def train(loader, num_classes, device, net, optimizer, criterion):
num_samples = 0
running_loss = 0
metrics = Metrics()
net.train()
for images, masks, _tile in tqdm(loader,
desc="Train",
unit="batch",
ascii=True):
images = images.to(device)
masks = masks.to(device)
assert images.size()[2:] == masks.size(
)[1:], "resolutions for images and masks are in sync"
num_samples += int(images.size(0))
optimizer.zero_grad()
outputs = net(images)
assert outputs.size()[2:] == masks.size(
)[1:], "resolutions for predictions and masks are in sync"
assert outputs.size(
)[1] == num_classes, "classes for predictions and dataset are in sync"
loss = criterion(outputs, masks)
loss.backward()
optimizer.step()
running_loss += loss.item()
for mask, output in zip(masks, outputs):
prediction = output.detach()
metrics.add(mask, prediction)
assert num_samples > 0, "dataset contains training images and labels"
class_stats = metrics.get_classification_stats()
return {
"loss": running_loss / num_samples,
"miou": metrics.get_miou(),
"fg_iou": metrics.get_fg_iou(),
"mcc": metrics.get_mcc(),
"accuracy": class_stats['accuracy'],
"precision": class_stats['precision'],
"recall": class_stats['recall'],
"f1": class_stats['f1']
}
def validate(loader, num_classes, device, net, criterion):
num_samples = 0
running_loss = 0
metrics = Metrics()
net.eval()
with torch.no_grad():
for images, masks, _tile in tqdm(loader,
desc="Validate",
unit="batch",
ascii=True):
images = images.to(device)
masks = masks.to(device)
assert images.size()[2:] == masks.size(
)[1:], "resolutions for images and masks are in sync"
num_samples += int(images.size(0))
outputs = net(images)
assert outputs.size()[2:] == masks.size(
)[1:], "resolutions for predictions and masks are in sync"
assert outputs.size(
)[1] == num_classes, "classes for predictions and dataset are in sync"
loss = criterion(outputs, masks)
running_loss += loss.item()
for mask, output in zip(masks, outputs):
metrics.add(mask, output)
assert num_samples > 0, "dataset contains validation images and labels"
class_stats = metrics.get_classification_stats()
return {
"loss": running_loss / num_samples,
"miou": metrics.get_miou(),
"fg_iou": metrics.get_fg_iou(),
"mcc": metrics.get_mcc(),
"accuracy": class_stats['accuracy'],
"precision": class_stats['precision'],
"recall": class_stats['recall'],
"f1": class_stats['f1']
}
def train(loader, config, log, device, nn, optimizer, criterion):
num_samples = 0
running_loss = 0
metrics = Metrics()
nn.train()
for images, masks, tiles in tqdm(loader,
desc="Train",
unit="batch",
ascii=True):
images = images.to(device)
masks = masks.to(device)
assert images.size()[2:] == masks.size(
)[1:], "resolutions for images and masks are in sync"
num_samples += int(images.size(0))
optimizer.zero_grad()
outputs = nn(images)
assert outputs.size()[2:] == masks.size(
)[1:], "resolutions for predictions and masks are in sync"
assert outputs.size()[1] == len(
config["classes"]
), "classes for predictions and dataset are in sync"
loss = criterion(outputs, masks, config)
loss.backward()
optimizer.step()
running_loss += loss.item()
for mask, output in zip(masks, outputs):
prediction = output.detach()
metrics.add(mask, prediction)
assert num_samples > 0, "dataset contains training images and labels"
log.log("{}{:.3f}".format("Loss:".ljust(25, " "),
running_loss / num_samples))
for classe in config["classes"][1:]:
log.log("{}{:.3f}".format((classe["title"] + " IoU:").ljust(25, " "),
metrics.get_fg_iou()))
def compare(masks, labels, tile, classes):
x, y, z = list(map(str, tile))
label = np.array(Image.open(os.path.join(labels, z, x, "{}.png".format(y))))
mask = np.array(Image.open(os.path.join(masks, z, x, "{}.png".format(y))))
assert label.shape == mask.shape
assert len(label.shape) == 2 and len(classes) == 2 # Still binary centric
metrics = Metrics()
metrics.add(torch.from_numpy(label), torch.from_numpy(mask), is_prob=False)
fg_iou = metrics.get_fg_iou()
fg_ratio = 100 * max(np.sum(mask != 0), np.sum(label != 0)) / mask.size
dist = 0.0 if math.isnan(fg_iou) else 1.0 - fg_iou
qod = 100 - (dist * (math.log(fg_ratio + 1.0) + np.finfo(float).eps) * (100 / math.log(100)))
qod = 0.0 if qod < 0.0 else qod # Corner case prophilaxy
return dist, fg_ratio, qod
def validate(loader, config, log, device, nn, criterion):
num_samples = 0
running_loss = 0
metrics = Metrics()
nn.eval()
with torch.no_grad():
for images, masks, tiles in tqdm(loader,
desc="Validate",
unit="batch",
ascii=True):
images = images.to(device)
masks = masks.to(device)
assert images.size()[2:] == masks.size(
)[1:], "resolutions for images and masks are in sync"
num_samples += int(images.size(0))
outputs = nn(images)
assert outputs.size()[2:] == masks.size(
)[1:], "resolutions for predictions and masks are in sync"
assert outputs.size()[1] == len(
config["classes"]
), "classes for predictions and dataset are in sync"
loss = criterion(outputs, masks, config)
running_loss += loss.item()
for mask, output in zip(masks, outputs):
metrics.add(mask, output)
assert num_samples > 0, "dataset contains validation images and labels"
log.log("{}{:.3f}".format("Loss:".ljust(25, " "),
running_loss / num_samples))
for classe in config["classes"][1:]:
log.log("{}{:.3f}".format((classe["title"] + " IoU:").ljust(25, " "),
metrics.get_fg_iou()))