Given groups=64, weight of size [64, 1, 5, 5], expected input[1, 1, 84, 128] to have 64 channels, but got 1 channels instead

[manish686]

Failing for gray scale images

[mnxtr]

Here’s a concise comment you can post on that issue:

---

Thanks for reporting this. The error indicates a channel mismatch: the grouped convolution layer is configured with groups=64 (so it expects 64 input channels), but the input tensor is grayscale with only 1 channel: input[1, 1, 84, 128].

If you’re using grayscale images, a practical workaround is to convert them to 3‑channel RGB before feeding them to the model, for example:

• With PIL:

from PIL import Image

img = Image.open(path).convert("RGB")

• With OpenCV:

import cv2

img = cv2.imread(path, cv2.IMREAD_GRAYSCALE)
img = cv2.merge([img, img, img])  # [H, W] -> [H, W, 3]

This ensures the input has the expected number of channels and avoids the groups/channel mismatch. If desired, the repo could also add a short note in the docs that grayscale images should be converted to RGB, or add an explicit preprocessing step in the dataset loader.