Skip to content

P
Paddle

PaddlePaddle / Paddle
大约 2 年 前同步成功

通知 2325
Star 20933
Fork 5424
P
Paddle
提交 047f3a76 编写于 作者: Y Yancey1989
浏览文件
操作

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into develop

上级 f8bc4132 0d784405
隐藏空白更改
内联 并排
Showing with 1525 addition and 586 deletion
.gitmodules 0 → 100644
浏览文件 @ 047f3a76
[submodule "book"]
path = book
url = https://github.com/PaddlePaddle/book.git
.pre-commit-config.yaml
浏览文件 @ 047f3a76
......@@ -2,12 +2,12 @@
sha: c25201a00e6b0514370501050cf2a8538ac12270
hooks:
- id: remove-crlf
files: (?!.*third_party)^.*$
files: (?!.*third_party)^.*$ | (?!.*book)^.*$
- repo: https://github.com/reyoung/mirrors-yapf.git
sha: v0.13.2
hooks:
- id: yapf
files: (.*\.(py|bzl)|BUILD|.*\.BUILD|WORKSPACE)$ # Bazel BUILD files follow Python syntax.
files: (.*\.(py|bzl)|BUILD|.*\.BUILD|WORKSPACE)$
- repo: https://github.com/pre-commit/pre-commit-hooks
sha: 7539d8bd1a00a3c1bfd34cdb606d3a6372e83469
hooks:
......@@ -15,7 +15,7 @@
- id: check-merge-conflict
- id: check-symlinks
- id: detect-private-key
files: (?!.*third_party)^.*$
files: (?!.*third_party)^.*$ | (?!.*book)^.*$
- id: end-of-file-fixer
- repo: https://github.com/PaddlePaddle/clang-format-pre-commit-hook.git
sha: 28c0ea8a67a3e2dbbf4822ef44e85b63a0080a29
......
authors
浏览文件 @ 047f3a76
......@@ -29,13 +29,16 @@ Luo, Tao
Lyu, Qin
Mao, Hongyue
Qian, Xiaojun
Qiao, Longfei
Qi, Jun
Qin, Duohao
Shen, Guolong
Shi, Guangchuan
Song, Xiang
Wang, Helin
Wang, Jiang
Wang, Yanfei
Wang, Yi
Wang, Yong
Weng, Renliang
Xu, Tianbing
......
book @ 22ed2a01
Subproject commit 22ed2a01aee872f055b5f5f212428f481cefc10d
cmake/external/protobuf.cmake
浏览文件 @ 047f3a76
......@@ -14,7 +14,7 @@
INCLUDE(ExternalProject)
FIND_PACKAGE(Protobuf)
FIND_PACKAGE(Protobuf 3.1)
IF(NOT PROTOBUF_FOUND)
SET(PROTOBUF_SOURCES_DIR ${THIRD_PARTY_PATH}/protobuf)
......
demo/image_classification/api_v2_train.py
浏览文件 @ 047f3a76
......@@ -13,9 +13,10 @@
# limitations under the License
import sys
import paddle.v2 as paddle
from api_v2_vgg import vgg_bn_drop
from api_v2_resnet import resnet_cifar10
def main():
......@@ -23,16 +24,16 @@ def main():
classdim = 10
# PaddlePaddle init
paddle.init(use_gpu=True, trainer_count=1)
paddle.init(use_gpu=False, trainer_count=1)
image = paddle.layer.data(
name="image", type=paddle.data_type.dense_vector(datadim))
# Add neural network config
# option 1. resnet
net = resnet_cifar10(image, depth=32)
# net = resnet_cifar10(image, depth=32)
# option 2. vgg
# net = vgg_bn_drop(image)
net = vgg_bn_drop(image)
out = paddle.layer.fc(input=net,
size=classdim,
......@@ -68,8 +69,8 @@ def main():
result = trainer.test(
reader=paddle.batch(
paddle.dataset.cifar.test10(), batch_size=128),
reader_dict={'image': 0,
'label': 1})
feeding={'image': 0,
'label': 1})
print "\nTest with Pass %d, %s" % (event.pass_id, result.metrics)
# Create trainer
......@@ -83,8 +84,8 @@ def main():
batch_size=128),
num_passes=5,
event_handler=event_handler,
reader_dict={'image': 0,
'label': 1})
feeding={'image': 0,
'label': 1})
if __name__ == '__main__':
......
demo/introduction/api_train_v2.py
浏览文件 @ 047f3a76
......@@ -30,26 +30,26 @@ def main():
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 100 == 0:
print "Pass %d, Batch %d, Cost %f, %s" % (
event.pass_id, event.batch_id, event.cost, event.metrics)
print "Pass %d, Batch %d, Cost %f" % (
event.pass_id, event.batch_id, event.cost)
if isinstance(event, paddle.event.EndPass):
result = trainer.test(
reader=paddle.reader.batched(
uci_housing.test(), batch_size=2),
reader_dict={'x': 0,
if (event.pass_id + 1) % 10 == 0:
result = trainer.test(
reader=paddle.batch(
uci_housing.test(), batch_size=2),
feeding={'x': 0,
'y': 1})
if event.pass_id % 10 == 0:
print "Test %d, %s" % (event.pass_id, result.metrics)
print "Test %d, %.2f" % (event.pass_id, result.cost)
# training
trainer.train(
reader=paddle.reader.batched(
reader=paddle.batch(
paddle.reader.shuffle(
uci_housing.train(), buf_size=500),
batch_size=2),
reader_dict={'x': 0,
'y': 1},
feeding={'x': 0,
'y': 1},
event_handler=event_handler,
num_passes=30)
......
demo/mnist/.gitignore
浏览文件 @ 047f3a76
......@@ -5,3 +5,6 @@ plot.png
train.log
*pyc
.ipynb_checkpoints
params.pkl
params.tar
params.tar.gz
demo/mnist/api_train_v2.py
浏览文件 @ 047f3a76
import paddle.v2 as paddle
import gzip
def softmax_regression(img):
......@@ -71,7 +72,11 @@ def main():
cost = paddle.layer.classification_cost(input=predict, label=label)
parameters = paddle.parameters.create(cost)
try:
with gzip.open('params.tar.gz', 'r') as f:
parameters = paddle.parameters.Parameters.from_tar(f)
except IOError:
parameters = paddle.parameters.create(cost)
optimizer = paddle.optimizer.Momentum(
learning_rate=0.1 / 128.0,
......@@ -86,11 +91,15 @@ def main():
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 100 == 0:
if event.batch_id % 1000 == 0:
print "Pass %d, Batch %d, Cost %f, %s" % (
event.pass_id, event.batch_id, event.cost, event.metrics)
if isinstance(event, paddle.event.EndPass):
result = trainer.test(reader=paddle.reader.batched(
with gzip.open('params.tar.gz', 'w') as f:
parameters.to_tar(f)
elif isinstance(event, paddle.event.EndPass):
result = trainer.test(reader=paddle.batch(
paddle.dataset.mnist.test(), batch_size=128))
print "Test with Pass %d, Cost %f, %s\n" % (
event.pass_id, result.cost, result.metrics)
......@@ -110,17 +119,16 @@ def main():
print 'Best pass is %s, testing Avgcost is %s' % (best[0], best[1])
print 'The classification accuracy is %.2f%%' % (100 - float(best[2]) * 100)
test_creator = paddle.dataset.mnist.test()
test_data = []
for item in test_creator():
test_data.append(item[0])
if len(test_data) == 100:
break
# output is a softmax layer. It returns probabilities.
# Shape should be (100, 10)
probs = paddle.infer(
output=predict,
parameters=parameters,
reader=paddle.batch(
paddle.reader.firstn(
paddle.reader.map_readers(lambda item: (item[0], ),
paddle.dataset.mnist.test()),
n=100),
batch_size=32))
probs = paddle.infer(output=predict, parameters=parameters, input=test_data)
print probs.shape
......
demo/recommendation/api_train_v2.py 0 → 100644
浏览文件 @ 047f3a76
import paddle.v2 as paddle
import cPickle
import copy
def main():
paddle.init(use_gpu=False)
movie_title_dict = paddle.dataset.movielens.get_movie_title_dict()
uid = paddle.layer.data(
name='user_id',
type=paddle.data_type.integer_value(
paddle.dataset.movielens.max_user_id() + 1))
usr_emb = paddle.layer.embedding(input=uid, size=32)
usr_gender_id = paddle.layer.data(
name='gender_id', type=paddle.data_type.integer_value(2))
usr_gender_emb = paddle.layer.embedding(input=usr_gender_id, size=16)
usr_age_id = paddle.layer.data(
name='age_id',
type=paddle.data_type.integer_value(
len(paddle.dataset.movielens.age_table)))
usr_age_emb = paddle.layer.embedding(input=usr_age_id, size=16)
usr_job_id = paddle.layer.data(
name='job_id',
type=paddle.data_type.integer_value(paddle.dataset.movielens.max_job_id(
) + 1))
usr_job_emb = paddle.layer.embedding(input=usr_job_id, size=16)
usr_combined_features = paddle.layer.fc(
input=[usr_emb, usr_gender_emb, usr_age_emb, usr_job_emb],
size=200,
act=paddle.activation.Tanh())
mov_id = paddle.layer.data(
name='movie_id',
type=paddle.data_type.integer_value(
paddle.dataset.movielens.max_movie_id() + 1))
mov_emb = paddle.layer.embedding(input=mov_id, size=32)
mov_categories = paddle.layer.data(
name='category_id',
type=paddle.data_type.sparse_binary_vector(
len(paddle.dataset.movielens.movie_categories())))
mov_categories_hidden = paddle.layer.fc(input=mov_categories, size=32)
mov_title_id = paddle.layer.data(
name='movie_title',
type=paddle.data_type.integer_value_sequence(len(movie_title_dict)))
mov_title_emb = paddle.layer.embedding(input=mov_title_id, size=32)
mov_title_conv = paddle.networks.sequence_conv_pool(
input=mov_title_emb, hidden_size=32, context_len=3)
mov_combined_features = paddle.layer.fc(
input=[mov_emb, mov_categories_hidden, mov_title_conv],
size=200,
act=paddle.activation.Tanh())
inference = paddle.layer.cos_sim(
a=usr_combined_features, b=mov_combined_features, size=1, scale=5)
cost = paddle.layer.regression_cost(
input=inference,
label=paddle.layer.data(
name='score', type=paddle.data_type.dense_vector(1)))
parameters = paddle.parameters.create(cost)
trainer = paddle.trainer.SGD(cost=cost,
parameters=parameters,
update_equation=paddle.optimizer.Adam(
learning_rate=1e-4))
feeding = {
'user_id': 0,
'gender_id': 1,
'age_id': 2,
'job_id': 3,
'movie_id': 4,
'category_id': 5,
'movie_title': 6,
'score': 7
}
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 100 == 0:
print "Pass %d Batch %d Cost %.2f" % (
event.pass_id, event.batch_id, event.cost)
trainer.train(
reader=paddle.batch(
paddle.reader.shuffle(
paddle.dataset.movielens.train(), buf_size=8192),
batch_size=256),
event_handler=event_handler,
feeding=feeding,
num_passes=1)
user_id = 234
movie_id = 345
user = paddle.dataset.movielens.user_info()[user_id]
movie = paddle.dataset.movielens.movie_info()[movie_id]
feature = user.value() + movie.value()
def reader():
yield feature
infer_dict = copy.copy(feeding)
del infer_dict['score']
prediction = paddle.infer(
output=inference,
parameters=parameters,
reader=paddle.batch(
reader, batch_size=32),
feeding=infer_dict)
print(prediction + 5) / 2
if __name__ == '__main__':
main()
demo/semantic_role_labeling/api_train_v2.py
浏览文件 @ 047f3a76
......@@ -163,11 +163,11 @@ def main():
update_equation=optimizer)
parameters.set('emb', load_parameter(conll05.get_embedding(), 44068, 32))
trn_reader = paddle.reader.batched(
trn_reader = paddle.batch(
paddle.reader.shuffle(
conll05.test(), buf_size=8192), batch_size=10)
reader_dict = {
feeding = {
'word_data': 0,
'ctx_n2_data': 1,
'ctx_n1_data': 2,
......@@ -183,7 +183,7 @@ def main():
reader=trn_reader,
event_handler=event_handler,
num_passes=10000,
reader_dict=reader_dict)
feeding=feeding)
if __name__ == '__main__':
......
demo/sentiment/train_v2.py
浏览文件 @ 047f3a76
......@@ -18,11 +18,7 @@ from paddle.trainer_config_helpers.poolings import MaxPooling
import paddle.v2 as paddle
def convolution_net(input_dim,
class_dim=2,
emb_dim=128,
hid_dim=128,
is_predict=False):
def convolution_net(input_dim, class_dim=2, emb_dim=128, hid_dim=128):
data = paddle.layer.data("word",
paddle.data_type.integer_value_sequence(input_dim))
emb = paddle.layer.embedding(input=data, size=emb_dim)
......@@ -42,8 +38,7 @@ def stacked_lstm_net(input_dim,
class_dim=2,
emb_dim=128,
hid_dim=512,
stacked_num=3,
is_predict=False):
stacked_num=3):
"""
A Wrapper for sentiment classification task.
