def pre_process(self, images):
mean = config.image_mean.reshape(1, 3, 1, 1).astype(np.float32)
std = config.image_std.reshape(1, 3, 1, 1).astype(np.float32)
mean = mge.tensor(mean).to(images.device)
std = mge.tensor(std).to(images.device)
normed_images = (images - mean) / std
normed_images = get_padded_tensor(normed_images, 64)
return normed_images
def forward(self, inputs):
images = inputs['image']
im_info = inputs['im_info']
gt_boxes = inputs['gt_boxes']
# process the images
normed_images = (images - config.image_mean[None, :, None, None]
) / config.image_std[None, :, None, None]
normed_images = get_padded_tensor(normed_images, 64)
if self.training:
return self._forward_train(normed_images, im_info, gt_boxes)
else:
return self._forward_test(normed_images, im_info)