David310/Detect_AI-generated_Image

1# Copyright (c) 2019, Adobe Inc. All rights reserved.2#3# This work is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike4# 4.0 International Public License. To view a copy of this license, visit5# https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode.6 7import torch8import torch.nn.parallel9import numpy as np10import torch.nn as nn11import torch.nn.functional as F12from IPython import embed13 14class Downsample(nn.Module):15    def __init__(self, pad_type='reflect', filt_size=3, stride=2, channels=None, pad_off=0):16        super(Downsample, self).__init__()17        self.filt_size = filt_size18        self.pad_off = pad_off19        self.pad_sizes = [int(1.*(filt_size-1)/2), int(np.ceil(1.*(filt_size-1)/2)), int(1.*(filt_size-1)/2), int(np.ceil(1.*(filt_size-1)/2))]20        self.pad_sizes = [pad_size+pad_off for pad_size in self.pad_sizes]21        self.stride = stride22        self.off = int((self.stride-1)/2.)23        self.channels = channels24 25        # print('Filter size [%i]'%filt_size)26        if(self.filt_size==1):27            a = np.array([1.,])28        elif(self.filt_size==2):29            a = np.array([1., 1.])30        elif(self.filt_size==3):31            a = np.array([1., 2., 1.])32        elif(self.filt_size==4):    33            a = np.array([1., 3., 3., 1.])34        elif(self.filt_size==5):    35            a = np.array([1., 4., 6., 4., 1.])36        elif(self.filt_size==6):    37            a = np.array([1., 5., 10., 10., 5., 1.])38        elif(self.filt_size==7):    39            a = np.array([1., 6., 15., 20., 15., 6., 1.])40 41        filt = torch.Tensor(a[:,None]*a[None,:])42        filt = filt/torch.sum(filt)43        self.register_buffer('filt', filt[None,None,:,:].repeat((self.channels,1,1,1)))44 45        self.pad = get_pad_layer(pad_type)(self.pad_sizes)46 47    def forward(self, inp):48        if(self.filt_size==1):49            if(self.pad_off==0):50                return inp[:,:,::self.stride,::self.stride]    51            else:52                return self.pad(inp)[:,:,::self.stride,::self.stride]53        else:54            return F.conv2d(self.pad(inp), self.filt, stride=self.stride, groups=inp.shape[1])55 56def get_pad_layer(pad_type):57    if(pad_type in ['refl','reflect']):58        PadLayer = nn.ReflectionPad2d59    elif(pad_type in ['repl','replicate']):60        PadLayer = nn.ReplicationPad2d61    elif(pad_type=='zero'):62        PadLayer = nn.ZeroPad2d63    else:64        print('Pad type [%s] not recognized'%pad_type)65    return PadLayer66 67 68class Downsample1D(nn.Module):69    def __init__(self, pad_type='reflect', filt_size=3, stride=2, channels=None, pad_off=0):70        super(Downsample1D, self).__init__()71        self.filt_size = filt_size72        self.pad_off = pad_off73        self.pad_sizes = [int(1. * (filt_size - 1) / 2), int(np.ceil(1. * (filt_size - 1) / 2))]74        self.pad_sizes = [pad_size + pad_off for pad_size in self.pad_sizes]75        self.stride = stride76        self.off = int((self.stride - 1) / 2.)77        self.channels = channels78 79        # print('Filter size [%i]' % filt_size)80        if(self.filt_size == 1):81            a = np.array([1., ])82        elif(self.filt_size == 2):83            a = np.array([1., 1.])84        elif(self.filt_size == 3):85            a = np.array([1., 2., 1.])86        elif(self.filt_size == 4):87            a = np.array([1., 3., 3., 1.])88        elif(self.filt_size == 5):89            a = np.array([1., 4., 6., 4., 1.])90        elif(self.filt_size == 6):91            a = np.array([1., 5., 10., 10., 5., 1.])92        elif(self.filt_size == 7):93            a = np.array([1., 6., 15., 20., 15., 6., 1.])94 95        filt = torch.Tensor(a)96        filt = filt / torch.sum(filt)97        self.register_buffer('filt', filt[None, None, :].repeat((self.channels, 1, 1)))98 99        self.pad = get_pad_layer_1d(pad_type)(self.pad_sizes)100 101    def forward(self, inp):102        if(self.filt_size == 1):103            if(self.pad_off == 0):104                return inp[:, :, ::self.stride]105            else:106                return self.pad(inp)[:, :, ::self.stride]107        else:108            return F.conv1d(self.pad(inp), self.filt, stride=self.stride, groups=inp.shape[1])109 110 111def get_pad_layer_1d(pad_type):112    if(pad_type in ['refl', 'reflect']):113        PadLayer = nn.ReflectionPad1d114    elif(pad_type in ['repl', 'replicate']):115        PadLayer = nn.ReplicationPad1d116    elif(pad_type == 'zero'):117        PadLayer = nn.ZeroPad1d118    else:119        print('Pad type [%s] not recognized' % pad_type)120    return PadLayer121