def gen_layout(self, bbox_sampled, label_original, opt):
# crop canvas
input_dict = crop_canvas(bbox_sampled, label_original, opt)
# generate layout
label_generated = self.G_box2mask.evaluate(
{
'label_map':
Variable(input_dict['label'], volatile=True),
'mask_ctx_in':
Variable(input_dict['mask_ctx_in'], volatile=True),
'mask_out':
Variable(input_dict['mask_out'], volatile=True),
'mask_in':
Variable(input_dict['mask_in'], volatile=True),
'cls':
Variable(input_dict['cls'], volatile=True),
'label_map_orig':
Variable(input_dict['label_orig'], volatile=True),
'mask_ctx_in_orig':
Variable(input_dict['mask_ctx_in_orig'], volatile=True),
'mask_out_orig':
Variable(input_dict['mask_out_orig'], volatile=True)
},
target_size=(input_dict['label_orig'].size()[2:4]))
# paste canvas
label_canvas = paste_canvas(label_original, label_generated.data, \
input_dict, resize=False)
return label_canvas, input_dict, label_generated.data
def gen_image(self, bbox_sampled, img_original, label_generated, opt):
# crop canvas
input_dict = crop_canvas(bbox_sampled, label_generated, opt, \
img_original=img_original, transform_img=True)
# generate layout
img_generated = self.G_mask2img.inference(
Variable(input_dict['label'], volatile=True),
Variable(torch.zeros_like(input_dict['label']), volatile=True),
Variable(input_dict['image'], volatile=True),
Variable(input_dict['mask_in'], volatile=True),
Variable(input_dict['mask_out'], volatile=True))
# paste canvas
img_canvas = paste_canvas(img_original, (img_generated.data+1)/2, \
input_dict, method=Image.BICUBIC, is_img=True)
return img_canvas, input_dict, img_generated.data