A repository for organizing contributions related to finetuning and related approaches using OpenAI's CLIP for the Dictionary Augmented Translations project in C4GT'23.
Run main.py with the proper arguments. Here are some details and the various options related to the arguments:
--method: either "base" (without training) or "coop" or "cocoop" (these are based on https://github.com/KaiyangZhou/CoOp).--dataset-dir: a directory to read the dataset from. It should be in standard imagefolder format. Implementation of a dataset loader is scheduled for later.--device: the device used for training the models--clip-model: the clip model used. For example, "RN50x16" or "ViT-B/16". You can use any of the officially supported OpenAI-CLIP models.--batch-size: the batch size used by the dataloader.--epochs: number of epochs.--lr: the learning rate for finetuning or training.--n-ctx: the number of learnable context vectors. Refer to the CoOp paper / repository.--checkpoint-dir: the directory to which the models are saved to every epoch. They are saved as "checkpoint-dir/0.pt", "checkpoint-dir/1.pt"... and so on.