def eval(args):
model_list = [m for m in dir(models) if "__" not in m]
assert args.model in model_list, "{} is not in lists: {}".format(
args.model, model_list)
assert os.path.isdir(
args.pretrained_model
), "{} doesn't exist, please load right pretrained model path for eval".format(
args.pretrained_model)
assert args.image_shape[
1] <= args.resize_short_size, "Please check the args:image_shape and args:resize_short_size, The croped size(image_shape[1]) must smaller than or equal to the resized length(resize_short_size) "
# check gpu: when using gpu, the number of visible cards should divide batch size
if args.use_gpu:
assert args.batch_size % fluid.core.get_cuda_device_count(
) == 0, "please support correct batch_size({}), which can be divided by available cards({}), you can change the number of cards by indicating: export CUDA_VISIBLE_DEVICES= ".format(
args.batch_size, fluid.core.get_cuda_device_count())
image = fluid.data(name='image',
shape=[None] + args.image_shape,
dtype='float32')
label = fluid.data(name='label', shape=[None, 1], dtype='int64')
# model definition
if args.model.startswith('EfficientNet'):
model = models.__dict__[args.model](is_test=True,
padding_type=args.padding_type,
use_se=args.use_se)
elif "ACNet" in args.model:
model = models.__dict__[args.model](deploy=args.deploy)
else:
model = models.__dict__[args.model]()
if args.model == "GoogLeNet":
out0, out1, out2 = model.net(input=image, class_dim=args.class_dim)
cost0 = fluid.layers.cross_entropy(input=out0, label=label)
cost1 = fluid.layers.cross_entropy(input=out1, label=label)
cost2 = fluid.layers.cross_entropy(input=out2, label=label)
avg_cost0 = fluid.layers.mean(x=cost0)
avg_cost1 = fluid.layers.mean(x=cost1)
avg_cost2 = fluid.layers.mean(x=cost2)
avg_cost = avg_cost0 + 0.3 * avg_cost1 + 0.3 * avg_cost2
acc_top1 = fluid.layers.accuracy(input=out0, label=label, k=1)
acc_top5 = fluid.layers.accuracy(input=out0, label=label, k=5)
else:
out = model.net(input=image, class_dim=args.class_dim)
cost, pred = fluid.layers.softmax_with_cross_entropy(
out, label, return_softmax=True)
avg_cost = fluid.layers.mean(x=cost)
acc_top1 = fluid.layers.accuracy(input=pred, label=label, k=1)
acc_top5 = fluid.layers.accuracy(input=pred, label=label, k=5)
test_program = fluid.default_main_program().clone(for_test=True)
fetch_list = [avg_cost.name, acc_top1.name, acc_top5.name]
gpu_id = int(os.environ.get('FLAGS_selected_gpus', 0))
place = fluid.CUDAPlace(gpu_id) if args.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
if args.use_gpu:
places = fluid.framework.cuda_places()
else:
places = fluid.framework.cpu_places()
compiled_program = fluid.compiler.CompiledProgram(
test_program).with_data_parallel(places=places)
fluid.io.load_persistables(exe, args.pretrained_model)
imagenet_reader = reader.ImageNetReader()
val_reader = imagenet_reader.val(settings=args)
# set places to run on the multi-card
feeder = fluid.DataFeeder(place=places, feed_list=[image, label])
test_info = [[], [], []]
cnt = 0
parallel_data = []
parallel_id = []
place_num = paddle.fluid.core.get_cuda_device_count(
) if args.use_gpu else int(os.environ.get('CPU_NUM', 1))
real_iter = 0
info_dict = {}
for batch_id, data in enumerate(val_reader()):
#image data and label
image_data = [items[0:2] for items in data]
image_id = [items[2] for items in data]
parallel_id.append(image_id)
parallel_data.append(image_data)
if place_num == len(parallel_data):
t1 = time.time()
loss_set, acc1_set, acc5_set = exe.run(
compiled_program,
fetch_list=fetch_list,
feed=list(feeder.feed_parallel(parallel_data, place_num)))
t2 = time.time()
period = t2 - t1
loss = np.mean(loss_set)
acc1 = np.mean(acc1_set)
acc5 = np.mean(acc5_set)
test_info[0].append(loss * len(data))
test_info[1].append(acc1 * len(data))
test_info[2].append(acc5 * len(data))
cnt += len(data)
if batch_id % args.print_step == 0:
info = "Testbatch {0},loss {1}, acc1 {2},acc5 {3},time {4}".format(real_iter, \
