deletion_batch_v4_recuperacion.py

import argparse
import os
import random
import shutil
import time
import sys
import warnings
from srblib import abs_path
from PIL import ImageFilter, Image
import torch
import torch.nn as nn
import torch.nn.parallel
import torch.backends.cudnn as cudnn
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
import torchvision.transforms as transforms
import torchvision.datasets as datasets
import torchvision.models as models
import numpy as np
import matplotlib.pyplot as plt
from tqdm import tqdm, trange
import skimage

# Fixing for deterministic results
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False

# bibliotecas inpainter
sys.path.insert(0, './generativeimptorch')
from utils.tools import get_config, get_model_list
from model.networks import Generator

imagenet_val_xml_path = './val_bb'
imagenet_val_path = './val/'
base_img_dir = abs_path(imagenet_val_path)
input_dir_path = 'images_list.txt'
text_file = abs_path(input_dir_path)
imagenet_class_mappings = './imagenet_class_mappings'

torch.manual_seed(0)
learning_rate = 0.1 * 0.8  # orig (0.3) 0.1 (preservation sparser) 0.3 (preservation dense)
max_iterations = 228  # 130 *2
l1_coeff = 0.01e-5 * 2  # *2 *4 *0.5 (robusto)
size = 224

tv_beta = 3
tv_coeff = 1e-2
factorTV = 1 * 0.5 * 0.005  # 1(dense) o 0.5 (sparser/sharp)   #0.5 (preservation)


def inpainter(img, mask):
    config = get_config('./generativeimptorch/configs/config.yaml')
    checkpoint_path = os.path.join('./generativeimptorch/checkpoints',
                                   config['dataset_name'],
                                   config['mask_type'] + '_' + config['expname'])
    cuda = config['cuda']
    device_ids = config['gpu_ids']

    with torch.no_grad():  # enter no grad context
        # Test a single masked image with a given mask
        x = img
        # denormaliza imagenet y se normaliza a inpainter [-1,1] mean=0.5, std=0.5
        x = transforms.Normalize(mean=[0.015 / 0.229, 0.044 / 0.224, 0.094 / 0.225],
                                 std=[0.5 / 0.229, 0.5 / 0.224, 0.5 / 0.225])(x)
        x = x * (mask)
        # Define the trainer
        netG = Generator(config['netG'], cuda, device_ids)
        # Resume weight
        last_model_name = get_model_list(checkpoint_path, "gen", iteration=0)
        netG.load_state_dict(torch.load(last_model_name))

        # netG = torch.nn.parallel.DataParallel(netG, device_ids=[0, 1])
        netG.cuda()
        # Inference
        x1, x2, offset_flow = netG(x, (1. - mask))

    return x2


def tv_norm(input, tv_beta):
    img = input[:, 0, :]
    row_grad = torch.abs((img[:, :-1, :] - img[:, 1:, :])).pow(tv_beta).sum(dim=(1, 2))
    col_grad = torch.abs((img[:, :, :-1] - img[:, :, 1:])).pow(tv_beta).sum(dim=(1, 2))
    return row_grad + col_grad


torch.cuda.set_device(1)  # especificar cual gpu 0 o 1
model = models.googlenet(pretrained=True)
# model = models.resnet50(pretrained=True)
# model = models.vgg16(pretrained=True)
# model = models.alexnet(pretrained=True)
model.cuda()
model.eval()

for param in model.parameters():
    param.requires_grad = False

print('GPU 0 Metod. Recuperacion ver 4')


def imagenet_label_mappings():
    fileName = os.path.join(imagenet_class_mappings, 'imagenet_label_mapping')
    with open(fileName, 'r') as f:
        image_label_mapping = {int(x.split(":")[0]): x.split(":")[1].strip()
                               for x in f.readlines() if len(x.strip()) > 0}
        return image_label_mapping

im_label_map = imagenet_label_mappings()


transform_val = transforms.Compose([
    transforms.Resize((256, 256)),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225]),
])


# Plots image from tensor
def tensor_imshow(inp, title=None, **kwargs):
    """Imshow for Tensor."""
    inp = inp.numpy().transpose((1, 2, 0))
    # Mean and std for ImageNet
    mean = np.array([0.485, 0.456, 0.406])
    std = np.array([0.229, 0.224, 0.225])
    inp = std * inp + mean
    inp = np.clip(inp, 0, 1)
    plt.imshow(inp, **kwargs)
    if title is not None:
        plt.title(title)
    plt.show()


list_of_layers = ['conv1',
                  'conv2',
                  'conv3',
                  'inception3a',
                  'inception3b',
                  'inception4a',
                  'inception4b',
                  'inception4c',
                  'inception4d',
                  'inception4e',
                  'inception5a',
                  'inception5b',
                  'fc'
                  ]

