Method encode_text

clip/model.py:343–356

(self, text)

        return self.visual(image.type(self.dtype))

    def encode_text(self, text):
        x = self.token_embedding(text).type(self.dtype)  # [batch_size, n_ctx, d_model]

        x = x + self.positional_embedding.type(self.dtype)
        x = x.permute(1, 0, 2)  # NLD -> LND
        x = self.transformer(x)
        x = x.permute(1, 0, 2)  # LND -> NLD
        x = self.ln_final(x).type(self.dtype)

        # x.shape = [batch_size, n_ctx, transformer.width]
        # take features from the eot embedding (eot_token is the highest number in each sequence)
        x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection

        return x

    def forward(self, image, text):
        image_features = self.encode_image(image)