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(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):
# 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)