def main(data_args, use_cuda, num_passes, lr):
i_shape = [3, 224, 224]
image = fluid.layers.data(name="image", shape=i_shape, dtype='float32')
Net = fvnet.Net()
out = Net.inference(input=image)
label = fluid.layers.data(name="label", shape=[2], dtype='float32')
sec = fluid.layers.square_error_cost(input=out, label=label)
avg_cost = fluid.layers.mean(x=sec)
optimizer = fluid.optimizer.Adam(learning_rate=lr) #Adam
optimizer.minimize(avg_cost)
fluid.memory_optimize(fluid.default_main_program())
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
layer_load = [
"conv1_1", "conv1_2", "conv2_1", "conv2_2", "conv3_1", "conv3_2",
"conv3_3", "conv4_1", "conv4_2", "conv4_3"
]
Net.load_weights(pretrained_npy, exe, place, layer_load)
# if pretrained_model:
# fluid.io.load_vars(exe, pretrained_model, predicate=if_exist)
model_path = "./pre_weight/"
fluid.io.save_inference_model(model_path, ['image'], [out], exe)
def infer(args):
# parameters from arguments
class_dim = args.class_dim
model_name = args.model
save_inference = args.save_inference
pretrained_model = args.pretrained_model
with_memory_optimization = args.with_mem_opt
image_shape = [int(m) for m in args.image_shape.split(",")]
model_list = [m for m in dir(models) if "__" not in m]
assert model_name in model_list, "{} is not in lists: {}".format(
args.model, model_list)
image = fluid.layers.data(name='image', shape=image_shape, dtype='float32')
# model definition
model = models.__dict__[model_name]()
if model_name == "GoogleNet":
out, _, _ = model.net(input=image, class_dim=class_dim)
else:
out = model.net(input=image, class_dim=class_dim)
out = fluid.layers.softmax(out)
test_program = fluid.default_main_program().clone(for_test=True)
fetch_list = [out.name]
if with_memory_optimization and not save_inference:
fluid.memory_optimize(fluid.default_main_program(),
skip_opt_set=set(fetch_list))
place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
fluid.io.load_persistables(exe, pretrained_model)
if save_inference:
fluid.io.save_inference_model(dirname=model_name,
feeded_var_names=['image'],
main_program=test_program,
target_vars=out,
executor=exe,
model_filename='model',
params_filename='params')
print("model: ", model_name, " is already saved")
exit(0)
test_batch_size = 1
test_reader = reader.test(settings=args, batch_size=test_batch_size)
feeder = fluid.DataFeeder(place=place, feed_list=[image])
TOPK = 1
for batch_id, data in enumerate(test_reader()):
result = exe.run(test_program,
fetch_list=fetch_list,
feed=feeder.feed(data))
result = result[0][0]
pred_label = np.argsort(result)[::-1][:TOPK]
print("Test-{0}-score: {1}, class {2}".format(batch_id,
result[pred_label],
pred_label))
sys.stdout.flush()
def check_decay_with_place(self, place, python_decay_fn, fluid_decay_fn,
kwargs):
main_prog = fluid.Program()
startup_prog = fluid.Program()
with fluid.program_guard(main_prog, startup_prog):
decayed_lr = fluid_decay_fn(**kwargs)
place = fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(startup_prog)
fluid.memory_optimize(main_prog)
for step in range(10):
lr_val, = exe.run(main_prog, feed={}, fetch_list=[decayed_lr])
python_decayed_lr = python_decay_fn(global_step=float(step),
**kwargs)
self.assertAlmostEqual(
python_decayed_lr,
lr_val[0],
msg='Failed fn is {0}, Python result is {1}, Fluid result is {2}'
.format(python_decay_fn.__name__, str(python_decayed_lr),
str(lr_val[0])))
def infer():
# parameters from arguments
use_gpu = False
class_dim = 5
model_name = "ResNet50"
pretrained_model = "./output_indoor/ResNet50/61"
with_memory_optimization = True
image_shape = [3, 224, 224]
# assert model_name in model_list, "{} is not in lists: {}".format(args.model,
# model_list)
image = fluid.layers.data(name='image', shape=image_shape, dtype='float32')
# model definition
model = mo.__dict__[model_name]()
if model_name is "GoogleNet":
out, _, _ = model.net(input=image, class_dim=class_dim)
else:
out = model.net(input=image, class_dim=class_dim)
test_program = fluid.default_main_program().clone(for_test=True)
if with_memory_optimization:
fluid.memory_optimize(fluid.default_main_program())
place = fluid.CUDAPlace(0) if use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
if pretrained_model:
def if_exist(var):
return os.path.exists(os.path.join(pretrained_model, var.name))
fluid.io.load_vars(exe, pretrained_model, predicate=if_exist)
test_batch_size = 1
test_reader = paddle.batch(reader.test(), batch_size=test_batch_size)
feeder = fluid.DataFeeder(place=place, feed_list=[image])
fetch_list = [out.name]
TOPK = 1
for batch_id, data in enumerate(test_reader()):
result = exe.run(test_program,
fetch_list=fetch_list,
feed=feeder.feed(data))
result = result[0][0]
pred_label = np.argsort(result)[::-1][:TOPK]
#print("Test-{0}-score: {1}, class {2}"
# .format(batch_id, result[pred_label], pred_label))
result = pred_label
sys.stdout.flush()
return result
def main():
args = parse_args()
print_arguments(args)
print_paddle_envs()
if args.no_random:
fluid.default_startup_program().random_seed = 1
# the unique trainer id, starting from 0, needed by trainer
# only
nccl_id_var, num_trainers, trainer_id = (None, 1,
int(
os.getenv(
"PADDLE_TRAINER_ID",
"0")))
if args.use_cprof:
pr = cProfile.Profile()
pr.enable()
model_def = __import__("%s" % args.model, fromlist=["models"])
train_args = list(model_def.get_model(args))
train_args.append(args)
# Run optimizer.minimize(avg_loss)
train_args[2].minimize(train_args[0])
if args.memory_optimize:
fluid.memory_optimize(fluid.default_main_program())
if args.update_method == "pserver":
train_prog, startup_prog = dist_transpile(trainer_id, args)
if not train_prog:
raise Exception(
"Must configure correct environments to run dist train.")
train_args.extend([train_prog, startup_prog])
if args.gpus > 1 and os.getenv("TRAINING_ROLE") == "TRAINER":
train_args.extend([nccl_id_var, num_trainers, trainer_id])
train_parallel(*train_args)
exit(0)
train(*train_args)
exit(0)
# for other update methods, use default programs
train_args.append(fluid.default_main_program())
train_args.append(fluid.default_startup_program())
if args.update_method == "nccl2":
nccl_id_var, num_trainers, trainer_id = append_nccl2_prepare(
trainer_id)
if args.gpus == 1:
# NOTE: parallel executor use profiler interanlly
if args.use_nvprof and args.device == 'GPU':
with profiler.cuda_profiler("cuda_profiler.txt", 'csv') as nvprof:
train(*train_args)
else:
train(*train_args)
else:
if args.device == "CPU":
raise Exception("Only support GPU perf with parallel exe")
train_args.extend([nccl_id_var, num_trainers, trainer_id])
train_parallel(*train_args)
def main(data_args,use_cuda,num_passes,lr,json_path1,json_path2):
i_shape=[3,224, 224]
image = fluid.layers.data(name="image", shape=i_shape, dtype='float32')
Net = fvnet.vgg_fluid()
out = Net.net(input=image)
label = fluid.layers.data(name="label", shape=[28*28], dtype='float32')#28*28 2
sec = fluid.layers.square_error_cost(input=out, label=label)
avg_cost = fluid.layers.mean(x=sec)
optimizer = fluid.optimizer.Adam(learning_rate=lr)#Adam
optimizer.minimize(avg_cost)
fluid.memory_optimize(fluid.default_main_program())
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
if pretrained_model:
fluid.io.load_vars(exe, pretrained_model, predicate=if_exist)
panda_cl = provider.Pandas_fm(json_path1,json_path2)
train_reader = paddle.batch(provider.read_train_for(data_args,panda_cl), batch_size=1)
Loss = []
ttime = 0
for pass_id in range(num_passes):
Loss_A = []
for batch_id , data in enumerate(train_reader()):
start = datetime.datetime.now()
i_datas = np.array(data[0][0])
i_labels = np.array(data[0][1])
loss = exe.run(fluid.default_main_program(),
feed={"image":i_datas,"label":i_labels},fetch_list = [avg_cost])
loss = np.mean(np.array(loss))
Loss_A.append(loss)
end = datetime.datetime.now()
ttime += (end-start).total_seconds()
if batch_id % 100 == 0:
print("Pass {0}, trainbatch {1}, loss {2}, time {3}".format(pass_id, \
batch_id, loss,ttime))
sys.stdout.flush()
ttime = 0
Loss.append(np.mean(Loss_A))
np.save('./models/loss/'+str(pass_id)+'_loss.npy',np.array(Loss))
model_path = os.path.join("./models/",str(pass_id))
if not os.path.isdir(model_path):
os.makedirs(model_path)
fluid.io.save_inference_model(model_path, ['image'], [out], exe)
np.save('./models/loss.npy',np.array(Loss))
def train(config):
""" model training """
config.vocab_size = len(open(config.vocab_path).readlines())
bow_loss, kl_loss, nll_loss, final_loss = knowledge_seq2seq(config)
bow_loss.persistable = True
kl_loss.persistable = True
nll_loss.persistable = True
final_loss.persistable = True
main_program = fluid.default_main_program()
inference_program = fluid.default_main_program().clone(for_test=True)
fluid.clip.set_gradient_clip(clip=fluid.clip.GradientClipByGlobalNorm(
clip_norm=config.grad_clip))
optimizer = fluid.optimizer.Adam(learning_rate=config.lr)
if config.stage == 0:
print("stage 0")
optimizer.minimize(bow_loss)
else:
print("stage 1")
optimizer.minimize(final_loss)
fluid.memory_optimize(main_program)
opt_var_name_list = optimizer.get_opti_var_name_list()
if config.use_gpu:
place = fluid.CUDAPlace(0)
else:
place = fluid.CPUPlace()
exe = Executor(place)
exe.run(framework.default_startup_program())
param_list = main_program.block(0).all_parameters()
param_name_list = [p.name for p in param_list]
init_model(config, param_name_list, place)
processors = KnowledgeCorpus(data_dir=config.data_dir,
data_prefix=config.data_prefix,
vocab_path=config.vocab_path,
min_len=config.min_len,
max_len=config.max_len)
train_generator = processors.data_generator(batch_size=config.batch_size,
phase="train",
shuffle=True)
valid_generator = processors.data_generator(batch_size=config.batch_size,
phase="dev",
shuffle=False)
model_handle = [exe, place, bow_loss, kl_loss, nll_loss, final_loss]
train_loop(config, train_generator, valid_generator, main_program,
inference_program, model_handle, param_name_list,
opt_var_name_list)
def eval(args):
# parameters from arguments
model_name = args.model
pretrained_model = args.pretrained_model
with_memory_optimization = args.with_mem_opt
image_shape = [int(m) for m in args.image_shape.split(",")]
assert model_name in model_list, "{} is not in lists: {}".format(args.model,
model_list)
image = fluid.layers.data(name='image', shape=image_shape, dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
# model definition
model = models.__dict__[model_name]()
out = model.net(input=image, embedding_size=args.embedding_size)
test_program = fluid.default_main_program().clone(for_test=True)
if with_memory_optimization:
fluid.memory_optimize(fluid.default_main_program())
place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
if pretrained_model:
def if_exist(var):
return os.path.exists(os.path.join(pretrained_model, var.name))
fluid.io.load_vars(exe, pretrained_model, predicate=if_exist)
test_reader = paddle.batch(reader.test(args), batch_size=args.batch_size, drop_last=False)
feeder = fluid.DataFeeder(place=place, feed_list=[image, label])
fetch_list = [out.name]
f, l = [], []
for batch_id, data in enumerate(test_reader()):
t1 = time.time()
[feas] = exe.run(test_program, fetch_list=fetch_list, feed=feeder.feed(data))
label = np.asarray([x[1] for x in data])
f.append(feas)
l.append(label)
t2 = time.time()
period = t2 - t1
if batch_id % 20 == 0:
print("[%s] testbatch %d, time %2.2f sec" % \
(fmt_time(), batch_id, period))
f = np.vstack(f)
l = np.hstack(l)
recall = recall_topk(f, l, k=1)
print("[%s] End test %d, test_recall %.5f" % (fmt_time(), len(f), recall))
sys.stdout.flush()
def infer(args):
# parameters from arguments
seg_num = args.seg_num
class_dim = args.class_dim
num_layers = args.num_layers
test_model = args.test_model
if test_model == None:
print('Please specify the test model ...')
