def test_prune(self):
main_program = fluid.Program()
startup_program = fluid.Program()
# X X O X O
# conv1-->conv2-->sum1-->conv3-->conv4-->sum2-->conv5-->conv6
# | ^ | ^
# |____________| |____________________|
#
# X: prune output channels
# O: prune input channels
with fluid.unique_name.guard():
with fluid.program_guard(main_program, startup_program):
input = fluid.data(name="image", shape=[None, 3, 16, 16])
label = fluid.data(name='label',
shape=[None, 1],
dtype='int64')
conv1 = conv_bn_layer(input, 8, 3, "conv1", act='relu')
conv2 = conv_bn_layer(conv1, 8, 3, "conv2", act='leaky_relu')
sum1 = conv1 + conv2
conv3 = conv_bn_layer(sum1, 8, 3, "conv3", act='relu6')
conv4 = conv_bn_layer(conv3, 8, 3, "conv4")
sum2 = conv4 + sum1
conv5 = conv_bn_layer(sum2, 8, 3, "conv5")
flag = fluid.layers.fill_constant([1], value=1, dtype='int32')
rand_flag = paddle.randint(2, dtype='int32')
cond = fluid.layers.less_than(x=flag, y=rand_flag)
cond_output = fluid.layers.create_global_var(
shape=[1],
value=0.0,
dtype='float32',
persistable=False,
name='cond_output')
def cond_block1():
cond_conv = conv_bn_layer(conv5, 8, 3, "conv_cond1_1")
fluid.layers.assign(input=cond_conv, output=cond_output)
def cond_block2():
cond_conv1 = conv_bn_layer(conv5, 8, 3, "conv_cond2_1")
cond_conv2 = conv_bn_layer(cond_conv1, 8, 3,
"conv_cond2_2")
fluid.layers.assign(input=cond_conv2, output=cond_output)
fluid.layers.cond(cond, cond_block1, cond_block2)
sum3 = fluid.layers.sum([sum2, cond_output])
conv6 = conv_bn_layer(sum3, 8, 3, "conv6")
sub1 = conv6 - sum3
mult = sub1 * sub1
conv7 = conv_bn_layer(mult,
8,
3,
"Depthwise_Conv7",
groups=8,
use_cudnn=False)
floored = fluid.layers.floor(conv7)
scaled = fluid.layers.scale(floored)
concated = fluid.layers.concat([scaled, mult], axis=1)
conv8 = conv_bn_layer(concated, 8, 3, "conv8")
predict = fluid.layers.fc(input=conv8, size=10, act='softmax')
cost = fluid.layers.cross_entropy(input=predict, label=label)
adam_optimizer = fluid.optimizer.AdamOptimizer(0.01)
avg_cost = fluid.layers.mean(cost)
adam_optimizer.minimize(avg_cost)
params = []
for param in main_program.all_parameters():
if 'conv' in param.name:
params.append(param.name)
#TODO: To support pruning convolution before fc layer.
params.remove('conv8_weights')
place = fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(startup_program)
x = np.random.random(size=(10, 3, 16, 16)).astype('float32')
label = np.random.random(size=(10, 1)).astype('int64')
loss_data, = exe.run(main_program,
feed={
"image": x,
"label": label
},
fetch_list=[cost.name])
pruner = Pruner()
main_program, _, _ = pruner.prune(main_program,
fluid.global_scope(),
params=params,
ratios=[0.5] * len(params),
place=place,
lazy=False,
only_graph=False,
param_backup=None,
param_shape_backup=None)
loss_data, = exe.run(main_program,
feed={
"image": x,
"label": label
},
fetch_list=[cost.name])
def fast_infer(args):
"""
Inference by beam search decoder based solely on Fluid operators.
"""
out_ids, out_scores, pyreader = fast_decoder(
ModelHyperParams.src_vocab_size,
ModelHyperParams.trg_vocab_size,
ModelHyperParams.phone_vocab_size,
ModelHyperParams.max_length + 1,
ModelHyperParams.n_layer,
ModelHyperParams.n_head,
ModelHyperParams.d_key,
ModelHyperParams.d_value,
ModelHyperParams.d_model,
ModelHyperParams.d_inner_hid,
ModelHyperParams.prepostprocess_dropout,
ModelHyperParams.attention_dropout,
ModelHyperParams.relu_dropout,
ModelHyperParams.preprocess_cmd,
ModelHyperParams.postprocess_cmd,
ModelHyperParams.weight_sharing,
InferTaskConfig.beam_size,
InferTaskConfig.max_out_len,
ModelHyperParams.bos_idx,
ModelHyperParams.eos_idx,
beta=ModelHyperParams.beta,
use_py_reader=args.use_py_reader)
# This is used here to set dropout to the test mode.
infer_program = fluid.default_main_program().clone(for_test=True)
if InferTaskConfig.use_gpu:
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()))
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
fluid.io.load_vars(exe,
InferTaskConfig.model_path,
vars=[
var for var in infer_program.list_vars()
if isinstance(var, fluid.framework.Parameter)
])
exec_strategy = fluid.ExecutionStrategy()
# For faster executor
exec_strategy.use_experimental_executor = True
exec_strategy.num_threads = 1
build_strategy = fluid.BuildStrategy()
infer_exe = fluid.ParallelExecutor(use_cuda=TrainTaskConfig.use_gpu,
main_program=infer_program,
build_strategy=build_strategy,
exec_strategy=exec_strategy)
# data reader settings for inference
args.train_file_pattern = args.test_file_pattern
args.use_token_batch = False
args.sort_type = reader.SortType.NONE
args.shuffle = False
args.shuffle_batch = False
test_data = prepare_data_generator(
args,
is_test=False,
count=dev_count,
pyreader=pyreader,
py_reader_provider_wrapper=py_reader_provider_wrapper,
place=place)
if args.use_py_reader:
pyreader.start()
data_generator = None
else:
data_generator = test_data()
trg_idx2word = reader.DataReader.load_dict(dict_path=args.trg_vocab_fpath,
reverse=True)
while True:
try:
feed_dict_list = prepare_feed_dict_list(data_generator, dev_count,
place)
if args.use_parallel_exe:
seq_ids, seq_scores = infer_exe.run(
fetch_list=[out_ids.name, out_scores.name],
feed=feed_dict_list,
return_numpy=False)
else:
seq_ids, seq_scores = exe.run(
program=infer_program,
fetch_list=[out_ids.name, out_scores.name],
feed=feed_dict_list[0]
if feed_dict_list is not None else None,
return_numpy=False,
use_program_cache=True)
seq_ids_list, seq_scores_list = [
seq_ids
], [seq_scores] if isinstance(
seq_ids, paddle.fluid.LoDTensor) else (seq_ids, seq_scores)
for seq_ids, seq_scores in zip(seq_ids_list, seq_scores_list):
# How to parse the results:
# Suppose the lod of seq_ids is:
# [[0, 3, 6], [0, 12, 24, 40, 54, 67, 82]]
# then from lod[0]:
# there are 2 source sentences, beam width is 3.
# from lod[1]:
# the first source sentence has 3 hyps; the lengths are 12, 12, 16
# the second source sentence has 3 hyps; the lengths are 14, 13, 15
hyps = [[] for i in range(len(seq_ids.lod()[0]) - 1)]
scores = [[] for i in range(len(seq_scores.lod()[0]) - 1)]
for i in range(len(seq_ids.lod()[0]) -
1): # for each source sentence
start = seq_ids.lod()[0][i]
end = seq_ids.lod()[0][i + 1]
for j in range(end - start): # for each candidate
sub_start = seq_ids.lod()[1][start + j]
sub_end = seq_ids.lod()[1][start + j + 1]
hyps[i].append(" ".join([
trg_idx2word[idx] for idx in post_process_seq(
np.array(seq_ids)[sub_start:sub_end])
]))
scores[i].append(np.array(seq_scores)[sub_end - 1])
print(hyps[i][-1])
if len(hyps[i]) >= InferTaskConfig.n_best:
break
except (StopIteration, fluid.core.EOFException):
# The data pass is over.
if args.use_py_reader:
pyreader.reset()
break
def do_train(args):
train_prog = fluid.default_main_program()
startup_prog = fluid.default_startup_program()
with fluid.program_guard(train_prog, startup_prog):
train_prog.random_seed = args.random_seed
startup_prog.random_seed = args.random_seed
with fluid.unique_name.guard():
# define input and reader
input_slots = [{
"name": "src_ids",
"shape": (-1, args.max_seq_len, 1),
"dtype": "int64"
}, {
"name": "pos_ids",
"shape": (-1, args.max_seq_len, 1),
"dtype": "int64"
}, {
"name": "sent_ids",
"shape": (-1, args.max_seq_len, 1),
"dtype": "int64"
}, {
"name": "input_mask",
"shape": (-1, args.max_seq_len, 1),
"dtype": "float32"
}, {
"name": "input_span_mask",
"shape": (-1, args.max_seq_len),
"dtype": "float32"
}, {
"name": "start_positions",
"shape": (-1, 1),
"dtype": "int64"
}, {
"name": "end_positions",
"shape": (-1, 1),
"dtype": "int64"
}, {
"name": "is_null_answer",
"shape": (-1, 1),
"dtype": "int64"
}]
input_field = InputField(input_slots)
input_field.build(build_pyreader=True)
# define the network
loss = create_net(is_training=True,
model_input=input_field,
args=args)
loss.persistable = True
# define the optimizer
if args.use_cuda:
dev_count = fluid.core.get_cuda_device_count()
else:
dev_count = int(
os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
# as we need to get the max training steps for warmup training,
# we define the data processer in advance
# usually, we can declare data processor later, outsides the program_gurad scope
processor = DataProcessor(vocab_path=args.vocab_path,
do_lower_case=args.do_lower_case,
max_seq_length=args.max_seq_len,
in_tokens=args.in_tokens,
doc_stride=args.doc_stride,
do_stride=args.do_stride,
max_query_length=args.max_query_len)
## define the data generator
batch_generator = processor.data_generator(
data_path=args.training_file,
batch_size=args.batch_size,
phase="train",
shuffle=True,
dev_count=dev_count,
epoch=args.epoch)
num_train_examples = processor.get_num_examples(phase='train')
max_train_steps = args.epoch * num_train_examples // dev_count // args.batch_size
warmup_steps = int(max_train_steps * args.warmup_proportion)
print(max_train_steps, warmup_steps, num_train_examples)
optimizor = optimization(loss=loss,
warmup_steps=warmup_steps,
num_train_steps=max_train_steps,
learning_rate=args.learning_rate,
train_program=train_prog,
startup_prog=startup_prog,
weight_decay=args.weight_decay,
scheduler=args.lr_scheduler,
use_fp16=args.use_fp16,
loss_scaling=args.loss_scaling)
# prepare training
## decorate the pyreader with batch_generator
input_field.reader.decorate_batch_generator(batch_generator)
## define the executor and program for training
if args.use_cuda:
place = fluid.CUDAPlace(0)
else:
place = fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(startup_prog)
assert (args.init_from_checkpoint == "") or (args.init_from_pretrain_model
== "")
## init from some checkpoint, to resume the previous training
if args.init_from_checkpoint:
init_from_checkpoint(args, exe, train_prog)
## init from some pretrain models, to better solve the current task
if args.init_from_pretrain_model:
init_from_pretrain_model(args, exe, train_prog)
build_strategy = fluid.compiler.BuildStrategy()
build_strategy.enable_inplace = True
compiled_train_prog = fluid.CompiledProgram(train_prog).with_data_parallel(
loss_name=loss.name, build_strategy=build_strategy)
# start training
step = 0
for epoch_step in range(args.epoch):
input_field.reader.start()
while True:
try:
