def run(dataloader, exe, program, fetchs, epoch=0, mode='train'):
"""
Feed data to the model and fetch the measures and loss
Args:
dataloader(fluid dataloader):
exe():
program():
fetchs(dict): dict of measures and the loss
epoch(int): epoch of training or validation
model(str): log only
Returns:
"""
fetch_list = [f[0] for f in fetchs.values()]
metric_list = [f[1] for f in fetchs.values()]
for m in metric_list:
m.reset()
batch_time = AverageMeter('cost', ':6.3f')
tic = time.time()
trainer_id = int(os.getenv("PADDLE_TRAINER_ID", 0))
for idx, batch in enumerate(dataloader()):
metrics = exe.run(program=program, feed=batch, fetch_list=fetch_list)
batch_time.update(time.time() - tic)
tic = time.time()
for i, m in enumerate(metrics):
metric_list[i].update(m[0], len(batch[0]))
fetchs_str = ''.join([str(m) for m in metric_list] + [str(batch_time)])
if trainer_id == 0:
logger.info("[epoch:%3d][%s][step:%4d]%s" %
(epoch, mode, idx, fetchs_str))
if trainer_id == 0:
logger.info("END [epoch:%3d][%s]%s" % (epoch, mode, fetchs_str))
def run(dataloader, exe, program, fetchs, epoch=0, mode='train'):
"""
Feed data to the model and fetch the measures and loss
Args:
dataloader(fluid dataloader):
exe():
program():
fetchs(dict): dict of measures and the loss
epoch(int): epoch of training or validation
model(str): log only
Returns:
"""
fetch_list = [f[0] for f in fetchs.values()]
metric_list = [f[1] for f in fetchs.values()]
for m in metric_list:
m.reset()
batch_time = AverageMeter('cost', '.3f')
tic = time.time()
for idx, batch in enumerate(dataloader()):
metrics = exe.run(program=program, feed=batch, fetch_list=fetch_list)
batch_time.update(time.time() - tic)
tic = time.time()
for i, m in enumerate(metrics):
metric_list[i].update(m[0], len(batch[0]))
fetchs_str = ''.join([m.value
for m in metric_list] + [batch_time.value])
logger.info("[epoch:{:3d}][{:s}][step:{:4d}]{:s}".format(
epoch, mode, idx, fetchs_str))
end_str = ''.join([m.mean for m in metric_list] + [batch_time.total])
logger.info("END [epoch:{:3d}][{:s}]{:s}".format(epoch, mode, end_str))
def run(dataloader, exe, program, fetchs, epoch=0, mode='train', vdl_writer=None):
"""
Feed data to the model and fetch the measures and loss
Args:
dataloader(fluid dataloader):
exe():
program():
fetchs(dict): dict of measures and the loss
epoch(int): epoch of training or validation
model(str): log only
Returns:
"""
fetch_list = [f[0] for f in fetchs.values()]
metric_list = [f[1] for f in fetchs.values()]
for m in metric_list:
m.reset()
batch_time = AverageMeter('elapse', '.3f')
tic = time.time()
for idx, batch in enumerate(dataloader()):
metrics = exe.run(program=program, feed=batch, fetch_list=fetch_list)
batch_time.update(time.time() - tic)
tic = time.time()
for i, m in enumerate(metrics):
metric_list[i].update(m[0], len(batch[0]))
fetchs_str = ''.join([str(m.value) + ' '
for m in metric_list] + [batch_time.value]) + 's'
if vdl_writer:
global total_step
logger.scaler('loss', metrics[0][0], total_step, vdl_writer)
total_step += 1
if mode == 'eval':
logger.info("{:s} step:{:<4d} {:s}s".format(mode, idx, fetchs_str))
else:
epoch_str = "epoch:{:<3d}".format(epoch)
step_str = "{:s} step:{:<4d}".format(mode, idx)
logger.info("{:s} {:s} {:s}".format(
logger.coloring(epoch_str, "HEADER")
if idx == 0 else epoch_str,
logger.coloring(step_str, "PURPLE"),
logger.coloring(fetchs_str, 'OKGREEN')))
end_str = ''.join([str(m.mean) + ' '