# capas para resnet50
# list_of_layers = ['relu',
#                   'layer1.0',
#                   'layer1.1',
#                   'layer1.2',
#                   'layer2.0',
#                   'layer2.1',
#                   'layer2.2',
#                   'layer2.3',
#                   'layer3.0',
#                   'layer3.1',
#                   'layer3.2',
#                   'layer3.3',
#                   'layer3.4',
#                   'layer3.5',
#                   'layer4.0',
#                   'layer4.1',
#                   'layer4.2',
#                   ]

# capas para vgg16
# list_of_layers = ['features.1',
#                   'features.3',
#                   'features.6',
#                   'features.8',
#                   'features.11',
#                   'features.13',
#                   'features.15',
#                   'features.18',
#                   'features.20',
#                   'features.22',
#                   'features.25',
#                   'features.27',
#                   'features.29'
#                   ]

# capas para alexnet
# list_of_layers = ['features.1',
#                   'features.4',
#                   'features.7',
#                   'features.9',
#                   'features.11',
#                   'classifier.2',
#                   'classifier.5'
#                   ]

activation_orig = {}


def get_activation_orig(name):
    def hook(model, input, output):
        activation_orig[name] = output

    return hook


def get_activation_mask(name):
    def hook(model, input, output):
        act_mask = output
        # print(act_mask.shape). #debug
        # print(activation_orig[name].shape) #debug
        limite_sup = (act_mask <= torch.fmax(torch.tensor(0), activation_orig[name]))
        limite_inf = (act_mask >= torch.fmin(torch.tensor(0), activation_orig[name]))
        oper = limite_sup * limite_inf
        # print('oper shape=',oper.shape). #debug
        act_mask.requires_grad_(True)
        act_mask.retain_grad()
        h = act_mask.register_hook(lambda grad: grad * oper)
        # x.register_hook(update_gradients(2))
        # activation[name]=act_mask
        # h.remove()

    return hook


def my_explanation(img_batch, max_iterations, gt_category):
    F_hook = []
    exp_hook = []
    #
    # for module_name, module in model.named_modules():
    #     if module_name in list_of_layers:
    #         F_hook.append(module.register_forward_hook(get_activation_orig(module_name)))

    for name, layer in model.named_children():
        if name in list_of_layers:
            F_hook.append(layer.register_forward_hook(get_activation_orig(name)))

    # se calculan las activaciones para el batch de imágenes y se almacenan en la lista activation_orig
    # la funcion "feed forward" registra los hook

    org_softmax = torch.nn.Softmax(dim=1)(model(img_batch))

    # se borran los hook registrados en Feed Forward
    for fh in F_hook:
        fh.remove()
    #
    # for module_name, module in model.named_modules():
    #     if module_name in list_of_layers:
    #         exp_hook.append(module.register_forward_hook(get_activation_mask(module_name)))

    for name, layer in model.named_children():
        if name in list_of_layers:
            exp_hook.append(layer.register_forward_hook(get_activation_mask(name)))

    for param in model.parameters():
        param.requires_grad = False

    np.random.seed(seed=0)
    mask = torch.from_numpy(np.float32(np.random.uniform(0, 0.01, size=(1, 1, 224, 224))))
    mask = mask.expand(img_batch.size(0), 1, 224, 224)
    mask = mask.cuda()
    mask.requires_grad = True

    # null_img = torch.zeros(img_batch.size(0), 3, 224, 224).cuda()
    # null_img_blur = transforms.GaussianBlur(kernel_size=223, sigma=10)(img_batch)
    # null_img_blur.requires_grad = False
    # null_img = null_img_blur.cuda()

    optimizer = torch.optim.Adam([mask], lr=learning_rate)

    for i in trange(max_iterations):
        extended_mask = mask.expand(img_batch.size(0), 3, 224, 224)

        img_inpainted = inpainter(img_batch, mask)
        img_inpainted = transforms.Normalize(mean=-1, std=2)(img_inpainted)
        img_inpainted = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                             std=[0.229, 0.224, 0.225])(img_inpainted)

        perturbated_input = img_batch.mul(extended_mask) + img_inpainted.mul(1 - extended_mask)
        # perturbated_input = perturbated_input.to(torch.float32)
        optimizer.zero_grad()
        outputs = torch.nn.Softmax(dim=1)(model(perturbated_input))  # (3,1000)

        preds = outputs[torch.arange(0, img_batch.size(0)).tolist(), gt_category.tolist()]

        loss = l1_coeff * torch.sum(torch.abs(1 - mask), dim=(1, 2, 3)) + preds + \
               factorTV * tv_coeff * tv_norm(mask, tv_beta)

        loss.backward(gradient=torch.ones_like(loss).cuda())
        optimizer.step()
        mask.data.clamp_(0, 1)

    for eh in exp_hook:
        eh.remove()

