No Expected Model Output dimensions #1367
Description
Hello,
I am using an autoencoder but the output channel dimension of the model does not correspond with the out-channel variable.
Here is my code:
# create a training data loader
train_ds = ArrayDataset(img_train[0], train_imtrans, img_train[1], train_segtrans)
train_loader = DataLoader(train_ds, batch_size=16, num_workers=8, pin_memory=torch.cuda.is_available())
# create a validation data loader
val_ds = ArrayDataset(img_val[0], val_imtrans, img_val[1], val_segtrans)
val_loader = DataLoader(val_ds, batch_size=1, num_workers=4, pin_memory=torch.cuda.is_available())
# create AE, MSELoss and Adam optimizer
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = monai.networks.nets.AutoEncoder(
dimensions=2,
in_channels=1,
out_channels=4,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units = 2
)
loss_function = torch.nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), 1e-3)
lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=3, verbose=True)
# start a typical PyTorch training
val_interval = 2
best_metric = -1
best_metric_epoch = -1
best_val_loss = 10000
best_val_loss_epoch = 10000
epoch_loss_values = list()
val_loss_values = list()
metric_values = list()
writer = SummaryWriter()
post_pred = AsDiscrete(argmax=True, to_onehot=True, n_classes=4)
post_label = AsDiscrete(to_onehot=True, n_classes=4)
for epoch in range(200):
print("-" * 10)
print('Epoch {}/{}'.format(epoch + 1, 200))
model.train()
epoch_loss = 0
step = 0
for batch_data in train_loader:
metric_sum = 0.0
metric_count = 0
metric_batch = 0
step += 1
inputs, labels = batch_data[0], batch_data[1]
optimizer.zero_grad()
outputs = model(inputs)
loss = loss_function(outputs, labels)
loss.backward()
optimizer.step()
epoch_loss += loss.item()
epoch_len = len(train_ds) // train_loader.batch_size
outputs1 = post_pred(outputs)
labels1 = post_label(labels)
value_train = monai.metrics.compute_meandice(y_pred=outputs1, y=labels1, include_background=False)
metric_count += value_train.shape[1]
metric_batch += value_train.shape[0]
metric_sum += value_train.sum().item()
writer.add_scalar("train_loss", loss.item(), epoch_len * epoch + step)
metric1 = metric_sum / (metric_count * metric_batch)
epoch_loss /= step
epoch_loss_values.append(epoch_loss)
print(f"epoch {epoch + 1} average loss: {epoch_loss:.4f} mean dice: {metric1:.4f}")
I am getting this error:
RuntimeError Traceback (most recent call last)
in
224
225 with tempfile.TemporaryDirectory() as tempdir:
--> 226 outputs, labels, outputs1, labels1, inputs, labels = main(tempdir, img_train, img_val)
in main(tempdir, img_train, img_val)
121 epoch_loss += loss.item()
122 epoch_len = len(train_ds) // train_loader.batch_size
--> 123 outputs1 = post_pred(outputs)
124 labels1 = post_label(labels)
125 value_train = monai.metrics.compute_meandice(y_pred=outputs1, y=labels1, include_background=False)
~\AppData\Local\Programs\Python\Python36\lib\site-packages\monai\transforms\post\array.py in call(self, img, argmax, to_onehot, n_classes, threshold_values, logit_thresh)
156 _nclasses = self.n_classes if n_classes is None else n_classes
157 assert isinstance(_nclasses, int), "One of self.n_classes or n_classes must be an integer"
--> 158 img = one_hot(img, _nclasses)
159
160 if threshold_values or self.threshold_values:
~\AppData\Local\Programs\Python\Python36\lib\site-packages\monai\networks\utils.py in one_hot(labels, num_classes, dtype, dim)
45
46 o = torch.zeros(size=sh, dtype=dtype, device=labels.device)
---> 47 labels = o.scatter_(dim=dim, index=labels.long(), value=1)
48
49 return labels
RuntimeError: index 5 is out of bounds for dimension 1 with size 4
I have checked the input and output shapes of the model and I got the next:
inputs: torch.Size([16, 1, 256, 256])
outputs: torch.Size([16, 16, 256, 256])
Why if I set the out-channel to 4 I got a channel of 16 in the outputs?