"%.5f"%loss,"%.5f"%acc1, "%.5f"%acc5, \
"%2.2f sec" % period)
logger.info(info)
sys.stdout.flush()
parallel_id = []
parallel_data = []
real_iter += 1
test_loss = np.sum(test_info[0]) / cnt
test_acc1 = np.sum(test_info[1]) / cnt
test_acc5 = np.sum(test_info[2]) / cnt
info = "Test_loss {0}, test_acc1 {1}, test_acc5 {2}".format(
"%.5f" % test_loss, "%.5f" % test_acc1, "%.5f" % test_acc5)
if args.save_json_path:
info_dict = {
"Test_loss": test_loss,
"test_acc1": test_acc1,
"test_acc5": test_acc5
}
save_json(info_dict, args.save_json_path)
logger.info(info)
sys.stdout.flush()
def train(args):
"""Train model
Args:
args: all arguments.
"""
startup_prog = fluid.Program()
train_prog = fluid.Program()
train_out = build_program(
is_train=True,
main_prog=train_prog,
startup_prog=startup_prog,
args=args)
train_data_loader = train_out[-1]
if args.use_ema:
train_fetch_vars = train_out[:-2]
ema = train_out[-2]
else:
train_fetch_vars = train_out[:-1]
train_fetch_list = [var.name for var in train_fetch_vars]
if args.validate:
test_prog = fluid.Program()
test_out = build_program(
is_train=False,
main_prog=test_prog,
startup_prog=startup_prog,
args=args)
test_data_loader = test_out[-1]
test_fetch_vars = test_out[:-1]
test_fetch_list = [var.name for var in test_fetch_vars]
#Create test_prog and set layers' is_test params to True
test_prog = test_prog.clone(for_test=True)
gpu_id = int(os.environ.get('FLAGS_selected_gpus', 0))
place = fluid.CUDAPlace(gpu_id) if args.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(startup_prog)
trainer_id = int(os.getenv("PADDLE_TRAINER_ID", 0))
#init model by checkpoint or pretrianed model.
init_model(exe, args, train_prog)
num_trainers = int(os.environ.get('PADDLE_TRAINERS_NUM', 1))
if args.use_dali:
import dali
train_iter = dali.train(settings=args)
if trainer_id == 0:
test_iter = dali.val(settings=args)
else:
imagenet_reader = reader.ImageNetReader(0 if num_trainers > 1 else None)
train_reader = imagenet_reader.train(settings=args)
if args.use_gpu:
if num_trainers <= 1:
places = fluid.framework.cuda_places()
else:
places = place
else:
if num_trainers <= 1:
places = fluid.framework.cpu_places()
else:
places = place
train_data_loader.set_sample_list_generator(train_reader, places)
if args.validate:
test_reader = imagenet_reader.val(settings=args)
test_data_loader.set_sample_list_generator(test_reader, places)
compiled_train_prog = best_strategy_compiled(args, train_prog,
train_fetch_vars[0], exe)
#NOTE: this for benchmark
total_batch_num = 0
for pass_id in range(args.num_epochs):
if num_trainers > 1 and not args.use_dali:
imagenet_reader.set_shuffle_seed(pass_id + (
args.random_seed if args.random_seed else 0))
train_batch_id = 0
train_batch_time_record = []
train_batch_metrics_record = []
if not args.use_dali:
train_iter = train_data_loader()
if args.validate:
test_iter = test_data_loader()
t1 = time.time()
for batch in train_iter:
#NOTE: this is for benchmark
if args.max_iter and total_batch_num == args.max_iter:
return
train_batch_metrics = exe.run(compiled_train_prog,
feed=batch,
fetch_list=train_fetch_list)
t2 = time.time()
train_batch_elapse = t2 - t1
train_batch_time_record.append(train_batch_elapse)
train_batch_metrics_avg = np.mean(
np.array(train_batch_metrics), axis=1)
train_batch_metrics_record.append(train_batch_metrics_avg)
if trainer_id == 0:
print_info("batch", train_batch_metrics_avg, train_batch_elapse,
pass_id, train_batch_id, args.print_step)
sys.stdout.flush()
train_batch_id += 1
t1 = time.time()
#NOTE: this for benchmark profiler
total_batch_num = total_batch_num + 1
if args.is_profiler and pass_id == 0 and train_batch_id == args.print_step:
profiler.start_profiler("All")
elif args.is_profiler and pass_id == 0 and train_batch_id == args.print_step + 5:
profiler.stop_profiler("total", args.profiler_path)
return
if args.use_dali:
train_iter.reset()
if trainer_id == 0 and args.validate:
if args.use_ema:
logger.info('ExponentialMovingAverage validate start...')
