def val_step(model, val_loader, device, num_batches=None,
log_interval: int = 100):
"""
Performs one step of validation. Calculates loss, forward pass and returns metrics.
Args:
model : PyTorch FasterRCNN Model.
val_loader : Validation loader.
device : "cuda" or "cpu"
num_batches : (optional) Integer To limit validation to certain number of batches.
log_interval : (optional) Defualt 100. Integer to Log after specified batch ids in every batch.
"""
model = model.to(device)
start_val_step = time.time()
last_idx = len(val_loader) - 1
batch_time_m = utils.AverageMeter()
cnt = 0
model.eval()
batch_start = time.time()
metrics = OrderedDict()
with torch.no_grad():
for batch_idx, (inputs, targets) in enumerate(val_loader):
last_batch = batch_idx == last_idx
images = list(image.to(device) for image in inputs)
targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
out = model(images)
iou = torch.stack([_evaluate_iou(t, o) for t, o in zip(targets, out)]).mean()
giou = torch.stack([_evaluate_giou(t, o) for t, o in zip(targets, out)]).mean()
cnt += 1
batch_time_m.update(time.time() - batch_start)
batch_start = time.time()
if last_batch or batch_idx % log_interval == 0: # If we reach the log intervel
print("Batch Validation Time: {batch_time.val:.3f} ({batch_time.avg:.3f}) ".format(
batch_time=batch_time_m,))
if num_batches is not None:
if cnt >= num_batches:
avg_iou = torch.stack([iou]).mean()
avg_giou = torch.stack([giou]).mean()
metrics["iou"] = avg_iou
metrics["giou"] = avg_giou
print(f"Done till {num_batches} Validation batches")
end_val_step = time.time()
print(f"Time taken for validation step = {end_val_step - start_val_step} sec")
return metrics
avg_iou = torch.stack([iou]).mean()
avg_giou = torch.stack([giou]).mean()
metrics["iou"] = avg_iou
metrics["giou"] = avg_giou
end_val_step = time.time()
print(f"Time taken for validation step = {end_val_step - start_val_step} sec")
return metrics
def train_step(
model: nn.Module,
train_loader,
device: str,
optimizer,
scheduler=None,
num_batches: int = None,
log_interval: int = 100,
scaler=None,
):
"""
Performs one step of training. Calculates loss, forward pass, computes gradient and returns metrics.
Args:
model : PyTorch Faster RCNN Model.
train_loader : Train loader.
device : "cuda" or "cpu"
optimizer : Torch optimizer to train.
scheduler : Learning rate scheduler.
num_batches : (optional) Integer To limit training to certain number of batches.
log_interval : (optional) Defualt 100. Integer to Log after specified batch ids in every batch.
grad_penalty : (optional) To penalize with l2 norm for big gradients.
scaler: (optional) Pass torch.cuda.amp.GradScaler() for fp16 precision Training.
