CODEWITHSUNDEEP
pythondeep-learningpytorchmlp
Siddeshwar Raghavan
Siddeshwar Raghavan

Reputation: 87

Confused about stacked tensor input to Multilayer perceptron in Pytorch

I have an input sequence in the form sequence_len x C x H x W = [10, 3, 16, 16] (Assuming batch size = 1). These are 10 images stacked together in a torch tensor. I wish to pass this to an MLP and obtain the next 10 as predictions from the MLP. The structure of MLP has one hidden layer with 32 units. If I flatten the input from dimension 1 - [10, 768]

My current code looks like this:

class MLP3(nn.Module):
    
    def __init__(self, ip_layers):
        super().__init__()
        self.layers = nn.Sequential(
        nn.Linear(ip_layers, 32),
        nn.ReLU(),
        nn.Linear(32, 32),
        nn.ReLU(),
        nn.Linear(32, 10)
    )

    def forward(self, x):
        #forward
        return self.layers(x)

However, I am unable to pass the entire tensor and not sure how I can obtain 10 outputs from the MLP. Any help will be highly appreciated. TIA

Upvotes: 0

Views: 408

Answers (1)

Abhishek Prajapat
Abhishek Prajapat

Reputation: 1888

To get 10 outputs at the end you have to use a multi-channel model. For that, you have to define the base_model in your __init__ function and modify your other datasets and the forward function and write loss for every class. I will write a boilerplate for you

class SDataset(torch.nn.utils.Dataset):
  def __init__(self):
    ....
  def __getitem__(self, idx):
    ....
    '''
    Till now everything was the same as before.
    Now the below thing is something that you have to change.
    '''
    outimg1 = self.image_seq[idx+1]
    outimg2 = self.image_seq[idx+2]
    .
    .
    outimg10 = self.image_seq[idx+10]

    return X, (outimg1, outimg2 ..... outimg10)

The model

class Predictor(torch.nn.Module):
  def __init__(self):
    super(Predictor, self).__init__()
    self.base_model = somebasemodel
    self.out1 = # Takes input from base_model and output an image
    self.out2 = # ....
    .
    .
    self.out10 = # ...

  def forward(self, x):
    features = self.base_model(x)
    res1 = self.out1(features)
    .
    .
    res10 = self.out10(features)
    return res1, res2, .... res10

The training step

inputs, outputs = batch
preds = model(inputs)
optimizer.zero_grad()
loss1 = criterion(preds[0], outputs[0])
.
.
loss10 = criterion(preds[9], outputs[9])
total_loss = loss1 + loss2 + ... + loss10
total_loss.backward()
optimizer.step()

Now that was the most of it.

Upvotes: 0

Related Questions