sfqcyy
/
mindspore
forked from OSSInnovation/mindspore
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

!6174 delete transformer's enable_data_sink option 

Merge pull request !6174 from yuchaojie/transformer2
This commit is contained in:
mindspore-ci-bot 2020-09-15 10:41:13 +08:00 committed by Gitee
parent b83d921735 786de3252e
commit 122e966277
6 changed files with 69 additions and 58 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
model_zoo/official/nlp/transformer/README.md
View File

@ -101,10 +101,9 @@ python eval.py > eval.log 2>&1 &
usage: train.py [--distribute DISTRIBUTE] [--epoch_size N] [----device_num N] [--device_id N]
[--enable_save_ckpt ENABLE_SAVE_CKPT]
[--enable_lossscale ENABLE_LOSSSCALE] [--do_shuffle DO_SHUFFLE]
[--enable_data_sink ENABLE_DATA_SINK] [--save_checkpoint_steps N]
[--save_checkpoint_num N] [--save_checkpoint_path SAVE_CHECKPOINT_PATH]
[--data_path DATA_PATH]
[--bucket_boundaries BUCKET_LENGTH]
[--save_checkpoint_steps N] [--save_checkpoint_num N]
[--save_checkpoint_path SAVE_CHECKPOINT_PATH]
[--data_path DATA_PATH] [--bucket_boundaries BUCKET_LENGTH]
options:
--distribute pre_training by serveral devices: "true"(training by more than 1 device) | "false", default is "false"
@ -114,7 +113,6 @@ options:
--enable_save_ckpt enable save checkpoint: "true" | "false", default is "true"
--enable_lossscale enable lossscale: "true" | "false", default is "true"
--do_shuffle enable shuffle: "true" | "false", default is "true"
--enable_data_sink enable data sink: "true" | "false", default is "false"
--checkpoint_path path to load checkpoint files: PATH, default is ""
--save_checkpoint_steps steps for saving checkpoint files: N, default is 2500
--save_checkpoint_num number for saving checkpoint files: N, default is 30
@ -143,7 +141,7 @@ eval_config.py:
```
Parameters for dataset and network (Training/Evaluation):
batch_size batch size of input dataset: N, default is 96
seq_length length of input sequence: N, default is 128
seq_length max length of input sequence: N, default is 128
vocab_size size of each embedding vector: N, default is 36560
hidden_size size of Transformer encoder layers: N, default is 1024
num_hidden_layers number of hidden layers: N, default is 6
@ -181,7 +179,7 @@ Parameters for learning rate:
``` bash
paste train.tok.clean.bpe.32000.en train.tok.clean.bpe.32000.de > train.all
python create_data.py --input_file train.all --vocab_file vocab.bpe.32000 --output_file /path/ende-l128-mindrecord --max_seq_length 128 --bucket [16, 32, 48, 64, 128]
python create_data.py --input_file train.all --vocab_file vocab.bpe.32000 --output_file /path/ende-l128-mindrecord --max_seq_length 128 --bucket [16,32,48,64,128]
```
- Convert the original data to mindrecord for evaluation:

87
model_zoo/official/nlp/transformer/create_data.py
View File

@ -19,6 +19,7 @@ from __future__ import division
from __future__ import print_function
import argparse
import ast
import collections
import logging
import numpy as np
@ -51,23 +52,23 @@ class SampleInstance():
return self.__str__()
def write_instance_to_file(writer, instance, tokenizer, max_seq_length, bucket):
"""Create files from `SampleInstance`s."""
def _find_bucket_length(num):
def get_instance_features(instance, tokenizer, max_seq_length, bucket):
"""Get features from `SampleInstance`s."""