    # Para visualizar las máscaras
    mask_np = (mask.cpu().detach().numpy())

    for i in range(mask_np.shape[0]):
        fig = plt.figure()
        fig.subplots_adjust(left=0.03, bottom=0, right=0.97, top=1, wspace=0.1, hspace=0.1)
        fig.set_size_inches(12, 5)
        ax = fig.add_subplot(1, 3, 1)
        inp = img_batch[i].cpu().numpy().transpose((1, 2, 0))
        mean = np.array([0.485, 0.456, 0.406])
        std = np.array([0.229, 0.224, 0.225])
        inp = std * inp + mean
        inp = np.clip(inp, 0, 1)
        ax.imshow(inp)
        ax.set_xticks([])
        ax.set_yticks([])
        ax = fig.add_subplot(1, 3, 2)
        mask_rz = mask_np[i, 0, :, :]
        ax.imshow(mask_rz)
        ax.set_xticks([])
        ax.set_yticks([])
        ax = fig.add_subplot(1, 3, 3)
        img_masked = np.multiply(inp, np.repeat(np.expand_dims(mask_rz, axis=2), 3, axis=2))
        ax.imshow(img_masked)
        ax.set_xticks([])
        ax.set_yticks([])
        # fig.tight_layout()
        plt.show()

    return mask



init_time = time.time()

########## Se carga el batch de imágenes adversarias ############
imgs_adv = np.load('adv_im_MRC_strong.npy')
adv_batch = torch.from_numpy(imgs_adv)
adv_batch = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                 std=[0.229, 0.224, 0.225])(adv_batch)

# preds_adv = torch.nn.Softmax(dim=1)(model(adv_batch))  # tensor(200,1000)
# probs_adv, labels_adv = torch.max(preds_adv, 1)  # Top 1 predicciones adversarias (200, 1)
##################################################################
# orig_labels = np.load('adv_orig_labels.npy')  # (200,)

img_name_list = []
with open(text_file, 'r') as f:
    for line in f:
        img_name_list.append(line.split('\n')[0])


class DataProcessing:
    def __init__(self, data_path, transform, img_idxs=[0, 1], if_noise=0, noise_var=0.0):
        self.data_path = data_path
        self.transform = transform
        self.if_noise = if_noise
        self.noise_mean = 0
        self.noise_var = noise_var
        self.img_idxs = img_idxs

        img_list = img_name_list[img_idxs[0]:img_idxs[1]]
        self.img_filenames = [os.path.join(data_path, f'{i}.JPEG') for i in img_list]
        # self.img_filenames.sort()

    def __getitem__(self, index):
        img = Image.open(os.path.join(self.data_path, self.img_filenames[index])).convert('RGB')
        target = self.get_image_class(os.path.join(self.data_path, self.img_filenames[index]))
        img_adv = adv_batch[index + self.img_idxs[0]]
        if self.if_noise == 1:
            img = skimage.util.random_noise(np.asarray(img), mode='gaussian',
                                            mean=self.noise_mean, var=self.noise_var,
                                            )  # numpy, dtype=float64,range (0, 1)
            img = Image.fromarray(np.uint8(img * 255))

        img = self.transform(img)
        return img, img_adv, target, os.path.join(self.data_path, self.img_filenames[index])
        # return img, target

    def __len__(self):
        return len(self.img_filenames)

    def get_image_class(self, filepath):
        # ImageNet 2012 validation set images?
        with open(os.path.join(imagenet_class_mappings, "ground_truth_val2012")) as f:
            ground_truth_val2012 = {x.split()[0]: int(x.split()[1])
                                    for x in f.readlines() if len(x.strip()) > 0}

        def get_class(f):
            ret = ground_truth_val2012.get(f, None)
            return ret

        image_class = get_class(filepath.split('/')[-1])
        return image_class


transform_val = transforms.Compose([
    transforms.Resize((256, 256)),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225]),
])

val_dataset = DataProcessing(base_img_dir, transform_val, img_idxs=[0, 200], if_noise=0, noise_var=0.0)
val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=25, shuffle=False, num_workers=10,
                                         pin_memory=True)

iterator = tqdm(enumerate(val_loader), total=len(val_loader), desc='batch')

recov_top_cnt_full = []
pertb_top_cnt_full = []
pertb_radio = []
recov_radio = []

for i, (images, imgs_adv, target, file_names) in iterator:
    imgs_adv.requires_grad = False
    images_orig = images.cuda()
    imgs_adv = imgs_adv.cuda()

    # predicciones originales para comparar
    preds_orig = torch.nn.Softmax(dim=1)(model(images_orig))
    probs_orig, labels_orig = torch.topk(preds_orig, 10)  # Top 10 predicciones originales

    preds_adv = torch.nn.Softmax(dim=1)(model(imgs_adv))  # tensor(n_batch,1000)
    probs_adv, labels_adv = torch.topk(preds_adv, 1000)  # Top 1 predicciones adversarias
    _, labels_adv_top = torch.max(preds_adv, 1)

