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(config, device, logger, vdl_writer):
global_config = config['Global']
# build post process
post_process_class = build_post_process(config['PostProcess'],
global_config)
# build model
if hasattr(post_process_class, 'character'):
config['Architecture']["Head"]['out_channels'] = len(
getattr(post_process_class, 'character'))
model = build_model(config['Architecture'])
init_model(config, model, logger)
# create data ops
transforms = []
use_padding = False
for op in config['Eval']['dataset']['transforms']:
op_name = list(op)[0]
if 'Label' in op_name:
continue
if op_name == 'KeepKeys':
op[op_name]['keep_keys'] = ['image']
if op_name == "ResizeTableImage":
use_padding = True
padding_max_len = op['ResizeTableImage']['max_len']
transforms.append(op)
global_config['infer_mode'] = True
ops = create_operators(transforms, global_config)
model.eval()
for file in get_image_file_list(config['Global']['infer_img']):
logger.info("infer_img: {}".format(file))
with open(file, 'rb') as f:
img = f.read()
data = {'image': img}
batch = transform(data, ops)
images = np.expand_dims(batch[0], axis=0)
images = paddle.to_tensor(images)
preds = model(images)
post_result = post_process_class(preds)
res_html_code = post_result['res_html_code']
res_loc = post_result['res_loc']
img = cv2.imread(file)
imgh, imgw = img.shape[0:2]
res_loc_final = []
for rno in range(len(res_loc[0])):
x0, y0, x1, y1 = res_loc[0][rno]
left = max(int(imgw * x0), 0)
top = max(int(imgh * y0), 0)
right = min(int(imgw * x1), imgw - 1)
bottom = min(int(imgh * y1), imgh - 1)
cv2.rectangle(img, (left, top), (right, bottom), (0, 0, 255), 2)
res_loc_final.append([left, top, right, bottom])
res_loc_str = json.dumps(res_loc_final)
logger.info("result: {}, {}".format(res_html_code, res_loc_final))
logger.info("success!")
def main():
startup_prog, eval_program, place, config, train_alg_type = program.preprocess()
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)
if train_alg_type == 'det':
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}
metrics = eval_det_run(exe, config, eval_info_dict, "eval")
logger.info("Eval result: {}".format(metrics))
else:
reader_type = config['Global']['reader_yml']
if "benchmark" not in reader_type:
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}
metrics = eval_rec_run(exe, config, eval_info_dict, "eval")
logger.info("Eval result: {}".format(metrics))
else:
eval_info_dict = {'program':eval_program,\
'fetch_name_list':eval_fetch_name_list,\
'fetch_varname_list':eval_fetch_varname_list}
test_rec_benchmark(exe, config, 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(True)
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()
feeded_var_names, target_vars, fetches_var_name = program.build_export(
config, eval_program, startup_prog)
eval_program = eval_program.clone(for_test=True)
exe = fluid.Executor(place)
exe.run(startup_prog)
init_model(config, eval_program, exe)
fluid.io.save_inference_model(dirname="./output/",
feeded_var_names=feeded_var_names,
main_program=eval_program,
target_vars=target_vars,
executor=exe,
model_filename='model',
params_filename='params')
print("save success, output_name_list:", fetches_var_name)
def main():
# build train 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]
model_average = train_build_outputs[-1]
# build eval program
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)
# initialize train reader
train_reader = reader_main(config=config, mode="train")
train_loader.set_sample_list_generator(train_reader, places=place)
# initialize eval reader
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 train_alg_type == 'rec' and '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,\
'model_average': model_average}
eval_info_dict = {'program':eval_program,\
'reader':eval_reader,\
'fetch_name_list':eval_fetch_name_list,\
'fetch_varname_list':eval_fetch_varname_list}
if train_alg_type == 'det':
program.train_eval_det_run(config, exe, train_info_dict, eval_info_dict)
elif train_alg_type == 'rec':
