Code issues

[kalpa61]

hi, you did a great job, but I ran into some problems when I tried to reproduce it. First, there was something wrong with the code in casmeii.py, and it wouldn't work without some changes, so I made some changes, as follows:

    if self.phase=='train':
        image_on = cv2.cvtColor(image_on0, cv2.COLOR_BGR2GRAY)
        image_apex = cv2.cvtColor(image_apex0, cv2.COLOR_BGR2GRAY)
        flow_1 = cv2.calcOpticalFlowFarneback(image_on, image_apex, None, 0.5, 3, 15, 3, 5, 1.2, 0)
        label = self.label[idx]
        flow_1 = self.transform_flow(flow_1)
        image_on0 = self.transform_onset(image_on0)
        if self.transform_aug is not None:
            flow_1 = self.transform_aug(flow_1)
            image_on0 = self.transform_aug(image_on0)
        return flow_1, image_on0, label
    else:
        image_on = cv2.cvtColor(image_on0, cv2.COLOR_BGR2GRAY)
        image_apex = cv2.cvtColor(image_apex0, cv2.COLOR_BGR2GRAY)
        flow_1 = cv2.calcOpticalFlowFarneback(image_on, image_apex, None, 0.5, 3, 15, 3, 5, 1.2, 0)
        label = self.label[idx]
        flow_1 = self.transform_flow(flow_1)
        image_on0 = self.transform_onset(image_on0)
        if self.transform_aug is not None:
            #flow_1 = self.transform_aug(flow_1)
            image_on0 = self.transform_aug(image_on0)
            image_apex0 = self.transform_aug(image_apex0)
        return flow_1,image_on0,label

Secondly, the training result is very unsatisfactory. The Acc of casme2 dataset is only 77% when it is trained under the condition of 5 classification. I am very confused about the problem, and my training environment is python3.8 + torch 2.1