"""
model = model.to(device)
start_train_step = time.time()
model.train()
last_idx = len(train_loader) - 1
batch_time_m = utils.AverageMeter()
cnt = 0
batch_start = time.time()
metrics = OrderedDict()
total_loss = utils.AverageMeter()
loss_classifier = utils.AverageMeter()
loss_box_reg = utils.AverageMeter()
loss_objectness = utils.AverageMeter()
loss_rpn_box_reg = utils.AverageMeter()
for batch_idx, (inputs, targets) in enumerate(train_loader):
last_batch = batch_idx == last_idx
images = list(image.to(device) for image in inputs)
targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
# zero the parameter gradients
optimizer.zero_grad()
if scaler is not None:
with amp.autocast():
loss_dict = model(images, targets)
loss = sum(loss_v for loss_v in loss_dict.values())
scaler.scale(loss).backward()
# Step using scaler.step()
scaler.step(optimizer)
# Update for next iteration
scaler.update()
else:
loss_dict = model(images, targets)
loss = sum(loss_v for loss_v in loss_dict.values())
loss.backward()
optimizer.step()
if scheduler is not None:
scheduler.step()
cnt += 1
total_loss.update(loss.item())
loss_classifier.update(loss_dict["loss_classifier"].item())
loss_box_reg.update(loss_dict["loss_box_reg"].item())
loss_objectness.update(loss_dict["loss_objectness"].item())
loss_rpn_box_reg.update(loss_dict["loss_rpn_box_reg"].item())
batch_time_m.update(time.time() - batch_start)
batch_start = time.time()
if last_batch or batch_idx % log_interval == 0: # If we reach the log intervel
print(
"Batch Train Time: {batch_time.val:.3f} ({batch_time.avg:.3f}) "
.format(batch_time=batch_time_m, ))
# "loss classifier: {loss_d.loss_classifier:>7.4f} "
# "loss box_reg: {loss_d.loss_box_reg:>7.4f} "
# "loss objectness: {loss_d.loss_objectness:>7.4f} "
# "loss rpn box reg: {loss_d.loss_rpn_box_reg:>7.4f}"
if num_batches is not None:
if cnt >= num_batches:
end_train_step = time.time()
metrics["total_loss"] = total_loss.avg
metrics["loss_classifier"] = loss_classifier.avg
metrics["loss_box_reg"] = loss_box_reg.avg
metrics["loss_objectness"] = loss_objectness.avg
metrics["loss_rpn_box_reg"] = loss_rpn_box_reg.avg
print(f"Done till {num_batches} train batches")
print(
f"Time taken for Training step = {end_train_step - start_train_step} sec"
)
return metrics
end_train_step = time.time()
metrics["total_loss"] = total_loss.avg
metrics["loss_classifier"] = loss_classifier.avg
metrics["loss_box_reg"] = loss_box_reg.avg
metrics["loss_objectness"] = loss_objectness.avg
metrics["loss_rpn_box_reg"] = loss_rpn_box_reg.avg
print(
f"Time taken for Training step = {end_train_step - start_train_step} sec"
)
return metrics
def train_step(
model,
train_loader,
criterion,
device,
optimizer,
scheduler=None,
num_batches: int = None,
log_interval: int = 100,
grad_penalty: bool = False,
scaler=None,
):
"""
Performs one step of training. Calculates loss, forward pass, computes gradient and returns metrics.
Args:
model : A pytorch CNN Model.
train_loader : Train loader.
criterion : Loss function to be optimized.
device : "cuda" or "cpu"
optimizer : Torch optimizer to train.
scheduler : Learning rate scheduler.
num_batches : (optional) Integer To limit training to certain number of batches.
log_interval : (optional) Defualt 100. Integer to Log after specified batch ids in every batch.
grad_penalty : (optional) To penalize with l2 norm for big gradients.
scaler: (optional) Pass torch.cuda.amp.GradScaler() for fp16 precision Training.
