def test_one_dataset(params, file_name, test_q_data, test_qa_data, best_epoch):
print "\n\nStart testing ......................\n Best epoch:", best_epoch
g_model = MODEL(n_question=params.n_question,
seqlen=params.seqlen,
batch_size=params.batch_size,
q_embed_dim=params.q_embed_dim,
qa_embed_dim=params.qa_embed_dim,
memory_size=params.memory_size,
memory_key_state_dim=params.memory_key_state_dim,
memory_value_state_dim=params.memory_value_state_dim,
final_fc_dim=params.final_fc_dim)
# create a module by given a Symbol
test_net = mx.mod.Module(symbol=g_model.sym_gen(),
data_names=['q_data', 'qa_data'],
label_names=['target'],
context=params.ctx)
# cresate memory by given input shapes
test_net.bind(data_shapes=[
mx.io.DataDesc(name='q_data', shape=(params.seqlen, params.batch_size), layout='SN'),
mx.io.DataDesc(name='qa_data', shape=(params.seqlen, params.batch_size), layout='SN')],
label_shapes=[mx.io.DataDesc(name='target', shape=(params.seqlen, params.batch_size), layout='SN')])
arg_params, aux_params = load_params(prefix=os.path.join('model', params.load, file_name),
epoch=best_epoch)
test_net.init_params(arg_params=arg_params, aux_params=aux_params,
allow_missing=False)
test_loss, test_accuracy, test_auc = test(test_net, params, test_q_data, test_qa_data, label='Test')
print "\ntest_auc\t", test_auc
print "test_accuracy\t", test_accuracy
print "test_loss\t", test_loss
class Trainer:
def __init__(self):
self.model = MODEL()
self.transform = transforms.Compose([
transforms.Resize((100, 100), interpolation=3),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
self.optimizer = torch.optim.Adam(self.model.parameters(), lr=0.0001)
self.logsoftmax = nn.LogSoftmax(dim=1)
def preprocess(self, img):
return self.transform(img)
def loss(
self,
output,
label,
epsilon=1e-6,
):
log_probs = self.logsoftmax(output)
targets = torch.zeros(log_probs.size()).scatter_(
1,
label.unsqueeze(0).unsqueeze(1).data, 1)
targets = (1 - epsilon) * targets + epsilon / 2
loss = (-targets * log_probs).mean(0).sum()
return loss
def train(self, x, label):
self.model.train()
x = self.preprocess(x).unsqueeze(0)
x, label = Variable(x), Variable(torch.tensor(label))
output = self.model(x)
loss = self.loss(output, label)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
return loss.item()
def loss(self, output, label, epsilon=1e-6):
log_probs = self.logsoftmax(output)
targets = torch.zeros(log_probs.size()).scatter_(
1,
label.unsqueeze(0).unsqueeze(1).data, 1)
targets = (1 - epsilon) * targets + epsilon / 2
loss = (-targets * log_probs).mean(0).sum()
return loss
def evaluate(self, x):
self.model.eval()
x = self.preprocess(x).unsqueeze(0)
x = Variable(x)
output = self.model(x)
return output.argmax().item()
def package(self, path):
state_dict = self.model.state_dict()
# result = json.dumps(state_dict)
torch.save(state_dict, path)
def load_package(self, path):
checkpoint = torch.load(path)
self.model.load_state_dict(checkpoint, strict=True)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', type=int, default=-1, help='the gpu will be used, e.g "0,1,2,3"')
parser.add_argument('--max_iter', type=int, default=10, help='number of iterations')
parser.add_argument('--decay_epoch', type=int, default=20, help='number of iterations')
parser.add_argument('--test', type=bool, default=False, help='enable testing')
parser.add_argument('--train_test', type=bool, default=True, help='enable testing')
parser.add_argument('--show', type=bool, default=True, help='print progress')
parser.add_argument('--init_std', type=float, default=0.1, help='weight initialization std')
parser.add_argument('--init_lr', type=float, default=0.01, help='initial learning rate')
parser.add_argument('--lr_decay', type=float, default=0.75, help='learning rate decay')
parser.add_argument('--final_lr', type=float, default=1E-5,
help='learning rate will not decrease after hitting this threshold')
parser.add_argument('--momentum', type=float, default=0.9, help='momentum rate')
parser.add_argument('--max_grad_norm', type=float, default=3.0, help='maximum gradient norm')
parser.add_argument('--hidden_dim', type=int, default=128, help='hidden layer dimension')
parser.add_argument('--n_hidden', type=int, default=2, help='hidden numbers')
dataset = 'oj'
if dataset == 'oj':
parser.add_argument('--batch_size', type=int, default=5, help='the batch size')
parser.add_argument('--qa_embed_dim', type=int, default=50, help='answer and question embedding dimensions')
parser.add_argument('--n_question', type=int, default=68, help='the number of unique questions in the dataset')
parser.add_argument('--seqlen', type=int, default=200, help='the allowed maximum length of a sequence')
parser.add_argument('--data_dir', type=str, default='./data/oj', help='data directory')
parser.add_argument('--data_name', type=str, default='oj', help='data set name')
parser.add_argument('--load', type=str, default='oj', help='model file to load')
parser.add_argument('--save', type=str, default='oj', help='path to save model')
elif dataset == 'assistments':
parser.add_argument('--batch_size', type=int, default=32, help='the batch size')
parser.add_argument('--qa_embed_dim', type=int, default=200, help='answer and question embedding dimensions')
parser.add_argument('--n_question', type=int, default=124, help='the number of unique questions in the dataset')
parser.add_argument('--seqlen', type=int, default=200, help='the allowed maximum length of a sequence')
parser.add_argument('--data_dir', type=str, default='./data/assistments', help='data directory')
parser.add_argument('--data_name', type=str, default='assistments', help='data set name')
parser.add_argument('--load', type=str, default='assistments', help='model file to load')
parser.add_argument('--save', type=str, default='assistments', help='path to save model')
elif dataset == 'STATICS':
parser.add_argument('--batch_size', type=int, default=10, help='the batch size')
parser.add_argument('--qa_embed_dim', type=int, default=100, help='answer and question embedding dimensions')
parser.add_argument('--n_question', type=int, default=1223,
help='the number of unique questions in the dataset')
parser.add_argument('--seqlen', type=int, default=800, help='the allowed maximum length of a sequence')
parser.add_argument('--data_dir', type=str, default='./data/STATICS', help='data directory')
parser.add_argument('--data_name', type=str, default='STATICS', help='data set name')
parser.add_argument('--load', type=str, default='STATICS', help='model file to load')
parser.add_argument('--save', type=str, default='STATICS', help='path to save model')
params = parser.parse_args()
params.lr = params.init_lr
print(params)
dat = DATA(n_question=params.n_question, seqlen=params.seqlen, separate_char=',')
# train_data_path = params.data_dir + "/" + "builder_train.csv"
# valid_data_path = params.data_dir + "/" + "builder_test.csv"
train_data_path = params.data_dir + "/" + params.data_name + "_train.csv"
valid_data_path = params.data_dir + "/" + params.data_name + "_valid.csv"
# test_data_path = params.data_dir + "/" + params.data_name + "_test.csv"
train_q_data, train_q_t_data, train_answer_data = dat.load_data(train_data_path)
valid_q_data, valid_q_t_data, valid_answer_data = dat.load_data(valid_data_path)
# test_q_data, test_q_t_data, test_answer_data = dat.load_data(test_data_path)
model = MODEL(n_question=params.n_question,
hidden_dim=params.hidden_dim,
x_embed_dim=params.qa_embed_dim,
hidden_layers=params.n_hidden,
gpu=params.gpu)
model.init_embeddings()
model.init_params()
# model = torch.load(params.data_dir + "/save/"+params.save)
# optimizer = optim.SGD(params=model.parameters(), lr=params.lr, momentum=params.momentum)
optimizer = optim.Adam(params=model.parameters(), lr=params.lr, betas=(0.9, 0.9))
if params.gpu >= 0:
print('device: ' + str(params.gpu))
torch.cuda.set_device(params.gpu)
model.cuda()
all_train_loss = {}
all_train_accuracy = {}
all_train_auc = {}
all_valid_loss = {}
all_valid_accuracy = {}
all_valid_auc = {}
best_valid_auc = 0
for idx in range(params.max_iter):
train_loss, train_accuracy, train_auc = train(model, idx, params, optimizer, train_q_data, train_q_t_data,
train_answer_data)
print('Epoch %d/%d, loss : %3.5f, auc : %3.5f, accuracy : %3.5f' % (
idx + 1, params.max_iter, train_loss, train_auc, train_accuracy))
valid_loss, valid_accuracy, valid_auc = test(model, params, optimizer, valid_q_data, valid_q_t_data,
valid_answer_data)
print('Epoch %d/%d, valid auc : %3.5f, valid accuracy : %3.5f' % (
idx + 1, params.max_iter, valid_auc, valid_accuracy))
# test_loss, test_accuracy, test_auc = test(model, params, optimizer, test_q_data, test_q_t_data,
# test_answer_data)
# print('Epoch %d/%d, test auc : %3.5f, test accuracy : %3.5f' % (
# idx + 1, params.max_iter, test_auc, test_accuracy))
all_train_auc[idx + 1] = train_auc
all_train_accuracy[idx + 1] = train_accuracy
all_train_loss[idx + 1] = train_loss
all_valid_loss[idx + 1] = valid_loss
all_valid_accuracy[idx + 1] = valid_accuracy
all_valid_auc[idx + 1] = valid_auc
#
# output the epoch with the best validation auc
if valid_auc > best_valid_auc:
print('%3.4f to %3.4f' % (best_valid_auc, valid_auc))
best_valid_auc = valid_auc
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--gpu',
type=int,
default=0,
