vasgaowei

Hi, thank you for your interest in our work. Could you please show me the training log?

Hi, the log seems all right. I have three questions: 1. First you can try our `pretrained` model to test the localization accuracy. 2. May I know the environment, e.g....

The following is my training log: {'BASIC': {'BACKUP_CODES': True, 'BACKUP_LIST': ['lib', 'tools_cam', 'configs'], 'DISP_FREQ': 10, 'GPU_ID': [0], 'NUM_WORKERS': 40, 'ROOT_DIR': './tools_cam/..', 'SAVE_DIR': 'ckpt/CUB/deit_tscam_small_patch16_224_CAM-NORMAL_SEED26_CAM-THR0.1_BS128_2021-07-29-18-30', 'SEED': 26, 'TIME': '2021-07-29-18-30'}, 'CUDNN': {'BENCHMARK': False,...

I didn't fully understand what you mean. But I guess if you want to use **joint_attns** as attention map but not **attn_weights**, you can uncomment line 481 as following picture.

Maybe the learning rate is larger. And you can try learning rate 0.001 which is the same as my settings'.

Make sure you download the right version of pretrained 'vgg16.pth' file. And I give the link of the version of pretrained 'vgg16.pth' file which woks fine.

In network.py, modify a line of code in `return_ss_boxes` function. The original code is following: `def return_ss_boxes(self, boxes_index, mode='TRAIN'): if mode == 'TEST': return boxes_index box_num = min(500, len(boxes_index)) indexes...

I'm sorry, i will update a new vision to fix many existed problems. And as for the problem you met, i suggest use version tensorboardX 1.4 and tensorboard 1.11.0.

I have another project [COCO](https://github.com/vasgaowei/D-MIL.pytorch) which can run coco datasets. You can check the [COCO](https://github.com/vasgaowei/D-MIL.pytorch/tree/D-MIL-COCO.pytorch) branch. And it's simple to modify codes for running **MELM**. And if you have any...

The number of **Region Proposals** used in training can affect the detection performance. So you may try adjust the number of proposals. It can be done by modifying the code...