# this is for minimizing the fetching op, saving the training speed.
if step % args.print_step == 0:
fetch_list = [loss.name]
else:
fetch_list = []
output = exe.run(compiled_train_prog, fetch_list=fetch_list)
if step % args.print_step == 0:
print("step: %d, loss: %.4f" % (step, np.sum(output[0])))
if step % args.save_step == 0 and step != 0:
if args.save_checkpoint:
save_checkpoint(args, exe, train_prog,
"step_" + str(step))
if args.save_param:
save_param(args, exe, train_prog, "step_" + str(step))
step += 1
except fluid.core.EOFException:
input_field.reader.reset()
break
if args.save_checkpoint:
save_checkpoint(args, exe, train_prog, "step_final")
if args.save_param:
save_param(args, exe, train_prog, "step_final")
def fit():
role = role_maker.UserDefinedRoleMaker(
current_id=current_id,
role=role_maker.Role.WORKER if bool(1==int(roles)) else role_maker.Role.SERVER,
worker_num=2,
server_endpoints=["127.0.0.1:36011"])
fleet.init(role)
BATCH_SIZE = 128
type_size=createDataList(model_file_path,model_file_path+'.data'+"/")
# 用于训练的数据提供器
train_reader=paddle.batch(reader=paddle.reader.shuffle(reader=dataReader(in_file_path+".data/trainer.list"),buf_size=BATCH_SIZE*100), batch_size=BATCH_SIZE)
test_reader=paddle.batch(reader=paddle.reader.shuffle(reader=dataReader(in_file_path+".data/test.list"),buf_size=BATCH_SIZE*100), batch_size=BATCH_SIZE)
data_shape = [3, 32, 32]
images = fluid.layers.data(name='images', shape=data_shape, dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
# 获取分类器
predict = networkConfiguration(images,type_size)
# 定义损失函数和准确率
cost = fluid.layers.cross_entropy(input=predict, label=label) # 交叉熵
avg_cost = fluid.layers.mean(cost) # 计算cost中所有元素的平均值
acc = fluid.layers.accuracy(input=predict, label=label) # 使用输入和标签计算准确率
# 定义优化方法
test_program = fluid.default_main_program().clone(for_test=True) # 获取测试程序
optimizer = fluid.optimizer.Adam(learning_rate=0.001)
strategy = DistributeTranspilerConfig()
strategy.sync_mode = True
optimizer = fleet.distributed_optimizer(optimizer,strategy)
# 定义优化方法
optimizer.minimize(avg_cost)
if fleet.is_server():
print("启动server")
fleet.init_server()
fleet.run_server()
elif fleet.is_worker():
print("启动worker")
########## 模型训练＆模型评估 ##########
# 创建Executor
use_cuda = False # 定义使用CPU还是GPU，使用CPU时use_cuda=False
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
print("cpu")
# 定义数据映射器
feeder = fluid.DataFeeder(feed_list=[images, label], place=place)
print("数据映射")
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
fleet.init_worker()
print(fleet.worker_endpoints())
for pass_id in range(EPOCH_NUM):
print(pass_id)
# 开始训练
for batch_id, data in enumerate(train_reader()): # 遍历train_reader
train_cost, train_acc = exe.run(program=fluid.default_main_program(), # 运行主程序
feed=feeder.feed(data), # 喂入一个batch的数据
fetch_list=[avg_cost, acc]) # fetch均方误差和准确率 # fetch均方误差和准确率
# 每100次batch打印一次训练、进行一次测试
if batch_id % 20 == 0:
print('Pass:%d, Batch:%d, Cost:%0.5f, Accuracy:%0.5f' %(pass_id, batch_id, train_cost[0], train_acc[0]))
# 开始测试
test_costs = [] # 测试的损失值
test_accs = [] # 测试的准确率
for batch_id, data in enumerate(test_reader()):
test_cost, test_acc = exe.run(program=test_program, # 执行训练程序
feed=feeder.feed(data), # 喂入数据
fetch_list=[avg_cost, acc]) # fetch误差、准确率
test_costs.append(test_cost[0]) # 记录每个batch的损失值
test_accs.append(test_acc[0]) # 记录每个batch的准确率
test_cost = (sum(test_costs) / len(test_costs)) # 计算误差平均值
test_acc = (sum(test_accs) / len(test_accs)) # 计算准确率平均值
print('Test:%d, Cost:%0.5f, ACC:%0.5f' % (pass_id, test_cost, test_acc))
save(predict,model_file_path,exe)
fleet.stop_worker()
def eval():
args = parse_args()
print_arguments(args)
# check whether the installed paddle is compiled with GPU
# PointRCNN model can only run on GPU
check_gpu(True)
load_config(args.cfg)
if args.set_cfgs is not None:
set_config_from_list(args.set_cfgs)
if not os.path.isdir(args.output_dir):
os.makedirs(args.output_dir)
if args.eval_mode == 'rpn':
cfg.RPN.ENABLED = True
cfg.RCNN.ENABLED = False
elif args.eval_mode == 'rcnn':
cfg.RCNN.ENABLED = True
cfg.RPN.ENABLED = cfg.RPN.FIXED = True
assert args.batch_size, "batch size must be 1 in rcnn evaluation"
elif args.eval_mode == 'rcnn_offline':
cfg.RCNN.ENABLED = True
cfg.RPN.ENABLED = False
assert args.batch_size, "batch size must be 1 in rcnn_offline evaluation"
else:
raise NotImplementedError("unkown eval mode: {}".format(
args.eval_mode))
place = fluid.CUDAPlace(0)
exe = fluid.Executor(place)
# build model
startup = fluid.Program()
eval_prog = fluid.Program()
with fluid.program_guard(eval_prog, startup):
with fluid.unique_name.guard():
eval_model = PointRCNN(cfg, args.batch_size, True, 'TEST')
eval_model.build()
eval_loader = eval_model.get_loader()
eval_feeds = eval_model.get_feeds()
eval_outputs = eval_model.get_outputs()
eval_prog = eval_prog.clone(True)
extra_keys = []
if args.eval_mode == 'rpn':
extra_keys.extend(['sample_id', 'rpn_cls_label', 'gt_boxes3d'])
if args.save_rpn_feature:
extra_keys.extend([
'pts_rect',
'pts_features',
'pts_input',
])
eval_keys, eval_values = parse_outputs(eval_outputs,
prog=eval_prog,
extra_keys=extra_keys)
eval_compile_prog = fluid.compiler.CompiledProgram(
eval_prog).with_data_parallel()
exe.run(startup)
# load weights
if not os.path.isdir(args.weights):
assert os.path.exists("{}.pdparams".format(args.weights)), \
"Given resume weight {}.pdparams not exist.".format(args.weights)
fluid.load(eval_prog, args.weights, exe)
kitti_feature_dir = os.path.join(args.output_dir, 'features')
kitti_output_dir = os.path.join(args.output_dir, 'detections', 'data')
seg_output_dir = os.path.join(args.output_dir, 'seg_result')
if args.save_rpn_feature:
if os.path.exists(kitti_feature_dir):
shutil.rmtree(kitti_feature_dir)
os.makedirs(kitti_feature_dir)
if os.path.exists(kitti_output_dir):
shutil.rmtree(kitti_output_dir)
os.makedirs(kitti_output_dir)
if os.path.exists(seg_output_dir):
shutil.rmtree(seg_output_dir)
os.makedirs(seg_output_dir)
# must make sure these dirs existing
roi_output_dir = os.path.join('./result_dir', 'roi_result', 'data')
refine_output_dir = os.path.join('./result_dir', 'refine_result', 'data')
final_output_dir = os.path.join("./result_dir", 'final_result', 'data')
if not os.path.exists(final_output_dir):
os.makedirs(final_output_dir)
if args.save_result:
if not os.path.exists(roi_output_dir):
os.makedirs(roi_output_dir)
if not os.path.exists(refine_output_dir):
os.makedirs(refine_output_dir)
# get reader
kitti_rcnn_reader = KittiRCNNReader(
data_dir=args.data_dir,
npoints=cfg.RPN.NUM_POINTS,
split=cfg.TEST.SPLIT,
mode='EVAL',
classes=cfg.CLASSES,
rcnn_eval_roi_dir=args.rcnn_eval_roi_dir,
rcnn_eval_feature_dir=args.rcnn_eval_feature_dir)
eval_reader = kitti_rcnn_reader.get_multiprocess_reader(
args.batch_size, eval_feeds)
eval_loader.set_sample_list_generator(eval_reader, place)
thresh_list = [0.1, 0.3, 0.5, 0.7, 0.9]
queue = multiprocessing.Queue(128)
mgr = multiprocessing.Manager()
lock = multiprocessing.Lock()
mdict = mgr.dict()
if cfg.RPN.ENABLED:
mdict['exit_proc'] = 0
mdict['total_gt_bbox'] = 0
mdict['total_cnt'] = 0
mdict['total_rpn_iou'] = 0
for i in range(len(thresh_list)):
mdict['total_recalled_bbox_list_{}'.format(i)] = 0
p_list = []
for i in range(METRIC_PROC_NUM):
p_list.append(
multiprocessing.Process(target=rpn_metric,
args=(queue, mdict, lock, thresh_list,
args.save_rpn_feature,
kitti_feature_dir,
seg_output_dir, kitti_output_dir,
kitti_rcnn_reader, cfg.CLASSES)))
p_list[-1].start()
if cfg.RCNN.ENABLED:
for i in range(len(thresh_list)):
mdict['total_recalled_bbox_list_{}'.format(i)] = 0
mdict['total_roi_recalled_bbox_list_{}'.format(i)] = 0
mdict['exit_proc'] = 0
mdict['total_cls_acc'] = 0
mdict['total_cls_acc_refined'] = 0
mdict['total_det_num'] = 0
mdict['total_gt_bbox'] = 0
p_list = []
for i in range(METRIC_PROC_NUM):
p_list.append(
multiprocessing.Process(
target=rcnn_metric,
args=(queue, mdict, lock, thresh_list, kitti_rcnn_reader,
roi_output_dir, refine_output_dir, final_output_dir,
args.save_result)))
p_list[-1].start()
try:
eval_loader.start()
eval_iter = 0
start_time = time.time()
cur_time = time.time()
while True:
eval_outs = exe.run(eval_compile_prog,
fetch_list=eval_values,
return_numpy=False)
rets_dict = {
k: (np.array(v), v.recursive_sequence_lengths())
for k, v in zip(eval_keys, eval_outs)
}
run_time = time.time() - cur_time
cur_time = time.time()
queue.put(rets_dict)
eval_iter += 1
logger.info("[EVAL] iter {}, time: {:.2f}".format(
eval_iter, run_time))
except fluid.core.EOFException:
# terminate metric process
for i in range(METRIC_PROC_NUM):
queue.put(None)
while mdict['exit_proc'] < METRIC_PROC_NUM:
time.sleep(1)
for p in p_list:
if p.is_alive():
p.join()
end_time = time.time()
logger.info(
"[EVAL] total {} iter finished, average time: {:.2f}".format(
eval_iter, (end_time - start_time) / float(eval_iter)))
if cfg.RPN.ENABLED:
avg_rpn_iou = mdict['total_rpn_iou'] / max(len(kitti_rcnn_reader),
1.)