This network uses bi-directional recurrent network,
......@@ -110,7 +105,7 @@ def stacked_lstm_net(input_dim,
if __name__ == '__main__':
# init
paddle.init(use_gpu=True, trainer_count=4)
paddle.init(use_gpu=False, trainer_count=4)
# network config
print 'load dictionary...'
......@@ -143,11 +138,11 @@ if __name__ == '__main__':
sys.stdout.flush()
if isinstance(event, paddle.event.EndPass):
result = trainer.test(
reader=paddle.reader.batched(
reader=paddle.batch(
lambda: paddle.dataset.imdb.test(word_dict),
batch_size=128),
reader_dict={'word': 0,
'label': 1})
feeding={'word': 0,
'label': 1})
print "\nTest with Pass %d, %s" % (event.pass_id, result.metrics)
# create trainer
......@@ -156,11 +151,11 @@ if __name__ == '__main__':
update_equation=adam_optimizer)
trainer.train(
reader=paddle.reader.batched(
reader=paddle.batch(
paddle.reader.shuffle(
lambda: paddle.dataset.imdb.train(word_dict), buf_size=1000),
batch_size=100),
event_handler=event_handler,
reader_dict={'word': 0,
'label': 1},
feeding={'word': 0,
'label': 1},
num_passes=10)
demo/seqToseq/api_train_v2.py
浏览文件 @ 047f3a76
import os
import paddle.v2 as paddle
from seqToseq_net_v2 import seqToseq_net_v2
# Data Definiation.
# TODO:This code should be merged to dataset package.
data_dir = "./data/pre-wmt14"
src_lang_dict = os.path.join(data_dir, 'src.dict')
trg_lang_dict = os.path.join(data_dir, 'trg.dict')
source_dict_dim = len(open(src_lang_dict, "r").readlines())
target_dict_dim = len(open(trg_lang_dict, "r").readlines())
def read_to_dict(dict_path):
with open(dict_path, "r") as fin:
out_dict = {
line.strip(): line_count
for line_count, line in enumerate(fin)
}
return out_dict
src_dict = read_to_dict(src_lang_dict)
trg_dict = read_to_dict(trg_lang_dict)
train_list = os.path.join(data_dir, 'train.list')
test_list = os.path.join(data_dir, 'test.list')
UNK_IDX = 2
START = "<s>"
END = "<e>"
def _get_ids(s, dictionary):
words = s.strip().split()
return [dictionary[START]] + \
[dictionary.get(w, UNK_IDX) for w in words] + \
[dictionary[END]]
def train_reader(file_name):
def reader():
with open(file_name, 'r') as f:
for line_count, line in enumerate(f):
line_split = line.strip().split('\t')
if len(line_split) != 2:
continue
src_seq = line_split[0] # one source sequence
src_ids = _get_ids(src_seq, src_dict)
trg_seq = line_split[1] # one target sequence
trg_words = trg_seq.split()
trg_ids = [trg_dict.get(w, UNK_IDX) for w in trg_words]
# remove sequence whose length > 80 in training mode
if len(src_ids) > 80 or len(trg_ids) > 80:
continue
trg_ids_next = trg_ids + [trg_dict[END]]
trg_ids = [trg_dict[START]] + trg_ids
yield src_ids, trg_ids, trg_ids_next
return reader
def seqToseq_net(source_dict_dim, target_dict_dim):
### Network Architecture
word_vector_dim = 512 # dimension of word vector
decoder_size = 512 # dimension of hidden unit in GRU Decoder network
encoder_size = 512 # dimension of hidden unit in GRU Encoder network
#### Encoder
src_word_id = paddle.layer.data(
name='source_language_word',
type=paddle.data_type.integer_value_sequence(source_dict_dim))
src_embedding = paddle.layer.embedding(
input=src_word_id,
size=word_vector_dim,
param_attr=paddle.attr.ParamAttr(name='_source_language_embedding'))
src_forward = paddle.networks.simple_gru(
input=src_embedding, size=encoder_size)
src_backward = paddle.networks.simple_gru(
input=src_embedding, size=encoder_size, reverse=True)
encoded_vector = paddle.layer.concat(input=[src_forward, src_backward])
#### Decoder
with paddle.layer.mixed(size=decoder_size) as encoded_proj:
encoded_proj += paddle.layer.full_matrix_projection(
input=encoded_vector)
backward_first = paddle.layer.first_seq(input=src_backward)
with paddle.layer.mixed(
size=decoder_size, act=paddle.activation.Tanh()) as decoder_boot:
decoder_boot += paddle.layer.full_matrix_projection(
input=backward_first)
def gru_decoder_with_attention(enc_vec, enc_proj, current_word):
decoder_mem = paddle.layer.memory(
name='gru_decoder', size=decoder_size, boot_layer=decoder_boot)
context = paddle.networks.simple_attention(
encoded_sequence=enc_vec,
encoded_proj=enc_proj,
decoder_state=decoder_mem)
with paddle.layer.mixed(size=decoder_size * 3) as decoder_inputs:
decoder_inputs += paddle.layer.full_matrix_projection(input=context)
decoder_inputs += paddle.layer.full_matrix_projection(
input=current_word)
gru_step = paddle.layer.gru_step(
name='gru_decoder',
input=decoder_inputs,
output_mem=decoder_mem,
size=decoder_size)
with paddle.layer.mixed(
size=target_dict_dim,
bias_attr=True,
act=paddle.activation.Softmax()) as out:
out += paddle.layer.full_matrix_projection(input=gru_step)
return out
decoder_group_name = "decoder_group"
group_input1 = paddle.layer.StaticInputV2(input=encoded_vector, is_seq=True)
group_input2 = paddle.layer.StaticInputV2(input=encoded_proj, is_seq=True)
group_inputs = [group_input1, group_input2]
trg_embedding = paddle.layer.embedding(
input=paddle.layer.data(
name='target_language_word',
type=paddle.data_type.integer_value_sequence(target_dict_dim)),
size=word_vector_dim,
param_attr=paddle.attr.ParamAttr(name='_target_language_embedding'))
group_inputs.append(trg_embedding)
# For decoder equipped with attention mechanism, in training,
# target embeding (the groudtruth) is the data input,
# while encoded source sequence is accessed to as an unbounded memory.
# Here, the StaticInput defines a read-only memory
# for the recurrent_group.
decoder = paddle.layer.recurrent_group(
name=decoder_group_name,
step=gru_decoder_with_attention,
input=group_inputs)
lbl = paddle.layer.data(
name='target_language_next_word',
type=paddle.data_type.integer_value_sequence(target_dict_dim))
cost = paddle.layer.classification_cost(input=decoder, label=lbl)
return cost
def main():
paddle.init(use_gpu=False, trainer_count=1)
# source and target dict dim.
dict_size = 30000
source_dict_dim = target_dict_dim = dict_size
# define network topology
cost = seqToseq_net_v2(source_dict_dim, target_dict_dim)
cost = seqToseq_net(source_dict_dim, target_dict_dim)
parameters = paddle.parameters.create(cost)
optimizer = paddle.optimizer.Adam(learning_rate=1e-4)
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 10 == 0:
print "Pass %d, Batch %d, Cost %f, %s" % (
event.pass_id, event.batch_id, event.cost, event.metrics)
# define optimize method and trainer
optimizer = paddle.optimizer.Adam(learning_rate=1e-4)
trainer = paddle.trainer.SGD(cost=cost,
parameters=parameters,
update_equation=optimizer)
reader_dict = {
# define data reader
feeding = {
'source_language_word': 0,
'target_language_word': 1,
'target_language_next_word': 2
}
trn_reader = paddle.reader.batched(
wmt14_reader = paddle.batch(
paddle.reader.shuffle(
train_reader("data/pre-wmt14/train/train"), buf_size=8192),
paddle.dataset.wmt14.train(dict_size=dict_size), buf_size=8192),
batch_size=5)
# define event_handler callback
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 10 == 0:
print "Pass %d, Batch %d, Cost %f, %s" % (
event.pass_id, event.batch_id, event.cost, event.metrics)
# start to train
trainer.train(
reader=trn_reader,
reader=wmt14_reader,
event_handler=event_handler,
num_passes=10000,
reader_dict=reader_dict)
feeding=feeding)
if __name__ == '__main__':
......
demo/seqToseq/seqToseq_net_v2.py 已删除 100644 → 0
浏览文件 @ f8bc4132
import paddle.v2.activation as activation
import paddle.v2.attr as attr
import paddle.v2.data_type as data_type
import paddle.v2.layer as layer
import paddle.v2.networks as networks
def seqToseq_net_v2(source_dict_dim, target_dict_dim):
### Network Architecture
word_vector_dim = 512 # dimension of word vector
decoder_size = 512 # dimension of hidden unit in GRU Decoder network
encoder_size = 512 # dimension of hidden unit in GRU Encoder network
#### Encoder
src_word_id = layer.data(
name='source_language_word',
type=data_type.integer_value_sequence(source_dict_dim))
src_embedding = layer.embedding(
input=src_word_id,
size=word_vector_dim,
param_attr=attr.ParamAttr(name='_source_language_embedding'))
src_forward = networks.simple_gru(input=src_embedding, size=encoder_size)
src_backward = networks.simple_gru(
input=src_embedding, size=encoder_size, reverse=True)
encoded_vector = layer.concat(input=[src_forward, src_backward])
#### Decoder
with layer.mixed(size=decoder_size) as encoded_proj:
encoded_proj += layer.full_matrix_projection(input=encoded_vector)
backward_first = layer.first_seq(input=src_backward)
with layer.mixed(size=decoder_size, act=activation.Tanh()) as decoder_boot:
decoder_boot += layer.full_matrix_projection(input=backward_first)
def gru_decoder_with_attention(enc_vec, enc_proj, current_word):
decoder_mem = layer.memory(
name='gru_decoder', size=decoder_size, boot_layer=decoder_boot)
context = networks.simple_attention(
encoded_sequence=enc_vec,
encoded_proj=enc_proj,
decoder_state=decoder_mem)
with layer.mixed(size=decoder_size * 3) as decoder_inputs:
decoder_inputs += layer.full_matrix_projection(input=context)
decoder_inputs += layer.full_matrix_projection(input=current_word)
gru_step = layer.gru_step(
name='gru_decoder',
input=decoder_inputs,
output_mem=decoder_mem,
size=decoder_size)
with layer.mixed(
size=target_dict_dim, bias_attr=True,
act=activation.Softmax()) as out:
out += layer.full_matrix_projection(input=gru_step)
return out
decoder_group_name = "decoder_group"
group_input1 = layer.StaticInputV2(input=encoded_vector, is_seq=True)
group_input2 = layer.StaticInputV2(input=encoded_proj, is_seq=True)
group_inputs = [group_input1, group_input2]
trg_embedding = layer.embedding(
input=layer.data(
name='target_language_word',
type=data_type.integer_value_sequence(target_dict_dim)),
size=word_vector_dim,
param_attr=attr.ParamAttr(name='_target_language_embedding'))
group_inputs.append(trg_embedding)
