Open jinglin80 opened 8 months ago
Thank you for updating! May you also share the tsne_visualization.py that was used to produce Figure 11?
Hi~Have updated. This implementation is borrowed from MMPretrain.
Thank you! However, when trying to replicate figure 11(e), I got the following error. May you help?
python main_tsne.py --checkpoint CMID_ResNet50_bk_200ep --data-root /datasets/UCMerced_LandUse/
Traceback (most recent call last):
File "official-CMID/Pretrain/main_tsne.py", line 8, in
Hi~It's may be related to version of MMCV. Please make sure the version in your environment is same as the project environment file.
Yes, it looks like an MMCV issue, as shown in the following warning.
import mmcv /opt/conda/lib/python3.7/site-packages/mmcv/init.py:21: UserWarning: On January 1, 2023, MMCV will release v2.0.0, in which it will remove components related to the training process and add a data transformation module. In addition, it will rename the package names mmcv to mmcv-lite and mmcv-full to mmcv. See https://github.com/open-mmlab/mmcv/blob/master/docs/en/compatibility.md for more details. 'On January 1, 2023, MMCV will release v2.0.0, in which it will remove '
Could you share the checkpoint file used to generate Figure 11?
plz refer to the CMID models pretrained on MillionAID in project main page.
I used the full UCM dataset and CMID_ResNet50_200ep.pth from the main page. However, the feat1.png I got does not resemble Figure 11(e). python main_tsne.py --checkpoint CMID_ResNet50_200ep.pth --data-root UCMerced_LandUse/Images
We also tried to reproduce Figure 11 (f) but got the following error message:
Traceback (most recent call last):
File "mmrotate/official-CMID/Pretrain/main_tsne.py", line 145, in
To check the content of config, we add a print statement on line 15 of clas_model.py print(config) Output: config checkpoint: /mmrotate/models/backbones/pretrain/CMID_Swin-B_200ep.pth config: null data_root: UCMerced_LandUse/Images deterministic: false early_exaggeration: 12.0 init: random is_distribute: false layer_ind: 0,1,2,3,4 learning_rate: 200.0 local_rank: 0 max_num_class: 20 n_components: 2 n_iter: 1000 n_iter_without_progress: 300 output: /mnt/in-house/python/Q1/model/full2 perplexity: 30.0 pool_type: specified rank: 0 seed: 0 world_size: 1
What did we miss? Furthermore, we made the following changes:
official-CMID/Pretrain/models/swin_transformer.py
from mmengine.model.weight_init import constant_init, truncnormal, trunc_normal_init
from mmengine.model import BaseModule, ModuleList
from mmengine.utils.dl_utils.parrots_wrapper import _BatchNorm
official-CMID/Pretrain/models/backbone_wrapper.py
from mmengine.model.weight_init import constant_init, truncnormal, trunc_normal_init
official-CMID/Pretrain/utils/tsne_utils.py, line 5
from mmengine.model import BaseModule
official-CMID/Pretrain/models/mim_utils.py, line 12
from mmengine.utils import digit_version from mmengine.utils.dl_utils import TORCH_VERSION
Thank you for the great work!
May you share the code for reproducing the figure 13 (Visualization of feature correspondence) of the paper? We have a bit difficult with it.
Thanks, Jing