program.train_eval_rec_run(config, exe, train_info_dict, eval_info_dict)
else:
program.train_eval_cls_run(config, exe, train_info_dict, eval_info_dict)
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'])
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 main():
global_config = config['Global']
# build model
model = build_model(config['Architecture'])
init_model(config, model, logger)
# build post process
post_process_class = build_post_process(config['PostProcess'],
global_config)
# create data ops
transforms = []
for op in config['Eval']['dataset']['transforms']:
op_name = list(op)[0]
if 'Label' in op_name:
continue
elif op_name == 'KeepKeys':
op[op_name]['keep_keys'] = ['image', 'shape']
transforms.append(op)
ops = create_operators(transforms, global_config)
save_res_path = config['Global']['save_res_path']
if not os.path.exists(os.path.dirname(save_res_path)):
os.makedirs(os.path.dirname(save_res_path))
model.eval()
with open(save_res_path, "wb") as fout:
for file in get_image_file_list(config['Global']['infer_img']):
logger.info("infer_img: {}".format(file))
with open(file, 'rb') as f:
img = f.read()
data = {'image': img}
batch = transform(data, ops)
images = np.expand_dims(batch[0], axis=0)
shape_list = np.expand_dims(batch[1], axis=0)
images = paddle.to_tensor(images)
preds = model(images)
post_result = post_process_class(preds, shape_list)
points, strs = post_result['points'], post_result['texts']
# write resule
dt_boxes_json = []
for poly, str in zip(points, strs):
tmp_json = {"transcription": str}
tmp_json['points'] = poly.tolist()
dt_boxes_json.append(tmp_json)
otstr = file + "\t" + json.dumps(dt_boxes_json) + "\n"
fout.write(otstr.encode())
src_img = cv2.imread(file)
draw_e2e_res(points, strs, config, src_img, file)
logger.info("success!")
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_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)
if alg in ['EAST', 'DB']:
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}
metrics = eval_det_run(exe, config, eval_info_dict, "eval")
logger.info("Eval result: {}".format(metrics))
else:
reader_type = config['Global']['reader_yml']
if "benchmark" not in reader_type:
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}
metrics = eval_rec_run(exe, config, eval_info_dict, "eval")
logger.info("Eval result: {}".format(metrics))
else:
eval_info_dict = {'program':eval_program,\
'fetch_name_list':eval_fetch_name_list,\
'fetch_varname_list':eval_fetch_varname_list}
test_rec_benchmark(exe, config, eval_info_dict)
def main():
global_config = config['Global']
# build dataloader
valid_dataloader = build_dataloader(config, 'Eval', device, logger)
# build post process
post_process_class = build_post_process(config['PostProcess'],
global_config)
# build model
# for rec algorithm
if hasattr(post_process_class, 'character'):
config['Architecture']["Head"]['out_channels'] = len(
getattr(post_process_class, 'character'))
model = build_model(config['Architecture'])
use_srn = config['Architecture']['algorithm'] == "SRN"
best_model_dict = init_model(config, model, logger)
if len(best_model_dict):
logger.info('metric in ckpt ***************')
for k, v in best_model_dict.items():
logger.info('{}:{}'.format(k, v))
# build metric
eval_class = build_metric(config['Metric'])
# start eval
metirc = program.eval(model, valid_dataloader, post_process_class,
eval_class, use_srn)
logger.info('metric eval ***************')
for k, v in metirc.items():
logger.info('{}:{}'.format(k, v))
def main(config, device, logger, vdl_writer):
# init dist environment
if config['Global']['distributed']:
dist.init_parallel_env()
global_config = config['Global']
# build dataloader
train_dataloader = build_dataloader(config, 'Train', device, logger)
if len(train_dataloader) == 0:
logger.error(
"No Images in train dataset, please ensure\n" +
"\t1. The images num in the train label_file_list should be larger than or equal with batch size.\n"
+
"\t2. The annotation file and path in the configuration file are provided normally."
