search

HuiZhang0812 / DiffusionAD

148 stars 16 forks source link

Why is recon_con still a noisy image? #39

Closed boxbox2 closed 7 months ago

boxbox2 commented 7 months ago

image

HuiZhang0812 commented 7 months ago

May I ask how many epochs did you train for?

boxbox2 commented 7 months ago

May I ask how many epochs did you train for?

3000 train epoch I use the args1 argument you gave but training_dataset_loader i set batch_size = 6,nums_workers = 8 and I trained only the carpet categories in the mvtec dataset { "img_size": [256,256], "Batch_Size": 6, "EPOCHS": 3000, "T": 1000, "base_channels": 128, "beta_schedule": "linear", "loss_type": "l2", "diffusion_lr": 1e-4, "seg_lr": 1e-5, "random_slice": true, "weight_decay": 0.0, "save_imgs":true, "save_vids":false, "dropout":0, "attention_resolutions":"32,16,8", "num_heads":4, "num_head_channels":-1, "noise_fn":"gauss", "channels":3, "mvtec_root_path":"datasets/mvtec", "visa_root_path":"datasets/VisA/visa", "dagm_root_path":"datasets/dagm", "mpdd_root_path":"datasets/mpdd", "anomaly_source_path":"datasets/dtd", "noisier_t_range":600, "less_t_range":300, "condition_w":1, "eval_normal_t":200, "eval_noisier_t":400, "output_path":"outputs"

}

HuiZhang0812 commented 7 months ago

Which "checkpoint_type" are you using in eval.py?

boxbox2 commented 7 months ago

Which "checkpoint_type" are you using in eval.py?

best

HuiZhang0812 commented 7 months ago

Try it with "checkpoint_type=last".

boxbox2 commented 7 months ago

Try it with "checkpoint_type=last".

image It's the same as the original

HuiZhang0812 commented 7 months ago

Have you recorded the loss curve?

boxbox2 commented 7 months ago

Have you recorded the loss curve?

I haven't done anything special. I just ran the cloned code step by step. After running train.py, I ran eval.py

HuiZhang0812 commented 7 months ago

I will run your config to verify, thank you for your patience. By the way, I found that the results in #40 were as expected.

azadfatemeh commented 7 months ago

I also have the same problem, did you find any solution for that?

boxbox2 commented 7 months ago

I also have the same problem, did you find any solution for that?

not yet

HuiZhang0812 commented 7 months ago

Thank you for your patience. We used the same configuration as yours, and our reconstruction and segmentation results were accurate. thread_000

May I ask if you are using the latest code? We recently released an update that fixes the issue of loss becoming "nan" when the batch size is small. Could you please try our latest code again?

boxbox2 commented 7 months ago

Thank you for your patience. We used the same configuration as yours, and our reconstruction and segmentation results were accurate. thread_000

May I ask if you are using the latest code? We recently released an update that fixes the issue of loss becoming "nan" when the batch size is small. Could you please try our latest code again?

Thank you for your reply. I did indeed use skipping method to handle NaN values before. I will re-download your updated code for experimentation. Thanks again for your patient response.

azadfatemeh commented 7 months ago

Thank you for your patience. We used the same configuration as yours, and our reconstruction and segmentation results were accurate. thread_000

May I ask if you are using the latest code? We recently released an update that fixes the issue of loss becoming "nan" when the batch size is small. Could you please try our latest code again?

Thanks for your prompt reply. My problem is solved now. Since I did not use the dataset you provided, I had some mistakes in normalization that may lead to this problem.