def train(epoches):
# 训练数据
train_reader = create_reader(True)
# 测试数据
test_reader = create_reader(False)
train_batch = paddle.batch(reader=paddle.reader.shuffle(train_reader,
buf_size=32),
batch_size=32)
test_batch = paddle.batch(reader=paddle.reader.shuffle(test_reader,
buf_size=32),
batch_size=32)
feeding = {'image': 0, 'target': 1, 'label': 2}
cost = train_caption_net(dict_dim=DICT_DIM,
target=target,
label=label,
input_images=image)
adam_optimizer = paddle.optimizer.Adam(
learning_rate=1e-4,
regularization=paddle.optimizer.L2Regularization(rate=8e-4),
model_average=paddle.optimizer.ModelAverage(average_window=0.5))
# 初始化参数，从预训练文件中加载ResNet模型参数
parameters = paddle.parameters.create(cost)
parameters.init_from_tar(gzip.open('params/Paddle_ResNet50.tar.gz'))
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 5 == 0:
line = "Pass %d, Batch %d, Cost %f, %s\n" % (
event.pass_id, event.batch_id, event.cost, event.metrics)
print(line)
if isinstance(event, paddle.event.EndPass):
with gzip.open('params/params_pass_%d.tar.gz' % event.pass_id,
'w') as f:
parameters.to_tar(f)
result = trainer.test(reader=test_batch)
line = "Test with Pass %d, %s" % (event.pass_id, result.metrics)
print(line)
trainer = paddle.trainer.SGD(cost=cost,
parameters=parameters,
update_equation=adam_optimizer)
trainer.train(reader=train_batch,
event_handler=event_handler,
feeding=feeding,
num_passes=epoches)
def train():
paddle.init(use_gpu=False, trainer_count=2)
image = paddle.layer.data(name='image',
type=paddle.data_type.dense_vector(IMG_HEIGHT *
IMG_WIDTH),
height=IMG_HEIGHT,
width=IMG_WIDTH)
output = model(image)
label = paddle.layer.data(
name='label', type=paddle.data_type.integer_value_sequence(NUM_CLASS))
loss = paddle.layer.ctc(input=output, label=label, size=NUM_CLASS + 1)
feeding = {'image': 0, 'label': 1}
parameters = paddle.parameters.create(loss)
optimizer = paddle.optimizer.RMSProp(
learning_rate=0.01,
regularization=paddle.optimizer.L2Regularization(rate=8e-4))
trainer = paddle.trainer.SGD(cost=loss,
parameters=parameters,
update_equation=optimizer)
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 5 == 0:
print("\npass %d, Batch: %d cost: %f" %
(event.pass_id + 1, event.batch_id + 1, event.cost))
else:
sys.stdout.write('*')
sys.stdout.flush()
if isinstance(event, paddle.event.EndPass):
# save parameters
with gzip.open('output/params_pass_%d.tar.gz' % event.pass_id,
'w') as f:
parameters.to_tar(f)
test_reader = data_reader.create_reader('test')
result = trainer.test(reader=paddle.batch(test_reader,
batch_size=32),
feeding=feeding)
print("\nTest with Pass %d, cost: %s" %
(event.pass_id + 1, result.cost))
train_reader = data_reader.create_reader('train')
reader = paddle.batch(reader=train_reader, batch_size=32)
trainer.train(reader=reader,
feeding=feeding,
num_passes=50,
event_handler=event_handler)
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 50 == 0:
print ("\npass %d, Batch: %d cost: %f metrics: %s" % (event.pass_id, event.batch_id, event.cost, event.metrics))
else:
sys.stdout.write('.')