)
return
if config['Eval']:
valid_dataloader = build_dataloader(config, 'Eval', device, logger)
else:
valid_dataloader = None
# build post process
post_process_class = build_post_process(config['PostProcess'],
global_config)
# 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'])
if config['Global']['distributed']:
model = paddle.DataParallel(model)
# build loss
loss_class = build_loss(config['Loss'])
# build optim
optimizer, lr_scheduler = build_optimizer(
config['Optimizer'],
epochs=config['Global']['epoch_num'],
step_each_epoch=len(train_dataloader),
parameters=model.parameters())
# build metric
eval_class = build_metric(config['Metric'])
# load pretrain model
pre_best_model_dict = init_model(config, model, logger, optimizer)
logger.info('train dataloader has {} iters'.format(len(train_dataloader)))
if valid_dataloader is not None:
logger.info('valid dataloader has {} iters'.format(
len(valid_dataloader)))
# start train
program.train(config, train_dataloader, valid_dataloader, device, model,
loss_class, optimizer, lr_scheduler, post_process_class,
eval_class, pre_best_model_dict, logger, vdl_writer)
def main():
global_config = config['Global']
# build post process
post_process_class = build_post_process(config['PostProcess'],
global_config)
# build model
if hasattr(post_process_class, 'character'):
config['Architecture']["Head"]['out_channels'] = len(
getattr(post_process_class, 'character'))
model = build_model(config['Architecture'])
init_model(config, model, logger)
# create data ops
transforms = []
for op in config['Eval']['dataset']['transforms']:
op_name = list(op)[0]
if 'Label' in op_name:
continue
elif op_name in ['RecResizeImg']:
op[op_name]['infer_mode'] = True
elif op_name == 'KeepKeys':
op[op_name]['keep_keys'] = ['image']
transforms.append(op)
global_config['infer_mode'] = True
ops = create_operators(transforms, global_config)
model.eval()
for file in get_image_file_list(config['Global']['infer_img']):
logger.info("infer_img: {}".format(file))
with open(file, 'rb') as f:
img = f.read()
data = {'image': img}
batch = transform(data, ops)
images = np.expand_dims(batch[0], axis=0)
images = paddle.to_tensor(images)
preds = model(images)
post_result = post_process_class(preds)
for rec_reuslt in post_result:
logger.info('\t result: {}'.format(rec_reuslt))
logger.info("success!")
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"])
init_model(config, model)
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(config, device, logger, vdl_writer):
# init dist environment
if config['Global']['distributed']:
dist.init_parallel_env()
global_config = config['Global']
# build dataloader
train_dataloader = build_dataloader(config, 'Train', device, logger)
if config['Eval']:
valid_dataloader = build_dataloader(config, 'Eval', device, logger)
else:
valid_dataloader = None
# build post process
post_process_class = build_post_process(config['PostProcess'],
global_config)
# 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'])
if config['Global']['distributed']:
model = paddle.DataParallel(model)
# build loss
loss_class = build_loss(config['Loss'])
# build optim
optimizer, lr_scheduler = build_optimizer(
config['Optimizer'],
epochs=config['Global']['epoch_num'],
step_each_epoch=len(train_dataloader),
parameters=model.parameters())
# build metric
eval_class = build_metric(config['Metric'])
# load pretrain model
pre_best_model_dict = init_model(config, model, logger, optimizer)
logger.info(
'train dataloader has {} iters, valid dataloader has {} iters'.format(
len(train_dataloader), len(valid_dataloader)))
quanter = QAT(config=quant_config, act_preprocess=PACT)
quanter.quantize(model)
# start train
program.train(config, train_dataloader, valid_dataloader, device, model,
loss_class, optimizer, lr_scheduler, post_process_class,
eval_class, pre_best_model_dict, logger, vdl_writer)
