Question about `align_corners=True`

Hi, dear authors! Sorry to bother you. I have a question about align_corners=True I want to set align_corners=True to improve Dice metric，but when I set align_corners=True, the following warning always appears：

train: UserWarning: When align_corners=True, the output would more aligned if input size (80, 80) is x+1 and out size (320, 320) is nx+1

val: UserWarning: When align_corners=True, the output would more aligned if input size (104, 104) is x+1 and out size (416, 416) is nx+1

Here is my config about train_pipeline and test_pipeline:

train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations'),
    dict(type='Resize',
         img_scale=[(320, 320), (416, 416)],
         keep_ratio=True),
    dict(type='RandomCrop',
         crop_size=crop_size,
         cat_max_ratio=0.9),
    dict(type='RandomFlip', prob=0.5),
    dict(
        type='Albu',
        transforms=albu_train_transforms,
        keymap=dict(img='image', gt_semantic_seg='mask'),
        update_pad_shape=False),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
]
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=(416, 416),
        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=True),
            dict(type='RandomFlip'),
            dict(type='Normalize', **img_norm_cfg),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img']),
        ])
]

Could you tell me how to usealign_corners=True correctly? How to fix this config up? Thank you!

[HarborYuan]

If you can read Chinese, this link may help you: https://zhuanlan.zhihu.com/p/87572724 In short, use align_corners=True when you rescale image from x+1 to nx+1, and use align_corners=False when you rescale image from x to nx.

[HarborYuan]

I am not an expert on this, it seems that you also need to ensure that the original image is with odd number of resolution. I usally use align_coners=False with even number of resolution. Some papers, e.g. max-deeplab will use 1025x1025 and 641x641 as the resolution. You may need to check their implementations to figure it out. This will emprically improve the performance in dense predictions.