def test_retinanet_eval(self):
checkpoint = os.path.join(self.MMDET_PATH, "work_dirs",
"retinanet_r50_fpn_1x_int8", "latest.pth")
comm_line = "tools/test.py configs/retinanet/retinanet_r50_fpn_1x_int8.py {} --eval bbox".format(
checkpoint)
runner = Command(
create_command_line(comm_line, self.activate_venv,
self.mmdet_python), self.MMDET_PATH)
runner.run()
def test_ssd300_eval(self):
checkpoint = os.path.join(self.MMDET_PATH, "work_dirs",
"ssd300_voc_int8", "latest.pth")
comm_line = "tools/test.py configs/pascal_voc/ssd300_voc_int8.py {} --eval mAP".format(
checkpoint)
runner = Command(
create_command_line(comm_line, self.activate_venv,
self.mmdet_python), self.MMDET_PATH)
runner.run()
def test_xnli_eval(self, temp_folder):
com_line = "examples/text-classification/run_xnli.py --model_name_or_path {output}" \
" --language zh --do_eval --data_dir {} --learning_rate 5e-5 --max_seq_length 128 --output_dir" \
" {output} --nncf_config nncf_bert_config_xnli.json --per_gpu_eval_batch_size 24" \
.format(DATASET_PATH, output=os.path.join(temp_folder["models"], "xnli"))
runner = Command(
create_command_line(com_line, self.activate_venv,
self.trans_python, self.cuda_visible_string),
self.TRANS_PATH)
runner.run()
def test_convert_to_onnx(self, temp_folder):
com_line = "examples/question-answering/run_squad.py --model_type bert --model_name_or_path {output}" \
" --output_dir {output}" \
" --to_onnx {output}/model.onnx".format(output=os.path.join(temp_folder["models"], "squad"))
runner = Command(
create_command_line(com_line, self.activate_venv,
self.trans_python, self.cuda_visible_string),
self.TRANS_PATH)
runner.run()
assert os.path.exists(
os.path.join(temp_folder["models"], "squad", "model.onnx"))
def test_glue_distilbert_eval(self, temp_folder):
com_line = "examples/text-classification/run_glue.py --model_name_or_path {output}" \
" --task_name SST-2 --do_eval --max_seq_length 128" \
" --output_dir {output} --data_dir {}/glue/glue_data/SST-2" \
" --nncf_config nncf_distilbert_config_sst2.json" \
.format(DATASET_PATH, output=os.path.join(temp_folder["models"], "distilbert_output"))
runner = Command(
create_command_line(com_line, self.activate_venv,
self.trans_python, self.cuda_visible_string),
self.TRANS_PATH)
runner.run()
def test_glue_eval(self, temp_folder):
com_line = "examples/text-classification/run_glue.py --model_name_or_path {output}" \
" --task_name mnli --do_eval --data_dir {}/glue/glue_data/MNLI --learning_rate 2e-5" \
" --num_train_epochs 1.0 --max_seq_length 128 --output_dir {output}" \
" --nncf_config nncf_roberta_config_mnli.json" \
.format(DATASET_PATH, output=os.path.join(temp_folder["models"], "roberta_mnli"))
runner = Command(
create_command_line(com_line, self.activate_venv,
self.trans_python, self.cuda_visible_string),
self.TRANS_PATH)
runner.run()
def test_squad_eval(self, temp_folder):
com_line = "examples/question-answering/run_squad.py --model_type bert --model_name_or_path {output}" \
" --do_eval --do_lower_case --predict_file {}/squad/dev-v1.1.json --learning_rate 3e-5" \
" --max_seq_length 384 --doc_stride 128 --per_gpu_eval_batch_size=4 --output_dir {output} " \
"--nncf_config nncf_bert_config_squad.json" \
.format(DATASET_PATH, output=os.path.join(temp_folder["models"], "squad"))
runner = Command(
create_command_line(com_line, self.activate_venv,
self.trans_python, self.cuda_visible_string),
self.TRANS_PATH)
runner.run()
def test_retinanet_export2onnx(self):
checkpoint = os.path.join(self.MMDET_PATH, "work_dirs",
"retinanet_r50_fpn_1x_int8", "latest.pth")
comm_line = "tools/pytorch2onnx.py configs/retinanet/retinanet_r50_fpn_1x_int8.py {} --output-file retinanet_r50_fpn_1x_int8.onnx".format(
checkpoint)
runner = Command(
create_command_line(comm_line, self.activate_venv,
self.mmdet_python), self.MMDET_PATH)
runner.run()
assert os.path.exists(
os.path.join(self.MMDET_PATH, "retinanet_r50_fpn_1x_int8.onnx"))
def test_lm_eval(self, temp_folder):