def main():
startup_prog, eval_program, place, config, _ = program.preprocess()
feeded_var_names, target_vars, fetches_var_name = program.build_export(
config, eval_program, startup_prog)
eval_program = eval_program.clone(for_test=True)
exe = fluid.Executor(place)
exe.run(startup_prog)
init_model(config, eval_program, exe)
save_inference_dir = config['Global']['save_inference_dir']
if not os.path.exists(save_inference_dir):
os.makedirs(save_inference_dir)
fluid.io.save_inference_model(dirname=save_inference_dir,
feeded_var_names=feeded_var_names,
main_program=eval_program,
target_vars=target_vars,
executor=exe,
model_filename='model',
params_filename='params')
print("inference model saved in {}/model and {}/params".format(
save_inference_dir, save_inference_dir))
print("save success, output_name_list:", fetches_var_name)
def main():
global_config = config['Global']
# build dataloader
valid_dataloader = build_dataloader(config, 'Eval', device, logger)
# build post process
post_process_class = build_post_process(config['PostProcess'],
global_config)
# 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'])
use_srn = config['Architecture']['algorithm'] == "SRN"
if "model_type" in config['Architecture'].keys():
model_type = config['Architecture']['model_type']
else:
model_type = None
best_model_dict = init_model(config, model)
if len(best_model_dict):
logger.info('metric in ckpt ***************')
for k, v in best_model_dict.items():
logger.info('{}:{}'.format(k, v))
# build metric
eval_class = build_metric(config['Metric'])
# 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))
def main():
global_config = config['Global']
# build post process
post_process_class = build_post_process(config['PostProcess'],
global_config)
# build model
if hasattr(post_process_class, 'character'):
config['Architecture']["Head"]['out_channels'] = len(
getattr(post_process_class, 'character'))
model = build_model(config['Architecture'])
init_model(config, model, logger)
# create data ops
transforms = []
for op in config['Eval']['dataset']['transforms']:
op_name = list(op)[0]
if 'Label' in op_name:
continue
elif op_name in ['RecResizeImg']:
op[op_name]['infer_mode'] = True
elif op_name == 'KeepKeys':
if config['Architecture']['algorithm'] == "SRN":
op[op_name]['keep_keys'] = [
'image', 'encoder_word_pos', 'gsrm_word_pos',
'gsrm_slf_attn_bias1', 'gsrm_slf_attn_bias2'
]
else:
op[op_name]['keep_keys'] = ['image']
transforms.append(op)
global_config['infer_mode'] = True
ops = create_operators(transforms, global_config)
save_res_path = config['Global'].get('save_res_path',
"./output/rec/predicts_rec.txt")
if not os.path.exists(os.path.dirname(save_res_path)):
os.makedirs(os.path.dirname(save_res_path))
model.eval()
with open(save_res_path, "w") as fout:
for file in get_image_file_list(config['Global']['infer_img']):
logger.info("infer_img: {}".format(file))
with open(file, 'rb') as f:
img = f.read()
data = {'image': img}
batch = transform(data, ops)
if config['Architecture']['algorithm'] == "SRN":
encoder_word_pos_list = np.expand_dims(batch[1], axis=0)
gsrm_word_pos_list = np.expand_dims(batch[2], axis=0)
gsrm_slf_attn_bias1_list = np.expand_dims(batch[3], axis=0)
gsrm_slf_attn_bias2_list = np.expand_dims(batch[4], axis=0)
others = [
paddle.to_tensor(encoder_word_pos_list),
paddle.to_tensor(gsrm_word_pos_list),
paddle.to_tensor(gsrm_slf_attn_bias1_list),
paddle.to_tensor(gsrm_slf_attn_bias2_list)
]
images = np.expand_dims(batch[0], axis=0)
images = paddle.to_tensor(images)
if config['Architecture']['algorithm'] == "SRN":
preds = model(images, others)
else:
preds = model(images)
post_result = post_process_class(preds)
for rec_result in post_result:
logger.info('\t result: {}'.format(rec_result))
if len(rec_result) >= 2:
fout.write(file + "\t" + rec_result[0] + "\t" +
str(rec_result[1]) + "\n")
logger.info("success!")
