import os import time import string import argparse import re import sys import random import pickle import torch import torch.backends.cudnn as cudnn import torch.utils.data import torch.nn.functional as F from torch.autograd import Variable import numpy as np from skimage.color import gray2rgb from nltk.metrics.distance import edit_distance import cv2 from utils import CTCLabelConverter, AttnLabelConverter, Averager from dataset_trba import hierarchical_dataset, AlignCollate from model_trba import Model, SuperPixler, CastNumpy, STRScore # import hiddenlayer as hl from lime import lime_image from lime.wrappers.scikit_image import SegmentationAlgorithm import matplotlib.pyplot as plt import random from captum._utils.models.linear_model import SkLearnLinearModel, SkLearnRidge import statistics import settings import sys import copy from captum_test import acquire_average_auc, saveAttrData from captum_improve_vitstr import rankedAttributionsBySegm from matplotlib import pyplot as plt from captum.attr._utils.visualization import visualize_image_attr device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') from captum.attr import ( GradientShap, DeepLift, DeepLiftShap, IntegratedGradients, LayerConductance, NeuronConductance, NoiseTunnel, Saliency, InputXGradient, GuidedBackprop, Deconvolution, GuidedGradCam, FeatureAblation, ShapleyValueSampling, Lime, KernelShap ) from captum.metrics import ( infidelity, sensitivity_max ) def getPredAndConf(opt, model, scoring, image, converter, labels): batch_size = image.size(0) length_for_pred = torch.IntTensor([opt.batch_max_length] * batch_size).to(device) text_for_pred = torch.LongTensor(batch_size, opt.batch_max_length + 1).fill_(0).to(device) text_for_loss, length_for_loss = converter.encode(labels, batch_max_length=opt.batch_max_length) if 'CTC' in opt.Prediction: preds = model(image, text_for_pred) confScore = scoring(preds) confScore = confScore.detach().cpu().numpy() # Calculate evaluation loss for CTC deocder. preds_size = torch.IntTensor([preds.size(1)] * batch_size) # Select max probabilty (greedy decoding) then decode index to character if opt.baiduCTC: _, preds_index = preds.max(2) preds_index = preds_index.view(-1) else: _, preds_index = preds.max(2) preds_str = converter.decode(preds_index.data, preds_size.data)[0] else: preds = model(image, text_for_pred, is_train=False) confScore = scoring(preds) confScore = confScore.detach().cpu().numpy() preds = preds[:, :text_for_loss.shape[1] - 1, :] target = text_for_loss[:, 1:] # without [GO] Symbol # cost = criterion(preds.contiguous().view(-1, preds.shape[-1]), target.contiguous().view(-1)) # select max probabilty (greedy decoding) then decode index to character _, preds_index = preds.max(2) preds_str = converter.decode(preds_index, length_for_pred) ### Remove all chars after '[s]' preds_str = preds_str[0] preds_str = preds_str[:preds_str.find('[s]')] # pred = pred[:pred_EOS] return preds_str, confScore ### Output and save segmentations only for one dataset only def outputSegmOnly(opt): ### targetDataset - one dataset only, SVTP-645, CUTE80-288images targetDataset = "CUTE80" # ['IIIT5k_3000', 'SVT', 'IC03_867', 'IC13_1015', 'IC15_2077', 'SVTP', 'CUTE80'] targetHeight = 32 targetWidth = 100 segmRootDir = "/home/uclpc1/Documents/STR/datasets/segmen"\ "tations/{}X{}/{}/".format(targetHeight, targetWidth, targetDataset) if not os.path.exists(segmRootDir): os.makedirs(segmRootDir) opt.eval = True ### Only IIIT5k_3000 eval_data_list = [targetDataset] target_output_orig = opt.outputOrigDir ### Taken from LIME segmentation_fn = SegmentationAlgorithm('quickshift', kernel_size=4, max_dist=200, ratio=0.2, random_seed=random.randint(0, 1000)) for eval_data in eval_data_list: eval_data_path = os.path.join(opt.eval_data, eval_data) AlignCollate_evaluation = AlignCollate(imgH=opt.imgH, imgW=opt.imgW, keep_ratio_with_pad=opt.PAD) eval_data, eval_data_log = hierarchical_dataset(root=eval_data_path, opt=opt, targetDir=target_output_orig) evaluation_loader = torch.utils.data.DataLoader( eval_data, batch_size=1, shuffle=False, num_workers=int(opt.workers), collate_fn=AlignCollate_evaluation, pin_memory=True) for i, (image_tensors, labels) in enumerate(evaluation_loader): image_tensors = ((image_tensors + 1.0) / 2.0) * 255.0 imgDataDict = {} img_numpy = image_tensors.cpu().detach().numpy()[0] ### Need to set batch size to 1 only if img_numpy.shape[0] == 1: img_numpy = gray2rgb(img_numpy[0]) # print("img_numpy shape: ", img_numpy.shape) # (32,100,3) segmOutput = segmentation_fn(img_numpy) # print("segmOutput unique: ", len(np.unique(segmOutput))) imgDataDict['segdata'] = segmOutput imgDataDict['label'] = labels[0] outputPickleFile = segmRootDir + "{}.pkl".format(i) with open(outputPickleFile, 'wb') as f: pickle.dump(imgDataDict, f) ### Returns the mean for each segmentation having shape as the same as the input ### This function can only one attribution image at a time def averageSegmentsOut(attr, segments): averagedInput = torch.clone(attr) sortedDict = {} for x in np.unique(segments): segmentMean = torch.mean(attr[segments == x][:]) sortedDict[x] = float(segmentMean.detach().cpu().numpy()) averagedInput[segments == x] = segmentMean return averagedInput, sortedDict def acquireSelectivityHit(origImg, attributions, segmentations, model, converter, labels, scoring): # print("segmentations unique len: ", np.unique(segmentations)) aveSegmentations, sortedDict = averageSegmentsOut(attributions[0,0], segmentations) sortedKeys = [k for k, v in sorted(sortedDict.items(), key=lambda item: item[1])] sortedKeys = sortedKeys[::-1] ### A list that should contain largest to smallest score # print("sortedDict: ", sortedDict) # {0: -5.51e-06, 1: -1.469e-05, 2: -3.06e-05,...