with ema.apply(exe):
validate(args, test_iter, exe, test_prog, test_fetch_list,
pass_id, train_batch_metrics_record,
compiled_train_prog)
logger.info('ExponentialMovingAverage validate over!')
validate(args, test_iter, exe, test_prog, test_fetch_list, pass_id,
train_batch_metrics_record, train_batch_time_record,
compiled_train_prog)
if args.use_dali:
test_iter.reset()
if pass_id % args.save_step == 0:
save_model(args, exe, train_prog, pass_id)
def train_resnet():
epoch = args.epoch
if not args.use_gpu:
place = fluid.CPUPlace()
elif not args.use_data_parallel:
place = fluid.CUDAPlace(0)
else:
place = fluid.CUDAPlace(fluid.dygraph.parallel.Env().dev_id)
with fluid.dygraph.guard(place):
if args.ce:
print("ce mode")
seed = 33
np.random.seed(seed)
fluid.default_startup_program().random_seed = seed
fluid.default_main_program().random_seed = seed
if args.use_data_parallel:
strategy = fluid.dygraph.parallel.prepare_context()
resnet = ResNet(class_dim=args.class_dim)
optimizer = optimizer_setting(parameter_list=resnet.parameters())
if args.use_data_parallel:
resnet = fluid.dygraph.parallel.DataParallel(resnet, strategy)
if args.use_imagenet_data:
imagenet_reader = reader.ImageNetReader(0)
train_reader = imagenet_reader.train(settings=args)
else:
train_reader = paddle.batch(reader_decorator(
paddle.dataset.flowers.train(use_xmap=True)),
batch_size=batch_size,
drop_last=True)
if args.use_imagenet_data:
test_reader = imagenet_reader.val(settings=args)
else:
test_reader = paddle.batch(reader_decorator(
paddle.dataset.flowers.test(use_xmap=True)),
batch_size=batch_size,
drop_last=True)
train_loader = fluid.io.DataLoader.from_generator(
capacity=32,
use_double_buffer=True,
iterable=True,
return_list=True,
use_multiprocess=True)
train_loader.set_sample_list_generator(train_reader, places=place)
test_loader = fluid.io.DataLoader.from_generator(
capacity=64,
use_double_buffer=True,
iterable=True,
return_list=True,
use_multiprocess=True)
test_loader.set_sample_list_generator(test_reader, places=place)
#NOTE: used in benchmark
total_batch_num = 0
for eop in range(epoch):
epoch_start = time.time()
resnet.train()
total_loss = 0.0
total_acc1 = 0.0
total_acc5 = 0.0
total_sample = 0
train_batch_cost_avg = TimeCostAverage()
train_reader_cost_avg = TimeCostAverage()
batch_start = time.time()
for batch_id, data in enumerate(train_loader()):
#NOTE: used in benchmark
if args.max_iter and total_batch_num == args.max_iter:
return
train_reader_cost = time.time() - batch_start
img, label = data
label.stop_gradient = True
out = resnet(img)
loss = fluid.layers.cross_entropy(input=out, label=label)
avg_loss = fluid.layers.mean(x=loss)
acc_top1 = fluid.layers.accuracy(input=out, label=label, k=1)
acc_top5 = fluid.layers.accuracy(input=out, label=label, k=5)
dy_out = avg_loss.numpy()
if args.use_data_parallel:
avg_loss = resnet.scale_loss(avg_loss)
avg_loss.backward()
resnet.apply_collective_grads()
else:
avg_loss.backward()
optimizer.minimize(avg_loss)
resnet.clear_gradients()
total_loss += dy_out
total_acc1 += acc_top1.numpy()
total_acc5 += acc_top5.numpy()
total_sample += 1
train_batch_cost = time.time() - batch_start
train_batch_cost_avg.record(train_batch_cost)
train_reader_cost_avg.record(train_reader_cost)
total_batch_num = total_batch_num + 1 #this is for benchmark
if batch_id % 10 == 0:
ips = float(
args.batch_size) / train_batch_cost_avg.get_average()
print(
