search

Zasder3 / train-CLIP

A PyTorch Lightning solution to training OpenAI's CLIP from scratch.
MIT License
653 stars 78 forks source link

ModuleNotFoundError: No module named 'clip' #4

Closed Jiang15 closed 3 years ago

Jiang15 commented 3 years ago

Thanks for your great work! when I call train.py with my dataset. I have ModuleNotFoundError: No module named 'clip'. Do I miss something?

Zasder3 commented 3 years ago

Sorry I haven't made a file stating the dependencies; it requires the original CLIP repository's pip installation protocol. Simply enter the following into your terminal and it will install the clip module:

$ pip install git+https://github.com/openai/CLIP.git
Jiang15 commented 3 years ago

Thank you very much! Yes I figured out after I found the openai clip repo and made it run. I trained on 9000 image text pairs(each image has 5 pieces of texts) but I got accuracy closed to 0 in first five epochs:/-------- Original message --------From: Cade Gordon @.>Date: Tue, Jun 1, 2021, 9:27 PMTo: Zasder3/train-CLIP @.>Cc: Wei Jiang @.>, Author @.>Subject: Re: [Zasder3/train-CLIP] ModuleNotFoundError: No module named 'clip' (#4) Sorry I haven't made a file stating the dependencies, it requires the original CLIP repositories pip installation protocol. Simply enter the following into your terminal and it will install the clip module: $ pip install git+https://github.com/openai/CLIP.git

—You are receiving this because you authored the thread.Reply to this email directly, view it on GitHub, or unsubscribe.

Zasder3 commented 3 years ago

This is an important point. Our repo starts from random inits and CLIP used 400M text image pairs. They don't even state what their accuracy is during a batch... One thing that could be helpful is using a pretrained initialization and see what happens from there. I'll make sure I double check to ensure the accuracy is correct though.

Jiang15 commented 3 years ago

Thanks! I went through the paper and find the accuracy curve. I think it is reasonable for almost 0 accuracy :D [image: image.png]

On Tue, Jun 1, 2021 at 11:59 PM Cade Gordon @.***> wrote:

This is an important point. Our repo starts from random inits and CLIP used 400M text image pairs. They don't even state what their accuracy is during a batch... One thing that could be helpful is using a pretrained initialization and see what happens from there. I'll make sure I double check to ensure the accuracy is correct though.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/Zasder3/train-CLIP/issues/4#issuecomment-852477529, or unsubscribe https://github.com/notifications/unsubscribe-auth/AKBQHPNV6IF3DCM2FR4GOD3TQVJ4JANCNFSM455DU4GA .

Zasder3 commented 3 years ago

Looks like the accuracy calculation was incorrect! I'll update the repo.

Jiang15 commented 3 years ago

Hi, many thanks for the update! I trained my model and now I get accuracy up to 85%. But when I load the model by suing the function “build_model”. I encounter another error showed in picture. Do you know what is the issue of it?

Zasder3 commented 3 years ago

Hi @Jiang15, the photo you sent over sadly isn't visible. Try using the GitHub website or app to insert it.

Although I can't see your error I do have some useful feedback on reloading the weights. Lightning requires you to actually use the checkpoints that the training run spits out. There's likely folder called lightning_Logs that will have your model saved. Checkout the documentation here and use the class CLIPWrapper to receive the checkpoint.

Jiang15 commented 3 years ago
Hi, What I tried is to use CLIPWrapper.load_from_checkpoint(myPath.ckpt). But it shows to me that “_init__() missing 3 required positional arguments: 'model_name', 'config', and 'minibatch_size'”. I went through the internet and I guess the hyperparameter is not saved?Have a nice day!  Sent from Mail for Windows 10 From: Cade GordonSent: Sunday, 6 June 2021 12:39 amTo: Zasder3/train-CLIPCc: Wei Jiang; MentionSubject: Re: [Zasder3/train-CLIP] ModuleNotFoundError: No module named 'clip' (#4) Hi @Jiang15, the photo you sent over sadly isn't visible. Try using the GitHub website or app to insert it.Although I can't see your error I do have some useful feedback on reloading the weights. Lightning requires you to actually use the checkpoints that the training run spits out. There's likely folder called lightning_Logs that will have your model saved. Checkout the documentation here and use the class CLIPWrapper to receive the checkpoint.—You are receiving this because you were mentioned.Reply to this email directly, view it on GitHub, or unsubscribe.