p85971406
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

!10654 gnmt add hub_config 

From: @gaojing22
Reviewed-by: @c_34,@wuxuejian
Signed-off-by: @c_34
This commit is contained in:
mindspore-ci-bot 2020-12-27 17:42:41 +08:00 committed by Gitee
parent 0c42c5b979 074964acc4
commit e817019141
3 changed files with 82 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
model_zoo/official/nlp/gnmt_v2/README.md
View File

@ -135,11 +135,13 @@ The GNMT network script and code result are as follows:
│ ├──lr_scheduler.py // Learning rate scheduler.
│ ├──optimizer.py // Optimizer.
├── scripts
│ ├──run_distributed_train_ascend.sh // shell script for distributed train on ascend.
│ ├──run_standalone_eval_ascend.sh // shell script for standalone eval on ascend.
│ ├──run_standalone_train_ascend.sh // shell script for standalone eval on ascend.
├── create_dataset.py // dataset preparation.
│ ├──run_distributed_train_ascend.sh // Shell script for distributed train on ascend.
│ ├──run_standalone_eval_ascend.sh // Shell script for standalone eval on ascend.
│ ├──run_standalone_train_ascend.sh // Shell script for standalone eval on ascend.
├── create_dataset.py // Dataset preparation.
├── eval.py // Infer API entry.
├── export.py // Export checkpoint file into air models.
├── mindspore_hub_conf.py // Hub config.
├── requirements.txt // Requirements of third party package.
├── train.py // Train API entry.
```

40
model_zoo/official/nlp/gnmt_v2/mindspore_hub_conf.py Normal file
View File

@ -0,0 +1,40 @@
# Copyright 2020 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""hub config."""
import mindspore.common.dtype as mstype
from config import GNMTConfig
from src.gnmt_model import GNMTNetworkWithLoss, GNMT
def get_config(config):
config = GNMTConfig.from_json_file(config)
config.compute_type = mstype.float16
config.dtype = mstype.float32
return config
def create_network(name, *args, **kwargs):
"""create gnmt network."""
if name == "gnmt":
if "config" in kwargs:
config = get_config(kwargs["config"])
else:
raise NotImplementedError(f"Please make sure the configuration file path is correct")
is_training = kwargs.get("is_training", False)
if is_training:
return GNMTNetworkWithLoss(config, is_training=is_training, *args)
return GNMT(config, *args)
raise NotImplementedError(f"{name} is not implemented in the repo")

45
model_zoo/official/nlp/gnmt_v2/src/gnmt_model/beam_search.py
View File

@ -172,6 +172,7 @@ class BeamSearchDecoder(nn.Cell):
max_decode_length=64,
sos_id=2,
eos_id=3,
is_using_while=False,
compute_type=mstype.float32):
super(BeamSearchDecoder, self).__init__()
@ -185,6 +186,7 @@ class BeamSearchDecoder(nn.Cell):
self.cov_penalty_factor = cov_penalty_factor
self.max_decode_length = max_decode_length
self.decoder = decoder
self.is_using_while = is_using_while
self.add = P.TensorAdd()
self.expand = P.ExpandDims()
@ -215,7 +217,12 @@ class BeamSearchDecoder(nn.Cell):
self.gather_nd = P.GatherNd()
self.start_ids = Tensor(np.full([batch_size * beam_width, 1], sos_id), mstype.int32)
self.init_seq = Tensor(np.full([batch_size, beam_width, 1], sos_id), mstype.int32)
if self.is_using_while:
self.start = Tensor(0, dtype=mstype.int32)
self.init_seq = Tensor(np.full([batch_size, beam_width, self.max_decode_length], sos_id),
mstype.int32)
else:
self.init_seq = Tensor(np.full([batch_size, beam_width, 1], sos_id), mstype.int32)
init_scores = np.tile(np.array([[0.] + [-INF] * (beam_width - 1)]), [batch_size, 1])
self.init_scores = Tensor(init_scores, mstype.float32)
@ -259,7 +266,7 @@ class BeamSearchDecoder(nn.Cell):
self.sub = P.Sub()
def one_step(self, cur_input_ids, enc_states, enc_attention_mask, state_log_probs,
state_seq, state_length, decoder_hidden_state=None, accu_attn_scores=None,
state_seq, state_length, idx=None, decoder_hidden_state=None, accu_attn_scores=None,
state_finished=None):
"""
Beam search one_step output.
@ -359,7 +366,13 @@ class BeamSearchDecoder(nn.Cell):
self.hidden_size))
# update state_seq
state_seq = self.concat((seq, self.expand(word_indices, -1)))
if self.is_using_while:
state_seq_new = self.cast(seq, mstype.float32)
word_indices_fp32 = self.cast(word_indices, mstype.float32)
state_seq_new[:, :, idx] = word_indices_fp32
state_seq = self.cast(state_seq_new, mstype.int32)
else:
state_seq = self.concat((seq, self.expand(word_indices, -1)))
cur_input_ids = self.reshape(word_indices, (-1, 1))
state_log_probs = topk_scores
@ -388,11 +401,22 @@ class BeamSearchDecoder(nn.Cell):
decoder_hidden_state = self.decoder_hidden_state
accu_attn_scores = self.accu_attn_scores
for _ in range(self.max_decode_length + 1):
cur_input_ids, state_log_probs, state_seq, state_length, decoder_hidden_state, accu_attn_scores, \
state_finished = self.one_step(cur_input_ids, enc_states, enc_attention_mask, state_log_probs,
state_seq, state_length, decoder_hidden_state, accu_attn_scores,
state_finished)
if not self.is_using_while:
for _ in range(self.max_decode_length + 1):
cur_input_ids, state_log_probs, state_seq, state_length, decoder_hidden_state, accu_attn_scores, \
state_finished = self.one_step(cur_input_ids, enc_states, enc_attention_mask, state_log_probs,
state_seq, state_length, None, decoder_hidden_state, accu_attn_scores,
state_finished)
else:
idx = self.start + 1
ends = self.start + self.max_decode_length + 1
while idx < ends:
cur_input_ids, state_log_probs, state_seq, state_length, decoder_hidden_state, accu_attn_scores, \
state_finished = self.one_step(cur_input_ids, enc_states, enc_attention_mask, state_log_probs,
state_seq, state_length, idx, decoder_hidden_state, accu_attn_scores,
state_finished)
idx = idx + 1
# add length penalty scores
penalty_len = self.length_penalty(state_length)
# return penalty_len
@ -408,6 +432,9 @@ class BeamSearchDecoder(nn.Cell):
gather_indices = self.concat((self.expand(self.batch_ids, -1), self.expand(top_beam_indices, -1)))
# sort sequence and attention scores
predicted_ids = self.gather_nd(state_seq, gather_indices)
predicted_ids = predicted_ids[:, 0:1, 1:(self.max_decode_length + 1)]
if not self.is_using_while:
predicted_ids = predicted_ids[:, 0:1, 1:(self.max_decode_length + 1)]
else:
predicted_ids = predicted_ids[:, 0:1, :self.max_decode_length]
return predicted_ids