def main():
# Run code with static graph mode.
try:
paddle.enable_static()
except:
pass
config = program.load_config(FLAGS.config)
program.merge_config(FLAGS.opt)
logger.info(config)
# check if set use_gpu=True in paddlepaddle cpu version
use_gpu = config['Global']['use_gpu']
program.check_gpu(use_gpu)
alg = config['Global']['algorithm']
assert alg in ['EAST', 'DB', 'Rosetta', 'CRNN', 'STARNet', 'RARE']
if alg in ['Rosetta', 'CRNN', 'STARNet', 'RARE']:
config['Global']['char_ops'] = CharacterOps(config['Global'])
place = fluid.CUDAPlace(0) if use_gpu else fluid.CPUPlace()
startup_prog = fluid.Program()
eval_program = fluid.Program()
eval_build_outputs = program.build(config,
eval_program,
startup_prog,
mode='test')
eval_fetch_name_list = eval_build_outputs[1]
eval_fetch_varname_list = eval_build_outputs[2]
eval_program = eval_program.clone(for_test=True)
exe = fluid.Executor(place)
exe.run(startup_prog)
init_model(config, eval_program, exe)
eval_reader = reader_main(config=config, mode="eval")
eval_info_dict = {'program':eval_program,\
'reader':eval_reader,\
'fetch_name_list':eval_fetch_name_list,\
'fetch_varname_list':eval_fetch_varname_list}
eval_args = dict()
eval_args = {
'exe': exe,
'config': config,
'eval_info_dict': eval_info_dict
}
metrics = eval_function(eval_args)
print("Baseline: {}".format(metrics))
params = get_pruned_params(eval_program)
print('Start to analyze')
sens_0 = slim.prune.sensitivity(
eval_program,
place,
params,
eval_function,
sensitivities_file="sensitivities_0.data",
pruned_ratios=[0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8],
eval_args=eval_args,
criterion='geometry_median')
def main():
FLAGS = ArgsParser().parse_args()
config = load_config(FLAGS.config)
merge_config(FLAGS.opt)
logger = get_logger()
# build post process
post_process_class = build_post_process(config['PostProcess'],
config['Global'])
# build model
# for rec algorithm
if hasattr(post_process_class, 'character'):
char_num = len(getattr(post_process_class, 'character'))
config['Architecture']["Head"]['out_channels'] = char_num
model = build_model(config['Architecture'])
init_model(config, model, logger)
model.eval()
save_path = '{}/inference'.format(config['Global']['save_inference_dir'])
infer_shape = [3, int(FLAGS.height), int(FLAGS.width)]
if config['Architecture']['model_type'] == "rec":
infer_shape = [3, 32, -1]
model = to_static(
model,
input_spec=[
paddle.static.InputSpec(
shape=[None] + infer_shape, dtype='float32')
])
paddle.jit.save(model, save_path)
logger.info('inference model is saved to {}'.format(save_path))
def main():
FLAGS = ArgsParser().parse_args()
config = load_config(FLAGS.config)
merge_config(FLAGS.opt)
logger = get_logger()
# build post process
post_process_class = build_post_process(config['PostProcess'],
config['Global'])
# build model
# for rec algorithm
if hasattr(post_process_class, 'character'):
char_num = len(getattr(post_process_class, 'character'))
config['Architecture']["Head"]['out_channels'] = char_num
model = build_model(config['Architecture'])
init_model(config, model, logger)
model.eval()
save_path = '{}/inference'.format(config['Global']['save_inference_dir'])
if config['Architecture']['algorithm'] == "SRN":
max_text_length = config['Architecture']['Head']['max_text_length']
other_shape = [
paddle.static.InputSpec(shape=[None, 1, 64, 256], dtype='float32'),
[
paddle.static.InputSpec(shape=[None, 256, 1], dtype="int64"),
paddle.static.InputSpec(shape=[None, max_text_length, 1],
dtype="int64"),
paddle.static.InputSpec(
shape=[None, 8, max_text_length, max_text_length],
dtype="int64"),
paddle.static.InputSpec(
shape=[None, 8, max_text_length, max_text_length],
dtype="int64")
]
]
model = to_static(model, input_spec=other_shape)
else:
infer_shape = [3, -1, -1]
if config['Architecture']['model_type'] == "rec":
infer_shape = [3, 32, -1] # for rec model, H must be 32
if 'Transform' in config['Architecture'] and config[
'Architecture']['Transform'] is not None and config[
'Architecture']['Transform']['name'] == 'TPS':
logger.info(
'When there is tps in the network, variable length input is not supported, and the input size needs to be the same as during training'