logger.info("average rpn iou: {:.3f}".format(avg_rpn_iou))
total_gt_bbox = float(max(mdict['total_gt_bbox'], 1.0))
for idx, thresh in enumerate(thresh_list):
recall = mdict['total_recalled_bbox_list_{}'.format(
idx)] / total_gt_bbox
logger.info(
"total bbox recall(thresh={:.3f}): {} / {} = {:.3f}".
format(thresh,
mdict['total_recalled_bbox_list_{}'.format(idx)],
mdict['total_gt_bbox'], recall))
if cfg.RCNN.ENABLED:
cnt = float(max(eval_iter, 1.0))
avg_cls_acc = mdict['total_cls_acc'] / cnt
avg_cls_acc_refined = mdict['total_cls_acc_refined'] / cnt
avg_det_num = mdict['total_det_num'] / cnt
logger.info("avg_cls_acc: {}".format(avg_cls_acc))
logger.info("avg_cls_acc_refined: {}".format(avg_cls_acc_refined))
logger.info("avg_det_num: {}".format(avg_det_num))
total_gt_bbox = float(max(mdict['total_gt_bbox'], 1.0))
for idx, thresh in enumerate(thresh_list):
cur_roi_recall = mdict['total_roi_recalled_bbox_list_{}'.
format(idx)] / total_gt_bbox
logger.info(
'total roi bbox recall(thresh=%.3f): %d / %d = %f' %
(thresh,
mdict['total_roi_recalled_bbox_list_{}'.format(idx)],
total_gt_bbox, cur_roi_recall))
for idx, thresh in enumerate(thresh_list):
cur_recall = mdict['total_recalled_bbox_list_{}'.format(
idx)] / total_gt_bbox
logger.info(
'total bbox recall(thresh=%.2f) %d / %.2f = %.4f' %
(thresh, mdict['total_recalled_bbox_list_{}'.format(idx)],
total_gt_bbox, cur_recall))
split_file = os.path.join('./data/KITTI', 'ImageSets', 'val.txt')
image_idx_list = [x.strip() for x in open(split_file).readlines()]
for k in range(image_idx_list.__len__()):
cur_file = os.path.join(final_output_dir,
'%s.txt' % image_idx_list[k])
if not os.path.exists(cur_file):
with open(cur_file, 'w') as temp_f:
pass
if float(sys.version[:3]) >= 3.6:
label_dir = os.path.join('./data/KITTI/object/training',
'label_2')
split_file = os.path.join('./data/KITTI', 'ImageSets',
'val.txt')
final_output_dir = os.path.join("./result_dir", 'final_result',
'data')
name_to_class = {'Car': 0, 'Pedestrian': 1, 'Cyclist': 2}
from tools.kitti_object_eval_python.evaluate import evaluate as kitti_evaluate
ap_result_str, ap_dict = kitti_evaluate(
label_dir,
final_output_dir,
label_split_file=split_file,
current_class=name_to_class["Car"])
logger.info("KITTI evaluate: {}, {}".format(
ap_result_str, ap_dict))
else:
logger.info(
"KITTI mAP only support python version >= 3.6, users can "
"run 'python3 tools/kitti_eval.py' to evaluate KITTI mAP.")
finally:
eval_loader.reset()
def run_trainer(self, args):
self.lr = args.lr
if args.nccl2_reduce_layer_local_run:
test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \
self.get_model(batch_size=args.batch_size, single_device=True)
elif args.use_dgc:
test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \
self.get_model(batch_size=args.batch_size, use_dgc=args.use_dgc)
else:
test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \
self.get_model(batch_size=args.batch_size)
if args.update_method == "pserver":
print_to_err(
type(self).__name__,
"begin to run transpile on trainer with pserver mode")
t = self.get_transpiler(
trainer_id=args.trainer_id,
main_program=fluid.default_main_program(),
pserver_endpoints=args.endpoints,
trainers=args.trainers,
sync_mode=args.sync_mode,
dc_asgd=args.dc_asgd,
hogwild_mode=args.hogwild)
trainer_prog = t.get_trainer_program()
print_to_err(
type(self).__name__,
"get trainer program done with pserver mode.")
elif args.update_method == "nccl2" or args.update_method == "nccl2_reduce_layer":
# transpile for nccl2
config = fluid.DistributeTranspilerConfig()
config.mode = "nccl2"
config.nccl_comm_num = args.nccl_comm_num
if args.use_hallreduce:
config.use_hierarchical_allreduce = True
config.hierarchical_allreduce_inter_nranks = args.hallreduce_inter_nranks
print_to_err(
type(self).__name__,
"begin to run transpile on trainer with nccl2 mode")
nccl2_t = fluid.DistributeTranspiler(config=config)
nccl2_t.transpile(
args.trainer_id,
program=fluid.default_main_program(),
startup_program=fluid.default_startup_program(),
trainers=args.endpoints,
current_endpoint=args.current_endpoint)
print_to_err(
type(self).__name__,
"get trainer program done. with nccl2 mode")
trainer_prog = fluid.default_main_program()
else:
print_to_err(
type(self).__name__,
"do nothing about main program, just use it")
trainer_prog = fluid.default_main_program()
print_to_err(type(self).__name__, "use main program done.")
if args.use_cuda:
device_id = int(os.getenv("FLAGS_selected_gpus", "0"))
place = fluid.CUDAPlace(device_id)
else:
place = fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
print_to_err(type(self).__name__, "run worker startup program done.")
exec_strategy = fluid.ExecutionStrategy()
exec_strategy.num_threads = 1
build_stra = fluid.BuildStrategy()
# FIXME force disable enable_inplace and memory_optimize
build_stra.enable_inplace = False
build_stra.memory_optimize = False
if args.hogwild:
build_stra.async_mode = True
if args.enable_backward_deps:
build_stra.enable_backward_optimizer_op_deps = True
if args.use_reduce:
build_stra.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
else:
build_stra.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce
pass_builder = None
if args.batch_merge_repeat > 1:
pass_builder = build_stra._finalize_strategy_and_create_passes()
mypass = pass_builder.insert_pass(0, "multi_batch_merge_pass")
mypass.set("num_repeats", args.batch_merge_repeat)
if args.update_method == "nccl2" or args.update_method == "nccl2_reduce_layer":
build_stra.num_trainers = len(args.endpoints.split(","))
build_stra.trainer_id = args.trainer_id
else:
# case args.update_method == "nccl2_reduce_layer":
build_stra.num_trainers = 1
build_stra.trainer_id = 0
print_to_err(type(self).__name__, "begin to compile with data parallel")
binary = compiler.CompiledProgram(trainer_prog).with_data_parallel(
loss_name=avg_cost.name,
build_strategy=build_stra,
exec_strategy=exec_strategy)
print_to_err(type(self).__name__, "program compiled with data parallel")
feed_var_list = [
var for var in trainer_prog.global_block().vars.values()
if var.is_data
]
feeder = fluid.DataFeeder(feed_var_list, place)
reader_generator = train_reader()
def get_data():
origin_batch = next(reader_generator)
if args.update_method != "local" and args.use_reader_alloc:
new_batch = []
for offset, item in enumerate(origin_batch):
if offset % 2 == args.trainer_id:
new_batch.append(item)
return new_batch
else:
return origin_batch
print_to_err(type(self).__name__, "begin to train on trainer")
out_losses = []
for i in six.moves.xrange(RUN_STEP):
loss, = exe.run(binary,
fetch_list=[avg_cost.name],
feed=feeder.feed(get_data()))
out_losses.append(loss[0])
print_to_err(type(self).__name__, "run step %d finished" % i)
print_to_err(type(self).__name__, "trainer run finished")
print_to_out(out_losses)
def compress(args):
############################################################################################################
# 1. quantization configs
############################################################################################################
quant_config = {
# weight quantize type, default is 'channel_wise_abs_max'
'weight_quantize_type': 'channel_wise_abs_max',
# activation quantize type, default is 'moving_average_abs_max'
'activation_quantize_type': 'moving_average_abs_max',
# weight quantize bit num, default is 8
'weight_bits': 8,
# activation quantize bit num, default is 8
'activation_bits': 8,
# ops of name_scope in not_quant_pattern list, will not be quantized
'not_quant_pattern': ['skip_quant'],
# ops of type in quantize_op_types, will be quantized
'quantize_op_types': ['conv2d', 'depthwise_conv2d', 'mul'],
# data type after quantization, such as 'uint8', 'int8', etc. default is 'int8'
'dtype': 'int8',
# window size for 'range_abs_max' quantization. defaulf is 10000
'window_size': 10000,
# The decay coefficient of moving average, default is 0.9
'moving_rate': 0.9,
}
train_reader = None
test_reader = None
if args.data == "mnist":
import paddle.dataset.mnist as reader
train_reader = reader.train()
val_reader = reader.test()
class_dim = 10
image_shape = "1,28,28"
elif args.data == "imagenet":
import imagenet_reader as reader
train_reader = reader.train()
val_reader = reader.val()
class_dim = 1000
image_shape = "3,224,224"
else:
raise ValueError("{} is not supported.".format(args.data))
image_shape = [int(m) for m in image_shape.split(",")]
assert args.model 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__[args.model]()
out = model.net(input=image, class_dim=class_dim)
cost = fluid.layers.cross_entropy(input=out, label=label)
avg_cost = fluid.layers.mean(x=cost)
acc_top1 = fluid.layers.accuracy(input=out, label=label, k=1)
acc_top5 = fluid.layers.accuracy(input=out, label=label, k=5)
train_prog = fluid.default_main_program()
val_program = fluid.default_main_program().clone(for_test=True)
place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
############################################################################################################
