NaN probabilities for step_sample about diffuseq HOT 4 CLOSED

Usually the code won't get NaN, but it's ok to add this line.

from diffuseq.

Hello！I met the same error, and I use your method. But It makes the sum of p is not 1, there is another error, because of np.random.choice . so I change the code to:

        p = w / np.sum(w)
        if np.sum(np.isnan(p)) > 0:
            p[np.isnan(p)] = 0
            p = p / np.sum(p)
        indices_np = np.random.choice(len(p), size=(batch_size,), p=p)
        indices = th.from_numpy(indices_np).long().to(device)

But I got the NAN error again.

This is my train.sh, I just use my dataset and change the dim

Hope you can give me some advice

from diffuseq.

Hello, I also meet Nan error when using my own dataset.
One advice is to set you batch size smaller, I set --bsz 64.(It works for me)
Another advice is to change the code,

        if hasNan:
            print("has Nan prob p=",p)
            size = len(p)
            p = np.ones(p.shape) * (1/size)
            print("new p =",p)

from diffuseq.

Ok, I'll try your method, thank you!!!!

commented

from diffuseq.