sys.stdout.flush()
if isinstance(event, paddle.event.EndPass):
# save parameters
with gzip.open('output/params_pass_%d.tar.gz' % event.pass_id, 'w') as f:
parameters.to_tar(f)
test_reader = data_reader.create_reader('test')
result = trainer.test(
reader=paddle.batch(test_reader, batch_size=128),
feeding=feeding)
class_error_rate = result.metrics['classification_error_evaluator']
print ("\nTest with Pass %d, cost: %s ratio: %f" % (event.pass_id, result.cost,class_error_rate))
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 5 == 0:
print("\npass %d, Batch: %d cost: %f" %
(event.pass_id + 1, event.batch_id + 1, event.cost))
else:
sys.stdout.write('*')
sys.stdout.flush()
if isinstance(event, paddle.event.EndPass):
# save parameters
with gzip.open('output/params_pass_%d.tar.gz' % event.pass_id,
'w') as f:
parameters.to_tar(f)
test_reader = data_reader.create_reader('test')
result = trainer.test(reader=paddle.batch(test_reader, batch_size=32),
feeding=feeding)
print("\nTest with Pass %d, cost: %s" %
(event.pass_id + 1, result.cost))
output = model(image)
# CTC LOSS
loss = paddle.layer.ctc(input=output, label=label, size=NUM_CLASS + 1)
parameters = paddle.parameters.create(loss)
optimizer = paddle.optimizer.RMSProp(
learning_rate=0.01,
regularization=paddle.optimizer.L2Regularization(rate=8e-4))
trainer = paddle.trainer.SGD(cost=loss,
parameters=parameters,
update_equation=optimizer)
train_reader = data_reader.create_reader('train')
feeding = {'image': 0, 'label': 1}
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 5 == 0:
print("\npass %d, Batch: %d cost: %f" %
(event.pass_id + 1, event.batch_id + 1, event.cost))
else:
sys.stdout.write('*')
sys.stdout.flush()
if isinstance(event, paddle.event.EndPass):
# save parameters
with gzip.open('output/params_pass_%d.tar.gz' % event.pass_id,
def train():
paddle.init(use_gpu=False, trainer_count=1)
x = paddle.layer.data(
name='image',
type=paddle.data_type.dense_vector(IMAGE_SIZE),
height=32,
width=80
)
label = []
for i in range(4):
label_tmp = paddle.layer.data(
name='label_part_%s' % i,
type=paddle.data_type.integer_value(10)
)
label.append(label_tmp)
output = model(x)
loss = []
for i in range(4):
loss_tmp = paddle.layer.classification_cost(
input=output[i],
label=label[i]
)
loss.append(loss_tmp)
loss = paddle.layer.addto(
input=loss,
bias_attr=False,
act=paddle.activation.Linear()
)
parameters = paddle.parameters.create(loss)
optimizer = paddle.optimizer.Adam(
learning_rate=1e-3,
regularization=paddle.optimizer.L2Regularization(rate=8e-4)
)
trainer = paddle.trainer.SGD(
cost=loss,
parameters=parameters,
update_equation=optimizer
)
feeding = {
'image': 0,
'label_part_0': 1,
'label_part_1': 2,
'label_part_2': 3,
'label_part_3': 4,
}
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 50 == 0:
print ("\npass %d, Batch: %d cost: %f metrics: %s" % (event.pass_id, event.batch_id, event.cost, event.metrics))
else:
sys.stdout.write('.')
sys.stdout.flush()
if isinstance(event, paddle.event.EndPass):
# save parameters
with gzip.open('output/params_pass_%d.tar.gz' % event.pass_id, 'w') as f:
parameters.to_tar(f)
test_reader = data_reader.create_reader('test')
result = trainer.test(
reader=paddle.batch(test_reader, batch_size=128),
feeding=feeding)
class_error_rate = result.metrics['classification_error_evaluator']
print ("\nTest with Pass %d, cost: %s ratio: %f" % (event.pass_id, result.cost,class_error_rate))
train_reader = data_reader.create_reader('train')
reader = paddle.batch(
reader=train_reader,
batch_size=128
)
trainer.train(
reader=reader,
feeding=feeding,
num_passes=10,
event_handler=event_handler
)