)
infer_shape[-1] = 100
model = to_static(model,
input_spec=[
paddle.static.InputSpec(shape=[None] +
infer_shape,
dtype='float32')
])
paddle.jit.save(model, save_path)
logger.info('inference model is saved to {}'.format(save_path))
def main():
config = program.load_config(FLAGS.config)
program.merge_config(FLAGS.opt)
logger.info(config)
# check if set use_gpu=True in paddlepaddle cpu version
use_gpu = config['Global']['use_gpu']
# check_gpu(use_gpu)
place = fluid.CUDAPlace(0) if use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
rec_model = create_module(
config['Architecture']['function'])(params=config)
startup_prog = fluid.Program()
eval_prog = fluid.Program()
with fluid.program_guard(eval_prog, startup_prog):
with fluid.unique_name.guard():
_, outputs = rec_model(mode="test")
fetch_name_list = list(outputs.keys())
fetch_varname_list = [outputs[v].name for v in fetch_name_list]
eval_prog = eval_prog.clone(for_test=True)
exe.run(startup_prog)
init_model(config, eval_prog, exe)
blobs = reader_main(config, 'test')()
infer_img = config['Global']['infer_img']
infer_list = get_image_file_list(infer_img)
max_img_num = len(infer_list)
if len(infer_list) == 0:
logger.info("Can not find img in infer_img dir.")
for i in range(max_img_num):
logger.info("infer_img:%s" % infer_list[i])
img = next(blobs)
predict = exe.run(program=eval_prog,
feed={"image": img},
fetch_list=fetch_varname_list,
return_numpy=False)
scores = np.array(predict[0])
label = np.array(predict[1])
if len(label.shape) != 1:
label, scores = scores, label
logger.info('\t scores: {}'.format(scores))
logger.info('\t label: {}'.format(label))
# save for inference model
target_var = []
for key, values in outputs.items():
target_var.append(values)
fluid.io.save_inference_model("./output",
feeded_var_names=['image'],
target_vars=target_var,
executor=exe,
main_program=eval_prog,
model_filename="model",
params_filename="params")
def test_reader():
config = program.load_config(FLAGS.config)
program.merge_config(FLAGS.opt)
logger.info(config)
train_reader = reader_main(config=config, mode="train")
import time
starttime = time.time()
count = 0
try:
for data in train_reader():
count += 1
if count % 1 == 0:
batch_time = time.time() - starttime
starttime = time.time()
logger.info("reader:", count, len(data), batch_time)
except Exception as e:
logger.info(e)
logger.info("finish reader: {}, Success!".format(count))
def main():
FLAGS = ArgsParser().parse_args()
config = load_config(FLAGS.config)
merge_config(FLAGS.opt)
logger = get_logger()
# build post process
post_process_class = build_post_process(config["PostProcess"],
config["Global"])
# build model
# for rec algorithm
if hasattr(post_process_class, "character"):
char_num = len(getattr(post_process_class, "character"))
if config["Architecture"]["algorithm"] in [
"Distillation",
]: # distillation model
for key in config["Architecture"]["Models"]:
config["Architecture"]["Models"][key]["Head"][
"out_channels"] = char_num
# just one final tensor needs to to exported for inference
config["Architecture"]["Models"][key][
"return_all_feats"] = False
else: # base rec model
config["Architecture"]["Head"]["out_channels"] = char_num
model = build_model(config["Architecture"])
_ = load_dygraph_params(config, model, logger, None)
model.eval()
save_path = config["Global"]["save_inference_dir"]
arch_config = config["Architecture"]
if arch_config["algorithm"] in [
"Distillation",
]: # distillation model
archs = list(arch_config["Models"].values())
for idx, name in enumerate(model.model_name_list):
sub_model_save_path = os.path.join(save_path, name, "inference")
export_single_model(model.model_list[idx], archs[idx],
sub_model_save_path, logger)
else:
save_path = os.path.join(save_path, "inference")
export_single_model(model, arch_config, save_path, logger)
def main():
############################################################################################################
# 1. quantization configs
############################################################################################################
quant_config = {
# weight preprocess type, default is None and no preprocessing is performed.
'weight_preprocess_type': None,
# activation preprocess type, default is None and no preprocessing is performed.
'activation_preprocess_type': None,
# 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,
# data type after quantization, such as 'uint8', 'int8', etc. default is 'int8'
'dtype': 'int8',
# window size for 'range_abs_max' quantization. default is 10000
'window_size': 10000,
# The decay coefficient of moving average, default is 0.9
'moving_rate': 0.9,
# for dygraph quantization, layers of type in quantizable_layer_type will be quantized
'quantizable_layer_type': ['Conv2D', 'Linear'],
}
FLAGS = ArgsParser().parse_args()
config = load_config(FLAGS.config)
merge_config(FLAGS.opt)
logger = get_logger()
# build post process
post_process_class = build_post_process(config['PostProcess'],
config['Global'])
# build model
# for rec algorithm
if hasattr(post_process_class, 'character'):
char_num = len(getattr(post_process_class, 'character'))
if config['Architecture']["algorithm"] in [
"Distillation",
]: # distillation model
for key in config['Architecture']["Models"]:
config['Architecture']["Models"][key]["Head"][
'out_channels'] = char_num
else: # base rec model
config['Architecture']["Head"]['out_channels'] = char_num
model = build_model(config['Architecture'])
# get QAT model
quanter = QAT(config=quant_config)
quanter.quantize(model)
init_model(config, model)
model.eval()
# build metric
eval_class = build_metric(config['Metric'])
# build dataloader
valid_dataloader = build_dataloader(config, 'Eval', device, logger)
use_srn = config['Architecture']['algorithm'] == "SRN"
model_type = config['Architecture']['model_type']
# start eval
metric = program.eval(model, valid_dataloader, post_process_class,
eval_class, model_type, use_srn)
logger.info('metric eval ***************')
for k, v in metric.items():
logger.info('{}:{}'.format(k, v))
infer_shape = [
3, 32, 100
] if config['Architecture']['model_type'] != "det" else [3, 640, 640]
save_path = config["Global"]["save_inference_dir"]
arch_config = config["Architecture"]
if arch_config["algorithm"] in [
"Distillation",
]: # distillation model
for idx, name in enumerate(model.model_name_list):
sub_model_save_path = os.path.join(save_path, name, "inference")
export_single_model(quanter, model.model_list[idx], infer_shape,
sub_model_save_path, logger)
else:
save_path = os.path.join(save_path, "inference")
export_single_model(quanter, model, infer_shape, save_path, logger)
def main():
config = program.load_config(FLAGS.config)
program.merge_config(FLAGS.opt)
logger.info(config)
char_ops = CharacterOps(config['Global'])
loss_type = config['Global']['loss_type']
config['Global']['char_ops'] = char_ops
# check if set use_gpu=True in paddlepaddle cpu version
use_gpu = config['Global']['use_gpu']
# check_gpu(use_gpu)
place = fluid.CUDAPlace(0) if use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
rec_model = create_module(
config['Architecture']['function'])(params=config)
startup_prog = fluid.Program()
eval_prog = fluid.Program()
with fluid.program_guard(eval_prog, startup_prog):
with fluid.unique_name.guard():
_, outputs = rec_model(mode="test")
fetch_name_list = list(outputs.keys())
fetch_varname_list = [outputs[v].name for v in fetch_name_list]
eval_prog = eval_prog.clone(for_test=True)
exe.run(startup_prog)
init_model(config, eval_prog, exe)
blobs = reader_main(config, 'test')()
infer_img = config['Global']['infer_img']
infer_list = get_image_file_list(infer_img)
max_img_num = len(infer_list)
if len(infer_list) == 0:
logger.info("Can not find img in infer_img dir.")