com_line = "examples/language-modeling/run_language_modeling.py --model_type gpt2 " \
"--model_name_or_path {output} --do_eval " \
" --output_dir {output} --eval_data_file {}/wikitext-2-raw/wiki.train.raw" \
" --nncf_config nncf_gpt2_config_wikitext_hw_config.json" \
.format(DATASET_PATH, output=os.path.join(temp_folder["models"], "lm_output"))
runner = Command(
create_command_line(com_line, self.activate_venv,
self.trans_python, self.cuda_visible_string),
self.TRANS_PATH)
runner.run()
def test_xnli_train(self, temp_folder):
com_line = "examples/text-classification/run_xnli.py --model_name_or_path bert-base-chinese" \
" --language zh --train_language zh --do_train --data_dir {} --per_gpu_train_batch_size 24" \
" --learning_rate 5e-5 --num_train_epochs 1.0 --max_seq_length 128 --output_dir {}" \
" --save_steps 200 --nncf_config nncf_bert_config_xnli.json" \
.format(DATASET_PATH, os.path.join(temp_folder["models"], "xnli"))
runner = Command(
create_command_line(com_line, self.VENV_TRANS_PATH,
self.trans_python, self.cuda_visible_string),
self.TRANS_PATH)
runner.run()
assert os.path.exists(
os.path.join(temp_folder["models"], "xnli", "pytorch_model.bin"))
def test_squad_train(self, temp_folder):
com_line = "examples/question-answering/run_squad.py --model_type bert --model_name_or_path " \
"bert-large-uncased-whole-word-masking-finetuned-squad --do_train --do_lower_case " \
"--train_file {}/squad/train-v1.1.json" \
" --learning_rate 3e-5 --num_train_epochs 1 --max_seq_length 384 --doc_stride 128 --output_dir " \
"{} --per_gpu_train_batch_size=1 --save_steps=200 --nncf_config" \
" nncf_bert_config_squad.json".format(DATASET_PATH, os.path.join(temp_folder["models"], "squad"))
runner = Command(
create_command_line(com_line, self.activate_venv,
self.trans_python, self.cuda_visible_string),
self.TRANS_PATH)
runner.run()
assert os.path.exists(
os.path.join(temp_folder["models"], "squad", "pytorch_model.bin"))
def test_maskrcnn_train(self):
subprocess.run(
"wget http://download.openmmlab.com/mmdetection/v2.0/mask_rcnn/mask_rcnn_r50_caffe_fpn_mstrain-poly_3x_coco/mask_rcnn_r50_caffe_fpn_mstrain-poly_3x_coco_bbox_mAP-0.408__segm_mAP-0.37_20200504_163245-42aa3d00.pth",
check=True,
shell=True,
cwd=self.MMDET_PATH)
comm_line = "tools/train.py configs/mask_rcnn/mask_rcnn_r50_caffe_fpn_1x_coco_int8.py"
runner = Command(
create_command_line(comm_line, self.activate_venv,
self.mmdet_python), self.MMDET_PATH)
runner.run()
assert os.path.exists(
os.path.join(self.MMDET_PATH, "work_dirs",
"mask_rcnn_r50_caffe_fpn_1x_coco_int8", "latest.pth"))
def test_retinanet_train(self):
subprocess.run(
"wget https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmdetection/models/retinanet_r50_fpn_2x_20190616-75574209.pth",
check=True,
shell=True,
cwd=self.MMDET_PATH)
comm_line = "tools/train.py configs/retinanet/retinanet_r50_fpn_1x_int8.py"
runner = Command(
create_command_line(comm_line, self.activate_venv,
self.mmdet_python), self.MMDET_PATH)
runner.run()
assert os.path.exists(
os.path.join(self.MMDET_PATH, "work_dirs",
"retinanet_r50_fpn_1x_int8", "latest.pth"))
def test_ssd300_train(self):
subprocess.run(
"wget https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/ssd300_voc_vgg16_caffe_240e_20190501-7160d09a.pth",
check=True,
shell=True,
cwd=self.MMDET_PATH)
comm_line = "tools/train.py configs/pascal_voc/ssd300_voc_int8.py"
runner = Command(
create_command_line(comm_line, self.activate_venv,
self.mmdet_python), self.MMDET_PATH)
runner.run()
assert os.path.exists(
os.path.join(self.MMDET_PATH, "work_dirs", "ssd300_voc_int8",
"latest.pth"))
def test_glue_train(self, temp_folder):
com_line = "examples/text-classification/run_glue.py --model_name_or_path" \
" roberta-large-mnli --task_name mnli --do_train --data_dir {}/glue/glue_data/MNLI" \
" --per_gpu_train_batch_size 4 --learning_rate 2e-5 --num_train_epochs 1.0 --max_seq_length 128 " \
"--output_dir {} --save_steps 200 --nncf_config" \
" nncf_roberta_config_mnli.json" \
