Nigel
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

!18868 CleanCode for cpm networks. 

Merge pull request !18868 from casgj/master_0625
This commit is contained in:
i-robot 2021-06-26 02:09:06 +00:00 committed by Gitee
parent 69cedc28f4 b3d486eef7
commit a52b732a8b
12 changed files with 16 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
model_zoo/official/nlp/cpm/README.md
View File

@ -413,7 +413,7 @@ The finetune performance and accuracy of 4 machines and 32 cards are as follows:
| Optimizer | Adam |
| Accuracy | 81.4% |
| Speed | 2740ms/step (32pcs) |
| Loss | 0.008 |
| Loss | 0.03 |
| Params (M) | 2597.1 |
| Checkpoint for inference | 57G （.ckpt file） |
| Scripts | <https://gitee.com/mindspore/mindspore/tree/master/model_zoo/official/nlp/cpm> |

2
model_zoo/official/nlp/cpm/README_CN.md
View File

@ -415,7 +415,7 @@ Zero-shot单机双卡推理性能和精度如下：
| 优化器 | Adam |
| 准确率 | 81.4% |
| 速度 | 2740毫秒/步(32卡) |
| 损失 | 0.08 |
| 损失 | 0.03 |
| 参数 (M) | 2597.1 |
| 推理检查点 | 57G （.ckpt文件） |
| 脚本 | <https://gitee.com/mindspore/mindspore/tree/master/model_zoo/official/nlp/cpm> |

3
model_zoo/official/nlp/cpm/eval.py
View File

@ -273,6 +273,8 @@ if __name__ == '__main__':
parser.add_argument("--has_train_strategy", type=ast.literal_eval, default=True,
help='Model has distributed training strategy.')
args_eval = parser.parse_args()
if args_eval.distribute:
set_parallel_env()
ckpt_file_list_test = None
if args_eval.has_train_strategy:
@ -286,7 +288,6 @@ if __name__ == '__main__':
result_accuracy = 0.0
if args_eval.distribute:
set_parallel_env()
print("Start validation on 2 devices with model parallel.")
result_accuracy = run_eval(args_eval, finetune_test_distrubute, ckpt_file_list_test)
else:

2
model_zoo/official/nlp/cpm/export.py
View File

@ -57,7 +57,7 @@ if __name__ == '__main__':
cpm_model = CPM_LAYER(config_eval)
if not args.has_train_strategy:
weights = load_checkpoint(args.ckpt_path)
weights = load_checkpoint(args.ckpt_path_doc)
can_be_loaded = {}
print("+++++++loading weights+++++")
for name, _ in weights.items():

2
model_zoo/official/nlp/cpm/scripts/run_eval_distribute_ascend.sh
View File

@ -59,7 +59,7 @@ do
export DEVICE_ID=$i
echo "start eval for rank $RANK_ID, device $DEVICE_ID"
env > env.log
python ../../eval.py --dataset $DATASET --data_path $LABEL --ckpt_path $MODEL_CKPT --ckpt_partition $CKPT_NUMBER --distribute True --has_train_strategy True> log_cpm.log 2>&1 &
python ../../eval.py --dataset $DATASET --data_path $LABEL --ckpt_path_doc $MODEL_CKPT --ckpt_partition $CKPT_NUMBER --distribute True --has_train_strategy True> log_cpm.log 2>&1 &
cd ${current_exec_path}
done
cd ${current_exec_path}

2
model_zoo/official/nlp/cpm/scripts/run_test_distribute_ascend.sh
View File

@ -69,7 +69,7 @@ do
env > env.log
python ../../test.py --dev_dataset $DEV_DATASET --dev_data_path $DEV_LABEL \
--test_dataset $TEST_DATASET --test_data_path $TEST_LABEL \
--ckpt_path $MODEL_CKPT --ckpt_partition $CKPT_NUMBER \
--ckpt_path_doc $MODEL_CKPT --ckpt_partition $CKPT_NUMBER \
--distribute True --has_train_strategy True> log_cpm.log 2>&1 &
cd ${current_exec_path}

