-
Notifications
You must be signed in to change notification settings - Fork 5.5k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Simplify Machine Translation demo by using Trainer API #10895
Changes from all commits
File filter
Filter by extension
Conversations
Jump to
Diff view
Diff view
There are no files selected for viewing
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,7 @@ | ||
file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") | ||
string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") | ||
|
||
# default test | ||
foreach(src ${TEST_OPS}) | ||
py_test(${src} SRCS ${src}.py) | ||
endforeach() |
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,319 @@ | ||
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. | ||
# | ||
# 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. | ||
import contextlib | ||
|
||
import numpy as np | ||
import paddle | ||
import paddle.fluid as fluid | ||
import paddle.fluid.framework as framework | ||
import paddle.fluid.layers as pd | ||
from paddle.fluid.executor import Executor | ||
from functools import partial | ||
import unittest | ||
import os | ||
|
||
dict_size = 30000 | ||
source_dict_dim = target_dict_dim = dict_size | ||
hidden_dim = 32 | ||
word_dim = 16 | ||
batch_size = 2 | ||
max_length = 8 | ||
topk_size = 50 | ||
trg_dic_size = 10000 | ||
beam_size = 2 | ||
|
||
decoder_size = hidden_dim | ||
|
||
|
||
def encoder(is_sparse): | ||
# encoder | ||
src_word_id = pd.data( | ||
name="src_word_id", shape=[1], dtype='int64', lod_level=1) | ||
src_embedding = pd.embedding( | ||
input=src_word_id, | ||
size=[dict_size, word_dim], | ||
dtype='float32', | ||
is_sparse=is_sparse, | ||
param_attr=fluid.ParamAttr(name='vemb')) | ||
|
||
fc1 = pd.fc(input=src_embedding, size=hidden_dim * 4, act='tanh') | ||
lstm_hidden0, lstm_0 = pd.dynamic_lstm(input=fc1, size=hidden_dim * 4) | ||
encoder_out = pd.sequence_last_step(input=lstm_hidden0) | ||
return encoder_out | ||
|
||
|
||
def decoder_train(context, is_sparse): | ||
# decoder | ||
trg_language_word = pd.data( | ||
name="target_language_word", shape=[1], dtype='int64', lod_level=1) | ||
trg_embedding = pd.embedding( | ||
input=trg_language_word, | ||
size=[dict_size, word_dim], | ||
dtype='float32', | ||
is_sparse=is_sparse, | ||
param_attr=fluid.ParamAttr(name='vemb')) | ||
|
||
rnn = pd.DynamicRNN() | ||
with rnn.block(): | ||
current_word = rnn.step_input(trg_embedding) | ||
pre_state = rnn.memory(init=context) | ||
current_state = pd.fc(input=[current_word, pre_state], | ||
size=decoder_size, | ||
act='tanh') | ||
|
||
current_score = pd.fc(input=current_state, | ||
size=target_dict_dim, | ||
act='softmax') | ||
rnn.update_memory(pre_state, current_state) | ||
rnn.output(current_score) | ||
|
||
return rnn() | ||
|
||
|
||
def decoder_decode(context, is_sparse): | ||
init_state = context | ||
array_len = pd.fill_constant(shape=[1], dtype='int64', value=max_length) | ||
counter = pd.zeros(shape=[1], dtype='int64', force_cpu=True) | ||
|
||
# fill the first element with init_state | ||
state_array = pd.create_array('float32') | ||
pd.array_write(init_state, array=state_array, i=counter) | ||
|
||
# ids, scores as memory | ||
ids_array = pd.create_array('int64') | ||
scores_array = pd.create_array('float32') | ||
|
||
init_ids = pd.data(name="init_ids", shape=[1], dtype="int64", lod_level=2) | ||
init_scores = pd.data( | ||
name="init_scores", shape=[1], dtype="float32", lod_level=2) | ||
|
||
pd.array_write(init_ids, array=ids_array, i=counter) | ||
pd.array_write(init_scores, array=scores_array, i=counter) | ||
|
||
cond = pd.less_than(x=counter, y=array_len) | ||
|
||
while_op = pd.While(cond=cond) | ||
with while_op.block(): | ||
pre_ids = pd.array_read(array=ids_array, i=counter) | ||
pre_state = pd.array_read(array=state_array, i=counter) | ||
pre_score = pd.array_read(array=scores_array, i=counter) | ||
