[C3W2B] Issues with UNQ_2 cell

[Jaspier]

When I run the test for UNQ_2 cell, the notebook prints this output:

This is kinda of a pesudocode of what I’m doing:

• First I produce fake_images passing condition as the generator’s parameter
• Then, I get predictions after running the discriminator having fake_images, and condition as parameters
• I proceed to calculate the adv_loss using adv_criterion using predictions and real as parameters
• and calculate the recon_loss using recon_criterion having fake_images and real as parameters
• Finally, I calculate gen_loss as adv_loss + lambda * (recon_loss)

Is there something wrong? What I can do to solve this issue ?

Thanks!
p.s:
hope this pseudocode doesn’t hurt the community’s criteria . Feel free to remove if it does.

[paulinpaloalto]

Here is the complete code for that test cell:

# UNIT TEST
def test_gen_reasonable(num_images=10):
    gen = torch.zeros_like
    disc = lambda x, y: torch.ones(len(x), 1)
    real = None
    condition = torch.ones(num_images, 3, 10, 10)
    adv_criterion = torch.mul
    recon_criterion = lambda x, y: torch.tensor(0)
    lambda_recon = 0
    assert get_gen_loss(gen, disc, real, condition, adv_criterion, recon_criterion, lambda_recon).sum() == num_images

    disc = lambda x, y: torch.zeros(len(x), 1)
    assert torch.abs(get_gen_loss(gen, disc, real, condition, adv_criterion, recon_criterion, lambda_recon)).sum() == 0

    adv_criterion = lambda x, y: torch.tensor(0)
    recon_criterion = lambda x, y: torch.abs(x - y).max()
    real = torch.randn(num_images, 3, 10, 10)
    lambda_recon = 2
    gen = lambda x: real + 1
    assert torch.abs(get_gen_loss(gen, disc, real, condition, adv_criterion, recon_criterion, lambda_recon) - 2) < 1e-4

    adv_criterion = lambda x, y: (x + y).max() + x.max()
    assert torch.abs(get_gen_loss(gen, disc, real, condition, adv_criterion, recon_criterion, lambda_recon) - 3) < 1e-4
test_gen_reasonable()
print("Success!")

You can see that the test cell is playing some pretty sophisticated games using the fact that the Function Under Test takes functions as some of its arguments. It looks like your code is failing the very first test. Notice in that case that real is in fact set to None and the adv_criterion function is torch.mul.

It’s fine to use pseudo-code to describe your solution. Your first two steps agree with mine. I think your mistake is that you should not be passing real “as is” to the adv_criterion function. My reading of the instructions is that gets fed to the reconstruction criterion function first. My code from that line on looks quite a bit different than yours.

[Jaspier]

Thank you! I figured out the issue was in this part:

@Jaspier

I proceed to calculate the adv_loss using adv_criterion using predictions and real as parameters

Instead of comparing with predictions vs real, It should be predictions vs tensor with the same shape as predictions, but whose all values are = 1, as showed in the video lecture.

[Hannah_Xu1]

Thank you for giving the hints. I stucked here for hours.