Can you guide me for creating self-supervised .ckpt file for custom dataset?

[saifhassan-phd]

Thank you @XzwHan @JegZheng for this wonderful contribution, I have implemented this and want to use this model for my research project. So can you guide me for creating self-supervised .ckpt file for custom dataset (Any image classification dataset)?

I have gone through paper and code but unable to understand how you have developed .ckpt file for self-supervised model.

Thank you in advance.

[JegZheng]

Thank you @XzwHan @JegZheng for this wonderful contribution, I have implemented this and want to use this model for my research project. So can you guide me for creating self-supervised .ckpt file for custom dataset (Any image classification dataset)?

I have gone through paper and code but unable to understand how you have developed .ckpt file for self-supervised model.

Thank you in advance.

Hi @saifhassan-phd ,

Our self-supervised model is adapted from our self-supervised learning work CACR. Please feel free to check this repo, where you can find how to use CACR to learn the self-supervised model on your dataset: https://github.com/JegZheng/CACR-SSL

[saifhassan-phd]

Thank you @XzwHan @JegZheng for this wonderful contribution, I have implemented this and want to use this model for my research project. So can you guide me for creating self-supervised .ckpt file for custom dataset (Any image classification dataset)?
I have gone through paper and code but unable to understand how you have developed .ckpt file for self-supervised model.
Thank you in advance.

Hi @saifhassan-phd ,

Our self-supervised model is adapted from our self-supervised learning work CACR. Please feel free to check this repo, where you can find how to use CACR to learn the self-supervised model on your dataset: https://github.com/JegZheng/CACR-SSL

ok, Thanks for your quick response.

[saifhassan-phd]

Thank you @XzwHan @JegZheng for this wonderful contribution, I have implemented this and want to use this model for my research project. So can you guide me for creating self-supervised .ckpt file for custom dataset (Any image classification dataset)?
I have gone through paper and code but unable to understand how you have developed .ckpt file for self-supervised model.
Thank you in advance.

Hi @saifhassan-phd ,

Our self-supervised model is adapted from our self-supervised learning work CACR. Please feel free to check this repo, where you can find how to use CACR to learn the self-supervised model on your dataset: https://github.com/JegZheng/CACR-SSL

@JegZheng @XzwHan , I have gone through https://github.com/JegZheng/CACR-SSL repo and completed cifar10 experiment and it generated encoder.pth file using main_ours.py file and also tested with linear.py file, now can you guide me regarding how can I use this encoder.pth with card model? Because in cifar10.yml config file, it mentions .ckpt file.

Thanks

[saifhassan-phd]

Thank you @XzwHan @JegZheng for this wonderful contribution, I have implemented this and want to use this model for my research project. So can you guide me for creating self-supervised .ckpt file for custom dataset (Any image classification dataset)?
I have gone through paper and code but unable to understand how you have developed .ckpt file for self-supervised model.
Thank you in advance.

Hi @saifhassan-phd ,
Our self-supervised model is adapted from our self-supervised learning work CACR. Please feel free to check this repo, where you can find how to use CACR to learn the self-supervised model on your dataset: https://github.com/JegZheng/CACR-SSL

@JegZheng @XzwHan , I have gone through https://github.com/JegZheng/CACR-SSL repo and completed cifar10 experiment and it generated encoder.pth file using main_ours.py file and also tested with linear.py file, now can you guide me regarding how can I use this encoder.pth with card model? Because in cifar10.yml config file, it mentions .ckpt file.

Thanks

Specifically, can you guide, how you generate snapshot_supervised.ckpt and snapshot_self_supervised.ckpt files, I tried using https://github.com/JegZheng/CACR-SSL, but it produced .pth file?

[JegZheng]

Thank you @XzwHan @JegZheng for this wonderful contribution, I have implemented this and want to use this model for my research project. So can you guide me for creating self-supervised .ckpt file for custom dataset (Any image classification dataset)?
I have gone through paper and code but unable to understand how you have developed .ckpt file for self-supervised model.
Thank you in advance.

Hi @saifhassan-phd ,
Our self-supervised model is adapted from our self-supervised learning work CACR. Please feel free to check this repo, where you can find how to use CACR to learn the self-supervised model on your dataset: https://github.com/JegZheng/CACR-SSL

@JegZheng @XzwHan , I have gone through https://github.com/JegZheng/CACR-SSL repo and completed cifar10 experiment and it generated encoder.pth file using main_ours.py file and also tested with linear.py file, now can you guide me regarding how can I use this encoder.pth with card model? Because in cifar10.yml config file, it mentions .ckpt file.
Thanks

Specifically, can you guide, how you generate snapshot_supervised.ckpt and snapshot_self_supervised.ckpt files, I tried using https://github.com/JegZheng/CACR-SSL, but it produced .pth file?

We have provided the proxy function to load .pth file. Please check our code here (CARD/classification/card_classification.py, L208): https://github.com/XzwHan/CARD/blob/ebe2f2ae95dc7a7a95e6a71c0c8e1cabf8451087/classification/card_classification.py#L208C4-L208C5

Just put the .pth file in the corresponding folder and rename it as aux_ckpt.pth and it shall be ready.
Alternatively, you may create a loading function with the following step:

1. Call torch.load() to load the .pth file (the .pth file from CACR repo is already a state_dict class)
2. Call model.load_state_dict() to paste the weight to the model.