} # print("aveSegmentations unique len: ", np.unique(aveSegmentations)) # print("aveSegmentations device: ", aveSegmentations.device) # cuda:0 # print("aveSegmentations shape: ", aveSegmentations.shape) # (224,224) # print("aveSegmentations: ", aveSegmentations) n_correct = [] confidenceList = [] # First index is one feature removed, second index two features removed, and so on... clonedImg = torch.clone(origImg) gt = str(labels[0]) for totalSegToHide in range(0, len(sortedKeys)): ### Acquire LIME prediction result currentSegmentToHide = sortedKeys[totalSegToHide] clonedImg[0,0][segmentations == currentSegmentToHide] = 0.0 pred, confScore = getPredAndConf(opt, model, scoring, clonedImg, converter, np.array([gt])) # To evaluate 'case sensitive model' with alphanumeric and case insensitve setting. if opt.sensitive and opt.data_filtering_off: pred = pred.lower() gt = gt.lower() alphanumeric_case_insensitve = '0123456789abcdefghijklmnopqrstuvwxyz' out_of_alphanumeric_case_insensitve = f'[^{alphanumeric_case_insensitve}]' pred = re.sub(out_of_alphanumeric_case_insensitve, '', pred) gt = re.sub(out_of_alphanumeric_case_insensitve, '', gt) if pred == gt: n_correct.append(1) else: n_correct.append(0) confScore = confScore[0][0]*100 confidenceList.append(confScore) return n_correct, confidenceList def main(opt): # 'IIIT5k_3000', 'SVT', 'IC03_860', 'IC03_867', 'IC13_857', 'IC13_1015', 'IC15_1811', 'IC15_2077', 'SVTP', 'CUTE80' datasetName = settings.TARGET_DATASET custom_segm_dataroot = "{}/{}X{}/{}/".format(settings.SEGM_DIR, opt.imgH, opt.imgW, datasetName) outputSelectivityPkl = "strexp_ave_{}_{}.pkl".format(settings.MODEL, datasetName) outputDir = "./attributionImgs/{}/{}/".format(settings.MODEL, datasetName) attrOutputDir = "./attributionData/{}/{}/".format(settings.MODEL, datasetName) acquireSelectivity = True acquireInfidelity = False acquireSensitivity = False ### GPU error imgHeight = 32 imgWidth = 100 if not os.path.exists(outputDir): os.makedirs(outputDir) if not os.path.exists(attrOutputDir): os.makedirs(attrOutputDir) """ model configuration """ if 'CTC' in opt.Prediction: converter = CTCLabelConverter(opt.character) else: converter = AttnLabelConverter(opt.character) opt.num_class = len(converter.character) if opt.rgb: opt.input_channel = 3 model_obj = Model(opt, device) print('model input parameters', opt.imgH, opt.imgW, opt.num_fiducial, opt.input_channel, opt.output_channel, opt.hidden_size, opt.num_class, opt.batch_max_length, opt.Transformation, opt.FeatureExtraction, opt.SequenceModeling, opt.Prediction) model = torch.nn.DataParallel(model_obj).to(device) # load model print('loading pretrained model from %s' % opt.saved_model) model.load_state_dict(torch.load(opt.saved_model, map_location=device)) opt.exp_name = '_'.join(opt.saved_model.split('/')[1:]) modelCopy = copy.deepcopy(model) scoring_singlechar = STRScore(opt=opt, converter=converter, device=device, enableSingleCharAttrAve=True) super_pixel_model_singlechar = torch.nn.Sequential( # super_pixler, # numpy2torch_converter, modelCopy, scoring_singlechar ).to(device) modelCopy.train() scoring_singlechar.train() super_pixel_model_singlechar.train() scoring = STRScore(opt=opt, converter=converter, device=device) super_pixel_model = torch.nn.Sequential( model, scoring ) model.train() scoring.train() super_pixel_model.train() """ keep evaluation model and result logs """ os.makedirs(f'./result/{opt.exp_name}', exist_ok=True) os.system(f'cp {opt.saved_model} ./result/{opt.exp_name}/') """ setup loss """ if 'CTC' in opt.Prediction: criterion = torch.nn.CTCLoss(zero_infinity=True).to(device) else: criterion = torch.nn.CrossEntropyLoss(ignore_index=0).to(device) # ignore [GO] token = ignore index 0 """Output shap values""" """ evaluation with 10 benchmark evaluation datasets """ # The evaluation datasets, dataset order is same with Table 1 in our paper. # eval_data_list = ['IIIT5k_3000', 'IC03_860', 'IC03_867', 'IC15_1811'] target_output_orig = opt.outputOrigDir # eval_data_list = ['IIIT5k_3000', 'SVT', 'IC03_860', 'IC03_867', 'IC13_857', # 'IC13_1015', 'IC15_1811', 'IC15_2077', 'SVTP', 'CUTE80'] # eval_data_list = ['IIIT5k_3000'] eval_data_list = [datasetName] # # To easily compute the total accuracy of our paper. # eval_data_list = ['IIIT5k_3000', 'SVT', 'IC03_867', # 'IC13_1015', 'IC15_2077', 'SVTP', 'CUTE80'] list_accuracy = [] total_forward_time = 0 total_evaluation_data_number = 0 total_correct_number = 0 log = open(f'./result/{opt.exp_name}/log_all_evaluation.txt', 