Batch_size Modification

[WuXueQiCai-DaJiangShen]

It seems that when we use train_on_batch to train our Faster RCNN model, we couldn't change the batch_size directly as what in model.fit( ). So I just simply concatenated a mini-batch of X, X2, Y, Y1 and Y2 and put them into the training process. However, the results turned out pretty bad. What's the reason of that and how could I modify the batch_size in training of Faster RCNN model?

            ...
            if len(rpn_accuracy_rpn_monitor) == epoch_length and C.verbose:
                mean_overlapping_bboxes = float(sum(rpn_accuracy_rpn_monitor))/len(rpn_accuracy_rpn_monitor)
                rpn_accuracy_rpn_monitor = []
                print('Average number of overlapping bounding boxes from RPN = {} for {} previous iterations'.format(mean_overlapping_bboxes, epoch_length))
                if mean_overlapping_bboxes == 0:
                    print('RPN is not producing bounding boxes that overlap the ground truth boxes. Check RPN settings or keep training.')

#            X, Y, img_data = next(data_gen_train)

            batch_size = 4
            X = []
            Y = [[],[]]
            img_datas = []
            iter_batch = 0
            for iter_batch in range(batch_size):
                X_s, Y_s, img_data = next(data_gen_train)
                X.append(X_s[0])
                Y[0].append(Y_s[0][0])
                Y[1].append(Y_s[1][0])
                img_datas.append(img_data)
            X = np.array(X)
            Y = [np.array(Y[0]),np.array(Y[1])]

            loss_rpn = model_rpn.train_on_batch(X, Y)
            P_rpn = model_rpn.predict_on_batch(X)
            ...

            ...
            X2s = []
            Y1s = []
            Y2s = []
            X_classifier = []
            Y_classifier = []
            iter_batch = 0
            for iter_batch in range(batch_size):

                # note: calc_iou converts from (x1,y1,x2,y2) to (x,y,w,h) format
                X2, Y1, Y2 = roi_helpers.calc_iou(R[iter_batch], img_datas[iter_batch], C, class_mapping)

                if X2 is None:
                    rpn_accuracy_rpn_monitor.append(0)
                    rpn_accuracy_for_epoch.append(0)
                    continue

                neg_samples = np.where(Y1[0, :, -1] == 1)
                pos_samples = np.where(Y1[0, :, -1] == 0)

                if len(neg_samples) > 0:
                    neg_samples = neg_samples[0]
                else:
                    neg_samples = []

                if len(pos_samples) > 0:
                    pos_samples = pos_samples[0]
                else:
                    pos_samples = []

                rpn_accuracy_rpn_monitor.append(len(pos_samples))
                rpn_accuracy_for_epoch.append((len(pos_samples)))
                if C.num_rois > 1:
                    if len(pos_samples) < C.num_rois//2:
                        selected_pos_samples = pos_samples.tolist()
                    else:
                        selected_pos_samples = np.random.choice(pos_samples, C.num_rois//2, replace=False).tolist()
                    try:
                        selected_neg_samples = np.random.choice(neg_samples, C.num_rois - len(selected_pos_samples), replace=False).tolist()
                    except:
                        selected_neg_samples = np.random.choice(neg_samples, C.num_rois - len(selected_pos_samples), replace=True).tolist()
    
                    sel_samples = selected_pos_samples + selected_neg_samples
                else:
                    # in the extreme case where num_rois = 1, we pick a random pos or neg sample
                    selected_pos_samples = pos_samples.tolist()
                    selected_neg_samples = neg_samples.tolist()
                    if np.random.randint(0, 2):
                        sel_samples = random.choice(neg_samples)
                    else:
                        sel_samples = random.choice(pos_samples)
                
                X2s.append(X2[:, sel_samples, :][0])
                Y1s.append(Y1[:, sel_samples, :][0])
                Y2s.append(Y2[:, sel_samples, :][0])

            X2s = np.array(X2s)
            Y1s = np.array(Y1s)
            Y2s = np.array(Y2s)

#            loss_class = model_classifier.train_on_batch([X, X2[:, sel_samples, :]], [Y1[:, sel_samples, :], Y2[:, sel_samples, :]])
            
            if X2s.shape[0] != batch_size:
                continue
            loss_class = model_classifier.train_on_batch([X, X2s], [Y1s, Y2s])
            ...