"[Epoch %d, batch %d] loss: %.5f, acc1: %.5f, acc5: %.5f, batch_cost: %.5f sec, reader_cost: %.5f sec, ips: %.5f images/sec"
% (eop, batch_id, total_loss / total_sample,
total_acc1 / total_sample, total_acc5 /
total_sample, train_batch_cost_avg.get_average(),
train_reader_cost_avg.get_average(), ips))
train_batch_cost_avg.reset()
train_reader_cost_avg.reset()
batch_start = time.time()
if args.ce:
print("kpis\ttrain_acc1\t%0.3f" % (total_acc1 / total_sample))
print("kpis\ttrain_acc5\t%0.3f" % (total_acc5 / total_sample))
print("kpis\ttrain_loss\t%0.3f" % (total_loss / total_sample))
train_epoch_cost = time.time() - epoch_start
print(
"[Epoch %d], loss %.5f, acc1 %.5f, acc5 %.5f, epoch_cost: %.5f s"
% (eop, total_loss / total_sample, total_acc1 / total_sample,
total_acc5 / total_sample, train_epoch_cost))
resnet.eval()
eval(resnet, test_loader)
save_parameters = (not args.use_data_parallel) or (
args.use_data_parallel
and fluid.dygraph.parallel.Env().local_rank == 0)
if save_parameters:
fluid.save_dygraph(resnet.state_dict(), 'resnet_params')
def infer(args):
model_list = [m for m in dir(models) if "__" not in m]
assert args.model in model_list, "{} is not in lists: {}".format(
args.model, model_list)
assert os.path.isdir(args.pretrained_model
), "please load right pretrained model path for infer"
assert args.image_shape[
1] <= args.resize_short_size, "Please check the args:image_shape and args:resize_short_size, The croped size(image_shape[1]) must smaller than or equal to the resized length(resize_short_size) "
if args.image_path:
assert os.path.isfile(
args.image_path
), "Please check the args:image_path, it should be a path to single image."
if args.use_gpu:
assert fluid.core.get_cuda_device_count(
) == 1, "please set \"export CUDA_VISIBLE_DEVICES=\" available single card"
else:
assert int(os.environ.get('CPU_NUM',
1)) == 1, "please set CPU_NUM as 1"
image = fluid.data(name='image',
shape=[None] + args.image_shape,
dtype='float32')
if args.model.startswith('EfficientNet'):
model = models.__dict__[args.model](is_test=True,
padding_type=args.padding_type,
use_se=args.use_se)
else:
model = models.__dict__[args.model]()
if args.model == "GoogLeNet":
out, _, _ = model.net(input=image, class_dim=args.class_dim)
else:
out = model.net(input=image, class_dim=args.class_dim)
out = fluid.layers.softmax(out)
test_program = fluid.default_main_program().clone(for_test=True)
fetch_list = [out.name]
gpu_id = int(os.environ.get('FLAGS_selected_gpus', 0))
place = fluid.CUDAPlace(gpu_id) if args.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
if args.use_gpu:
places = fluid.framework.cuda_places()
else:
places = fluid.framework.cpu_places()
compiled_program = fluid.compiler.CompiledProgram(
test_program).with_data_parallel(places=places)
fluid.io.load_persistables(exe, args.pretrained_model)
if args.save_inference:
fluid.io.save_inference_model(dirname=args.model,
feeded_var_names=['image'],
main_program=test_program,
target_vars=out,
executor=exe,
model_filename='model',
params_filename='params')
logger.info("model: {0} is already saved".format(args.model))
exit(0)
imagenet_reader = reader.ImageNetReader()
test_reader = imagenet_reader.test(settings=args)
feeder = fluid.DataFeeder(place=places, feed_list=[image])
TOPK = args.topk
if os.path.exists(args.class_map_path):
logger.info(
"The map of readable label and numerical label has been found!")
