def track(sequence: Iterable,
description: str = "Working...",
disable: bool = False,
style: Literal["rich", "tqdm"] = None,
**kwargs):
"""Progress bar with `'rich'` and `'tqdm'` styles."""
if style is None:
style = settings.progress_bar_style
if style not in ["rich", "tqdm"]:
raise ValueError("style must be one of ['rich', 'tqdm']")
if disable:
return sequence
if style == "tqdm":
# fixes repeated pbar in jupyter
# see https://github.com/tqdm/tqdm/issues/375
if hasattr(tqdm_base, "_instances"):
for instance in list(tqdm_base._instances):
tqdm_base._decr_instances(instance)
return tqdm_base(sequence, desc=description, file=sys.stdout, **kwargs)
else:
in_colab = "google.colab" in sys.modules
force_jupyter = None if not in_colab else True
console = Console(force_jupyter=force_jupyter)
return track_base(sequence,
description=description,
console=console,
**kwargs)
def tqdm(*args, **kwargs):
# get_ipython().events.register('post_execute', tqdm_clear)
if hasattr(tqdm_base, '_instances'):
for instance in list(tqdm_base._instances):
tqdm_base._decr_instances(instance)
return tqdm_base(*args, **kwargs)
def tqdm(*args, **kwargs):
"""Decorator of tqdm, to avoid some errors if tqdm
terminated unexpectedly
Returns
-------
an decorated `tqdm` class
"""
if hasattr(tqdm_base, '_instances'):
for instance in list(tqdm_base._instances):
tqdm_base._decr_instances(instance)
return tqdm_base(*args, **kwargs)
def track(sequence: Iterable,
description: str = "Working...",
disable: bool = False,
style: Literal["rich", "tqdm"] = None,
**kwargs):
"""
Progress bar with `'rich'` and `'tqdm'` styles.
Parameters
----------
sequence
Iterable sequence.
description
First text shown to left of progress bar.
disable
Switch to turn off progress bar.
style
One of ["rich", "tqdm"]. "rich" is interactive
and is not persistent after close.
**kwargs
Keyword args to tqdm or rich.
Examples
--------
>>> from scvi.utils import track
>>> my_list = [1, 2, 3]
>>> for i in track(my_list): print(i)
"""
if style is None:
style = settings.progress_bar_style
if style not in ["rich", "tqdm"]:
raise ValueError("style must be one of ['rich', 'tqdm']")
if disable:
return sequence
if style == "tqdm":
# fixes repeated pbar in jupyter
# see https://github.com/tqdm/tqdm/issues/375
if hasattr(tqdm_base, "_instances"):
for instance in list(tqdm_base._instances):
tqdm_base._decr_instances(instance)
return tqdm_base(sequence, desc=description, file=sys.stdout, **kwargs)
else:
in_colab = "google.colab" in sys.module
force_jupyter = None if not in_colab else True
console = Console(force_jupyter=force_jupyter)
return track_base(sequence,
description=description,
console=console,
**kwargs)
def track(sequence: Iterable,
description: str = "Working...",
disable: bool = False,
style: Literal["rich", "tqdm"] = None,
**kwargs):
"""Progress bar with `'rich'` and `'tqdm'` styles."""
if style is None:
style = settings.progress_bar_style
if style not in ["rich", "tqdm"]:
raise ValueError("style must be one of ['rich', 'tqdm']")
if disable:
return sequence
if style == "tqdm":
# fixes repeated pbar in jupyter
# see https://github.com/tqdm/tqdm/issues/375
if hasattr(tqdm_base, "_instances"):
for instance in list(tqdm_base._instances):
tqdm_base._decr_instances(instance)
return tqdm_base(sequence, desc=description, file=sys.stdout, **kwargs)
else:
with warnings.catch_warnings():
warnings.simplefilter("ignore")
return tqdm_rich(sequence, desc=description, **kwargs)
def reset_tqdm():
for instance in list(tqdm._instances):
tqdm._decr_instances(instance)
def tqdm(*args, **kwargs):
if hasattr(tqdm_base, '_instances'):
for instance in list(tqdm_base._instances):
tqdm_base._decr_instances(instance)
return tqdm_base(*args, **kwargs)
def main():
loss_history = {"epoch": [], "train": [], "val": []}
train_metrices_history = {
"epoch": [],
"acc": [],
"iou": [],
"dice": [],
"sens": [],
"spec": []
}
val_metrices_history = {
"epoch": [],
"acc": [],
"iou": [],
"dice": [],
"sens": [],
"spec": []
}
prev_loss = -100
loss_increase_counter = 0
early_stop = True
early_stop_threshold = 5
for e in range(epochs):
model.train()
running_loss = 0.0
train_accuracy = 0.0
train_sensitivity = 0.0
train_specificity = 0.0
train_iou = 0.0 # Jaccard Score