# 2. quantization transform programs (training aware)
# Make some quantization transforms in the graph before training and testing.
# According to the weight and activation quantization type, the graph will be added
# some fake quantize operators and fake dequantize operators.
############################################################################################################
val_program = quant_aware(val_program,
place,
quant_config,
scope=None,
for_test=True)
compiled_train_prog = quant_aware(train_prog,
place,
quant_config,
scope=None,
for_test=False)
opt = create_optimizer(args)
opt.minimize(avg_cost)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
if args.pretrained_model:
def if_exist(var):
return os.path.exists(os.path.join(args.pretrained_model,
var.name))
fluid.io.load_vars(exe, args.pretrained_model, predicate=if_exist)
val_reader = paddle.batch(val_reader, batch_size=args.batch_size)
train_reader = paddle.batch(train_reader,
batch_size=args.batch_size,
drop_last=True)
train_feeder = feeder = fluid.DataFeeder([image, label], place)
val_feeder = feeder = fluid.DataFeeder([image, label],
place,
program=val_program)
def test(epoch, program):
batch_id = 0
acc_top1_ns = []
acc_top5_ns = []
for data in val_reader():
start_time = time.time()
acc_top1_n, acc_top5_n = exe.run(
program,
feed=train_feeder.feed(data),
fetch_list=[acc_top1.name, acc_top5.name])
end_time = time.time()
if batch_id % args.log_period == 0:
_logger.info(
"Eval epoch[{}] batch[{}] - acc_top1: {}; acc_top5: {}; time: {}"
.format(epoch, batch_id, np.mean(acc_top1_n),
np.mean(acc_top5_n), end_time - start_time))
acc_top1_ns.append(np.mean(acc_top1_n))
acc_top5_ns.append(np.mean(acc_top5_n))
batch_id += 1
_logger.info(
"Final eval epoch[{}] - acc_top1: {}; acc_top5: {}".format(
epoch, np.mean(np.array(acc_top1_ns)),
np.mean(np.array(acc_top5_ns))))
return np.mean(np.array(acc_top1_ns))
def train(epoch, compiled_train_prog):
batch_id = 0
for data in train_reader():
start_time = time.time()
loss_n, acc_top1_n, acc_top5_n = exe.run(
compiled_train_prog,
feed=train_feeder.feed(data),
fetch_list=[avg_cost.name, acc_top1.name, acc_top5.name])
end_time = time.time()
loss_n = np.mean(loss_n)
acc_top1_n = np.mean(acc_top1_n)
acc_top5_n = np.mean(acc_top5_n)
if batch_id % args.log_period == 0:
_logger.info(
"epoch[{}]-batch[{}] - loss: {}; acc_top1: {}; acc_top5: {}; time: {}"
.format(epoch, batch_id, loss_n, acc_top1_n, acc_top5_n,
end_time - start_time))
batch_id += 1
build_strategy = fluid.BuildStrategy()
build_strategy.memory_optimize = False
build_strategy.enable_inplace = False
build_strategy.fuse_all_reduce_ops = False
build_strategy.sync_batch_norm = False
exec_strategy = fluid.ExecutionStrategy()
compiled_train_prog = compiled_train_prog.with_data_parallel(
loss_name=avg_cost.name,
build_strategy=build_strategy,
exec_strategy=exec_strategy)
############################################################################################################
# train loop
############################################################################################################
best_acc1 = 0.0
best_epoch = 0
for i in range(args.num_epochs):
train(i, compiled_train_prog)
acc1 = test(i, val_program)
fluid.io.save_persistables(exe,
dirname=os.path.join(
args.checkpoint_dir, str(i)),
main_program=val_program)
if acc1 > best_acc1:
best_acc1 = acc1
best_epoch = i
fluid.io.save_persistables(exe,
dirname=os.path.join(
args.checkpoint_dir, 'best_model'),
main_program=val_program)
fluid.io.load_persistables(exe,
dirname=os.path.join(args.checkpoint_dir,
'best_model'),
main_program=val_program)
############################################################################################################
# 3. Freeze the graph after training by adjusting the quantize
# operators' order for the inference.
# The dtype of float_program's weights is float32, but in int8 range.
############################################################################################################
float_program, int8_program = convert(val_program, place, quant_config, \
scope=None, \
save_int8=True)
print("eval best_model after convert")
final_acc1 = test(best_epoch, float_program)
############################################################################################################
# 4. Save inference model
############################################################################################################
model_path = os.path.join(
quantization_model_save_dir, args.model,
'act_' + quant_config['activation_quantize_type'] + '_w_' +
quant_config['weight_quantize_type'])
float_path = os.path.join(model_path, 'float')
int8_path = os.path.join(model_path, 'int8')
if not os.path.isdir(model_path):
os.makedirs(model_path)
fluid.io.save_inference_model(dirname=float_path,
feeded_var_names=[image.name],
target_vars=[out],
executor=exe,
main_program=float_program,
model_filename=float_path + '/model',
params_filename=float_path + '/params')
fluid.io.save_inference_model(dirname=int8_path,
feeded_var_names=[image.name],
target_vars=[out],
executor=exe,
main_program=int8_program,
model_filename=int8_path + '/model',
params_filename=int8_path + '/params')
def setUp(self):
self.places = [fluid.CPUPlace()]
if core.is_compiled_with_cuda():
self.places.append(fluid.CUDAPlace(0))
def main(args):
hidden_size = args.hidden_size
neg_num = args.neg_num
epoch = args.epoch
p = args.p
q = args.q
save_path = args.save_path
batch_size = args.batch_size
walk_len = args.walk_len
win_size = args.win_size
if not os.path.isdir(save_path):
os.makedirs(save_path)
dataset = load(args.dataset)
if args.offline_learning:
log.info("Start random walk on disk...")
walk_save_path = os.path.join(save_path, "walks")
if not os.path.isdir(walk_save_path):
os.makedirs(walk_save_path)
pool = Pool(args.processes)
args_list = [(x, dataset.graph, walk_save_path, 1, batch_size,
walk_len, p, q, np.random.randint(2**32))
for x in range(epoch)]
pool.map(process, args_list)
filelist = glob.glob(os.path.join(walk_save_path, "*"))
log.info("Random walk on disk Done.")
else:
filelist = None
train_steps = int(dataset.graph.num_nodes / batch_size) * epoch
place = fluid.CUDAPlace(0) if args.use_cuda else fluid.CPUPlace()
node2vec_prog = fluid.Program()
startup_prog = fluid.Program()
with fluid.program_guard(node2vec_prog, startup_prog):
with fluid.unique_name.guard():
node2vec_pyreader, node2vec_loss = node2vec_model(
dataset.graph, hidden_size=hidden_size, neg_num=neg_num)
lr = l.polynomial_decay(0.025, train_steps, 0.0001)
adam = fluid.optimizer.Adam(lr)
adam.minimize(node2vec_loss)
node2vec_pyreader.decorate_tensor_provider(
node2vec_generator(dataset.graph,
batch_size=batch_size,
walk_len=walk_len,
win_size=win_size,
epoch=epoch,
neg_num=neg_num,
p=p,
q=q,
filelist=filelist))
node2vec_pyreader.start()
exe = fluid.Executor(place)
exe.run(startup_prog)
prev_time = time.time()
step = 0
while 1:
try:
node2vec_loss_val = exe.run(node2vec_prog,
fetch_list=[node2vec_loss],
return_numpy=True)[0]
cur_time = time.time()
use_time = cur_time - prev_time
prev_time = cur_time
step += 1
log.info("Step %d " % step +
"Node2vec Loss: %f " % node2vec_loss_val +
" %f s/step." % use_time)
except fluid.core.EOFException:
node2vec_pyreader.reset()
break
fluid.io.save_persistables(exe, os.path.join(save_path, "paddle_model"),
node2vec_prog)
def eval(vocab, infer_progs, dev_count, logger, args):
infer_prog, infer_startup_prog, infer_model = infer_progs
feed_order = infer_model.feed_order
loss = infer_model.loss
# prepare device
place = core.CUDAPlace(0) if args.use_gpu else core.CPUPlace()
exe = Executor(place)
if not args.use_gpu:
place = fluid.CPUPlace()
import multiprocessing
dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
else:
place = fluid.CUDAPlace(0)
dev_count = fluid.core.get_cuda_device_count()
if args.para_print:
with open("infer_program.desc", 'w') as f:
print(str(infer_prog), file=f)
# todo infer_startup_prog is not used, implicitly training scope will be used
# Use test set as validation each pass
total_loss = 0.0
total_cnt = 0
n_batch_cnt = 0
n_batch_loss = 0.0
val_feed_list = [
infer_prog.global_block().var(var_name) for var_name in feed_order
]
val_feeder = fluid.DataFeeder(val_feed_list, place)
dev_data = data.BidirectionalLMDataset(
args.test_path, vocab, test=True, shuffle_on_load=False)
dev_data_iter = lambda: dev_data.iter_batches(args.batch_size * dev_count, args.num_steps)
dev_reader = read_multiple(dev_data_iter, args.batch_size, dev_count)
last_hidden_values = np.zeros(
(dev_count, args.num_layers * 2 * args.batch_size * args.embed_size),
dtype='float32')
last_cell_values = np.zeros(
(dev_count, args.num_layers * 2 * args.batch_size * args.hidden_size),
dtype='float32')
for batch_id, batch_list in enumerate(dev_reader(), 1):
feed_data = batch_reader(batch_list, args)
feed = list(val_feeder.feed_parallel(feed_data, dev_count))
for i in range(dev_count):
init_hidden_tensor = fluid.core.LoDTensor()
if args.use_gpu:
placex = fluid.CUDAPlace(i)
else:
placex = fluid.CPUPlace()
init_hidden_tensor.set(last_hidden_values[i], placex)
init_cell_tensor = fluid.core.LoDTensor()
init_cell_tensor.set(last_cell_values[i], placex)
feed[i]['init_hiddens'] = init_hidden_tensor
feed[i]['init_cells'] = init_cell_tensor
#todo test pe has bug in r1.3
#import pdb; pdb.set_trace()
last_hidden_values = []
last_cell_values = []
for i in range(dev_count):
val_fetch_outs = exe.run(
program=infer_prog,
feed=feed[i],
fetch_list=[
infer_model.loss.name, infer_model.last_hidden.name,
infer_model.last_cell.name
], # + [x[0] for x in names] + [x[0] for x in grad_names],
return_numpy=False)
last_hidden_values.append(np.array(val_fetch_outs[1]))
last_cell_values.append(np.array(val_fetch_outs[2]))
total_loss += np.array(val_fetch_outs[0]).sum()
n_batch_cnt += len(np.array(val_fetch_outs[0]))
total_cnt += len(np.array(val_fetch_outs[0]))
n_batch_loss += np.array(val_fetch_outs[0]).sum()
last_hidden_values = np.array(last_hidden_values).reshape((
dev_count, args.num_layers * 2 * args.batch_size * args.embed_size))
last_cell_values = np.array(last_cell_values).reshape(
(dev_count,
args.num_layers * 2 * args.batch_size * args.hidden_size))
log_every_n_batch = args.log_interval
if log_every_n_batch > 0 and batch_id % log_every_n_batch == 0:
logger.info('Average dev loss from batch {} to {} is {}'.format(
batch_id - log_every_n_batch + 1, batch_id, "%.10f" % (
n_batch_loss / n_batch_cnt)))
n_batch_loss = 0.0
n_batch_cnt = 0
batch_offset = 0
ppl = np.exp(total_loss / total_cnt)
return ppl
def train_loop(args,
logger,
vocab,
train_progs,
infer_progs,
optimizer,
nccl2_num_trainers=1,
nccl2_trainer_id=0,
worker_endpoints=None):
train_prog, train_startup_prog, train_model = train_progs
infer_prog, infer_startup_prog, infer_model = infer_progs
# prepare device
place = core.CUDAPlace(0) if args.use_gpu else core.CPUPlace()
exe = Executor(place)
if not args.use_gpu:
place = fluid.CPUPlace()
import multiprocessing
dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
else:
place = fluid.CUDAPlace(0)
dev_count = fluid.core.get_cuda_device_count()
if args.load_dir:
logger.info('load pretrained checkpoints from {}'.format(args.load_dir))
# Todo: why not need to run train_startup_prog before load_persistables
fluid.io.load_persistables(exe, args.load_dir, main_program=train_prog)
elif args.load_pretraning_params:
logger.info('load pretrained params from {}'.format(args.load_pretraning_params))
exe.run(train_startup_prog)
init_pretraining_params(exe, args.load_pretraning_params, main_program=train_prog)
else:
exe.run(train_startup_prog)
# prepare data
feed_list = [
train_prog.global_block().var(var_name)
for var_name in train_model.feed_order
]
feeder = fluid.DataFeeder(feed_list, place)
logger.info('Training the model...')