with open(args.class_map_path) as f:
label_dict = {}
strinfo = re.compile(r"\d+ ")
for item in f.readlines():
key = item.split(" ")[0]
value = [
strinfo.sub("", l).replace("\n", "")
for l in item.split(", ")
]
label_dict[key] = value
info = {}
parallel_data = []
parallel_id = []
place_num = paddle.fluid.core.get_cuda_device_count(
) if args.use_gpu else int(os.environ.get('CPU_NUM', 1))
if os.path.exists(args.save_json_path):
logger.warning("path: {} Already exists! will recover it\n".format(
args.save_json_path))
with open(args.save_json_path, "w") as fout:
for batch_id, data in enumerate(test_reader()):
image_data = [[items[0]] for items in data]
image_id = [items[1] for items in data]
parallel_id.append(image_id)
parallel_data.append(image_data)
if place_num == len(parallel_data):
result = exe.run(
compiled_program,
fetch_list=fetch_list,
feed=list(feeder.feed_parallel(parallel_data, place_num)))
for i, res in enumerate(result[0]):
pred_label = np.argsort(res)[::-1][:TOPK]
real_id = str(np.array(parallel_id).flatten()[i])
_, real_id = os.path.split(real_id)
if os.path.exists(args.class_map_path):
readable_pred_label = []
for label in pred_label:
readable_pred_label.append(label_dict[str(label)])
info[real_id] = {}
info[real_id]['score'], info[real_id]['class'], info[
real_id]['class_name'] = str(res[pred_label]), str(
pred_label), readable_pred_label
else:
info[real_id] = {}
info[real_id]['score'], info[real_id]['class'] = str(
res[pred_label]), str(pred_label)
logger.info("{}, {}".format(real_id, info[real_id]))
sys.stdout.flush()
fout.write(real_id + "\t" + json.dumps(info[real_id]) +
"\n")
parallel_data = []
parallel_id = []
os.remove(".tmp.txt")
def train_mobilenet():
if not args.use_gpu:
place = fluid.CPUPlace()
elif not args.use_data_parallel:
place = fluid.CUDAPlace(0)
else:
place = fluid.CUDAPlace(fluid.dygraph.parallel.Env().dev_id)
with fluid.dygraph.guard(place):
# 1. init net and optimizer
place_num = paddle.fluid.core.get_cuda_device_count(
) if args.use_gpu else int(os.environ.get('CPU_NUM', 1))
if args.ce:
print("ce mode")
seed = 33
np.random.seed(seed)
fluid.default_startup_program().random_seed = seed
fluid.default_main_program().random_seed = seed
if args.use_data_parallel:
strategy = fluid.dygraph.parallel.prepare_context()
if args.model == "MobileNetV1":
net = MobileNetV1(class_dim=args.class_dim, scale=1.0)
model_path_pre = 'mobilenet_v1'
elif args.model == "MobileNetV2":
net = MobileNetV2(class_dim=args.class_dim, scale=1.0)
model_path_pre = 'mobilenet_v2'
else:
print(
"wrong model name, please try model = MobileNetV1 or MobileNetV2"
)
exit()
optimizer = create_optimizer(args=args,
parameter_list=net.parameters())
if args.use_data_parallel:
net = fluid.dygraph.parallel.DataParallel(net, strategy)
# 2. load checkpoint
if args.checkpoint:
assert os.path.exists(args.checkpoint + ".pdparams"), \
"Given dir {}.pdparams not exist.".format(args.checkpoint)
assert os.path.exists(args.checkpoint + ".pdopt"), \
"Given dir {}.pdopt not exist.".format(args.checkpoint)
para_dict, opti_dict = fluid.dygraph.load_dygraph(args.checkpoint)