def _find_bucket_length(source_tokens, target_tokens):
source_ids = tokenizer.convert_tokens_to_ids(source_tokens)
target_ids = tokenizer.convert_tokens_to_ids(target_tokens)
num = max(len(source_ids), len(target_ids))
assert num <= bucket[-1]
for index in range(1, len(bucket)):
if bucket[index - 1] < num <= bucket[index]:
return bucket[index]
return bucket[0]
def _convert_ids_and_mask(input_tokens):
def _convert_ids_and_mask(input_tokens, seq_max_bucket_length):
input_ids = tokenizer.convert_tokens_to_ids(input_tokens)
input_mask = [1] * len(input_ids)
assert len(input_ids) <= max_seq_length
seq_max_bucket_length = _find_bucket_length(len(input_ids))
while len(input_ids) < seq_max_bucket_length:
input_ids.append(0)
input_mask.append(0)
@ -77,10 +78,11 @@ def write_instance_to_file(writer, instance, tokenizer, max_seq_length, bucket):
return input_ids, input_mask
source_sos_ids, source_sos_mask = _convert_ids_and_mask(instance.source_sos_tokens)
source_eos_ids, source_eos_mask = _convert_ids_and_mask(instance.source_eos_tokens)
target_sos_ids, target_sos_mask = _convert_ids_and_mask(instance.target_sos_tokens)
target_eos_ids, target_eos_mask = _convert_ids_and_mask(instance.target_eos_tokens)
seq_max_bucket_length = _find_bucket_length(instance.source_sos_tokens, instance.target_sos_tokens)
source_sos_ids, source_sos_mask = _convert_ids_and_mask(instance.source_sos_tokens, seq_max_bucket_length)
source_eos_ids, source_eos_mask = _convert_ids_and_mask(instance.source_eos_tokens, seq_max_bucket_length)
target_sos_ids, target_sos_mask = _convert_ids_and_mask(instance.target_sos_tokens, seq_max_bucket_length)
target_eos_ids, target_eos_mask = _convert_ids_and_mask(instance.target_eos_tokens, seq_max_bucket_length)
features = collections.OrderedDict()
features["source_sos_ids"] = np.asarray(source_sos_ids)
@ -92,8 +94,7 @@ def write_instance_to_file(writer, instance, tokenizer, max_seq_length, bucket):
features["target_eos_ids"] = np.asarray(target_eos_ids)
features["target_eos_mask"] = np.asarray(target_eos_mask)
writer.write_raw_data([features])
return features
return features, seq_max_bucket_length
def create_training_instance(source_words, target_words, max_seq_length, clip_to_max_len):
"""Creates `SampleInstance`s for a single sentence pair."""
@ -131,7 +132,8 @@ def main():
parser.add_argument("--clip_to_max_len", type=bool, default=False,
help='clip sequences to maximum sequence length.')
parser.add_argument("--max_seq_length", type=int, default=128, help='Maximum sequence length.')
parser.add_argument("--bucket", type=list, default=[16, 32, 48, 64, 128], help='bucket sequence length')
parser.add_argument("--bucket", type=ast.literal_eval, default=[16, 32, 48, 64, 128],
help='bucket sequence length')
args = parser.parse_args()
@ -141,29 +143,21 @@ def main():
for input_pattern in args.input_file.split(","):
input_files.append(input_pattern)
logging.info("*** Reading from input files ***")
logging.info("*** Read from input files ***")
for input_file in input_files:
logging.info(" %s", input_file)
output_file = args.output_file
logging.info("*** Writing to output files ***")
logging.info("*** Write to output files ***")
logging.info(" %s", output_file)
writer = FileWriter(output_file, args.num_splits)
data_schema = {"source_sos_ids": {"type": "int64", "shape": [-1]},
"source_sos_mask": {"type": "int64", "shape": [-1]},
"source_eos_ids": {"type": "int64", "shape": [-1]},
"source_eos_mask": {"type": "int64", "shape": [-1]},
"target_sos_ids": {"type": "int64", "shape": [-1]},
"target_sos_mask": {"type": "int64", "shape": [-1]},
"target_eos_ids": {"type": "int64", "shape": [-1]},
"target_eos_mask": {"type": "int64", "shape": [-1]}
}
writer.add_schema(data_schema, "tranformer hisi")
total_written = 0
total_read = 0
feature_dict = {}
for i in args.bucket:
feature_dict[i] = []
for input_file in input_files:
logging.info("*** Reading from %s ***", input_file)
with open(input_file, "r") as reader:
@ -174,7 +168,7 @@ def main():
total_read += 1
if total_read % 100000 == 0:
logging.info("%d ...", total_read)
logging.info("Read %d ...", total_read)
source_line, target_line = line.strip().split("\t")
source_tokens = tokenizer.tokenize(source_line)
@ -189,10 +183,13 @@ def main():
if instance is None:
continue
features = write_instance_to_file(writer, instance, tokenizer, args.max_seq_length, args.bucket)
total_written += 1
features, seq_max_bucket_length = get_instance_features(instance, tokenizer, args.max_seq_length,
args.bucket)
for key in feature_dict:
if key == seq_max_bucket_length:
feature_dict[key].append(features)
if total_written <= 20:
if total_read <= 10:
logging.info("*** Example ***")
logging.info("source tokens: %s", " ".join(
[tokenization.convert_to_printable(x) for x in instance.source_eos_tokens]))
@ -203,9 +200,33 @@ def main():
feature = features[feature_name]
logging.info("%s: %s", feature_name, feature)
writer.commit()
for i in args.bucket:
if args.num_splits == 1:
output_file_name = output_file
else:
output_file_name = output_file + '_' + str(i) + '_'
writer = FileWriter(output_file_name, args.num_splits)
data_schema = {"source_sos_ids": {"type": "int64", "shape": [-1]},
"source_sos_mask": {"type": "int64", "shape": [-1]},
"source_eos_ids": {"type": "int64", "shape": [-1]},
"source_eos_mask": {"type": "int64", "shape": [-1]},
"target_sos_ids": {"type": "int64", "shape": [-1]},
"target_sos_mask": {"type": "int64", "shape": [-1]},
"target_eos_ids": {"type": "int64", "shape": [-1]},
"target_eos_mask": {"type": "int64", "shape": [-1]}
}
writer.add_schema(data_schema, "tranformer")
features_ = feature_dict[i]
logging.info("Bucket length %d has %d samples, start writing...", i, len(features_))
for item in features_:
writer.write_raw_data([item])
total_written += 1
writer.commit()
logging.info("Wrote %d total instances", total_written)
if __name__ == "__main__":
logging.basicConfig(level=logging.INFO)
main()

4
model_zoo/official/nlp/transformer/scripts/run_distribute_train_ascend.sh
View File

@ -52,11 +52,11 @@ do
--enable_save_ckpt="true" \
--enable_lossscale="true" \
--do_shuffle="true" \
--enable_data_sink="false" \
--checkpoint_path="" \
--save_checkpoint_steps=2500 \
--save_checkpoint_num=30 \
--data_path=$DATA_PATH > log.txt 2>&1 &
--data_path=$DATA_PATH \
--bucket_boundaries=[16,32,48,64,128] > log.txt 2>&1 &
cd ../
done
cd ..