exe_strategy = fluid.parallel_executor.ExecutionStrategy()
if args.para_print:
exe_strategy.num_threads = 1
debug_init(train_prog, train_model.grad_vars,
train_model.grad_vars_name)
with open("program.desc", 'w') as f:
print(str(train_prog), file=f)
parallel_executor = fluid.ParallelExecutor(
loss_name=train_model.loss.name,
main_program=train_prog,
use_cuda=bool(args.use_gpu),
exec_strategy=exe_strategy,
num_trainers=nccl2_num_trainers,
trainer_id=nccl2_trainer_id)
load_params(train_prog, parallel_executor, place, logger, args)
print_para(train_prog, parallel_executor, logger, optimizer, args)
logger.info("begin to load data")
train_data = data.BidirectionalLMDataset(
args.train_path,
vocab,
test=(not args.shuffle),
shuffle_on_load=args.shuffle)
logger.info("finished load vocab")
# get train epoch size
log_interval = args.log_interval
total_time = 0.0
batch_size = args.batch_size
hidden_size = args.hidden_size
custom_samples_array = np.zeros(
(batch_size, args.num_steps, args.n_negative_samples_batch + 1),
dtype='int64')
custom_probabilities_array = np.zeros(
(batch_size, args.num_steps, args.n_negative_samples_batch + 1),
dtype='float32')
for i in range(batch_size):
for j in range(0, args.num_steps):
for k in range(0, args.n_negative_samples_batch + 1):
custom_samples_array[i][j][k] = k
custom_probabilities_array[i][j][k] = 1.0
for epoch_id in range(args.max_epoch):
start_time = time.time()
logger.info("epoch id {}".format(epoch_id))
train_data_iter = lambda: train_data.iter_batches(batch_size * dev_count, args.num_steps)
train_reader = read_multiple(train_data_iter, batch_size, dev_count)
total_num = 0
n_batch_loss = 0.0
n_batch_cnt = 0
last_hidden_values = np.zeros(
(dev_count, args.num_layers * 2 * batch_size * args.embed_size),
dtype='float32')
last_cell_values = np.zeros(
(dev_count, args.num_layers * 2 * batch_size * hidden_size),
dtype='float32')
begin_time = time.time()
for batch_id, batch_list in enumerate(train_reader(), 1):
feed_data = batch_reader(batch_list, args)
feed = list(feeder.feed_parallel(feed_data, dev_count))
for i in range(dev_count):
init_hidden_tensor = fluid.core.LoDTensor()
if args.use_gpu:
placex = fluid.CUDAPlace(i)
else:
placex = fluid.CPUPlace()
init_hidden_tensor.set(last_hidden_values[i], placex)
init_cell_tensor = fluid.core.LoDTensor()
init_cell_tensor.set(last_cell_values[i], placex)
feed[i]['init_hiddens'] = init_hidden_tensor
feed[i]['init_cells'] = init_cell_tensor
fetch_outs = parallel_executor.run(
feed=feed,
fetch_list=[
train_model.loss.name, train_model.last_hidden.name,
train_model.last_cell.name
], # + [x[0] for x in names] + [x[0] for x in grad_names],
return_numpy=False)
cost_train = np.array(fetch_outs[0]).mean()
#import pdb; pdb.set_trace()
last_hidden_values = np.array(fetch_outs[1])
last_hidden_values = last_hidden_values.reshape(
(dev_count, args.num_layers * 2 * batch_size * args.embed_size))
last_cell_values = np.array(fetch_outs[2])
last_cell_values = last_cell_values.reshape((
dev_count, args.num_layers * 2 * batch_size * args.hidden_size))
#vars = fetch_outs[2:2+len(names)]
#grad_vars = fetch_outs[2+len(names):]
total_num += args.batch_size * dev_count
n_batch_loss += np.array(fetch_outs[0]).sum()
#logger.info("n_batch_loss from {} to {} is {}, {} ".format(
# batch_id - log_interval, batch_id, n_batch_loss,
# np.array(fetch_outs[0]).sum()))
n_batch_cnt += len(np.array(fetch_outs[0]))
if batch_id > 0 and batch_id % log_interval == 0:
#vars_print(logger, args, vars=(vars, names), grad_vars=(grad_vars, grad_names))
print_para(train_prog, parallel_executor, logger, optimizer,
args)
smoothed_ppl = np.exp(n_batch_loss / n_batch_cnt)
ppl = np.exp(
np.array(fetch_outs[0]).sum() /
len(np.array(fetch_outs[0])))
used_time = time.time() - begin_time
speed = log_interval / used_time
logger.info(
"[train] epoch:{}, step:{}, loss:{:.3f}, ppl:{:.3f}, smoothed_ppl:{:.3f}, speed:{:.3f}".
format(epoch_id, batch_id, n_batch_loss / n_batch_cnt, ppl,
smoothed_ppl, speed))
n_batch_loss = 0.0
n_batch_cnt = 0
begin_time = time.time()
if batch_id > 0 and batch_id % args.dev_interval == 0:
valid_ppl = eval(vocab, infer_progs, dev_count, logger, args)
logger.info("valid ppl {}".format(valid_ppl))
if batch_id > 0 and batch_id % args.save_interval == 0:
model_path = os.path.join(args.para_save_dir,
str(batch_id + epoch_id))
if not os.path.isdir(model_path):
os.makedirs(model_path)
fluid.io.save_persistables(
executor=exe, dirname=model_path, main_program=train_prog)
if args.detail and batch_id > 100:
exit()
end_time = time.time()
total_time += end_time - start_time
logger.info("train ppl {}".format(ppl))
if epoch_id == args.max_epoch - 1 and args.enable_ce:
logger.info("lstm_language_model_duration\t%s" %
(total_time / args.max_epoch))
logger.info("lstm_language_model_loss\t%s" % ppl[0])
model_path = os.path.join(args.para_save_dir, str(epoch_id))
if not os.path.isdir(model_path):
os.makedirs(model_path)
fluid.io.save_persistables(
executor=exe, dirname=model_path, main_program=train_prog)
valid_ppl = eval(vocab, infer_progs, dev_count, logger, args)
logger.info("valid ppl {}".format(valid_ppl))
test_ppl = eval(vocab, infer_progs, dev_count, place, logger, args)
logger.info("test ppl {}".format(test_ppl))
# 使用交叉熵函数，描述真实样本标签与预测概率之间的误差 cost = fluid.layers.cross_entropy(input=predict, label=label) avg_cost = fluid.layers.mean(cost) # 计算分类准确率 acc = fluid.layers.accuracy(input=predict, label=label) # 定义优化函数 # 使用adam算法优化，学习率设定为0.001 optimizer = fluid.optimizer.AdamaxOptimizer(learning_rate=0.001) opts = optimizer.minimize(avg_cost) # 模型训练 # 创建训练的Executor use_cuda = False place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() test_program = fluid.default_main_program().clone(for_test=True) exe = fluid.Executor(place) exe.run(fluid.default_main_program()) # 告知网络传入数据分两部分，image，label feeder = fluid.DataFeeder(place=place, feed_list=[image, label]) # 展示模型曲线 all_train_iter = 0 all_train_iters = [] all_train_costs = [] all_train_accs = [] def draw_train_process(title, iters, costs, accs, label_cost, label_acc):
def train():
learning_rate = cfg.learning_rate
image_shape = [3, 512, 512]
if cfg.enable_ce:
fluid.default_startup_program().random_seed = 1000
fluid.default_main_program().random_seed = 1000
import random
random.seed(0)
np.random.seed(0)
devices_num = get_device_num()
total_batch_size = devices_num * cfg.TRAIN.im_per_batch
use_random = True
startup_prog = fluid.Program()
train_prog = fluid.Program()
with fluid.program_guard(train_prog, startup_prog):
with fluid.unique_name.guard():
model = model_builder.EAST(
add_conv_body_func=resnet.ResNet(),
use_random=use_random)
model.build_model(image_shape)
losses, keys = model.loss()
loss = losses[0]
fetch_list = losses
boundaries = cfg.lr_steps
gamma = cfg.lr_gamma
step_num = len(cfg.lr_steps)
values = [learning_rate * (gamma**i) for i in range(step_num + 1)]
lr = exponential_with_warmup_decay(
learning_rate=learning_rate,
boundaries=boundaries,
values=values,
warmup_iter=cfg.warm_up_iter,
warmup_factor=cfg.warm_up_factor)
optimizer = fluid.optimizer.AdamOptimizer(learning_rate=lr, regularization=fluid.regularizer.L2Decay(cfg.weight_decay))
optimizer.minimize(loss)
fetch_list = fetch_list + [lr]
for var in fetch_list:
var.persistable = True
gpu_id = int(os.environ.get('FLAGS_selected_gpus', 0))
place = fluid.CUDAPlace(gpu_id) if cfg.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
build_strategy = fluid.BuildStrategy()
build_strategy.fuse_all_optimizer_ops = False
build_strategy.fuse_elewise_add_act_ops = True
build_strategy.sync_batch_norm=True
exec_strategy = fluid.ExecutionStrategy()
exec_strategy.num_iteration_per_drop_scope = 1
exe.run(startup_prog)
if cfg.pretrained_model:
def if_exist(var):
return os.path.exists(os.path.join(cfg.pretrained_model, var.name))
fluid.io.load_vars(exe, cfg.pretrained_model, predicate=if_exist)
compiled_train_prog = fluid.CompiledProgram(train_prog).with_data_parallel(
loss_name=loss.name,
build_strategy=build_strategy,
exec_strategy=exec_strategy)
dataset = icdar.ICDAR2015Dataset()
data_generator = dataset.get_batch(num_workers=24,
input_size=512,
batch_size=14)
def train_loop():
start_time = time.time()
prev_start_time = start_time
start = start_time
train_stats = TrainingStats(cfg.log_window, keys)
#for iter_id, data in enumerate(next(data_generator)):
for iter_id in range(100000):
data = next(data_generator)
#for data in data_list:
prev_start_time = start_time
start_time = time.time()
outs = exe.run(compiled_train_prog, fetch_list=[v.name for v in fetch_list],
feed={"input_images": data[0],
"input_score_maps": data[2],
"input_geo_maps": data[3],
"input_training_masks": data[4]})
stats = {k: np.array(v).mean() for k, v in zip(keys, outs[:-1])}
train_stats.update(stats)
logs = train_stats.log()
strs = '{}, batch: {}, lr: {:.5f}, {}, time: {:.3f}'.format(
now_time(), iter_id,
np.mean(outs[-1]), logs, start_time - prev_start_time)
if iter_id % 10 == 0:
print(strs)
sys.stdout.flush()
if (iter_id + 1) % cfg.TRAIN.snapshot_iter == 0:
save_model(exe, "model_iter{}".format(iter_id), train_prog)
if (iter_id + 1) == cfg.max_iter:
break
end_time = time.time()
total_time = end_time - start_time
last_loss = np.array(outs[0]).mean()
train_loop()
def test_case(self):
self._test_case(fluid.CPUPlace())
if fluid.is_compiled_with_cuda():
self._test_case(fluid.CUDAPlace(0))
'models/lm/zh_giga.no_cna_cmn.prune01244.klm',
"Filepath for language model.")