help='the gpu will be used, e.g "0,1,2,3"')
parser.add_argument('--max_iter',
type=int,
default=30,
help='number of iterations')
parser.add_argument('--decay_epoch',
type=int,
default=20,
help='number of iterations')
parser.add_argument('--test',
type=bool,
default=False,
help='enable testing')
parser.add_argument('--train_test',
type=bool,
default=True,
help='enable testing')
parser.add_argument('--show',
type=bool,
default=True,
help='print progress')
parser.add_argument('--init_std',
type=float,
default=0.1,
help='weight initialization std')
parser.add_argument('--init_lr',
type=float,
default=0.01,
help='initial learning rate')
parser.add_argument('--lr_decay',
type=float,
default=0.75,
help='learning rate decay')
parser.add_argument(
'--final_lr',
type=float,
default=1E-5,
help='learning rate will not decrease after hitting this threshold')
parser.add_argument('--momentum',
type=float,
default=0.9,
help='momentum rate')
parser.add_argument('--maxgradnorm',
type=float,
default=50.0,
help='maximum gradient norm')
parser.add_argument('--final_fc_dim',
type=float,
default=50,
help='hidden state dim for final fc layer')
dataset = 'assist2009_updated'
if dataset == 'assist2009_updated':
parser.add_argument('--q_embed_dim',
type=int,
default=50,
help='question embedding dimensions')
parser.add_argument('--batch_size',
type=int,
default=32,
help='the batch size')
parser.add_argument('--qa_embed_dim',
type=int,
default=200,
help='answer and question embedding dimensions')
parser.add_argument('--memory_size',
type=int,
default=20,
help='memory size')
parser.add_argument(
'--n_question',
type=int,
default=110,
help='the number of unique questions in the dataset')
parser.add_argument('--seqlen',
type=int,
default=200,
help='the allowed maximum length of a sequence')
parser.add_argument('--data_dir',
type=str,
default='./data/assist2009_updated',
help='data directory')
parser.add_argument('--data_name',
type=str,
default='assist2009_updated',
help='data set name')
parser.add_argument('--load',
type=str,
default='data/assist2009_updated',
help='model file to load')
parser.add_argument('--save',
type=str,
default='data/assist2009_updated/model',
help='path to save model')
elif dataset == 'STATICS':
parser.add_argument('--batch_size',
type=int,
default=10,
help='the batch size')
parser.add_argument('--q_embed_dim',
type=int,
default=50,
help='question embedding dimensions')
parser.add_argument('--qa_embed_dim',
type=int,
default=100,
help='answer and question embedding dimensions')
parser.add_argument('--memory_size',
type=int,
default=50,
help='memory size')
parser.add_argument(
'--n_question',
type=int,
default=1223,
help='the number of unique questions in the dataset')
parser.add_argument('--seqlen',
type=int,
default=6,
help='the allowed maximum length of a sequence')
parser.add_argument('--data_dir',
type=str,
default='./data/STATICS',
help='data directory')
parser.add_argument('--data_name',
type=str,
default='STATICS',
help='data set name')
parser.add_argument('--load',
type=str,
default='STATICS',
help='model file to load')
parser.add_argument('--save',
type=str,
default='STATICS',
help='path to save model')
params = parser.parse_args()
params.lr = params.init_lr
params.memory_key_state_dim = params.q_embed_dim # 50
params.memory_value_state_dim = params.qa_embed_dim # 200
print(params)
dat = DATA(n_question=params.n_question,
seqlen=params.seqlen,
separate_char=',')
# train_data_path = params.data_dir + "/" + "test5.1.txt"
train_data_path = params.data_dir + "/" + params.data_name + "_train2.csv"
valid_data_path = params.data_dir + "/" + params.data_name + "_valid2.csv"
test_data_path = params.data_dir + "/" + params.data_name + "_test.csv"
train_q_data, train_qa_data, _ = dat.load_data(train_data_path)
valid_q_data, valid_qa_data, _ = dat.load_data(valid_data_path)
test_q_data, test_qa_data, _ = dat.load_data(test_data_path)
params.memory_key_state_dim = params.q_embed_dim # 记忆key = 问题embedding的维度
params.memory_value_state_dim = params.qa_embed_dim # 记忆val = 回ç”embedding的维度
model = MODEL(n_question=params.n_question,
batch_size=params.batch_size,
q_embed_dim=params.q_embed_dim,
qa_embed_dim=params.qa_embed_dim,
memory_size=params.memory_size,
memory_key_state_dim=params.memory_key_state_dim,
memory_value_state_dim=params.memory_value_state_dim,
final_fc_dim=params.final_fc_dim)
model.init_embeddings()
model.init_params()
# optimizer = optim.SGD(params=model.parameters(), lr=params.lr, momentum=params.momentum)
optimizer = optim.Adam(params=model.parameters(),
lr=params.lr,
betas=(0.9, 0.9))
if params.gpu >= 0:
print('device: ' + str(params.gpu))
torch.cuda.set_device(params.gpu)
model.cuda()
all_train_loss = {}
all_train_accuracy = {}
all_train_auc = {}
all_valid_loss = {}
all_valid_accuracy = {}
all_valid_auc = {}
best_valid_auc = 0
train_shuffle_index = np.random.permutation(train_q_data.shape[0])
valid_shuffie_index = np.random.permutation(valid_q_data.shape[0])
train_q_data_shuffled = train_q_data[train_shuffle_index]
train_qa_data_shuffled = train_qa_data[train_shuffle_index]
valid_q_data_shuffled = valid_q_data[valid_shuffie_index]
valid_qa_data_shuffled = valid_qa_data[valid_shuffie_index]
train_loss_list, train_acc_list, train_auc_list, train_f1_list = [], [], [], []
valid_loss_list, valid_acc_list, valid_auc_list, valid_f1_list = [], [], [], []
for idx in range(params.max_iter):
# train_loss, train_accuracy, train_auc = train(idx, model, params, optimizer, train_q_data, train_qa_data)
train_loss, train_accuracy, train_auc, train_f1 = train(
idx, model, params, optimizer, train_q_data_shuffled,
train_qa_data_shuffled)
train_loss_list.append(train_loss)
train_acc_list.append(train_accuracy)
train_auc_list.append(train_auc)
train_f1_list.append(train_f1)
print(
'Epoch %d/%d, loss : %3.5f, auc : %3.5f, accuracy : %3.5f, f1 : %.4f'
% (idx + 1, params.max_iter, train_loss, train_auc, train_accuracy,
train_f1))
# valid_loss, valid_accuracy, valid_auc = test(model, params, optimizer, valid_q_data, valid_qa_data)
valid_loss, valid_accuracy, valid_auc, valid_f1 = test(
model, params, optimizer, valid_q_data_shuffled,
valid_qa_data_shuffled)
valid_loss_list.append(valid_loss)
valid_acc_list.append(valid_accuracy)
valid_auc_list.append(valid_auc)
valid_f1_list.append(valid_f1)
print(
'Epoch %d/%d, valid auc : %3.5f, valid accuracy : %3.5f, f1 : %.4f'
% (idx + 1, params.max_iter, valid_auc, valid_accuracy, valid_f1))
all_train_auc[idx + 1] = train_auc
all_train_accuracy[idx + 1] = train_accuracy
all_train_loss[idx + 1] = train_loss
all_valid_loss[idx + 1] = valid_loss
all_valid_accuracy[idx + 1] = valid_accuracy
all_valid_auc[idx + 1] = valid_auc
#
# output the epoch with the best validation auc
generate_dir(params.save)
if valid_auc > best_valid_auc:
print('%3.4f to %3.4f' % (best_valid_auc, valid_auc))
best_valid_auc = valid_auc
best_epoch = idx + 1
best_valid_acc = valid_accuracy
best_valid_loss = valid_loss
test_loss, test_accuracy, test_auc, test_f1 = test(
model, params, optimizer, test_q_data, test_qa_data)
print(
"test_auc: %.4f\ttest_accuracy: %.4f\ttest_loss: %.4f\ttest_f1: %.4f"
% (test_auc, test_accuracy, test_loss, test_f1))
# save_checkpoint(model, memory, params.save + "/Epoch%d-test_auc%.2f-val_auc%.2f-loss%.2f.pt"%(best_epoch, test_auc, valid_auc, test_loss))
save_path = params.save + "/Epoch%d-test_auc%.2f-val_auc%.2f-loss%.2f.pt" % (
best_epoch, test_auc, valid_auc, test_loss)
torch.save(model, save_path)
print(save_path + " save to " + params.save)
print("best outcome: best epoch: %.4f" % (best_epoch))
os.system(f"mv {save_path} {params.save}/best.pt")
print("valid_auc: %.4f\tvalid_accuracy: %.4f\tvalid_loss: %.4f\t" %
(best_valid_auc, best_valid_acc, best_valid_loss))
print("test_auc: %.4f\ttest_accuracy: %.4f\ttest_loss: %.4f\t" %
(test_auc, test_accuracy, test_loss))
print(train_loss_list, train_acc_list, train_auc_list, train_f1_list,
valid_loss_list, valid_acc_list, valid_auc_list, valid_f1_list)
train_log = 'log.txt'
with open(train_log, "w") as f:
f.write("train_loss_list================")
f.write(str(train_loss_list))
f.write("train_acc_list================")
f.write(str(train_acc_list))
f.write("train_auc_list================")
f.write(str(train_auc_list))
f.write("train_f1_lis================")
f.write(str(train_f1_list))
f.write("valid_loss_list================")
f.write(str(valid_loss_list))
f.write("valid_acc_list================")
f.write(str(valid_acc_list))
f.write("================")
f.write(str(valid_auc_list))
f.write("================")
f.write(str(valid_f1_list))
np.save("train_loss_list.npy", np.array(train_loss_list))
np.save("train_acc_list.npy", np.array(train_acc_list))
np.save("train_auc_list.npy", np.array(train_auc_list))
np.save("train_f1_list.npy", np.array(train_f1_list))
np.save("valid_loss_list.npy", np.array(valid_loss_list))
np.save("valid_acc_list.npy", np.array(valid_acc_list))
np.save("valid_auc_list.npy", np.array(valid_auc_list))
np.save("valid_f1_list.npy", np.array(valid_f1_list))
def train(args):
log_file = open('../data/logs/log.txt', 'a')
log_file.write('\n' + '\n')
log_file.write('-------------------------------------------\n')
log_file.write(str(args) + '\n')
log_file.write('-------------------------------------------\n')
print('model initializing..')
