EVMET.Architecture(
models.resnet50(pretrained=True).eval(), 'resnet50', 'layer4'),
'vgg16':
EVMET.Architecture(
models.vgg16(pretrained=True).eval(), 'vgg16', 'features_29')
}
arch = arch_dict[arch_name]
#arch=EVMET.Architecture(models.resnet101(pretrained=True).eval(),'resnet101','layer4')
#arch=EVMET.Architecture(models.resnet152(pretrained=True).eval(),'resnet152','layer4')
#arch=EVMET.Architecture(models.vgg16(pretrained=True).eval(),'vgg16','features_29')
metrics_dict = {
'average_drop':
ADIC.AverageDrop('average_drop', arch),
'average_increase':
ADIC.IncreaseInConfidence('average_increase', arch),
'increase_in_confidence':
ADIC.IncreaseInConfidence('increase_in_confidence', arch),
'deletion':
DAI.Deletion('deletion', arch),
'insertion':
DAI.Insertion('insertion', arch),
'average_complexity':
COMPLEXITY.Complexity('Average complexity', arch),
'average_coherency':
COHERENCY.Coherency('Average coherency', arch),
'average_score_variance':
ASV.AverageScoreVariance('Average score variance', arch)
}
metrics = {m: metrics_dict[m] for m in metric_names}
with open('filter.txt', 'r') as f:
txt = f.read()
img_list = [p.split()[0] for p in txt.strip().split('\n')[base:base + window]]
img_list = {get_num_img(p.split()[0]): p.split()[0] for p in img_list}
try:
os.mkdir(f'{outpath_root}filter/')
except:
pass
#arch=EVMET.Architecture(models.resnet18(pretrained=True).eval(),'resnet18','layer4')
arch = EVMET.Architecture(
models.vgg16(pretrained=True).eval(), 'vgg16', 'features_29')
avg_drop = ADIC.AverageDrop('average_drop', arch)
inc_conf = ADIC.IncreaseInConfidence('increase_in_confidence', arch)
deletion = DAI.Deletion('deletion', arch)
insertion = DAI.Insertion('insertion', arch)
img_dict = IMUT.IMG_list(path=p, outpath_root='out/filter/', GT=GT,
labs=labs).select_imgs(img_list)
em = EVMET.MetricsEvaluator(img_dict,
saliency_map_extractor=run,
model=arch,
metrics=[avg_drop, inc_conf, deletion, insertion])
start = time.time()
now = start
path0 = img_dict.get_outpath_root()
conv_layer = arch.layer #MODEL_CONFIG[arch.name]['conv_layer']