for i in range(max_img_num):
logger.info("infer_img:%s" % infer_list[i])
img = next(blobs)
if loss_type != "srn":
predict = exe.run(program=eval_prog,
feed={"image": img},
fetch_list=fetch_varname_list,
return_numpy=False)
else:
encoder_word_pos_list = []
gsrm_word_pos_list = []
gsrm_slf_attn_bias1_list = []
gsrm_slf_attn_bias2_list = []
encoder_word_pos_list.append(img[1])
gsrm_word_pos_list.append(img[2])
gsrm_slf_attn_bias1_list.append(img[3])
gsrm_slf_attn_bias2_list.append(img[4])
encoder_word_pos_list = np.concatenate(encoder_word_pos_list,
axis=0).astype(np.int64)
gsrm_word_pos_list = np.concatenate(gsrm_word_pos_list,
axis=0).astype(np.int64)
gsrm_slf_attn_bias1_list = np.concatenate(gsrm_slf_attn_bias1_list,
axis=0).astype(
np.float32)
gsrm_slf_attn_bias2_list = np.concatenate(gsrm_slf_attn_bias2_list,
axis=0).astype(
np.float32)
predict = exe.run(program=eval_prog, \
feed={'image': img[0], 'encoder_word_pos': encoder_word_pos_list,
'gsrm_word_pos': gsrm_word_pos_list, 'gsrm_slf_attn_bias1': gsrm_slf_attn_bias1_list,
'gsrm_slf_attn_bias2': gsrm_slf_attn_bias2_list}, \
fetch_list=fetch_varname_list, \
return_numpy=False)
if loss_type == "ctc":
preds = np.array(predict[0])
preds = preds.reshape(-1)
preds_lod = predict[0].lod()[0]
preds_text = char_ops.decode(preds)
probs = np.array(predict[1])
ind = np.argmax(probs, axis=1)
blank = probs.shape[1]
valid_ind = np.where(ind != (blank - 1))[0]
if len(valid_ind) == 0:
continue
score = np.mean(probs[valid_ind, ind[valid_ind]])
elif loss_type == "attention":
preds = np.array(predict[0])
probs = np.array(predict[1])
end_pos = np.where(preds[0, :] == 1)[0]
if len(end_pos) <= 1:
preds = preds[0, 1:]
score = np.mean(probs[0, 1:])
else:
preds = preds[0, 1:end_pos[1]]
score = np.mean(probs[0, 1:end_pos[1]])
preds = preds.reshape(-1)
preds_text = char_ops.decode(preds)
elif loss_type == "srn":
char_num = char_ops.get_char_num()
preds = np.array(predict[0])
preds = preds.reshape(-1)
probs = np.array(predict[1])
ind = np.argmax(probs, axis=1)
valid_ind = np.where(preds != int(char_num - 1))[0]
if len(valid_ind) == 0:
continue
score = np.mean(probs[valid_ind, ind[valid_ind]])
preds = preds[:valid_ind[-1] + 1]
preds_text = char_ops.decode(preds)
logger.info("\t index: {}".format(preds))
logger.info("\t word : {}".format(preds_text))
logger.info("\t score: {}".format(score))
# save for inference model
target_var = []
for key, values in outputs.items():
target_var.append(values)
fluid.io.save_inference_model("./output/",
feeded_var_names=['image'],
target_vars=target_var,
executor=exe,
main_program=eval_prog,
model_filename="model",
params_filename="params")
def main():
config = program.load_config(FLAGS.config)
program.merge_config(FLAGS.opt)
logger.info(config)
# check if set use_gpu=True in paddlepaddle cpu version
use_gpu = config['Global']['use_gpu']
program.check_gpu(use_gpu)
alg = config['Global']['algorithm']
assert alg in ['EAST', 'DB', 'Rosetta', 'CRNN', 'STARNet', 'RARE']
if alg in ['Rosetta', 'CRNN', 'STARNet', 'RARE']:
config['Global']['char_ops'] = CharacterOps(config['Global'])
place = fluid.CUDAPlace(0) if use_gpu else fluid.CPUPlace()
startup_program = fluid.Program()
train_program = fluid.Program()
train_build_outputs = program.build(config,
train_program,
startup_program,
mode='train')