4
model_zoo/official/nlp/transformer/scripts/run_standalone_train_ascend.sh
View File

@ -37,9 +37,9 @@ python train.py \
--enable_save_ckpt="true" \
--enable_lossscale="true" \
--do_shuffle="true" \
--enable_data_sink="false" \
--checkpoint_path="" \
--save_checkpoint_steps=2500 \
--save_checkpoint_num=30 \
--data_path=$DATA_PATH > log.txt 2>&1 &
--data_path=$DATA_PATH \
--bucket_boundaries=[16,32,48,64,128] > log.txt 2>&1 &
cd ..

4
model_zoo/official/nlp/transformer/src/dataset.py
View File

@ -19,8 +19,8 @@ import mindspore.dataset as de
import mindspore.dataset.transforms.c_transforms as deC
from .config import transformer_net_cfg
de.config.set_seed(1)
def create_transformer_dataset(epoch_count=1, rank_size=1, rank_id=0, do_shuffle="true", enable_data_sink="true",
dataset_path=None, bucket_boundaries=None):
def create_transformer_dataset(epoch_count=1, rank_size=1, rank_id=0, do_shuffle="true", dataset_path=None,
bucket_boundaries=None):
"""create dataset"""
def batch_per_bucket(bucket_len, dataset_path):
dataset_path = dataset_path + "_" + str(bucket_len) + "_00"

16
model_zoo/official/nlp/transformer/train.py
View File

@ -16,6 +16,7 @@
import time
import argparse
import ast
import mindspore.common.dtype as mstype
from mindspore.common.tensor import Tensor
@ -94,8 +95,6 @@ def argparse_init():
help="Use lossscale or not, default is true.")
parser.add_argument("--do_shuffle", type=str, default="true", choices=['true', 'false'],
help="Enable shuffle for dataset, default is true.")
parser.add_argument("--enable_data_sink", type=str, default="false", choices=['true', 'false'],
help="Enable data sink, default is false.")
parser.add_argument("--checkpoint_path", type=str, default="", help="Checkpoint file path")
parser.add_argument("--enable_save_ckpt", type=str, default="true", choices=['true', 'false'],
help="Enable save checkpoint, default is true.")
@ -105,8 +104,8 @@ def argparse_init():
parser.add_argument("--save_checkpoint_path", type=str, default="./checkpoint/", help="Save checkpoint file path, "
"default is ./checkpoint/")
parser.add_argument("--data_path", type=str, default="", help="Data path, it is better to use absolute path")
parser.add_argument("--bucket_boundaries", type=list, default=[16, 32, 48, 64, 128], help="sequence length for "
"different bucket")
parser.add_argument("--bucket_boundaries", type=ast.literal_eval, default=[16, 32, 48, 64, 128],
help="sequence length for different bucket")
return parser
@ -131,7 +130,6 @@ def run_transformer_train():
rank_id = 0
dataset = create_transformer_dataset(epoch_count=1, rank_size=device_num,
rank_id=rank_id, do_shuffle=args.do_shuffle,
enable_data_sink=args.enable_data_sink,
dataset_path=args.data_path,
bucket_boundaries=args.bucket_boundaries)
@ -171,13 +169,7 @@ def run_transformer_train():
netwithgrads.set_train(True)
model = Model(netwithgrads)
enable_sink = (args.enable_data_sink == "true")
if enable_sink:
sink_size = args.save_checkpoint_steps
model.train(args.epoch_size*dataset.get_dataset_size()//sink_size, dataset, callbacks=callbacks,
dataset_sink_mode=enable_sink, sink_size=sink_size)
else:
model.train(args.epoch_size, dataset, callbacks=callbacks, dataset_sink_mode=enable_sink)
model.train(args.epoch_size, dataset, callbacks=callbacks, dataset_sink_mode=False)
if __name__ == '__main__':
run_transformer_train()