'a') dashed_line = '-' * 80 print(dashed_line) log.write(dashed_line + '\n') selectivity_eval_results = [] imageData = [] targetText = "all" middleMaskThreshold = 5 testImgCount = 0 imgResultDir = str(opt.Transformation) + "-" + str(opt.FeatureExtraction) + "-" + str(opt.SequenceModeling) + "-" + str(opt.Prediction) + "-" + str(opt.scorer) # define a perturbation function for the input (used for calculating infidelity) def perturb_fn(modelInputs): noise = torch.tensor(np.random.normal(0, 0.003, modelInputs.shape)).float() noise = noise.to(device) return noise, modelInputs - noise if opt.blackbg: shapImgLs = np.zeros(shape=(1, 1, 32, 100)).astype(np.float32) trainList = np.array(shapImgLs) background = torch.from_numpy(trainList).to(device) if imgResultDir != "": if not os.path.exists(imgResultDir): os.makedirs(imgResultDir) for eval_data in eval_data_list: eval_data_path = os.path.join(opt.eval_data, eval_data) AlignCollate_evaluation = AlignCollate(imgH=opt.imgH, imgW=opt.imgW, keep_ratio_with_pad=opt.PAD) eval_data, eval_data_log = hierarchical_dataset(root=eval_data_path, opt=opt, targetDir=target_output_orig) evaluation_loader = torch.utils.data.DataLoader( eval_data, batch_size=1, shuffle=False, num_workers=int(opt.workers), collate_fn=AlignCollate_evaluation, pin_memory=True) # image_tensors, labels = next(iter(evaluation_loader)) ### Iterate one batch only for i, (orig_img_tensors, labels) in enumerate(evaluation_loader): # img_rgb *= 255.0 # img_rgb = img_rgb.astype('int') # print("img_rgb max: ", img_rgb.max()) ### 255 # img_rgb = np.asarray(orig_img_tensors) # segmentations = segmentation_fn(img_rgb) # print("segmentations shape: ", segmentations.shape) # (224, 224) # print("segmentations min: ", segmentations.min()) 0 # print("Unique: ", len(np.unique(segmentations))) # (70) # print("target: ", target) tensor([[ 0, 29, 26, 25, 12 results_dict = {} pklFilename = custom_segm_dataroot + "{}.pkl".format(i) with open(pklFilename, 'rb') as f: pklData = pickle.load(f) segmDataNP = pklData["segdata"] # print("segmDataNP unique: ", len(np.unique(segmDataNP))) assert pklData["label"] == labels[0] segmTensor = torch.from_numpy(segmDataNP).unsqueeze(0).unsqueeze(0) # print("segmTensor min: ", segmTensor.min()) # 0 starting segmentation segmTensor = segmTensor.to(device) # print("segmTensor shape: ", segmTensor.shape) # img1 = np.asarray(imgPIL.convert('L')) # sys.exit() # img1 = img1 / 255.0 # img1 = torch.from_numpy(img1).unsqueeze(0).unsqueeze(0).type(torch.FloatTensor).to(device) img1 = orig_img_tensors.to(device) img1.requires_grad = True bgImg = torch.zeros(img1.shape).to(device) # preds = model(img1, seqlen=converter.batch_max_length) target = converter.encode(labels) target = target[0][:, 1:] charOffset = 0 input = img1 origImgNP = torch.clone(orig_img_tensors).detach().cpu().numpy()[0][0] # (1, 1, 224, 224) origImgNP = gray2rgb(origImgNP) # preds = model(input) # preds_prob = F.softmax(preds, dim=2) # preds_max_prob, preds_max_idx = preds_prob.max(dim=2) # print("preds_max_idx: ", preds_max_idx) tensor([[14, 26, 25, 12 ### Captum test collectedAttributions = [] for charIdx in range(0, len(labels)): scoring_singlechar.setSingleCharOutput(charIdx + charOffset) gtClassNum = target[0][charIdx + charOffset] ### Shapley Value Sampling svs = ShapleyValueSampling(super_pixel_model_singlechar) # attr = svs.attribute(input, target=0, n_samples=200) ### Individual pixels, too long to calculate attributions = svs.attribute(input, target=gtClassNum, feature_mask=segmTensor) collectedAttributions.append(attributions) aveAttributions = torch.mean(torch.cat(collectedAttributions,dim=0), dim=0).unsqueeze(0) rankedAttr = rankedAttributionsBySegm(aveAttributions, segmDataNP) rankedAttr = rankedAttr.detach().cpu().numpy()[0][0] rankedAttr = gray2rgb(rankedAttr) mplotfig, _ = visualize_image_attr(rankedAttr, origImgNP, method='blended_heat_map') mplotfig.savefig(outputDir + '{}_shapley_l.png'.format(i)) mplotfig.clear() plt.close(mplotfig) saveAttrData(attrOutputDir + f'{i}_shapley_l.pkl', aveAttributions, segmDataNP, origImgNP) if acquireSelectivity: n_correct, confidenceList = acquireSelectivityHit(img1, aveAttributions, segmDataNP, modelCopy, converter, labels, scoring_singlechar) results_dict["shapley_local_acc"] = n_correct results_dict["shapley_local_conf"] = confidenceList if acquireInfidelity: infid = float(infidelity(super_pixel_model_singlechar, perturb_fn, img1, aveAttributions).detach().cpu().numpy()) results_dict["shapley_local_infid"] = infid if acquireSensitivity: sens = float(sensitivity_max(svs.attribute, img1, target=0).detach().cpu().numpy()) results_dict["shapley_local_sens"] = sens ### Shapley Value Sampling svs = ShapleyValueSampling(super_pixel_model) # attr = svs.attribute(input, target=0, n_samples=200) ### Individual pixels, too long to calculate attributions = svs.attribute(input, target=0, feature_mask=segmTensor) collectedAttributions.append(attributions) rankedAttr = rankedAttributionsBySegm(attributions, segmDataNP) rankedAttr = rankedAttr.detach().cpu().numpy()[0][0] rankedAttr = gray2rgb(rankedAttr) mplotfig, _ = visualize_image_attr(rankedAttr, origImgNP, method='blended_heat_map') mplotfig.savefig(outputDir + '{}_shapley.png'.format(i)) mplotfig.clear() plt.close(mplotfig) saveAttrData(attrOutputDir + f'{i}_shapley.pkl', attributions, segmDataNP, origImgNP) if acquireSelectivity: n_correct, confidenceList = acquireSelectivityHit(img1, attributions, segmDataNP, model, converter, labels, scoring) results_dict["shapley_acc"] = n_correct results_dict["shapley_conf"] = confidenceList if acquireInfidelity: infid = float(infidelity(super_pixel_model, perturb_fn, img1, attributions, normalize=True).detach().cpu().numpy()) results_dict["shapley_infid"] = infid if acquireSensitivity: sens = float(sensitivity_max(svs.attribute, img1, target=0).detach().cpu().numpy()) results_dict["shapley_sens"] = sens ### Global + Local context aveAttributions = torch.mean(torch.cat(collectedAttributions,dim=0), dim=0).unsqueeze(0) rankedAttr = rankedAttributionsBySegm(aveAttributions, segmDataNP) rankedAttr = rankedAttr.detach().cpu().numpy()[0][0] rankedAttr = gray2rgb(rankedAttr) mplotfig, _ = visualize_image_attr(rankedAttr, origImgNP, method='blended_heat_map') mplotfig.savefig(outputDir + '{}_shapley_gl.png'.format(i)) mplotfig.clear() plt.close(mplotfig) saveAttrData(attrOutputDir + f'{i}_shapley_gl.pkl', aveAttributions, segmDataNP, origImgNP) if acquireSelectivity: n_correct, confidenceList = acquireSelectivityHit(img1, aveAttributions, segmDataNP, modelCopy, converter, labels, scoring_singlechar) results_dict["shapley_global_local_acc"] = n_correct results_dict["shapley_global_local_conf"] = confidenceList if acquireInfidelity: infid = float(infidelity(super_pixel_model_singlechar, perturb_fn, img1, aveAttributions).detach().cpu().numpy()) results_dict["shapley_global_local_infid"] = infid if acquireSensitivity: sens = float(sensitivity_max(svs.attribute, img1, target=0).detach().cpu().numpy()) results_dict["shapley_global_local_sens"] = sens # Baselines ### Integrated Gradients ig = IntegratedGradients(super_pixel_model) attributions = ig.attribute(input, target=0) rankedAttr = rankedAttributionsBySegm(attributions, segmDataNP) rankedAttr = rankedAttr.detach().cpu().numpy()[0][0] rankedAttr = gray2rgb(rankedAttr) mplotfig, _ = visualize_image_attr(rankedAttr, origImgNP, method='blended_heat_map') mplotfig.savefig(outputDir + '{}_intgrad.png'.format(i)) mplotfig.clear() plt.close(mplotfig) saveAttrData(attrOutputDir + f'{i}_intgrad.pkl', attributions, segmDataNP, origImgNP) if acquireSelectivity: n_correct, confidenceList = acquireSelectivityHit(img1, attributions, segmDataNP, model, converter, labels, scoring) results_dict["intgrad_acc"] = n_correct results_dict["intgrad_conf"] = confidenceList if acquireInfidelity: infid = float(infidelity(super_pixel_model, perturb_fn, img1, attributions, normalize=True).detach().cpu().numpy()) results_dict["intgrad_infid"] = infid if acquireSensitivity: sens = float(sensitivity_max(ig.attribute, img1, target=0).detach().cpu().numpy()) results_dict["intgrad_sens"] = sens ### Gradient SHAP using zero-background gs = GradientShap(super_pixel_model) # We define a distribution of baselines and draw `n_samples` from that # distribution in order to estimate the expectations of gradients across all baselines baseline_dist = torch.zeros((1, 1, imgHeight, imgWidth)) baseline_dist = baseline_dist.to(device) attributions = gs.attribute(input, baselines=baseline_dist, target=0) rankedAttr = rankedAttributionsBySegm(attributions, segmDataNP) rankedAttr = rankedAttr.detach().cpu().numpy()[0][0] rankedAttr = gray2rgb(rankedAttr) mplotfig, _ = visualize_image_attr(rankedAttr, origImgNP, method='blended_heat_map') mplotfig.savefig(outputDir + '{}_gradshap.png'.format(i)) mplotfig.clear() plt.close(mplotfig) saveAttrData(attrOutputDir + f'{i}_gradshap.pkl', attributions, segmDataNP, origImgNP) if acquireSelectivity: n_correct, confidenceList = acquireSelectivityHit(img1, attributions, segmDataNP, model, converter, labels, scoring) results_dict["gradshap_acc"] = n_correct results_dict["gradshap_conf"] = confidenceList if acquireInfidelity: infid = float(infidelity(super_pixel_model, perturb_fn, img1, attributions, normalize=True).detach().cpu().numpy()) results_dict["gradshap_infid"] = infid if acquireSensitivity: sens = float(sensitivity_max(gs.attribute, img1, target=0).detach().cpu().numpy()) results_dict["gradshap_sens"] = sens ### DeepLift using zero-background dl = DeepLift(super_pixel_model) attributions = dl.attribute(input, target=0) rankedAttr = rankedAttributionsBySegm(attributions, segmDataNP) rankedAttr = rankedAttr.detach().cpu().numpy()[0][0] rankedAttr = gray2rgb(rankedAttr) mplotfig, _ = visualize_image_attr(rankedAttr, origImgNP, method='blended_heat_map') mplotfig.savefig(outputDir + '{}_deeplift.png'.format(i)) mplotfig.clear() plt.close(mplotfig) saveAttrData(attrOutputDir + f'{i}_deeplift.pkl', attributions, segmDataNP, origImgNP) if acquireSelectivity: n_correct, confidenceList = acquireSelectivityHit(img1, attributions, segmDataNP, model, converter, labels, scoring) results_dict["deeplift_acc"] = n_correct