return
image_shape = [int(m) for m in args.image_shape.split(",")]
image_shape = [seg_num] + image_shape
# model definition
model = TSN_ResNet(layers=num_layers, seg_num=seg_num)
image = fluid.layers.data(name='image', shape=image_shape, dtype='float32')
out = model.net(input=image, class_dim=class_dim)
# for test
inference_program = fluid.default_main_program().clone(for_test=True)
if args.with_mem_opt:
fluid.memory_optimize(fluid.default_main_program())
place = fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
def is_parameter(var):
if isinstance(var, Parameter):
return isinstance(var, Parameter)
if test_model is not None:
vars = filter(is_parameter, inference_program.list_vars())
fluid.io.load_vars(exe, test_model, vars=vars)
# reader
test_reader = paddle.batch(reader.infer(seg_num), batch_size=1)
feeder = fluid.DataFeeder(place=place, feed_list=[image])
fetch_list = [out.name]
# test
TOPK = 1
for batch_id, data in enumerate(test_reader()):
data, vid = data[0]
data = [[data]]
result = exe.run(inference_program,
fetch_list=fetch_list,
feed=feeder.feed(data))
result = result[0][0]
pred_label = np.argsort(result)[::-1][:TOPK]
print("Test sample: {0}, score: {1}, class {2}".format(
vid, result[pred_label], pred_label))
sys.stdout.flush()
def infer(model):
predict = create_model(model=model)
place = fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
fluid.memory_optimize(fluid.default_main_program())
load_model(exe,fluid.default_main_program(),model=model)
print("load model succeed")
while True:
ret,frame = cap.read()
cv2.imshow('frame',frame)#一个窗口用以显示原视频
gray = cv2.cvtColor(frame,cv2.COLOR_BGR2GRAY)
# 探测图片中的人脸
faces = face_cascade.detectMultiScale(
gray,
scaleFactor = 1.5,
minNeighbors = 5,
minSize = (5,5)
)
for (x, y, w, h) in faces:
w = w*1.1
h = h*1.3
w = int(w)
h = int(h)
frame = cv2.rectangle(frame, (x, y), (x+w, y+h), (255, 0, 0), 2)
xmin = x
xmax = x + w
ymin = y
ymax = y + h
slip_image = frame[ymin:ymax,xmin:xmax]
slip_image_224 = cv2.resize(slip_image, (224,224), interpolation=cv2.INTER_CUBIC)
slip_image_224 = slip_image_224.transpose((2,0,1))
imgs = []
imgs.append(slip_image_224)
imgs = np.array(imgs)
imgs = imgs.astype(np.float32)
imgs /= 255.0
result = exe.run(fluid.default_main_program(),
feed={'img': imgs},
fetch_list=[predict])
points = result[0]
points = points.reshape(-1,2)
draw_landmark_point(slip_image,points)
cv2.imshow("image!",slip_image)
cv2.imshow('origin image',frame)#一个窗口用以显示原视频
if cv2.waitKey(1) &0xFF == ord('q'):
break
cap.release()
cv2.destroyAllWindows()
def add_optimizer(args, avg_cost):
#optimizer = fluid.optimizer.SGD(learning_rate=0.002)
optimizer = fluid.optimizer.Momentum(
learning_rate=fluid.layers.piecewise_decay(
boundaries=[100], values=[0.1, 0.2]),
momentum=0.9,
regularization=fluid.regularizer.L2Decay(1e-4))
optimizer.minimize(avg_cost)
if args.use_mem_opt:
fluid.memory_optimize(fluid.default_main_program())
def main():
rnn_out = encoder_decoder()
label = layers.data(
name="target_language_next_word", shape=[1], dtype='int64', lod_level=1)
cost = layers.cross_entropy(input=rnn_out, label=label)
avg_cost = fluid.layers.mean(cost)
optimizer = fluid.optimizer.Adagrad(learning_rate=1e-4)
optimizer.minimize(avg_cost)
fluid.memory_optimize(fluid.default_main_program())
# fluid.release_memory(fluid.default_main_program())
# fix the order of training data
train_data = paddle.batch(
paddle.dataset.wmt14.train(dict_size), batch_size=batch_size)
# train_data = paddle.batch(
# paddle.reader.shuffle(
# paddle.dataset.wmt14.train(dict_size), buf_size=1000),
# batch_size=batch_size)
place = core.CPUPlace()
exe = Executor(place)
exe.run(framework.default_startup_program())
feed_order = [
'src_word_id', 'target_language_word', 'target_language_next_word'
]
feed_list = [
fluid.default_main_program().global_block().var(var_name)
for var_name in feed_order
]
feeder = fluid.DataFeeder(feed_list, place)
batch_id = 0
for pass_id in range(10):
for data in train_data():
outs = exe.run(fluid.default_main_program(),
feed=feeder.feed(data),
fetch_list=[avg_cost])
avg_cost_val = np.array(outs[0])
print('pass_id=' + str(pass_id) + ' batch=' + str(batch_id) +
" avg_cost=" + str(avg_cost_val))
if batch_id > 2:
exit(0)
if math.isnan(float(avg_cost_val)):
sys.exit("got NaN loss, training failed.")
batch_id += 1
def main():
args = parse_args()
print_arguments(args)
# the unique trainer id, starting from 0, needed by trainer
# only
nccl_id_var, num_trainers, trainer_id = (
None, 1, int(os.getenv("PADDLE_TRAINER_ID", "-1")))
if args.use_cprof:
pr = cProfile.Profile()
pr.enable()
model_def = __import__("models.%s" % args.model, fromlist=["models"])
train_args = list(model_def.get_model(args))
train_args.append(args)
# Run optimizer.minimize(avg_loss)
train_args[2].minimize(train_args[0])
if args.memory_optimize:
fluid.memory_optimize(fluid.default_main_program())
if args.update_method == "pserver":
train_prog, startup_prog = dist_transpile(trainer_id)
if not train_prog:
raise Exception(
"Must configure correct environments to run dist train.")
train_args.extend([train_prog, startup_prog])
if args.gpus > 1 and os.getenv("PADDLE_TRAINING_ROLE") == "TRAINER":
train_args.extend([nccl_id_var, num_trainers, trainer_id])
train_parallel(*train_args)
train(*train_args)
exit(0)
# for other update methods, use default programs
train_args.append(fluid.default_main_program())
train_args.append(fluid.default_startup_program())
if args.update_method == "nccl2":
nccl_id_var, num_trainers, trainer_id = append_nccl2_prepare(trainer_id)
if args.gpus == 1:
# NOTE: parallel executor use profiler interanlly
if args.use_nvprof and args.device == 'GPU':
with profiler.cuda_profiler("cuda_profiler.txt", 'csv') as nvprof:
train(*train_args)
else:
train(*train_args)
else:
if args.device == "CPU":
raise Exception("Only support GPU perf with parallel exe")
train_args.extend([nccl_id_var, num_trainers, trainer_id])
train_parallel(*train_args)
def get_model(args, is_train, main_prog, startup_prog):
