increase batch size 1 to 16, it made wrong result. about dsntnn HOT 7 CLOSED

You haven't provided much detail here. What is the complete shape of the tensor that you are passing into the DSNT function when using batch size 16? How many keypoints are you trying to predict?

from dsntnn.

I want to predict 16 points.
I edit FCN's input channel 3 to 1.

in my code.
...

print(img.shape)
coords, heatmaps = model(img)

...

torch.Size([16, 1, 224, 224])

from dsntnn.

What is the shape of the heatmaps (FCN output)?

from dsntnn.

print(heatmaps.shape)

torch.Size([16, 16, 224, 224])

from dsntnn.

You probably have a problem somewhere in your code. I recommend trying a batch size that is quite different from your number of keypoints (e.g. 5) since you may be hiding an incorrect permutation of axes somewhere. Then take a careful look at any view/reshape operations that you might be using.

from dsntnn.

Ok i set batch size = 8

now

coords, heatmaps = model(img)

print(img.shape)
print(coords.shape)
print(heatmaps.shape)

torch.Size([8, 1, 224, 224])
torch.Size([8, 16, 2])
torch.Size([8, 16, 224, 224])

but i don't know what this means.

 Then take a careful look at any view/reshape operations that you might be using.

May you explain more thoroughly to me?

Which part should I watch in my code?
(e.g. dataset? loss function?)

from dsntnn.

I'm sorry, but unless you can produce a specific example of an issue with DSNTNN I'm going to have to close this issue. There are just too many possible reasons why your particular project might be failing.

from dsntnn.