model = MODEL(args)
# CUDA_VISIBLE_DEVICES=args.GPU_DEVICE
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
saver = tf.train.Saver(write_version=tf.train.SaverDef.V2,
max_to_keep=10)
print('loading data..')
load_file = args.file_directory + str(args.first_train_file)
train_data = load_data(args, load_file)
load_file = args.file_directory + str(args.first_test_file)
test_data = load_data(args, load_file)
for step in range(args.n_epochs):
# training
print('epoch:' + str(step))
args.is_train = True
if step is not 0:
for i in range(args.train_file_num):
start_list = list(
range(0, train_data.size, args.batch_size))
np.random.shuffle(start_list)
for start in start_list:
end = start + args.batch_size
model.train(
sess, get_feed_dict(model, train_data, start, end))
args.is_train = False
train_loss = 0
train_time_loss = 0
train_batchs = 0
train_labels = []
train_scores = []
train_time_true = []
train_time_pre = []
for i in range(args.train_file_num):
#load_file=args.file_directory+str(args.first_train_file+i)
#train_data=load_data(args,load_file)
start_list = list(
range(0, train_data.size, args.batch_size * 10))
train_batchs = train_batchs + len(start_list)
for start in start_list:
end = start + args.batch_size * 10
loss, labels, scores, time_loss, time_pre, time_true, time_label = model.test(
sess, get_feed_dict(model, train_data, start, end))
train_labels.extend(labels)
train_scores.extend(scores)
for k in range(len(time_label)):
if time_label[k] == 0:
continue
train_time_true.append(time_true[k])
train_time_pre.append(time_pre[k])
train_loss = train_loss + loss
train_time_loss = train_time_loss + time_loss
train_auc, train_f1, train_pre, train_recall, train_time_f1 = model.eval(
args, train_labels, train_scores, train_time_true,
train_time_pre)
train_loss = train_loss / train_batchs
train_time_loss = train_time_loss / train_batchs
time_stamp = datetime.datetime.now().strftime('%Y.%m.%d-%H:%M:%S')
log_str = 'time:' + time_stamp + ' epoch %d \ntrain_loss:%.4f train_time_loss:%.4f train_time_f1:%.4f train_auc:%.4f train_f1_score:%.4f'\
% (step, train_loss, train_time_loss, train_time_f1, train_auc, train_f1)
print(log_str)
log_file.write(log_str)
log_file.write('\n')
log_file.close()
def run_ES(parameters, model: MODEL, utils, root):
n_step = parameters['n_step']
n_protocol = 2**n_step
exact_data = np.zeros((n_protocol, 2),
dtype=np.float64) # 15 digits precision
b2_array = lambda n10: np.array(list(np.binary_repr(n10, width=n_step)),
dtype=np.int)
st = time.time()
# ---> measuring estimated time <---
model.update_protocol(b2_array(0))
psi = model.compute_evolved_state()
model.compute_fidelity(psi_evolve=psi)
model.compute_energy(psi_evolve=psi)
print("Est. run time : \t %.3f s" % (0.5 * n_protocol *
(time.time() - st)))
# ---> Starting real calculation <---
st = time.time()
for p in range(n_protocol):
model.update_protocol(b2_array(p))
psi = model.compute_evolved_state()
exact_data[p] = (model.compute_fidelity(psi_evolve=psi),
model.compute_energy(psi_evolve=psi))
outfile = utils.make_file_name(parameters, root=root)
with open(outfile, 'wb') as f:
pickle.dump(exact_data, f, protocol=4)
print("Total run time : \t %.3f s" % (time.time() - st))
print("\n Thank you and goodbye !")
f.close()
def SD_2SF(param, model: MODEL, init=False):
if model.n_h_field > 2:
assert False, 'This works only for bang-bang protocols'
n_step = param['n_step']
n_fid_eval = 0
if init:
# Random initialization
tmp = np.ones(n_step, dtype=int) # m = 0 state ...
tmp[0:n_step // 2] = 0
np.random.shuffle(tmp)
model.update_protocol(tmp)
old_fid = model.compute_fidelity()
best_protocol = np.copy(model.protocol())
else:
# So user can feed in data say from a specific protocol
old_fid = model.compute_fidelity()
best_protocol = np.copy(model.protocol())
x1_ar, x2_ar = np.triu_indices(n_step, 1)
order = np.arange(0, x1_ar.shape[0], dtype=np.int)
while True: # careful with this. For binary actions, this is guaranteed to break
np.random.shuffle(order)
local_minima = True
for pos in order:
t1, t2 = (x1_ar[pos], x2_ar[pos])
model.swap(t1, t2)
new_fid = model.compute_fidelity()
n_fid_eval += 1
if new_fid > old_fid: # accept descent
#print("%.15f"%new_fid,'\t',n_fid_eval)
old_fid = new_fid
best_protocol = np.copy(model.protocol())
local_minima = False # will exit for loop before it ends ... local update accepted
break
else:
model.swap(t1, t2)
if local_minima:
break
return old_fid, best_protocol, n_fid_eval
def run_SD(parameters, model: MODEL, utils, root, save=True):
if parameters['verbose'] == 0:
blockPrint()
outfile = utils.make_file_name(parameters, root=root)
n_exist_sample, all_result = utils.read_current_results(outfile)
n_sample = parameters['n_sample']
if n_exist_sample >= n_sample:
print(
"\n\n-----------> Samples already computed in file -- terminating ... <-----------"
)
return all_result
print("\n\n-----------> Starting stochastic descent <-----------")
n_iteration_left = n_sample - n_exist_sample # data should be saved 10 times --> no more (otherwise things are way too slow !)
n_mod = max([1, n_iteration_left // 10])
for it in range(n_iteration_left):
start_time = time.time()
if parameters['task'] == 'SD':
best_fid, best_protocol, n_fid_eval = SD(
parameters, model,
init=True) # -- --> performing stochastic descent here <-- --
elif parameters['task'] == 'SD2':
best_fid, best_protocol, n_fid_eval = SD_2SF(
parameters, model, init=True
) # -- --> performing 2 spin flip stochastic descent here <-- --
else:
assert False, 'Error in task specification'
energy, delta_energy, Sent = model.compute_observables(
protocol=best_protocol)
result = [
n_fid_eval, best_fid, energy, delta_energy, Sent, best_protocol
]
print(
"\n----------> RESULT FOR STOCHASTIC DESCENT NO %i <-------------"
% (it + 1))
print("Number of fidelity eval \t%i" % n_fid_eval)
print("Best fidelity \t\t\t%.16f" % best_fid)
print("Best hx_protocol\t\t", list(best_protocol))
all_result.append(result)
if save and it % n_mod == 0:
with open(outfile, 'wb') as f:
pickle.dump([parameters, all_result], f)
f.close()
print("Saved iteration --> %i to %s" %
(it + n_exist_sample, outfile))
print("Iteration run time --> %.4f s" % (time.time() - start_time))
print("\n Thank you and goodbye !")
