Skip to content

Use 128*128 for training #7

@leyuan-sun

Description

@leyuan-sun

When I used 128*128 for training, the following errors have been shown

[rank0]: context_feats, context, rgb_features = self.encoder.encode_images(
[rank0]: File "/home/sly/FlowRAM/diffuser_actor/utils/encoder.py", line 185, in encode_images
[rank0]: rgb_features = rgb_features[self.feature_map_pyramid[0]] # flow-8 都是256的 original
[rank0]: File "/home/sly/anaconda3/envs/flowram/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1962, in getattr
[rank0]: raise AttributeError(
[rank0]: AttributeError: 'Encoder' object has no attribute 'feature_map_pyramid'. Did you mean: 'feature_pyramid'?

In encoder.py if 128128,then feature_map_pyramid realled coarse_feature_map, but even I used coarse_feature_map, still new errors about tensor shape existed, could you plz try using 128128 with your code?
if self.image_size == (128, 128):
print("self.image_size=",self.image_size)
# Coarse RGB features are the 2nd layer of the feature pyramid
# at 1/4 resolution (32x32)
# Fine RGB features are the 1st layer of the feature pyramid
# at 1/2 resolution (64x64)
self.coarse_feature_map = ['res5', 'res1', 'res1', 'res1']
self.downscaling_factor_pyramid = [4, 2, 2, 2]
elif self.image_size == (256, 256):
# Coarse RGB features are the 3rd layer of the feature pyramid
# at 1/8 resolution (32x32)
# Fine RGB features are the 1st layer of the feature pyramid
# at 1/2 resolution (128x128)
self.feature_map_pyramid = ['res3', 'res1', 'res1', 'res1']
self.downscaling_factor_pyramid = [8, 2, 2, 2]

Or could you plz provide 256*256 dataset, only one task data is fine, I just want to run your code no need to check the final performance. Thanks in advance!!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions