search

Haoqing-Wang / LocalMIM

[CVPR 2023 Highlight] Masked Image Modeling with Local Multi-Scale Reconstruction
42 stars 5 forks source link

ADE20k results reproduction #3

Open rayleizhu opened 1 year ago

rayleizhu commented 1 year ago

I cannot reproduce ADE20k performance (49.5 mIoU) shown in Table 2 of the paper.

Using your released checkpoint and semantic segmentation code, I got 48.65 mIoU (see the log in onedrive).

Any suggestions for the reproduction?

Haoqing-Wang commented 1 year ago

I suggest to adjust the learning rate lr from [1e-4, 2e-4, 3e-4, 4e-4, 5e-4], and the layer_decay_rate can also be adjusted. Please check different learning rate first. https://github.com/Haoqing-Wang/LocalMIM/blob/314a6e558a8c47bf0ef146235e2ced8d01e27fe6/semantic_segmentation/configs/upernet_vit_base_12_512_slide_160k_ade20k.py#L23

rayleizhu commented 1 year ago

The current config is consistent with the description in your paper (p13).

image

  1. Can you release the training log? If not,
  2. Can you release the config you used to produce the result in your paper? If not,
  3. Can you recall the critical hyperparameters you actually used?