enablePrint()
if save:
with open(outfile, 'wb') as f:
pickle.dump([parameters, all_result], f)
f.close()
return all_result
def run_GRAPE(parameters, model: MODEL, utils, root, save=True):
if parameters['verbose'] == 0:
blockPrint()
outfile = utils.make_file_name(parameters, root=root)
n_exist_sample, optimal_index, all_result = utils.read_current_results(
outfile)
n_sample = parameters['n_sample']
if optimal_index is not None:
best_seen_fid = all_result[optimal_index][1]
else:
best_seen_fid = 0.0
if n_exist_sample >= n_sample:
print(
"\n\n-----------> Samples already computed in file -- terminating ... <-----------"
)
print(
"\n\n-------> Best encountered fidelity over all samples is %0.8f <-------"
% (best_seen_fid))
print("\n\n-------> Best encountered hx_protocol over all samples:")
print(list(all_result[optimal_index][5]))
print("<-------")
return all_result
print("\n\n-----------> Starting stochastic descent <-----------")
n_iteration_left = n_sample - n_exist_sample # data should be saved 10 times --> no more (otherwise things are way too slow !)
n_mod = max([1, n_iteration_left // 10])
for it in range(n_iteration_left):
start_time = time.time()
if parameters['task'] == 'GRAPE':
best_fid, best_protocol, n_fid_eval = GRAPE(
parameters, model,
init=True) # -- --> performing stochastic descent here <-- --
else:
assert False, 'Error in task specification'
energy, delta_energy, Sent = model.compute_observables(
protocol=best_protocol, discrete=False)
result = [
n_fid_eval, best_fid, energy, delta_energy, Sent, best_protocol
]
print("\n----------> RESULT FOR GRAPE NO %i <-------------" % (it + 1))
print("Number of fidelity eval \t%i" % n_fid_eval)
print("Best fidelity \t\t\t%.16f" % best_fid)
print("Best hx_protocol\t\t", list(best_protocol))
all_result.append(result)
# check if a better fidelity has been seen in previous samples
if best_fid > best_seen_fid:
best_seen_fid = best_fid
optimal_index = len(all_result) - 1
if save and it % n_mod == 0:
with open(outfile, 'wb') as f:
pickle.dump([parameters, all_result], f)
f.close()
print("Saved iteration --> %i to %s" %
(it + n_exist_sample, outfile))
print("Iteration run time --> %.4f s" % (time.time() - start_time))
# print best seen fidelity and protocol over all times
print(
"\n\n-------> Best encountered fidelity over all samples is %0.8f <-------"
% (best_seen_fid))
print("\n\n-------> Best encountered hx_protocol over all samples:")
print(list(all_result[optimal_index][5]))
print("<-------")
print("\n Thank you and goodbye !")
enablePrint()
if save:
with open(outfile, 'wb') as f:
pickle.dump([parameters, all_result], f)
f.close()
return all_result
def GRAPE(param, model: MODEL, init=False):
n_step = param['n_step']
n_fid_eval = 0
if init:
# Random initialization
model.update_protocol(
np.random.uniform(-1.0, 1.0, size=param['n_step']))
old_fid = model.compute_fidelity(protocol=model.protocol(),
discrete=False)
best_protocol = np.copy(model.protocol())
else:
# So user can feed in data say from a specific protocol
raise NotImplementedError("yet to be implemented.")
#old_fid = model.compute_fidelity(discrete=False)
#best_protocol = np.copy(model.protocol())
eta = 0.6 # initial learning rate
n_fid_eval = 0
fid_diff = 1.0
while np.abs(fid_diff) > 1E-9 and n_fid_eval < param[
'n_quench']: # guaranteed to break but after very long time
# compute protocol gradient
protocol_gradient = model.compute_protocol_gradient()
# normalise gradient
protocol_gradient /= np.max(np.abs(protocol_gradient))
# GD update rule
new_protocol = model.protocol() + eta * protocol_gradient
# impose boundedness condition
ind_max = np.where(new_protocol > param['hx_max'])
new_protocol[ind_max] = param['hx_max']
ind_min = np.where(new_protocol < param['hx_min'])
new_protocol[ind_min] = param['hx_min']
###
fid = model.compute_fidelity(protocol=new_protocol, discrete=False)
# if we overshoot, decrease step size, otherwise update protocol
fid_diff = fid - old_fid
if fid_diff < 0:
eta *= 0.5
print('overshot minimum:', n_fid_eval, eta, np.abs(fid_diff))
else:
# update protocol
model.update_protocol(new_protocol)
old_fid = fid.copy()
n_fid_eval += 1
print(fid)
print(n_fid_eval, eta, fid_diff)
return old_fid, model.protocol(), n_fid_eval
def train_one_dataset(params, file_name, train_q_data, train_qa_data, valid_q_data, valid_qa_data, valid_tf_data):
### ================================== model initialization ==================================
g_model = MODEL(n_question=params.n_question,
seqlen=params.seqlen,
batch_size=params.batch_size,
q_embed_dim=params.q_embed_dim,
qa_embed_dim=params.qa_embed_dim,
memory_size=params.memory_size,
memory_key_state_dim=params.memory_key_state_dim,
memory_value_state_dim=params.memory_value_state_dim,
final_fc_dim=params.final_fc_dim)
# create a module by given a Symbol
net = mx.mod.Module(symbol=g_model.sym_gen(),
data_names=['q_data', 'qa_data'],
label_names=['target'],
context=params.ctx, )
# create memory by given input shapes
net.bind(data_shapes=[mx.io.DataDesc(name='q_data', shape=(params.seqlen, params.batch_size), layout='SN'),
mx.io.DataDesc(name='qa_data', shape=(params.seqlen, params.batch_size), layout='SN')],
label_shapes=[mx.io.DataDesc(name='target', shape=(params.seqlen, params.batch_size), layout='SN')])
# initial parameters with the default DKVMN initializer
init_dkvmn_param_file_name = params.save + '-dkvmn_initialization'
arg_params, aux_params = load_params(prefix=os.path.join('model', params.load, init_dkvmn_param_file_name), epoch=30)
net.init_params(arg_params=arg_params, aux_params=aux_params, allow_missing=False)
'''
# initial parameters with the default random initializer
net.init_params(initializer=mx.init.Normal(sigma=params.init_std), force_init=True)
'''
# decay learning rate in the lr_scheduler
lr_scheduler = mx.lr_scheduler.FactorScheduler(step=20 * (train_q_data.shape[0] / params.batch_size),
factor=0.667, stop_factor_lr=1e-5)
net.init_optimizer(optimizer='sgd', optimizer_params={'learning_rate': params.lr, 'momentum': params.momentum,
'lr_scheduler': lr_scheduler})
'''
for parameters in net.get_params()[0]:
print(parameters, net.get_params()[0][parameters].asnumpy().shape)
#print(parameters, net.get_params()[0][parameters])
print("\n")
'''
### ================================== start training ==================================
all_train_loss = {}
all_train_acc = {}
all_train_auc = {}
all_valid_loss = {}
all_valid_acc = {}
all_valid_auc = {}
best_valid_acc = -1
best_valid_loss = -1
for idx in range(params.max_iter):
train_loss, train_acc = run.train(net, params, train_q_data, train_qa_data, label='Train')
pred_list, target_list = run.test(net, params, valid_q_data, valid_qa_data, valid_tf_data, label='Valid')
all_pred = np.concatenate(pred_list, axis=0)
all_target = np.concatenate(target_list, axis=0)
valid_loss = run.binaryEntropy(all_target, all_pred)
valid_acc = run.compute_accuracy(all_target, all_pred)
'''
print('epoch', idx + 1)
print("valid_auc\t", valid_auc, "\ttrain_auc\t", train_auc)
print("valid_acc\t", valid_acc, "\ttrain_acc\t", train_acc)
print("valid_loss\t", valid_loss, "\ttrain_loss\t", train_loss)
'''
all_valid_loss[idx + 1] = valid_loss
all_train_loss[idx + 1] = train_loss
all_valid_acc[idx + 1] = valid_acc
all_train_acc[idx + 1] = train_acc
# output the epoch with the best validation auc
if valid_acc > best_valid_acc:
best_valid_acc = valid_acc
best_valid_loss = valid_loss
if not os.path.isdir('model'):
os.makedirs('model')
if not os.path.isdir(os.path.join('model', params.save)):
os.makedirs(os.path.join('model', params.save))
net.save_checkpoint(prefix=os.path.join('model', params.save, file_name), epoch=30)
if not os.path.isdir('result'):
os.makedirs('result')