net.set_dict(para_dict)
optimizer.set_dict(opti_dict)
# 3. reader
test_data_loader = utility.create_data_loader(is_train=False,
args=args)
num_trainers = int(os.environ.get('PADDLE_TRAINERS_NUM', 1))
imagenet_reader = reader.ImageNetReader(seed=0, place_num=place_num)
train_dataset = ImageNetDataset(os.path.join(args.data_dir, "train"),
mode='train')
train_data_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
places=place,
shuffle=True,
drop_last=True,
num_workers=10)
test_dataset = ImageNetDataset(os.path.join(args.data_dir, "val"),
mode='val')
test_data_loader = DataLoader(test_dataset,
batch_size=args.batch_size,
places=place,
shuffle=True,
drop_last=True,
num_workers=1)
# 4. train loop
total_batch_num = 0 #this is for benchmark
for eop in range(args.num_epochs):
epoch_start = time.time()
if num_trainers > 1:
imagenet_reader.set_shuffle_seed(
eop + (args.random_seed if args.random_seed else 0))
net.train()
total_loss = 0.0
total_acc1 = 0.0
total_acc5 = 0.0
total_sample = 0
batch_id = 0
t_last = 0
# 4.1 for each batch, call net() , backward(), and minimize()
batch_cost_avg = TimeCostAverage()
batch_reader_avg = TimeCostAverage()
batch_net_avg = TimeCostAverage()
batch_backward_avg = TimeCostAverage()
batch_start = time.time()
for img, label in train_data_loader():
if args.max_iter and total_batch_num == args.max_iter:
return
batch_reader_end = time.time()
# 4.1.1 call net()
out = net(img)
softmax_out = fluid.layers.softmax(out, use_cudnn=False)
loss = fluid.layers.cross_entropy(input=softmax_out,
label=label)
avg_loss = fluid.layers.mean(x=loss)
acc_top1 = fluid.layers.accuracy(input=out, label=label, k=1)
acc_top5 = fluid.layers.accuracy(input=out, label=label, k=5)
batch_net_end = time.time()
# 4.1.2 call backward()
if args.use_data_parallel:
avg_loss = net.scale_loss(avg_loss)
avg_loss.backward()
net.apply_collective_grads()
else:
avg_loss.backward()
batch_backward_end = time.time()
# 4.1.3 call minimize()
optimizer.minimize(avg_loss)
net.clear_gradients()
t2 = time.time()
avg_loss_value = avg_loss.numpy()
acc_top1_value = acc_top1.numpy()
acc_top5_value = acc_top5.numpy()
total_loss += avg_loss_value
total_acc1 += acc_top1_value
total_acc5 += acc_top5_value
total_sample += 1
batch_id += 1
# NOTE: used for benchmark
train_batch_cost = time.time() - batch_start
batch_cost_avg.record(train_batch_cost)
batch_reader_avg.record(batch_reader_end - batch_start)
batch_net_avg.record(batch_net_end - batch_reader_end)
batch_backward_avg.record(batch_backward_end - batch_net_end)
total_batch_num = total_batch_num + 1
if batch_id % args.print_step == 0:
ips = float(args.batch_size) / batch_cost_avg.get_average()
print(
"[Epoch %d, batch %d], avg_loss %.5f, acc_top1 %.5f, acc_top5 %.5f, batch_cost: %.5f sec, net_cost: %.5f sec, backward_cost: %.5f sec, reader_cost: %.5f sec, ips: %.5f images/sec"
% (eop, batch_id, avg_loss_value, acc_top1_value,
acc_top5_value, batch_cost_avg.get_average(),
batch_net_avg.get_average(),
batch_backward_avg.get_average(),
batch_reader_avg.get_average(), ips))
sys.stdout.flush()
batch_cost_avg.reset()
batch_net_avg.reset()
