def _pixel_selector_grad(op, grad): """The gradients for 'pixel_selector'. Args: op: The 'pixel_selector' operation we want to differentiate. grad: Gradient with respect to the output of the 'pixel_selector' op. Returns: Gradients with respect to the coordinates of points of interest for 'pixel_selector'. """ input = op.inputs[0] coord = op.inputs[1] strides = op.inputs[2] coord_grad = ops.zeros_like((NUM_POINTS, 3), tf.float32) back_grad = ops.reshape(grad, [-1]) coord_grad_tmp = np.zeros((NUM_POINTS, 3), np.float32) for i in range(0, NUM_POINTS): for j in range(0, 3): coord_tmp = np.zeros((NUM_POINTS, 3), np.float32) coord_tmp[i, j] = 1.0 coord_tmp = coord + coord_tmp tmp_1 = ops.reshape( select_module.pixel_selector(input, coord_tmp, strides), [-1]) coord_tmp = np.zeros((NUM_POINTS, 3), np.float32) coord_tmp[i, j] = -1.0 coord_tmp = coord + coord_tmp tmp_2 = ops.reshape( select_module.pixel_selector(input, coord_tmp, strides), [-1]) tmp = ops.subtract(tmp_1, tmp_2) tmp = ops.divide(tmp, 2) tmp = ops.multiply(tmp, back_grad) tmp_3 = np.zeros((NUM_POINTS, 3), np.float32) tmp_3[i, j] = 1.0 coord_grad_tmp = coord_grad_tmp + tmp_3 * ops.reduce_sum(tmp) coord_grad = coord_grad_tmp return [None, coord_grad, None]
def calc_loss(logits: tf.Tensor, caps_out: tf.Tensor, x: tf.Tensor, y: tf.Tensor, decoded: tf.Tensor): with tf.variable_scope('calc_loss'): # margin loss 中调节上margin和下margind的权重 lambda_val = 0.5 # 上margin与下margin的参数值 m_plus = 0.95 m_minus = 0.05 max_l = tf.square(tf.maximum(0., m_plus-logits)) max_r = tf.square(tf.maximum(0., logits-m_minus)) margin_loss = tf.reduce_mean(tf.reduce_sum(y * max_l + lambda_val * (1. - y) * max_r, axis=-1)) orgin = tf.reshape(x, (x.shape[0], -1)) reconstruct_loss = 0.0005*tf.reduce_mean(tf.square(orgin-decoded)) total_loss = margin_loss+reconstruct_loss return total_loss
def main(ckpt_weights, image_size, output_size, model_def, class_num, depth_multiplier, obj_thresh, iou_thresh, train_set, test_image): h = Helper(None, class_num, f'data/{train_set}_anchor.npy', np.reshape(np.array(image_size), (-1, 2)), np.reshape(np.array(output_size), (-1, 2))) network = eval(model_def) # type :yolo_mobilev2 yolo_model, yolo_model_warpper = network([image_size[0], image_size[1], 3], len(h.anchors[0]), class_num, alpha=depth_multiplier) yolo_model_warpper.load_weights(str(ckpt_weights)) print(INFO, f' Load CKPT {str(ckpt_weights)}') orig_img = h._read_img(str(test_image)) image_shape = orig_img.shape[0:2] img, _ = h._process_img(orig_img, true_box=None, is_training=False, is_resize=True) """ load images """ img = tf.expand_dims(img, 0) y_pred = yolo_model_warpper.predict(img) """ box list """ _yxyx_box = [] _yxyx_box_scores = [] """ preprocess label """ for l, pred_label in enumerate(y_pred): """ split the label """ pred_xy = pred_label[..., 0:2] pred_wh = pred_label[..., 2:4] pred_confidence = pred_label[..., 4:5] pred_cls = pred_label[..., 5:] # box_scores = obj_score * class_score box_scores = tf.sigmoid(pred_cls) * tf.sigmoid(pred_confidence) # obj_mask = pred_confidence_score[..., 0] > obj_thresh """ reshape box """ # NOTE tf_xywh_to_all will auto use sigmoid function pred_xy_A, pred_wh_A = tf_xywh_to_all(pred_xy, pred_wh, l, h) boxes = correct_box(pred_xy_A, pred_wh_A, image_size, image_shape) boxes = tf.reshape(boxes, (-1, 4)) box_scores = tf.reshape(box_scores, (-1, class_num)) """ append box and scores to global list """ _yxyx_box.append(boxes) _yxyx_box_scores.append(box_scores) yxyx_box = tf.concat(_yxyx_box, axis=0) yxyx_box_scores = tf.concat(_yxyx_box_scores, axis=0) mask = yxyx_box_scores >= obj_thresh """ do nms