if not os.path.isdir(os.path.join('result', params.save)):
os.makedirs(os.path.join('result', params.save))
f_save_log = open(os.path.join('result', params.save, file_name), 'w')
f_save_log.write("valid_auc:\n" + str(all_valid_auc) + "\n\n")
f_save_log.write("train_auc:\n" + str(all_train_auc) + "\n\n")
f_save_log.write("valid_loss:\n" + str(all_valid_loss) + "\n\n")
f_save_log.write("train_loss:\n" + str(all_train_loss) + "\n\n")
f_save_log.write("valid_acc:\n" + str(all_valid_acc) + "\n\n")
f_save_log.write("train_acc:\n" + str(all_train_acc) + "\n\n")
f_save_log.close()
return best_valid_acc, best_valid_loss
def train_one_dataset(params, file_name, train_q_data, train_qa_data, valid_q_data, valid_qa_data):
### ================================== model initialization ==================================
g_model = MODEL(n_question=params.n_question,
seqlen=params.seqlen,
batch_size=params.batch_size,
q_embed_dim=params.q_embed_dim,
qa_embed_dim=params.qa_embed_dim,
memory_size=params.memory_size,
memory_key_state_dim=params.memory_key_state_dim,
memory_value_state_dim=params.memory_value_state_dim,
final_fc_dim = params.final_fc_dim)
# create a module by given a Symbol
net = mx.mod.Module(symbol=g_model.sym_gen(),
data_names = ['q_data', 'qa_data'],
label_names = ['target'],
context=params.ctx)
# create memory by given input shapes
net.bind(data_shapes=[mx.io.DataDesc(name='q_data', shape=(params.seqlen, params.batch_size), layout='SN'),
mx.io.DataDesc(name='qa_data', shape=(params.seqlen, params.batch_size), layout='SN')],
label_shapes=[mx.io.DataDesc(name='target', shape=(params.seqlen, params.batch_size), layout='SN')])
# initial parameters with the default random initializer
net.init_params(initializer=mx.init.Normal(sigma=params.init_std))
# decay learning rate in the lr_scheduler
lr_scheduler = mx.lr_scheduler.FactorScheduler(step=20*(train_q_data.shape[0]/params.batch_size), factor=0.667, stop_factor_lr=1e-5)
net.init_optimizer(optimizer='sgd', optimizer_params={'learning_rate': params.lr, 'momentum':params.momentum,'lr_scheduler': lr_scheduler})
for parameters in net.get_params()[0]:
print parameters, net.get_params()[0][parameters].asnumpy().shape
print "\n"
### ================================== start training ==================================
all_train_loss = {}
all_train_accuracy = {}
all_train_auc = {}
all_valid_loss = {}
all_valid_accuracy = {}
all_valid_auc = {}
best_valid_auc = 0
for idx in xrange(params.max_iter):
train_loss, train_accuracy, train_auc = train(net, params, train_q_data, train_qa_data, label='Train')
valid_loss, valid_accuracy, valid_auc = test(net, params, valid_q_data, valid_qa_data, label='Valid')
print 'epoch', idx + 1
print "valid_auc\t", valid_auc, "\ttrain_auc\t", train_auc
print "valid_accuracy\t", valid_accuracy, "\ttrain_accuracy\t", train_accuracy
print "valid_loss\t", valid_loss, "\ttrain_loss\t", train_loss
net.save_checkpoint(prefix=os.path.join('model', params.save, file_name), epoch=idx+1)
# output the epoch with the best validation auc
if valid_auc > best_valid_auc:
best_valid_auc = valid_auc
best_epoch = idx+1
f_save_log = open(os.path.join('result', params.save, file_name), 'w')
f_save_log.write("valid_auc:\n" + str(all_valid_auc) + "\n\n")
f_save_log.write("train_auc:\n" + str(all_train_auc) + "\n\n")
f_save_log.write("valid_loss:\n" + str(all_valid_loss) + "\n\n")
f_save_log.write("train_loss:\n" + str(all_train_loss) + "\n\n")
f_save_log.write("valid_accuracy:\n" + str(all_valid_accuracy) + "\n\n")
f_save_log.write("train_accuracy:\n" + str(all_train_accuracy) + "\n\n")
f_save_log.close()
return best_epoch
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--gpu",
type=int,
default=0,
help="the gpu will be used")
parser.add_argument("--max_iter",
type=int,
default=300,
help="number of iterations")
parser.add_argument("--decay_epoch",
type=int,
default=20,
help="number of iterations")
parser.add_argument("--test",
type=bool,
default=False,
help="enable tesing")
parser.add_argument("--train_test",
type=bool,
default=True,
help="enable testing")
parser.add_argument("-show",
type=bool,
default=True,
help="print progress")
parser.add_argument("--init_std",
type=float,
default=0.1,
help="weight initial std")
parser.add_argument("--init_lr",
type=float,
default=0.01,
help="initial learning rate")
parser.add_argument("--lr_decay",
type=float,
default=0.75,
help="learning rate decay")
parser.add_argument(
"--final_lr",
type=float,
default=1e-5,
help=
"learning rate will not decrease after hitting the threshold final_lr")
parser.add_argument("--momentum",
type=float,
default=0.9,
help="momentum rate")
parser.add_argument("--maxgradnorm",
type=float,
default=50.0,
help="maximum gradient norm")
parser.add_argument("--final_fc_dim",
type=float,
default=50,
help="hidden state dim for final fc layer")
dataset = "assist2009_updated"
if dataset == "assist2009_updated":
parser.add_argument("--q_embed_dim",
type=int,
default=50,
help="question embedding dimensions")
parser.add_argument("--batch_size",
type=int,
default=32,
help="the batch size")
parser.add_argument("--qa_embed_dim",
type=int,
default=200,
help="answer and question embedding dimensions")
parser.add_argument("--memory_size",
type=int,
default=20,
help="memory_size")
parser.add_argument(
"--n_question",
type=int,
default=110,
help="the number of unique questions in the database")
parser.add_argument("--seqlen",
type=int,
default=200,
help="the allowed maximum length of a seqence")
parser.add_argument("--data_dir",
type=str,
default="./data/assist2009_updated")
parser.add_argument("--data_name",
type=str,
default="assist2009_updated")
parser.add_argument("--load",
type=str,
default="assist2009_updated",
help="model file to load")
parser.add_argument("--save",
type=str,
default="assist2009_updated",
help="path to save model")
params = parser.parse_args()
params.lr = params.init_lr
params.memory_key_state_dim = params.q_embed_dim
params.memory_value_state_dim = params.qa_embed_dim
print(params)
dat = Data(params.n_question, params.seqlen, ",")
train_data_path = params.data_dir + "/" + params.data_name + "_train1.csv"
valid_data_path = params.data_dir + "/" + params.data_name + "_valid1.csv"
test_data_path = params.data.dir + "/" + params.data_name + "_test.csv"
train_q_data, train_qa_data = dat.load_data(train_data_path)
valid_q_data, valid_qa_data = dat.load_data(valid_data_path)
test_q_data, test_qa_data = dat.load_data(test_data_path)
model = MODEL(params.n_question, params.batch_size, params.q_embed_dim,
params.qa_embed_dim, params.memory_size,
params.memory_key_state_dim, params.memory_value_state_dim,
params.final_fc_dim)
model.init_embedding()
model.init_params()
optimizer = optim.Adam(params=model.parameters(),
lr=params.lr,
betas=(0.9, 0.9))
if params.gpu >= 0:
print("device: " + str(params.gpu))
torch.cuda.set_device(params.gpu)
model.cuda()
all_train_loss = {}
all_train_accuracy = {}
all_train_auc = {}
all_valid_loss = {}
all_valid_accuracy = {}
all_valid_auc = {}
best_valid_auc = 0
for idx in range(params.max_iter):
train_loss, train_accuracy, train_auc = train(idx, model, params,
optimizer, train_q_data,
train_qa_data)
print(f"Epoch {idx + 1}/{params.max_iter}, loss: ")
valid_loss, valid_accuracy, valid_auc = test(model, params,
valid_q_data,
valid_qa_data)
all_train_auc[idx + 1] = train_auc
all_train_accuracy[idx + 1] = train_accuracy
all_train_loss[idx + 1] = train_loss
all_valid_auc[idx + 1] = valid_auc
all_valid_accuracy[idx + 1] = valid_accuracy
all_valid_loss[idx + 1] = valid_loss
if valid_auc > best_valid_auc:
best_valid_auc = valid_auc
print(f"best_auc: ${best_valid_auc}")
# Rescale data
data.rescale()
# Visualize rescaled data
data.show()