train_loader = train_build_outputs[0]
train_fetch_name_list = train_build_outputs[1]
train_fetch_varname_list = train_build_outputs[2]
train_opt_loss_name = train_build_outputs[3]
eval_program = fluid.Program()
eval_build_outputs = program.build(config,
eval_program,
startup_program,
mode='eval')
eval_fetch_name_list = eval_build_outputs[1]
eval_fetch_varname_list = eval_build_outputs[2]
eval_program = eval_program.clone(for_test=True)
train_reader = reader_main(config=config, mode="train")
train_loader.set_sample_list_generator(train_reader, places=place)
eval_reader = reader_main(config=config, mode="eval")
exe = fluid.Executor(place)
exe.run(startup_program)
# compile program for multi-devices
train_compile_program = program.create_multi_devices_program(
train_program, train_opt_loss_name)
# dump mode structure
if config['Global']['debug']:
if 'attention' in config['Global']['loss_type']:
logger.warning('Does not suport dump attention...')
else:
summary(train_program)
init_model(config, train_program, exe)
train_info_dict = {'compile_program':train_compile_program,\
'train_program':train_program,\
'reader':train_loader,\
'fetch_name_list':train_fetch_name_list,\
'fetch_varname_list':train_fetch_varname_list}
eval_info_dict = {'program':eval_program,\
'reader':eval_reader,\
'fetch_name_list':eval_fetch_name_list,\
'fetch_varname_list':eval_fetch_varname_list}
if alg in ['EAST', 'DB']:
program.train_eval_det_run(config, exe, train_info_dict,
eval_info_dict)
else:
program.train_eval_rec_run(config, exe, train_info_dict,
eval_info_dict)
def main():
config = program.load_config(FLAGS.config)
program.merge_config(FLAGS.opt)
logger.info(config)
# check if set use_gpu=True in paddlepaddle cpu version
use_gpu = config['Global']['use_gpu']
program.check_gpu(use_gpu)
alg = config['Global']['algorithm']
assert alg in ['EAST', 'DB', 'Rosetta', 'CRNN', 'STARNet', 'RARE']
if alg in ['Rosetta', 'CRNN', 'STARNet', 'RARE']:
config['Global']['char_ops'] = CharacterOps(config['Global'])
place = fluid.CUDAPlace(0) if use_gpu else fluid.CPUPlace()
startup_program = fluid.Program()
train_program = fluid.Program()
train_build_outputs = program.build(config,
train_program,
startup_program,
mode='train')
train_loader = train_build_outputs[0]
train_fetch_name_list = train_build_outputs[1]
train_fetch_varname_list = train_build_outputs[2]
train_opt_loss_name = train_build_outputs[3]
eval_program = fluid.Program()
eval_build_outputs = program.build(config,
eval_program,
startup_program,
mode='eval')
eval_fetch_name_list = eval_build_outputs[1]
eval_fetch_varname_list = eval_build_outputs[2]
eval_program = eval_program.clone(for_test=True)
train_reader = reader_main(config=config, mode="train")
train_loader.set_sample_list_generator(train_reader, places=place)
eval_reader = reader_main(config=config, mode="eval")
exe = fluid.Executor(place)
exe.run(startup_program)
# compile program for multi-devices
init_model(config, train_program, exe)
sen = load_sensitivities("sensitivities_0.data")
for i in skip_list:
if i in sen.keys():
sen.pop(i)
back_bone_list = ['conv' + str(x) for x in range(1, 5)]
for i in back_bone_list:
for key in list(sen.keys()):
if i + '_' in key:
sen.pop(key)
ratios = get_ratios_by_loss(sen, 0.03)
logger.info("FLOPs before pruning: {}".format(flops(eval_program)))
pruner = Pruner(criterion='geometry_median')