# NOTE: mnist is small, we don't implement data sharding yet.
opt = None
data_file_handle = None
with fluid.program_guard(main_prog, startup_prog):
if args.use_reader_op:
filelist = [
os.path.join(args.data_path, f)
for f in os.listdir(args.data_path)
]
data_file_handle = fluid.layers.open_files(
filenames=filelist,
shapes=[[-1, 1, 28, 28], (-1, 1)],
lod_levels=[0, 0],
dtypes=["float32", "int64"],
thread_num=1,
pass_num=1)
data_file = fluid.layers.double_buffer(
fluid.layers.batch(
data_file_handle, batch_size=args.batch_size))
with fluid.unique_name.guard():
if args.use_reader_op:
input, label = fluid.layers.read_file(data_file)
else:
images = fluid.layers.data(
name='pixel', shape=[1, 28, 28], dtype='float32')
label = fluid.layers.data(
name='label', shape=[1], dtype='int64')
predict = cnn_model(images)
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(x=cost)
# Evaluator
batch_acc = fluid.layers.accuracy(input=predict, label=label)
# Optimization
if is_train:
opt = fluid.optimizer.AdamOptimizer(
learning_rate=0.001, beta1=0.9, beta2=0.999)
opt.minimize(avg_cost)
if args.memory_optimize:
fluid.memory_optimize(main_prog)
# Reader
if is_train:
reader = paddle.dataset.mnist.train()
else:
reader = paddle.dataset.mnist.test()
batched_reader = paddle.batch(
reader, batch_size=args.batch_size * args.gpus)
return avg_cost, opt, [batch_acc], batched_reader, data_file_handle
def infer(args):
# parameters from arguments
model_name = args.model
pretrained_model = args.pretrained_model
with_memory_optimization = args.with_mem_opt
image_shape = [int(m) for m in args.image_shape.split(",")]
assert model_name in model_list, "{} is not in lists: {}".format(
args.model, model_list)
image = fluid.layers.data(name='image', shape=image_shape, dtype='float32')
# model definition
model = models.__dict__[model_name]()
out = model.net(input=image, embedding_size=args.embedding_size)
test_program = fluid.default_main_program().clone(for_test=True)
if with_memory_optimization:
fluid.memory_optimize(fluid.default_main_program())
place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
if pretrained_model:
def if_exist(var):
return os.path.exists(os.path.join(pretrained_model, var.name))
fluid.io.load_vars(exe, pretrained_model, predicate=if_exist)
infer_reader = paddle.batch(reader.infer(args),
batch_size=args.batch_size,
drop_last=False)
feeder = fluid.DataFeeder(place=place, feed_list=[image])
fetch_list = [out.name]
for batch_id, data in enumerate(infer_reader()):
result = exe.run(test_program,
fetch_list=fetch_list,
feed=feeder.feed(data))
result = result[0][0].reshape(-1)
print("Test-{0}-feature: {1}".format(batch_id, result[:5]))
sys.stdout.flush()
def construct_resnet(self, depth, learning_rate, momentum):
input = fluid.layers.data(
name='data', shape=self.data_shape, dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
predict = self.model_dict[self.data_set](input, self.class_dim, depth)
cost = fluid.layers.cross_entropy(input=predict, label=label)
self.avg_cost = fluid.layers.mean(x=cost)
self.accuracy = fluid.layers.accuracy(input=predict, label=label)
# inference program
self.inference_program = fluid.default_main_program().clone(
for_test=True)
optimizer = fluid.optimizer.Momentum(
learning_rate=learning_rate, momentum=momentum)
opts = optimizer.minimize(self.avg_cost)
fluid.memory_optimize(fluid.default_main_program())
self.train_program = fluid.default_main_program().clone()
def test_main(use_cuda, use_py_func_op, use_parallel_executor):
if use_cuda and not fluid.core.is_compiled_with_cuda():
return None
with fluid.program_guard(fluid.Program(), fluid.Program()):
with fluid.scope_guard(fluid.core.Scope()):
fluid.default_main_program().random_seed = 1
fluid.default_startup_program().random_seed = 1
np.random.seed(1)
img = fluid.layers.data(name='image', shape=[784], dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
loss = simple_fc_net(img, label, use_py_func_op)
optimizer = fluid.optimizer.SGD(learning_rate=1e-3)
optimizer.minimize(loss)
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
feeder = fluid.DataFeeder(feed_list=[img, label], place=place)
r = paddle.batch(reader, batch_size=10)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
#FIXME force use old memory optimzie strategy here to pass the unittest
#since open the new strategy will crash the unittest
fluid.memory_optimize(fluid.default_main_program())
train_cp = compiler.CompiledProgram(fluid.default_main_program())
if use_parallel_executor:
train_cp = train_cp.with_data_parallel(loss_name=loss.name)
fetch_list = [loss.name]
else:
fetch_list = [loss]
ret = []
for epoch_id in six.moves.range(2):
for d in r():
L, = exe.run(train_cp,
feed=feeder.feed(d),
fetch_list=fetch_list)
ret.append(L)
return np.array(ret)
def construct_vgg16_net(self, learning_rate):
images = fluid.layers.data(name='pixel',
shape=self.data_shape,
dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
net = self.vgg16_bn_drop(images)
predict = fluid.layers.fc(input=net,
size=self.class_dim,
act='softmax')
cost = fluid.layers.cross_entropy(input=predict, label=label)
self.avg_cost = fluid.layers.mean(x=cost)
self.accuracy = fluid.layers.accuracy(input=predict, label=label)
self.inference_program = fluid.default_main_program().clone(
for_test=True)
optimizer = fluid.optimizer.Adam(learning_rate=learning_rate)
opts = optimizer.minimize(self.avg_cost)
fluid.memory_optimize(fluid.default_main_program())
self.train_program = fluid.default_main_program().clone()
# lr = fluid.layers.polynomial_decay(base_lr, total_step, end_learning_rate=0, power=0.9)
# area = fluid.layers.elementwise_max(
# fluid.layers.reduce_mean(mask),
# fluid.layers.assign(np.array(
# [0.1], dtype=np.float32)))
# loss_mean = fluid.layers.reduce_mean(loss) / area
# opt = fluid.optimizer.Adam(learning_rate=2e-4)
# opt = fluid.optimizer.Momentum(
# lr,
# momentum=0.9,
# regularization=fluid.regularizer.L2DecayRegularizer(
# regularization_coeff=weight_decay), )
# retv = opt.minimize(loss_mean, startup_program=sp, no_grad_set=no_grad_set)
# tp = tp.clone(True)
fluid.memory_optimize(tp, print_log=False, skip_opt_set=[pred.name], level=1)
place = fluid.CPUPlace()
if use_gpu:
place = fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(sp)
load_model(recover_path)
if parallel_flag:
exe_p = fluid.ParallelExecutor(use_cuda=True,
loss_name=loss_mean.name,
main_program=tp)
early_stopcount = 0
def train(args):
"""Train
"""
if not os.path.exists(args.save_path):
os.makedirs(args.save_path)
net = Network(args.vocab_size, args.emb_size, args.hidden_size)
train_program = fluid.Program()
train_startup = fluid.Program()
if "CE_MODE_X" in os.environ:
train_program.random_seed = 110
train_startup.random_seed = 110
with fluid.program_guard(train_program, train_startup):
with fluid.unique_name.guard():
logits, loss = net.network(args.loss_type)
loss.persistable = True
logits.persistable = True
# gradient clipping
fluid.clip.set_gradient_clip(
clip=fluid.clip.GradientClipByValue(max=1.0, min=-1.0))
optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate)
optimizer.minimize(loss)
print("begin memory optimization ...")
fluid.memory_optimize(train_program)
print("end memory optimization ...")
test_program = fluid.Program()
test_startup = fluid.Program()
if "CE_MODE_X" in os.environ:
test_program.random_seed = 110
test_startup.random_seed = 110
with fluid.program_guard(test_program, test_startup):
with fluid.unique_name.guard():
logits, loss = net.network(args.loss_type)
loss.persistable = True
logits.persistable = True
test_program = test_program.clone(for_test=True)
if args.use_cuda:
place = fluid.CUDAPlace(0)
dev_count = fluid.core.get_cuda_device_count()
else:
place = fluid.CPUPlace()
dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
print("device count %d" % dev_count)
print("theoretical memory usage: ")
print(
fluid.contrib.memory_usage(program=train_program,
batch_size=args.batch_size))
exe = fluid.Executor(place)
exe.run(train_startup)
exe.run(test_startup)
train_exe = fluid.ParallelExecutor(use_cuda=args.use_cuda,
loss_name=loss.name,
main_program=train_program)
test_exe = fluid.ParallelExecutor(use_cuda=args.use_cuda,
main_program=test_program,
share_vars_from=train_exe)
if args.word_emb_init is not None:
print("start loading word embedding init ...")
if six.PY2:
word_emb = np.array(pickle.load(open(args.word_emb_init,
'rb'))).astype('float32')
else:
word_emb = np.array(
pickle.load(open(args.word_emb_init, 'rb'),
encoding="bytes")).astype('float32')
net.set_word_embedding(word_emb, place)
print("finish init word embedding ...")
print("start loading data ...")
