Open mikeizbicki opened 4 years ago
@jwsbrennan There's two possibilities: either you're not creating your scores
vector properly (most likely), or you're not calling the line2img
function correctly.
You should print out the actual values of your scores
vector to inspect what they're equal to. Based on the fact that they get lighter as the index moves to the right, my guess is that somehow the formula you are using to calculate them contains the index in it in a way that it shouldn't be.
I'm seeing a similar issue with the visualization, except that even the blanks are all colored differently. I've attached my code below, and I'm wondering if you can give me a hint on how to debug this code. Thank you!
line = line.strip()
if args.case_insensitive:
line = line.lower()
input_tensor = str_to_tensor([line],args.input_length) # size = [x, 1, 76]
output_class,output_nextchars = model(input_tensor)
probs = softmax(output_class)
scores = torch.zeros([len(line)])
print("input_tensor: ", input_tensor)
for i in range(len(line)):
new_tensor = input_tensor
new_tensor[i+1,0] = torch.zeros([len(vocabulary)]) # +1 to exclude the starting char
new_output,new_nextchars = model(new_tensor)
new_probs = softmax(new_output)
scores[i] = torch.dist(probs, new_probs)
line2img(line,scores,filename)
@keweizhou1999 The first thing I notice is that torch.dist
does not calculate the l2 form between two vectors, so that could be messing up your results.
I've tried using torch.dist(probs, new_probs, 2) and torch.norm(), but it's still producing some sort of all green figure
For anyone else who has this same problem, we looked at it in office hours and realized that the problem was that line
new_tensor = input_tensor
does not create a copy of the tensor, and you need to explicitly call the clone
function.
When I'm running my code, the output I get looks like this:
Any idea what I'm doing wrong? Side note: the pattern of starting dark green and getting slightly lighter as it goes on is true for the character-by-character analysis as well
Originally posted by @jwsbrennan in https://github.com/mikeizbicki/cmc-csci181/issues/22#issuecomment-615332150