I have implemented the combining as I did in the assignment C1W3_WGAN_GP:
I created a tensor of booleans named epsilon by assigning the result of comparing a tensor of torch.rand (of real image size) to be less than the given probability of real images.
I then assigned to target_images the sum of real images times epsilon and fake images times logical not epsilon
I think that the assert “# Make sure that no mixing happened” might mean that tensor target_images is the correct length but I don’t know what “mixing” means.
I used torch.multinomial to randomly select a proportion of uniform index indicators (1s). This produced an index at random positions of the correct proportions. I then used the index to select a subset from the sample. I did this for the real and fake indices. I concatenated the selected samples. I also concatenated the selected indices. I sorted the combined indices in ascending order and used these to select the concatenated samples.
I am however, not getting the order right and am incurring the following assert error:
It may be that I’m simply not understanding what you mean, but that all sounds a bit worrying. The point is to do things “in place” and also “row-wise”, not “element-wise”. You clone the “reals” and then replace some of those samples (in-place) with fakes.
I also did not see any reference to torch.multinomial in the instructions. They do discuss using torch.rand or torch.bernoulli. I used torch.rand FWIW …
I had a look at the other thread and it makes much more sense now. It did say there were multiple ways to sort this, but what I did was wrong. I randomly sampled from each of real and fake and concatenated them in a way that the indexing wouldn’t have worked either. The solution described on the thread that you linked works. It swaps elements nicely preserving the order. Thanks again for your help.
Need some help as I am getting 2 different assert errors after going
through each assert one by one.
I have also read another lenghty thread as indicated by Paul to OP but
don’t believe the issues are similar over there.
Any suggestion will be greatly appreciated!
Not sure what it means by “mixing” in this context:
AssertionError Traceback (most recent call last)
Input In [123], in <cell line: 27>()
25 assert tuple(test_combination.shape) == (n_test_samples, 10, 10)
26 # Make sure that no mixing happened
---> 27 assert torch.abs((test_combination.sum([1, 2]).median()) - 100) < 1e-5
29 test_reals = torch.arange(n_test_samples)[:, None].float()
30 test_fakes = torch.zeros(n_test_samples, 1)
AssertionError:
Error about using 2 different devices. Don’alet suppose we should use the global variable device defined earlier, right?
RuntimeError Traceback (most recent call last)
Input In [125], in <cell line: 48>()
47 assert torch.abs(test_combination - test_reals).sum() < 1e-4
48 if torch.cuda.is_available():
49 # Check that the solution matches the input device
---> 50 assert str(combine_sample(
51 torch.ones(n_test_samples, 10, 10).cuda(),
52 torch.zeros(n_test_samples, 10, 10).cuda(),
53 0.8
54 ).device).startswith("cuda")
55 print("Success!")
Input In [117], in combine_sample(real, fake, p_real)
25 # print( f"mask.shape: {mask.shape}" )
26 # print( mask )
---> 27 target_images = ### I am using torch.where( ) here ###
29 print( target_images.shape )
31 #### END CODE HERE ####
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!
The mixing comment most likely has to do with the points about preserving the order of the reals as was discussed in detail on that other thread that I linked. But let’s fix the two device problem first and then see what happens. Did you use torch.ones anywhere? I did not need that, but if you do use it, try using torch.ones_like instead to make sure the device is the same. Or did you explicitly assign the device anywhere in your logic? That should also not be necessary.
Did you try running the experiment I showed on this post in that other thread with your code? Please show us the output you get from that.
I figured out the bit about different device. I started with torch.where() and
that appears to be the main culprit. No matter what I tried the cpu & cuda
issue persists.
I then looked at your replies to that other very long topic thread and started
experimenting with boolean indexing of torch tensor. Although with wrong
answer to begin with but at finally overcome the issue of cpu and cuda!
So … I stuck with that and also re-read your replies many many many times to try to clarify my understanding. And then tried many many many
different things and I finally figured out the way I did with torch.where()
is just not going to work.
I got it now … re-iterating the fact that your replies and also the suggestion
about the array indexing short-cut way. Thanks man!!!
I know I am not suppsoed to post codes, so I will just elaborate the logic
here.
Since p_real should be the probability for sampling from real, so I build a
“mask” following the hint that suggests:
torch.rand(length_reals) > p
This actually should give the mask to get distribution from the fakes, since
the logic is ‘> p_real’. I got this bit wrong as well before.
So, the rest then became a bit easier once I tested many combos of
your array indexing hint. All I needed to do is just applying the same “fake”
mask to both real and fake to replace the position(s) in real tensor with
content from the fake at same position(s). it’s the tensor position that really
is the key here.
It’s great to hear that you were able to get this to work using the suggestions on that other thread. There may well be a way to get this to work with torch.where, but I think the “logical indexing” approach is just easier to implement and is just clearer code. Clarity matters!