add_arg('decoding_method', str,
'ctc_beam_search',
# 'ctc_greedy',
"Decoding method. Options: ctc_beam_search, ctc_greedy",
choices=['ctc_beam_search', 'ctc_greedy'])
add_arg('specgram_type', str,
'linear',
"Audio feature type. Options: linear, mfcc.",
choices=['linear', 'mfcc'])
args = parser.parse_args()
if args.use_gpu:
place = fluid.CUDAPlace(0)
else:
place = fluid.CPUPlace()
data_generator = DataGenerator(
vocab_filepath=args.vocab_path,
mean_std_filepath=args.mean_std_path,
augmentation_config='{}',
specgram_type=args.specgram_type,
keep_transcription_text=True,
place=place,
is_training=False)
# prepare ASR model
ds2_model = DeepSpeech2Model(
vocab_size=data_generator.vocab_size,
num_conv_layers=args.num_conv_layers,
num_rnn_layers=args.num_rnn_layers,
def main():
batch_size = 20
if args.enable_ce:
train_reader = paddle.batch(
paddle.dataset.uci_housing.train(), batch_size=batch_size)
test_reader = paddle.batch(
paddle.dataset.uci_housing.test(), batch_size=batch_size)
else:
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.uci_housing.train(), buf_size=500),
batch_size=batch_size)
test_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.uci_housing.test(), buf_size=500),
batch_size=batch_size)
# feature vector of length 13
x = fluid.data(name='x', shape=[None, 13], dtype='float32')
y = fluid.data(name='y', shape=[None, 1], dtype='float32')
main_program = fluid.default_main_program()
startup_program = fluid.default_startup_program()
if args.enable_ce:
main_program.random_seed = 90
startup_program.random_seed = 90
y_predict = fluid.layers.fc(input=x, size=1, act=None)
cost = fluid.layers.square_error_cost(input=y_predict, label=y)
avg_loss = fluid.layers.mean(cost)
test_program = main_program.clone(for_test=True)
sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
sgd_optimizer.minimize(avg_loss)
# can use CPU or GPU
use_cuda = args.use_gpu
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
exe = fluid.Executor(place)
# Specify the directory to save the parameters
params_dirname = "fit_a_line.inference.model"
num_epochs = args.num_epochs
# main train loop.
feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
exe.run(startup_program)
train_prompt = "Train cost"
test_prompt = "Test cost"
step = 0
exe_test = fluid.Executor(place)
for pass_id in range(num_epochs):
for data_train in train_reader():
avg_loss_value, = exe.run(
main_program,
feed=feeder.feed(data_train),
fetch_list=[avg_loss])
if step % 10 == 0: # record a train cost every 10 batches
print("%s, Step %d, Cost %f" %
(train_prompt, step, avg_loss_value[0]))
if step % 100 == 0: # record a test cost every 100 batches
test_metics = train_test(
executor=exe_test,
program=test_program,
reader=test_reader,
fetch_list=[avg_loss],
feeder=feeder)
print("%s, Step %d, Cost %f" %
(test_prompt, step, test_metics[0]))
# If the accuracy is good enough, we can stop the training.
if test_metics[0] < 10.0:
break
step += 1
if math.isnan(float(avg_loss_value[0])):
sys.exit("got NaN loss, training failed.")
if params_dirname is not None:
# We can save the trained parameters for the inferences later
fluid.io.save_inference_model(params_dirname, ['x'], [y_predict],
exe)
if args.enable_ce and pass_id == args.num_epochs - 1:
print("kpis\ttrain_cost\t%f" % avg_loss_value[0])
print("kpis\ttest_cost\t%f" % test_metics[0])
infer_exe = fluid.Executor(place)
inference_scope = fluid.core.Scope()
# infer
with fluid.scope_guard(inference_scope):
[inference_program, feed_target_names, fetch_targets
] = fluid.io.load_inference_model(params_dirname, infer_exe)
batch_size = 10
infer_reader = paddle.batch(
paddle.dataset.uci_housing.test(), batch_size=batch_size)
infer_data = next(infer_reader())
infer_feat = numpy.array(
[data[0] for data in infer_data]).astype("float32")
infer_label = numpy.array(
[data[1] for data in infer_data]).astype("float32")
assert feed_target_names[0] == 'x'
results = infer_exe.run(
inference_program,
feed={feed_target_names[0]: numpy.array(infer_feat)},
fetch_list=fetch_targets)
print("infer results: (House Price)")
for idx, val in enumerate(results[0]):
print("%d: %.2f" % (idx, val))
print("\nground truth:")
for idx, val in enumerate(infer_label):
print("%d: %.2f" % (idx, val))
save_result(results[0], infer_label)
def run_gpu_fleet_api_trainer(self, args):
assert args.update_method == "nccl2"
self.lr = args.lr
exec_strategy = fluid.ExecutionStrategy()
exec_strategy.num_threads = 1
dist_strategy = DistributedStrategy()
dist_strategy.exec_strategy = exec_strategy
dist_strategy.fuse_memory_size = 1 # MB
dist_strategy.fuse_laryer_size = 1
if args.use_local_sgd:
dist_strategy.use_local_sgd = True
if args.ut4grad_allreduce:
dist_strategy._ut4grad_allreduce = True
role = role_maker.PaddleCloudRoleMaker(is_collective=True)
fleet.init(role)
print_to_err("gpu_fleet", "fleet.node_num:")
# "fleet.node_id:", fleet.node_id(),
# "fleet.trainer_num:", fleet.worker_num())
test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \
self.get_model(batch_size=args.batch_size, dist_strategy=dist_strategy)
trainer_prog = fleet._origin_program
dist_prog = fleet.main_program
device_id = int(os.getenv("FLAGS_selected_gpus", "0"))
place = fluid.CUDAPlace(device_id)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
eprint(type(self).__name__, "run worker startup program done.")
feed_var_list = [
var for var in trainer_prog.global_block().vars.values()
if var.is_data
]
feeder = fluid.DataFeeder(feed_var_list, place)
reader_generator = train_reader()
def get_data():
origin_batch = next(reader_generator)
if args.update_method != "local" and args.use_reader_alloc:
new_batch = []
for offset, item in enumerate(origin_batch):
if offset % 2 == args.trainer_id:
new_batch.append(item)
return new_batch
else:
return origin_batch
print_to_err(type(self).__name__, "begin to train on trainer")
out_losses = []
for i in six.moves.xrange(RUN_STEP):
loss, = exe.run(dist_prog,
fetch_list=[avg_cost.name],
feed=feeder.feed(get_data()))
out_losses.append(loss[0])
print_to_err(type(self).__name__, "run step %d finished" % i)
print_to_err(type(self).__name__, "trainer run finished")
if six.PY2:
print(pickle.dumps(out_losses))
else:
sys.stdout.buffer.write(pickle.dumps(out_losses))
def get_places():
places = [fluid.CPUPlace()]
if fluid.core.is_compiled_with_cuda():
places.append(fluid.CUDAPlace(0))
return places
def run_trainer(self, args):
seed = 90
device_id = int(os.getenv("FLAGS_selected_gpus", "0"))
place = fluid.CUDAPlace(device_id)
def _get_data(batch):
if args.update_method != "local":
new_batch = []
for offset, item in enumerate(batch):
if offset % 2 == args.trainer_id:
new_batch.append(item)
return new_batch
else:
return batch
with fluid.dygraph.guard(place):
fluid.default_startup_program().random_seed = seed
fluid.default_main_program().random_seed = seed
np.random.seed(seed)
import random
random.seed = seed
model, train_reader, opt = self.get_model()
nranks = len(args.endpoints.split(",")) if args.endpoints else 1
if args.update_method == "nccl2":
strategy = dygraph.parallel.ParallelStrategy()
strategy.nranks = nranks
strategy.local_rank = args.trainer_id
strategy.trainer_endpoints = args.endpoints.split(",")
strategy.current_endpoint = args.current_endpoint
print_to_err(
type(self).__name__,
"begin to prepare context in dygraph with nccl2")
dygraph.parallel.prepare_context(strategy)
model = dygraph.parallel.DataParallel(model, strategy)
print_to_err(type(self).__name__, "model built in dygraph")
out_losses = []
print_to_err(type(self).__name__, "begin to run dygraph training")
for step_id, data in enumerate(train_reader()):
data = _get_data(data)
if step_id == RUN_STEP:
break
loss = self.run_one_loop(model, opt, data)
if step_id % 10 == 0:
print_to_err(
type(self).__name__,
"loss at step %d: %f" % (step_id, loss.numpy()))
out_losses.append(loss.numpy())
# FIXME(Yancey1989): scale the loss inplace
if args.update_method == "nccl2":
loss = model.scale_loss(loss)
loss.backward()
if args.update_method == "nccl2":
model.apply_collective_grads()
opt.minimize(loss)
model.clear_gradients()
print_to_out(out_losses)
def __next__(self):
if self._first_batch is not None:
batch = self._first_batch
self._first_batch = None
return batch
if self._counter >= self._size and self._size > 0:
if self._auto_reset:
self.reset()
raise StopIteration
# Gather outputs
outputs = []
for p in self._pipes:
with p._check_api_type_scope(types.PipelineAPIType.ITERATOR):
outputs.append(p.share_outputs())
for i in range(self._num_gpus):
dev_id = self._pipes[i].device_id
# Initialize dict for all output categories
category_outputs = dict()
# Segregate outputs into categories
for j, out in enumerate(outputs[i]):
category_outputs[self.output_map[j]] = out
pd_gpu_place = fluid.CUDAPlace(dev_id)
pd_cpu_place = fluid.CPUPlace()
category_pd_type = dict()
category_place = dict()
category_tensors = dict()
category_shapes = dict()
category_lengths = dict()
for cat, out in category_outputs.items():
lod = self.normalized_map[cat]
assert out.is_dense_tensor() or lod > 0, \
"non-dense tensor lists must have LoD > 0"
if lod > 0:
# +1 for batch dim
seq_len = recursive_length(out, lod + 1)[1:]
shape = out.at(0).shape
if callable(shape):
shape = shape()
shape = [sum(seq_len[-1])] + list(shape[lod:])
category_shapes[cat] = shape
category_lengths[cat] = seq_len
else:
out = out.as_tensor()
category_shapes[cat] = out.shape()
category_lengths[cat] = []
category_tensors[cat] = out
category_pd_type[cat] = to_paddle_type(out)
if isinstance(out, (TensorGPU, TensorListGPU)):
category_place[cat] = pd_gpu_place
else:
category_place[cat] = pd_cpu_place
if self._data_batches[i] is None:
pd_tensors = {}
for cat, lod in self.normalized_map.items():
lod_tensor = fluid.core.LoDTensor()
lod_tensor._set_dims(category_shapes[cat])
pd_tensors[cat] = lod_tensor
self._data_batches[i] = pd_tensors
else:
pd_tensors = self._data_batches[i]
# Copy data from DALI Tensors to LoDTensors
for cat, tensor in category_tensors.items():
if hasattr(tensor, 'shape'): # could be tensor list
assert self._dynamic_shape or \
tensor.shape() == pd_tensors[cat].shape(), \
("Shapes do not match: DALI tensor has size {0}, "
"but LoDTensor has size {1}".format(
tensor.shape(), pd_tensors[cat].shape()))
lod_tensor = pd_tensors[cat]
lod_tensor._set_dims(category_shapes[cat])
seq_len = category_lengths[cat]
lod_tensor.set_recursive_sequence_lengths(seq_len)
ptr = lod_tensor._mutable_data(category_place[cat],
category_pd_type[cat])
feed_ndarray(tensor, ptr)
for p in self._pipes:
with p._check_api_type_scope(types.PipelineAPIType.ITERATOR):
p.release_outputs()
p.schedule_run()
self._counter += self._num_gpus * self.batch_size
if (not self._fill_last_batch) and (self._counter >
self._size) and self._size > 0:
# First calculate how much data is required to
# return exactly self._size entries.
diff = self._num_gpus * self.batch_size - (self._counter -
self._size)
# Figure out how many GPUs to grab from.
num_gpus_to_grab = int(math.ceil(diff / self.batch_size))