# For decoder equipped with attention mechanism, in training,
# target embeding (the groudtruth) is the data input,
# while encoded source sequence is accessed to as an unbounded memory.
# Here, the StaticInput defines a read-only memory
# for the recurrent_group.
decoder = layer.recurrent_group(
name=decoder_group_name,
step=gru_decoder_with_attention,
input=group_inputs)
lbl = layer.data(
name='target_language_next_word',
type=data_type.integer_value_sequence(target_dict_dim))
cost = layer.classification_cost(input=decoder, label=lbl)
return cost
demo/word2vec/train_v2.py 0 → 100644
浏览文件 @ 047f3a76
import math
import paddle.v2 as paddle
dictsize = 1953
embsize = 32
hiddensize = 256
N = 5
def wordemb(inlayer):
wordemb = paddle.layer.table_projection(
input=inlayer,
size=embsize,
param_attr=paddle.attr.Param(
name="_proj",
initial_std=0.001,
learning_rate=1,
l2_rate=0, ))
return wordemb
def main():
paddle.init(use_gpu=False, trainer_count=1)
word_dict = paddle.dataset.imikolov.build_dict()
dict_size = len(word_dict)
firstword = paddle.layer.data(
name="firstw", type=paddle.data_type.integer_value(dict_size))
secondword = paddle.layer.data(
name="secondw", type=paddle.data_type.integer_value(dict_size))
thirdword = paddle.layer.data(
name="thirdw", type=paddle.data_type.integer_value(dict_size))
fourthword = paddle.layer.data(
name="fourthw", type=paddle.data_type.integer_value(dict_size))
nextword = paddle.layer.data(
name="fifthw", type=paddle.data_type.integer_value(dict_size))
Efirst = wordemb(firstword)
Esecond = wordemb(secondword)
Ethird = wordemb(thirdword)
Efourth = wordemb(fourthword)
contextemb = paddle.layer.concat(input=[Efirst, Esecond, Ethird, Efourth])
hidden1 = paddle.layer.fc(input=contextemb,
size=hiddensize,
act=paddle.activation.Sigmoid(),
layer_attr=paddle.attr.Extra(drop_rate=0.5),
bias_attr=paddle.attr.Param(learning_rate=2),
param_attr=paddle.attr.Param(
initial_std=1. / math.sqrt(embsize * 8),
learning_rate=1))
predictword = paddle.layer.fc(input=hidden1,
size=dict_size,
bias_attr=paddle.attr.Param(learning_rate=2),
act=paddle.activation.Softmax())
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 100 == 0:
result = trainer.test(
paddle.batch(
paddle.dataset.imikolov.test(word_dict, N), 32))
print "Pass %d, Batch %d, Cost %f, %s, Testing metrics %s" % (
event.pass_id, event.batch_id, event.cost, event.metrics,
result.metrics)
cost = paddle.layer.classification_cost(input=predictword, label=nextword)
parameters = paddle.parameters.create(cost)
adam_optimizer = paddle.optimizer.Adam(
learning_rate=3e-3,
regularization=paddle.optimizer.L2Regularization(8e-4))
trainer = paddle.trainer.SGD(cost, parameters, adam_optimizer)
trainer.train(
paddle.batch(paddle.dataset.imikolov.train(word_dict, N), 32),
num_passes=30,
event_handler=event_handler)
if __name__ == '__main__':
main()
doc/api/index_cn.rst
浏览文件 @ 047f3a76
API
===
\ No newline at end of file
===
模型配置 API
------------
.. toctree::
:maxdepth: 1
v2/model_configs.rst
数据 API
--------
.. toctree::
:maxdepth: 1
v2/data.rst
训练 API
--------
.. toctree::
:maxdepth: 1
v2/run_logic.rst
\ No newline at end of file
doc/api/index_en.rst
浏览文件 @ 047f3a76
......@@ -7,4 +7,20 @@ Model Config API
.. toctree::
:maxdepth: 1
v2/model_configs.rst
\ No newline at end of file
v2/model_configs.rst
Data API
--------
.. toctree::
:maxdepth: 1
v2/data.rst
Train API
---------
.. toctree::
:maxdepth: 1
v2/run_logic.rst
\ No newline at end of file
doc/api/v2/data.rst 0 → 100644
浏览文件 @ 047f3a76
================
Data Related API
================
#########
DataTypes
#########
.. automodule:: paddle.v2.data_type
:members:
##########
DataFeeder
##########
.. automodule:: paddle.v2.data_feeder
:members:
######
Reader
######
.. automodule:: paddle.v2.reader
:members:
.. automodule:: paddle.v2.reader.creator
:members:
#########
minibatch
#########
.. automodule:: paddle.v2.minibatch
:members:
#######
Dataset
#######
.. automodule:: paddle.v2.dataset
:members:
mnist
+++++
.. automodule:: paddle.v2.dataset.mnist
:members:
cifar
+++++
.. automodule:: paddle.v2.dataset.cifar
:members:
conll05
+++++++
.. automodule:: paddle.v2.dataset.conll05
:members:
imdb
++++
.. automodule:: paddle.v2.dataset.imdb
:members:
imikolov
++++++++
.. automodule:: paddle.v2.dataset.imikolov
:members:
movielens
+++++++++
.. automodule:: paddle.v2.dataset.movielens
:members:
sentiment
+++++++++
.. automodule:: paddle.v2.dataset.sentiment
:members:
uci_housing
+++++++++++
.. automodule:: paddle.v2.dataset.uci_housing
:members:
doc/api/v2/model_configs.rst
浏览文件 @ 047f3a76
#########################
Configuration Related API
#########################
======
Layers
======
.. automodule:: paddle.v2.layer
:members:
==========
Attributes
==========
.. automodule:: paddle.v2.attr
:members:
===========
Activations
===========
.. automodule:: paddle.v2.activation
:members:
========
Poolings
========
.. automodule:: paddle.v2.pooling
:members:
========
Networks
========
.. automodule:: paddle.v2.networks
:members:
==========
Optimizers
==========
.. automodule:: paddle.v2.optimizer
:members:
doc/api/v2/run_logic.rst 0 → 100644
浏览文件 @ 047f3a76
###########
Trainer API
###########
==========
Parameters
==========
.. automodule:: paddle.v2.parameters
:members:
=======
Trainer
=======
.. automodule:: paddle.v2.trainer
:members:
=====
Event
=====
.. automodule:: paddle.v2.event
:members:
=========
Inference
=========
.. autofunction:: paddle.v2.infer
\ No newline at end of file
doc/design/reader/README.md
浏览文件 @ 047f3a76
......@@ -23,19 +23,19 @@ An example implementation for single item data reader creator:
```python
def reader_creator_random_image(width, height):
def reader():
while True:
yield numpy.random.uniform(-1, 1, size=width*height)
return reader
def reader():
while True:
yield numpy.random.uniform(-1, 1, size=width*height)
return reader
```
An example implementation for multiple item data reader creator:
```python
def reader_creator_random_imageand_label(widht, height, label):
def reader():
while True:
yield numpy.random.uniform(-1, 1, size=width*height), label
return reader
def reader_creator_random_image_and_label(width, height, label):
def reader():
while True:
yield numpy.random.uniform(-1, 1, size=width*height), label
return reader
```
## Batch Reader Interface
......@@ -74,11 +74,11 @@ mnist_train_batch_reader = paddle.batch(mnist_train, 128)
Also easy to create custom batch reader:
```python
def custom_batch_reader():
while True:
batch = []
for i in xrange(128):
batch.append((numpy.random.uniform(-1, 1, 28*28),)) # note that it's a tuple being appended.
yield batch
while True:
batch = []
for i in xrange(128):
batch.append((numpy.random.uniform(-1, 1, 28*28),)) # note that it's a tuple being appended.
yield batch
mnist_random_image_batch_reader = custom_batch_reader
```
......@@ -123,16 +123,16 @@ We can do:
```python
def reader_creator_random_image(width, height):
def reader():
while True:
yield numpy.random.uniform(-1, 1, size=width*height)
return reader
def reader():
while True:
yield numpy.random.uniform(-1, 1, size=width*height)
return reader
def reader_creator_bool(t):
def reader:
while True:
yield t
return reader
def reader:
while True:
yield t
return reader
true_reader = reader_creator_bool(True)
false_reader = reader_creator_bool(False)
......@@ -172,18 +172,18 @@ We decided to use dictionary (`{"image":0, "label":1}`) instead of list (`["imag
```python
def image_reader_creator(image_path, label_path, n):
def reader():
f = open(image_path)
l = open(label_path)
images = numpy.fromfile(
f, 'ubyte', count=n * 28 * 28).reshape((n, 28 * 28)).astype('float32')
images = images / 255.0 * 2.0 - 1.0
labels = numpy.fromfile(l, 'ubyte', count=n).astype("int")
for i in xrange(n):
yield images[i, :], labels[i] # a single entry of data is created each time
f.close()
l.close()
return reader
def reader():
f = open(image_path)
l = open(label_path)
images = numpy.fromfile(
f, 'ubyte', count=n * 28 * 28).reshape((n, 28 * 28)).astype('float32')
images = images / 255.0 * 2.0 - 1.0
labels = numpy.fromfile(l, 'ubyte', count=n).astype("int")
for i in xrange(n):
yield images[i, :], labels[i] # a single entry of data is created each time
f.close()
l.close()
return reader
# images_reader_creator creates a reader
reader = image_reader_creator("/path/to/image_file", "/path/to/label_file", 1024)
......@@ -196,7 +196,7 @@ An example implementation of paddle.train could be:
```python
def train(batch_reader, mapping, batch_size, total_pass):
for pass_idx in range(total_pass):
for mini_batch in batch_reader(): # this loop will never end in online learning.
do_forward_backward(mini_batch, mapping)
for pass_idx in range(total_pass):
for mini_batch in batch_reader(): # this loop will never end in online learning.
do_forward_backward(mini_batch, mapping)
```
paddle/gserver/gradientmachines/MultiGradientMachine.cpp
浏览文件 @ 047f3a76
......@@ -346,7 +346,9 @@ Evaluator* MultiGradientMachine::makeEvaluator() const {
void MultiGradientMachine::eval(Evaluator* evaluator) const {
for (auto& thread : threads_) {
SetDevice device(thread->getDeviceId());
thread->getGradientMachine()->eval(evaluator);
if (thread->hasInputData()) {
thread->getGradientMachine()->eval(evaluator);
}
}
}
......@@ -356,14 +358,19 @@ void MultiGradientMachine::getOutArgs(std::vector<Argument>* outArgs,
REGISTER_TIMER("waitOutArgs");
thread->waitOutArgsReady();
}
outArgs_.resize(threads_[0]->getOutArgs().size());
outArgs_.resize(threads_[threads_.size() - 1]->getOutArgs().size());
REGISTER_TIMER("copyOutArgs");
for (size_t i = 0; i < outArgs_.size(); ++i) {
std::vector<Argument> args;
args.reserve(threads_.size());
for (auto& thread : threads_) {
args.push_back(thread->getOutArgs()[i]);
// If the thread input is empty, then the output is empty.