def train_with_feed(batch_data):
"""
Train on one batch
"""
#to do get_feed_names
feed_dict = dict(zip(net.get_feed_names(), batch_data))
cost = train_exe.run(feed=feed_dict, fetch_list=[loss.name])
return cost[0]
def test_with_feed(batch_data):
"""
Test on one batch
"""
feed_dict = dict(zip(net.get_feed_names(), batch_data))
score = test_exe.run(feed=feed_dict, fetch_list=[logits.name])
return score[0]
def evaluate():
"""
Evaluate to choose model
"""
val_batches = reader.batch_reader(args.val_path, args.batch_size,
place, args.max_len, 1)
scores = []
labels = []
for batch in val_batches:
scores.extend(test_with_feed(batch))
labels.extend([x[0] for x in batch[2]])
return eva.evaluate_Recall(zip(scores, labels))
def save_exe(step, best_recall):
"""
Save exe conditional
"""
recall_dict = evaluate()
print('evaluation recall result:')
print('1_in_2: %s\t1_in_10: %s\t2_in_10: %s\t5_in_10: %s' %
(recall_dict['1_in_2'], recall_dict['1_in_10'],
recall_dict['2_in_10'], recall_dict['5_in_10']))
if recall_dict['1_in_10'] > best_recall and step != 0:
fluid.io.save_inference_model(args.save_path,
net.get_feed_inference_names(),
logits,
exe,
main_program=train_program)
print("Save model at step %d ... " % step)
print(
time.strftime('%Y-%m-%d %H:%M:%S',
time.localtime(time.time())))
best_recall = recall_dict['1_in_10']
return best_recall
# train over different epoches
global_step, train_time = 0, 0.0
best_recall = 0
for epoch in six.moves.xrange(args.num_scan_data):
train_batches = reader.batch_reader(args.train_path, args.batch_size,
place, args.max_len,
args.sample_pro)
begin_time = time.time()
sum_cost = 0
ce_cost = 0
for batch in train_batches:
if (args.save_path is not None) and (global_step % args.save_step
== 0):
best_recall = save_exe(global_step, best_recall)
cost = train_with_feed(batch)
global_step += 1
sum_cost += cost.mean()
ce_cost = cost.mean()
if global_step % args.print_step == 0:
print('training step %s avg loss %s' %
(global_step, sum_cost / args.print_step))
sum_cost = 0
pass_time_cost = time.time() - begin_time
train_time += pass_time_cost
print("Pass {0}, pass_time_cost {1}".format(
epoch, "%2.2f sec" % pass_time_cost))
if "CE_MODE_X" in os.environ and epoch == args.num_scan_data - 1:
card_num = get_cards()
print("kpis\ttrain_duration_card%s\t%s" %
(card_num, pass_time_cost))
print("kpis\ttrain_loss_card%s\t%s" % (card_num, ce_cost))
def finetune(args):
"""
Finetune
"""
if not os.path.exists(args.save_path):
os.makedirs(args.save_path)
net = Network(args.vocab_size, args.emb_size, args.hidden_size)
train_program = fluid.Program()
train_startup = fluid.Program()
if "CE_MODE_X" in os.environ:
train_program.random_seed = 110
train_startup.random_seed = 110
with fluid.program_guard(train_program, train_startup):
with fluid.unique_name.guard():
logits, loss = net.network(args.loss_type)
loss.persistable = True
logits.persistable = True
# gradient clipping
fluid.clip.set_gradient_clip(
clip=fluid.clip.GradientClipByValue(max=1.0, min=-1.0))
optimizer = fluid.optimizer.Adam(
learning_rate=fluid.layers.exponential_decay(
learning_rate=args.learning_rate,
decay_steps=400,
decay_rate=0.9,
staircase=True))
optimizer.minimize(loss)
print("begin memory optimization ...")
fluid.memory_optimize(train_program)
print("end memory optimization ...")
test_program = fluid.Program()
test_startup = fluid.Program()
if "CE_MODE_X" in os.environ:
test_program.random_seed = 110
test_startup.random_seed = 110
with fluid.program_guard(test_program, test_startup):
with fluid.unique_name.guard():
logits, loss = net.network(args.loss_type)
loss.persistable = True
logits.persistable = True
test_program = test_program.clone(for_test=True)
if args.use_cuda:
place = fluid.CUDAPlace(0)
dev_count = fluid.core.get_cuda_device_count()
else:
place = fluid.CPUPlace()
dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
print("device count %d" % dev_count)
print("theoretical memory usage: ")
print(
fluid.contrib.memory_usage(program=train_program,
batch_size=args.batch_size))
exe = fluid.Executor(place)
exe.run(train_startup)
exe.run(test_startup)
train_exe = fluid.ParallelExecutor(use_cuda=args.use_cuda,
loss_name=loss.name,
main_program=train_program)
test_exe = fluid.ParallelExecutor(use_cuda=args.use_cuda,
main_program=test_program,
share_vars_from=train_exe)
if args.init_model:
init.init_pretraining_params(exe,
args.init_model,
main_program=train_startup)
print('sccuess init %s' % args.init_model)
print("start loading data ...")
def train_with_feed(batch_data):
"""
Train on one batch
"""
#to do get_feed_names
feed_dict = dict(zip(net.get_feed_names(), batch_data))
cost = train_exe.run(feed=feed_dict, fetch_list=[loss.name])
return cost[0]
def test_with_feed(batch_data):
"""
Test on one batch
"""
feed_dict = dict(zip(net.get_feed_names(), batch_data))
score = test_exe.run(feed=feed_dict, fetch_list=[logits.name])
return score[0]
def evaluate():
"""
Evaluate to choose model
"""
val_batches = reader.batch_reader(args.val_path, args.batch_size,
place, args.max_len, 1)
scores = []
labels = []
for batch in val_batches:
scores.extend(test_with_feed(batch))
labels.extend([x[0] for x in batch[2]])
scores = [x[0] for x in scores]
return eva.evaluate_cor(scores, labels)
def save_exe(step, best_cor):
"""
Save exe conditional
"""
cor = evaluate()
print('evaluation cor relevance %s' % cor)
if cor > best_cor and step != 0:
fluid.io.save_inference_model(args.save_path,
net.get_feed_inference_names(),
logits,
exe,
main_program=train_program)
print("Save model at step %d ... " % step)
print(
time.strftime('%Y-%m-%d %H:%M:%S',
time.localtime(time.time())))
best_cor = cor
return best_cor
# train over different epoches
global_step, train_time = 0, 0.0
best_cor = 0.0
pre_index = -1
for epoch in six.moves.xrange(args.num_scan_data):
train_batches = reader.batch_reader(args.train_path, args.batch_size,
place, args.max_len,
args.sample_pro)
begin_time = time.time()
sum_cost = 0
for batch in train_batches:
if (args.save_path is not None) and (global_step % args.save_step
== 0):
best_cor = save_exe(global_step, best_cor)
cost = train_with_feed(batch)
global_step += 1
sum_cost += cost.mean()
if global_step % args.print_step == 0:
print('training step %s avg loss %s' %
(global_step, sum_cost / args.print_step))
sum_cost = 0
pass_time_cost = time.time() - begin_time
train_time += pass_time_cost
print("Pass {0}, pass_time_cost {1}".format(
epoch, "%2.2f sec" % pass_time_cost))
def main():
args = parse_args()
lstm_size = args.hidden_dim
data = fluid.layers.data(name="words",
shape=[1],
lod_level=1,
dtype='int64')
sentence = fluid.layers.embedding(input=data,
size=[len(word_dict), args.emb_dim])
sentence = fluid.layers.fc(input=sentence, size=lstm_size, act='tanh')
rnn = fluid.layers.DynamicRNN()
with rnn.block():
word = rnn.step_input(sentence)
prev_hidden = rnn.memory(value=0.0, shape=[lstm_size])
prev_cell = rnn.memory(value=0.0, shape=[lstm_size])
def gate_common(
ipt,
hidden,
size,
):
gate0 = fluid.layers.fc(input=ipt, size=size, bias_attr=True)
gate1 = fluid.layers.fc(input=hidden, size=size, bias_attr=False)
gate = fluid.layers.sums(input=[gate0, gate1])
return gate
forget_gate = fluid.layers.sigmoid(
x=gate_common(word, prev_hidden, lstm_size))
input_gate = fluid.layers.sigmoid(
x=gate_common(word, prev_hidden, lstm_size))
output_gate = fluid.layers.sigmoid(
x=gate_common(word, prev_hidden, lstm_size))
cell_gate = fluid.layers.tanh(
x=gate_common(word, prev_hidden, lstm_size))
cell = fluid.layers.sums(input=[
fluid.layers.elementwise_mul(x=forget_gate, y=prev_cell),
fluid.layers.elementwise_mul(x=input_gate, y=cell_gate)
])
hidden = fluid.layers.elementwise_mul(x=output_gate,
y=fluid.layers.tanh(x=cell))
rnn.update_memory(prev_cell, cell)
rnn.update_memory(prev_hidden, hidden)
rnn.output(hidden)
last = fluid.layers.sequence_pool(rnn(), 'last')
logit = fluid.layers.fc(input=last, size=2, act='softmax')
loss = fluid.layers.cross_entropy(input=logit,
label=fluid.layers.data(name='label',
shape=[1],
dtype='int64'))
loss = fluid.layers.mean(x=loss)
# add acc
batch_size_tensor = fluid.layers.create_tensor(dtype='int64')
batch_acc = fluid.layers.accuracy(input=logit, label=fluid.layers.data(name='label', \
shape=[1], dtype='int64'), total=batch_size_tensor)
adam = fluid.optimizer.Adam()
adam.minimize(loss)
fluid.memory_optimize(fluid.default_main_program())
place = fluid.CPUPlace() if args.device == 'CPU' else fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
train_reader = batch(paddle.reader.shuffle(crop_sentence(
imdb.train(word_dict), args.crop_size),
buf_size=25000),
batch_size=args.batch_size)
train_acc_kpi = None
for kpi in tracking_kpis:
if kpi.name == 'imdb_%s_train_acc' % (args.batch_size):
train_acc_kpi = kpi
train_speed_kpi = None
for kpi in tracking_kpis:
if kpi.name == 'imdb_%s_train_speed' % (args.batch_size):
train_speed_kpi = kpi
iters, num_samples, start_time = 0, 0, time.time()
for pass_id in range(args.pass_num):
train_accs = []
train_losses = []
for batch_id, data in enumerate(train_reader()):
if iters == args.skip_batch_num:
start_time = time.time()
num_samples = 0
if iters == args.iterations:
break
tensor_words = to_lodtensor([x[0] for x in data], place)
label = numpy.array([x[1] for x in data]).astype("int64")
label = label.reshape((-1, 1))
loss_np, acc, weight = exe.run(
fluid.default_main_program(),
feed={
"words": tensor_words,
"label": label
},
fetch_list=[loss, batch_acc, batch_size_tensor])
iters += 1
for x in data:
num_samples += len(x[0])
print(
"Pass = %d, Iter = %d, Loss = %f, Accuracy = %f" %
(pass_id, iters, loss_np, acc)
) # The accuracy is the accumulation of batches, but not the current batch.
