megatron_patch/model/llava/clip_encoder.py

# Copyright (c) 2023 Alibaba PAI and Nvidia Megatron-LM Team.
#
# 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 torch
import torch.nn as nn

from transformers import CLIPVisionModel, CLIPImageProcessor, CLIPVisionConfig

import warnings
try:
    import cvcuda
    from megatron_patch.data.llava.cvcuda_image_processing_clip import CLIPCVCUDAImageProcessor
    warnings.warn("The cvcuda environment exists, use the cvcuda operator for preprocessing")
except:
    warnings.warn("The cvcuda environment does not exist. Install cvcuda and use it")


class CLIPVisionTower(nn.Module):
    def __init__(self, vision_tower,
                 cvcuda_image_processing=False,
                 mm_vision_select_layer=-2,
                 mm_vision_select_feature='patch',
                 delay_load=False):
        super().__init__()

        self.is_loaded = False

        self.vision_tower_name = vision_tower
        self.cvcuda_image_processing = cvcuda_image_processing
        self.select_layer = mm_vision_select_layer
        self.select_feature = mm_vision_select_feature

        if not delay_load:
            self.load_model()
        else:
            self.cfg_only = CLIPVisionConfig.from_pretrained(self.vision_tower_name)

    def load_model(self):
        if self.cvcuda_image_processing:
            self.image_processor = CLIPCVCUDAImageProcessor.from_pretrained(self.vision_tower_name)
        else:
            self.image_processor = CLIPImageProcessor.from_pretrained(self.vision_tower_name)
        self.vision_tower = CLIPVisionModel.from_pretrained(self.vision_tower_name)

        self.is_loaded = True

    def feature_select(self, image_forward_outs):
        image_features = image_forward_outs.hidden_states[self.select_layer]
        if self.select_feature == 'patch':
            image_features = image_features[:, 1:]
        elif self.select_feature == 'cls_patch':
            image_features = image_features
        else:
            raise ValueError(f'Unexpected select feature: {self.select_feature}')
        return image_features

    @torch.no_grad()
    def forward(self, images):
        if type(images) is list:
            image_features = []
            for image in images:
                image_forward_out = self.vision_tower(image.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True)
                image_feature = self.feature_select(image_forward_out).to(image.dtype)
                image_features.append(image_feature)
        else:
            image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True)
            image_features = self.feature_select(image_forward_outs).to(images.dtype)

        return image_features

    @property
    def dummy_feature(self):
        return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)

    @property
    def dtype(self):
        return self.vision_tower.dtype

    @property
    def device(self):
        return self.vision_tower.device

    @property
    def config(self):
        if self.is_loaded:
            return self.vision_tower.config
        else:
            return self.cfg_only

    @property
    def hidden_size(self):
        return self.config.hidden_size

    @property
    def num_patches(self):
        return (self.config.image_size // self.config.patch_size) ** 2