Back propogation from 2 different positions of network

[mfatih7]

Hello

I want to construct a network architecture that will be trained with respect to 2 different losses.

One of the losses will be computed with respect to the output(y) of the network.
The other loss will be computed with respect to some variables(w) that are not output.
I have both ground truth labels for y and w variables.

I have a network(F) that is trained successfully. Its input is x and output is w.
I can generate a G network that instantiates F network.
G network does not need any learnable parameters.
G network will take the w outputs of F network and compute y values.
All of the calculations of G network can be implemented in its forward() function.
All of the calculation of G network is differentiable.

If I generate a network like G( F(x) ) = y is it possible to backpropagate both

• From w to x
• From y to x

Any code example will be very helpful

thank you

[mfatih7]

Hello

Let me ask another question

criterion=torch.nn.MSELoss(size_average=True, reduce=True, reduction='mean')
optimizer=torch.optim.Adam(model.parameters(), lr=learning_rate)

…

outputs_1, outputs_2 = model(inputs_1, inputs_2)
loss_train = criterion(outputs_1, labels_1)
optimizer.zero_grad()
loss_train.backward()
optimizer.step()  

In the code above, how can Pytorch understand that loss_train.backward() operates with respect to outputs_1 not outputs_2 ?

outputs_1 is a tensor. Isn’t it similar to a variable like numpy array?
Or does it carry information about the model architecture?

[mfatih7]

Hi

I think the answers below have sufficient information.

A model with multiple outputs

I have to implement a Convolutional Neural Network, that takes a kinect image (1640480) and return a 1 x8 tensor predicting the class to which the object belongs and a 1 x 4 tensor, predicting the bounding box around the image, if its present. Please help me how can I implement a suitable model to give two outputs and how to calculate loss and backpropagate in that case? Also, i have just around 6000 training images, how can I achieve the best possible results with limited number of training i…

How to weight my own multiple losses?

I have multiple losses, and i want to weight them seperately. More specifically, i want to set the weight of loss1 to 1, and the weight of loss2 to 0.5, is that correct? loss1 = nn.L1Loss(B, A) loss2 = nn.L1Loss(C, A) loss2 = loss2*0.5 loss1.backward(retain_variables=True) loss2.backward() optimizer.step()

On the other hand, If Pytorch publishes more tutorials, especially non-vanilla learning models and mechanisms it will be beneficial in terms of practicality.