train_elapsed = time.time() - start_time
examples_per_sec = num_samples / train_elapsed
print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' %
(num_samples, train_elapsed, examples_per_sec))
train_speed_kpi.add_record(np.array(examples_per_sec, dtype='float32'))
break
train_speed_kpi.persist()
def run_benchmark(model, args):
dshape, class_dim = get_data_shape(args)
input = fluid.layers.data(name='data', shape=dshape, dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
predict = model(input, class_dim)
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(x=cost)
fluid.default_main_program().seed = 1
fluid.default_startup_program().seed = 1
batch_size_tensor = fluid.layers.create_tensor(dtype='int64')
batch_acc = fluid.layers.accuracy(input=predict,
label=label,
total=batch_size_tensor)
inference_program = fluid.default_main_program().clone()
with fluid.program_guard(inference_program):
inference_program = fluid.io.get_inference_program(
target_vars=[batch_acc, batch_size_tensor])
optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
opts = optimizer.minimize(avg_cost)
fluid.memory_optimize(fluid.default_main_program())
# Init ParallelExecutor
train_exe, test_exe = get_parallel_executor(args, avg_cost,
fluid.default_main_program(),
inference_program)
# Prepare reader
train_reader, test_reader = init_reader(args)
def test(test_exe):
test_accuracy = fluid.average.WeightedAverage()
for batch_id, data in enumerate(test_reader()):
if batch_id == args.iterations:
break
img_data = np.array(map(lambda x: x[0].reshape(dshape),
data)).astype("float32")
y_data = np.array(map(lambda x: x[1],
data)).astype("int64").reshape([-1, 1])
acc, weight = test_exe.run(
fetch_list=[batch_acc.name, batch_size_tensor.name],
feed={
"data": img_data,
"label": y_data
})
acc = float((acc * weight).sum() / weight.sum())
weight = int(weight.sum())
test_accuracy.add(value=acc, weight=weight)
return test_accuracy.eval()
im_num, total_train_time, total_iters = 0, 0.0, 0
accuracy = fluid.average.WeightedAverage()
fetch_list = [avg_cost.name, batch_acc.name, batch_size_tensor.name]
for pass_id in range(args.pass_num):
every_pass_loss = []
accuracy.reset()
iter, pass_duration = 0, 0.0
for batch_id, data in enumerate(train_reader()):
batch_start = time.time()
if iter == args.iterations:
break
image = np.array(map(lambda x: x[0].reshape(dshape),
data)).astype('float32')
label = np.array(map(lambda x: x[1],
data)).astype('int64').reshape([-1, 1])
loss, acc, weight = train_exe.run(fetch_list=fetch_list,
feed={
'data': image,
'label': label
})
acc = float((acc * weight).sum() / weight.sum())
loss = (loss * weight).sum() / weight.sum()
weight = int(weight.sum())
accuracy.add(value=acc, weight=weight)
if iter >= args.skip_batch_num or pass_id != 0:
batch_duration = time.time() - batch_start
pass_duration += batch_duration
im_num += label.shape[0]
every_pass_loss.append(loss)
# print("Pass: %d, Iter: %d, loss: %s, acc: %s" %
# (pass_id, iter, str(loss), str(acc)))
iter += 1
total_iters += 1
total_train_time += pass_duration
pass_train_acc = accuracy.eval()
pass_test_acc = test(test_exe)
print(
"Pass:%d, Loss:%f, Train Accuray:%f, Test Accuray:%f, Handle Images Duration: %f\n"
% (pass_id, np.mean(every_pass_loss), pass_train_acc,
pass_test_acc, pass_duration))
record_kpi(pass_id, iter, pass_train_acc, total_train_time, im_num)
examples_per_sec = im_num / total_train_time
sec_per_batch = total_train_time / \
(iter * args.pass_num - args.skip_batch_num)
print('\nTotal examples: %d, total time: %.5f' %
(im_num, total_train_time))
print('%.5f examples/sec, %.5f sec/batch \n' %
(examples_per_sec, sec_per_batch))
def check_network_convergence(self,
method,
memory_opt=True,
iter=50,
batch_size=None,
allow_op_delay=False,
feed_dict=None,
seed=None,
use_parallel_executor=True,
balance_parameter_opt_between_cards=False):
def run_executor(exe, feed, fetch_list, program=None):
if isinstance(exe, fluid.ParallelExecutor):
res = exe.run(fetch_list=fetch_list, feed=feed)
elif isinstance(exe, fluid.Executor):
if program is None:
program = fluid.default_main_program()
res = exe.run(program=program, feed=feed, fetch_list=fetch_list)
else:
raise ValueError('Unkown type exe')
return res
main = fluid.Program()
startup = fluid.Program()
startup.random_seed = 1 # Fix random seed
with fluid.program_guard(main, startup):
if seed is not None:
startup.random_seed = seed
loss = method(use_feed=feed_dict is not None)
adam = fluid.optimizer.Adam()
adam.minimize(loss)
if memory_opt:
fluid.memory_optimize(main)
place = fluid.CUDAPlace(0)
startup_exe = fluid.Executor(place)
startup_exe.run(startup)
exec_strategy = fluid.ExecutionStrategy()
exec_strategy.allow_op_delay = allow_op_delay
build_strategy = fluid.BuildStrategy()
build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce if balance_parameter_opt_between_cards else fluid.BuildStrategy.ReduceStrategy.AllReduce
if use_parallel_executor:
exe = fluid.ParallelExecutor(
True,
loss_name=loss.name,
exec_strategy=exec_strategy,
build_strategy=build_strategy)
else:
exe = fluid.Executor(place=place)
if batch_size is not None:
batch_size *= fluid.core.get_cuda_device_count()
begin = time.time()
first_loss, = run_executor(
exe=exe, feed=feed_dict, fetch_list=[loss.name])
first_loss = np.array(first_loss)
for i in xrange(iter):
run_executor(exe=exe, feed=feed_dict, fetch_list=[])
last_loss, = run_executor(
exe=exe, feed=feed_dict, fetch_list=[loss.name])
end = time.time()
if batch_size is not None:
print "%.4f Instance per second" % (
(batch_size * iter + 2) / (end - begin))
last_loss = np.array(last_loss)
print first_loss, last_loss
# self.assertGreater(first_loss[0], last_loss[0])
return first_loss, last_loss
def train(args):
# parse config
config = parse_config(args.config)
train_config = merge_configs(config, 'train', vars(args))
valid_config = merge_configs(config, 'valid', vars(args))
train_model = models.get_model(args.model_name, train_config, mode='train')
valid_model = models.get_model(args.model_name, valid_config, mode='valid')
# build model
startup = fluid.Program()
train_prog = fluid.Program()
with fluid.program_guard(train_prog, startup):
with fluid.unique_name.guard():
train_model.build_input(not args.no_use_pyreader)
train_model.build_model()
# for the input, has the form [data1, data2,..., label], so train_feeds[-1] is label
train_feeds = train_model.feeds()
train_feeds[-1].persistable = True
# for the output of classification model, has the form [pred]
train_outputs = train_model.outputs()
for output in train_outputs:
output.persistable = True
train_loss = train_model.loss()
train_loss.persistable = True
# outputs, loss, label should be fetched, so set persistable to be true
optimizer = train_model.optimizer()
optimizer.minimize(train_loss)
train_pyreader = train_model.pyreader()
if not args.no_memory_optimize:
fluid.memory_optimize(train_prog)
valid_prog = fluid.Program()
with fluid.program_guard(valid_prog, startup):
with fluid.unique_name.guard():
valid_model.build_input(not args.no_use_pyreader)
valid_model.build_model()
valid_feeds = valid_model.feeds()
valid_outputs = valid_model.outputs()
valid_loss = valid_model.loss()
valid_pyreader = valid_model.pyreader()
place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(startup)
if args.resume:
# if resume weights is given, load resume weights directly
assert os.path.exists(args.resume), \
"Given resume weight dir {} not exist.".format(args.resume)
def if_exist(var):
return os.path.exists(os.path.join(args.resume, var.name))
fluid.io.load_vars(exe, args.resume, predicate=if_exist, main_program=train_prog)
else:
# if not in resume mode, load pretrain weights
if args.pretrain:
assert os.path.exists(args.pretrain), \
"Given pretrain weight dir {} not exist.".format(args.pretrain)
pretrain = args.pretrain or train_model.get_pretrain_weights()
if pretrain:
train_model.load_pretrain_params(exe, pretrain, train_prog, place)
train_exe = fluid.ParallelExecutor(
use_cuda=args.use_gpu,
loss_name=train_loss.name,
main_program=train_prog)
valid_exe = fluid.ParallelExecutor(
use_cuda=args.use_gpu,
share_vars_from=train_exe,
main_program=valid_prog)
# get reader
bs_denominator = 1
if (not args.no_use_pyreader) and args.use_gpu:
bs_denominator = train_config.TRAIN.num_gpus
train_config.TRAIN.batch_size = int(train_config.TRAIN.batch_size /
bs_denominator)
valid_config.VALID.batch_size = int(valid_config.VALID.batch_size /
bs_denominator)
train_reader = get_reader(args.model_name.upper(), 'train', train_config)
valid_reader = get_reader(args.model_name.upper(), 'valid', valid_config)
# get metrics
train_metrics = get_metrics(args.model_name.upper(), 'train', train_config)
valid_metrics = get_metrics(args.model_name.upper(), 'valid', valid_config)
train_fetch_list = [train_loss.name] + [x.name for x in train_outputs
] + [train_feeds[-1].name]
valid_fetch_list = [valid_loss.name] + [x.name for x in valid_outputs
] + [valid_feeds[-1].name]
epochs = args.epoch_num or train_model.epoch_num()
if args.no_use_pyreader:
train_feeder = fluid.DataFeeder(place=place, feed_list=train_feeds)
valid_feeder = fluid.DataFeeder(place=place, feed_list=valid_feeds)
train_without_pyreader(exe, train_prog, train_exe, train_reader, train_feeder,
train_fetch_list, train_metrics, epochs = epochs,
log_interval = args.log_interval, valid_interval = args.valid_interval,
save_dir = args.save_dir, save_model_name = args.model_name,
test_exe = valid_exe, test_reader = valid_reader, test_feeder = valid_feeder,
test_fetch_list = valid_fetch_list, test_metrics = valid_metrics)
else:
train_pyreader.decorate_paddle_reader(train_reader)
valid_pyreader.decorate_paddle_reader(valid_reader)
train_with_pyreader(exe, train_prog, train_exe, train_pyreader, train_fetch_list, train_metrics,
epochs = epochs, log_interval = args.log_interval,
valid_interval = args.valid_interval,
save_dir = args.save_dir, save_model_name = args.model_name,
test_exe = valid_exe, test_pyreader = valid_pyreader,
test_fetch_list = valid_fetch_list, test_metrics = valid_metrics)
def train(conf_dict):
"""
train process
"""
# Get data layer
data = layers.DataLayer()
# Load network structure dynamically
net = utils.import_class(
"nets", conf_dict["net"]["module_name"], conf_dict["net"]["class_name"])(conf_dict)
# Load loss function dynamically
loss = utils.import_class(
"losses", conf_dict["loss"]["module_name"], conf_dict["loss"]["class_name"])(conf_dict)
# Load Optimization method
optimizer = utils.import_class(
"optimizers", "paddle_optimizers", conf_dict["optimizer"]["class_name"])(conf_dict)
# Get service
place = fluid.core.CPUPlace()
if conf_dict["task_mode"] == "pairwise":
# Build network
left = data.ops(name="left", shape=[1], dtype="int64", lod_level=1)
pos_right = data.ops(name="right", shape=[
1], dtype="int64", lod_level=1)
neg_right = data.ops(name="neg_right", shape=[
1], dtype="int64", lod_level=1)
left_feat, pos_score = net.predict(left, pos_right)
_, neg_score = net.predict(left, neg_right)
avg_cost = loss.compute(pos_score, neg_score)
# Get Feeder and Reader
feeder = fluid.DataFeeder(place=place, feed_list=[
left.name, pos_right.name, neg_right.name])
reader = data_reader.get_reader(conf_dict, False, None)
else:
# Build network
left = data.ops(name="left", shape=[1], dtype="int64", lod_level=1)
right = data.ops(name="right", shape=[1], dtype="int64", lod_level=1)
label = data.ops(name="label", shape=[1], dtype="int64", lod_level=0)
left_feat, pred = net.predict(left, right)
avg_cost = loss.compute(pred, label)
# Get Feeder and Reader
feeder = fluid.DataFeeder(place=place, feed_list=[
left.name, right.name, label.name])
reader = data_reader.get_reader(conf_dict, False, None)
# Save Infer model
infer_program = fluid.default_main_program().clone()
# operate Optimization
optimizer.ops(avg_cost)
# optimize memory
fluid.memory_optimize(fluid.default_main_program())
executor = fluid.Executor(place)
executor.run(fluid.default_startup_program())
# Get and run executor
parallel_executor = fluid.ParallelExecutor(
use_cuda=False, loss_name=avg_cost.name,
main_program=fluid.default_main_program())
# Get device number
device_count = parallel_executor.device_count
logging.info("device count: %d" % device_count)
# run train
logging.info("start train process ...")