batch_backward_avg.reset()
batch_reader_avg.reset()
batch_start = time.time()
if args.ce:
print("kpis\ttrain_acc1\t%0.3f" % (total_acc1 / total_sample))
print("kpis\ttrain_acc5\t%0.3f" % (total_acc5 / total_sample))
print("kpis\ttrain_loss\t%0.3f" % (total_loss / total_sample))
train_epoch_cost = time.time() - epoch_start
print(
"[Epoch %d], loss %.5f, acc1 %.5f, acc5 %.5f, epoch_cost: %.5f s"
% (eop, total_loss / total_sample, total_acc1 / total_sample,
total_acc5 / total_sample, train_epoch_cost))
# 4.2 save checkpoint
save_parameters = (not args.use_data_parallel) or (
args.use_data_parallel
and fluid.dygraph.parallel.Env().local_rank == 0)
if save_parameters:
if not os.path.isdir(args.model_save_dir):
os.makedirs(args.model_save_dir)
model_path = os.path.join(
args.model_save_dir,
"_" + model_path_pre + "_epoch{}".format(eop))
fluid.dygraph.save_dygraph(net.state_dict(), model_path)
fluid.dygraph.save_dygraph(optimizer.state_dict(), model_path)
# 4.3 validation
net.eval()
eval(net, test_data_loader, eop)
# 5. save final results
save_parameters = (not args.use_data_parallel) or (
args.use_data_parallel
and fluid.dygraph.parallel.Env().local_rank == 0)
if save_parameters:
model_path = os.path.join(args.model_save_dir,
"_" + model_path_pre + "_final")
fluid.dygraph.save_dygraph(net.state_dict(), model_path)
def train_mobilenet():
if not args.use_gpu:
place = fluid.CPUPlace()
elif not args.use_data_parallel:
place = fluid.CUDAPlace(0)
else:
place = fluid.CUDAPlace(fluid.dygraph.parallel.Env().dev_id)
with fluid.dygraph.guard(place):
# 1. init net and optimizer
place_num = paddle.fluid.core.get_cuda_device_count(
) if args.use_gpu else int(os.environ.get('CPU_NUM', 1))
if args.ce:
print("ce mode")
seed = 33
np.random.seed(seed)
fluid.default_startup_program().random_seed = seed
fluid.default_main_program().random_seed = seed
if args.use_data_parallel:
strategy = fluid.dygraph.parallel.prepare_context()
if args.model == "MobileNetV1":
net = MobileNetV1(class_dim=args.class_dim, scale=1.0)
model_path_pre = 'mobilenet_v1'
elif args.model == "MobileNetV2":
net = MobileNetV2(class_dim=args.class_dim, scale=1.0)
model_path_pre = 'mobilenet_v2'
else:
print(
"wrong model name, please try model = MobileNetV1 or MobileNetV2"
)
exit()
optimizer = create_optimizer(args=args,
parameter_list=net.parameters())
if args.use_data_parallel:
net = fluid.dygraph.parallel.DataParallel(net, strategy)
# 2. load checkpoint
if args.checkpoint:
assert os.path.exists(args.checkpoint + ".pdparams"), \
"Given dir {}.pdparams not exist.".format(args.checkpoint)
assert os.path.exists(args.checkpoint + ".pdopt"), \
"Given dir {}.pdopt not exist.".format(args.checkpoint)
para_dict, opti_dict = fluid.dygraph.load_dygraph(args.checkpoint)
net.set_dict(para_dict)
optimizer.set_dict(opti_dict)
# 3. reader
train_data_loader, train_data = utility.create_data_loader(
is_train=True, args=args)
test_data_loader, test_data = utility.create_data_loader(
is_train=False, args=args)
num_trainers = int(os.environ.get('PADDLE_TRAINERS_NUM', 1))
imagenet_reader = reader.ImageNetReader(seed=0, place_num=place_num)
