Error: Roboflow-Custom-Scaled-YOLOv4

[Mohamed]

I found the Issue also reported in the YOLOv7 repo (creator of Scaled-YOLOv4 also made YOLOv7).

It seems they introduced the same error in YOLOv7. This comment notes how to fix it:

• RuntimeError: result type Float can't be cast to the desired output type long int · Issue #35 · WongKinYiu/yolov7 · GitHub

The fix is to convert to long the "gain" of the gridspace on declaration.

In all lines with
gain = torch.ones(7, device=targets.device)

change to
gain = torch.ones(7, device=targets.device).long()

For some reason newest versions of cuda/torch do not do this cast automatically when needed

gain = torch.ones(7, device=targets.device) is located here utils/general.py#L512

*** try updating utils/general.py#L512 to read:
gain = torch.ones(7, device=targets.device).long()

and utils/general.py#L1101 to read:

targets.append([i, cls, float(x.cpu()),   
              float(y.cpu()),   
              float(w.cpu()),   
              float(h.cpu()),   
              float(conf.cpu())])

^^ To do this, open the notebook in Colab, then “Save a Copy in Drive”

1. Ensure your Runtime is set to GPU:
   

2. Run the first cell:

# clone Scaled_YOLOv4
!git clone https://github.com/roboflow-ai/ScaledYOLOv4.git  # clone repo
%cd /content/ScaledYOLOv4/
#checkout the yolov4-large branch
!git checkout yolov4-large

3. Edit ScaledYOLOv4/utils/general.py by navigating to it in your Colab directory, and double-clicking on general.py to open it:
   

   

4. Edit Line 512 to this:
   gain = torch.ones(7, device=targets.device).long() # normalized to gridspace gain
   

and Edit Line 1101 to this:

targets.append([i, cls, float(x.cpu()),   
              float(y.cpu()),   
              float(w.cpu()),   
              float(h.cpu()),   
              float(conf.cpu())])

• original Line 1101

5. CTRL+s to save your changes

6. Begin running the cells the rest of the notebook, starting with this cell:

import torch
print('Using torch %s %s' % (torch.__version__, torch.cuda.get_device_properties(0) if torch.cuda.is_available() else 'CPU'))

7. Be sure you generate your dataset version, ensure Resize is to 416x416 (default input size for the model, and offering the fastest inference speeds) – then export your dataset in YOLOv5 PyTorch TXT format, and copy/paste your code snippet in the cell highlighted in my screenshot, below (you can skip the cell just above it):
   

   

8. Run the training cell, and it works! (at least it did for me, just now):
   

   

• My shared notebook