|
||
# expand the lod of pre_state to be the same with pre_score | ||
pre_state_expanded = pd.sequence_expand(pre_state, pre_score) | ||
|
||
pre_ids_emb = pd.embedding( | ||
input=pre_ids, | ||
size=[dict_size, word_dim], | ||
dtype='float32', | ||
is_sparse=is_sparse) | ||
|
||
# use rnn unit to update rnn | ||
current_state = pd.fc(input=[pre_state_expanded, pre_ids_emb], | ||
size=decoder_size, | ||
act='tanh') | ||
current_state_with_lod = pd.lod_reset(x=current_state, y=pre_score) | ||
# use score to do beam search | ||
current_score = pd.fc(input=current_state_with_lod, | ||
size=target_dict_dim, | ||
act='softmax') | ||
topk_scores, topk_indices = pd.topk(current_score, k=topk_size) | ||
selected_ids, selected_scores = pd.beam_search( | ||
pre_ids, topk_indices, topk_scores, beam_size, end_id=10, level=0) | ||
|
||
pd.increment(x=counter, value=1, in_place=True) | ||
|
||
# update the memories | ||
pd.array_write(current_state, array=state_array, i=counter) | ||
pd.array_write(selected_ids, array=ids_array, i=counter) | ||
pd.array_write(selected_scores, array=scores_array, i=counter) | ||
|
||
pd.less_than(x=counter, y=array_len, cond=cond) | ||
|
||
translation_ids, translation_scores = pd.beam_search_decode( | ||
ids=ids_array, scores=scores_array) | ||
|
||
# return init_ids, init_scores | ||
|
||
return translation_ids, translation_scores | ||
|
||
|
||
def set_init_lod(data, lod, place): | ||
res = fluid.LoDTensor() | ||
res.set(data, place) | ||
res.set_lod(lod) | ||
return res | ||
|
||
|
||
def to_lodtensor(data, place): | ||
seq_lens = [len(seq) for seq in data] | ||
cur_len = 0 | ||
lod = [cur_len] | ||
for l in seq_lens: | ||
cur_len += l | ||
lod.append(cur_len) | ||
flattened_data = np.concatenate(data, axis=0).astype("int64") | ||
flattened_data = flattened_data.reshape([len(flattened_data), 1]) | ||
res = fluid.LoDTensor() | ||
res.set(flattened_data, place) | ||
res.set_lod([lod]) | ||
return res | ||
|
||
|
||
def train_program(is_sparse): | ||
context = encoder(is_sparse) | ||
rnn_out = decoder_train(context, is_sparse) | ||
label = pd.data( | ||
name="target_language_next_word", shape=[1], dtype='int64', lod_level=1) | ||
cost = pd.cross_entropy(input=rnn_out, label=label) | ||
avg_cost = pd.mean(cost) | ||
return avg_cost | ||
|
||
|
||
def train(use_cuda, is_sparse, is_local=True): | ||
EPOCH_NUM = 1 | ||
|
||
if use_cuda and not fluid.core.is_compiled_with_cuda(): | ||
return | ||
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() | ||
|
||
train_reader = paddle.batch( | ||
paddle.reader.shuffle( | ||
paddle.dataset.wmt14.train(dict_size), buf_size=1000), | ||
batch_size=batch_size) | ||
|
||
feed_order = [ | ||
'src_word_id', 'target_language_word', 'target_language_next_word' | ||
] | ||
|
||
def event_handler(event): | ||
if isinstance(event, fluid.EndStepEvent): | ||
print('pass_id=' + str(event.epoch) + ' batch=' + str(event.step)) | ||
if event.step == 10: | ||
trainer.stop() | ||
|
||
trainer = fluid.Trainer( | ||
train_func=partial(train_program, is_sparse), | ||
optimizer=fluid.optimizer.Adagrad( | ||
learning_rate=1e-4, | ||
regularization=fluid.regularizer.L2DecayRegularizer( | ||
regularization_coeff=0.1)), | ||
place=place) | ||
|
||
trainer.train( | ||
reader=train_reader, | ||
num_epochs=EPOCH_NUM, | ||
event_handler=event_handler, | ||
feed_order=feed_order) | ||
|
||
|
||
def decode_main(use_cuda, is_sparse): | ||
|
||
if use_cuda and not fluid.core.is_compiled_with_cuda(): | ||
return | ||
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() | ||
|
||
context = encoder(is_sparse) | ||
translation_ids, translation_scores = decoder_decode(context, is_sparse) | ||
|
||
exe = Executor(place) | ||
exe.run(framework.default_startup_program()) | ||
|
||
init_ids_data = np.array([1 for _ in range(batch_size)], dtype='int64') | ||
init_scores_data = np.array( | ||
