我想自己訓練自己的資料集
參考github上的教學，我已經跑完diffusion model,guided classifier
但我進行Strength-controllable Diffusion Anomaly Synthesis時
程式可以正常執行完成
torchrun --nproc_per_node=1 sample.py
sample wheel: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████| 16/16 [03:37<00:00, 13.60s/it]
[rank0]:[W1205 18:00:08.749365132 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
但我去查看我程式執行的結果，會是一張雜訊的照片（參考MVTec-AD/sdas/的結果，應該是正常的照片）
想請問這是哪裡有問題？
我的diffusion model,guided classifier跑出來的checkpoint副檔名都是.pth.tar
我有修改sample.yaml的內容
如：pertrain_path: experiments/DATA/wheel_diffusion_checkpoints/ckpt_1000.pth.tar
pertrain_path: experiments/DATA/wheel_classifier_checkpoints/ckpt_1000.pth.tar
這是我執行train_diffusion.py結果，請問這樣是對的嗎？
這是我執行train_classifier.py結果，請問這樣是對的嗎？
I want to train my own dataset.
Following the instructions on the GitHub repository, I have already completed the diffusion model and guided classifier training steps.
However, when I run Strength-controllable Diffusion Anomaly Synthesis using:：torchrun --nproc_per_node=1 sample.py
the program finishes without any errors:
sample wheel: 100%|███████████████████████████████████████████| 16/16 [03:37<00:00, 13.60s/it]
[rank0]:[W1205 18:00:08.749365132 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit...
But the output images are ust noisy images
(According to the reference output in MVTec-AD/sdas/, they should be normal-looking images instead of pure noise.)
I would like to know what might be causing this issue.
My diffusion model and classifier checkpoints were saved as .pth.tar
So I modified my sample.yaml accordingly, for example:
like：
pertrain_path: experiments/DATA/wheel_diffusion_checkpoints/ckpt_1000.pth.tar
pertrain_path: experiments/DATA/wheel_classifier_checkpoints/ckpt_1000.pth.tar
Here are the results from running train_diffusion.py. Is this correct?
Here are the results from running train_classifier.py. Is this correct?"
Is this the correct way to load these checkpoints, or could this be the reason why my results become noise?
我想自己訓練自己的資料集
參考github上的教學，我已經跑完diffusion model,guided classifier
但我進行Strength-controllable Diffusion Anomaly Synthesis時
程式可以正常執行完成
torchrun --nproc_per_node=1 sample.py
sample wheel: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████| 16/16 [03:37<00:00, 13.60s/it]
[rank0]:[W1205 18:00:08.749365132 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
但我去查看我程式執行的結果，會是一張雜訊的照片（參考MVTec-AD/sdas/的結果，應該是正常的照片）
想請問這是哪裡有問題？
我的diffusion model,guided classifier跑出來的checkpoint副檔名都是.pth.tar
我有修改sample.yaml的內容
如：pertrain_path: experiments/DATA/wheel_diffusion_checkpoints/ckpt_1000.pth.tar
pertrain_path: experiments/DATA/wheel_classifier_checkpoints/ckpt_1000.pth.tar
這是我執行train_diffusion.py結果，請問這樣是對的嗎？
這是我執行train_classifier.py結果，請問這樣是對的嗎？
I want to train my own dataset.
Following the instructions on the GitHub repository, I have already completed the diffusion model and guided classifier training steps.
However, when I run Strength-controllable Diffusion Anomaly Synthesis using:：torchrun --nproc_per_node=1 sample.py
the program finishes without any errors:
sample wheel: 100%|███████████████████████████████████████████| 16/16 [03:37<00:00, 13.60s/it]
[rank0]:[W1205 18:00:08.749365132 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit...
But the output images are ust noisy images
(According to the reference output in MVTec-AD/sdas/, they should be normal-looking images instead of pure noise.)
I would like to know what might be causing this issue.
My diffusion model and classifier checkpoints were saved as .pth.tar
So I modified my sample.yaml accordingly, for example:
like：
pertrain_path: experiments/DATA/wheel_diffusion_checkpoints/ckpt_1000.pth.tar
pertrain_path: experiments/DATA/wheel_classifier_checkpoints/ckpt_1000.pth.tar
Here are the results from running train_diffusion.py. Is this correct?
Here are the results from running train_classifier.py. Is this correct?"
Is this the correct way to load these checkpoints, or could this be the reason why my results become noise?