for every classes""" _boxes = [] _scores = [] _classes = [] for c in range(class_num): class_boxes = tf.boolean_mask(yxyx_box, mask[:, c]) class_box_scores = tf.boolean_mask(yxyx_box_scores[:, c], mask[:, c]) select = tf.image.non_max_suppression(class_boxes, scores=class_box_scores, max_output_size=30, iou_threshold=iou_thresh) class_boxes = tf.gather(class_boxes, select) class_box_scores = tf.gather(class_box_scores, select) _boxes.append(class_boxes) _scores.append(class_box_scores) _classes.append(tf.ones_like(class_box_scores) * c) boxes = tf.concat(_boxes, axis=0) classes = tf.concat(_classes, axis=0) scores = tf.concat(_scores, axis=0) """ draw box """ font = ImageFont.truetype(font='asset/FiraMono-Medium.otf', size=tf.cast( tf.floor(3e-2 * image_shape[0] + 0.5), tf.int32).numpy()) thickness = (image_shape[0] + image_shape[1]) // 300 """ show result """ if len(classes) > 0: pil_img = Image.fromarray(orig_img) print(f'[top\tleft\tbottom\tright\tscore\tclass]') for i, c in enumerate(classes): box = boxes[i] score = scores[i] label = '{:2d} {:.2f}'.format(int(c.numpy()), score.numpy()) draw = ImageDraw.Draw(pil_img) label_size = draw.textsize(label, font) top, left, bottom, right = box print( f'[{top:.1f}\t{left:.1f}\t{bottom:.1f}\t{right:.1f}\t{score:.2f}\t{int(c):2d}]' ) top = max(0, tf.cast(tf.floor(top + 0.5), tf.int32)) left = max(0, tf.cast(tf.floor(left + 0.5), tf.int32)) bottom = min(image_shape[0], tf.cast(tf.floor(bottom + 0.5), tf.int32)) right = min(image_shape[1], tf.cast(tf.floor(right + 0.5), tf.int32)) if top - image_shape[0] >= 0: text_origin = tf.convert_to_tensor([left, top - label_size[1]]) else: text_origin = tf.convert_to_tensor([left, top + 1]) for j in range(thickness): draw.rectangle([left + j, top + j, right - j, bottom - j], outline=h.colormap[c]) draw.rectangle( [tuple(text_origin), tuple(text_origin + label_size)], fill=h.colormap[c]) draw.text(text_origin, label, fill=(0, 0, 0), font=font) del draw pil_img.show() else: print(NOTE, ' no boxes detected')
if __name__ == "__main__": g = tf.get_default_graph() ds, ds_val = mnist_dataset() iterator = ds.make_one_shot_iterator() next_x, next_y = iterator.get_next() batch_x = tf.placeholder_with_default(next_x, shape=[100, 28, 28, 1]) batch_y = tf.placeholder_with_default(next_y, shape=[100, 10]) logits, caps_out = capsnet(batch_x) decoded = decoder(caps_out, batch_y) """ define loss """ loss = calc_loss(logits, caps_out, batch_x, batch_y, decoded) """ define summary """ acc_op, acc = tf.metrics.accuracy(tf.argmax(batch_y, -1), tf.argmax(logits, -1)) tf.summary.scalar('loss', loss) tf.summary.scalar('acc', acc) tf.summary.image('reconstruction_img', tf.reshape(decoded, (100, 28, 28, 1))) summ = tf.summary.merge_all() """ define train op """ steps = tf.train.get_or_create_global_step(g) train_op = tf.train.AdamOptimizer().minimize(loss, global_step=steps) config = tf.ConfigProto() config.gpu_options.allow_growth = True with tf.Session(config=config) as sess: writer = tf.summary.FileWriter('log', g) sess.run([tf.global_variables_initializer(), tf.local_variables_initializer()]) for i in range(10): with tqdm(total=60000//100, bar_format='{n_fmt}/{total_fmt} |{bar}| {rate_fmt}{postfix}', unit=' batch', dynamic_ncols=True) as t: for j in range(60000//100): _, summ_, steps_, loss_, acc_ = sess.run([train_op, summ, steps, loss, acc]) t.set_postfix(loss='{:<5.3f}'.format(loss_), acc='{:<4.2f}%'.format(acc_*100))
def flatten(previous_layer): return tf.reshape(previous_layer, shape=[-1, (previous_layer.get_shape()[1] * previous_layer.get_shape()[2] * previous_layer.get_shape()[3])])