"""
model = model.to(device)
start_train_step = time.time()
metrics = OrderedDict()
model.train()
last_idx = len(train_loader) - 1
batch_time_m = utils.AverageMeter()
# data_time_m = utils.AverageMeter()
losses_m = utils.AverageMeter()
top1_m = utils.AverageMeter()
top5_m = utils.AverageMeter()
cnt = 0
batch_start = time.time()
# num_updates = epoch * len(loader)
for batch_idx, (inputs, target) in enumerate(train_loader):
last_batch = batch_idx == last_idx
# data_time_m.update(time.time() - batch_start)
inputs = inputs.to(device)
target = target.to(device)
# zero the parameter gradients
optimizer.zero_grad()
if scaler is not None:
with amp.autocast():
output = model(inputs)
loss = criterion(output, target)
# Scale the loss using Grad Scaler
if grad_penalty is True:
# Scales the loss for autograd.grad's backward pass, resulting in scaled grad_params
scaled_grad_params = torch.autograd.grad(scaler.scale(loss),
model.parameters(),
create_graph=True)
# Creates unscaled grad_params before computing the penalty. scaled_grad_params are
# not owned by any optimizer, so ordinary division is used instead of scaler.unscale_:
inv_scale = 1.0 / scaler.get_scale()
grad_params = [p * inv_scale for p in scaled_grad_params]
# Computes the penalty term and adds it to the loss
with amp.autocast():
grad_norm = 0
for grad in grad_params:
grad_norm += grad.pow(2).sum()
grad_norm = grad_norm.sqrt()
loss = loss + grad_norm
scaler.scale(loss).backward()
# Step using scaler.step()
scaler.step(optimizer)
# Update for next iteration
scaler.update()
else:
output = model(inputs)
loss = criterion(output, target)
if grad_penalty is True:
# Create gradients
grad_params = torch.autograd.grad(loss,
model.parameters(),
create_graph=True)
# Compute the L2 Norm as penalty and add that to loss
grad_norm = 0
for grad in grad_params:
grad_norm += grad.pow(2).sum()
grad_norm = grad_norm.sqrt()
loss = loss + grad_norm
loss.backward()
optimizer.step()
if scheduler is not None:
scheduler.step()
cnt += 1
acc1, acc5 = accuracy(output, target, topk=(1, 5))
top1_m.update(acc1.item(), output.size(0))
top5_m.update(acc5.item(), output.size(0))
losses_m.update(loss.item(), inputs.size(0))
batch_time_m.update(time.time() - batch_start)
batch_start = time.time()
if last_batch or batch_idx % log_interval == 0: # If we reach the log intervel
print(
"Batch Train Time: {batch_time.val:.3f} ({batch_time.avg:.3f}) "
"Loss: {loss.val:>7.4f} ({loss.avg:>6.4f}) "
"Top 1 Accuracy: {top1.val:>7.4f} ({top1.avg:>7.4f}) "
"Top 5 Accuracy: {top5.val:>7.4f} ({top5.avg:>7.4f})".format(
batch_time=batch_time_m,
loss=losses_m,
top1=top1_m,
top5=top5_m))
if num_batches is not None:
if cnt >= num_batches:
end_train_step = time.time()
metrics["loss"] = losses_m.avg
metrics["top1"] = top1_m.avg
metrics["top5"] = top5_m.avg
print(f"Done till {num_batches} train batches")
print(
f"Time taken for train step = {end_train_step - start_train_step} sec"
)
return metrics
metrics["loss"] = losses_m.avg
metrics["top1"] = top1_m.avg
metrics["top5"] = top5_m.avg
end_train_step = time.time()
print(
f"Time taken for train step = {end_train_step - start_train_step} sec")
return metrics
def val_step(model,
val_loader,
criterion,
device,
num_batches=None,
log_interval: int = 100):
"""
Performs one step of validation. Calculates loss, forward pass and returns metrics.
Args:
model : A pytorch CNN Model.
val_loader : Validation loader.
criterion : Loss function to be optimized.
device : "cuda" or "cpu"
num_batches : (optional) Integer To limit validation to certain number of batches.
log_interval : (optional) Defualt 100. Integer to Log after specified batch ids in every batch.
