Understanding ops under reshaping scope

[mrgloom]

Under reshaping scope:

https://github.com/TropComplique/FaceBoxes-tensorflow/blob/545ec4f4f3c55c3592ee189ed56a11a3fd017194/src/detector.py#L286

Why is 2 reshapes is needed here:

https://github.com/TropComplique/FaceBoxes-tensorflow/blob/545ec4f4f3c55c3592ee189ed56a11a3fd017194/src/detector.py#L297

And why list is stacked via tf.stack?

y = tf.reshape(y, tf.stack([batch_size, height_i, width_i, num_predictions_per_location, 4]))

And here: https://github.com/TropComplique/FaceBoxes-tensorflow/blob/545ec4f4f3c55c3592ee189ed56a11a3fd017194/src/detector.py#L301

Order of with tf.stack and without is different.

[mrgloom]

It looks like it can be done in single step, tested in numpy:

    box_encoding_0, box_encoding_1, box_encoding_2, class_0, class_1, class_2 = outputs
    
    batch_size = 1
    num_predictions_per_location = [21, 1, 1]
    
    # V1 :
    print('box_encoding_0.shape', box_encoding_0.shape)
    _, height_i, width_i, _ = box_encoding_0.shape
    box_encoding_0_1 = box_encoding_0.reshape(batch_size, height_i, width_i, num_predictions_per_location[0], 4)
    print('box_encoding_0_1.shape', box_encoding_0_1.shape)
    num_anchors_on_feature_map = height_i * width_i * num_predictions_per_location[0]
    box_encoding_0_1 = box_encoding_0_1.reshape(batch_size, num_anchors_on_feature_map, 4)
    print('box_encoding_0_1.shape', box_encoding_0_1.shape)
    
    # V2 : 
    box_encoding_0_2 = box_encoding_0.reshape(batch_size, num_anchors_on_feature_map, 4)
    print('box_encoding_0_2.shape', box_encoding_0_2.shape)
    
    print('diff', np.max(np.abs(box_encoding_0_1-box_encoding_0_2)))