[saifhassan-phd]

Thank you @XzwHan @JegZheng for this wonderful contribution, I have implemented this and want to use this model for my research project. So can you guide me for creating self-supervised .ckpt file for custom dataset (Any image classification dataset)?
I have gone through paper and code but unable to understand how you have developed .ckpt file for self-supervised model.
Thank you in advance.

Hi @saifhassan-phd ,
Our self-supervised model is adapted from our self-supervised learning work CACR. Please feel free to check this repo, where you can find how to use CACR to learn the self-supervised model on your dataset: https://github.com/JegZheng/CACR-SSL

@JegZheng @XzwHan , I have gone through https://github.com/JegZheng/CACR-SSL repo and completed cifar10 experiment and it generated encoder.pth file using main_ours.py file and also tested with linear.py file, now can you guide me regarding how can I use this encoder.pth with card model? Because in cifar10.yml config file, it mentions .ckpt file.
Thanks

Specifically, can you guide, how you generate snapshot_supervised.ckpt and snapshot_self_supervised.ckpt files, I tried using https://github.com/JegZheng/CACR-SSL, but it produced .pth file?

We have provided the proxy function to load .pth file. Please check our code here (CARD/classification/card_classification.py, L208): https://github.com/XzwHan/CARD/blob/ebe2f2ae95dc7a7a95e6a71c0c8e1cabf8451087/classification/card_classification.py#L208C4-L208C5

Just put the .pth file in the corresponding folder and rename it as aux_ckpt.pth and it shall be ready. Alternatively, you may create a loading function with the following step:

1. Call torch.load() to load the .pth file (the .pth file from CACR repo is already a state_dict class)
2. Call model.load_state_dict() to paste the weight to the model.

I didn't understand, Let me rephrase, I want to know how you produced snapshot_self_supervised.ckpt file?

[saifhassan-phd]

I mean how can I generate these .ckpt files for cifar-10 dataset

[saifhassan-phd]

Thank you @XzwHan @JegZheng for this wonderful contribution, I have implemented this and want to use this model for my research project. So can you guide me for creating self-supervised .ckpt file for custom dataset (Any image classification dataset)?
I have gone through paper and code but unable to understand how you have developed .ckpt file for self-supervised model.
Thank you in advance.

Hi @saifhassan-phd ,
Our self-supervised model is adapted from our self-supervised learning work CACR. Please feel free to check this repo, where you can find how to use CACR to learn the self-supervised model on your dataset: https://github.com/JegZheng/CACR-SSL

@JegZheng @XzwHan , I have gone through https://github.com/JegZheng/CACR-SSL repo and completed cifar10 experiment and it generated encoder.pth file using main_ours.py file and also tested with linear.py file, now can you guide me regarding how can I use this encoder.pth with card model? Because in cifar10.yml config file, it mentions .ckpt file.
Thanks

Specifically, can you guide, how you generate snapshot_supervised.ckpt and snapshot_self_supervised.ckpt files, I tried using https://github.com/JegZheng/CACR-SSL, but it produced .pth file?

We have provided the proxy function to load .pth file. Please check our code here (CARD/classification/card_classification.py, L208): https://github.com/XzwHan/CARD/blob/ebe2f2ae95dc7a7a95e6a71c0c8e1cabf8451087/classification/card_classification.py#L208C4-L208C5

Just put the .pth file in the corresponding folder and rename it as aux_ckpt.pth and it shall be ready. Alternatively, you may create a loading function with the following step:

1. Call torch.load() to load the .pth file (the .pth file from CACR repo is already a state_dict class)
2. Call model.load_state_dict() to paste the weight to the model.

I have tried that proxy function, it causes keyerror on line 210, tried load_state_dict(aux_state[0]), load_state_dict(aux_state['state_dict']) and load_state_dict(aux_state), but again error.

[JegZheng]

The .pth file saved from CACR repo is directly the state_dict, and the .ckpt in CARD is a dictionary, where the state_dict is stored under the key 'state_dict'. This is why:

• In CACR, we directly load the weight to the model as model.load_state_dict(torch.load(file_path))
• In CARD, we load as model.load_state_dict(torch.load(file_path['state_dict'])

Let's do a simpler solution without changing the code. Could you please try:

1. We load the .pth file produced by CACR code
2. Resave this state_dict in a dictionary, and make the key named 'state_dict'.

For example,
model_weight = torch.load('encoder.pth')
weight_dict = {'state_dict': model_weight}
torch.save(weight_dict, 'snapshot_self_supervised.ckpt')

[saifhassan-phd]

The .pth file saved from CACR repo is directly the state_dict, and the .ckpt in CARD is a dictionary, where the state_dict is stored under the key 'state_dict'. This is why:

• In CACR, we directly load the weight to the model as model.load_state_dict(torch.load(file_path))
• In CARD, we load as model.load_state_dict(torch.load(file_path['state_dict'])

Let's do a simpler solution without changing the code. Could you please try:

1. We load the .pth file produced by CACR code
2. Resave this state_dict in a dictionary, and make the key named 'state_dict'.

For example, model_weight = torch.load('encoder.pth') weight_dict = {'state_dict': model_weight} torch.save(weight_dict, 'snapshot_self_supervised.ckpt')

I got it, I was doing mistake by using different backbone while SSL and then in CARD, got it. Error Solved. Thanks for your quick response.

[saifhassan-phd]

Hey @JegZheng @XzwHan

I have successfully implemented small_experiments/cifar10 but could not find code for cifar10 using moco (self-supervised).

How can we pre-train cifar10 using moco and then use it with CARD model?