"""
model = model.to(device)
start_val_step = time.time()
last_idx = len(val_loader) - 1
batch_time_m = utils.AverageMeter()
# data_time_m = utils.AverageMeter()
losses_m = utils.AverageMeter()
top1_m = utils.AverageMeter()
top5_m = utils.AverageMeter()
cnt = 0
model.eval()
batch_start = time.time()
metrics = OrderedDict()
with torch.no_grad():
for batch_idx, (inputs, target) in enumerate(val_loader):
last_batch = batch_idx == last_idx
inputs = inputs.to(device)
target = target.to(device)
output = model(inputs)
if isinstance(output, (tuple, list)):
output = output[0]
loss = criterion(output, target)
cnt += 1
acc1, acc5 = accuracy(output, target, topk=(1, 5))
reduced_loss = loss.data
losses_m.update(reduced_loss.item(), inputs.size(0))
top1_m.update(acc1.item(), output.size(0))
top5_m.update(acc5.item(), output.size(0))
batch_time_m.update(time.time() - batch_start)
batch_start = time.time()
if (last_batch or batch_idx % log_interval
== 0): # If we reach the log intervel
print(
"Batch Inference Time: {batch_time.val:.3f} ({batch_time.avg:.3f}) "
"Loss: {loss.val:>7.4f} ({loss.avg:>6.4f}) "
"Top 1 Accuracy: {top1.val:>7.4f} ({top1.avg:>7.4f}) "
"Top 5 Accuracy: {top5.val:>7.4f} ({top5.avg:>7.4f})".
format(batch_time=batch_time_m,
loss=losses_m,
top1=top1_m,
top5=top5_m))
if num_batches is not None:
if cnt >= num_batches:
end_val_step = time.time()
metrics["loss"] = losses_m.avg
metrics["top1"] = top1_m.avg
metrics["top5"] = top5_m.avg
print(f"Done till {num_batches} validation batches")
print(
f"Time taken for validation step = {end_val_step - start_val_step} sec"
)
return metrics
metrics["loss"] = losses_m.avg
metrics["top1"] = top1_m.avg
metrics["top5"] = top5_m.avg
print("Finished the validation epoch")
end_val_step = time.time()
print(
f"Time taken for validation step = {end_val_step - start_val_step} sec"
)
return metrics
def train_step(
model: nn.Module,
train_loader,
criterion,
device: str,
optimizer,
scheduler=None,
num_batches: int = None,
log_interval: int = 100,
scaler=None,
):
"""
Performs one step of training. Calculates loss, forward pass, computes gradient and returns metrics.
Args:
model : PyTorch Detr Model.
train_loader : Train loader.
device : "cuda" or "cpu"
criterion : Detr Loss function to be optimized.
optimizer : Torch optimizer to train.
scheduler : Learning rate scheduler.
num_batches : (optional) Integer To limit training to certain number of batches.
log_interval : (optional) Defualt 100. Integer to Log after specified batch ids in every batch.
scaler: (optional) Pass torch.cuda.amp.GradScaler() for fp16 precision Training.
"""
model = model.to(device)
criterion = criterion.to(device)
start_train_step = time.time()
model.train()
last_idx = len(train_loader) - 1
batch_time_m = utils.AverageMeter()
criterion.train()
cnt = 0
batch_start = time.time()
metrics = OrderedDict()
total_loss = utils.AverageMeter()
bbox_loss = utils.AverageMeter()
giou_loss = utils.AverageMeter()
labels_loss = utils.AverageMeter()
for batch_idx, (inputs, targets) in enumerate(train_loader):
last_batch = batch_idx == last_idx
images = list(image.to(device) for image in inputs)
targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
optimizer.zero_grad()
if scaler is not None:
with amp.autocast():
outputs = model(images)
loss_dict = criterion(outputs, targets)
weight_dict = criterion.weight_dict
loss = sum(loss_dict[k] * weight_dict[k]
for k in loss_dict.keys() if k in weight_dict)
scaler.scale(loss).backward()
# Step using scaler.step()
scaler.step(optimizer)
# Update for next iteration
scaler.update()
else:
outputs = model(images)
loss_dict = criterion(outputs, targets)
weight_dict = criterion.weight_dict