for epoch_id in range(conf_dict["epoch_num"]):
losses = []
# Get batch data iterator
batch_data = paddle.batch(reader, conf_dict["batch_size"], drop_last=False)
start_time = time.time()
for iter, data in enumerate(batch_data()):
if len(data) < device_count:
continue
avg_loss = parallel_executor.run(
[avg_cost.name], feed=feeder.feed(data))
print("epoch: %d, iter: %d, loss: %f" %
(epoch_id, iter, np.mean(avg_loss[0])))
losses.append(np.mean(avg_loss[0]))
end_time = time.time()
print("epoch: %d, loss: %f, used time: %d sec" %
(epoch_id, np.mean(losses), end_time - start_time))
model_save_dir = conf_dict["model_path"]
model_path = os.path.join(model_save_dir, str(epoch_id))
if not os.path.exists(model_save_dir):
os.makedirs(model_save_dir)
if conf_dict["task_mode"] == "pairwise":
feed_var_names = [left.name, pos_right.name]
target_vars = [left_feat, pos_score]
else:
feed_var_names = [left.name, right.name]
target_vars = [left_feat, pred]
fluid.io.save_inference_model(
model_path, feed_var_names, target_vars, executor, infer_program)
def train():
avg_cost, feeding_list = seq_to_seq_net(args.embedding_dim,
args.encoder_size,
args.decoder_size,
args.dict_size,
args.dict_size,
False,
beam_size=args.beam_size,
max_length=args.max_length)
# clone from default main program
inference_program = fluid.default_main_program().clone()
optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate)
optimizer.minimize(avg_cost)
fluid.memory_optimize(fluid.default_main_program())
train_batch_generator = paddle.batch(paddle.reader.shuffle(
paddle.dataset.wmt14.train(args.dict_size), buf_size=1000),
batch_size=args.batch_size)
test_batch_generator = paddle.batch(paddle.reader.shuffle(
paddle.dataset.wmt14.test(args.dict_size), buf_size=1000),
batch_size=args.batch_size)
place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0)
exe = Executor(place)
exe.run(framework.default_startup_program())
def do_validation():
total_loss = 0.0
count = 0
for batch_id, data in enumerate(test_batch_generator()):
src_seq = to_lodtensor(map(lambda x: x[0], data), place)[0]
trg_seq = to_lodtensor(map(lambda x: x[1], data), place)[0]
lbl_seq = to_lodtensor(map(lambda x: x[2], data), place)[0]
fetch_outs = exe.run(inference_program,
feed={
feeding_list[0]: src_seq,
feeding_list[1]: trg_seq,
feeding_list[2]: lbl_seq
},
fetch_list=[avg_cost],
return_numpy=False)
total_loss += lodtensor_to_ndarray(fetch_outs[0])[0]
count += 1
return total_loss / count
iters, num_samples, start_time = 0, 0, time.time()
for pass_id in xrange(args.pass_num):
train_accs = []
train_losses = []
for batch_id, data in enumerate(train_batch_generator()):
if iters == args.skip_batch_num:
start_time = time.time()
num_samples = 0
if iters == args.iterations:
break
src_seq, word_num = to_lodtensor(map(lambda x: x[0], data), place)
num_samples += word_num
trg_seq, word_num = to_lodtensor(map(lambda x: x[1], data), place)
num_samples += word_num
lbl_seq, _ = to_lodtensor(map(lambda x: x[2], data), place)
fetch_outs = exe.run(framework.default_main_program(),
feed={
feeding_list[0]: src_seq,
feeding_list[1]: trg_seq,
feeding_list[2]: lbl_seq
},
fetch_list=[avg_cost])
iters += 1
loss = np.array(fetch_outs[0])
print(
"Pass = %d, Iter = %d, Loss = %f" % (pass_id, iters, loss)
) # The accuracy is the accumulation of batches, but not the current batch.
train_elapsed = time.time() - start_time
examples_per_sec = num_samples / train_elapsed
print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' %
(num_samples, train_elapsed, examples_per_sec))
# evaluation
if args.with_test:
test_loss = do_validation()
exit(0)
def train(args):
print("pretraining start")
ernie_config = ErnieConfig(args.ernie_config_path)
ernie_config.print_config()
train_program = fluid.Program()
startup_prog = fluid.Program()
with fluid.program_guard(train_program, startup_prog):
with fluid.unique_name.guard():
train_pyreader, next_sent_acc, mask_lm_loss, total_loss = create_model(
pyreader_name='train_reader', ernie_config=ernie_config)
scheduled_lr = optimization(loss=total_loss,
warmup_steps=args.warmup_steps,
num_train_steps=args.num_train_steps,
learning_rate=args.learning_rate,
train_program=train_program,
startup_prog=startup_prog,
weight_decay=args.weight_decay,
scheduler=args.lr_scheduler,
use_fp16=args.use_fp16,
loss_scaling=args.loss_scaling)
fluid.memory_optimize(input_program=train_program,
skip_opt_set=[
next_sent_acc.name, mask_lm_loss.name,
total_loss.name
])
test_prog = fluid.Program()
with fluid.program_guard(test_prog, startup_prog):
with fluid.unique_name.guard():
test_pyreader, next_sent_acc, mask_lm_loss, total_loss = create_model(
pyreader_name='test_reader', ernie_config=ernie_config)
test_prog = test_prog.clone(for_test=True)
if args.use_cuda:
place = fluid.CUDAPlace(0)
dev_count = fluid.core.get_cuda_device_count()
else:
place = fluid.CPUPlace()
dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
print("Device count %d" % dev_count)
print("theoretical memory usage: ")
if args.in_tokens:
print(
fluid.contrib.memory_usage(program=train_program,
batch_size=args.batch_size //
args.max_seq_len))
else:
print(
fluid.contrib.memory_usage(program=train_program,
batch_size=args.batch_size))
nccl2_num_trainers = 1
nccl2_trainer_id = 0
print("args.is_distributed:", args.is_distributed)
if args.is_distributed:
worker_endpoints_env = os.getenv("worker_endpoints")
worker_endpoints = worker_endpoints_env.split(",")
trainers_num = len(worker_endpoints)
current_endpoint = os.getenv("current_endpoint")
trainer_id = worker_endpoints.index(current_endpoint)
if trainer_id == 0:
print("train_id == 0, sleep 60s")
time.sleep(60)
print("worker_endpoints:{} trainers_num:{} current_endpoint:{} \
trainer_id:{}".format(worker_endpoints, trainers_num,
current_endpoint, trainer_id))