for m in metric_list] + [batch_time.total]) + 's'
if mode == 'eval':
logger.info("END {:s} {:s}s".format(mode, end_str))
else:
end_epoch_str = "END epoch:{:<3d}".format(epoch)
logger.info("{:s} {:s} {:s}".format(
logger.coloring(end_epoch_str, "RED"),
logger.coloring(mode, "PURPLE"),
logger.coloring(end_str, "OKGREEN")))
# return top1_acc in order to save the best model
if mode == 'valid':
return fetchs["top1"][1].avg
def run(dataloader, exe, program, fetchs, epoch=0, mode='train'):
"""
Feed data to the model and fetch the measures and loss
Args:
dataloader(fluid dataloader):
exe():
program():
fetchs(dict): dict of measures and the loss
epoch(int): epoch of training or validation
model(str): log only
Returns:
"""
fetch_list = [f[0] for f in fetchs.values()]
metric_list = [f[1] for f in fetchs.values()]
for m in metric_list:
m.reset()
batch_time = AverageMeter('elapse', '.3f')
tic = time.time()
for idx, batch in enumerate(dataloader()):
metrics = exe.run(program=program, feed=batch, fetch_list=fetch_list)
batch_time.update(time.time() - tic)
tic = time.time()
for i, m in enumerate(metrics):
metric_list[i].update(m[0], len(batch[0]))
fetchs_str = ''.join([str(m.value) + ' '
for m in metric_list] + [batch_time.value])
if mode == 'eval':
logger.info("{:s} step:{:<4d} {:s}s".format(mode, idx, fetchs_str))
else:
logger.info("epoch:{:<3d} {:s} step:{:<4d} {:s}s".format(
epoch, mode, idx, fetchs_str))
end_str = ''.join([str(m.mean) + ' '
for m in metric_list] + [batch_time.total])
if mode == 'eval':
logger.info("END {:s} {:s}s".format(mode, end_str))
else:
logger.info("END epoch:{:<3d} {:s} {:s}s".format(epoch, mode, end_str))
# return top1_acc in order to save the best model
if mode == 'valid':
return fetchs["top1"][1].avg
def run(dataloader,
exe,
program,
fetchs,
epoch=0,
mode='train',
config=None,
vdl_writer=None):
"""
Feed data to the model and fetch the measures and loss
Args:
dataloader(fluid dataloader):
exe():
program():
fetchs(dict): dict of measures and the loss
epoch(int): epoch of training or validation
model(str): log only
Returns:
"""
fetch_list = [f[0] for f in fetchs.values()]
metric_list = [f[1] for f in fetchs.values()]
for m in metric_list:
m.reset()
batch_time = AverageMeter('elapse', '.5f', need_avg=True)
tic = time.time()
dataloader = dataloader if config.get('use_dali') else dataloader()()
for idx, batch in enumerate(dataloader):
if idx == 10:
for m in metric_list:
m.reset()
batch_time.reset()
batch_size = batch[0]["feed_image"].shape()[0]
metrics = exe.run(program=program, feed=batch, fetch_list=fetch_list)
batch_time.update(time.time() - tic)
for i, m in enumerate(metrics):
metric_list[i].update(np.mean(m), batch_size)
fetchs_str = ''.join([str(m.value) + ' '
for m in metric_list] + [batch_time.mean]) + 's'
ips_info = " ips: {:.5f} images/sec.".format(batch_size /
batch_time.avg)
fetchs_str += ips_info
if vdl_writer:
global total_step
logger.scaler('loss', metrics[0][0], total_step, vdl_writer)
total_step += 1
if mode == 'eval':
if idx % config.get('print_interval', 10) == 0:
logger.info("{:s} step:{:<4d} {:s}".format(
mode, idx, fetchs_str))
else:
epoch_str = "epoch:{:<3d}".format(epoch)
step_str = "{:s} step:{:<4d}".format(mode, idx)
if idx % config.get('print_interval', 10) == 0:
logger.info("{:s} {:s} {:s}".format(
epoch_str if idx == 0 else epoch_str, step_str,
fetchs_str))
tic = time.time()
if config.get('use_dali'):
