def test_modelkey(self):
self.args.modelkey = 3213894809
model = forest.Victim(self.args, self.defs, setup=self.setup)
model.initialize(self.args.modelkey)
self.assertAlmostEqual(model.model.linear.weight[0][0].item(),
0.007770763710141182,
places=5)
if torch.distributed.get_rank() == 0:
print('Currently evaluating -------------------------------:')
print(datetime.datetime.now().strftime("%A, %d. %B %Y %I:%M%p"))
print(args)
print(repr(defs))
print(
f'CPUs: {torch.get_num_threads()}, GPUs: {torch.cuda.device_count()} on {socket.gethostname()}'
)
print(
f'Ensemble launched on {torch.distributed.get_world_size()} GPUs'
f' with backend {forest.consts.DISTRIBUTED_BACKEND}.')
if torch.cuda.is_available():
print(f'GPU : {torch.cuda.get_device_name(device=device)}')
model = forest.Victim(args, setup=setup)
data = forest.Kettle(args,
model.defs.batch_size,
model.defs.augmentations,
setup=setup)
witch = forest.Witch(args, setup=setup)
start_time = time.time()
if args.pretrained:
print('Loading pretrained model...')
stats_clean = None
else:
stats_clean = model.train(data, max_epoch=args.max_epoch)
train_time = time.time()
poison_delta = witch.brew(model, data)
targets.append(target)
indices.append(idx)
for enum, (img, target, idx) in enumerate(data.targetset):
targets.append(target)
indices.append('target')
img = img.unsqueeze(0).to(**data.setup)
f = feature_extractor(img).detach().cpu().numpy()
feats = np.append(feats, f, axis=0)
return feats, targets, indices
if __name__ == "__main__":
setup = forest.utils.system_startup(args)
model = forest.Victim(args, setup=setup)
data = forest.Kettle(args,
model.defs.batch_size,
model.defs.augmentations,
model.defs.mixing_method,
setup=setup)
witch = forest.Witch(args, setup=setup)
witch.patch_targets(data)
start_time = time.time()
if args.pretrained_model:
print('Loading pretrained model...')
stats_clean = None
elif args.skip_clean_training:
print('Skipping clean training...')
stats_clean = None