def attack_analysis(parameters, n, a_start, a_type):
N = np.shape(parameter.X_test)[0]
a_index = random.randint(0, 1)
a_end = a_start / 100
#print("###########")
tmp_array = np.copy(parameter.X_test)
#print (parameter.X_test[0:2, 1:2])
if (a_type == "additive"):
parameter.X_test[a_index:a_index + n +
1, :] = parameter.X_test[a_index:a_index + n +
1, :] + random.uniform(
a_start, a_end)
else:
parameter.X_test[a_index:a_index + n +
1, :] = parameter.X_test[a_index:a_index + n +
1, :] - random.uniform(
a_start, a_end)
rmse_error, average_error = predict_function.predict(
parameter.X_test, parameter.Y_test, parameters)
parameter.X_test = np.copy(tmp_array)
#print (parameter.X_test[0:2, 1:2])
return rmse_error, average_error
def diff_attack_analysis(parameters, n, a_start, a_type):
N = np.shape(parameter.X_test)[0]
#a_index = random.randint(0, N + 1 - n)
a_index = random.randint(0, 1)
a_end = a_start + a_start / 100
tmp_array = np.copy(parameter.X_test)
parameter.X_test[a_index:a_index + n +
1:2, :] = parameter.X_test[a_index:a_index + n +
1:2, :] + random.uniform(
a_start, a_end)
parameter.X_test[a_index + 1:a_index + n +
1:2, :] = parameter.X_test[a_index + 1:a_index + n +
1:2, :] - random.uniform(
a_start, a_end)
rmse_error, average_error = predict_function.predict(
parameter.X_test, parameter.Y_test, parameters)
parameter.X_test = np.copy(tmp_array)
return rmse_error, average_error
from predict_function import pred_args, read_json, load_checkpoint, predict import matplotlib.pyplot as plt # get parameters pred_arg = pred_args() # load a trained model model = load_checkpoint(pred_arg.checkpoint, pred_arg.gpu) # predict flower predict(pred_arg.image_path, model, pred_arg.gpu, pred_arg.top_k)
def main():
#parse arguments
config.parse()
args = config.args
for k, v in vars(args).items():
logger.info(f"{k}:{v}")
#set seeds
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed_all(args.random_seed)
np.random.seed(args.random_seed)
random.seed(args.random_seed)
#arguments check
device, n_gpu = args_check(args)
os.makedirs(args.output_dir, exist_ok=True)
forward_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
args.forward_batch_size = forward_batch_size
#load config
teachers_and_student = parse_model_config(args.model_config_json)
#Prepare GLUE task
processor = processors[args.task_name]()
args.output_mode = output_modes[args.task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
#read data
train_dataset = None
eval_datasets = None
num_train_steps = None
tokenizer_S = teachers_and_student['student']['tokenizer']
prefix_S = teachers_and_student['student']['prefix']
if args.do_train:
train_dataset = load_and_cache_examples(args,
args.task_name,
tokenizer_S,
prefix=prefix_S,
evaluate=False)
if args.do_predict:
eval_datasets = []
eval_task_names = ("mnli",
"mnli-mm") if args.task_name == "mnli" else (
args.task_name, )
for eval_task in eval_task_names:
eval_datasets.append(
load_and_cache_examples(args,
eval_task,
tokenizer_S,
prefix=prefix_S,
evaluate=True))
logger.info("Data loaded")
#Build Model and load checkpoint
if args.do_train:
model_Ts = []
for teacher in teachers_and_student['teachers']:
model_type_T = teacher['model_type']
model_config_T = teacher['config']
checkpoint_T = teacher['checkpoint']
_, _, model_class_T = MODEL_CLASSES[model_type_T]
model_T = model_class_T(model_config_T, num_labels=num_labels)
state_dict_T = torch.load(checkpoint_T, map_location='cpu')
missing_keys, un_keys = model_T.load_state_dict(state_dict_T,
strict=True)
logger.info(f"Teacher Model {model_type_T} loaded")
model_T.to(device)
model_T.eval()
