search

514flowey / JDet-cobb

The Jittor Implementation of COBB
Apache License 2.0
9 stars 0 forks source link

About code #6

Open UncleNiNi opened 1 month ago

UncleNiNi commented 1 month ago

Hello, I downloaded your code and the DOTA1.5 dataset. Following the documentation, I preprocessed the dataset. However, when running the code, I encountered the following yellow warning messages: Loading config from: configs/cobb/faster_rcnn_cobb_ln_r50_fpn_1x_dota.py [w 0714 16:05:48.414852 76 init.py:1507] load parameter fc.weight failed ... [w 0714 16:05:48.414931 76 init.py:1507] load parameter fc.bias failed ... [w 0714 16:05:48.416049 76 init.py:1526] load total 267 params, 2 failed After these warnings appear, the training process completes instantly and then the evaluation process begins. I'm sure it hasn't actually started training, which is very frustrating. I would be very grateful if you could help me with this!

514flowey commented 1 month ago

It's normal to see the yellow warnings. Besides, faster_rcnn_cobb_ln_r50_fpn_1x_dota.py is for dota1.0, not dota1.5. You may change the config files for this dataset.

UncleNiNi commented 1 month ago

Thank you for your reply! Currently, what I have done is creating a new configuration file for DOTA1.5 in the base folder, copying the contents of the DOTA configuration file, and then modifying the dataset path in the content. I also tried changing the image size parameters to 1024 or 600. After that, I placed the references to the configuration file and the path of schedule_1x at the beginning of the list in configs/cobb/faster_rcnn_cobb_ln_r50_fpn_1x_dota.py. I also tried putting the contents of these two files directly in configs/cobb/faster_rcnn_cobb_ln_r50_fpn_1x_dota.py, but the training program still cannot run. Could you please give me some tips on how to use DOTA1.5 for model training?

514flowey commented 1 month ago

It's possible that the labels.pkl for DOTA1.5 was not built successfully. you can refer to https://github.com/514flowey/JDet-cobb/blob/e552bc1bb35518b993c5d63ced14cde52867f325/python/jdet/data/custom.py#L49C1-L49C56 and check the img_infos. Besides, you may config the test dataset for testing like

    test=dict(
        type="ImageDataset",        
        images_dir='datasets/dotav1_5_test_1024_200_1.0/images',
        dataset_type='DOTA1_5',
        transforms=[