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(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():
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'])
extra_input = config['Architecture']['algorithm'] in [
"SRN", "NRTR", "SAR", "SEED"
]
if "model_type" in config['Architecture'].keys():
model_type = config['Architecture']['model_type']
else:
model_type = None
best_model_dict = load_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, extra_input)
logger.info('metric eval ***************')
for k, v in metric.items():
logger.info('{}:{}'.format(k, v))
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, 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'))
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'])
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 = load_model(config, model, 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)))
use_amp = config["Global"].get("use_amp", False)
if use_amp:
AMP_RELATED_FLAGS_SETTING = {
'FLAGS_cudnn_batchnorm_spatial_persistent': 1,
'FLAGS_max_inplace_grad_add': 8,
}
paddle.fluid.set_flags(AMP_RELATED_FLAGS_SETTING)
scale_loss = config["Global"].get("scale_loss", 1.0)
use_dynamic_loss_scaling = config["Global"].get(
"use_dynamic_loss_scaling", False)
scaler = paddle.amp.GradScaler(
init_loss_scaling=scale_loss,
use_dynamic_loss_scaling=use_dynamic_loss_scaling)
else:
scaler = None
# 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, scaler)
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, 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, 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['Architecture']['model_type'] == 'det':
input_shape = [1, 3, 640, 640]
elif config['Architecture']['model_type'] == 'rec':
input_shape = [1, 3, 32, 320]
flops = paddle.flops(model, input_shape)
logger.info("FLOPs before pruning: {}".format(flops))
from paddleslim.dygraph import FPGMFilterPruner
model.train()
pruner = FPGMFilterPruner(model, input_shape)
# 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 = load_model(config, model, 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, False)
if config['Architecture']['model_type'] == 'det':
main_indicator = 'hmean'
else:
main_indicator = 'acc'
logger.info("metric[{}]: {}".format(main_indicator,
metric[main_indicator]))
return metric[main_indicator]
run_sensitive_analysis = False
"""
run_sensitive_analysis=True:
Automatically compute the sensitivities of convolutions in a model.
The sensitivity of a convolution is the losses of accuracy on test dataset in
differenct pruned ratios. The sensitivities can be used to get a group of best
ratios with some condition.
run_sensitive_analysis=False:
Set prune trim ratio to a fixed value, such as 10%. The larger the value,
the more convolution weights will be cropped.
"""
if run_sensitive_analysis:
params_sensitive = pruner.sensitive(
eval_func=eval_fn,
sen_file="./deploy/slim/prune/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("{}, {}".format(key, params_sensitive[key]))
else:
params_sensitive = {}
for param in model.parameters():
if 'transpose' not in param.name and 'linear' not in param.name:
# set prune ratio as 10%. The larger the value, the more convolution weights will be cropped
params_sensitive[param.name] = 0.1
plan = pruner.prune_vars(params_sensitive, [0])
flops = paddle.flops(model, input_shape)
logger.info("FLOPs after pruning: {}".format(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():
############################################################################################################
# 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))