    # obtención de explicaciones para imágenes adversarias
    _, orig_labels = torch.max(preds_orig, 1) # labels originales para generar las explicaciones
    mask = my_explanation(imgs_adv, max_iterations, labels_adv_top)

    # reenmascaramiento de img adv con explicación
    adv_masked = imgs_adv.mul(mask)
    adv_masked = adv_masked.to(torch.float32)

    # prediccion reenmascaramiento
    preds_masked = torch.nn.Softmax(dim=1)(model(adv_masked))  # tensor(n_batch,1000)
    probs_masked, labels_masked = torch.topk(preds_masked, 1000)  # predicciones recuperadas y ordenadas (n_batch, 1000)

    for i in range(val_loader.batch_size):  # se itera en el tamaño del batch
        prob_orig = probs_orig.cpu().detach()[i]  # (10, ) topk = 10
        label_orig = labels_orig.cpu().detach()[i]  # (10, ) topk = 10

        prob_masked = probs_masked.cpu().detach()[i]  # (1000,)
        label_masked = labels_masked.cpu().detach()[i]  # (1000,)
        label_adv = labels_adv.cpu().detach()[i]

        prob_adv = probs_adv.cpu().detach()[i]

        # se buscan donde estan las etiquetas originales dentro de las perturbadas
        pertub_pos_list = [torch.where(label_adv == label_orig_item)[0].item() for label_orig_item in label_orig]
        pertub_pos_list_tensor = torch.tensor(pertub_pos_list)
        pertb_radio.append(pertub_pos_list_tensor)
        pertb_top_cnt = [torch.where(pertub_pos_list_tensor <= i)[0].nelement() >= 1 for i in range(10)]
        pertb_top_cnt_full.append(torch.tensor(pertb_top_cnt))

        # se buscan donde estan las etiquetas originales dentro de las recuperadas
        recov_pos_list = [torch.where(label_masked == label_orig_item)[0].item() for label_orig_item in label_orig]
        recov_pos_list_tensor = torch.tensor(recov_pos_list)
        recov_radio.append(recov_pos_list_tensor)
        # el indice de pos_list_tensor determina el rango de predicciones a incluir en el top 10
        # por ejemplo pos_list_tensor[0] analiza el top 10 de la primera prediccion original en el grupo de recuperados
        # pos_list_tensor[0:4] analiza el top 10 del grupo de las 3 primeras predicciones originales en el grupo recup
        recov_top_cnt = [torch.where(recov_pos_list_tensor <= i)[0].nelement() >= 1 for i in range(10)]
        recov_top_cnt_full.append(torch.tensor(recov_top_cnt))

        print('muestra ', file_names[i].split('/')[-1].split('.JPEG')[0])
        print('top 10 muestra original: ', label_orig.tolist())
        print('top 10 muestra original (decod): ', [im_label_map.get(label) for label in label_orig.tolist()])
        print('top 10 prob muestra original (%): ', [round(num * 100, 2) for num in prob_orig.tolist()])
        print('lista top 10 recuperados acum ', recov_top_cnt)
        print('top 10 perturbados    ', label_adv[0:10].tolist())
        print('top 10 perturbados (decod) ', [im_label_map.get(label) for label in label_adv[0:10].tolist()])
        print('top 10 prob muestra pertb (%)', [round(num * 100, 2) for num in prob_adv[0:10].tolist()])
        print('pos orig en perturbado  ', pertub_pos_list)
        print('top 10 recuperados    ', label_masked[0:10].tolist())
        print('top 10 recuperados (decod) ', [im_label_map.get(label) for label in label_masked[0:10].tolist()])
        print('top 10 prob muestra recup (%)', [round(num * 100, 2) for num in prob_masked[0:10].tolist()])
        print('pos orig en recuperado  ', recov_pos_list)
        print('')


recov_table = torch.stack(recov_top_cnt_full)
recov_stats = recov_table.sum(0) / (val_loader.batch_size * len(val_loader))

pertb_table = torch.stack(pertb_top_cnt_full)
pertb_stats = pertb_table.sum(0) / (val_loader.batch_size * len(val_loader))

print('estado despues de perturbar')
print(pertb_stats)
print('')

print('estado despues de recuperar')
print(recov_stats)
print('')

pr = torch.stack(pertb_radio).float()
print('radio perturbacion')
print(pr)
print('promedio')
print(pr.mean(0))
print('')

rr = torch.stack(recov_radio).float()
print('radio recuperacion')
print(rr)
print('promedio')
print(rr.mean(0))
print('')
print('Time taken: {:.3f}'.format(time.time() - init_time))