model_Ts.append(model_T)
student = teachers_and_student['student']
model_type_S = student['model_type']
model_config_S = student['config']
checkpoint_S = student['checkpoint']
_, _, model_class_S = MODEL_CLASSES[model_type_S]
model_S = model_class_S(model_config_S, num_labels=num_labels)
if checkpoint_S is not None:
state_dict_S = torch.load(checkpoint_S, map_location='cpu')
missing_keys, un_keys = model_S.load_state_dict(state_dict_S,
strict=False)
logger.info(f"missing keys:{missing_keys}")
logger.info(f"unexpected keys:{un_keys}")
else:
logger.warning("Initializing student randomly")
logger.info("Student Model loaded")
model_S.to(device)
if args.local_rank != -1 or n_gpu > 1:
if args.local_rank != -1:
raise NotImplementedError
elif n_gpu > 1:
if args.do_train:
model_Ts = [
torch.nn.DataParallel(model_T) for model_T in model_Ts
]
model_S = torch.nn.DataParallel(model_S) #,output_device=n_gpu-1)
if args.do_train:
#parameters
params = list(model_S.named_parameters())
all_trainable_params = divide_parameters(params, lr=args.learning_rate)
logger.info("Length of all_trainable_params: %d",
len(all_trainable_params))
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
raise NotImplementedError
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.forward_batch_size,
drop_last=True)
num_train_steps = int(
len(train_dataloader) // args.gradient_accumulation_steps *
args.num_train_epochs)
########## DISTILLATION ###########
train_config = TrainingConfig(
gradient_accumulation_steps=args.gradient_accumulation_steps,
ckpt_frequency=args.ckpt_frequency,
log_dir=args.output_dir,
output_dir=args.output_dir,
fp16=args.fp16,
device=args.device)
distill_config = DistillationConfig(temperature=args.temperature,
kd_loss_type='ce')
logger.info(f"{train_config}")
logger.info(f"{distill_config}")
adaptor_T = BertForGLUESimpleAdaptor
adaptor_S = BertForGLUESimpleAdaptor
distiller = MultiTeacherDistiller(train_config=train_config,
distill_config=distill_config,
model_T=model_Ts,
model_S=model_S,
adaptor_T=adaptor_T,
adaptor_S=adaptor_S)
optimizer = AdamW(all_trainable_params, lr=args.learning_rate)
scheduler_class = get_linear_schedule_with_warmup
scheduler_args = {
'num_warmup_steps': int(args.warmup_proportion * num_train_steps),
'num_training_steps': num_train_steps
}
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Forward batch size = %d", forward_batch_size)
logger.info(" Num backward steps = %d", num_train_steps)
callback_func = partial(predict,
eval_datasets=eval_datasets,
args=args)
with distiller:
distiller.train(optimizer,
scheduler_class=scheduler_class,
scheduler_args=scheduler_args,
dataloader=train_dataloader,
num_epochs=args.num_train_epochs,
callback=callback_func,
max_grad_norm=1)
if not args.do_train and args.do_predict:
res = predict(model_S, eval_datasets, step=0, args=args)
print(res)
print('Cost at initial theta (zeros): ', cost)
print('Gradient at initial theta (zeros): ')
print(grad)
result = opt.fmin_tnc(func=costfunction,
x0=initial_theta,
fprime=gradient,
args=(X2, Y))
theta = result[0]
cost = costfunction(theta, X2, Y)
print('Cost at theta found by fmin_tnc: ', cost)
print('Theta: ')
print(theta)
#Prediction
p = predict(theta, X2)
print('Prediction: ')
print(p)
#Accuracy
correct = [
1 if ((a == 1 and b == 1) or (a == 0 and b == 0)) else 0
for (a, b) in zip(p, Y)
]
accuracy = (sum(map(int, correct)) / len(correct))
print('Train Accuracy: {0}%'.format(accuracy * 100))
#Confusion Matrix
Y_predict = predict(theta, X2_test)
conf_metrics = metrics.confusion_matrix(Y_test, Y_predict)
print('Confusion Metrics: ')
