Object masks inversion using

[greg007-web]

I trained a custom model with 3 objects (A, B, C), and worked well, the predicition is correct. But when I get the maks using pred_masks, sometimes A is the number 0, 1, or 2, depends of the image. This happen for each object. Here is my code:

cfg = get_cfg()
cfg.merge_from_file("./modelo/config.yml")
cfg.MODEL.WEIGHTS = "./modelo/model_0209999.pth" # path for final model
cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.8
predictor = DefaultPredictor(cfg)
original = cv2.imread('foto_02.jpeg')
#original = Image.open('./in/t32 (2).jpg')
#plt.imshow(cv2.cvtColor(original, cv2.COLOR_BGR2RGB))
outputs = predictor(original)
MetadataCatalog.get(cfg.DATASETS.TRAIN[0]).thing_classes = ['A','B','C'] 
v = Visualizer(original[:, :, ::-1],
               metadata=MetadataCatalog.get(cfg.DATASETS.TRAIN[0]),
               #scale=0.8,
               instance_mode=ColorMode.SEGMENTATION)
out = v.draw_instance_predictions(outputs["instances"].to("cpu"))
)


### Taking Masks
mask_A = outputs['instances'].pred_masks.cpu().numpy()[0] ### )Here numpy()[?], sometimes is 0, 1 or 2, depend of the image
mask_B = outputs['instances'].pred_masks.cpu().numpy()[1] 
mask_C = outputs['instances'].pred_masks.cpu().numpy()[2]

Again, the final image with predictions, the name of detected objects are Ok, the problem is when I use pred_masks. I tried to remove MetadataCatalog.get, but the problem is the same. Is there any way to keep the same pattern? Or is a BUG?