zc/mlp-project
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Added transcrowd_gap

Browse source
This commit is contained in:
Zhengyi Chen 2024-02-05 14:01:27 +00:00
parent bcff06f9c2
commit 322d7f9ea5
2 changed files with 81 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

8
model/reverse_perspective.py
View file

@ -37,9 +37,8 @@ class PerspectiveEstimator(nn.Module):
After all, it is reasonable to say that you see more when you look at
faraway places.
This do imply that **we need to obtain a reasonably good feature extractor
from general images before training this submodule**. Hence, for now, we
prob. should work on transformer first.
The paper utilizes a unsupervised loss -- "row feature density" refers to
the density of features computed from ?
:param input_shape: (N, C, H, W)
:param conv_kernel_shape: Oriented as (H, W)
@ -63,6 +62,8 @@ class PerspectiveEstimator(nn.Module):
(_, _, height, width) = input_shape
# Sanity checking
# [TODO] Maybe this is unnecessary, maybe we can automatically suggest new params,
# but right now let's just do this...
(_conv_height, _conv_width) = (
np.floor(
(height + 2 * conv_padding - conv_dilation * (conv_kernel_shape[0] - 1) - 1)
@ -112,4 +113,5 @@ class PerspectiveEstimator(nn.Module):
return out
# def unsupervised_loss(predictions, targets):