def call(self, x): ''' To compute final output from classification branch and location branch Arguments x: Return tensor ''' assert len(x) == 2, 'OutputLayer must be passed 2 inputs: clzbbe_tensor and roi_tensor' clzbbe_3dtensor = x[0] # (batch_size, num_of_rois, num_of_classes+4), batch_size = 1 num_of_class = clzbbe_3dtensor.shape[2]-4 clz_2dtensor = clzbbe_3dtensor[0, :, :num_of_class] # (num_of_rois, num_of_classes) bbe_2dtensor = clzbbe_3dtensor[0, :, num_of_class:] # (num_of_rois, 4) roi_3dtensor = x[1] # (batch_size, num_of_rois, 4), batch_size = 1 roi_2dtensor = roi_3dtensor[0] # (num_of_rois, 4) pbox_2dtensor = bbe2box2d(box_2dtensor=roi_2dtensor, bbe_2dtensor=bbe_2dtensor) clz_3dtensor = tf.expand_dims(input=clz_2dtensor, axis=2) pclzprob = tf.math.reduce_max(input_tensor=clz_3dtensor, axis=1) pclz = tf.math.argmax(input=clz_3dtensor, axis=1) pclz = tf.cast(pclz, dtype='float32') t = tf.concat(values=[roi_2dtensor, pbox_2dtensor, pclz, pclzprob], axis=1) return t
# predict rois
roi2d = pad_roi2d(roi2d=bbox2d[:, :4],
max_num_of_rois=max_num_of_rois,
pad_roi=[0, 0, ishape[0], ishape[1]])
roi_2dtensor = tf.constant(value=roi2d, dtype='float32')
# predict object proposal
batch_x = [
tf.expand_dims(input=x, axis=0),
tf.expand_dims(input=roi_2dtensor, axis=0)
]
y_pred = detection_model.predict_on_batch(batch_x)
clzbbe_tensor = y_pred[0]
clz_2dtensor = clzbbe_tensor[:, :len(classes)]
bbe_2dtensor = clzbbe_tensor[:, len(classes):]
pred_bbox2d = bbe2box2d(box_2dtensor=roi_2dtensor,
bbe_2dtensor=bbe_2dtensor)
_, ax = plt.subplots(figsize=(15, 7.35))
ax.imshow(x / 255)
# iterate over rois
for k in range(pred_bbox2d.shape[0]):
print(np.array(clz_2dtensor[k]))
# print('{} -> {}'.format(k, classes[clz]))
clz = np.argmax(clz_2dtensor[k])
roi = roi_2dtensor[k]
bbox = pred_bbox2d[k]
rframe = box2frame(box=roi, apoint=[0, 0])