# RPN week13 實驗

## 1 epoch

train

val

time:
386

## 2 epoch

train

val

time:
703

# 4 epoch

train

val

time
1308.7086091041565

# IOU answer
def IOU_threshold(num_batch, label, max_ious_all, gt_argmax_ious_all):
    pos_iou_threshold  = 0.7
    neg_iou_threshold = 0.3  
    
    label_all = []
    for n in range(num_batch):
        l = copy.deepcopy(label)
        l[max_ious_all[n] < neg_iou_threshold] = 0
        l[gt_argmax_ious_all[n]] = 1
        l[max_ious_all[n] >= pos_iou_threshold] = 1
        label_all.append(l)
    #print ("label_all 0 and 1: ", sum(label_all[0]), sum(label_all[1]))     
    
    pos_ratio = 0.5
    n_sample = 256
    n_pos = int(pos_ratio * n_sample)
    #print(n_pos)    
    
    for n in range(num_batch):
        #print(np.sum((label_all[n] == 1)))
        pos_index = np.where(label_all[n] == 1)[0]
        #print(pos_index)
        if len(pos_index) > n_pos:
            disable_index = np.random.choice(pos_index, size=(len(pos_index) - n_pos), replace=False)
            label_all[n][disable_index] = -1
        #print(np.sum((label_all[n] == 1)))  

        n_neg = n_sample - np.sum(label_all[n] == 1)
        neg_index = np.where(label_all[n] == 0)[0]
        if len(neg_index) > n_neg:
            disable_index = np.random.choice(neg_index, size=(len(neg_index) - n_neg), replace = False)
            label_all[n][disable_index] = -1
        #print(np.sum((label_all[n] == 0))) 
    return label_all

# Implementation 2: count GTBbox answer
def d_calculatrion(base_ctr_y_all, base_ctr_x_all, ctr_y, ctr_x, base_height_all, base_width_all, height, width, num_batch):
    anchor_locs_all = []
    for n in range(num_batch):
        dy = (base_ctr_y_all[n].numpy() - ctr_y) / height
        dx = (base_ctr_x_all[n].numpy() - ctr_x) / width
        dh = np.log(base_height_all[n].numpy()/ height)
        dw = np.log(base_width_all[n].numpy() / width)
        anchor_locs_all.append(np.vstack((dy, dx,dh, dw)).transpose())
    return anchor_locs_all

epoch4

epoch2