train_reader = imagenet_reader.train(settings=args)
test_reader = imagenet_reader.val(settings=args)
train_data_loader.set_sample_list_generator(train_reader, place)
test_data_loader.set_sample_list_generator(test_reader, place)
# 4. train loop
for eop in range(args.num_epochs):
if num_trainers > 1:
imagenet_reader.set_shuffle_seed(
eop + (args.random_seed if args.random_seed else 0))
net.train()
total_loss = 0.0
total_acc1 = 0.0
total_acc5 = 0.0
total_sample = 0
batch_id = 0
t_last = 0
# 4.1 for each batch, call net() , backward(), and minimize()
for img, label in train_data_loader():
t1 = time.time()
label = to_variable(label.numpy().astype('int64').reshape(
int(args.batch_size // place_num), 1))
t_start = time.time()
# 4.1.1 call net()
out = net(img)
t_end = time.time()
softmax_out = fluid.layers.softmax(out, use_cudnn=False)
loss = fluid.layers.cross_entropy(input=softmax_out,
label=label)
avg_loss = fluid.layers.mean(x=loss)
acc_top1 = fluid.layers.accuracy(input=out, label=label, k=1)
acc_top5 = fluid.layers.accuracy(input=out, label=label, k=5)
t_start_back = time.time()
# 4.1.2 call backward()
if args.use_data_parallel:
avg_loss = net.scale_loss(avg_loss)
avg_loss.backward()
net.apply_collective_grads()
else:
avg_loss.backward()
t_end_back = time.time()
# 4.1.3 call minimize()
optimizer.minimize(avg_loss)
net.clear_gradients()
t2 = time.time()
train_batch_elapse = t2 - t1
if batch_id % args.print_step == 0:
print( "epoch id: %d, batch step: %d, avg_loss %0.5f acc_top1 %0.5f acc_top5 %0.5f %2.4f sec net_t:%2.4f back_t:%2.4f read_t:%2.4f" % \
(eop, batch_id, avg_loss.numpy(), acc_top1.numpy(), acc_top5.numpy(), train_batch_elapse,
t_end - t_start, t_end_back - t_start_back, t1 - t_last))
sys.stdout.flush()
total_loss += avg_loss.numpy()
total_acc1 += acc_top1.numpy()
total_acc5 += acc_top5.numpy()
total_sample += 1
batch_id += 1
t_last = time.time()
if args.ce:
print("kpis\ttrain_acc1\t%0.3f" % (total_acc1 / total_sample))
print("kpis\ttrain_acc5\t%0.3f" % (total_acc5 / total_sample))
print("kpis\ttrain_loss\t%0.3f" % (total_loss / total_sample))
print("epoch %d | batch step %d, loss %0.3f acc1 %0.3f acc5 %0.3f %2.4f sec" % \
(eop, batch_id, total_loss / total_sample, \
total_acc1 / total_sample, total_acc5 / total_sample, train_batch_elapse))
# 4.2 save checkpoint
save_parameters = (not args.use_data_parallel) or (
args.use_data_parallel
and fluid.dygraph.parallel.Env().local_rank == 0)
if save_parameters:
if not os.path.isdir(args.model_save_dir):
os.makedirs(args.model_save_dir)
model_path = os.path.join(
args.model_save_dir,
"_" + model_path_pre + "_epoch{}".format(eop))
fluid.dygraph.save_dygraph(net.state_dict(), model_path)
fluid.dygraph.save_dygraph(optimizer.state_dict(), model_path)
# 4.3 validation
net.eval()
eval(net, test_data_loader, eop)
# 5. save final results
save_parameters = (not args.use_data_parallel) or (
args.use_data_parallel
and fluid.dygraph.parallel.Env().local_rank == 0)
if save_parameters:
model_path = os.path.join(args.model_save_dir,
"_" + model_path_pre + "_final")
fluid.dygraph.save_dygraph(net.state_dict(), model_path)