# prepare nccl2 env.
config = fluid.DistributeTranspilerConfig()
config.mode = "nccl2"
t = fluid.DistributeTranspiler(config=config)
t.transpile(trainer_id,
trainers=worker_endpoints_env,
current_endpoint=current_endpoint,
program=train_program,
startup_program=startup_prog)
nccl2_num_trainers = trainers_num
nccl2_trainer_id = trainer_id
exe = fluid.Executor(place)
exe.run(startup_prog)
if args.init_checkpoint and args.init_checkpoint != "":
init_checkpoint(exe, args.init_checkpoint, train_program,
args.use_fp16)
data_reader = ErnieDataReader(filelist=args.train_filelist,
batch_size=args.batch_size,
vocab_path=args.vocab_path,
voc_size=ernie_config['vocab_size'],
epoch=args.epoch,
max_seq_len=args.max_seq_len,
generate_neg_sample=args.generate_neg_sample,
in_tokens=args.in_tokens,
is_bidirection=args.is_bidirection)
exec_strategy = fluid.ExecutionStrategy()
if args.use_fast_executor:
exec_strategy.use_experimental_executor = True
exec_strategy.num_threads = dev_count
exec_strategy.num_iteration_per_drop_scope = min(10, args.skip_steps)
build_strategy = fluid.BuildStrategy()
build_strategy.remove_unnecessary_lock = False
train_exe = fluid.ParallelExecutor(use_cuda=args.use_cuda,
loss_name=total_loss.name,
build_strategy=build_strategy,
exec_strategy=exec_strategy,
main_program=train_program,
num_trainers=nccl2_num_trainers,
trainer_id=nccl2_trainer_id)
if args.valid_filelist and args.valid_filelist != "":
predict = predict_wrapper(args,
exe,
ernie_config,
test_prog=test_prog,
pyreader=test_pyreader,
fetch_list=[
next_sent_acc.name, mask_lm_loss.name,
total_loss.name
])
train_pyreader.decorate_tensor_provider(data_reader.data_generator())
train_pyreader.start()
steps = 0
cost = []
lm_cost = []
acc = []
time_begin = time.time()
while steps < args.num_train_steps:
try:
steps += nccl2_num_trainers
skip_steps = args.skip_steps * nccl2_num_trainers
if nccl2_trainer_id != 0:
train_exe.run(fetch_list=[])
continue
if steps % skip_steps != 0:
train_exe.run(fetch_list=[])
else:
each_next_acc, each_mask_lm_cost, each_total_cost, np_lr = train_exe.run(
fetch_list=[
next_sent_acc.name, mask_lm_loss.name, total_loss.name,
scheduled_lr.name
])
acc.extend(each_next_acc)
lm_cost.extend(each_mask_lm_cost)
cost.extend(each_total_cost)
print("feed_queue size", train_pyreader.queue.size())
time_end = time.time()
used_time = time_end - time_begin
epoch, current_file_index, total_file, current_file, mask_type = data_reader.get_progress(
)
print("current learning_rate:%f" % np_lr[0])
print(
"epoch: %d, progress: %d/%d, step: %d, loss: %f, "
"ppl: %f, next_sent_acc: %f, speed: %f steps/s, file: %s, mask_type: %s"
%
(epoch, current_file_index, total_file, steps,
np.mean(np.array(cost)), np.mean(np.exp(
np.array(lm_cost))), np.mean(np.array(acc)),
skip_steps / used_time, current_file, mask_type))
cost = []
lm_cost = []
acc = []
time_begin = time.time()
if steps % args.save_steps == 0:
save_path = os.path.join(args.checkpoints,
"step_" + str(steps))
fluid.io.save_persistables(exe, save_path, train_program)
if args.valid_filelist and steps % args.validation_steps == 0:
vali_cost, vali_lm_cost, vali_acc, vali_steps, vali_speed = predict(
)
print("[validation_set] epoch: %d, step: %d, "
"loss: %f, global ppl: %f, batch-averged ppl: %f, "
"next_sent_acc: %f, speed: %f steps/s" %
(epoch, steps, np.mean(np.array(vali_cost) / vali_steps),
np.exp(np.mean(np.array(vali_lm_cost) / vali_steps)),
np.mean(np.exp(np.array(vali_lm_cost) / vali_steps)),
np.mean(np.array(vali_acc) / vali_steps), vali_speed))
except fluid.core.EOFException:
train_pyreader.reset()
break
places = fluid.layers.get_places(device_count=0, device_type=device_type)
pd = fluid.layers.ParallelDo(places, use_nccl=use_nccl)
with pd.do():
x_ = pd.read_input(x)
y_ = pd.read_input(y)
y_predict = fluid.layers.fc(input=x_, size=1, act=None)
cost = fluid.layers.square_error_cost(input=y_predict, label=y_)
avg_cost = fluid.layers.mean(x=cost)
pd.write_output(avg_cost)
cost = pd()
avg_cost = fluid.layers.mean(x=cost)
sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.01)
sgd_optimizer.minimize(avg_cost)
fluid.memory_optimize(fluid.default_main_program(), print_log=True)
# fluid.release_memory(fluid.default_main_program())
BATCH_SIZE = 200
# fix the order of training data
train_reader = paddle.batch(
paddle.dataset.uci_housing.train(), batch_size=BATCH_SIZE)
# train_reader = paddle.batch(
# paddle.reader.shuffle(
# paddle.dataset.uci_housing.train(), buf_size=500),
# batch_size=BATCH_SIZE)
feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
exe = fluid.Executor(place)
main_program=program)
if __name__ == '__main__':
parse = argparse.ArgumentParser(description='')
parse.add_argument('--model', help='model name', nargs='?')
args = parse.parse_args()
model = args.model
DataSet = create_reader(model)
predict = create_model(model=model)
place = fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
fluid.memory_optimize(fluid.default_main_program())
load_model(exe, fluid.default_main_program(), model=model)
batches = DataSet.get_batch_generator(1, 1234)
for i, imgs, names in batches:
result = exe.run(fluid.default_main_program(),
feed={'img': imgs},
fetch_list=[predict])
print(i)
path = path + 'data/unet/test/ColorImage/' + names[0].split(
"image/")[1]
picture = np.argmax(result[0], axis=1)
picture = picture.reshape((1024, 1024))
saveImage(picture, path)
def main(args):
ernie_config = ErnieConfig(args.ernie_config_path)
ernie_config.print_config()
if args.use_cuda:
place = fluid.CUDAPlace(int(os.getenv('FLAGS_selected_gpus', '0')))
dev_count = fluid.core.get_cuda_device_count()
else:
place = fluid.CPUPlace()
dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
exe = fluid.Executor(place)
reader = task_reader.ClassifyReader(
vocab_path=args.vocab_path,
label_map_config=args.label_map_config,
max_seq_len=args.max_seq_len,
do_lower_case=args.do_lower_case,
in_tokens=args.in_tokens,
random_seed=args.random_seed)
if not (args.do_train or args.do_val or args.do_test):
raise ValueError("For args `do_train`, `do_val` and `do_test`, at "
"least one of them must be True.")
startup_prog = fluid.Program()
if args.random_seed is not None:
startup_prog.random_seed = args.random_seed
if args.do_train:
train_data_generator = reader.data_generator(
input_file=args.train_set,
batch_size=args.batch_size,
epoch=args.epoch,
shuffle=True,
phase="train")
num_train_examples = reader.get_num_examples(args.train_set)
if args.in_tokens:
max_train_steps = args.epoch * num_train_examples // (
args.batch_size // args.max_seq_len) // dev_count
else:
max_train_steps = args.epoch * num_train_examples // args.batch_size // dev_count
warmup_steps = int(max_train_steps * args.warmup_proportion)
print("Device count: %d" % dev_count)
print("Num train examples: %d" % num_train_examples)
print("Max train steps: %d" % max_train_steps)
print("Num warmup steps: %d" % warmup_steps)
train_program = fluid.Program()
with fluid.program_guard(train_program, startup_prog):
with fluid.unique_name.guard():
train_pyreader, graph_vars = create_model(
args,
pyreader_name='train_reader',
ernie_config=ernie_config)
scheduled_lr = optimization(
loss=graph_vars["loss"],
warmup_steps=warmup_steps,
num_train_steps=max_train_steps,
learning_rate=args.learning_rate,
train_program=train_program,
startup_prog=startup_prog,
weight_decay=args.weight_decay,
scheduler=args.lr_scheduler,
use_fp16=args.use_fp16,
loss_scaling=args.loss_scaling)
fluid.memory_optimize(
input_program=train_program,
skip_opt_set=[
graph_vars["loss"].name,
graph_vars["probs"].name,
graph_vars["accuracy"].name,
graph_vars["num_seqs"].name,
])
if args.verbose:
if args.in_tokens:
lower_mem, upper_mem, unit = fluid.contrib.memory_usage(
program=train_program,
batch_size=args.batch_size // args.max_seq_len)
else:
lower_mem, upper_mem, unit = fluid.contrib.memory_usage(
program=train_program, batch_size=args.batch_size)
print("Theoretical memory usage in training: %.3f - %.3f %s" %
(lower_mem, upper_mem, unit))
if args.do_val or args.do_test:
test_prog = fluid.Program()
with fluid.program_guard(test_prog, startup_prog):
with fluid.unique_name.guard():
test_pyreader, graph_vars = create_model(
args,
pyreader_name='test_reader',
ernie_config=ernie_config)
test_prog = test_prog.clone(for_test=True)
exe.run(startup_prog)
if args.do_train:
if args.init_checkpoint and args.init_pretraining_params:
print(
"WARNING: args 'init_checkpoint' and 'init_pretraining_params' "
"both are set! Only arg 'init_checkpoint' is made valid.")
if args.init_checkpoint:
init_checkpoint(
exe,
args.init_checkpoint,
main_program=startup_prog,
use_fp16=args.use_fp16)
elif args.init_pretraining_params:
init_pretraining_params(
exe,
args.init_pretraining_params,
main_program=startup_prog,
use_fp16=args.use_fp16)
elif args.do_val or args.do_test:
if not args.init_checkpoint:
raise ValueError("args 'init_checkpoint' should be set if"
"only doing validation or testing!")
init_checkpoint(
exe,
args.init_checkpoint,
main_program=startup_prog,
use_fp16=args.use_fp16)
if args.do_train:
exec_strategy = fluid.ExecutionStrategy()
if args.use_fast_executor:
exec_strategy.use_experimental_executor = True
exec_strategy.num_threads = dev_count
exec_strategy.num_iteration_per_drop_scope = args.num_iteration_per_drop_scope
train_exe = fluid.ParallelExecutor(
use_cuda=args.use_cuda,
loss_name=graph_vars["loss"].name,
exec_strategy=exec_strategy,
main_program=train_program)
train_pyreader.decorate_tensor_provider(train_data_generator)
else:
train_exe = None
if args.do_train:
train_pyreader.start()
steps = 0
if warmup_steps > 0:
graph_vars["learning_rate"] = scheduled_lr
if args.save_log and args.log_path:
if os.path.exists(args.log_path):
raise FileExistsError("Logging file already exists!")
