augmentations_clip.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the Apache License, Version 2.0
# found in the LICENSE file in the root directory of this source tree.

# import logging

from torchvision import transforms

import torch 
import cv2
from PIL import Image
import numpy as np

from my_transforms import (
    GaussianBlur,
    make_normalize_transform,
    make_normalize_transform_clip,
)

def add_gaussian_noise(tensor, mean=0.0, std=0.1):
    noise = torch.randn(tensor.size()).cuda() * std + mean
    return tensor + noise




class DataAugmentationCLIP(object):
    def __init__(
        self,
        global_crops_scale,
        local_crops_scale,
        local_crops_number,
        global_crops_size=224,
        local_crops_size=96,
    ):

        self.source_trans = transforms.Compose([
            # transforms.RandomCrop(224),
            # transforms.CenterCrop(224),
            transforms.ToTensor(),
            make_normalize_transform_clip(),
        ])

        # self.crop = transforms.Compose([
        #     transforms.CenterCrop(224),
           
        # ])

        self.crop = transforms.Compose([
    transforms.Resize(224),  # 将短边缩放到 224，长边会按比例缩放
    transforms.RandomCrop(224),  # 然后裁剪到 224x224
])

        self.centercrop = transforms.Compose([
            transforms.CenterCrop(224),
           
        ])

        self.randomcrop = transforms.Compose([
            transforms.RandomCrop(224),
           
        ])

        self.local_crops_number = local_crops_number

    def __call__(self, image):
        output = {}
        output["source"] = []

        if np.array(image).shape[0]<224 or np.array(image).shape[1]<224:
            crops_all = [
             self.centercrop(image) for _ in range(self.local_crops_number)
         ]
        else:
            crops_all = [
                self.centercrop(image) for _ in range(self.local_crops_number)
            ]
            
        for crops_image in crops_all:
            output["source"].append(self.source_trans(crops_image))   #单独使用好一些
    

        output["offsets"] = ()

        return output
    

class DataAugmentationDINO(object):
    def __init__(
        self,
        global_crops_scale,
        local_crops_scale,
        local_crops_number,
        global_crops_size=224,
        local_crops_size=96,
    ):

        self.source_trans = transforms.Compose([
            # transforms.RandomCrop(224),
            # transforms.CenterCrop(224),
            transforms.ToTensor(),
            make_normalize_transform(),
        ])

        # self.crop = transforms.Compose([
        #     transforms.CenterCrop(224),
           
        # ])

        self.crop = transforms.Compose([
    transforms.Resize(224),  # 将短边缩放到 224，长边会按比例缩放
    transforms.CenterCrop(224),  # 然后裁剪到 224x224
])

        self.centercrop = transforms.Compose([
            transforms.CenterCrop(224),
           
        ])

        self.local_crops_number = local_crops_number

    def __call__(self, image):
        output = {}
        output["source"] = []

        if np.array(image).shape[0]<224 or np.array(image).shape[1]<224:
            crops_all = [
             self.centercrop(image) for _ in range(self.local_crops_number)
         ]
        else:
            crops_all = [
                self.centercrop(image) for _ in range(self.local_crops_number)
            ]
            
        for crops_image in crops_all:
            output["source"].append(self.source_trans(crops_image))   #单独使用好一些
    

        output["offsets"] = ()

        return output


class DataAugmentationResNet_test(object):
    def __init__(
        self,
        global_crops_scale,
        local_crops_scale,
        local_crops_number,
        global_crops_size=224,
        local_crops_size=96,
    ):

        self.source_trans = transforms.Compose([
            # transforms.RandomCrop(224),
            # transforms.CenterCrop(224),
            transforms.ToTensor(),
            make_normalize_transform(),
        ])

        # self.crop = transforms.Compose([
        #     transforms.CenterCrop(224),
           
        # ])

        self.crop = transforms.Compose([
    transforms.Resize(224),  # 将短边缩放到 224，长边会按比例缩放
    transforms.CenterCrop(224),  # 然后裁剪到 224x224
])

        self.centercrop = transforms.Compose([
            transforms.CenterCrop(224),
           
        ])

        self.local_crops_number = local_crops_number

    def __call__(self, image):
        output = {}
        output["source"] = []

        if np.array(image).shape[0]<224 or np.array(image).shape[1]<224:
            crops_all = [
             self.centercrop(image) for _ in range(self.local_crops_number)
         ]
        else:
            crops_all = [
                self.centercrop(image) for _ in range(self.local_crops_number)
            ]
            
        for crops_image in crops_all:
            output["source"].append(self.source_trans(crops_image))   #单独使用好一些
    

        output["offsets"] = ()

        return output
    


class DataAugmentationCLIP_gen(object):
    def __init__(
        self,
        global_crops_scale,
        local_crops_scale,
        local_crops_number,
        global_crops_size=224,
        local_crops_size=96,
    ):

        self.source_trans = transforms.Compose([
            # transforms.RandomCrop(224),
            # transforms.CenterCrop(224),
            transforms.ToTensor(),
            make_normalize_transform_clip(),
        ])

        # self.crop = transforms.Compose([
        #     transforms.RandomCrop(224),
           
        # ])

        self.crop = transforms.Compose([
    transforms.Resize(224),  # 将短边缩放到 224，长边会按比例缩放
    transforms.CenterCrop(224),  # 然后裁剪到 224x224
])

        self.centercrop = transforms.Compose([
            transforms.CenterCrop(224),
           
        ])

        self.local_crops_number = local_crops_number

    def __call__(self, image):
        output = {}
        output["source"] = []

        # if np.array(image).shape[0]<224 or np.array(image).shape[1]<224:
        #     crops_all = [
        #      self.crop(self.centercrop(image)) for _ in range(self.local_crops_number)
        #  ]
        # else:
        crops_all = [
            self.crop(image) for _ in range(self.local_crops_number)
        ]
            
        for crops_image in crops_all:
            output["source"].append(self.source_trans(crops_image))   #单独使用好一些
    

        output["offsets"] = ()

        return output