First of all, I’m new to deep learning and Pytorch, so I apologise in advance for my question.
I have a model that consists in 3 conv2D layers, and ReLU activations. It takes as an input grayscale images, where the dynamic range is between [0,255], that are normalised to the interval [0,1]. The input images have some black regions, some white ones, and others in between.
However, the outputs dynamic range is compressed to [0.4,0.41]. All images are grey, even after a renormalisation to bring them back to [0,255].
I’m a bit lost and I don’t find why this is the case.
What I tried:
- Plot histograms of gradients, although I’m not sure how to interpret this. It seems that for some parameters the model stops learning after a few epochs. For others one layer stops learning.
- Modify kernel sizes
- Add more layers
- Try quite a lot of different values of learning rate
Any help is appreciated. Please let me know if more information is needed. And if this is not the place to ask this kind of questions, please let me know where I can get some help.
Thank you for taking the time to help!
This is the model:
self.model = nn.Sequential(
nn.Conv2d(in_channels=1, out_channels=64, kernel_size=13, stride=1, padding=5),
nn.Conv2d(in_channels=64, out_channels=32, kernel_size=3, stride=1, padding=2),
nn.Conv2d(in_channels=32, out_channels=1, kernel_size=5, stride=1, padding=2),
def forward(self, input_image):
output_image = self.model(input_image)
Regarding my training script, for each epoch I do the following:
running_loss = 0
last_loss = 0
batch_index = 0
for batch, loader in enumerate(train_dataloader):
input_img = loader['input'].to(device, non_blocking=True)
gt_img = loader['gt'].to(device, non_blocking=True)
output = model(input_img)
loss = loss_fn(output, gt_img)
running_loss += loss.item()
if batch_index % 100 == 99:
last_loss = running_loss / 1000
running_loss = 0.0
batch_index += 1
Let me know if more information is needed