# %% [markdown]
# ## 3. Define the model architecture and learn, validation and test processes.
from model import MODEL
# %% [markdown]
# ## 4. Create the model with hyperparameters and start start training process.
# Instantiate the model
network = \
MODEL(input_size=12, output_size=4, hidden_size=32, num_layers=2, p=0.3)
print(network)
# Execute learning process
model_name = \
network.learn(data=data, batch_size=16, sequence_length=16, lr=0.01, \
num_epochs=100, validate_every=1, max_norm=4)
#%% [markdown]
# ## 5. Read best model from the file.
with open(model_name, 'rb') as f:
# Load checkpoint
checkpoint = torch.load(f)
def SD(param, model: MODEL, init_random=False):
""" Single spin flip stochastic descent
"""
n_step = param['n_step']
n_fid_eval = 1
n_visit = 1
if init_random:
# Random initialization
model.update_protocol(
np.random.randint(0, model.n_h_field, size=n_step))
old_fid = model.compute_fidelity()
best_protocol = np.copy(model.protocol())
else:
# So user can feed in data say from a specific protocol
old_fid = model.compute_fidelity()
best_protocol = np.copy(model.protocol())
random_position = np.arange(n_step, dtype=int)
fid_series = [old_fid]
while True:
np.random.shuffle(random_position)
local_minima_reached = True # trick
for t in random_position:
model.update_hx(t,
model.protocol_hx(t) ^ 1) # assumes binary fields
new_fid = model.compute_fidelity()
n_fid_eval += 1
fid_series.append(old_fid)
if new_fid > old_fid: # accept descent
old_fid = new_fid
n_visit += 1
local_minima_reached = False
break
else:
model.update_hx(t,
model.protocol_hx(t)
^ 1) # assumes binary fields
if local_minima_reached:
break
return old_fid, np.copy(model.protocol()), n_fid_eval, n_visit, fid_series
def SA(param, model: MODEL):
Ti = param['Ti']
n_quench = param['n_quench']
if n_quench == 0:
return
n_step = param['n_step']
# initial random protocol
model.update_protocol(np.random.randint(0, model.n_h_field, size=n_step))
old_fid = model.compute_fidelity()
best_fid = old_fid
best_protocol = np.copy(model.protocol())
T = Ti
step = 0
while T > 1e-12:
beta = 1. / T
# --- ---> single spin flip update <--- ---
random_time = np.random.randint(0, n_step)
current_hx = model.protocol_hx(random_time)
model.update_hx(random_time, model.random_flip(random_time))
# --- --- --- --- --- --- --- --- --- ---
new_fid = model.compute_fidelity()
if new_fid > best_fid:
best_fid = new_fid
best_protocol = np.copy(
model.protocol()) # makes an independent copy !
d_fid = new_fid - old_fid
if d_fid > 0.: # accept move
old_fid = new_fid
elif np.exp(beta * d_fid) > np.random.uniform(): # accept move
old_fid = new_fid
else: # reject move
model.update_hx(random_time, current_hx)
step += 1
T = Ti * (1.0 - step / n_quench)
return best_fid, best_protocol, n_quench
def SD_2SF(param, model: MODEL, init_random=False):
""" 2SF + 1 SF stochastic descent: all possible 2 spin-flip and
1 spin-flip moves are considered. Algorithm halts when all moves will decrease fidelity
"""
if model.n_h_field > 2:
assert False, 'This works only for bang-bang protocols'
n_step = param['n_step']
n_fid_eval = 0
n_visit = 1
if init_random:
# Random initialization
model.update_protocol(
np.random.randint(0, model.n_h_field, size=n_step))
old_fid = model.compute_fidelity()
best_protocol = np.copy(model.protocol())
else:
# So user can feed in data say from a specific protocol
old_fid = model.compute_fidelity()
best_protocol = np.copy(model.protocol())
x1_ar, x2_ar = np.triu_indices(n_step, 1)
n_2F_step = x1_ar.shape[0] # number of possible 2-flip updates
order2F = np.arange(
0, n_2F_step,
dtype=np.int) # ordering sequenc // to be randomly shuffled
n_1F_step = n_step
order1F = np.arange(
0, n_1F_step,
dtype=np.int) # ordering sequenc // to be randomly shuffled
order1F_vs_2F = 2 * np.ones(n_2F_step + n_1F_step, dtype=np.int)
order1F_vs_2F[:n_1F_step] = 1
############################
#########################
while True: # careful with this. For binary actions, this is guaranteed to break
np.random.shuffle(order1F)
np.random.shuffle(order2F)
np.random.shuffle(order1F_vs_2F)
idx_1F = 0
idx_2F = 0
local_minima_reached = True
for update_type in order1F_vs_2F:
if update_type == 1:
# perform 1 SF update
t = order1F[idx_1F]
model.update_hx(t,
model.protocol_hx(t)
^ 1) # assumes binary fields
new_fid = model.compute_fidelity()
n_fid_eval += 1
idx_1F += 1
if new_fid > old_fid: # accept descent
#print("%.15f"%new_fid,'\t',n_fid_eval)
n_visit += 1
old_fid = new_fid
local_minima_reached = False # will exit for loop before it ends ... local update accepted
break
else:
model.update_hx(t, model.protocol_hx(t) ^ 1)
else:
# perform 2 SF update
o2F = order2F[idx_2F]
t1, t2 = x1_ar[o2F], x2_ar[o2F]
model.update_hx(t1,
model.protocol_hx(t1)
^ 1) # assumes binary fields
model.update_hx(t2, model.protocol_hx(t2) ^ 1)
new_fid = model.compute_fidelity()
n_fid_eval += 1
idx_2F += 1
if new_fid > old_fid: # accept descent
print("%.15f" % new_fid, '\t', n_fid_eval)
n_visit += 1
old_fid = new_fid
local_minima_reached = False # will exit for loop before it ends ... local update accepted
break
else:
model.update_hx(t1,
model.protocol_hx(t1)
^ 1) # assumes binary fields
model.update_hx(t2, model.protocol_hx(t2) ^ 1)
if local_minima_reached:
break
return old_fid, np.copy(model.protocol()), n_fid_eval, n_visit
def SD(param, model: MODEL, init=False):
n_step = param['n_step']
n_fid_eval = 0
if init:
# Random initialization
model.update_protocol(
np.random.randint(0, model.n_h_field, size=n_step))
old_fid = model.compute_fidelity()
best_protocol = np.copy(model.protocol())
else:
# So user can feed in data say from a specific protocol
old_fid = model.compute_fidelity()
best_protocol = np.copy(model.protocol())
while True: # careful with this. For binary actions, this is guaranteed to break
random_position = np.arange(n_step, dtype=int)
np.random.shuffle(random_position)
local_minima = True
for t in random_position:
model.update_hx(t, model.random_flip(t))
new_fid = model.compute_fidelity()
n_fid_eval += 1
if new_fid > old_fid: # accept descent
old_fid = new_fid
best_protocol = np.copy(model.protocol())
local_minima = False # will exit for loop before it ends ... local update accepted
break
else:
model.update_hx(t, model.random_flip(t))
if local_minima:
break
return old_fid, best_protocol, n_fid_eval
return check_args(parser.parse_args())
def check_args(args):
# creates saving directory if necessary
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
# makes sure batch_size and epoch are positive
if args.epoch < 1:
raise Exception('Number of epochs must be larger than or equal to one')
if args.batch_size < 1:
raise Exception('Batch size must be larger than or equal to one')
return args
if __name__ == '__main__':
# parse arguments
args = parse_args()
# sets random seeds
np.random.seed(args.seed)
torch.manual_seed(args.seed)
# instanciates the model
model = MODEL(args)
# trains the model
train(model)
def Gibbs_Sampling(param, model: MODEL):
# should also measure acceptance rate
Ti = param['Ti']
beta = 1. / Ti
n_step = param['n_step']
n_equilibrate = 10000
n_auto_correlate = n_step * 10 # should look at auto-correlation time !