loss = sum(loss_dict[k] * weight_dict[k] for k in loss_dict.keys()
if k in weight_dict)
loss.backward()
optimizer.step()
if scheduler is not None:
scheduler.step()
cnt += 1
total_loss.update(loss.item())
bbox_loss.update(loss_dict["loss_bbox"].item())
giou_loss.update(loss_dict["loss_giou"].item())
labels_loss.update(loss_dict["loss_ce"].item())
batch_time_m.update(time.time() - batch_start)
batch_start = time.time()
if last_batch or batch_idx % log_interval == 0: # If we reach the log intervel
print(
"Batch Train Time: {batch_time.val:.3f} ({batch_time.avg:.3f}) "
.format(batch_time=batch_time_m, ))
if num_batches is not None:
if cnt >= num_batches:
end_train_step = time.time()
metrics["total_loss"] = total_loss.avg
metrics["bbox_loss"] = bbox_loss.avg
metrics["giou_loss"] = giou_loss.avg
metrics["labels_loss"] = labels_loss.avg
print(f"Done till {num_batches} train batches")
print(
f"Time taken for Training step = {end_train_step - start_train_step} sec"
)
return metrics
end_train_step = time.time()
metrics["total_loss"] = total_loss.avg
metrics["bbox_loss"] = bbox_loss.avg
metrics["giou_loss"] = giou_loss.avg
metrics["labels_loss"] = labels_loss.avg
print(
f"Time taken for Training step = {end_train_step - start_train_step} sec"
)
return metrics
def val_step(model: nn.Module,
val_loader,
criterion,
device,
num_batches: int = None,
log_interval: int = 100):
"""
Performs one step of validation. Calculates loss, forward pass and returns metrics.
Args:
model : PyTorch Detr Model.
val_loader : Validation loader.
criterion : Detr Loss function to be optimized.
device : "cuda" or "cpu"
num_batches : (optional) Integer To limit validation to certain number of batches.
log_interval : (optional) Defualt 100. Integer to Log after specified batch ids in every batch.
"""
model = model.to(device)
start_val_step = time.time()
last_idx = len(val_loader) - 1
batch_time_m = utils.AverageMeter()
cnt = 0
model.eval()
criterion.eval()
batch_start = time.time()
metrics = OrderedDict()
total_loss = utils.AverageMeter()
bbox_loss = utils.AverageMeter()
giou_loss = utils.AverageMeter()
labels_loss = utils.AverageMeter()
with torch.no_grad():
for batch_idx, (inputs, targets) in enumerate(val_loader):
last_batch = batch_idx == last_idx
images = list(image.to(device) for image in inputs)
targets = [{k: v.to(device)
for k, v in t.items()} for t in targets]
outputs = model(images)
loss_dict = criterion(outputs, targets)
weight_dict = criterion.weight_dict
loss = sum(loss_dict[k] * weight_dict[k] for k in loss_dict.keys()
if k in weight_dict)
cnt += 1
total_loss.update(loss.item())
bbox_loss.update(loss_dict["loss_bbox"].item())
giou_loss.update(loss_dict["loss_giou"].item())
labels_loss.update(loss_dict["loss_ce"].item())
batch_time_m.update(time.time() - batch_start)
batch_start = time.time()
if last_batch or batch_idx % log_interval == 0: # If we reach the log intervel
print(
"Batch Validation Time: {batch_time.val:.3f} ({batch_time.avg:.3f}) "
.format(batch_time=batch_time_m, ))
if num_batches is not None:
if cnt >= num_batches:
end_val_step = time.time()
metrics["total_loss"] = total_loss.avg
metrics["bbox_loss"] = bbox_loss.avg
metrics["giou_loss"] = giou_loss.avg
metrics["labels_loss"] = labels_loss.avg
print(f"Done till {num_batches} Validation batches")
print(
f"Time taken for validation step = {end_val_step - start_val_step} sec"
)
return metrics
end_val_step = time.time()
metrics["total_loss"] = total_loss.avg
metrics["bbox_loss"] = bbox_loss.avg
metrics["giou_loss"] = giou_loss.avg
metrics["labels_loss"] = labels_loss.avg
print(
f"Time taken for validation step = {end_val_step - start_val_step} sec"
)
return metrics