print("ratios: {}".format(ratios))
pruned_val_program, _, _ = pruner.prune(eval_program,
fluid.global_scope(),
params=ratios.keys(),
ratios=ratios.values(),
place=place,
only_graph=True)
pruned_program, _, _ = pruner.prune(train_program,
fluid.global_scope(),
params=ratios.keys(),
ratios=ratios.values(),
place=place)
logger.info("FLOPs after pruning: {}".format(flops(pruned_val_program)))
train_compile_program = program.create_multi_devices_program(
pruned_program, train_opt_loss_name)
train_info_dict = {'compile_program':train_compile_program,\
'train_program':pruned_program,\
'reader':train_loader,\
'fetch_name_list':train_fetch_name_list,\
'fetch_varname_list':train_fetch_varname_list}
eval_info_dict = {'program':pruned_val_program,\
'reader':eval_reader,\
'fetch_name_list':eval_fetch_name_list,\
'fetch_varname_list':eval_fetch_varname_list}
if alg in ['EAST', 'DB']:
program.train_eval_det_run(config,
exe,
train_info_dict,
eval_info_dict,
is_slim="prune")
else:
program.train_eval_rec_run(config, exe, train_info_dict,
eval_info_dict)
def main():
############################################################################################################
# 1. quantization configs
############################################################################################################
quant_config = {
# weight preprocess type, default is None and no preprocessing is performed.
'weight_preprocess_type': None,
# activation preprocess type, default is None and no preprocessing is performed.
'activation_preprocess_type': None,
# 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,
# data type after quantization, such as 'uint8', 'int8', etc. default is 'int8'
'dtype': 'int8',
# window size for 'range_abs_max' quantization. default is 10000
'window_size': 10000,
# The decay coefficient of moving average, default is 0.9
'moving_rate': 0.9,
# for dygraph quantization, layers of type in quantizable_layer_type will be quantized
'quantizable_layer_type': ['Conv2D', 'Linear'],
}
FLAGS = ArgsParser().parse_args()
config = load_config(FLAGS.config)
merge_config(FLAGS.opt)
logger = get_logger()
# build post process
post_process_class = build_post_process(config['PostProcess'],
config['Global'])
# build model
# for rec algorithm
if hasattr(post_process_class, 'character'):
char_num = len(getattr(post_process_class, 'character'))
config['Architecture']["Head"]['out_channels'] = char_num
model = build_model(config['Architecture'])
# get QAT model
quanter = QAT(config=quant_config)
quanter.quantize(model)
init_model(config, model, logger)
model.eval()
# build metric
eval_class = build_metric(config['Metric'])
# build dataloader
valid_dataloader = build_dataloader(config, 'Eval', device, logger)
# start eval
metirc = program.eval(model, valid_dataloader, post_process_class,
eval_class)
logger.info('metric eval ***************')
for k, v in metirc.items():
logger.info('{}:{}'.format(k, v))
save_path = '{}/inference'.format(config['Global']['save_inference_dir'])
infer_shape = [
3, 32, 100
] if config['Architecture']['model_type'] != "det" else [3, 640, 640]
quanter.save_quantized_model(model,
save_path,
input_spec=[
paddle.static.InputSpec(shape=[None] +
infer_shape,
dtype='float32')
])
logger.info('inference QAT model is saved to {}'.format(save_path))