# initial random protocol
model.update_protocol(np.random.randint(0, model.n_h_field, size=n_step))
old_fid = model.compute_fidelity()
best_fid = old_fid
for i in range(n_equilibrate):
random_time = np.random.randint(0, n_step)
current_hx = model.protocol_hx(random_time)
model.update_hx(random_time, model.random_flip(random_time))
new_fid = model.compute_fidelity()
d_fid = new_fid - old_fid
if d_fid > 0.: # accept move
old_fid = new_fid
elif np.exp(beta * d_fid) > np.random.uniform(): # accept move
old_fid = new_fid
else: # reject move
model.update_hx(random_time, current_hx)
samples = []
fid_samples = []
energy_samples = []
for i in range(n_sample):
for j in range(n_auto_correlate):
random_time = np.random.randint(0, n_step)
current_hx = model.protocol_hx(random_time)
model.update_hx(random_time, model.random_flip(random_time))
new_fid = model.compute_fidelity()
d_fid = new_fid - old_fid
if d_fid > 0.: # accept move
old_fid = new_fid
elif np.exp(beta * d_fid) > np.random.uniform(): # accept move
old_fid = new_fid
else: # reject move
model.update_hx(random_time, current_hx)
samples.append(np.copy(model.protocol()))
fid_samples.append(model.compute_fidelity())
energy_samples.append(model.compute_energy())
return samples, fid_samples, energy_samples
def __init__(self):
self.db = MODEL.ChangeInfoUser()
def run_SA(parameters, model: MODEL, utils, root, save=True):
if parameters['verbose'] == 0:
blockPrint()
outfile = utils.make_file_name(parameters, root=root)
n_exist_sample, all_result = utils.read_current_results(outfile)
n_sample = parameters['n_sample']
if parameters['Ti'] < 0.:
parameters['Ti'] = compute_initial_Ti(parameters, model, n_sample=1000)
print("Initial temperature Ti=%.3f" % parameters['Ti'])
if n_exist_sample >= n_sample:
print(
"\n\n-----------> Samples already computed in file -- terminating ... <-----------"
)
return
all_result
print("\n\n-----------> Starting simulated annealing <-----------")
n_iteration_left = n_sample - n_exist_sample # data should be saved 10 times --> no more (otherwise things are way too slow !)
n_mod = max([1, n_iteration_left // 10])
for it in range(n_iteration_left):
start_time = time.time()
best_fid, best_protocol, n_fid_eval = SA(
parameters, model) # -- --> performing annealing here <-- --
if parameters['task'] == 'SASD':
print(' -> Stochastic descent ... ')
model.update_protocol(best_protocol)
best_fid, best_protocol, n_fid_eval_SD = SD(parameters,
model,
init=False)
n_fid_eval += n_fid_eval_SD
energy = model.compute_energy(protocol=best_protocol)
result = [n_fid_eval, best_fid, energy, best_protocol]
print("\n----------> RESULT FOR ANNEALING NO %i <-------------" %
(it + 1))
print("Number of fidelity eval \t%i" % n_fid_eval)
print("Best fidelity \t\t\t%.4f" % best_fid)
print("Best hx_protocol\t\t", list(best_protocol))
all_result.append(result)
if save and it % n_mod == 0:
with open(outfile, 'wb') as f:
pickle.dump([parameters, all_result], f)
f.close()
print("Saved iteration --> %i to %s" %
(it + n_exist_sample, outfile))
print("Iteration run time --> %.4f s" % (time.time() - start_time))
print("\n Thank you and goodbye !")
enablePrint()
if save:
with open(outfile, 'wb') as f:
pickle.dump([parameters, all_result], f)
f.close()
return all_result
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', type=int, default=-1)
parser.add_argument('--max_iter',
type=int,
default=300,
help='number of iterations')
parser.add_argument('--decay_epoch',
type=int,
default=20,
help='number of iterations')
parser.add_argument('--test',
type=bool,
default=False,
help='enable testing')
parser.add_argument('--train_test',
type=bool,
default=True,
help='enable testing')
parser.add_argument('--show',
type=bool,
default=True,
help='print progress')
parser.add_argument('--init_std',
type=float,
default=0.1,
help='weight initialization std')
parser.add_argument('--init_lr',
type=float,
default=0.001,
help='initial learning rate')
parser.add_argument('--lr_decay',
type=float,
default=0.75,
help='learning rate decay')
parser.add_argument(
'--final_lr',
type=float,
default=1E-5,
help='learning rate will not decrease after hitting this threshold')
parser.add_argument('--momentum',
type=float,
default=0.9,
help='momentum rate')
parser.add_argument('--max_grad_norm',
type=float,
default=3.0,
help='maximum gradient norm')
parser.add_argument('--hidden_dim',
type=int,
default=200,
help='hidden layer dimension')
parser.add_argument('--n_hidden',
type=int,
default=2,
help='hidden numbers')
parser.add_argument('--dataset', type=str, default='assist2017')
if parser.parse_args().dataset == 'assist2009_updated':
parser.add_argument('--batch_size',
type=int,
default=100,
help='the batch size')
parser.add_argument('--qa_embed_dim',
type=int,
default=200,
help='answer and question embedding dimensions')
parser.add_argument(
'--n_question',
type=int,
default=110,
help='the number of unique questions in the dataset')
parser.add_argument('--seqlen',
type=int,
default=200,
help='the allowed maximum length of a sequence')
parser.add_argument('--data_dir',
type=str,
default='./data/assist2009_updated',
help='data directory')
parser.add_argument('--data_name',
type=str,
default='assist2009_updated',
help='data set name')
parser.add_argument('--load',
type=str,
default='assist2009_updated',
help='model file to load')
parser.add_argument('--save',
type=str,
default='assist2009_updated',
help='path to save model')
elif parser.parse_args().dataset == 'assist2015':
parser.add_argument('--batch_size',
type=int,
default=100,
help='the batch size')
parser.add_argument('--qa_embed_dim',
type=int,
default=200,
help='answer and question embedding dimensions')
parser.add_argument(
'--n_question',
type=int,
default=100,
help='the number of unique questions in the dataset')
parser.add_argument('--seqlen',
type=int,
default=200,
help='the allowed maximum length of a sequence')
parser.add_argument('--data_dir',
type=str,
default='./data/assist2015',
help='data directory')
parser.add_argument('--data_name',
type=str,
default='assist2015',
help='data set name')
parser.add_argument('--load',
type=str,
default='assist2015',
help='model file to load')
parser.add_argument('--save',
type=str,
default='assist2015',
help='path to save model')
elif parser.parse_args().dataset == 'assist2017':
parser.add_argument('--batch_size',
type=int,
default=32,
help='the batch size')
parser.add_argument('--qa_embed_dim',
type=int,
default=200,
help='answer and question embedding dimensions')
parser.add_argument(
'--n_question',
type=int,
default=102,
help='the number of unique questions in the dataset')
parser.add_argument('--seqlen',
type=int,
default=200,
help='the allowed maximum length of a sequence')
parser.add_argument('--data_dir',
type=str,
default='../dataset/assist2017/train_valid_test',
help='data directory')
parser.add_argument('--data_name',
type=str,
default='assist2017',
help='data set name')
parser.add_argument('--load',
type=str,
default='assist2017',
help='model file to load')
parser.add_argument('--save',
type=str,
default='assist2017',
help='path to save model')
elif parser.parse_args().dataset == 'STATICS':
parser.add_argument('--batch_size',
type=int,
default=100,
help='the batch size')
parser.add_argument('--qa_embed_dim',
type=int,
default=200,
help='answer and question embedding dimensions')
parser.add_argument(
'--n_question',
type=int,
default=1223,
help='the number of unique questions in the dataset')
parser.add_argument('--seqlen',
type=int,
default=200,
help='the allowed maximum length of a sequence')
parser.add_argument('--data_dir',
type=str,
default='./data/STATICS',
help='data directory')
parser.add_argument('--data_name',
type=str,
default='STATICS',
help='data set name')
parser.add_argument('--load',
type=str,
default='STATICS',
help='model file to load')
parser.add_argument('--save',
type=str,
default='STATICS',
help='path to save model')
elif parser.parse_args().dataset == 'synthetic':
parser.add_argument('--batch_size',
type=int,
default=100,
help='the batch size')
parser.add_argument('--qa_embed_dim',
type=int,
default=200,
help='answer and question embedding dimensions')
parser.add_argument(
'--n_question',
type=int,
default=50,
help='the number of unique questions in the dataset')
parser.add_argument('--seqlen',
type=int,
default=200,
help='the allowed maximum length of a sequence')
parser.add_argument('--data_dir',
type=str,
default='./data/synthetic',
help='data directory')
parser.add_argument('--data_name',
type=str,
default='synthetic',
help='data set name')
parser.add_argument('--load',
type=str,
default='synthetic',
help='model file to load')
parser.add_argument('--save',
type=str,
default='synthetic',
help='path to save model')
params = parser.parse_args()
params.lr = params.init_lr
print(params)
dat = DATA(n_question=params.n_question,
