evaluation

[nationalflag]

The batchsize of eval.py must be 1?

[zmonoid]

Same problem, trying to solve with larger batch size.

[Jumabek]

@zmonoid would you mind sharing your solution?

[zmonoid]

@Jumabek Yes, basically, it is like:

for i, (inputs, box_targets, label_targets) in enumerate(dataloader):
        print('%d/%d' % (i, len(dataloader)))
        loc_preds, cls_preds = net(Variable(inputs.cuda(), volatile=True))


        for j in range(inputs.size(0)):
        	gt_boxes.append(box_targets[j])
        	gt_labels.append(label_targets[j])
	        box_preds, label_preds, score_preds = box_coder.decode(
	            loc_preds[j].cpu().data.squeeze(),
	            F.softmax(cls_preds[j].squeeze(), dim=1).cpu().data,
	            score_thresh=0.01)

	        pred_boxes.append(box_preds)
        	pred_labels.append(label_preds)
        	pred_scores.append(score_preds)

Still it is very slow. Main time consumption is at box_coder.decode.

[Jumabek]

Thanks I changed box decoding to gpu and got improvement from 17sec to 7 sec for single image

On Thu, May 31, 2018, 7:13 PM ZHOU Bin [email protected] wrote:

@Jumabek https://github.com/Jumabek Yes, basically, it is like:

for i, (inputs, box_targets, label_targets) in enumerate(dataloader): print('%d/%d' % (i, len(dataloader))) loc_preds, cls_preds = net(Variable(inputs.cuda(), volatile=True))

    for j in range(inputs.size(0)):
    	gt_boxes.append(box_targets[j])
    	gt_labels.append(label_targets[j])
      box_preds, label_preds, score_preds = box_coder.decode(
          loc_preds[j].cpu().data.squeeze(),
          F.softmax(cls_preds[j].squeeze(), dim=1).cpu().data,
          score_thresh=0.01)

      pred_boxes.append(box_preds)
    	pred_labels.append(label_preds)
    	pred_scores.append(score_preds)

Still it is very slow. Main time consumption is at box_coder.decode.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/kuangliu/torchcv/issues/14#issuecomment-393483892, or mute the thread https://github.com/notifications/unsubscribe-auth/AEJIQa62FDSz19b0CcDjiyOOju7re_psks5t38ItgaJpZM4SiGcI .

[Jumabek]

you might find https://github.com/kuangliu/torchcv/issues/29 informative

[Jumabek]

Speed is a lot faster when I am evaluating later models like after 300K iter

[vaishnavm217]

@Jumabek can u share the gpu implementation? I'm always getting this: RuntimeError: Expected object of type torch.cuda.FloatTensor but found type torch.FloatTensor for argument #2 'other' Is this because my other variables are in CPU? Update: Figured it out. But still is very slow

[vaishnavm217]

Why is the score threshold kept very low?? The VOC function needs all the generated boxes? The decoding is very fast when the threshold is high