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))
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():
global_config = config['Global']
# build post process
post_process_class = build_post_process(config['PostProcess'],
global_config)
# build model
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'])
init_model(config, model)
# create data ops
transforms = []
for op in config['Eval']['dataset']['transforms']:
op_name = list(op)[0]
if 'Label' in op_name:
continue
elif op_name in ['RecResizeImg']:
op[op_name]['infer_mode'] = True
elif op_name == 'KeepKeys':
if config['Architecture']['algorithm'] == "SRN":
op[op_name]['keep_keys'] = [
'image', 'encoder_word_pos', 'gsrm_word_pos',
'gsrm_slf_attn_bias1', 'gsrm_slf_attn_bias2'
]
else:
op[op_name]['keep_keys'] = ['image']
transforms.append(op)
global_config['infer_mode'] = True
ops = create_operators(transforms, global_config)
save_res_path = config['Global'].get('save_res_path',
"./output/rec/predicts_rec.txt")
if not os.path.exists(os.path.dirname(save_res_path)):
os.makedirs(os.path.dirname(save_res_path))
model.eval()
with open(save_res_path, "w") as fout:
for file in get_image_file_list(config['Global']['infer_img']):
logger.info("infer_img: {}".format(file))
with open(file, 'rb') as f:
img = f.read()
data = {'image': img}
batch = transform(data, ops)
if config['Architecture']['algorithm'] == "SRN":
encoder_word_pos_list = np.expand_dims(batch[1], axis=0)
gsrm_word_pos_list = np.expand_dims(batch[2], axis=0)
gsrm_slf_attn_bias1_list = np.expand_dims(batch[3], axis=0)
gsrm_slf_attn_bias2_list = np.expand_dims(batch[4], axis=0)
others = [
paddle.to_tensor(encoder_word_pos_list),
paddle.to_tensor(gsrm_word_pos_list),
paddle.to_tensor(gsrm_slf_attn_bias1_list),
paddle.to_tensor(gsrm_slf_attn_bias2_list)
]
images = np.expand_dims(batch[0], axis=0)
images = paddle.to_tensor(images)
if config['Architecture']['algorithm'] == "SRN":
preds = model(images, others)
else:
preds = model(images)
post_result = post_process_class(preds)
info = None
if isinstance(post_result, dict):
rec_info = dict()
for key in post_result:
if len(post_result[key][0]) >= 2:
rec_info[key] = {
"label": post_result[key][0][0],
"score": float(post_result[key][0][1]),
}
info = json.dumps(rec_info)
else:
if len(post_result[0]) >= 2:
info = post_result[0][0] + "\t" + str(post_result[0][1])
if info is not None:
logger.info("\t result: {}".format(info))
fout.write(file + "\t" + info)
logger.info("success!")
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():
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]
model_average = train_build_outputs[-1]
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)