with open(args.log_path, 'w') as logfile:
logfile.write('%s\n' % time.asctime())
print('Writing logs into %s' % args.log_path)
time_begin = time.time()
while True:
try:
steps += 1
if steps % args.skip_steps != 0:
train_exe.run(fetch_list=[])
else:
outputs = evaluate(train_exe, train_program, train_pyreader,
graph_vars, "train")
if args.verbose:
verbose = "train pyreader queue size: %d, " % train_pyreader.queue.size(
)
verbose += "learning rate: %f" % (
outputs["learning_rate"]
if warmup_steps > 0 else args.learning_rate)
print(verbose)
current_example, current_epoch = reader.get_train_progress()
time_end = time.time()
used_time = time_end - time_begin
print("epoch: %d, progress: %d/%d, step: %d, "
"ave loss: %.4f, micro_f1: %.4f, micro_p: %.4f, micro_r: %.4f, "
"speed: %f steps/s" %
(current_epoch, current_example, num_train_examples, steps,
outputs["loss"], outputs["micro_f"], outputs["micro_p"], outputs["micro_r"],
args.skip_steps / used_time))
# Todo: complete logging function
# Todo: print more useful metrics: f1/p/r instead of acc
if args.save_log and args.log_path:
with open(args.log_path, 'a') as logfile:
logfile.write("epoch: %d, progress: %d/%d, step: %d, "
"ave loss: %.4f, ave_acc: %.4f, micro_f1: %.4f, micro_p: %.4f, micro_r: %.4f, "
"speed: %f steps/s\n" %
(current_epoch, current_example, num_train_examples, steps,
outputs["loss"], outputs["accuracy"], outputs["micro_f"], outputs["micro_p"], outputs["micro_r"],
args.skip_steps / used_time))
time_begin = time.time()
if steps % args.save_steps == 0:
save_path = os.path.join(args.checkpoints,
"step_" + str(steps))
fluid.io.save_persistables(exe, save_path, train_program)
if steps % args.validation_steps == 0:
# evaluate dev set
if args.do_val:
test_pyreader.decorate_tensor_provider(
reader.data_generator(
args.dev_set,
batch_size=args.batch_size,
epoch=1,
shuffle=False))
evaluate(exe, test_prog, test_pyreader, graph_vars,
"dev")
# evaluate test set
if args.do_test:
test_pyreader.decorate_tensor_provider(
reader.data_generator(
args.test_set,
batch_size=args.batch_size,
epoch=1,
shuffle=False))
evaluate(exe, test_prog, test_pyreader, graph_vars,
"test")
except fluid.core.EOFException:
save_path = os.path.join(args.checkpoints, "step_" + str(steps))
fluid.io.save_persistables(exe, save_path, train_program)
train_pyreader.reset()
break
# final eval on dev set
if args.do_val:
test_pyreader.decorate_tensor_provider(
reader.data_generator(
args.dev_set,
batch_size=args.batch_size,
epoch=1,
shuffle=False))
print("Final validation result:")
evaluate(exe, test_prog, test_pyreader, graph_vars, "dev")
# final eval on test set
if args.do_test:
test_pyreader.decorate_tensor_provider(
reader.data_generator(
args.test_set,
batch_size=args.batch_size,
epoch=1,
shuffle=False))
print("Final test result:")
evaluate(exe, test_prog, test_pyreader, graph_vars, "test")
def main():
if args.data_set == "cifar10":
classdim = 10
if args.data_format == 'NCHW':
data_shape = [3, 32, 32]
else:
data_shape = [32, 32, 3]
else:
classdim = 102
if args.data_format == 'NCHW':
data_shape = [3, 224, 224]
else:
data_shape = [224, 224, 3]
# Input data
images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
# Train program
net = vgg16_bn_drop(images)
predict = fluid.layers.fc(input=net, size=classdim, act='softmax')
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(x=cost)
# Evaluator
batch_size_tensor = fluid.layers.create_tensor(dtype='int64')
batch_acc = fluid.layers.accuracy(input=predict,
label=label,
total=batch_size_tensor)
# inference program
inference_program = fluid.default_main_program().clone()
with fluid.program_guard(inference_program):
inference_program = fluid.io.get_inference_program(
target_vars=[batch_acc, batch_size_tensor])
# Optimization
optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate)
opts = optimizer.minimize(avg_cost)
fluid.memory_optimize(fluid.default_main_program())
# Initialize executor
place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0)
exe = fluid.Executor(place)
# Parameter initialization
exe.run(fluid.default_startup_program())
# data reader
train_reader = paddle.batch(paddle.reader.shuffle(
paddle.dataset.cifar.train10()
if args.data_set == 'cifar10' else paddle.dataset.flowers.train(),
buf_size=5120),
batch_size=args.batch_size)
test_reader = paddle.batch(paddle.dataset.cifar.test10() if args.data_set
== 'cifar10' else paddle.dataset.flowers.test(),
batch_size=args.batch_size)
# test
def test(exe):
test_accuracy = fluid.average.WeightedAverage()
for batch_id, data in enumerate(test_reader()):
img_data = np.array(map(lambda x: x[0].reshape(data_shape),
data)).astype("float32")
y_data = np.array(map(lambda x: x[1], data)).astype("int64")
y_data = y_data.reshape([-1, 1])
acc, weight = exe.run(inference_program,
feed={
"pixel": img_data,
"label": y_data
},
fetch_list=[batch_acc, batch_size_tensor])
test_accuracy.add(value=acc, weight=weight)
return test_accuracy.eval()
iters, num_samples, start_time = 0, 0, time.time()
accuracy = fluid.average.WeightedAverage()
for pass_id in range(args.pass_num):
accuracy.reset()
train_accs = []
train_losses = []
for batch_id, data in enumerate(train_reader()):
if iters == args.skip_batch_num:
start_time = time.time()
num_samples = 0
if iters == args.iterations:
break
img_data = np.array(map(lambda x: x[0].reshape(data_shape),
data)).astype("float32")
y_data = np.array(map(lambda x: x[1], data)).astype("int64")
y_data = y_data.reshape([-1, 1])
loss, acc, weight = exe.run(
fluid.default_main_program(),
feed={
"pixel": img_data,
"label": y_data
},
fetch_list=[avg_cost, batch_acc, batch_size_tensor])
accuracy.add(value=acc, weight=weight)
iters += 1
num_samples += len(y_data)
print(
"Pass = %d, Iter = %d, Loss = %f, Accuracy = %f" %
(pass_id, iters, loss, acc)
) # The accuracy is the accumulation of batches, but not the current batch.
# pass_train_acc = accuracy.eval()
train_losses.append(loss)
train_accs.append(acc)
print("Pass: %d, Loss: %f, Train Accuray: %f\n" %
(pass_id, np.mean(train_losses), np.mean(train_accs)))
train_elapsed = time.time() - start_time
examples_per_sec = num_samples / train_elapsed
print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' %
(num_samples, train_elapsed, examples_per_sec))
# evaluation
if args.with_test:
pass_test_acc = test(exe)
exit(0)
def _build_env(self):
if self.env.is_inititalized:
return
self._build_env_start_event()
self.env.is_inititalized = True
self.env.main_program = clone_program(self._base_main_program,
for_test=False)
self.env.startup_program = fluid.Program()
with fluid.program_guard(self.env.main_program,
self._base_startup_program):
with fluid.unique_name.guard(self.env.UNG):
self.env.outputs = self._build_net()
if self.is_train_phase or self.is_test_phase:
self.env.labels = self._add_label()
self.env.loss = self._add_loss()
self.env.metrics = self._add_metrics()
if self.is_predict_phase or self.is_test_phase:
self.env.main_program = clone_program(self.env.main_program,
for_test=True)
hub.common.paddle_helper.set_op_attr(self.env.main_program,
is_test=True)
if self.config.use_pyreader:
t_program = fluid.Program()
with fluid.program_guard(t_program, self.env.startup_program):
self.env.py_reader = fluid.layers.py_reader(
capacity=64,
shapes=[var.shape for var in self.feed_var_list],
dtypes=[
dtype_map[var.dtype] for var in self.feed_var_list
],
lod_levels=[var.lod_level for var in self.feed_var_list],
use_double_buffer=False)
feed_var_list = self.feed_var_list
py_vars = fluid.layers.read_file(self.env.py_reader)
py_vars = to_list(py_vars)
input_dict = {
feed_var_list[index].name: py_var
for index, py_var in enumerate(py_vars)
}
hub.connect_program(pre_program=t_program,
next_program=self.env.main_program,
input_dict=input_dict,
need_log=False)
self.env.main_program = t_program
if not self.is_predict_phase:
self.env.loss = self.env.main_program.global_block().vars[
self.env.loss.name]
metrics_name = [var.name for var in self.env.metrics]
self.env.metrics = [
self.env.main_program.global_block().vars[name]
for name in metrics_name
]
outputs_name = [var.name for var in self.env.outputs]
self.env.outputs = [
self.env.main_program.global_block().vars[name]
for name in outputs_name
]
if self.config.enable_memory_optim:
for var_name in self.fetch_list:
var = self.env.main_program.global_block().vars[var_name]
var.persistable = True
if self.is_train_phase:
with fluid.program_guard(self.env.main_program,
self._base_startup_program):
with fluid.unique_name.guard(self.env.UNG):
self.config.strategy.execute(self.loss,
self._base_data_reader,
self.config)
if self.is_train_phase:
loss_name = self.env.loss.name
share_vars_from = None
else:
loss_name = None
if self._base_compiled_program is None:
share_vars_from = None
else:
share_vars_from = self._base_compiled_program
if not self.config.use_data_parallel:
if self.config.enable_memory_optim:
fluid.memory_optimize(self.env.main_program)
self.env.main_program_compiled = None
else:
self.env.main_program_compiled = fluid.CompiledProgram(
self.env.main_program).with_data_parallel(
loss_name=loss_name,
share_vars_from=share_vars_from,
build_strategy=self.build_strategy)
if self._base_compiled_program is None:
self._base_compiled_program = self.env.main_program_compiled
self.exe.run(self.env.startup_program)
self._build_env_end_event()
