# Generate X (x_img) and label Y ([y_rpn_cls, y_rpn_regr])
X, Y, img_data, debug_img, debug_num_pos = next(data_gen_train)
# Train rpn model and get loss value [_, loss_rpn_cls, loss_rpn_regr]
loss_rpn = model_rpn.train_on_batch(X, Y)
# Get predicted rpn from rpn model [rpn_cls, rpn_regr]
P_rpn = model_rpn.predict_on_batch(X)
# print(P_rpn[0])
# R: bboxes (shape=(300,4))
# Convert rpn layer to roi bboxes
R = rpn_to_roi(P_rpn[0],
P_rpn[1],
C,
'tf',
use_regr=True,
overlap_thresh=0.7,
max_boxes=300)
# note: calc_iou converts from (x1,y1,x2,y2) to (x,y,w,h) format
# X2: bboxes that iou > C.classifier_min_overlap for all gt bboxes in 300 non_max_suppression bboxes
# Y1: one hot code for bboxes from above => x_roi (X)
# Y2: corresponding labels and corresponding gt bboxes
X2, Y1, Y2, IouS = calc_iou(R, img_data, C, class_mapping)
# If X2 is None means there are no matching bboxes
if X2 is None:
rpn_accuracy_rpn_monitor.append(0)
rpn_accuracy_for_epoch.append(0)
continue
img = cv2.imread(filepath)
X, ratio_h, ratio_w = format_img(img, C)
X = np.transpose(X, (0, 2, 3, 1))
# get output layer Y1, Y2 from the RPN and the feature maps F
# Y1: y_rpn_cls
# Y2: y_rpn_regr
# for layer in model_rpn.layers:
# print(layer.get_config(), layer.get_weights())
[Y1, Y2, F] = model_rpn.predict(X)
# print(Y1.shape)
# Get bboxes by applying NMS
# R.shape = (300, 4)
R = rpn_to_roi(Y1, Y2, C, 'tf', overlap_thresh=0.7)
print(len(R))
# convert from (x1,y1,x2,y2) to (x,y,w,h)
R[:, 2] -= R[:, 0]
R[:, 3] -= R[:, 1]
# apply the spatial pyramid pooling to the proposed regions
bboxes = {}
probs = {}
for jk in range(R.shape[0] // C.num_rois + 1):
ROIs = np.expand_dims(R[C.num_rois * jk:C.num_rois * (jk + 1), :],
axis=0)
if ROIs.shape[1] == 0:
break