# 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,
}
# 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.
act_preprocess_func = pact
optimizer_func = get_optimizer
executor = exe
eval_program = quant_aware(eval_program,
place,
quant_config,
scope=None,
act_preprocess_func=act_preprocess_func,
optimizer_func=optimizer_func,
executor=executor,
for_test=True)
quant_train_program = quant_aware(train_program,
place,
quant_config,
scope=None,
act_preprocess_func=act_preprocess_func,
optimizer_func=optimizer_func,
executor=executor,
for_test=False)
# compile program for multi-devices
train_compile_program = program.create_multi_devices_program(
quant_train_program, train_opt_loss_name, for_quant=True)
init_model(config, train_program, exe)
train_info_dict = {'compile_program':train_compile_program,\
'train_program':quant_train_program,\
'reader':train_loader,\
'fetch_name_list':train_fetch_name_list,\
'fetch_varname_list':train_fetch_varname_list,\
'model_average': model_average}
eval_info_dict = {'program':eval_program,\
'reader':eval_reader,\
'fetch_name_list':eval_fetch_name_list,\
'fetch_varname_list':eval_fetch_varname_list}
if train_alg_type == 'det':
program.train_eval_det_run(config,
exe,
train_info_dict,
eval_info_dict,
is_slim="quant")
elif train_alg_type == 'rec':
program.train_eval_rec_run(config,
exe,
train_info_dict,
eval_info_dict,
is_slim="quant")
else:
program.train_eval_cls_run(config,
exe,
train_info_dict,
eval_info_dict,
is_slim="quant")
def main(config, device, logger, vdl_writer):
# init dist environment
if config['Global']['distributed']:
dist.init_parallel_env()
global_config = config['Global']
# build dataloader
train_dataloader = build_dataloader(config, 'Train', device, logger)
if config['Eval']:
valid_dataloader = build_dataloader(config, 'Eval', device, logger)
else:
valid_dataloader = None
# build post process
post_process_class = build_post_process(config['PostProcess'],
global_config)
# 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'])
flops = paddle.flops(model, [1, 3, 640, 640])
logger.info(f"FLOPs before pruning: {flops}")
from paddleslim.dygraph import FPGMFilterPruner
model.train()
pruner = FPGMFilterPruner(model, [1, 3, 640, 640])
# build loss
loss_class = build_loss(config['Loss'])
# build optim
optimizer, lr_scheduler = build_optimizer(
config['Optimizer'],
epochs=config['Global']['epoch_num'],
step_each_epoch=len(train_dataloader),
parameters=model.parameters())
# build metric
eval_class = build_metric(config['Metric'])
# load pretrain model
pre_best_model_dict = init_model(config, model, logger, optimizer)
logger.info(
'train dataloader has {} iters, valid dataloader has {} iters'.format(
len(train_dataloader), len(valid_dataloader)))
# build metric
eval_class = build_metric(config['Metric'])
logger.info(
'train dataloader has {} iters, valid dataloader has {} iters'.format(
len(train_dataloader), len(valid_dataloader)))
def eval_fn():
metric = program.eval(model, valid_dataloader, post_process_class,
eval_class)
logger.info(f"metric['hmean']: {metric['hmean']}")
return metric['hmean']
params_sensitive = pruner.sensitive(eval_func=eval_fn,
sen_file="./sen.pickle",
skip_vars=[
"conv2d_57.w_0",
"conv2d_transpose_2.w_0",
"conv2d_transpose_3.w_0"
])
logger.info(
"The sensitivity analysis results of model parameters saved in sen.pickle"
)
# calculate pruned params's ratio
params_sensitive = pruner._get_ratios_by_loss(params_sensitive, loss=0.02)
for key in params_sensitive.keys():
logger.info(f"{key}, {params_sensitive[key]}")
plan = pruner.prune_vars(params_sensitive, [0])
for param in model.parameters():
if ("weights" in param.name
and "conv" in param.name) or ("w_0" in param.name
and "conv2d" in param.name):
logger.info(f"{param.name}: {param.shape}")
flops = paddle.flops(model, [1, 3, 640, 640])
logger.info(f"FLOPs after pruning: {flops}")
# start train
program.train(config, train_dataloader, valid_dataloader, device, model,
loss_class, optimizer, lr_scheduler, post_process_class,
eval_class, pre_best_model_dict, logger, vdl_writer)
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, device, logger, vdl_writer):