results_dict["deeplift_conf"] = confidenceList if acquireInfidelity: infid = float(infidelity(super_pixel_model, perturb_fn, img1, attributions, normalize=True).detach().cpu().numpy()) results_dict["deeplift_infid"] = infid if acquireSensitivity: sens = float(sensitivity_max(dl.attribute, img1, target=0).detach().cpu().numpy()) results_dict["deeplift_sens"] = sens ### Saliency saliency = Saliency(super_pixel_model) attributions = saliency.attribute(input, target=0) ### target=class0 rankedAttr = rankedAttributionsBySegm(attributions, segmDataNP) rankedAttr = rankedAttr.detach().cpu().numpy()[0][0] rankedAttr = gray2rgb(rankedAttr) mplotfig, _ = visualize_image_attr(rankedAttr, origImgNP, method='blended_heat_map') mplotfig.savefig(outputDir + '{}_saliency.png'.format(i)) mplotfig.clear() plt.close(mplotfig) saveAttrData(attrOutputDir + f'{i}_saliency.pkl', attributions, segmDataNP, origImgNP) if acquireSelectivity: n_correct, confidenceList = acquireSelectivityHit(img1, attributions, segmDataNP, model, converter, labels, scoring) results_dict["saliency_acc"] = n_correct results_dict["saliency_conf"] = confidenceList if acquireInfidelity: infid = float(infidelity(super_pixel_model, perturb_fn, img1, attributions, normalize=True).detach().cpu().numpy()) results_dict["saliency_infid"] = infid if acquireSensitivity: sens = float(sensitivity_max(saliency.attribute, img1, target=0).detach().cpu().numpy()) results_dict["saliency_sens"] = sens ### InputXGradient input_x_gradient = InputXGradient(super_pixel_model) attributions = input_x_gradient.attribute(input, target=0) rankedAttr = rankedAttributionsBySegm(attributions, segmDataNP) rankedAttr = rankedAttr.detach().cpu().numpy()[0][0] rankedAttr = gray2rgb(rankedAttr) mplotfig, _ = visualize_image_attr(rankedAttr, origImgNP, method='blended_heat_map') mplotfig.savefig(outputDir + '{}_inpxgrad.png'.format(i)) mplotfig.clear() plt.close(mplotfig) saveAttrData(attrOutputDir + f'{i}_inpxgrad.pkl', attributions, segmDataNP, origImgNP) if acquireSelectivity: n_correct, confidenceList = acquireSelectivityHit(img1, attributions, segmDataNP, model, converter, labels, scoring) results_dict["inpxgrad_acc"] = n_correct results_dict["inpxgrad_conf"] = confidenceList if acquireInfidelity: infid = float(infidelity(super_pixel_model, perturb_fn, img1, attributions, normalize=True).detach().cpu().numpy()) results_dict["inpxgrad_infid"] = infid if acquireSensitivity: sens = float(sensitivity_max(input_x_gradient.attribute, img1, target=0).detach().cpu().numpy()) results_dict["inpxgrad_sens"] = sens ## GuidedBackprop gbp = GuidedBackprop(super_pixel_model) attributions = gbp.attribute(input, target=0) rankedAttr = rankedAttributionsBySegm(attributions, segmDataNP) rankedAttr = rankedAttr.detach().cpu().numpy()[0][0] rankedAttr = gray2rgb(rankedAttr) mplotfig, _ = visualize_image_attr(rankedAttr, origImgNP, method='blended_heat_map') mplotfig.savefig(outputDir + '{}_guidedbp.png'.format(i)) mplotfig.clear() plt.close(mplotfig) saveAttrData(attrOutputDir + f'{i}_guidedbp.pkl', attributions, segmDataNP, origImgNP) if acquireSelectivity: n_correct, confidenceList = acquireSelectivityHit(img1, attributions, segmDataNP, model, converter, labels, scoring) results_dict["guidedbp_acc"] = n_correct results_dict["guidedbp_conf"] = confidenceList if acquireInfidelity: infid = float(infidelity(super_pixel_model, perturb_fn, img1, attributions, normalize=True).detach().cpu().numpy()) results_dict["guidedbp_infid"] = infid if acquireSensitivity: sens = float(sensitivity_max(gbp.attribute, img1, target=0).detach().cpu().numpy()) results_dict["guidedbp_sens"] = sens # # ## Deconvolution deconv = Deconvolution(super_pixel_model) attributions = deconv.attribute(input, target=0) rankedAttr = rankedAttributionsBySegm(attributions, segmDataNP) rankedAttr = rankedAttr.detach().cpu().numpy()[0][0] rankedAttr = gray2rgb(rankedAttr) mplotfig, _ = visualize_image_attr(rankedAttr, origImgNP, method='blended_heat_map') mplotfig.savefig(outputDir + '{}_deconv.png'.format(i)) mplotfig.clear() plt.close(mplotfig) saveAttrData(attrOutputDir + f'{i}_deconv.pkl', attributions, segmDataNP, origImgNP) if acquireSelectivity: n_correct, confidenceList = acquireSelectivityHit(img1, attributions, segmDataNP, model, converter, labels, scoring) results_dict["deconv_acc"] = n_correct results_dict["deconv_conf"] = confidenceList if acquireInfidelity: infid = float(infidelity(super_pixel_model, perturb_fn, img1, attributions, normalize=True).detach().cpu().numpy()) results_dict["deconv_infid"] = infid if acquireSensitivity: sens = float(sensitivity_max(deconv.attribute, img1, target=0).detach().cpu().numpy()) results_dict["deconv_sens"] = sens ### Feature ablator ablator = FeatureAblation(super_pixel_model) attributions = ablator.attribute(input, target=0, feature_mask=segmTensor) rankedAttr = rankedAttributionsBySegm(attributions, segmDataNP) rankedAttr = rankedAttr.detach().cpu().numpy()[0][0] rankedAttr = gray2rgb(rankedAttr) mplotfig, _ = visualize_image_attr(rankedAttr, origImgNP, method='blended_heat_map') mplotfig.savefig(outputDir + '{}_featablt.png'.format(i)) mplotfig.clear() plt.close(mplotfig) saveAttrData(attrOutputDir + f'{i}_featablt.pkl', attributions, segmDataNP, origImgNP) if acquireSelectivity: n_correct, confidenceList = acquireSelectivityHit(img1, attributions, segmDataNP, model, converter, labels, scoring) results_dict["featablt_acc"] = n_correct results_dict["featablt_conf"] = confidenceList if acquireInfidelity: infid = float(infidelity(super_pixel_model, perturb_fn, img1, attributions, normalize=True).detach().cpu().numpy()) results_dict["featablt_infid"] = infid if acquireSensitivity: sens = float(sensitivity_max(ablator.attribute, img1, target=0).detach().cpu().numpy()) results_dict["featablt_sens"] = sens ## LIME interpretable_model = SkLearnRidge(alpha=1, fit_intercept=True) ### This is the default used by LIME lime = Lime(super_pixel_model, interpretable_model=interpretable_model) attributions = lime.attribute(input, target=0, feature_mask=segmTensor) rankedAttr = rankedAttributionsBySegm(attributions, segmDataNP) rankedAttr = rankedAttr.detach().cpu().numpy()[0][0] rankedAttr = gray2rgb(rankedAttr) mplotfig, _ = visualize_image_attr(rankedAttr, origImgNP, method='blended_heat_map') mplotfig.savefig(outputDir + '{}_lime.png'.format(i)) mplotfig.clear() plt.close(mplotfig) saveAttrData(attrOutputDir + f'{i}_lime.pkl', attributions, segmDataNP, origImgNP) if acquireSelectivity: n_correct, confidenceList = acquireSelectivityHit(img1, attributions, segmDataNP, model, converter, labels, scoring) results_dict["lime_acc"] = n_correct results_dict["lime_conf"] = confidenceList if acquireInfidelity: infid = float(infidelity(super_pixel_model, perturb_fn, img1, attributions, normalize=True).detach().cpu().numpy()) results_dict["lime_infid"] = infid if acquireSensitivity: sens = float(sensitivity_max(lime.attribute, img1, target=0).detach().cpu().numpy()) results_dict["lime_sens"] = sens ### KernelSHAP ks = KernelShap(super_pixel_model) attributions = ks.attribute(input, target=0, feature_mask=segmTensor) rankedAttr = rankedAttributionsBySegm(attributions, segmDataNP) rankedAttr = rankedAttr.detach().cpu().numpy()[0][0] rankedAttr = gray2rgb(rankedAttr) mplotfig, _ = visualize_image_attr(rankedAttr, origImgNP, method='blended_heat_map') mplotfig.savefig(outputDir + '{}_kernelshap.png'.format(i)) mplotfig.clear() plt.close(mplotfig) saveAttrData(attrOutputDir + f'{i}_kernelshap.pkl', attributions, segmDataNP, origImgNP) if acquireSelectivity: n_correct, confidenceList = acquireSelectivityHit(img1, attributions, segmDataNP, model, converter, labels, scoring) results_dict["kernelshap_acc"] = n_correct results_dict["kernelshap_conf"] = confidenceList if acquireInfidelity: infid = float(infidelity(super_pixel_model, perturb_fn, img1, attributions, normalize=True).detach().cpu().numpy()) results_dict["kernelshap_infid"] = infid if acquireSensitivity: sens = float(sensitivity_max(ks.attribute, img1, target=0).detach().cpu().numpy()) results_dict["kernelshap_sens"] = sens selectivity_eval_results.append(results_dict) with open(outputSelectivityPkl, 'wb') as f: pickle.dump(selectivity_eval_results, f) testImgCount += 1 print("testImgCount: ", testImgCount) def outputOrigImagesOnly(opt): datasetName = "CUTE80" # ['IIIT5k_3000', 'SVT', 'IC03_867', 'IC13_1015', 'IC15_2077', 'SVTP', 'CUTE80'] opt.outputOrigDir = "./datasetOrigImgs/{}/".format(datasetName) opt.output_orig = True opt.corruption_num = 0 opt.apply_corruptions = False opt.min_imgnum = 0 opt.max_imgnum = 1000 target_output_orig = opt.outputOrigDir if not os.path.exists(target_output_orig): os.makedirs(target_output_orig) """ model configuration """ if 'CTC' in opt.Prediction: converter = CTCLabelConverter(opt.character) else: converter = AttnLabelConverter(opt.character) opt.num_class = len(converter.character) if opt.rgb: opt.input_channel = 3 model_obj = Model(opt, device) print('model input parameters', opt.imgH, opt.imgW, opt.num_fiducial, opt.input_channel, opt.output_channel, opt.hidden_size, opt.num_class, opt.batch_max_length, opt.Transformation, opt.FeatureExtraction, opt.SequenceModeling, opt.Prediction) model = torch.nn.DataParallel(model_obj).to(device) # load model print('loading pretrained model from %s' % opt.saved_model) model.load_state_dict(torch.load(opt.saved_model, map_location=device)) opt.exp_name = '_'.join(opt.saved_model.split('/')[1:]) scoring = STRScore(opt=opt, converter=converter, device=device) ### super_pixel_model = torch.nn.Sequential( model, scoring ) model.train() scoring.train() super_pixel_model.train() # print(model) """ keep evaluation model and result logs """ os.makedirs(f'./result/{opt.exp_name}', exist_ok=True) os.system(f'cp {opt.saved_model} ./result/{opt.exp_name}/') """ setup loss """ if 'CTC' in opt.Prediction: criterion = torch.nn.CTCLoss(zero_infinity=True).to(device) else: criterion = torch.nn.CrossEntropyLoss(ignore_index=0).to(device) # ignore [GO] token = ignore index 0 """Output shap values""" """ evaluation with 10 benchmark evaluation datasets """ # The evaluation datasets, dataset order is same with Table 1 in our