auto tmp = thread->getOutArgs();
if (tmp.size() > 0) {
args.push_back(tmp[i]);
}
}
outArgs_[i].concat(args, useGpu_, outArgStream_, passType);
}
......@@ -534,7 +541,7 @@ void TrainerThread::prefetch() {
void TrainerThread::forward() {
if (!inArgsCopied_) {
REGISTER_TIMER("copyInArgs");
copyInArgs();
batchSize_ = copyInArgs();
} else {
inArgsCopied_ = false;
}
......@@ -564,7 +571,12 @@ void TrainerThread::forward() {
{
REGISTER_TIMER("thread_forward");
gradientMachine_->forward(inArgs_, &outArgs_, multiMachine_->getPassType());
if (batchSize_ > 0) {
gradientMachine_->forward(
inArgs_, &outArgs_, multiMachine_->getPassType());
} else {
outArgs_.clear();
}
}
outArgsReadySem_.post();
}
......@@ -574,7 +586,13 @@ void TrainerThread::backward() {
if (multiMachine_->isPassGrad()) {
copyOutputGrad();
}
gradientMachine_->backward(backwardCallback_);
if (batchSize_ > 0) {
gradientMachine_->backward(backwardCallback_);
} else {
for (size_t i = parameters_.size(); i > 0; i--) {
backwardCallback(parameters_[i - 1].get());
}
}
if (multiMachine_->hasNonstaticCpuParamters()) {
mergeCpuGradients();
}
......@@ -732,7 +750,7 @@ void TrainerThread::notifyValueReady(int paramId) {
notifyValueDispatch(paramId);
}
void TrainerThread::copyInArgs() {
int TrainerThread::copyInArgs() {
const std::vector<Argument>& fullInArgs = multiMachine_->getInArgs();
int numThreads = multiMachine_->getAllThreads().size();
int32_t numSequences = fullInArgs[0].getNumSequences();
......@@ -748,7 +766,7 @@ void TrainerThread::copyInArgs() {
}
if (copySize == 0) {
return;
return 0;
}
for (size_t i = 0; i < fullInArgs.size(); i++) {
......@@ -758,6 +776,7 @@ void TrainerThread::copyInArgs() {
copySize,
FLAGS_parallel_nn ? false : multiMachine_->useGpu());
}
return copySize;
}
void TrainerThread::mergeCpuGradients() {
......
paddle/gserver/gradientmachines/MultiGradientMachine.h
浏览文件 @ 047f3a76
......@@ -387,6 +387,9 @@ public:
/// copy the output gradient from the main GradientMachine.
void copyOutputGrad();
/// Whether the thread has input data.
bool hasInputData() { return batchSize_ != 0; }
protected:
void mergeCpuGradients();
......@@ -407,7 +410,7 @@ protected:
void copyGradToBufferThread();
void gradCollectThread();
void copyInArgs();
int copyInArgs();
void forward();
void backward();
void backwardCallback(Parameter* para);
......@@ -467,6 +470,7 @@ protected:
/// indicate whether inArgs is copied before forward()
bool inArgsCopied_;
int batchSize_;
};
} // namespace paddle
python/paddle/trainer/PyDataProvider2.py
浏览文件 @ 047f3a76
......@@ -45,6 +45,23 @@ class CacheType(object):
class InputType(object):
"""
InputType is the base class for paddle input types.
.. note::
this is a base class, and should never be used by user.
:param dim: dimension of input. If the input is an integer, it means the
value range. Otherwise, it means the size of layer.
:type dim: int
:param seq_type: sequence type of input. 0 means it is not a sequence. 1
means it is a variable length sequence. 2 means it is a
nested sequence.
:type seq_type: int
:param type: data type of input.
:type type: int
"""
__slots__ = ['dim', 'seq_type', 'type']
def __init__(self, dim, seq_type, tp):
......@@ -54,20 +71,61 @@ class InputType(object):
def dense_slot(dim, seq_type=SequenceType.NO_SEQUENCE):
"""
Dense Vector. It means the input feature is dense float vector. For example,
if the input is an image with 28*28 pixels, the input of Paddle neural
network should be a dense vector with dimension 784.
:param dim: dimension of this vector.
:type dim: int
:param seq_type: sequence type of input.
:type seq_type: int
:return: An input type object.
:rtype: InputType
"""
return InputType(dim, seq_type, DataType.Dense)
def sparse_non_value_slot(dim, seq_type=SequenceType.NO_SEQUENCE):
"""
Sparse binary vector. It means the input feature is a sparse vector and the
every element in this vector is either zero or one.
:param dim: dimension of this vector.
:type dim: int
:param seq_type: sequence type of this input.
:type seq_type: int
:return: An input type object.
:rtype: InputType
"""
return InputType(dim, seq_type, DataType.SparseNonValue)
def sparse_value_slot(dim, seq_type=SequenceType.NO_SEQUENCE):
"""
Sparse vector. It means the input feature is a sparse vector. Most of the
elements in this vector are zero, others could be any float value.
:param dim: dimension of this vector.
:type dim: int
:param seq_type: sequence type of this input.
:type seq_type: int
:return: An input type object.
:rtype: InputType
"""
return InputType(dim, seq_type, DataType.SparseValue)
def index_slot(value_range, seq_type=SequenceType.NO_SEQUENCE):
"""Data type of integer.
"""
Data type of integer.
:param seq_type: sequence type of this input.
:type seq_type: int
:param value_range: range of this integer.
:type value_range: int
:return: An input type object
:rtype: InputType
"""
return InputType(value_range, seq_type, DataType.Index)
......@@ -76,10 +134,17 @@ dense_vector = dense_slot
sparse_binary_vector = sparse_non_value_slot
sparse_vector = sparse_value_slot
integer_value = index_slot
integer_value.__doc__ = index_slot.__doc__
def dense_vector_sequence(dim):
"""
Data type of a sequence of dense vector.
:param dim: dimension of dense vector.
:type dim: int
:return: An input type object
:rtype: InputType
"""
return dense_vector(dim, seq_type=SequenceType.SEQUENCE)
......@@ -88,6 +153,15 @@ def dense_vector_sub_sequence(dim):
def sparse_binary_vector_sequence(dim):
"""
Data type of a sequence of sparse vector, which every element is either zero
or one.
:param dim: dimension of sparse vector.
:type dim: int
:return: An input type object
:rtype: InputType
"""
return sparse_binary_vector(dim, seq_type=SequenceType.SEQUENCE)
......@@ -96,6 +170,15 @@ def sparse_binary_vector_sub_sequence(dim):
def sparse_vector_sequence(dim):
"""
Data type of a sequence of sparse vector, which most elements are zero,
others could be any float value.
:param dim: dimension of sparse vector.
:type dim: int
:return: An input type object
:rtype: InputType
"""
return sparse_vector(dim, seq_type=SequenceType.SEQUENCE)
......@@ -104,8 +187,11 @@ def sparse_vector_sub_sequence(dim):
def integer_value_sequence(value_range):
"""Data type of a sequence of integer.
"""
Data type of a sequence of integer.
:param value_range: range of each element.
:type value_range: int
"""
return integer_value(value_range, seq_type=SequenceType.SEQUENCE)
......@@ -115,7 +201,6 @@ def integer_value_sub_sequence(dim):
integer_sequence = integer_value_sequence
integer_sequence.__doc__ = integer_value_sequence.__doc__
class SingleSlotWrapper(object):
......
python/paddle/trainer_config_helpers/layer_math.py
浏览文件 @ 047f3a76
......@@ -39,6 +39,7 @@ register_unary_math_op('abs', act.AbsActivation())
register_unary_math_op('sigmoid', act.SigmoidActivation())
register_unary_math_op('tanh', act.TanhActivation())
register_unary_math_op('square', act.SquareActivation())
register_unary_math_op('relu', act.ReluActivation())
def add(layeroutput, other):
......
python/paddle/trainer_config_helpers/layers.py
浏览文件 @ 047f3a76
......@@ -795,17 +795,16 @@ def data_layer(name, size, height=None, width=None, layer_attr=None):
.. code-block:: python
data = data_layer(name="input",
size=1000)
data = data_layer(name="input", size=1000)
:param name: Name of this data layer.
:type name: basestring
:param size: Size of this data layer.
:type size: int
:param height: Height of this data layer, used for image
:type size: int|None
:type height: int|None
:param width: Width of this data layer, used for image
:type size: int|None
:type width: int|None
:param layer_attr: Extra Layer Attribute.
:type layer_attr: ExtraLayerAttribute.
:return: LayerOutput object.
......
python/paddle/trainer_config_helpers/tests/configs/math_ops.py
浏览文件 @ 047f3a76
......@@ -7,8 +7,9 @@ x = layer_math.exp(x)
x = layer_math.log(x)
x = layer_math.abs(x)
x = layer_math.sigmoid(x)
x = layer_math.tanh(x)
x = layer_math.square(x)
x = layer_math.square(x)
x = layer_math.relu(x)
y = 1 + x
y = y + 1
y = x + y
......
python/paddle/trainer_config_helpers/tests/configs/protostr/math_ops.protostr
浏览文件 @ 047f3a76
......@@ -65,13 +65,28 @@ layers {
}
}
}
layers {
name: "__tanh_0__"
type: "mixed"
size: 100
active_type: "tanh"
inputs {
input_layer_name: "__sigmoid_0__"
proj_conf {
type: "identity"
name: "___tanh_0__.w0"
input_size: 100
output_size: 100
}
}
}
layers {
name: "__square_0__"
type: "mixed"
size: 100
active_type: "square"
inputs {
input_layer_name: "__sigmoid_0__"
input_layer_name: "__tanh_0__"
proj_conf {
type: "identity"
name: "___square_0__.w0"
......@@ -81,15 +96,15 @@ layers {
}
}
layers {
name: "__square_1__"
name: "__relu_0__"
type: "mixed"
size: 100
active_type: "square"
active_type: "relu"
inputs {
input_layer_name: "__square_0__"
proj_conf {
type: "identity"
name: "___square_1__.w0"
name: "___relu_0__.w0"
input_size: 100
output_size: 100
}
......@@ -101,7 +116,7 @@ layers {
size: 100
active_type: ""
inputs {
input_layer_name: "__square_1__"
input_layer_name: "__relu_0__"
}
slope: 1.0
intercept: 1
......@@ -123,7 +138,7 @@ layers {
size: 100
active_type: ""
inputs {
input_layer_name: "__square_1__"
input_layer_name: "__relu_0__"
proj_conf {
type: "identity"
name: "___mixed_0__.w0"
......@@ -147,7 +162,7 @@ layers {
size: 100
active_type: ""
inputs {
input_layer_name: "__square_1__"
input_layer_name: "__relu_0__"
}
slope: -1.0
intercept: 0.0
......@@ -339,8 +354,9 @@ sub_models {
layer_names: "__log_0__"
layer_names: "__abs_0__"
layer_names: "__sigmoid_0__"
layer_names: "__tanh_0__"
layer_names: "__square_0__"
layer_names: "__square_1__"
layer_names: "__relu_0__"
layer_names: "__slope_intercept_layer_0__"
layer_names: "__slope_intercept_layer_1__"
layer_names: "__mixed_0__"
......