dataloader.reset()
end_str = ''.join([str(m.mean) + ' '
for m in metric_list] + [batch_time.total]) + 's'
ips_info = "ips: {:.5f} images/sec.".format(batch_size * batch_time.count /
batch_time.sum)
if mode == 'eval':
logger.info("END {:s} {:s}s {:s}".format(mode, end_str, ips_info))
else:
end_epoch_str = "END epoch:{:<3d}".format(epoch)
logger.info("{:s} {:s} {:s} {:s}".format(end_epoch_str, mode, end_str,
ips_info))
# return top1_acc in order to save the best model
if mode == 'valid':
return fetchs["top1"][1].avg
def run(dataloader,
exe,
program,
feeds,
fetchs,
epoch=0,
mode='train',
config=None,
vdl_writer=None,
lr_scheduler=None):
"""
Feed data to the model and fetch the measures and loss
Args:
dataloader(paddle io dataloader):
exe():
program():
fetchs(dict): dict of measures and the loss
epoch(int): epoch of training or validation
model(str): log only
Returns:
"""
fetch_list = [f[0] for f in fetchs.values()]
metric_list = [f[1] for f in fetchs.values()]
if mode == "train":
metric_list.append(AverageMeter('lr', 'f', need_avg=False))
for m in metric_list:
m.reset()
batch_time = AverageMeter('elapse', '.3f')
use_dali = config.get('use_dali', False)
dataloader = dataloader if use_dali else dataloader()
tic = time.time()
for idx, batch in enumerate(dataloader):
# ignore the warmup iters
if idx == 5:
batch_time.reset()
if use_dali:
batch_size = batch[0]["feed_image"].shape()[0]
feed_dict = batch[0]
else:
batch_size = batch[0].shape()[0]
feed_dict = {
key.name: batch[idx]
for idx, key in enumerate(feeds.values())
}
metrics = exe.run(program=program,
feed=feed_dict,
fetch_list=fetch_list)
batch_time.update(time.time() - tic)
for i, m in enumerate(metrics):
metric_list[i].update(np.mean(m), batch_size)
if mode == "train":
metric_list[-1].update(lr_scheduler.get_lr())
fetchs_str = ''.join([str(m.value) + ' '
for m in metric_list] + [batch_time.mean]) + 's'
ips_info = " ips: {:.5f} images/sec.".format(batch_size /
batch_time.avg)
fetchs_str += ips_info
if lr_scheduler is not None:
if lr_scheduler.update_specified:
curr_global_counter = lr_scheduler.step_each_epoch * epoch + idx
update = max(
0, curr_global_counter - lr_scheduler.update_start_step
) % lr_scheduler.update_step_interval == 0
if update:
lr_scheduler.step()
else:
lr_scheduler.step()
if vdl_writer:
global total_step
logger.scaler('loss', metrics[0][0], total_step, vdl_writer)
total_step += 1
if mode == 'valid':
if idx % config.get('print_interval', 10) == 0:
logger.info("{:s} step:{:<4d} {:s}".format(
mode, idx, fetchs_str))
else:
epoch_str = "epoch:{:<3d}".format(epoch)
step_str = "{:s} step:{:<4d}".format(mode, idx)
if idx % config.get('print_interval', 10) == 0:
logger.info("{:s} {:s} {:s}".format(
logger.coloring(epoch_str, "HEADER")
if idx == 0 else epoch_str,
logger.coloring(step_str, "PURPLE"),
logger.coloring(fetchs_str, 'OKGREEN')))
tic = time.time()
end_str = ''.join([str(m.mean) + ' '
for m in metric_list] + [batch_time.total]) + 's'
ips_info = "ips: {:.5f} images/sec.".format(batch_size * batch_time.count /
batch_time.sum)
if mode == 'valid':
logger.info("END {:s} {:s}s {:s}".format(mode, end_str, ips_info))
else:
end_epoch_str = "END epoch:{:<3d}".format(epoch)
logger.info("{:s} {:s} {:s} {:s}".format(end_epoch_str, mode, end_str,
ips_info))
if use_dali:
dataloader.reset()
# return top1_acc in order to save the best model
if mode == 'valid':
return fetchs["top1"][1].avg