Error converting to onnx: forward function contains for loop

Hello, there is a for loop in my forward function. When I turned to onnx, the following error occurred:

[ONNXRuntimeError] : 1 : FAIL : Non-zero status code returned while running Split node. Name:'Split_ 1277' Status Message: Cannot split using values in 'split' attribute. Axis=0 Input shape={59} NumOutputs=17 Num entries in 'split' (must equal number of outputs) was 17 Sum of sizes in 'split' (must equal size of selected axis) was 17

Part of my forward code：

    y, ey, x, ex = pad(boxes, w, h)
    if len(boxes) > 0:
        im_data = []
        indx_y = torch.where(ey > y-1)[0]
        for ind in indx_y:
             img_k =  imgs[image_inds[ind],:, (y[ind] - 1).type(torch.int64):ey[ind].type(torch.int64), (x[ind]-1).type(torch.int64):ex[ind].type(torch.int64)].unsqueeze(0)
             im_data.append(imresample(img_k, (24, 24)))
        im_data = torch.cat(im_data, dim=0)
        return im_data

I found that during the first onnx conversion, the for loop was executed 17 times, but when I tested it, the for loop required 59 times, so there was an error. In the forward function, indx_y is dynamic, so the number of for loops is also dynamic. Is there any way to solve this problem？

cc @neginraoof

Start with the provided forward function and reproduce the ONNX conversion using inputs where indx_y produces different numbers of iterations. Compare the conversion-time and test-time Split shapes and confirm that the resulting model handles the dynamic loop count without the reported Split mismatch.