[1. for _ in range(batch_size)], dtype='float32') | ||
init_ids_data = init_ids_data.reshape((batch_size, 1)) | ||
init_scores_data = init_scores_data.reshape((batch_size, 1)) | ||
init_lod = [i for i in range(batch_size)] + [batch_size] | ||
init_lod = [init_lod, init_lod] | ||
|
||
train_data = paddle.batch( | ||
paddle.reader.shuffle( | ||
paddle.dataset.wmt14.train(dict_size), buf_size=1000), | ||
batch_size=batch_size) | ||
for _, data in enumerate(train_data()): | ||
init_ids = set_init_lod(init_ids_data, init_lod, place) | ||
init_scores = set_init_lod(init_scores_data, init_lod, place) | ||
|
||
src_word_data = to_lodtensor(map(lambda x: x[0], data), place) | ||
|
||
result_ids, result_scores = exe.run( | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I don't see any Inferencer. We should use the high level api. There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Or will there be any sub-sequence PR to add the infer part? There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Discussed with Nicky and Jeff. We could add some simple test to translate a sample sentence later. There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I will talk to Longfei to see if there is any solution to not expose executor, will try to update in next PR |
||
framework.default_main_program(), | ||
feed={ | ||
'src_word_id': src_word_data, | ||
'init_ids': init_ids, | ||
'init_scores': init_scores | ||
}, | ||
fetch_list=[translation_ids, translation_scores], | ||
return_numpy=False) | ||
print result_ids.lod() | ||
break | ||
|
||
|
||
class TestMachineTranslation(unittest.TestCase): | ||
pass | ||
|
||
|
||
@contextlib.contextmanager | ||
def scope_prog_guard(): | ||
prog = fluid.Program() | ||
startup_prog = fluid.Program() | ||
scope = fluid.core.Scope() | ||
with fluid.scope_guard(scope): | ||
with fluid.program_guard(prog, startup_prog): | ||
yield | ||
|
||
|
||
def inject_test_train(use_cuda, is_sparse): | ||
f_name = 'test_{0}_{1}_train'.format('cuda' if use_cuda else 'cpu', 'sparse' | ||
if is_sparse else 'dense') | ||
|
||
def f(*args): | ||
with scope_prog_guard(): | ||
train(use_cuda, is_sparse) | ||
|
||
setattr(TestMachineTranslation, f_name, f) | ||
|
||
|
||
def inject_test_decode(use_cuda, is_sparse, decorator=None): | ||
f_name = 'test_{0}_{1}_decode'.format('cuda' | ||
if use_cuda else 'cpu', 'sparse' | ||
if is_sparse else 'dense') | ||
|
||
def f(*args): | ||
with scope_prog_guard(): | ||
decode_main(use_cuda, is_sparse) | ||
|
||
if decorator is not None: | ||
f = decorator(f) | ||
|
||
setattr(TestMachineTranslation, f_name, f) | ||
|
||
|
||
for _use_cuda_ in (False, True): | ||
for _is_sparse_ in (False, True): | ||
inject_test_train(_use_cuda_, _is_sparse_) | ||
|
||
for _use_cuda_ in (False, True): | ||
for _is_sparse_ in (False, True): | ||
|
||
_decorator_ = None | ||
if _use_cuda_: | ||
_decorator_ = unittest.skip( | ||
reason='Beam Search does not support CUDA!') | ||
|
||
inject_test_decode( | ||
is_sparse=_is_sparse_, use_cuda=_use_cuda_, decorator=_decorator_) | ||
|
||
if __name__ == '__main__': | ||
unittest.main() |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We should move away from using executor, right?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yeah, I think we should not expose executor. Probably we can write
decode_main()
similar to thetrain()
method above?There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The decode and train in this example are not compatible with each other, we cannot provide the save model from train and use it in infer
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We cannot use trainer.train either because during decode, it is not using optimizer or backward pass, it is doing a beam search
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yeah, and I think the GPU implementation of beam search is still missing?