clip/clip.py

# Copyright 2021 Zilliz. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import sys
from pathlib import Path
import torch
from torchvision import transforms

from towhee.types.image_utils import to_pil
from towhee.operator.base import NNOperator, OperatorFlag
from towhee.types.arg import arg, to_image_color
from towhee import register
from transformers import CLIPTokenizer, CLIPTextModel ,CLIPModel,CLIPProcessor
from train_clip_with_hf_trainer import train_with_hf_trainer


@register(output_schema=['vec'])
class Clip(NNOperator): 
    """
    CLIP multi-modal embedding operator
    """
    def __init__(self, model_name: str, modality: str):
        self.modality = modality
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        cfg = self._configs()[model_name]
        self.model = CLIPModel.from_pretrained(cfg)
        self.tokenizer = CLIPTokenizer.from_pretrained(cfg)
        self.processor =  CLIPProcessor.from_pretrained(cfg)

    def inference_single_data(self, data):
        if self.modality == 'image':
            vec = self._inference_from_image(data)
        elif self.modality == 'text': 
            vec = self._inference_from_text(data)
        else:
            raise ValueError("modality[{}] not implemented.".format(self._modality))
        return vec.detach().cpu().numpy().flatten()

    def __call__(self, data):
        if not isinstance(data, list):
            data = [data]
        else:
            data = data
        results = []
        for single_data in data:
            result = self.inference_single_data(single_data)
            results.append(result)
        if len(data) == 1:        
            return results[0]
        else:
            return results

    def _inference_from_text(self, text):
        tokens = self.tokenizer([text], padding=True, return_tensors="pt")
        text_features = self.model.get_text_features(**tokens)
        return text_features

    @arg(1, to_image_color('RGB'))
    def _inference_from_image(self, img):
        img = to_pil(img)
        inputs = processor(images=img, return_tensors="pt")
        image_features = self.model.get_image_features(**inputs)
        return image_features

    def train(self, **kwargs):
        data_args = kwargs.pop('data_args', None)
        training_args = kwargs.pop('training_args', None)
        train_with_hf_trainer(self.model, self.tokenizer, data_args, training_args)
    
    def _configs(self):
        config = {}
        config['clip_vit_base_32'] = 'openai/clip-vit-base-patch16'
        config['clip_vit_base_16'] = 'openai/clip-vit-base-patch32'
        config['clip_vit_large_14'] = 'openai/clip-vit-large-patch14'
        config['clip_vit_large_14_336'] ='openai/clip-vit-large-patch14-336'
        return config