazy_stop_callback, + **kwargs): try: - return trainer.train_loop_origin(*args, **kwargs) + return trainer.train_loop_origin(*args_train, + **kwargs_train) except MsRegressTool.EarlyStopError: pass @@ -530,7 +537,8 @@ def compare_arguments_nested(print_content, ) return False if not all([ - compare_arguments_nested(None, sub_arg1, sub_arg2) + compare_arguments_nested( + None, sub_arg1, sub_arg2, rtol=rtol, atol=atol) for sub_arg1, sub_arg2 in zip(arg1, arg2) ]): if print_content is not None: @@ -551,7 +559,8 @@ def compare_arguments_nested(print_content, print(f'{print_content}, key diff:{set(keys1) - set(keys2)}') return False if not all([ - compare_arguments_nested(None, arg1[key], arg2[key]) + compare_arguments_nested( + None, arg1[key], arg2[key], rtol=rtol, atol=atol) for key in keys1 ]): if print_content is not None: @@ -574,7 +583,7 @@ def compare_arguments_nested(print_content, raise ValueError(f'type not supported: {type1}') -def compare_io_and_print(baseline_json, io_json, compare_fn=None): +def compare_io_and_print(baseline_json, io_json, compare_fn=None, **kwargs): if compare_fn is None: def compare_fn(*args, **kwargs): @@ -602,10 +611,10 @@ def compare_io_and_print(baseline_json, io_json, compare_fn=None): else: match = compare_arguments_nested( f'unmatched module {key} input args', v1input['args'], - v2input['args']) and match + v2input['args'], **kwargs) and match match = compare_arguments_nested( f'unmatched module {key} input kwargs', v1input['kwargs'], - v2input['kwargs']) and match + v2input['kwargs'], **kwargs) and match v1output = numpify_tensor_nested(v1['output']) v2output = numpify_tensor_nested(v2['output']) res = compare_fn(v1output, v2output, key, 'output') @@ -615,8 +624,11 @@ def compare_io_and_print(baseline_json, io_json, compare_fn=None): ) match = match and res else: - match = compare_arguments_nested(f'unmatched module {key} outputs', - v1output, v2output) and match + match = compare_arguments_nested( + f'unmatched module {key} outputs', + arg1=v1output, + arg2=v2output, + **kwargs) and match return match @@ -624,7 +636,8 @@ def compare_backward_and_print(baseline_json, bw_json, level, ignore_keys=None, - compare_fn=None): + compare_fn=None, + **kwargs): if compare_fn is None: def compare_fn(*args, **kwargs): @@ -653,18 +666,26 @@ def compare_backward_and_print(baseline_json, data2, grad2, data_after2 = bw_json[key]['data'], bw_json[key][ 'grad'], bw_json[key]['data_after'] match = compare_arguments_nested( - f'unmatched module {key} tensor data', data1, data2) and match + f'unmatched module {key} tensor data', + arg1=data1, + arg2=data2, + **kwargs) and match if level == 'strict': match = compare_arguments_nested( - f'unmatched module {key} grad data', grad1, - grad2) and match + f'unmatched module {key} grad data', + arg1=grad1, + arg2=grad2, + **kwargs) and match match = compare_arguments_nested( f'unmatched module {key} data after step', data_after1, - data_after2) and match + data_after2, **kwargs) and match return match -def compare_cfg_and_optimizers(baseline_json, cfg_json, compare_fn=None): +def compare_cfg_and_optimizers(baseline_json, + cfg_json, + compare_fn=None, + **kwargs): if compare_fn is None: def compare_fn(*args, **kwargs): @@ -686,12 +707,12 @@ def compare_cfg_and_optimizers(baseline_json, cfg_json, compare_fn=None): print( f"Optimizer type not equal:{optimizer1['type']} and {optimizer2['type']}" ) - match = compare_arguments_nested('unmatched optimizer defaults', - optimizer1['defaults'], - optimizer2['defaults']) and match - match = compare_arguments_nested('unmatched optimizer state_dict', - optimizer1['state_dict'], - optimizer2['state_dict']) and match + match = compare_arguments_nested( + 'unmatched optimizer defaults', optimizer1['defaults'], + optimizer2['defaults'], **kwargs) and match + match = compare_arguments_nested( + 'unmatched optimizer state_dict', optimizer1['state_dict'], + optimizer2['state_dict'], **kwargs) and match res = compare_fn(lr_scheduler1, lr_scheduler2, None, 'lr_scheduler') if res is not None: @@ -703,16 +724,17 @@ def compare_cfg_and_optimizers(baseline_json, cfg_json, compare_fn=None): print( f"Optimizer type not equal:{lr_scheduler1['type']} and {lr_scheduler2['type']}" ) - match = compare_arguments_nested('unmatched lr_scheduler state_dict', - lr_scheduler1['state_dict'], - lr_scheduler2['state_dict']) and match + match = compare_arguments_nested( + 'unmatched lr_scheduler state_dict', lr_scheduler1['state_dict'], + lr_scheduler2['state_dict'], **kwargs) and match res = compare_fn(cfg1, cfg2, None, 'cfg') if res is not None: print(f'cfg compared with user compare_fn with result:{res}\n') match = match and res else: - match = compare_arguments_nested('unmatched cfg', cfg1, cfg2) and match + match = compare_arguments_nested( + 'unmatched cfg', arg1=cfg1, arg2=cfg2, **kwargs) and match res = compare_fn(state1, state2, None, 'state') if res is not None: @@ -721,6 +743,6 @@ def compare_cfg_and_optimizers(baseline_json, cfg_json, compare_fn=None): match = match and res else: match = compare_arguments_nested('unmatched random state', state1, - state2) and match + state2, **kwargs) and match return match diff --git a/tests/msdatasets/test_ms_dataset.py b/tests/msdatasets/test_ms_dataset.py index 91a3b5c5..1e537e93 100644 --- a/tests/msdatasets/test_ms_dataset.py +++ b/tests/msdatasets/test_ms_dataset.py @@ -52,7 +52,8 @@ class MsDatasetTest(unittest.TestCase): @unittest.skipUnless(test_level() >= 1, 'skip test in current test level') def test_ms_csv_basic(self): ms_ds_train = MsDataset.load( - 'afqmc_small', namespace='userxiaoming', split='train') + 'clue', subset_name='afqmc', + split='train').to_hf_dataset().select(range(5)) print(next(iter(ms_ds_train))) @unittest.skipUnless(test_level() >= 1, 'skip test in current test level') diff --git a/tests/trainers/test_finetune_sequence_classification.py b/tests/trainers/test_finetune_sequence_classification.py index f2adfa22..27db1f18 100644 --- a/tests/trainers/test_finetune_sequence_classification.py +++ b/tests/trainers/test_finetune_sequence_classification.py @@ -16,7 +16,8 @@ from modelscope.trainers.optimizer.child_tuning_adamw_optimizer import \ calculate_fisher from modelscope.utils.constant import ModelFile, Tasks from modelscope.utils.data_utils import to_device -from modelscope.utils.regress_test_utils import MsRegressTool +from modelscope.utils.regress_test_utils import (MsRegressTool, + compare_arguments_nested) from modelscope.utils.test_utils import test_level @@ -41,6 +42,33 @@ class TestFinetuneSequenceClassification(unittest.TestCase): def test_trainer_repeatable(self): import torch # noqa + def compare_fn(value1, value2, key, type): + # Ignore the differences between optimizers of two torch versions + if type != 'optimizer': + return None + + match = (value1['type'] == value2['type']) + shared_defaults = set(value1['defaults'].keys()).intersection( + set(value2['defaults'].keys())) + match = all([ + compare_arguments_nested(f'Optimizer defaults {key} not match', + value1['defaults'][key], + value2['defaults'][key]) + for key in shared_defaults + ]) and match + match = (len(value1['state_dict']['param_groups']) == len( + value2['state_dict']['param_groups'])) and match + for group1, group2 in zip(value1['state_dict']['param_groups'], + value2['state_dict']['param_groups']): + shared_keys = set(group1.keys()).intersection( + set(group2.keys())) + match = all([ + compare_arguments_nested( + f'Optimizer param_groups {key} not match', group1[key], + group2[key]) for key in shared_keys + ]) and match + return match + def cfg_modify_fn(cfg): cfg.task = 'nli' cfg['preprocessor'] = {'type': 'nli-tokenizer'} @@ -98,7 +126,8 @@ class TestFinetuneSequenceClassification(unittest.TestCase): name=Trainers.nlp_base_trainer, default_args=kwargs) with self.regress_tool.monitor_ms_train( - trainer, 'sbert-base-tnews', level='strict'): + trainer, 'sbert-base-tnews', level='strict', + compare_fn=compare_fn): trainer.train() def finetune(self, diff --git a/tests/trainers/test_trainer_with_nlp.py b/tests/trainers/test_trainer_with_nlp.py index 6030ada9..8357e778 100644 --- a/tests/trainers/test_trainer_with_nlp.py +++ b/tests/trainers/test_trainer_with_nlp.py @@ -29,7 +29,8 @@ class TestTrainerWithNlp(unittest.TestCase): os.makedirs(self.tmp_dir) self.dataset = MsDataset.load( - 'afqmc_small', namespace='userxiaoming', split='train') + 