python/paddle/v2/activation.py
浏览文件 @ 047f3a76
......@@ -12,26 +12,15 @@
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle.trainer_config_helpers.activations import *
import paddle.trainer_config_helpers.activations
import copy
__all__ = [
"Base", "Tanh", "Sigmoid", "Softmax", "Identity", "Linear",
'SequenceSoftmax', "Exp", "Relu", "BRelu", "SoftRelu", "STanh", "Abs",
"Square", "Log"
]
__all__ = []
Base = BaseActivation
Tanh = TanhActivation
Sigmoid = SigmoidActivation
Softmax = SoftmaxActivation
SequenceSoftmax = SequenceSoftmaxActivation
Identity = IdentityActivation
Linear = Identity
Relu = ReluActivation
BRelu = BReluActivation
SoftRelu = SoftReluActivation
STanh = STanhActivation
Abs = AbsActivation
Square = SquareActivation
Exp = ExpActivation
Log = LogActivation
suffix = 'Activation'
for act in paddle.trainer_config_helpers.activations.__all__:
new_name = act[:-len(suffix)]
globals()[new_name] = copy.copy(
getattr(paddle.trainer_config_helpers.activations, act))
globals()[new_name].__name__ = new_name
__all__.append(new_name)
python/paddle/v2/attr.py
浏览文件 @ 047f3a76
......@@ -12,12 +12,16 @@
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle.trainer_config_helpers.attrs import *
import paddle.trainer_config_helpers.attrs
__all__ = [
"Param",
"Extra",
]
Param = ParameterAttribute
Extra = ExtraLayerAttribute
Param = paddle.trainer_config_helpers.attrs.ParameterAttribute
Extra = paddle.trainer_config_helpers.attrs.ExtraLayerAttribute
for each in paddle.trainer_config_helpers.attrs.__all__:
globals()[each] = getattr(paddle.trainer_config_helpers.attrs, each)
__all__.append(each)
python/paddle/v2/config_base.py
浏览文件 @ 047f3a76
......@@ -13,12 +13,55 @@
# limitations under the License.
import collections
import re
from paddle.trainer_config_helpers.default_decorators import wrap_name_default
import paddle.trainer_config_helpers as conf_helps
class LayerType(type):
def __new__(cls, name, bases, attrs):
method_name = attrs.get('METHOD_NAME', None)
if method_name is not None:
method = getattr(conf_helps, method_name)
if method.__doc__ is not None:
mapper = attrs.get("__map_docstr__", None)
if mapper is not None:
attrs['__doc__'] = LayerType.__map_docstr__(
mapper(method.__doc__),
method_name=method_name,
name=name)
else:
attrs['__doc__'] = LayerType.__map_docstr__(
method.__doc__, method_name=method_name, name=name)
return super(LayerType, cls).__new__(cls, name, bases, attrs)
@staticmethod
def __map_docstr__(doc, name, method_name):
assert isinstance(doc, basestring)
# replace LayerOutput to paddle.v2.config_base.Layer
doc = doc.replace("LayerOutput", "paddle.v2.config_base.Layer")
doc = doc.replace('ParameterAttribute',
'paddle.v2.attr.ParameterAttribute')
doc = re.sub(r'ExtraLayerAttribute[^\s]?',
'paddle.v2.attr.ExtraAttribute', doc)
# xxx_layer to xxx
doc = re.sub(r"(?P<name>[a-z]+)_layer", r"\g<name>", doc)
# XxxxActivation to paddle.v2.Activation.Xxxx
doc = re.sub(r"(?P<name>[A-Z][a-zA-Z]+)Activation",
r"paddle.v2.Activation.\g<name>", doc)
# TODO(yuyang18): Add more rules if needed.
return doc
class Layer(object):
__metaclass__ = LayerType
def __init__(self, name=None, parent_layers=None):
assert isinstance(parent_layers, dict)
self.name = name
......@@ -80,6 +123,8 @@ def __convert_to_v2__(method_name, parent_names, is_default_name=True):
wrapper = None
class V2LayerImpl(Layer):
METHOD_NAME = method_name
def __init__(self, **kwargs):
parent_layers = dict()
other_kwargs = dict()
......
python/paddle/v2/data_feeder.py
浏览文件 @ 047f3a76
......@@ -12,13 +12,20 @@
# See the License for the specific language governing permissions and
# limitations under the License.
from py_paddle import swig_paddle
from py_paddle import DataProviderConverter
import data_type
import paddle.trainer.PyDataProvider2 as pydp2
__all__ = ['DataFeeder']
def default_feeding_map(data_types):
reader_dict = dict()
for i, tp in enumerate(data_types):
reader_dict[tp[0]] = i
return reader_dict
class DataFeeder(DataProviderConverter):
"""
DataFeeder converts the data returned by paddle.reader into a data structure
......@@ -29,7 +36,10 @@ class DataFeeder(DataProviderConverter):
to feed it to C++ interface.
The example usage:
.. code-block:: python
data_types = [('image', paddle.data_type.dense_vector(784)),
('label', paddle.data_type.integer_value(10))]
reader_dict = {'image':0, 'label':1}
......@@ -43,49 +53,51 @@ class DataFeeder(DataProviderConverter):
# [ [1.0,2.0,3.0,4.0], 5, [6,7,8] ] # second sample
# ]
arg = feeder(minibatch_data)
.. note::
This module is for internal use only. Users should use the `reader`
interface.
:param data_types: A list to specify data name and type. Each item is
a tuple of (data_name, data_type).
:type data_types: list
:param reader_dict: A dictionary to specify the position of each data
in the input data.
:type feeding: dict
"""
def __init__(self, data_types, reader_dict):
"""
:param data_types: A list to specify data name and type. Each item is
a tuple of (data_name, data_type). For example:
[('image', paddle.data_type.dense_vector(784)),
('label', paddle.data_type.integer_value(10))]
:type data_types: A list of tuple
:param reader_dict: A dictionary to specify the position of each data
in the input data.
:type reader_dict: dict()
"""
def __init__(self, data_types, feeding=None):
self.input_names = []
input_types = []
self.reader_dict = reader_dict
if feeding is None:
feeding = default_feeding_map(data_types)
self.feeding = feeding
for each in data_types:
self.input_names.append(each[0])
assert isinstance(each[1], data_type.InputType)
if not isinstance(each[1], pydp2.InputType):
raise TypeError("second item in each data_type should be an "
"InputType")
input_types.append(each[1])
DataProviderConverter.__init__(self, input_types)
def __len__(self):
return len(self.input_names)
def convert(self, dat, argument=None):
"""
:param dat: A list of mini-batch data. Each sample is a list or tuple
one feature or multiple features.
for example:
[
([0.2, 0.2], ), # first sample
([0.8, 0.3], ), # second sample
]
or,
[
[[0.2, 0.2], ], # first sample
[[0.8, 0.3], ], # second sample
]
:type dat: List
:type dat: list
:param argument: An Arguments object contains this mini-batch data with
one or multiple features. The Arguments definition is
in the API.
:type argument: swig_paddle.Arguments
:type argument: py_paddle.swig_paddle.Arguments
"""
def reorder_data(data):
......@@ -93,7 +105,7 @@ class DataFeeder(DataProviderConverter):
for each in data:
reorder = []
for name in self.input_names:
reorder.append(each[self.reader_dict[name]])
reorder.append(each[self.feeding[name]])
retv.append(reorder)
return retv
......
python/paddle/v2/data_type.py
浏览文件 @ 047f3a76
......@@ -12,11 +12,15 @@
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle.trainer.PyDataProvider2 import \
InputType, DataType, dense_vector, sparse_binary_vector,\
sparse_vector, integer_value, integer_value_sequence
import paddle.trainer.PyDataProvider2 as pydp2
__all__ = [
'InputType', 'DataType', 'dense_vector', 'sparse_binary_vector',
'sparse_vector', 'integer_value', 'integer_value_sequence'
import_list = [
nm for nm in dir(pydp2)
if '_' in nm and nm[0] != '_' and ('value' in nm or 'vector' in nm)
]
import_list.extend(['InputType'])
for nm in import_list:
globals()[nm] = getattr(pydp2, nm)
__all__ = import_list
python/paddle/v2/dataset/__init__.py
浏览文件 @ 047f3a76
......@@ -11,6 +11,9 @@
# 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.
"""
Dataset package.
"""
import mnist
import imikolov
......
python/paddle/v2/dataset/cifar.py
浏览文件 @ 047f3a76
......@@ -13,6 +13,8 @@
# limitations under the License.
"""
CIFAR dataset: https://www.cs.toronto.edu/~kriz/cifar.html
TODO(yuyang18): Complete the comments.
"""
import cPickle
......
python/paddle/v2/dataset/conll05.py
浏览文件 @ 047f3a76
......@@ -16,15 +16,17 @@ import tarfile
import gzip
import itertools
from common import download
__all__ = ['test, get_dict', 'get_embedding']
"""
Conll 2005 dataset. Paddle semantic role labeling Book and demo use this
dataset as an example. Because Conll 2005 is not free in public, the default
downloaded URL is test set of Conll 2005 (which is public). Users can change
URL and MD5 to their Conll dataset.
TODO(yuyang18): Complete comments.
"""
__all__ = ['test, get_dict', 'get_embedding']
DATA_URL = 'http://www.cs.upc.edu/~srlconll/conll05st-tests.tar.gz'
DATA_MD5 = '387719152ae52d60422c016e92a742fc'
WORDDICT_URL = 'http://paddlepaddle.bj.bcebos.com/demo/srl_dict_and_embedding/wordDict.txt'
......
python/paddle/v2/dataset/imdb.py
浏览文件 @ 047f3a76
......@@ -13,6 +13,8 @@
# limitations under the License.
"""
IMDB dataset: http://ai.stanford.edu/%7Eamaas/data/sentiment/aclImdb_v1.tar.gz
TODO(yuyang18): Complete comments.
"""
import paddle.v2.dataset.common
......
python/paddle/v2/dataset/imikolov.py
浏览文件 @ 047f3a76
......@@ -13,6 +13,8 @@
# limitations under the License.
"""
imikolov's simple dataset: http://www.fit.vutbr.cz/~imikolov/rnnlm/
Complete comments.
"""
import paddle.v2.dataset.common
import tarfile
......
python/paddle/v2/dataset/mnist.py
浏览文件 @ 047f3a76
......@@ -13,6 +13,9 @@
# limitations under the License.
"""
MNIST dataset.
This module will download dataset from http://yann.lecun.com/exdb/mnist/ and
parse train set and test set into paddle reader creators.
"""
import paddle.v2.dataset.common
import subprocess
......@@ -72,6 +75,15 @@ def reader_creator(image_filename, label_filename, buffer_size):
def train():
"""
MNIST train set creator.
It returns a reader creator, each sample in the reader is image pixels in
[0, 1] and label in [0, 9].
:return: Train reader creator
:rtype: callable
"""
return reader_creator(
paddle.v2.dataset.common.download(TRAIN_IMAGE_URL, 'mnist',
TRAIN_IMAGE_MD5),
......@@ -80,6 +92,15 @@ def train():
def test():
"""
MNIST test set cretor.
It returns a reader creator, each sample in the reader is image pixels in
[0, 1] and label in [0, 9].
:return: Test reader creator.
:rtype: callable
"""
return reader_creator(
paddle.v2.dataset.common.download(TEST_IMAGE_URL, 'mnist',
TEST_IMAGE_MD5),
......
python/paddle/v2/dataset/movielens.py
浏览文件 @ 047f3a76
......@@ -11,6 +11,11 @@
# 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.
"""
Movielens 1-M dataset.
TODO(yuyang18): Complete comments.
"""
import zipfile
from common import download
......@@ -18,7 +23,12 @@ import re
import random
import functools
__all__ = ['train_creator', 'test_creator']
__all__ = [
'train', 'test', 'get_movie_title_dict', 'max_movie_id', 'max_user_id',
'age_table', 'movie_categories', 'max_job_id', 'user_info', 'movie_info'
]
age_table = [1, 18, 25, 35, 45, 50, 56]
class MovieInfo(object):
......@@ -33,17 +43,32 @@ class MovieInfo(object):
[MOVIE_TITLE_DICT[w.lower()] for w in self.title.split()]
]
def __str__(self):
return "<MovieInfo id(%d), title(%s), categories(%s)>" % (
self.index, self.title, self.categories)
def __repr__(self):
return self.__str__()
class UserInfo(object):
def __init__(self, index, gender, age, job_id):
self.index = int(index)
self.is_male = gender == 'M'
self.age = [1, 18, 25, 35, 45, 50, 56].index(int(age))
self.age = age_table.index(int(age))
self.job_id = int(job_id)
def value(self):
return [self.index, 0 if self.is_male else 1, self.age, self.job_id]
def __str__(self):
return "<UserInfo id(%d), gender(%s), age(%d), job(%d)>" % (
self.index, "M"
if self.is_male else "F", age_table[self.age], self.job_id)
def __repr__(self):
return str(self)
MOVIE_INFO = None
MOVIE_TITLE_DICT = None
......@@ -54,7 +79,8 @@ USER_INFO = None
def __initialize_meta_info__():
fn = download(
url='http://files.grouplens.org/datasets/movielens/ml-1m.zip',
md5='c4d9eecfca2ab87c1945afe126590906')
module_name='movielens',
md5sum='c4d9eecfca2ab87c1945afe126590906')
global MOVIE_INFO
if MOVIE_INFO is None:
pattern = re.compile(r'^(.*)\((\d+)\)$')
......@@ -117,14 +143,63 @@ def __reader_creator__(**kwargs):
return lambda: __reader__(**kwargs)
train_creator = functools.partial(__reader_creator__, is_test=False)
test_creator = functools.partial(__reader_creator__, is_test=True)
train = functools.partial(__reader_creator__, is_test=False)
test = functools.partial(__reader_creator__, is_test=True)
def get_movie_title_dict():
__initialize_meta_info__()
return MOVIE_TITLE_DICT
def __max_index_info__(a, b):
if a.index > b.index:
return a
else:
return b
def max_movie_id():
__initialize_meta_info__()
return reduce(__max_index_info__, MOVIE_INFO.viewvalues()).index
def max_user_id():
__initialize_meta_info__()
return reduce(__max_index_info__, USER_INFO.viewvalues()).index
def __max_job_id_impl__(a, b):
if a.job_id > b.job_id:
return a
else:
return b
def max_job_id():
__initialize_meta_info__()
return reduce(__max_job_id_impl__, USER_INFO.viewvalues()).job_id
def movie_categories():
__initialize_meta_info__()
return CATEGORIES_DICT
def user_info():
__initialize_meta_info__()
return USER_INFO
def movie_info():
__initialize_meta_info__()
return MOVIE_INFO
def unittest():
for train_count, _ in enumerate(train_creator()()):
for train_count, _ in enumerate(train()()):
pass
for test_count, _ in enumerate(test_creator()()):
for test_count, _ in enumerate(test()()):
pass
print train_count, test_count
......
python/paddle/v2/dataset/sentiment.py
浏览文件 @ 047f3a76
......@@ -15,18 +15,19 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""
The script fetch and preprocess movie_reviews data set
The script fetch and preprocess movie_reviews data set that provided by NLTK
that provided by NLTK
TODO(yuyang18): Complete dataset.
