val_generator

[truc-h-nguyen]

@nickvazz I finally get the model to work! I changed the activation for the last layer to sigmoid(so the range of values could be between 0 and 1 as you mentioned). I also tried "leakyrelu" in stead of "relu"`. The results look potential but I'm not sure if it goes well with IoU which needs positive values. I updated IoU metric with new code from here.

I had a trouble with val_generator. I got different images for train_imgs[14] and val_imgs[14], but train_generator[14] and val_generator[14] show the same picture.

And I believe the mask the model predicts is the combination of multiple objects.

[nickvazz]

I think the IOU will get values from the sigmoid, so that should be okay

[nickvazz]

Are the val/train generator being fed the same images with the same seed?

[truc-h-nguyen]

Yes, this is how I set a side a validation images:

val_samples = 60
train_imgs = coco_imgs[:-val_samples]
train_masks = coco_masks[:-val_samples]
val_imgs = coco_imgs[-val_samples:]
val_masks = coco_masks[-val_samples:]

Then fed the images to val/train generator:

train_generator = DataGenerator(input_img = train_imgs, input_mask = train_masks, image_size = img_size,
                                augmentation=True, batch_size = 5)
val_generator = DataGenerator(input_img = val_imgs, input_mask = val_masks, image_size = img_size,
                                augmentation=True, batch_size = 5)

train_imgs[14], train_generator[14], val_generator[14] show the same image, while val_imgs[14] shows a different image.

I had the seed before but dropped it because I think it's not necessary. Regardless the seed, train_generator[14], val_generator[14] still show the same image.