train_dice = 0.0
ts = time.time()
for i, d in tqdm(enumerate(trainloader),
total=len(trainloader),
leave=True,
position=0,
desc='Epoch: {}'.format(e)):
inputs_, targets_, _ = d
inputs, targets = inputs_.to(device), targets_.to(device)
opt.zero_grad()
outputs = model(inputs)
loss = loss_f(outputs, targets.long())
loss.backward()
opt.step()
running_loss += loss.item()
train_iou += IoU()(outputs, targets)
train_dice += Fscore()(outputs, targets)
train_accuracy += Accuracy()(outputs, targets)
train_sensitivity += Sensitivity()(outputs, targets)
train_specificity += Specificity()(outputs, targets)
state = {
'epoch': e,
'state_dict': model.state_dict(),
'optimizer': opt.state_dict(),
'loss': loss.item()
}
tr_loss = running_loss / len(trainloader)
vl_loss, val_acc, val_iou, val_dice, val_sens, val_spec = val()
loss_history["epoch"].append(e)
loss_history["train"].append(tr_loss)
loss_history["val"].append(vl_loss)
tr_acc = train_accuracy / len(trainloader)
tr_iou = train_iou / len(trainloader)
tr_dice = train_dice / len(trainloader)
tr_sens = train_sensitivity / len(trainloader)
tr_spec = train_specificity / len(trainloader)
val_metrices_history["epoch"].append(e)
val_metrices_history["acc"].append(val_acc)
val_metrices_history["iou"].append(val_iou)
val_metrices_history["dice"].append(val_dice)
val_metrices_history["sens"].append(val_sens)
val_metrices_history["spec"].append(val_spec)
train_metrices_history["epoch"].append(e)
train_metrices_history["acc"].append(tr_acc)
train_metrices_history["iou"].append(tr_iou)
train_metrices_history["dice"].append(tr_dice)
train_metrices_history["sens"].append(tr_sens)
train_metrices_history["spec"].append(tr_spec)
file_name = os.path.join(ckp_path +
'task_2_weights_epoch_{}'.format(e) + '.pt')
print(
"Finish Epoch {0},Time Elapsed {1}, train loss: {2:.6g}, val loss: {3:.6g}"
.format(e,
time.time() - ts, tr_loss, vl_loss))
print(
"Metrices Train : Acc {0:.6g},IOU {1:.6g}, Dice {2:.6g}, Sens {3:.6g}, Spec {4:.6g}"
.format(tr_acc, tr_iou, tr_dice, tr_sens, tr_spec))
print(
"Metrices Valid : Acc {0:.6g},IOU {1:.6g}, Dice {2:.6g}, Sens {3:.6g}, Spec {4:.6g}"
.format(val_acc, val_iou, val_dice, val_sens, val_spec))
print("-" * 60)
torch.save(state, file_name)
# Implemented early stopping
if vl_loss > prev_loss:
loss_increase_counter += 1
else:
loss_increase_counter = 0
if early_stop and loss_increase_counter > early_stop_threshold:
print("Early Stopping..")
break
prev_loss = vl_loss
torch.cuda.empty_cache()
list(getattr(tqdm, '_instances'))
for instance in list(tqdm._instances):
tqdm._decr_instances(instance)
print('Finished Training')
# Write the train loss to the csv
(pd.DataFrame.from_dict(data=loss_history, orient='columns').to_csv(
os.path.join(logs_path, 'loss.csv'),
header=['epoch', 'train_loss', 'val_loss']))
(pd.DataFrame.from_dict(
data=train_metrices_history, orient='columns').to_csv(
os.path.join(logs_path, 'train_metrices.csv'),
header=["epoch", "acc", "iou", "dice", "sens", "spec"]))
(pd.DataFrame.from_dict(
data=val_metrices_history, orient='columns').to_csv(
os.path.join(logs_path, 'val_metrices.csv'),
header=["epoch", "acc", "iou", "dice", "sens", "spec"]))
def tqdm_fixed(*args, **kwargs):
from tqdm import tqdm as tqdm_base
if hasattr(tqdm_base, '_instances'):
for instance in list(tqdm_base._instances):
tqdm_base._decr_instances(instance)
return tqdm_base(*args, **kwargs)
def clear_tqdm():
if hasattr(tqdm, '_instances'):
for instance in list(tqdm._instances):
tqdm._decr_instances(instance)
loss = criterion(output, target_tokens)
epoch_loss += loss.item()
return epoch_loss / len(iterator)
N_EPOCHS = args.epochs
CLIP = 1
best_valid_loss = float('inf')
for epoch in range(N_EPOCHS):
train_loss = train(model, train_iterator, criterion, optimizer, CLIP)
valid_loss = evaluate(model, valid_iterator, criterion)
if valid_loss < best_valid_loss:
best_valid_loss = valid_loss
torch.save(model.state_dict(), args.save+'/model.pth')
print('Epoch: ', epoch)
print('Train loss: ', train_loss)
print('Valid loss: ', valid_loss)
test_loss = evaluate(model, test_iterator, criterion)
print('Test Loss: {:.2f}'.format(test_loss))
for instance in list(tqdm._instances):
tqdm._decr_instances(instance)
