modules/language/fgclip.py

import torch.nn as nn
from transformers import SiglipTextModel
from modules.build import LANGUAGE_REGISTRY
import torch

@LANGUAGE_REGISTRY.register()
class FGCLIPLanguageEncoder(nn.Module):
    def __init__(self, cfg, weights="google/siglip-base-patch16-224", max_position_embeddings = 512):
        super().__init__()

        # Load tokenizer and model
        self.model = SiglipTextModel.from_pretrained(weights, max_position_embeddings = max_position_embeddings)

    def forward(self, txt_ids, **kwargs):
        # txt_ids: (B, L)
        caption_input = torch.tensor(txt_ids)
        outputs = self.model(input_ids=txt_ids).last_hidden_state
        return outputs