'clue', subset_name='afqmc', + split='train').to_hf_dataset().select(range(2)) def tearDown(self): shutil.rmtree(self.tmp_dir) @@ -73,7 +74,7 @@ class TestTrainerWithNlp(unittest.TestCase): output_dir = os.path.join(self.tmp_dir, ModelFile.TRAIN_OUTPUT_DIR) pipeline_sentence_similarity(output_dir) - @unittest.skipUnless(test_level() >= 1, 'skip test in current test level') + @unittest.skipUnless(test_level() >= 3, 'skip test in current test level') def test_trainer_with_backbone_head(self): model_id = 'damo/nlp_structbert_sentiment-classification_chinese-base' kwargs = dict( @@ -99,6 +100,8 @@ class TestTrainerWithNlp(unittest.TestCase): model_id = 'damo/nlp_structbert_sentiment-classification_chinese-base' cfg = read_config(model_id, revision='beta') cfg.train.max_epochs = 20 + cfg.preprocessor.train['label2id'] = {'0': 0, '1': 1} + cfg.preprocessor.val['label2id'] = {'0': 0, '1': 1} cfg.train.work_dir = self.tmp_dir cfg_file = os.path.join(self.tmp_dir, 'config.json') cfg.dump(cfg_file) @@ -120,22 +123,24 @@ class TestTrainerWithNlp(unittest.TestCase): checkpoint_path=os.path.join(self.tmp_dir, 'epoch_10.pth')) self.assertTrue(Metrics.accuracy in eval_results) - @unittest.skipUnless(test_level() >= 1, 'skip test in current test level') + @unittest.skipUnless(test_level() >= 2, 'skip test in current test level') def test_trainer_with_configured_datasets(self): model_id = 'damo/nlp_structbert_sentence-similarity_chinese-base' cfg: Config = read_config(model_id) cfg.train.max_epochs = 20 + cfg.preprocessor.train['label2id'] = {'0': 0, '1': 1} + cfg.preprocessor.val['label2id'] = {'0': 0, '1': 1} cfg.train.work_dir = self.tmp_dir cfg.dataset = { 'train': { - 'name': 'afqmc_small', + 'name': 'clue', + 'subset_name': 'afqmc', 'split': 'train', - 'namespace': 'userxiaoming' }, 'val': { - 'name': 'afqmc_small', + 'name': 'clue', + 'subset_name': 'afqmc', 'split': 'train', - 'namespace': 'userxiaoming' }, } cfg_file = os.path.join(self.tmp_dir, 'config.json') @@ -159,6 +164,11 @@ class TestTrainerWithNlp(unittest.TestCase): model_id = 'damo/nlp_structbert_sentence-similarity_chinese-base' cfg: Config = read_config(model_id) cfg.train.max_epochs = 3 + cfg.preprocessor.first_sequence = 'sentence1' + cfg.preprocessor.second_sequence = 'sentence2' + cfg.preprocessor.label = 'label' + cfg.preprocessor.train['label2id'] = {'0': 0, '1': 1} + cfg.preprocessor.val['label2id'] = {'0': 0, '1': 1} cfg.train.work_dir = self.tmp_dir cfg_file = os.path.join(self.tmp_dir, 'config.json') cfg.dump(cfg_file) From 8ec90ccbf8926235ee3176d4fd6b7f10dbdf09ad Mon Sep 17 00:00:00 2001 From: "xiangpeng.wxp" Date: Thu, 20 Oct 2022 17:35:27 +0800 Subject: [PATCH 10/10] [to #42322933] Add uttest for en2fr and fr2en tasks Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/10467797 * add uttest for en2fr and fr2en tasks --- tests/pipelines/test_csanmt_translation.py | 14 ++++++++++++++ 1 file changed, 14 insertions(+) diff --git a/tests/pipelines/test_csanmt_translation.py b/tests/pipelines/test_csanmt_translation.py index f7ec81cd..83827813 100644 --- a/tests/pipelines/test_csanmt_translation.py +++ b/tests/pipelines/test_csanmt_translation.py @@ -26,6 +26,20 @@ class TranslationTest(unittest.TestCase, DemoCompatibilityCheck): pipeline_ins = pipeline(self.task, model=model_id) print(pipeline_ins(input=inputs)) + @unittest.skipUnless(test_level() >= 0, 'skip test in current test level') + def test_run_with_model_name_for_en2fr(self): + model_id = 'damo/nlp_csanmt_translation_en2fr' + inputs = 'When I was in my 20s, I saw my very first psychotherapy client.' + pipeline_ins = pipeline(self.task, model=model_id) + print(pipeline_ins(input=inputs)) + + @unittest.skipUnless(test_level() >= 0, 'skip test in current test level') + def test_run_with_model_name_for_fr2en(self): + model_id = 'damo/nlp_csanmt_translation_fr2en' + inputs = "Quand j'avais la vingtaine, j'ai vu mes tout premiers clients comme psychothérapeute." + pipeline_ins = pipeline(self.task, model=model_id) + print(pipeline_ins(input=inputs)) + @unittest.skipUnless(test_level() >= 2, 'skip test in current test level') def test_run_with_default_model(self): inputs = '声明补充说，沃伦的同事都深感震惊，并且希望他能够投案自首。'