def evaluate():
setup = inversefed.utils.system_startup()
defs = inversefed.training_strategy('conservative')
defs.epochs = opt.epochs
loss_fn, trainloader, validloader = preprocess(opt, defs, valid=False)
model = create_model(opt)
model.to(**setup)
root = create_save_dir()
filename = os.path.join(root, '{}_{}.pth'.format(opt.arch, opt.epochs))
print(filename)
if not os.path.exists(filename):
assert False
print(filename)
model.load_state_dict(torch.load(filename))
model.eval()
stats = {'valid_losses': list(), 'valid_Accuracy': list()}
inversefed.training.training_routine.validate(model,
loss_fn,
validloader,
defs,
setup=setup,
stats=stats)
print(stats)
def main():
setup = inversefed.utils.system_startup()
defs = inversefed.training_strategy('conservative')
defs.epochs = opt.epochs
loss_fn, trainloader, validloader = preprocess(opt, defs)
# init model
model = create_model(opt)
model.to(**setup)
save_dir = create_save_dir()
if not os.path.exists(save_dir):
os.makedirs(save_dir)
file = f'{save_dir}/{arch}_{defs.epochs}.pth'
inversefed.train(model,
loss_fn,
trainloader,
validloader,
defs,
setup=setup,
save_dir=save_dir)
torch.save(model.state_dict(), f'{file}')
model.eval()
import numpy as np
from PIL import Image
import inversefed
torch.backends.cudnn.benchmark = inversefed.consts.BENCHMARK
from collections import defaultdict
import datetime
import time
import os
# Parse input arguments
args = inversefed.options().parse_args()
# Parse training strategy
defs = inversefed.training_strategy('conservative')
defs.epochs = args.epochs
# 100% reproducibility?
if args.deterministic:
image2graph2vec.utils.set_deterministic()
def boost_salient_filters(input_gradient, boost_const, salient_filters):
for idx in range(len(input_gradient)):
if len(input_gradient[idx].size()) == 4:
salient_filters -= input_gradient[idx].size()[0]
for ind in range(len(salient_filters)):
if salient_filters[ind] < 0:
input_gradient[idx][salient_filters[ind]] = input_gradient[
idx][salient_filters[ind]] * boost_const
parser.add_argument('--aug_list', default=None, required=True, type=str, help='Vision model.')
parser.add_argument('--optim', default=None, required=True, type=str, help='Vision model.')
parser.add_argument('--mode', default=None, required=True, type=str, help='Mode.')
parser.add_argument('--rlabel', default=False, type=bool, help='rlabel')
parser.add_argument('--arch', default=None, required=True, type=str, help='Vision model.')
parser.add_argument('--data', default=None, required=True, type=str, help='Vision dataset.')
parser.add_argument('--epochs', default=None, required=True, type=int, help='Vision epoch.')
parser.add_argument('--resume', default=0, type=int, help='rlabel')
opt = parser.parse_args()
num_images = 1
# init env
setup = inversefed.utils.system_startup()
defs = inversefed.training_strategy('conservative'); defs.epochs = opt.epochs
# init training
arch = opt.arch
trained_model = True
mode = opt.mode
assert mode in ['normal', 'aug', 'crop']
config = create_config(opt)
def create_save_dir():
return 'benchmark/images/data_{}_arch_{}_epoch_{}_optim_{}_mode_{}_auglist_{}_rlabel_{}'.format(opt.data, opt.arch, opt.epochs, opt.optim, opt.mode, \
opt.aug_list, opt.rlabel)