# Figure out how many results to grab from the last GPU
# (as a fractional GPU batch may be required to bring us
# right up to self._size).
mod_diff = diff % self.batch_size
data_from_last_gpu = mod_diff if mod_diff else self.batch_size
# Grab the relevant data.
# 1) Grab everything from the relevant GPUs.
# 2) Grab the right data from the last GPU.
# 3) Append data together correctly and return.
output = self._data_batches[0:num_gpus_to_grab]
output[-1] = output[-1].copy()
for cat in self.output_map:
lod_tensor = output[-1][cat]
output[-1][cat] = lod_tensor_clip(lod_tensor,
data_from_last_gpu)
return output
return self._data_batches
import os
import paddle
import paddle.fluid as fluid
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, Linear
import numpy as np
import Data_Load_Module
import Network
#调用加载数据的函数
train_loader = Data_Load_Module.load_data('train')
#在使用GPU机器时，可以将use_gpu变量设置成True
use_gpu = True
place = fluid.CUDAPlace(0) if use_gpu else fluid.CPUPlace()
with fluid.dygraph.guard(place):
model = Network.MNIST()
model.train()
#四种优化算法的设置方案，可以逐一尝试效果
#optimizer = fluid.optimizer.SGDOptimizer(learning_rate=0.01, parameter_list=model.parameters())
#optimizer = fluid.optimizer.MomentumOptimizer(learning_rate=0.01, momentum=0.9, parameter_list=model.parameters())
#optimizer = fluid.optimizer.AdagradOptimizer(learning_rate=0.01, parameter_list=model.parameters())
optimizer = fluid.optimizer.AdamOptimizer(learning_rate=0.01, parameter_list=model.parameters())
EPOCH_NUM = 5
for epoch_id in range(EPOCH_NUM):
for batch_id, data in enumerate(train_loader()):
#准备数据
def main():
batch_size = 32
num_epochs = args.num_epochs
net_type = args.net
# 训练ckpt和inf模型路径
param_base_dir = os.path.join(code_base_dir, 'params')
param_base_dir = os.path.join(param_base_dir, net_type)
infer_param_path = os.path.join(param_base_dir, "inf")
ckpt_param_path = os.path.join(param_base_dir, "ckpt")
print(infer_param_path)
print(ckpt_param_path)
train_reader = paddle.batch(
paddle.reader.shuffle(data_reader(), int(batch_size * 1.5)),
batch_size)
test_reader = paddle.batch(
paddle.reader.shuffle(data_reader(8, 10), int(batch_size * 1.5)),
batch_size)
train_program = fluid.Program()
train_init = fluid.Program()
with fluid.program_guard(train_program, train_init):
image = fluid.layers.data(name='image',
shape=[3, 512, 512],
dtype='float32')
label = fluid.layers.data(name='label',
shape=[1, 512, 512],
dtype='int32')
train_loader = fluid.io.DataLoader.from_generator(
feed_list=[image, label], capacity=batch_size)
test_loader = fluid.io.DataLoader.from_generator(
feed_list=[image, label], capacity=batch_size)
if net_type == "unet_simple":
prediction = unet_simple(image, 2, [512, 512])
elif net_type == "unet_base":
prediction = unet_base(image, 2, [512, 512])
elif net_type == "deeplabv3":
prediction = deeplabv3p(image, 2)
else:
print("错误的网络类型")
sys.exit(0)
avg_loss = create_loss(prediction, label, 2)
miou = mean_iou(prediction, label, 2)
# decay=paddle.fluid.regularizer.L2Decay(0.1)
# optimizer = fluid.optimizer.SGD(learning_rate=0.0005,regularization=decay)
# optimizer = fluid.optimizer.DecayedAdagradOptimizer(learning_rate=0.02,regularization=decay)
# optimizer = fluid.optimizer.RMSProp(learning_rate=0.1,momentum=0.8,centered=True, regularization=decay)
optimizer = fluid.optimizer.AdamOptimizer(learning_rate=0.003)
# optimizer = fluid.optimizer.MomentumOptimizer(learning_rate=0.006, momentum=0.8,regularization=decay)
optimizer.minimize(avg_loss)
place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(train_init)
# fluid.io.load_persistables(exe, ckpt_param_path, train_init)
exe_test = fluid.Executor(place)
test_program = train_program.clone(for_test=True)
# train_program=fluid.CompiledProgram(train_program).with_data_parallel(loss_name=avg_loss.name)
test_program = fluid.CompiledProgram(test_program).with_data_parallel(
loss_name=avg_loss.name)
train_loader.set_sample_list_generator(train_reader, places=place)
test_loader.set_sample_list_generator(test_reader, places=place)
feeder = fluid.DataFeeder(place=place, feed_list=[image, label])
step = 1
best_miou = 0
train_prompt = "Train_miou"
test_prompt = "Test_miou"
plot_prompt = Ploter(train_prompt, test_prompt)
for pass_id in range(num_epochs):
for data_train in train_loader():
avg_loss_value, miou_value = exe.run(train_program,
feed=data_train,
fetch_list=[avg_loss, miou])
if step % 10 == 0:
print("\t\tTrain pass %d, Step %d, Cost %f, Miou %f" %
(pass_id, step, avg_loss_value[0], miou_value[0]))
# if step % 10 ==0:
# plot_prompt.append(train_prompt, step, miou_value[0])
# plot_prompt.plot()
eval_miou = 0
if step % 100 == 0:
auc_metric = fluid.metrics.Auc("AUC")
test_losses = []
test_mious = []
for _, test_data in enumerate(test_loader()):
# print(test_data)
# input("pause")
_, test_loss, test_miou = exe_test.run(
test_program,
feed=test_data,
fetch_list=[prediction, avg_loss, miou])
test_losses.append(test_loss[0])
test_mious.append(test_miou[0])
eval_miou = np.average(np.array(test_mious))
# plot_prompt.append(test_prompt, step, eval_miou)
# plot_prompt.plot()
print("Test loss: %f ,miou: %f" %
(np.average(np.array(test_losses)), eval_miou))
if math.isnan(float(avg_loss_value[0])):
sys.exit("got NaN loss, training failed.")
if step % 100 == 0 and param_base_dir is not None and eval_miou > best_miou:
best_miou = eval_miou
print("Saving params of step: %d" % step)
fluid.io.save_inference_model(infer_param_path,
feeded_var_names=['image'],
target_vars=[prediction],
executor=exe,
main_program=train_program)
fluid.io.save_persistables(exe, ckpt_param_path, train_program)
step += 1
print(best_miou)
def get_all_places(self):
p = [fluid.CPUPlace()]
if fluid.is_compiled_with_cuda():
p.append(fluid.CUDAPlace(0))
return p
def main(use_cuda):
"""
Advbox demo which demonstrate how to use advbox.
"""
TOTAL_NUM = 500
IMG_NAME = 'img'
LABEL_NAME = 'label'
img = fluid.layers.data(name=IMG_NAME, shape=[1, 28, 28], dtype='float32')
# gradient should flow
img.stop_gradient = False
label = fluid.layers.data(name=LABEL_NAME, shape=[1], dtype='int64')
logits = mnist_cnn_model(img)
cost = fluid.layers.cross_entropy(input=logits, label=label)
avg_cost = fluid.layers.mean(x=cost)
#根据配置选择使用CPU资源还是GPU资源
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
exe = fluid.Executor(place)
BATCH_SIZE = 1
test_reader = paddle.batch(paddle.reader.shuffle(
paddle.dataset.mnist.test(), buf_size=128 * 10),
batch_size=BATCH_SIZE)
fluid.io.load_params(exe,
"./mnist/",
main_program=fluid.default_main_program())
# advbox demo
m = PaddleModel(fluid.default_main_program(),
IMG_NAME,
LABEL_NAME,
logits.name,
avg_cost.name, (-1, 1),
channel_axis=1)
#使用静态FGSM epsilon不可变
attack = FGSM_static(m)
attack_config = {"epsilon": 0.01}
# use test data to generate adversarial examples
total_count = 0
fooling_count = 0
for data in test_reader():
total_count += 1
adversary = Adversary(data[0][0], data[0][1])
# FGSM non-targeted attack
adversary = attack(adversary, **attack_config)
if adversary.is_successful():
fooling_count += 1
#print(
# 'attack success, original_label=%d, adversarial_label=%d, count=%d'
# % (data[0][1], adversary.adversarial_label, total_count))
else:
logger.info('attack failed, original_label=%d, count=%d' %
(data[0][1], total_count))
if total_count >= TOTAL_NUM:
print(
"[TEST_DATASET]: fooling_count=%d, total_count=%d, fooling_rate=%f"
% (fooling_count, total_count,
float(fooling_count) / total_count))
break
print("fgsm attack done without any defence")
#使用FeatureFqueezingDefence
# advbox FeatureFqueezingDefence demo
n = PaddleLabelSmoothingDefenceModel(fluid.default_main_program(),
IMG_NAME,
LABEL_NAME,
logits.name,
avg_cost.name, (-1, 1),
channel_axis=1,
preprocess=None,
smoothing=0.1)
attack_new = FGSM_static(n)
attack_config = {"epsilon": 0.01}
total_count = 0
fooling_count = 0
for data in test_reader():
total_count += 1
#不设置y 会自动获取
adversary = Adversary(data[0][0], None)
# FGSM non-targeted attack
adversary = attack_new(adversary, **attack_config)
if adversary.is_successful():
fooling_count += 1
logger.info(
'attack success, original_label=%d, adversarial_label=%d, count=%d'
% (data[0][1], adversary.adversarial_label, total_count))
else:
logger.info('attack failed, original_label=%d, count=%d' %
(data[0][1], total_count))
if total_count >= TOTAL_NUM:
print(
"[TEST_DATASET]: fooling_count=%d, total_count=%d, fooling_rate=%f"
% (fooling_count, total_count,
float(fooling_count) / total_count))
break
print("fgsm attack done with LabelSmoothingDefence")
def train(args):
config = parse_config(args.config)
train_config = merge_configs(config, 'train', vars(args))
valid_config = merge_configs(config, 'valid', vars(args))
print_configs(train_config, 'Train')
use_data_parallel = False
trainer_count = fluid.dygraph.parallel.Env().nranks
place = fluid.CUDAPlace(fluid.dygraph.parallel.Env().dev_id) \
if use_data_parallel else fluid.CUDAPlace(0)
with fluid.dygraph.guard(place):
if use_data_parallel:
strategy = fluid.dygraph.parallel.prepare_context()
video_model = TSM_ResNet("TSM", train_config)
optimizer = create_optimizer(train_config.TRAIN,
video_model.parameters())
if use_data_parallel:
video_model = fluid.dygraph.parallel.DataParallel(
video_model, strategy)
bs_denominator = 1
if args.use_gpu:
# check number of GPUs
gpus = os.getenv("CUDA_VISIBLE_DEVICES", "")
if gpus == "":
pass
else:
gpus = gpus.split(",")
num_gpus = len(gpus)
assert num_gpus == train_config.TRAIN.num_gpus, \
"num_gpus({}) set by CUDA_VISIBLE_DEVICES" \
"shoud be the same as that" \
"set in {}({})".format(
num_gpus, args.config, train_config.TRAIN.num_gpus)
bs_denominator = train_config.TRAIN.num_gpus
train_config.TRAIN.batch_size = int(train_config.TRAIN.batch_size /
bs_denominator)
train_reader = KineticsReader(mode="train", cfg=train_config)
train_reader = train_reader.create_reader()
if use_data_parallel:
train_reader = fluid.contrib.reader.distributed_batch_reader(
train_reader)
for epoch in range(train_config.TRAIN.epoch):
video_model.train()
total_loss = 0.0
total_acc1 = 0.0
total_acc5 = 0.0
total_sample = 0
for batch_id, data in enumerate(train_reader()):
x_data = np.array([item[0] for item in data])
y_data = np.array([item[1] for item in data]).reshape([-1, 1])
imgs = to_variable(x_data)
labels = to_variable(y_data)
labels.stop_gradient = True
outputs = video_model(imgs)
loss = fluid.layers.cross_entropy(input=outputs,
label=labels,
ignore_index=-1)
avg_loss = fluid.layers.mean(loss)
acc_top1 = fluid.layers.accuracy(input=outputs,
label=labels,
k=1)
acc_top5 = fluid.layers.accuracy(input=outputs,
label=labels,
k=5)
if use_data_parallel:
avg_loss = video_model.scale_loss(avg_loss)
avg_loss.backward()
video_model.apply_collective_grads()
else:
avg_loss.backward()
optimizer.minimize(avg_loss)
video_model.clear_gradients()
total_loss += avg_loss.numpy()[0]
total_acc1 += acc_top1.numpy()[0]
total_acc5 += acc_top5.numpy()[0]
total_sample += 1
print('TRAIN Epoch {}, iter {}, loss = {}, acc1 {}, acc5 {}'.
