eriklindernoren/PyTorch-YOLOv3

Two bugs need to be fixed for training.

zhaowt61 opened this issue · 4 comments

zhaowt61 commented

1 change the " ptr = 0" in the function "load_weights(self, weights_path):" to "ptr=5". Because they added header_info into the weights file, which makes the loading mismatched by 5 bits. This bug is hidden so deep!
2 remove "model.eval()" in "detect.py"

The above solutions worked perfectly for me.

BeibinLi commented

I do not think these two changes will make difference.

  1. The original code is correct. The five bits are read in "header = np.fromfile(fp, dtype=np.int32, count=5)" line.

  2. "detect.py" does not affect training. So, removing "model.eval()" would not affect the training result either. Moreover, "model.eval()" is strongly advised for testing.

  • 👍1
zhaowt61 commented

I do not think these two changes will make difference.

1. The original code is correct. The five bits are read in "header = np.fromfile(fp, dtype=np.int32, count=5)" line.

2. "detect.py" does not affect training. So, removing "model.eval()" would not affect the training result either. Moreover, "model.eval()" is strongly advised for testing.

The first change is necessary, because I found that the loading and saving of weights are mis-matched.

  • 👍1
stevehjc commented

1 change the " ptr = 0" in the function "load_weights(self, weights_path):" to "ptr=5". Because they added header_info into the weights file, which makes the loading mismatched by 5 bits. This bug is hidden so deep!
2 remove "model.eval()" in "detect.py"

The above solutions worked perfectly for me.

I think ptr=0 is right, but definition of self.header_info is wrong:
modify the
self.header_info = np.array([0, 0, 0, self.seen, 0])
to
self.header_info = np.array([0, 0, 0, self.seen, 0],dtype=np.int32),default np.int is np.int64. This is why these are 5bit offset.

ygliang2009 commented

I do not think these two changes will make difference.

1. The original code is correct. The five bits are read in "header = np.fromfile(fp, dtype=np.int32, count=5)" line.

2. "detect.py" does not affect training. So, removing "model.eval()" would not affect the training result either. Moreover, "model.eval()" is strongly advised for testing.

The first change is necessary, because I found that the loading and saving of weights are mis-matched.

when loading yolov3 official yolov3.weights, need to use ptr = 0, when loading the model from saving trained, need to modify ptr = 0 to ptr = 5

  • 👍5
  • 🎉1