"""
import common
import collections
import nltk
import numpy as np
from itertools import chain
import nltk
from nltk.corpus import movie_reviews
import common
__all__ = ['train', 'test', 'get_word_dict']
NUM_TRAINING_INSTANCES = 1600
NUM_TOTAL_INSTANCES = 2000
......
python/paddle/v2/dataset/uci_housing.py
浏览文件 @ 047f3a76
......@@ -11,6 +11,11 @@
# 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.
"""
UCI Housing dataset.
TODO(yuyang18): Complete comments.
"""
import numpy as np
import os
......
python/paddle/v2/dataset/wmt14.py
浏览文件 @ 047f3a76
......@@ -14,129 +14,92 @@
"""
wmt14 dataset
"""
import paddle.v2.dataset.common
import tarfile
import os.path
import itertools
import paddle.v2.dataset.common
__all__ = ['train', 'test', 'build_dict']
URL_DEV_TEST = 'http://www-lium.univ-lemans.fr/~schwenk/cslm_joint_paper/data/dev+test.tgz'
MD5_DEV_TEST = '7d7897317ddd8ba0ae5c5fa7248d3ff5'
URL_TRAIN = 'http://localhost:8000/train.tgz'
MD5_TRAIN = '72de99da2830ea5a3a2c4eb36092bbc7'
def word_count(f, word_freq=None):
add = paddle.v2.dataset.common.dict_add
if word_freq == None:
word_freq = {}
for l in f:
for w in l.strip().split():
add(word_freq, w)
add(word_freq, '<s>')
add(word_freq, '<e>')
return word_freq
def get_word_dix(word_freq):
TYPO_FREQ = 50
word_freq = filter(lambda x: x[1] > TYPO_FREQ, word_freq.items())
word_freq_sorted = sorted(word_freq, key=lambda x: (-x[1], x[0]))
words, _ = list(zip(*word_freq_sorted))
word_idx = dict(zip(words, xrange(len(words))))
word_idx['<unk>'] = len(words)
return word_idx
def get_word_freq(train, dev):
word_freq = word_count(train, word_count(dev))
if '<unk>' in word_freq:
# remove <unk> for now, since we will set it as last index
del word_freq['<unk>']
return word_freq
def build_dict():
base_dir = './wmt14-data'
train_en_filename = base_dir + '/train/train.en'
train_fr_filename = base_dir + '/train/train.fr'
dev_en_filename = base_dir + '/dev/ntst1213.en'
dev_fr_filename = base_dir + '/dev/ntst1213.fr'
if not os.path.exists(train_en_filename) or not os.path.exists(
train_fr_filename):
with tarfile.open(
paddle.v2.dataset.common.download(URL_TRAIN, 'wmt14',
MD5_TRAIN)) as tf:
tf.extractall(base_dir)
if not os.path.exists(dev_en_filename) or not os.path.exists(
dev_fr_filename):
with tarfile.open(
paddle.v2.dataset.common.download(URL_DEV_TEST, 'wmt14',
MD5_DEV_TEST)) as tf:
tf.extractall(base_dir)
f_en = open(train_en_filename)
f_fr = open(train_fr_filename)
f_en_dev = open(dev_en_filename)
f_fr_dev = open(dev_fr_filename)
word_freq_en = get_word_freq(f_en, f_en_dev)
word_freq_fr = get_word_freq(f_fr, f_fr_dev)
f_en.close()
f_fr.close()
f_en_dev.close()
f_fr_dev.close()
return get_word_dix(word_freq_en), get_word_dix(word_freq_fr)
def reader_creator(directory, path_en, path_fr, URL, MD5, dict_en, dict_fr):
# this is a small set of data for test. The original data is too large and will be add later.
URL_TRAIN = 'http://paddlepaddle.bj.bcebos.com/demo/wmt_shrinked_data/wmt14.tgz'
MD5_TRAIN = 'a755315dd01c2c35bde29a744ede23a6'
START = "<s>"
END = "<e>"
UNK = "<unk>"
UNK_IDX = 2
def __read_to_dict__(tar_file, dict_size):
def __to_dict__(fd, size):
out_dict = dict()
for line_count, line in enumerate(fd):
if line_count < size:
out_dict[line.strip()] = line_count
else:
break
return out_dict
with tarfile.open(tar_file, mode='r') as f:
names = [
each_item.name for each_item in f
if each_item.name.endswith("src.dict")
]
assert len(names) == 1
src_dict = __to_dict__(f.extractfile(names[0]), dict_size)
names = [
each_item.name for each_item in f
if each_item.name.endswith("trg.dict")
]
assert len(names) == 1
trg_dict = __to_dict__(f.extractfile(names[0]), dict_size)
return src_dict, trg_dict
def reader_creator(tar_file, file_name, dict_size):
def reader():
if not os.path.exists(path_en) or not os.path.exists(path_fr):
with tarfile.open(
paddle.v2.dataset.common.download(URL, 'wmt14', MD5)) as tf:
tf.extractall(directory)
f_en = open(path_en)
f_fr = open(path_fr)
UNK_en = dict_en['<unk>']
UNK_fr = dict_fr['<unk>']
for en, fr in itertools.izip(f_en, f_fr):
src_ids = [dict_en.get(w, UNK_en) for w in en.strip().split()]
tar_ids = [
dict_fr.get(w, UNK_fr)
for w in ['<s>'] + fr.strip().split() + ['<e>']
src_dict, trg_dict = __read_to_dict__(tar_file, dict_size)
with tarfile.open(tar_file, mode='r') as f:
names = [
each_item.name for each_item in f
if each_item.name.endswith(file_name)
]
# remove sequence whose length > 80 in training mode
if len(src_ids) == 0 or len(tar_ids) <= 1 or len(
src_ids) > 80 or len(tar_ids) > 80:
continue
yield src_ids, tar_ids[:-1], tar_ids[1:]
f_en.close()
f_fr.close()
for name in names:
for line in f.extractfile(name):
line_split = line.strip().split('\t')
if len(line_split) != 2:
continue
src_seq = line_split[0] # one source sequence
src_words = src_seq.split()
src_ids = [
src_dict.get(w, UNK_IDX)
for w in [START] + src_words + [END]
]
trg_seq = line_split[1] # one target sequence
trg_words = trg_seq.split()
trg_ids = [trg_dict.get(w, UNK_IDX) for w in trg_words]
# remove sequence whose length > 80 in training mode
if len(src_ids) > 80 or len(trg_ids) > 80:
continue
trg_ids_next = trg_ids + [trg_dict[END]]
trg_ids = [trg_dict[START]] + trg_ids
yield src_ids, trg_ids, trg_ids_next
return reader
def train(dict_en, dict_fr):
directory = './wmt14-data'
return reader_creator(directory, directory + '/train/train.en',
directory + '/train/train.fr', URL_TRAIN, MD5_TRAIN,
dict_en, dict_fr)
def train(dict_size):
return reader_creator(
paddle.v2.dataset.common.download(URL_TRAIN, 'wmt14', MD5_TRAIN),
'train/train', dict_size)
def test(dict_en, dict_fr):
directory = './wmt14-data'
return reader_creator(directory, directory + '/dev/ntst1213.en',
directory + '/dev/ntst1213.fr', URL_DEV_TEST,
MD5_DEV_TEST, dict_en, dict_fr)
def test(dict_size):
return reader_creator(
paddle.v2.dataset.common.download(URL_TRAIN, 'wmt14', MD5_TRAIN),
'test/test', dict_size)
python/paddle/v2/event.py
浏览文件 @ 047f3a76
......@@ -34,6 +34,10 @@ class WithMetric(object):
class TestResult(WithMetric):
"""
Result that trainer.test return.
"""
def __init__(self, evaluator, cost):
super(TestResult, self).__init__(evaluator)
self.cost = cost
......
python/paddle/v2/inference.py
浏览文件 @ 047f3a76
import numpy
import py_paddle.swig_paddle as api
import collections
import topology
import minibatch
from data_feeder import DataFeeder
import itertools
import numpy
__all__ = ['infer']
......@@ -21,10 +21,33 @@ class Inference(object):
self.__gradient_machine__ = gm
self.__data_types__ = topo.data_type()
def iter_infer(self, reader, reader_dict=None):
if reader_dict is None:
reader_dict = self.default_reader_dict()
feeder = DataFeeder(self.__data_types__, reader_dict)
def iter_infer(self, input=None, batch_size=None, reader=None,
feeding=None):
feeder = DataFeeder(self.__data_types__, feeding)
if reader is None:
assert input is not None and isinstance(input, collections.Iterable)
if not isinstance(input, collections.Iterable):
raise TypeError("When reader is None, input should be whole "
"inference data and should be iterable")
if batch_size is None:
if not hasattr(input, '__len__'):
raise ValueError("Should set batch size when input data "
"don't contain length.")
batch_size = len(input)
def __reader_impl__():
for each_sample in input:
if len(feeder) == 1:
yield [each_sample]
else:
yield each_sample
reader = minibatch.batch(__reader_impl__, batch_size=batch_size)
else:
if input is not None:
raise ValueError("User should set either input or reader, "
"should not set them both.")
self.__gradient_machine__.start()
for data_batch in reader():
yield self.__gradient_machine__.forwardTest(feeder(data_batch))
......@@ -47,13 +70,53 @@ class Inference(object):
else:
return retv
def default_reader_dict(self):
reader_dict = dict()
for i, tp in enumerate(self.__data_types__):
reader_dict[tp[0]] = i
return reader_dict
def infer(output,
parameters,
input=None,
batch_size=None,
reader=None,
feeding=None,
field='value'):
"""
Infer a neural network by given neural network output and parameters. The
user should pass either a batch of input data or reader method.
Example usages:
.. code-block:: python
result = paddle.infer(prediction, parameters, input=SomeData,
batch_size=32)
print result
:param output: output of the neural network that would be inferred
:type output: paddle.v2.config_base.Layer
:param parameters: parameters of the neural network.
:type parameters: paddle.v2.parameters.Parameters
:param input: input data batch. Should be a python iterable object, and each
element is the data batch.
:type input: collections.Iterable
:param batch_size: the batch size when perform inference. Default is the
length of input.
:type batch_size: int
:param reader: input data reader creator in batch. If this field is set, the
`input` and `batch_size` will be ignored.
:type reader: callable
:param feeding: Reader dictionary. Default could generate from input
value.