paper. # eval_data_list = ['IIIT5k_3000', 'IC03_860', 'IC03_867', 'IC15_1811'] # eval_data_list = ['IIIT5k_3000', 'SVT', 'IC03_860', 'IC03_867', 'IC13_857', # 'IC13_1015', 'IC15_1811', 'IC15_2077', 'SVTP', 'CUTE80'] # eval_data_list = ['IIIT5k_3000'] eval_data_list = [datasetName] # # To easily compute the total accuracy of our paper. # eval_data_list = ['IIIT5k_3000', 'SVT', 'IC03_867', # 'IC13_1015', 'IC15_2077', 'SVTP', 'CUTE80'] list_accuracy = [] total_forward_time = 0 total_evaluation_data_number = 0 total_correct_number = 0 log = open(f'./result/{opt.exp_name}/log_all_evaluation.txt', 'a') dashed_line = '-' * 80 print(dashed_line) log.write(dashed_line + '\n') selectivity_eval_results = [] imageData = [] targetText = "all" middleMaskThreshold = 5 testImgCount = 0 imgResultDir = str(opt.Transformation) + "-" + str(opt.FeatureExtraction) + "-" + str(opt.SequenceModeling) + "-" + str(opt.Prediction) + "-" + str(opt.scorer) if opt.blackbg: shapImgLs = np.zeros(shape=(1, 1, 32, 100)).astype(np.float32) trainList = np.array(shapImgLs) background = torch.from_numpy(trainList).to(device) if imgResultDir != "": if not os.path.exists(imgResultDir): os.makedirs(imgResultDir) for eval_data in eval_data_list: eval_data_path = os.path.join(opt.eval_data, eval_data) AlignCollate_evaluation = AlignCollate(imgH=opt.imgH, imgW=opt.imgW, keep_ratio_with_pad=opt.PAD) eval_data, eval_data_log = hierarchical_dataset(root=eval_data_path, opt=opt, targetDir=target_output_orig) evaluation_loader = torch.utils.data.DataLoader( eval_data, batch_size=1, shuffle=False, num_workers=int(opt.workers), collate_fn=AlignCollate_evaluation, pin_memory=True) # image_tensors, labels = next(iter(evaluation_loader)) ### Iterate one batch only for i, (orig_img_tensors, labels) in enumerate(evaluation_loader): testImgCount += 1 print("testImgCount: ", testImgCount) ### Use to check if the model predicted the image or not. Output a pickle file with the image index. def modelDatasetPredOnly(opt): ### targetDataset - one dataset only, CUTE80 has 288 samples targetDataset = "CUTE80" # ['IIIT5k_3000', 'SVT', 'IC03_867', 'IC13_1015', 'IC15_2077', 'SVTP', 'CUTE80'] outputSelectivityPkl = "metrics_predictonly_results_{}.pkl".format(targetDataset) start_time = time.time() """ model configuration """ if 'CTC' in opt.Prediction: converter = CTCLabelConverter(opt.character) else: converter = AttnLabelConverter(opt.character) opt.num_class = len(converter.character) if opt.rgb: opt.input_channel = 3 model_obj = Model(opt, device) print('model input parameters', opt.imgH, opt.imgW, opt.num_fiducial, opt.input_channel, opt.output_channel, opt.hidden_size, opt.num_class, opt.batch_max_length, opt.Transformation, opt.FeatureExtraction, opt.SequenceModeling, opt.Prediction) model = torch.nn.DataParallel(model_obj).to(device) # load model print('loading pretrained model from %s' % opt.saved_model) model.load_state_dict(torch.load(opt.saved_model, map_location=device)) opt.exp_name = '_'.join(opt.saved_model.split('/')[1:]) scoring = STRScore(opt=opt, converter=converter, device=device) ### super_pixel_model = torch.nn.Sequential( model, scoring ) model.train() scoring.train() super_pixel_model.train() if opt.blackbg: shapImgLs = np.zeros(shape=(1, 1, 224, 224)).astype(np.float32) trainList = np.array(shapImgLs) background = torch.from_numpy(trainList).to(device) opt.eval = True eval_data_list = [targetDataset] testImgCount = 0 list_accuracy = [] total_forward_time = 0 total_evaluation_data_number = 0 total_correct_number = 0 log = open(f'./result/{opt.exp_name}/log_all_evaluation.txt', 'a') dashed_line = '-' * 80 print(dashed_line) log.write(dashed_line + '\n') target_output_orig = opt.outputOrigDir predOutput = [] for eval_data in eval_data_list: eval_data_path = os.path.join(opt.eval_data, eval_data) AlignCollate_evaluation = AlignCollate(imgH=opt.imgH, imgW=opt.imgW, keep_ratio_with_pad=opt.PAD) eval_data, eval_data_log = hierarchical_dataset(root=eval_data_path, opt=opt, targetDir=target_output_orig) evaluation_loader = torch.utils.data.DataLoader( eval_data, batch_size=1, shuffle=False, num_workers=int(opt.workers), collate_fn=AlignCollate_evaluation, pin_memory=True) testImgCount = 0 for i, (orig_img_tensors, labels) in enumerate(evaluation_loader): image = orig_img_tensors.to(device) batch_size = 1 length_for_pred = torch.IntTensor([opt.batch_max_length] * batch_size).to(device) text_for_pred = torch.LongTensor(batch_size, opt.batch_max_length + 1).fill_(0).to(device) text_for_loss, length_for_loss = converter.encode(labels, batch_max_length=opt.batch_max_length) if 'CTC' in opt.Prediction: preds = model(image, text_for_pred) confScore = scoring(preds) confScore = confScore.detach().cpu().numpy() # Calculate evaluation loss for CTC deocder. preds_size = torch.IntTensor([preds.size(1)] * batch_size) # Select max probabilty (greedy decoding) then decode index to character if opt.baiduCTC: _, preds_index = preds.max(2) preds_index = preds_index.view(-1) else: _, preds_index = preds.max(2) preds_str = converter.decode(preds_index.data, preds_size.data)[0] else: preds = model(image, text_for_pred, is_train=False) confScore = scoring(preds) confScore = confScore.detach().cpu().numpy() preds = preds[:, :text_for_loss.shape[1] - 1, :] target = text_for_loss[:, 1:] # without [GO] Symbol # cost = criterion(preds.contiguous().view(-1, preds.shape[-1]), target.contiguous().view(-1)) # select max probabilty (greedy decoding) then decode index to character _, preds_index = preds.max(2) preds_str = converter.decode(preds_index, length_for_pred) ### Remove all chars after '[s]' preds_str = preds_str[0] preds_str = preds_str[:preds_str.find('[s]')] # print("preds_str: ", preds_str) # lowercased prediction # print("labels: ", labels[0]) # gt already in lowercased if preds_str==labels[0]: predOutput.append(1) else: predOutput.append(0) with open(outputSelectivityPkl, 'wb') as f: pickle.dump(predOutput, f) if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--eval_data', required=True, help='path to evaluation dataset') parser.add_argument('--benchmark_all_eval', action='store_true', help='evaluate 10 benchmark evaluation datasets') parser.add_argument('--workers', type=int, help='number of data loading workers', default=4) parser.add_argument('--batch_size', type=int, default=192, help='input batch size') parser.add_argument('--saved_model', required=True, help="path to saved_model to evaluation") """ Data processing """ parser.add_argument('--batch_max_length', type=int, default=25, help='maximum-label-length') parser.add_argument('--imgH', type=int, default=32, help='the height of the input image') parser.add_argument('--imgW', type=int, default=100, help='the width of the input image') parser.add_argument('--superHeight', type=int, default=5, help='the height of the superpixel') parser.add_argument('--superWidth', type=int, default=2, help='the width of the superpixel') parser.add_argument('--min_imgnum', type=int, default=0, help='set this to skip for loop index of specific image number') parser.add_argument('--max_imgnum', type=int, default=2, help='set this to skip for loop index of specific image number') parser.add_argument('--severity', type=int, default=1, help='severity level if apply corruptions') parser.add_argument('--scorer', type=str, default='cumprod', help='See STRScore: cumprod | mean') parser.add_argument('--corruption_num', type=int, default=0, help='corruption to apply') parser.add_argument('--confidence_mode', type=int, default=0, help='0-sum of argmax; 1-edit distance') parser.add_argument('--outputOrigDir', type=str, default="output_orig/", help='output directory to save original \ images. This will be automatically created. Needs --output_orig too.') parser.add_argument('--output_orig', action='store_true', help='if true, output first original rgb image of each batch') parser.add_argument('--compare_corrupt', action='store_true', help='set to true to output results across corruptions') parser.add_argument('--is_shap', action='store_true', help='no need to call in command line') parser.add_argument('--blackbg', action='store_true', help='if True, background color for covering features will be black(0)') parser.add_argument('--rgb', action='store_true', help='use rgb input') parser.add_argument('--character', type=str, default='0123456789abcdefghijklmnopqrstuvwxyz', help='character label') parser.add_argument('--sensitive', action='store_true', help='for sensitive character mode') parser.add_argument('--PAD', action='store_true', help='whether to keep ratio then pad for image resize') parser.add_argument('--data_filtering_off', action='store_true', help='for data_filtering_off mode') parser.add_argument('--apply_corruptions', action='store_true', help='apply corruptions to images') parser.add_argument('--output_feat_maps', action='store_true', help='toggle this to output images of featmaps') parser.add_argument('--baiduCTC', action='store_true', help='for data_filtering_off mode') """ Model Architecture """ parser.add_argument('--Transformation', type=str, required=True, help='Transformation stage. None|TPS') parser.add_argument('--FeatureExtraction', type=str, required=True, help='FeatureExtraction stage. VGG|RCNN|ResNet') parser.add_argument('--SequenceModeling', type=str, required=True, help='SequenceModeling stage. None|BiLSTM') parser.add_argument('--Prediction', type=str, required=True, help='Prediction stage. CTC|Attn') parser.add_argument('--num_fiducial', type=int, default=20, help='number of fiducial points of TPS-STN') parser.add_argument('--input_channel', type=int, default=1, help='the number of input channel of Feature extractor') parser.add_argument('--output_channel', type=int, default=512, help='the number of output channel of Feature extractor') parser.add_argument('--hidden_size', type=int, default=256, help='the size of the LSTM hidden state') opt = parser.parse_args() """ vocab / character number configuration """ if opt.sensitive: opt.character = string.printable[:-6] # same with ASTER setting (use 94 char). cudnn.benchmark = True cudnn.deterministic = True opt.num_gpu = torch.cuda.device_count() # acquire_average_auc(opt) main(opt) # outputOrigImagesOnly(opt)