global_config = config['Global']
# build dataloader
valid_dataloader = build_dataloader(config, 'Eval', device, logger)
# build post process
post_process_class = build_post_process(config['PostProcess'],
global_config)
# 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'])
flops = paddle.flops(model, [1, 3, 640, 640])
logger.info(f"FLOPs before pruning: {flops}")
from paddleslim.dygraph import FPGMFilterPruner
model.train()
pruner = FPGMFilterPruner(model, [1, 3, 640, 640])
# build metric
eval_class = build_metric(config['Metric'])
def eval_fn():
metric = program.eval(model, valid_dataloader, post_process_class,
eval_class)
logger.info(f"metric['hmean']: {metric['hmean']}")
return metric['hmean']
params_sensitive = pruner.sensitive(eval_func=eval_fn,
sen_file="./sen.pickle",
skip_vars=[
"conv2d_57.w_0",
"conv2d_transpose_2.w_0",
"conv2d_transpose_3.w_0"
])
logger.info(
"The sensitivity analysis results of model parameters saved in sen.pickle"
)
# calculate pruned params's ratio
params_sensitive = pruner._get_ratios_by_loss(params_sensitive, loss=0.02)
for key in params_sensitive.keys():
logger.info(f"{key}, {params_sensitive[key]}")
plan = pruner.prune_vars(params_sensitive, [0])
flops = paddle.flops(model, [1, 3, 640, 640])
logger.info(f"FLOPs after pruning: {flops}")
# load pretrain model
pre_best_model_dict = init_model(config, model, logger, None)
metric = program.eval(model, valid_dataloader, post_process_class,
eval_class)
logger.info(f"metric['hmean']: {metric['hmean']}")
# start export model
from paddle.jit import to_static
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')
])
save_path = '{}/inference'.format(config['Global']['save_inference_dir'])
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)
char_ops = CharacterOps(config['Global'])
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')
imgs = next(blobs())
for img in imgs:
predict = exe.run(program=eval_prog,
feed={"image": img},
fetch_list=fetch_varname_list,
return_numpy=False)
preds = np.array(predict[0])
if preds.shape[1] == 1:
preds = preds.reshape(-1)
preds_lod = predict[0].lod()[0]
preds_text = char_ops.decode(preds)
else:
end_pos = np.where(preds[0, :] == 1)[0]
if len(end_pos) <= 1:
preds_text = preds[0, 1:]
else:
preds_text = preds[0, 1:end_pos[1]]
preds_text = preds_text.reshape(-1)
preds_text = char_ops.decode(preds_text)
print(preds)
print(preds_text)
# 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():
# 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,
}
startup_prog, eval_program, place, config, alg_type = program.preprocess()
feeded_var_names, target_vars, fetches_var_name = program.build_export(
config, eval_program, startup_prog)
eval_program = eval_program.clone(for_test=True)
exe = fluid.Executor(place)
exe.run(startup_prog)
eval_program = quant_aware(
eval_program, place, quant_config, scope=None, for_test=True)
init_model(config, eval_program, exe)
# 2. Convert the program before save inference program
# The dtype of eval_program's weights is float32, but in int8 range.
eval_program = convert(eval_program, place, quant_config, scope=None)
eval_fetch_name_list = fetches_var_name
eval_fetch_varname_list = [v.name for v in target_vars]
eval_reader = reader_main(config=config, mode="eval")
quant_info_dict = {'program':eval_program,\
'reader':eval_reader,\
'fetch_name_list':eval_fetch_name_list,\
'fetch_varname_list':eval_fetch_varname_list}
if alg_type == 'det':
final_metrics = eval_det_run(exe, config, quant_info_dict, "eval")
else:
final_metrics = eval_rec_run(exe, config, quant_info_dict, "eval")
print(final_metrics)
# 3. Save inference model
model_path = "./quant_model"
if not os.path.isdir(model_path):
os.makedirs(model_path)
fluid.io.save_inference_model(
dirname=model_path,
feeded_var_names=feeded_var_names,
target_vars=target_vars,
executor=exe,
main_program=eval_program,
model_filename=model_path + '/model',
params_filename=model_path + '/params')
print("model saved as {}".format(model_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']
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)