:param field: The prediction field. It should in [`value`, `ids`]. `value`
means return the prediction probabilities, `ids` means return
the prediction labels. Default is `value`
:type field: str
:return: a numpy array
:rtype: numpy.ndarray
"""
def infer(output, parameters, reader, reader_dict=None, field='value'):
inferer = Inference(output=output, parameters=parameters)
return inferer.infer(field=field, reader=reader, reader_dict=reader_dict)
return inferer.infer(
field=field,
input=input,
batch_size=batch_size,
reader=reader,
feeding=feeding)
python/paddle/v2/layer.py
浏览文件 @ 047f3a76
......@@ -28,7 +28,7 @@ The primary usage shows below.
act=paddle.activation.Softmax())
# use prediction instance where needed.
parameters = paddle.v2.parameters.create(cost)
parameters = paddle.parameters.create(cost)
"""
import collections
......@@ -47,26 +47,32 @@ from paddle.trainer.config_parser import \
RecurrentLayerGroupEnd, model_type
import activation
import re
import data_type
__all__ = ['parse_network', 'data']
__projection_names__ = filter(lambda x: x.endswith('_projection'),
dir(conf_helps))
__all__ += __projection_names__
__operator_names__ = filter(lambda x: x.endswith('_operator'), dir(conf_helps))
__all__ += __operator_names__
def parse_network(*outputs):
"""
parse all output layers and then generate a model config proto.
:param outputs:
:return:
Parse all output layers and then generate a ModelConfig object.
.. note::
This function is used internally in paddle.v2 module. User should never
invoke this method.
:param outputs: Output layers.
:type outputs: Layer
:return: A ModelConfig object instance.
:rtype: ModelConfig
"""
def __real_func__():
"""
__real_func__ is the function that config_parser.parse invoked. It is
the plain old paddle configuration function.
"""
context = dict()
real_output = [each.to_proto(context=context) for each in outputs]
conf_helps.outputs(real_output)
......@@ -81,6 +87,8 @@ So we also need to implement some special LayerV2.
class DataLayerV2(Layer):
METHOD_NAME = 'data_layer'
def __init__(self, name, type, **kwargs):
assert isinstance(type, data_type.InputType)
......@@ -99,6 +107,17 @@ class DataLayerV2(Layer):
args[each] = self.__kwargs__[each]
return getattr(conf_helps, self.__method_name__)(name=self.name, **args)
def __map_docstr__(doc):
doc = re.sub(r'(data = [^\)]+)\).*',
"data = paddle.layer.data(name=\"input\", "
"type=paddle.data_type.dense_vector(1000))", doc)
doc = re.sub(r':param size:.*',
':param type: Data type of this data layer', doc)
doc = re.sub(r':type size:.*',
":type size: paddle.v2.data_type.InputType", doc)
return doc
class WithExtraParent(Layer):
def extra_parent(self):
......@@ -347,6 +366,7 @@ class RecurrentLayerOutput(Layer):
LayerV2 = Layer
data = DataLayerV2
data.__name__ = 'data'
AggregateLevel = conf_helps.layers.AggregateLevel
ExpandLevel = conf_helps.layers.ExpandLevel
memory = MemoryV2
......@@ -386,6 +406,7 @@ def __convert_layer__(_new_name_, _old_name_, _parent_names_):
global __all__
__all__.append(_new_name_)
globals()[new_name] = __convert_to_v2__(_old_name_, _parent_names_)
globals()[new_name].__name__ = new_name
for each_layer_name in dir(conf_helps):
......@@ -399,21 +420,6 @@ del parent_names
del new_name
del each_layer_name
# convert projection
for prj in __projection_names__:
globals()[prj] = __convert_to_v2__(
prj, parent_names=['input'], is_default_name=False)
# convert operator
operator_list = [
# [V1_method_name, parent_names],
['dotmul_operator', ['a', 'b']],
['conv_operator', ['img', 'filter']]
]
for op in operator_list:
globals()[op[0]] = __convert_to_v2__(
op[0], parent_names=op[1], is_default_name=False)
@wrap_name_default()
def recurrent_group(step, input, name=None):
......@@ -464,3 +470,29 @@ def recurrent_group(step, input, name=None):
return retv[0]
else:
return retv
__projection_names__ = filter(lambda x: x.endswith('_projection'),
dir(conf_helps))
__all__ += __projection_names__
__operator_names__ = filter(lambda x: x.endswith('_operator'), dir(conf_helps))
__all__ += __operator_names__
# convert projection
for prj in __projection_names__:
globals()[prj] = __convert_to_v2__(
prj, parent_names=['input'], is_default_name=False)
globals()[prj].__name__ = prj
# convert operator
operator_list = [
# [V1_method_name, parent_names],
['dotmul_operator', ['a', 'b']],
['conv_operator', ['img', 'filter']]
]
for op in operator_list:
globals()[op[0]] = __convert_to_v2__(
op[0], parent_names=op[1], is_default_name=False)
globals()[op[0]].__name__ = op[0]
python/paddle/v2/minibatch.py
浏览文件 @ 047f3a76
......@@ -12,24 +12,30 @@
# See the License for the specific language governing permissions and
# limitations under the License.
__all__ = ['batch']
def batch(reader, batch_size):
"""
Create a batch reader.
Create a batched reader.
:param reader: the data reader to read from.
:param batch_size: batch_size
:return: the batch reader.
:type reader: callable
:param batch_size: size of each mini-batch
:type batch_size: int
:return: the batched reader.
:rtype: callable
"""
def batch_reader():
r = reader()
batch = []
b = []
for instance in r:
batch.append(instance)
if len(batch) == batch_size:
yield batch
batch = []
if batch:
yield batch
b.append(instance)
if len(b) == batch_size:
yield b
b = []
if b:
yield b
return batch_reader
python/paddle/v2/networks.py
浏览文件 @ 047f3a76
......@@ -38,6 +38,7 @@ def __initialize__():
parent_names=parents,
is_default_name='name' in argspec.args)
globals()[each_subnetwork] = v2_subnet
globals()[each_subnetwork].__name__ = each_subnetwork
global __all__
__all__.append(each_subnetwork)
......
python/paddle/v2/optimizer.py
浏览文件 @ 047f3a76
import py_paddle.swig_paddle as swig_api
import paddle.trainer_config_helpers.optimizers as v1_optimizers
import paddle.trainer_config_helpers.config_parser_utils as config_parser_utils
import paddle.v2
import paddle.trainer_config_helpers.optimizers as v1_optimizers
"""
Optimizers(update equation) for SGD method.
TODO(yuyang18): Complete comments.
"""
__all__ = [
'Momentum', 'Adam', 'Adamax', 'AdaGrad', 'DecayedAdaGrad', 'AdaDelta',
......@@ -44,7 +49,7 @@ class Optimizer(object):
class Momentum(Optimizer):
def __init__(self, momentum=None, sparse=False, **kwargs):
learning_method = v1_optimizers.MomentumOptimizer(
momentum=None, sparse=False)
momentum=momentum, sparse=sparse)
super(Momentum, self).__init__(
learning_method=learning_method, **kwargs)
......
python/paddle/v2/parameters.py
浏览文件 @ 047f3a76
import numpy as np
import py_paddle.swig_paddle as api
from paddle.proto.ParameterConfig_pb2 import ParameterConfig
import struct
import tarfile
import cStringIO
from topology import Topology
__all__ = ['Parameters', 'create']
......@@ -10,6 +12,7 @@ __all__ = ['Parameters', 'create']
def create(layers):
"""
Create parameter pool by topology.
:param layers:
:return:
"""
......@@ -67,6 +70,7 @@ class Parameters(object):
def keys(self):
"""
keys are the names of each parameter.
:return: list of parameter name
:rtype: list
"""
......@@ -75,6 +79,7 @@ class Parameters(object):
def names(self):
"""
names of each parameter.
:return: list of parameter name
:rtype: list
"""
......@@ -83,6 +88,7 @@ class Parameters(object):
def has_key(self, key):
"""
has_key return true if there are such parameter name == key
:param key: Parameter name
:type key: basestring
:return: True if contains such key
......@@ -118,6 +124,12 @@ class Parameters(object):
if len(self.__gradient_machines__) == 0:
# create new parameter in python numpy.
if len(self.__tmp_params__) != 0:
ret_list = [
mat for name, mat in self.__tmp_params__ if name == key
]
if len(ret_list) == 1:
return ret_list[0]
return np.ndarray(shape=shape, dtype=np.float32)
else:
for each_gradient_machine in self.__gradient_machines__:
......@@ -136,6 +148,7 @@ class Parameters(object):
def get_shape(self, key):
"""
get shape of the parameter.
:param key: parameter name
:type key: basestring
:return: parameter's shape
......@@ -190,6 +203,7 @@ class Parameters(object):
def set(self, parameter_name, value):
"""
Set parameter by parameter name & matrix.
:param parameter_name: parameter name
:type parameter_name: basestring
:param value: parameter matrix
......@@ -222,6 +236,67 @@ class Parameters(object):
self.__gradient_machines__.append(gradient_machine)
def serialize(self, name, f):
"""
:param name:
:param f:
:type f: file
:return:
"""
param = self.get(name)
size = reduce(lambda a, b: a * b, param.shape)
f.write(struct.pack("IIQ", 0, 4, size))
param = param.astype(np.float32)
f.write(param.tobytes())
def deserialize(self, name, f):
"""
:param name:
:param f:
:type f: file
:return:
"""
f.read(16) # header
arr = np.frombuffer(f.read(), dtype=np.float32)
self.set(name, arr.reshape(self.get_shape(name)))
def to_tar(self, f):
tar = tarfile.TarFile(fileobj=f, mode='w')
for nm in self.names():
buf = cStringIO.StringIO()
self.serialize(nm, buf)
tarinfo = tarfile.TarInfo(name=nm)
buf.seek(0)
tarinfo.size = len(buf.getvalue())
tar.addfile(tarinfo, buf)
conf = self.__param_conf__[nm]
confStr = conf.SerializeToString()
tarinfo = tarfile.TarInfo(name="%s.protobuf" % nm)
tarinfo.size = len(confStr)
buf = cStringIO.StringIO(confStr)
buf.seek(0)
tar.addfile(tarinfo, fileobj=buf)
@staticmethod
def from_tar(f):
params = Parameters()
tar = tarfile.TarFile(fileobj=f, mode='r')
for finfo in tar:
assert isinstance(finfo, tarfile.TarInfo)
if finfo.name.endswith('.protobuf'):
f = tar.extractfile(finfo)
conf = ParameterConfig()
conf.ParseFromString(f.read())
params.__append_config__(conf)
for param_name in params.names():
f = tar.extractfile(param_name)
params.deserialize(param_name, f)
return params
def __get_parameter_in_gradient_machine__(gradient_machine, name):
"""
......
python/paddle/v2/pooling.py
浏览文件 @ 047f3a76
......@@ -12,13 +12,15 @@
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle.trainer_config_helpers.poolings import *
import paddle.trainer_config_helpers.poolings
import copy
__all__ = ["Max", "CudnnMax", "Avg", "CudnnAvg", "Sum", "SquareRootN"]
__all__ = []
suffix = 'Pooling'
Max = MaxPooling
CudnnMax = CudnnMaxPooling
Avg = AvgPooling
CudnnAvg = CudnnAvgPooling
Sum = SumPooling
SquareRootN = SquareRootNPooling
for name in paddle.trainer_config_helpers.poolings.__all__:
new_name = name[:-len(suffix)]
globals()[new_name] = copy.copy(
getattr(paddle.trainer_config_helpers.poolings, name))
globals()[new_name].__name__ = new_name
__all__.append(new_name)
python/paddle/v2/reader/__init__.py
浏览文件 @ 047f3a76
......@@ -11,15 +11,64 @@
# 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.
"""
At training and testing time, PaddlePaddle programs need to read data. To ease
the users' work to write data reading code, we define that
# It would be too lengthy to require our users to prefix decorators with `decorator`.
# For example, we want the following line
#
# r = paddle.reader.decorator.bufferd(paddle.reader.creator.text("hello.txt"))
#
# to be a shorter version:
#
# r = paddle.reader.buffered(paddle.reader.creator.text("hello.txt"))
- A *reader* is a function that reads data (from file, network, random number
generator, etc) and yields data items.