.format(DATASET_PATH, os.path.join(temp_folder["models"], "roberta_mnli"))
runner = Command(
create_command_line(com_line, self.activate_venv,
self.trans_python, self.cuda_visible_string),
self.TRANS_PATH)
runner.run()
assert os.path.exists(
os.path.join(temp_folder["models"], "roberta_mnli",
"pytorch_model.bin"))
def test_lm_train(self, temp_folder):
com_line = "examples/language-modeling/run_language_modeling.py --model_type gpt2 --model_name_or_path gpt2" \
" --do_train --per_gpu_train_batch_size 8" \
" --train_data_file {}/wikitext-2-raw/wiki.train.raw " \
" --output_dir {} --nncf_config" \
" nncf_gpt2_config_wikitext_hw_config.json".format(DATASET_PATH, os.path.join(temp_folder["models"],
"lm_output"))
runner = Command(
create_command_line(com_line, self.activate_venv,
self.trans_python, self.cuda_visible_string),
self.TRANS_PATH)
runner.run()
assert os.path.exists(
os.path.join(temp_folder["models"], "lm_output",
"pytorch_model.bin"))
def test_loaded_model_evals_according_to_saved_acc(_params, tmp_path):
p = _params
config_path = p['sample_config_path']
checkpoint_path = p['checkpoint_path']
tmp_path = str(tmp_path)
args = {}
args['data'] = tmp_path + '/' + p['dataset']
args['dataset'] = p['dataset']
args['config'] = str(config_path)
args['mode'] = 'test'
args['log-dir'] = tmp_path
args['workers'] = 4
args['seed'] = 1
args['resume'] = checkpoint_path
if p['execution_mode'] == ExecutionMode.MULTIPROCESSING_DISTRIBUTED:
args['multiprocessing-distributed'] = ''
else:
pytest.skip("DataParallel eval takes too long for this test to be run during pre-commit")
runner = Command(create_command_line(get_cli_dict_args(args), "classification"))
res = runner.run()
assert res == 0
acc1 = parse_best_acc1(tmp_path)
assert torch.load(checkpoint_path)['best_acc1'] == pytest.approx(acc1)
def test_compression_eval_trained(_params, tmp_path):
p = _params
args = p['args']
tc = p['test_config']
args['mode'] = 'test'
args['log-dir'] = tmp_path
args['workers'] = 4
args['seed'] = 1
checkpoint_path = os.path.join(args['checkpoint-save-dir'],
tc['checkpoint_name'] + '_best.pth')
args['resume'] = checkpoint_path
if 'weights' in args:
del args['weights']
reset_context('orig')
reset_context('quantized_graphs')
runner = Command(
create_command_line(get_cli_dict_args(args), tc['sample_type']))
res = runner.run(timeout=tc['timeout'])
assert res == 0
acc1 = parse_best_acc1(tmp_path)
assert torch.load(checkpoint_path)['best_acc1'] == approx(
acc1, abs=tc['absolute_tolerance_eval'])
def test_compression_train(_params, tmp_path):
p = _params
args = p['args']
tc = p['test_config']
args['mode'] = 'train'
args['log-dir'] = tmp_path
args['workers'] = 4
args['seed'] = 1
reset_context('orig')
reset_context('quantized_graphs')
runner = Command(
create_command_line(get_cli_dict_args(args), tc['sample_type']))
res = runner.run(timeout=tc['timeout'])
assert res == 0
checkpoint_path = os.path.join(args['checkpoint-save-dir'],
tc['checkpoint_name'] + '_best.pth')
assert os.path.exists(checkpoint_path)
actual_acc = torch.load(checkpoint_path)['best_acc1']
ref_acc = tc['expected_accuracy']
better_accuracy_tolerance = 3
tolerance = tc[
'absolute_tolerance_train'] if actual_acc < ref_acc else better_accuracy_tolerance
assert actual_acc == approx(ref_acc, abs=tolerance)
def test_compression_eval_trained(_params, tmp_path):
p = _params
args = p['args']
tc = p['test_config']
args['mode'] = 'test'
args['log-dir'] = tmp_path
args['workers'] = 4
args['seed'] = 1
checkpoint_path = os.path.join(args['checkpoint-save-dir'],
tc['checkpoint_name'] + '_best.pth')
args['resume'] = checkpoint_path
if 'weights' in args:
del args['weights']
reset_context('orig')
reset_context('quantized_graphs')
runner = Command(
create_command_line(get_cli_dict_args(args), tc['sample_type']))
res = runner.run(timeout=tc['timeout'])
assert res == 0
output_path = None
for root, _, names in os.walk(str(tmp_path)):
for name in names:
if 'output' in name:
output_path = os.path.join(root, name)
assert os.path.exists(output_path)
with open(output_path, "r") as f:
last_line = f.readlines()[-1]
acc1 = float(re.findall("\\d+\\.\\d+", last_line)[0])
assert torch.load(checkpoint_path)['best_acc1'] == approx(
acc1, abs=tc['absolute_tolerance_eval'])
def test_loaded_model_evals_according_to_saved_acc(_params, tmp_path, dataset_dir):
p = _params
config_path = p['sample_config_path']
checkpoint_path = p['checkpoint_path']
metrics_path = str(tmp_path.joinpath('metrics.json'))
tmp_path = str(tmp_path)
args = {}
if not dataset_dir:
dataset_dir = tmp_path
args['data'] = dataset_dir
args['dataset'] = p['dataset']
args['config'] = str(config_path)
args['mode'] = 'test'
args['log-dir'] = tmp_path
args['workers'] = 0 # Workaroundr the PyTorch MultiProcessingDataLoader issue
args['seed'] = 1
args['resume'] = checkpoint_path
args['metrics-dump'] = metrics_path
if p['execution_mode'] == ExecutionMode.MULTIPROCESSING_DISTRIBUTED:
args['multiprocessing-distributed'] = ''
else:
pytest.skip("DataParallel eval takes too long for this test to be run during pre-commit")
runner = Command(create_command_line(get_cli_dict_args(args), "classification"))
res = runner.run()
assert res == 0
with open(metrics_path) as metric_file:
metrics = json.load(metric_file)
assert torch.load(checkpoint_path)['best_acc1'] == pytest.approx(metrics['Accuracy'])
def test_glue_distilbert_train(self, temp_folder):
com_line = "examples/text-classification/run_glue.py --model_name_or_path" \
" distilbert-base-uncased" \
" --task_name SST-2 --do_train --max_seq_length 128 --per_gpu_train_batch_size 8" \
" --data_dir {}/glue/glue_data/SST-2 --learning_rate 5e-5 --num_train_epochs 3.0" \
" --output_dir {} --save_steps 200 --nncf_config" \
" nncf_distilbert_config_sst2.json".format(DATASET_PATH, os.path.join(temp_folder["models"],
"distilbert_output"))
runner = Command(
create_command_line(com_line, self.activate_venv,
self.trans_python, self.cuda_visible_string),
self.TRANS_PATH)
runner.run()
assert os.path.exists(
os.path.join(temp_folder["models"], "distilbert_output",
"pytorch_model.bin"))
def test_compression_eval_trained(_params, tmp_path):
p = _params
args = p['args']
tc = p['test_config']
args['mode'] = 'test'
args['log-dir'] = tmp_path
args['workers'] = 0 # Workaround for PyTorch MultiprocessingDataLoader issues
args['seed'] = 1
checkpoint_path = os.path.join(args['checkpoint-save-dir'], tc['checkpoint_name'] + '_best.pth')
args['resume'] = checkpoint_path
if 'weights' in args:
del args['weights']
runner = Command(create_command_line(get_cli_dict_args(args), tc['sample_type']))
runner.run(timeout=tc['timeout'])
acc1 = parse_best_acc1(tmp_path)
assert torch.load(checkpoint_path)['best_acc1'] == approx(acc1, abs=tc['absolute_tolerance_eval'])
def test_maskrcnn_eval(self):
checkpoint = os.path.join(self.MMDET_PATH, "work_dirs",
"mask_rcnn_r50_caffe_fpn_1x_coco_int8",
"latest.pth")
comm_line = "tools/test.py configs/mask_rcnn/mask_rcnn_r50_caffe_fpn_1x_coco_int8.py {} --eval bbox".format(
checkpoint)
runner = Command(
create_command_line(comm_line, self.activate_venv,
self.mmdet_python), self.MMDET_PATH)
res = runner.run()
assert res == 0
def test_compression_train(_params, tmp_path):
p = _params
args = p['args']
tc = p['test_config']
args['mode'] = 'train'
args['log-dir'] = tmp_path
args['workers'] = 0 # Workaround for PyTorch MultiprocessingDataLoader issues
args['seed'] = 1
runner = Command(create_command_line(get_cli_dict_args(args), tc['sample_type']))
runner.run(timeout=tc['timeout'])
checkpoint_path = os.path.join(args['checkpoint-save-dir'], tc['checkpoint_name'] + '_best.pth')
assert os.path.exists(checkpoint_path)
actual_acc = torch.load(checkpoint_path)['best_acc1']
ref_acc = tc['expected_accuracy']
better_accuracy_tolerance = 3
tolerance = tc['absolute_tolerance_train'] if actual_acc < ref_acc else better_accuracy_tolerance
assert actual_acc == approx(ref_acc, abs=tolerance)