format(epoch, batch_id,
avg_loss.numpy()[0],
acc_top1.numpy()[0],
acc_top5.numpy()[0]))
print(
'TRAIN End, Epoch {}, avg_loss= {}, avg_acc1= {}, avg_acc5= {}'
.format(epoch, total_loss / total_sample,
total_acc1 / total_sample, total_acc5 / total_sample))
video_model.eval()
val(epoch, video_model, valid_config, args)
if fluid.dygraph.parallel.Env().local_rank == 0:
fluid.dygraph.save_dygraph(video_model.state_dict(), "final")
logger.info('[TRAIN] training finished')
def test_grad(self):
places = [fluid.CPUPlace()]
if core.is_compiled_with_cuda():
places.append(fluid.CUDAPlace(0))
for p in places:
self.func(p)
def main(args):
ernie_config = ErnieConfig(args.ernie_config_path)
ernie_config.print_config()
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=False,
is_inference=True)
predict_prog = fluid.Program()
predict_startup = fluid.Program()
with fluid.program_guard(predict_prog, predict_startup):
with fluid.unique_name.guard():
predict_pyreader, probs, feed_target_names = create_model(
args,
pyreader_name='predict_reader',
ernie_config=ernie_config,
is_classify=True,
is_prediction=True,
ernie_version=args.ernie_version)
predict_prog = predict_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()))
place = fluid.CUDAPlace(0) if args.use_cuda == True else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(predict_startup)
if args.init_checkpoint:
init_pretraining_params(exe, args.init_checkpoint, predict_prog)
else:
raise ValueError(
"args 'init_checkpoint' should be set for prediction!")
#保存模型
assert args.save_inference_model_path, "args save_inference_model_path should be set for prediction"
_, ckpt_dir = os.path.split(args.init_checkpoint.rstrip('/'))
dir_name = ckpt_dir + '_inference_model'
model_path = os.path.join(args.save_inference_model_path, dir_name)
print("save inference model to %s" % model_path)
fluid.io.save_inference_model(model_path,
feed_target_names, [probs],
exe,
main_program=predict_prog)
#载入模型
print("load inference model from %s" % model_path)
infer_program, feed_target_names, probs = fluid.io.load_inference_model(
model_path, exe)
src_ids = feed_target_names[0]
sent_ids = feed_target_names[1]
pos_ids = feed_target_names[2]
input_mask = feed_target_names[3]
if args.ernie_version == "2.0":
task_ids = feed_target_names[4]
#计算相似度
predict_data_generator = reader.data_generator(input_file=args.predict_set,
batch_size=args.batch_size,
epoch=1,
shuffle=False)
print("-------------- prediction results --------------")
np.set_printoptions(precision=4, suppress=True)
index = 0
for sample in predict_data_generator():
src_ids_data = sample[0]
sent_ids_data = sample[1]
pos_ids_data = sample[2]
task_ids_data = sample[3]
input_mask_data = sample[4]
if args.ernie_version == "1.0":
output = exe.run(infer_program,
feed={
src_ids: src_ids_data,
sent_ids: sent_ids_data,
pos_ids: pos_ids_data,
input_mask: input_mask_data
},
fetch_list=probs)
elif args.ernie_version == "2.0":
output = exe.run(infer_program,
feed={
src_ids: src_ids_data,
sent_ids: sent_ids_data,
pos_ids: pos_ids_data,
task_ids: task_ids_data,
input_mask: input_mask_data
},
fetch_list=probs)
else:
raise ValueError("ernie_version must be 1.0 or 2.0")
for single_result in output[0]:
print("example_index:{}\t{}".format(index, single_result))
index += 1
def train(use_cuda):
classdim = 102
data_shape = [3, 224, 224]
images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
net = resnet_cifar10(images, 50)
predict = fluid.layers.fc(input=net, size=classdim, act='softmax')
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(cost)
acc = fluid.layers.accuracy(input=predict, label=label)
# Test program
test_program = fluid.default_main_program().clone(for_test=True)
optimizer = fluid.optimizer.Adam(learning_rate=0.001)
optimizer.minimize(avg_cost)
BATCH_SIZE = 20
PASS_NUM = 50
train_reader = paddle.batch(paddle.reader.shuffle(
paddle.dataset.flowers.train(), buf_size=128 * 10),
batch_size=BATCH_SIZE)
test_reader = paddle.batch(paddle.dataset.flowers.test(),
batch_size=BATCH_SIZE)
num_trainers = 1
trainer_id = 0
# ========================= for nccl2 dist train =================================
if os.getenv("PADDLE_INIT_TRAINER_ID", None) != None:
# append gen_nccl_id at the end of startup program
trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
print("using trainer_id: ", trainer_id)
port = os.getenv("PADDLE_INIT_PORT")
worker_ips = os.getenv("PADDLE_WORKERS")
worker_endpoints = []
for ip in worker_ips.split(","):
print(ip)
worker_endpoints.append(':'.join([ip, port]))
num_trainers = len(worker_endpoints)
current_endpoint = os.getenv("POD_IP") + ":" + port
worker_endpoints.remove(current_endpoint)
nccl_id_var = fluid.default_startup_program().global_block(
).create_var(name="NCCLID",
persistable=True,
type=fluid.core.VarDesc.VarType.RAW)
fluid.default_startup_program().global_block().append_op(
type="gen_nccl_id",
inputs={},
outputs={"NCCLID": nccl_id_var},
attrs={
"endpoint": current_endpoint,
"endpoint_list": worker_endpoints,
"trainer_id": trainer_id
})
# ========================= for nccl2 dist train =================================
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
feeder = fluid.DataFeeder(place=place, feed_list=[images, label])
def test(pass_id, exe):
acc_list = []
avg_loss_list = []
for tid, test_data in enumerate(test_reader()):
loss_t, acc_t = exe.run(program=test_program,
feed=feeder.feed(test_data),
fetch_list=[avg_cost, acc])
if math.isnan(float(loss_t)):
sys.exit("got NaN loss, training failed.")
acc_list.append(float(acc_t))
avg_loss_list.append(float(loss_t))
acc_value = numpy.array(acc_list).mean()
avg_loss_value = numpy.array(avg_loss_list).mean()
print('PassID {0:1}, Test Loss {1:2.2}, Acc {2:2.2}'.format(
pass_id, float(avg_loss_value), float(acc_value)))
def train_loop(main_program):
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
for pass_id in range(PASS_NUM):
pass_start = time.time()
num_samples = 0
for batch_id, data in enumerate(train_reader()):
exe.run(main_program, feed=feeder.feed(data))
num_samples += len(data)
pass_spent = time.time() - pass_start
print("Pass id %d, train speed %f" %
(pass_id, num_samples / pass_spent))
test(pass_id, exe)
def train_loop_parallel(main_program):
startup_exe = fluid.Executor(place)
startup_exe.run(fluid.default_startup_program())
exe = fluid.ParallelExecutor(True,
avg_cost.name,
num_threads=1,
allow_op_delay=False,
num_trainers=num_trainers,
trainer_id=trainer_id)
feeder = fluid.DataFeeder(place=place, feed_list=[images, label])
for pass_id in range(PASS_NUM):
num_samples = 0
start_time = time.time()
for batch_id, data in enumerate(train_reader()):
loss, = exe.run([avg_cost.name], feed=feeder.feed(data))
num_samples += len(data)
if batch_id % 1 == 0:
print("Pass %d, batch %d, loss %s" %
(pass_id, batch_id, np.array(loss)))
pass_elapsed = time.time() - start_time
print("Pass = %d, Training performance = %f imgs/s\n" %
(pass_id, num_samples / pass_elapsed))
test(pass_id, startup_exe)
train_loop_parallel(fluid.default_main_program())
def setUp(self):
self.input = numpy.ones(5).astype("int32")
self.place = fluid.CUDAPlace(
0) if fluid.is_compiled_with_cuda() else fluid.CPUPlace()
self.init_test_func()
def main(use_cuda):
"""
Advbox demo which demonstrate how to use advbox.
"""
class_dim = 1000
IMG_NAME = 'img'
LABEL_NAME = 'label'
#模型路径
pretrained_model = "models/resnet_50/115"
with_memory_optimization = 1
image_shape = [3,224,224]
image = fluid.layers.data(name=IMG_NAME, shape=image_shape, dtype='float32')
# gradient should flow
image.stop_gradient = False
label = fluid.layers.data(name=LABEL_NAME, shape=[1], dtype='int64')
# model definition
model = resnet.ResNet50()
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())
# 根据配置选择使用CPU资源还是GPU资源
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
#加载模型参数
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)
cost = fluid.layers.cross_entropy(input=out, label=label)
avg_cost = fluid.layers.mean(x=cost)
# advbox demo
m = PaddleModel(
fluid.default_main_program(),
IMG_NAME,
LABEL_NAME,
out.name,
avg_cost.name, (-1, 1),
channel_axis=3)
attack = FGSM(m)
attack_config = {"epsilons": 0.3}
test_data = get_image("cat.jpg")
# 猫对应的标签
test_label = 285
adversary = Adversary(test_data, test_label)
# FGSM non-targeted attack
adversary = attack(adversary, **attack_config)
if adversary.is_successful():
print(
'attack success, original_label=%d, adversarial_label=%d'
% (test_label, adversary.adversarial_label))
else:
print('attack failed, original_label=%d, ' % (test_label))
print("fgsm attack done")