I have a neural network which outputs output. I want to transform output before the loss and backpropogation happen.
Here is my general code:
with torch.set_grad_enabled(training):
outputs = net(x_batch[:, 0], x_batch[:, 1]) # the prediction of the NN
# My issue is here:
outputs = transform_torch(outputs)
loss = my_loss(outputs, y_batch)
if training:
scheduler.step()
loss.backward()
optimizer.step()
Following the advice in How to transform output of neural network and still train? , I have a transformation function which I put my output through:
def transform_torch(predictions):
new_tensor = []
for i in range(int(len(predictions))):
arr = predictions[i]
a = arr.clone().detach()
# My transformation, which results in a positive first element, and the other elements represent decrements of the first positive element.
b = torch.negative(a)
b[0] = abs(b[0])
new_tensor.append(torch.cumsum(b, dim = 0))
# new_tensor[i].requires_grad = True
new_tensor = torch.stack(new_tensor, 0)
return new_tensor
Note: In addition to clone().detach(), I also tried the methods described in Pytorch preferred way to copy a tensor, to similar result.
My problem is that no training actually happens with this tensor that is tranformed.
If I try to modify the tensor in-place (e.g. directly modify arr), then Torch complains that I can't modify a tensor in-place with a gradient attached to it.
Any suggestions?
Calling detach on your predictions stops gradient propagation to your model. Nothing you do after that can change your parameters.
How about modifying your code to avoid this:
def transform_torch(predictions):
b = torch.cat([predictions[:, :1, ...].abs(), -predictions[:, 1:, ...]], dim=1)
new_tensor = torch.cumsum(b, dim=1)
return new_tensor
A little test you can run, to verify that gradients do propagate through this transformation is:
# start with some random tensor representing the input predictions
# make sure it requires_grad
pred = torch.rand((4, 5, 2, 3)).requires_grad_(True)
# transform it
tpred = transform_torch(pred)
# make up some "default" loss function and back-prop
tpred.mean().backward()
# check to see all gradients of the original prediction:
pred.grad
# as you can see, all gradients are non-zero
Out[]:
tensor([[[[ 0.0417, 0.0417, 0.0417],
[ 0.0417, 0.0417, 0.0417]],
[[-0.0333, -0.0333, -0.0333],
[-0.0333, -0.0333, -0.0333]],
[[-0.0250, -0.0250, -0.0250],
[-0.0250, -0.0250, -0.0250]],
[[-0.0167, -0.0167, -0.0167],
[-0.0167, -0.0167, -0.0167]],
[[-0.0083, -0.0083, -0.0083],
[-0.0083, -0.0083, -0.0083]]],
[[[ 0.0417, 0.0417, 0.0417],
[ 0.0417, 0.0417, 0.0417]],
[[-0.0333, -0.0333, -0.0333],
[-0.0333, -0.0333, -0.0333]],
[[-0.0250, -0.0250, -0.0250],
[-0.0250, -0.0250, -0.0250]],
[[-0.0167, -0.0167, -0.0167],
[-0.0167, -0.0167, -0.0167]],
[[-0.0083, -0.0083, -0.0083],
[-0.0083, -0.0083, -0.0083]]],
[[[ 0.0417, 0.0417, 0.0417],
[ 0.0417, 0.0417, 0.0417]],
[[-0.0333, -0.0333, -0.0333],
[-0.0333, -0.0333, -0.0333]],
[[-0.0250, -0.0250, -0.0250],
[-0.0250, -0.0250, -0.0250]],
[[-0.0167, -0.0167, -0.0167],
[-0.0167, -0.0167, -0.0167]],
[[-0.0083, -0.0083, -0.0083],
[-0.0083, -0.0083, -0.0083]]],
[[[ 0.0417, 0.0417, 0.0417],
[ 0.0417, 0.0417, 0.0417]],
[[-0.0333, -0.0333, -0.0333],
[-0.0333, -0.0333, -0.0333]],
[[-0.0250, -0.0250, -0.0250],
[-0.0250, -0.0250, -0.0250]],
[[-0.0167, -0.0167, -0.0167],
[-0.0167, -0.0167, -0.0167]],
[[-0.0083, -0.0083, -0.0083],
[-0.0083, -0.0083, -0.0083]]]])
If you'll try this little test with your original code you'll either get an error that you are trying to propagate through tensors that do not require_grad, or you'll get no grads for the input pred.