- A *reader creator* is a function that returns a reader function.
- A *reader decorator* is a function, which accepts one or more readers, and
returns a reader.
- A *batch reader* is a function that reads data (from *reader*, file, network,
random number generator, etc) and yields a batch of data items.
#####################
Data Reader Interface
#####################
Indeed, *data reader* doesn't have to be a function that reads and yields data
items. It can be any function with no parameter that creates a iterable
(anything can be used in :code:`for x in iterable`)\:
.. code-block:: python
iterable = data_reader()
Element produced from the iterable should be a **single** entry of data,
**not** a mini batch. That entry of data could be a single item, or a tuple of
items.
Item should be of `supported type <http://www.paddlepaddle.org/doc/ui/data_provider
/pydataprovider2.html?highlight=dense_vector#input-types>`_ (e.g., numpy 1d
array of float32, int, list of int)
An example implementation for single item data reader creator:
.. code-block:: python
def reader_creator_random_image(width, height):
def reader():
while True:
yield numpy.random.uniform(-1, 1, size=width*height)
return reader
An example implementation for multiple item data reader creator:
.. code-block:: python
def reader_creator_random_image_and_label(width, height, label):
def reader():
while True:
yield numpy.random.uniform(-1, 1, size=width*height), label
return reader
TODO(yuyang18): Should we add whole design doc here?
"""
import decorator
from decorator import *
import creator
__all__ = decorator.__all__ + ['creator']
python/paddle/v2/reader/creator.py
浏览文件 @ 047f3a76
......@@ -11,6 +11,10 @@
# 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.
"""
Creator package contains some simple reader creator, which could be used in user
program.
"""
__all__ = ['np_array', 'text_file']
......@@ -38,7 +42,7 @@ def np_array(x):
def text_file(path):
"""
Creates a data reader that outputs text line by line from given text file.
Trailing new line ('\n') of each line will be removed.
Trailing new line ('\\\\n') of each line will be removed.
:path: path of the text file.
:returns: data reader of text file
......
python/paddle/v2/reader/decorator.py
浏览文件 @ 047f3a76
......@@ -28,9 +28,11 @@ def map_readers(func, *readers):
Creates a data reader that outputs return value of function using
output of each data readers as arguments.
:param func: function to use.
:param *readers: readers whose outputs will be used as arguments of func.
:returns: the created data reader.
:param func: function to use. The type of func should be (Sample) => Sample
:type: callable
:param readers: readers whose outputs will be used as arguments of func.
:return: the created data reader.
:rtype: callable
"""
def reader():
......@@ -45,16 +47,19 @@ def map_readers(func, *readers):
def shuffle(reader, buf_size):
"""
Creates a data reader whose data output is suffled.
Creates a data reader whose data output is shuffled.
Output from the iterator that created by original reader will be
buffered into shuffle buffer, and then shuffled. The size of shuffle buffer
is determined by argument buf_size.
:param reader: the original reader whose output will be shuffled.
:type reader: callable
:param buf_size: shuffle buffer size.
:type buf_size: int
:returns:the new reader whose output is shuffled.
:return: the new reader whose output is shuffled.
:rtype: callable
"""
def data_reader():
......@@ -88,7 +93,8 @@ def chain(*readers):
[0, 0, 0, 1, 1, 1, 2, 2, 2]
:param readers: input readers.
:returns: the new data reader.
:return: the new data reader.
:rtype: callable
"""
def reader():
......@@ -115,12 +121,13 @@ def compose(*readers, **kwargs):
The composed reader will output:
(1, 2, 3, 4, 5)
:*readers: readers that will be composed together.
:check_alignment: if True, will check if input readers are aligned
:param readers: readers that will be composed together.
:param check_alignment: if True, will check if input readers are aligned
correctly. If False, will not check alignment and trailing outputs
will be discarded. Defaults to True.
:type check_alignment: bool
:returns: the new data reader.
:return: the new data reader.
:raises ComposeNotAligned: outputs of readers are not aligned.
Will not raise when check_alignment is set to False.
......@@ -161,7 +168,9 @@ def buffered(reader, size):
as the buffer is not empty.
:param reader: the data reader to read from.
:type reader: callable
:param size: max buffer size.
:type size: int
:returns: the buffered data reader.
"""
......@@ -196,6 +205,13 @@ def buffered(reader, size):
def firstn(reader, n):
"""
Limit the max number of samples that reader could return.
:param reader: the data reader to read from.
:type reader: callable
:param n: the max number of samples that return.
:type n: int
:return: the decorated reader.
:rtype: callable
"""
# TODO(yuyang18): Check if just drop the reader, could clean the opened
......
python/paddle/v2/tests/run_tests.sh
浏览文件 @ 047f3a76
......@@ -22,7 +22,7 @@ cd $SCRIPTPATH
$1 -m pip install ../../../../paddle/dist/*.whl
test_list="test_data_feeder.py"
test_list="test_data_feeder.py test_parameters.py"
export PYTHONPATH=$PWD/../../../../python/
......
python/paddle/v2/tests/test_parameters.py 0 → 100644
浏览文件 @ 047f3a76
import unittest
import sys
try:
import py_paddle
del py_paddle
except ImportError:
print >> sys.stderr, "It seems swig of Paddle is not installed, this " \
"unittest will not be run."
sys.exit(0)
import paddle.v2.parameters as parameters
from paddle.proto.ParameterConfig_pb2 import ParameterConfig
import random
import cStringIO
import numpy
def __rand_param_config__(name):
conf = ParameterConfig()
conf.name = name
size = 1
for i in xrange(2):
dim = random.randint(1, 1000)
conf.dims.append(dim)
size *= dim
conf.size = size
assert conf.IsInitialized()
return conf
class TestParameters(unittest.TestCase):
def test_serialization(self):
params = parameters.Parameters()
params.__append_config__(__rand_param_config__("param_0"))
params.__append_config__(__rand_param_config__("param_1"))
for name in params.names():
param = params.get(name)
param[:] = numpy.random.uniform(
-1.0, 1.0, size=params.get_shape(name))
params.set(name, param)
tmp_file = cStringIO.StringIO()
params.to_tar(tmp_file)
tmp_file.seek(0)
params_dup = parameters.Parameters.from_tar(tmp_file)
self.assertEqual(params_dup.names(), params.names())
for name in params.names():
self.assertEqual(params.get_shape(name), params_dup.get_shape(name))
p0 = params.get(name)
p1 = params_dup.get(name)
self.assertTrue(numpy.isclose(p0, p1).all())
if __name__ == '__main__':
unittest.main()
python/paddle/v2/tests/test_topology.py
浏览文件 @ 047f3a76
......@@ -16,6 +16,7 @@ import paddle.v2.layer as layer
import paddle.v2.topology as topology
import paddle.v2.data_type as data_type
import paddle.trainer_config_helpers as conf_helps
import paddle.trainer.PyDataProvider2 as pydp2
class TestTopology(unittest.TestCase):
......@@ -35,13 +36,13 @@ class TestTopology(unittest.TestCase):
pixel_data_type = filter(lambda type: type[0] == "pixel", data_types)
self.assertEqual(len(pixel_data_type), 1)
pixel_data_type = pixel_data_type[0]
self.assertEqual(pixel_data_type[1].type, data_type.DataType.Dense)
self.assertEqual(pixel_data_type[1].type, pydp2.DataType.Dense)
self.assertEqual(pixel_data_type[1].dim, 784)
label_data_type = filter(lambda type: type[0] == "label", data_types)
self.assertEqual(len(label_data_type), 1)
label_data_type = label_data_type[0]
self.assertEqual(label_data_type[1].type, data_type.DataType.Index)
self.assertEqual(label_data_type[1].type, pydp2.DataType.Index)
self.assertEqual(label_data_type[1].dim, 10)
def test_get_layer(self):
......
python/paddle/v2/trainer.py
浏览文件 @ 047f3a76
......@@ -9,6 +9,10 @@ from . import optimizer as v2_optimizer
from . import parameters as v2_parameters
__all__ = ['SGD']
"""
Trainer package
TODO(yuyang18): Complete comments.
"""
def default_event_handler(event):
......@@ -22,14 +26,20 @@ def default_event_handler(event):
pass
class SGD():
def __init__(self, cost, parameters, update_equation):
"""
Simple SGD Trainer.
class SGD(object):
"""
Simple SGD Trainer.
TODO(yuyang18): Complete comments
:param update_equation: The optimizer object.
:type update_equation: paddle.v2.optimizer.Optimizer
:param cost: Target cost that neural network should be optimized.
:type cost: paddle.v2.config_base.Layer
:param parameters: The parameters dictionary.
:type parameters: paddle.v2.parameters.Parameters
"""
:param update_equation: The optimizer object.
:type update_equation: v2_optimizer.Optimizer
"""
def __init__(self, cost, parameters, update_equation):
if not isinstance(parameters, v2_parameters.Parameters):
raise TypeError('parameters should be parameters')
......@@ -47,29 +57,26 @@ class SGD():
self.__topology_in_proto__, api.CREATE_MODE_NORMAL,
self.__optimizer__.enable_types())
assert isinstance(gm, api.GradientMachine)
parameters.append_gradient_machine(gm)
self.__gradient_machine__ = gm
self.__gradient_machine__.randParameters()
parameters.append_gradient_machine(gm)
def train(self, reader, num_passes=1, event_handler=None, reader_dict=None):
def train(self, reader, num_passes=1, event_handler=None, feeding=None):
"""
Training method. Will train num_passes of input data.
:param reader:
:param topology: Network Topology, use one or more Layers to represent it.
:param parameters: The parameter pools.
:param num_passes: The total train passes.
:param event_handler: Event handler. A method will be invoked when event
occurred.
:type event_handler: (BaseEvent) => None
:param feeding: Feeding is a map of neural network input name and array
index that reader returns.
:type feeding: dict
:return:
"""
if event_handler is None:
event_handler = default_event_handler
if reader_dict is None:
reader_dict = self.default_reader_dict()
__check_train_args__(**locals())
updater = self.__optimizer__.create_local_updater()
......@@ -81,9 +88,7 @@ class SGD():
pass_evaluator = self.__gradient_machine__.makeEvaluator()
assert isinstance(pass_evaluator, api.Evaluator)
out_args = api.Arguments.createArguments(0)
feeder = DataFeeder(self.__data_types__, reader_dict)
feeder = DataFeeder(self.__data_types__, feeding)
for pass_id in xrange(num_passes):
event_handler(v2_event.BeginPass(pass_id))
pass_evaluator.start()
......@@ -101,7 +106,7 @@ class SGD():
for each_param in self.__gradient_machine__.getNonStaticParameters(
):
updater.update(each_param)
cost_sum = out_args.sumCosts()
cost_sum = out_args.sum()
cost = cost_sum / len(data_batch)
updater.finishBatch(cost)
batch_evaluator.finish()
......@@ -117,17 +122,8 @@ class SGD():
event_handler(v2_event.EndPass(pass_id, evaluator=pass_evaluator))
self.__gradient_machine__.finish()
def default_reader_dict(self):
reader_dict = dict()
for i, tp in enumerate(self.__data_types__):
reader_dict[tp[0]] = i
return reader_dict
def test(self, reader, reader_dict=None):
if reader_dict is None:
reader_dict = self.default_reader_dict()
feeder = DataFeeder(self.__data_types__, reader_dict)
def test(self, reader, feeding=None):
feeder = DataFeeder(self.__data_types__, feeding)
evaluator = self.__gradient_machine__.makeEvaluator()
out_args = api.Arguments.createArguments(0)
evaluator.start()
......@@ -137,7 +133,7 @@ class SGD():
num_samples += len(data_batch)
self.__gradient_machine__.forward(
feeder(data_batch), out_args, api.PASS_TEST)
total_cost += out_args.sumCosts()
total_cost += out_args.sum()
self.__gradient_machine__.eval(evaluator)
evaluator.finish()
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或