Open Trevor-Philips-cbd opened 4 months ago
Hi! We actually modified the original CLIP model (in the code) with a controller for image degradation. The controller is retrained and saved as the weight "daclip_ViT-B-32.pt". So you need to load our specific weight for degraded images.
Hi! We actually modified the original CLIP model (in the code) with a controller for image degradation. The controller is retrained and saved as the weight "daclip_ViT-B-32.pt". So you need to load our specific weight for degraded images.
Hello, if I want to retrain your controller, what should I do? I saw in your documentation that I need to load the “laion2b_s34b_b79k” weights, but I couldn’t find the relevant weights. Could you please provide a download link?
The training instructions are in this readme file. Usually, the model weights can be automatically downloaded (or you can manually download it via Huggingface).
The training instructions are in this readme file. Usually, the model weights can be automatically downloaded (or you can manually download it via Huggingface).
Hello, I have loaded the weights for the clip according to the instructions, but there is an error saying that the weights do not match. I have followed the instructions in your project and openclip to download the weights of “laion2b_s34b_b79k” and load them into the model. I am not sure which step I got wrong, and I hope you can give me some guidance.
Can you show your test code? Since we changed the openclip code for image degradation, you need to load the weight with our own da-clip code, as in the evaluate.py script.
Hello, could you explain how the pre-trained model you loaded during the training of DA-Clip was obtained? Because I noticed that the model “laion2b_s34b_b79k” is used in your training command, but in your readme, you are asked to download the “daclip_ViT-B-32.pt” file. What is the difference between these two? According to the description in your paper, the training weights of Clip should be loaded here and remain unchanged throughout the training process. I don’t understand the difference between these two sets of model weights, and I hope you can explain it, thank you.