2
model_zoo/official/nlp/cpm/scripts/run_test_standalone_ascend.sh
View File

@ -71,7 +71,7 @@ do
env > env.log
python ../../test.py --dev_dataset $DEV_DATASET --dev_data_path $DEV_LABEL \
--test_dataset $TEST_DATASET --test_data_path $TEST_LABEL \
--ckpt_path $MODEL_CKPT --ckpt_partition $CKPT_NUMBER \
--ckpt_path_doc $MODEL_CKPT --ckpt_partition $CKPT_NUMBER \
--ckpt_epoch $ckptepoch --result_path $result_path \
--distribute False --has_train_strategy True> log_cpm.log 2>&1

2
model_zoo/official/nlp/cpm/scripts/run_zero-shot_inference_distribute_ascend.sh
View File

@ -58,7 +58,7 @@ do
export DEVICE_ID=$i
echo "start eval for rank $RANK_ID, device $DEVICE_ID"
env > env.log
python ../../zero-shot.py --dataset $DATASET --truth_labels_path $LABEL --ckpt_path $MODEL_CKPT --distribute True --has_train_strategy False> log_cpm.log 2>&1 &
python ../../zero-shot.py --dataset $DATASET --truth_labels_path $LABEL --ckpt_path_doc $MODEL_CKPT --distribute True --has_train_strategy False> log_cpm.log 2>&1 &
cd ${current_exec_path}
done
cd ${current_exec_path}

2
model_zoo/official/nlp/cpm/scripts/run_zero-shot_inference_standalone_ascend.sh
View File

@ -54,6 +54,6 @@ cp -r ../scripts/*.sh ./eval
cd ./eval || exit
echo "start training for device $DEVICE_ID"
env > env.log
python ../../zero-shot.py --dataset $DATASET --truth_labels_path $LABEL --ckpt_path $MODEL_CKPT --has_train_strategy False > log_cpm.log 2>&1 &
python ../../zero-shot.py --dataset $DATASET --truth_labels_path $LABEL --ckpt_path_doc $MODEL_CKPT --has_train_strategy False > log_cpm.log 2>&1 &
cd ..

2
model_zoo/official/nlp/cpm/sort.py
View File

@ -18,7 +18,7 @@ import argparse
if __name__ == '__main__':
parser = argparse.ArgumentParser(description="Accuracy sort.")
parser.add_argument("--result_path", type=str, default="scripts/result.txt",
parser.add_argument("--result_path", type=str, default="/home/result.txt",
help='Text save address.')
args_eval = parser.parse_args()

4
model_zoo/official/nlp/cpm/train.py
View File

@ -262,8 +262,8 @@ if __name__ == '__main__':
args = parser.parse_args()
if args.multi_machine:
print("Training on multiple machines")
print("Training on multiple machines.")
train_paralle(args.dataset, args.pretrain_ckpt_path, config_train_multi_machine)
else:
print("Training on single machine and using 8 cards.")
print("Training on single machine.")
train_paralle(args.dataset, args.pretrain_ckpt_path, config_train_single_machine)

4
model_zoo/official/nlp/cpm/zero-shot.py
View File

@ -271,6 +271,9 @@ if __name__ == '__main__':
help='Whether the loaded checkpoints have distributed training strategy.')
parser.add_argument('--ckpt_partition', type=int, default=1, help="Number of checkpoint partition.")
args_parse = parser.parse_args()
if args_parse.distribute:
set_parallel_env()
ckpt_file_list_test = None
if args_parse.has_train_strategy:
# Get the checkpoint with train strategy.
@ -281,7 +284,6 @@ if __name__ == '__main__':
ckpt_file_list_test = create_ckpt_file_list(args_parse)
print("Get checkpoint file lists++++", ckpt_file_list_test, flush=True)
if args_parse.distribute:
set_parallel_env()
print("Staring evaluating on 2 devices with model parallel.")
run_eval(args_parse, config_zero_shot_distrubute, ckpt_file_list_test)
else: