how to predict?

[ansalrv2e]

can you provide prediction code to detect

[explainingai-code]

Hello @ansalrv2e ,

The prediction code is in infer.py.
Specifically evaluate_map allows you to compute average precision scores

SSD-PyTorch/tools/infer.py

Line 277 in 41b3090

def evaluate_map(args):

And infer has the logic to run inference on images and draw prediction boxes for visualization. If you want it to run for images outside of your dataset, for which you dont have ground truth, then additionally you would have to do the following changes.

1. Pass the image tensor as usual but targets as None in below line
   

   SSD-PyTorch/tools/infer.py

   Line 211 in 41b3090

   _, ssd_detections = model(im_tensor.unsqueeze(0).to(device), [target])

   
   

   SSD-PyTorch/model/ssd.py

   Line 455 in 41b3090

   def forward(self, x, targets=None):

2. Comment this entire block of infer method
   

   SSD-PyTorch/tools/infer.py

   Lines 216 to 239 in 41b3090

   	# Saving images with ground truth boxes
   	for idx, box in enumerate(target['bboxes']):
   	x1, y1, x2, y2 = box.detach().cpu().numpy()
   	x1, y1, x2, y2 = int(w*x1), int(h*y1), int(w*x2), int(h*y2)
   	cv2.rectangle(gt_im, (x1, y1), (x2, y2), thickness=2, color=[0, 255, 0])
   	cv2.rectangle(gt_im_copy, (x1, y1), (x2, y2), thickness=2, color=[0, 255, 0])
   	text = voc.idx2label[target['labels'][idx].detach().cpu().item()]
   	text_size, _ = cv2.getTextSize(text, cv2.FONT_HERSHEY_PLAIN, 1, 1)
   	text_w, text_h = text_size
   	cv2.rectangle(gt_im_copy, (x1, y1), (x1 + 10 + text_w, y1 + 10 + text_h), [255, 255, 255], -1)
   	cv2.putText(gt_im, text=voc.idx2label[target['labels'][idx].detach().cpu().item()],
   	org=(x1 + 5, y1 + 15),
   	thickness=1,
   	fontScale=1,
   	color=[0, 0, 0],
   	fontFace=cv2.FONT_HERSHEY_PLAIN)
   	cv2.putText(gt_im_copy, text=text,
   	org=(x1 + 5, y1 + 15),
   	thickness=1,
   	fontScale=1,
   	color=[0, 0, 0],
   	fontFace=cv2.FONT_HERSHEY_PLAIN)
   	cv2.addWeighted(gt_im_copy, 0.7, gt_im, 0.3, 0, gt_im)
   	cv2.imwrite('samples/output_ssd_gt_{}.png'.format(i), gt_im)

This should allow you to infer on images for which you don't have ground truth.