def loop_unet(dl, model, losses_stats, device):
samples_plus_losses = []
loss_func = torch.nn.CrossEntropyLoss()
with torch.no_grad():
for (x, y), sample in pbar(dl):
torch.manual_seed(0)
x, y = _move_to_device(x, y, device)
preds = model(x)
loss = loss_func(preds, y)
loss = {
"loss_unet": float(loss.cpu().numpy()),
"loss_total": float(loss.cpu().numpy()),
}
for l in losses_stats.keys():
losses_stats[l].append(loss[l])
loss_comp = LossesRecordComponent()
loss_comp.set_losses(loss)
sample[0].add_component(loss_comp)
sample[0].set_img(tensor_to_image(x[0]))
sample[0].segmentation.set_mask_array(
MaskArray(y[0].detach().cpu().numpy()))
samples_plus_losses.append(sample[0])
return samples_plus_losses, losses_stats
def loop_mmdet(dl, model, losses_stats, device):
samples_plus_losses = []
with torch.no_grad():
for data, sample in pbar(dl):
torch.manual_seed(0)
_, data = _move_to_device(None, data, device)
loss = model(**data)
loss = sum_losses_mmdet(loss)
for l in losses_stats.keys():
losses_stats[l].append(loss[l])
sample[0].update(loss)
samples_plus_losses.append(sample[0])
return samples_plus_losses, losses_stats
def loop_mmdet(dl, model, losses_stats, device):
samples_plus_losses = []
with torch.no_grad():
for data, sample in pbar(dl):
torch.manual_seed(0)
_, data = _move_to_device(None, data, device)
loss = model(**data)
loss = sum_losses_mmdet(loss)
for l in losses_stats.keys():
losses_stats[l].append(loss[l])
loss_comp = LossesRecordComponent()
loss_comp.set_losses(loss)
sample[0].add_component(loss_comp)
sample[0].set_img(mmdet_tensor_to_image(data["img"][0]))
samples_plus_losses.append(sample[0])
return samples_plus_losses, losses_stats
def loop_yolo(dl, model, losses_stats, device):
samples_plus_losses = []
compute_loss = ComputeLoss(model)
with torch.no_grad():
for (x, y), sample in pbar(dl):
torch.manual_seed(0)
x, y = _move_to_device(x, y, device)
preds = model(x)
loss = compute_loss(preds, y)[0]
loss = {
"loss_yolo": float(loss.cpu().numpy()),
"loss_total": float(loss.cpu().numpy()),
}
for l in losses_stats.keys():
losses_stats[l].append(loss[l])
loss_comp = LossesRecordComponent()
loss_comp.set_losses(loss)
sample[0].add_component(loss_comp)
sample[0].set_img(tensor_to_image(x[0]))
samples_plus_losses.append(sample[0])
return samples_plus_losses, losses_stats