Trained Retinanet Object Detector Yields too Large Bounding Boxes

[tango (Lasse Karlsson)]

Using the notebook from the 2019 Part 2 course where Jeremy trained a retinanet object dector using the Pascal 2007 dataset, I tried to make an object detector of my own. I have changed a few things to make it work for my purposes. 1: I replaced the resnet model from the notebook with resnet18 (I haven’t changed the retinanet split, could this be a problem? Edit: resnet50 doesn’t make a difference). I have of course also changed the databunch to my own data.

After getting poor results where the bounding boxes are way too large, I tried normalizing using imagenet_stats as I thought this could differ from the Pascal dataset but this seemed to have little effect. My loss is steadily going down and with the loss function from the notebook and I achieve less than 1 (about 0.6).

Altogether I am training on about 150 images from my own hand-labelled dataset. Could the dataset size be the problem?

This is the notebook:
https://colab.research.google.com/drive/1ljLfGHGWtxkun2q4XdXRbhrntELex0ur

[almarca]

I have the same problem. Do you know something new?

[tango (Lasse Karlsson)]

No, on Monday I will try to Pretrain the head on the pascal dataset and fine tune it with my own dataset

[blueharen]

I have been trying to do the same thing as you. However, when I get to the actual training part with the bounding boxes, it requires an absurd amount of memory and I can’t run it, even on an AWS server. Do you have any idea how to combat this?

Thanks

[tango (Lasse Karlsson)]

Maybe lower your batch size? Are you scaling down the images?

[almarca]

I’ve solved my problem changed a single line in code.

def show_preds(img, output, idx, detect_thresh=0.25, classes=None, ax=None):
bbox_pred, scores, preds = process_output(output, idx, detect_thresh)
if len(scores) != 0:
    to_keep = nms(bbox_pred, scores)
    bbox_pred, preds, scores = bbox_pred[to_keep].cpu(), preds[to_keep].cpu(), scores[to_keep].cpu()
    t_sz = torch.Tensor([*img.size])[None].float()
    bbox_pred[:,:2] = bbox_pred[:,:2] - bbox_pred[:,2:]/2
    bbox_pred[:,:2] = (bbox_pred[:,:2] + 1) * t_sz/2
    bbox_pred[:,2:] = bbox_pred[:,2:] * t_sz/2
    bbox_pred = bbox_pred.long()
if ax is None: _, ax = plt.subplots(1,1)
img.show(ax=ax)
for bbox, c, scr in zip(bbox_pred, preds, scores):
    txt = str(c.item()) if classes is None else classes[c.item()+1]
    draw_rect(ax, [bbox[1],bbox[0],bbox[3],bbox[2]], text=f'{txt} {scr:.2f}')

Look bbox_pred[:,2:] = bbox_pred[:,2:] * t_sz/2, that works for me.

[blueharen]

@tango Thanks, I’m an idiot and forgot to size down the images.

Which brings me to my next issue. I was able to train a model (still had high losses of about 3.5 each) but now I am unable to view any predictions.
When I run show_preds(img, output, idx, detect_threshold=0.25, classes=data.classes , I get the following error.

I wanted to see what was going on in the show_preds function so I added 3 print statements and ran it.

It seems that all of my predictions are way below 0.25, and my bbox_pred is completely empty.

So I figured that changing the threshold to something lower would find these prediciton values as they are all under .25. Changing this to 0.07 was the highest I could go without throwing the RuntimeError. However, doing this makes the kernel run forever with no output as shown below. Interrupting it kills the kernel every time.

Any thoughts? Thanks

EDIT: Raising batch size fixed this.