undertakerweian
commented
5 years ago Hi, deraltefritz!
you have changing line 58.59.60.119.120.121.122 in "generator.py" at first,
in my case,
i want changing the number of anchors box from 18 into 6, so i have modify
"generator.py" line58.59.60 from len(self.anchors)//3 into len(self.anchors)//1
and line 119.120.121.122 from max_index%3 into max_index%1
afterall,
you have changing "yolo.py"
in line 296 from anchors[12:] into anchors[4:]
in line 234.235.236 from len(anchors)//6 into len(anchors)//2
in line 322 from anchors[6:12] into anchors[2:4]
in line 346 from anchors[:6] into anchors[:2]
look4pritam
Hi,
has anyone ever tried to use more or less anchors than the default (=9)?
Even though "gen_anchors.py" allows for more anchors to be computed, "yolo.py" seems to be hard coded to exactly 9 anchors.
It's clear that the number of anchors must be divisible by 3 (each 'YoloLayer' is passed one third of the anchors). So I tried changing the calls to YoloLayer:
loss_yolo_1 = YoloLayer(anchors_yolo1, #anchors[12:] ...loss_yolo_2 = YoloLayer(anchors_yolo2,#anchors[6:12] ...loss_yolo_3 = YoloLayer(anchors_yolo3,#anchors[:6] ...and made sure that "anchors_yolo1" etc. contain one third of the anchors:anchors = [4, 11, 4, 72, 14, 113, 17, 329, 18, 22, 27, 48, 40, 88, 54, 278, 85, 36, 92, 99, 259, 107, 321, 313]anchors_yolo1 = [85, 36, 92, 99, 259, 107, 321, 313]anchors_yolo2 = [18, 22, 27, 48, 40, 88, 54, 278]anchors_yolo3 = [4, 11, 4, 72, 14, 113, 17, 329]But then I ran into the following error:
InvalidArgumentError: Dimensions must be equal, but are 3 and 4 for 'yolo_layer_1/mul' (op: 'Mul') with input shapes: [?,?,?,3,?], [1,1,1,4,2]. ---> 94 pred_wh = tf.expand_dims(tf.exp(pred_box_wh) * self.anchors / net_factor, 4)I have a hunch that the 3 in this line
\# adjust the shape of the y_predict [batch, grid_h, grid_w, 3, 4+1+nb_class]y_pred = tf.reshape(y_pred, tf.concat([tf.shape(y_pred)[:3], tf.constant([3, -1])], axis=0))relates to the default 3 anchors per layer, and tried changing it to the number of anchors per layer (4 in the example above). But that just leads to further errors:InvalidArgumentError: Dimensions must be equal, but are 3 and 4 for 'yolo_layer_1/add' (op: 'Add') with input shapes: [4,?,?,3,2], [?,?,?,18,4,?]. ---> 66 pred_box_xy = (self.cell_grid[:,:grid_h,:grid_w,:,:] + tf.sigmoid(y_pred[..., :2])) # sigma(t_xy) + c_xyCould anyone tell me what changes are necessary to have an arbitrary (divisible by 3) number of anchors?
Thank you!