seqlen=params.seqlen,
separate_char=',')
if params.dataset != 'synthetic':
train_data_path = params.data_dir + "/" + params.data_name + "_train1.csv"
valid_data_path = params.data_dir + "/" + params.data_name + "_valid1.csv"
test_data_path = params.data_dir + "/" + params.data_name + "_test.csv"
else:
train_data_path = params.data_dir + "/" + "naive_c5_q50_s4000_v0_train1.csv"
valid_data_path = params.data_dir + "/" + "naive_c5_q50_s4000_v0_valid1.csv"
test_data_path = params.data_dir + "/" + "naive_c5_q50_s4000_v0_test.csv"
train_q_data, train_q_t_data, train_answer_data = dat.load_data(
train_data_path)
valid_q_data, valid_q_t_data, valid_answer_data = dat.load_data(
valid_data_path)
test_q_data, test_q_t_data, test_answer_data = dat.load_data(
test_data_path)
model = MODEL(n_question=params.n_question,
hidden_dim=params.hidden_dim,
x_embed_dim=params.qa_embed_dim,
hidden_layers=params.n_hidden,
gpu=params.gpu)
model.init_embeddings()
model.init_params()
optimizer = optim.Adam(params=model.parameters(),
lr=params.lr,
betas=(0.9, 0.9))
if params.gpu >= 0:
print('device: ' + str(params.gpu))
torch.cuda.set_device(params.gpu)
model.cuda()
# all_train_loss = {}
# all_train_accuracy = {}
# all_train_auc = {}
# all_valid_loss = {}
# all_valid_accuracy = {}
# all_valid_auc = {}
best_valid_auc = 0
for idx in range(params.max_iter):
train_loss, train_accuracy, train_auc = train(model, params, optimizer,
train_q_data,
train_q_t_data,
train_answer_data)
print(
'Epoch %d/%d, loss : %3.5f, auc : %3.5f, accuracy : %3.5f' %
(idx + 1, params.max_iter, train_loss, train_auc, train_accuracy))
valid_loss, valid_accuracy, valid_auc = test(model, params, optimizer,
valid_q_data,
valid_q_t_data,
valid_answer_data)
print('Epoch %d/%d, valid auc : %3.5f, valid accuracy : %3.5f' %
(idx + 1, params.max_iter, valid_auc, valid_accuracy))
test_loss, test_accuracy, test_auc = test(model, params, optimizer,
test_q_data, test_q_t_data,
test_answer_data)
print('Epoch %d/%d, test auc : %3.5f, test accuracy : %3.5f' %
(idx + 1, params.max_iter, test_auc, test_accuracy))
# all_train_auc[idx + 1] = train_auc
# all_train_accuracy[idx + 1] = train_accuracy
# all_train_loss[idx + 1] = train_loss
# all_valid_loss[idx + 1] = valid_loss
# all_valid_accuracy[idx + 1] = valid_accuracy
# all_valid_auc[idx + 1] = valid_auc
if valid_auc > best_valid_auc:
print('%3.4f to %3.4f' % (best_valid_auc, valid_auc))
best_valid_auc = valid_auc
print('Q Sentence Len :', max_len[1])
print('Ans Sentence Len :', max_len[2])
print('CNN Filter Size :', CNN_FILTER_SIZE)
print('CNN Filter Size2 :', CNN_FILTER_SIZE2)
print('CNN Filter Num :', CNN_FILTER_NUM)
print('DNN Output Size :', DNN_WIDTH, '->', 1)
print('###############################################################', '\n')
with open('output_data/question/qa.train.json') as data_file:
train_q = json.load(data_file)
with open('output_data/question/qa.val.json') as data_file:
val_q = json.load(data_file)
start = time.time()
acm_net = MODEL.MODEL(BATCHSIZE, X_DIMENSION, DNN_WIDTH, CNN_FILTER_SIZE,
CNN_FILTER_SIZE2, CNN_FILTER_NUM, CNN_FILTER_NUM2,
LEARNING_RATE, DROPOUT, choice, max_plot_num, max_len,
parameterPath)
acm_net.initialize()
accuracy = utility.train(train_q, val_q, plotFilePath, acm_net, EPOCH,
LEARNING_RATE, BATCHSIZE, DROPOUT, choice, max_len)
accuracy_train = np.array(accuracy['train'])
accuracy_val = np.array(accuracy['val'])
print('val: ', accuracy_val, np.amax(accuracy_val))
print('train: ', accuracy_train, np.amax(accuracy_train))
eval_time = time.time() - start
print('use time: ', eval_time)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', type=int, default=0, help='the gpu will be used, e.g "0,1,2,3"')
parser.add_argument('--max_iter', type=int, default=300, help='number of iterations')
parser.add_argument('--decay_epoch', type=int, default=20, help='number of iterations')
parser.add_argument('--test', type=bool, default=False, help='enable testing')
parser.add_argument('--train_test', type=bool, default=True, help='enable testing')
parser.add_argument('--show', type=bool, default=True, help='print progress')
parser.add_argument('--init_std', type=float, default=0.1, help='weight initialization std')
parser.add_argument('--init_lr', type=float, default=0.01, help='initial learning rate')
parser.add_argument('--lr_decay', type=float, default=0.75, help='learning rate decay')
parser.add_argument('--final_lr', type=float, default=1E-5,
help='learning rate will not decrease after hitting this threshold')
parser.add_argument('--momentum', type=float, default=0.9, help='momentum rate')
parser.add_argument('--maxgradnorm', type=float, default=50.0, help='maximum gradient norm')
parser.add_argument('--final_fc_dim', type=float, default=50, help='hidden state dim for final fc layer')
dataset = 'assist2009_updated'
if dataset == 'assist2009_updated':
parser.add_argument('--q_embed_dim', type=int, default=50, help='question embedding dimensions')
parser.add_argument('--batch_size', type=int, default=32, help='the batch size')
parser.add_argument('--qa_embed_dim', type=int, default=200, help='answer and question embedding dimensions')
parser.add_argument('--memory_size', type=int, default=20, help='memory size')
parser.add_argument('--n_question', type=int, default=110, help='the number of unique questions in the dataset')
parser.add_argument('--seqlen', type=int, default=200, help='the allowed maximum length of a sequence')
parser.add_argument('--data_dir', type=str, default='./data/assist2009_updated', help='data directory')
parser.add_argument('--data_name', type=str, default='assist2009_updated', help='data set name')
parser.add_argument('--load', type=str, default='assist2009_updated', help='model file to load')
parser.add_argument('--save', type=str, default='assist2009_updated', help='path to save model')
elif dataset == 'STATICS':
parser.add_argument('--batch_size', type=int, default=10, help='the batch size')
parser.add_argument('--q_embed_dim', type=int, default=50, help='question embedding dimensions')
parser.add_argument('--qa_embed_dim', type=int, default=100, help='answer and question embedding dimensions')
parser.add_argument('--memory_size', type=int, default=50, help='memory size')
parser.add_argument('--n_question', type=int, default=1223, help='the number of unique questions in the dataset')
parser.add_argument('--seqlen', type=int, default=6, help='the allowed maximum length of a sequence')
parser.add_argument('--data_dir', type=str, default='./data/STATICS', help='data directory')
parser.add_argument('--data_name', type=str, default='STATICS', help='data set name')
parser.add_argument('--load', type=str, default='STATICS', help='model file to load')
parser.add_argument('--save', type=str, default='STATICS', help='path to save model')
params = parser.parse_args()
params.lr = params.init_lr
params.memory_key_state_dim = params.q_embed_dim
params.memory_value_state_dim = params.qa_embed_dim
print(params)
dat = DATA(n_question=params.n_question, seqlen=params.seqlen, separate_char=',')
# train_data_path = params.data_dir + "/" + "test5.1.txt"
train_data_path = params.data_dir + "/" + params.data_name + "_train1.csv"
valid_data_path = params.data_dir + "/" + params.data_name + "_valid1.csv"
test_data_path = params.data_dir + "/" + params.data_name + "_test.csv"
train_q_data, train_qa_data = dat.load_data(train_data_path)
valid_q_data, valid_qa_data = dat.load_data(valid_data_path)
test_q_data, test_qa_data = dat.load_data(test_data_path)
params.memory_key_state_dim = params.q_embed_dim
params.memory_value_state_dim = params.qa_embed_dim
model = MODEL(n_question=params.n_question,
batch_size=params.batch_size,
q_embed_dim=params.q_embed_dim,
qa_embed_dim=params.qa_embed_dim,
memory_size=params.memory_size,
memory_key_state_dim=params.memory_key_state_dim,
memory_value_state_dim=params.memory_value_state_dim,
final_fc_dim=params.final_fc_dim)
model.init_embeddings()
model.init_params()
# optimizer = optim.SGD(params=model.parameters(), lr=params.lr, momentum=params.momentum)
optimizer = optim.Adam(params=model.parameters(), lr=params.lr, betas=(0.9, 0.9))
if params.gpu >= 0:
print('device: ' + str(params.gpu))
torch.cuda.set_device(params.gpu)
model.cuda()
all_train_loss = {}
all_train_accuracy = {}
all_train_auc = {}
all_valid_loss = {}
all_valid_accuracy = {}
all_valid_auc = {}
best_valid_auc = 0
# shuffle_index = np.random.permutation(train_q_data.shape[0])
# q_data_shuffled = train_q_data[shuffle_index]
# qa_data_shuffled = train_qa_data[shuffle_index]
for idx in range(params.max_iter):
train_loss, train_accuracy, train_auc = train(idx, model, params, optimizer, train_q_data, train_qa_data)
print('Epoch %d/%d, loss : %3.5f, auc : %3.5f, accuracy : %3.5f' % (idx + 1, params.max_iter, train_loss, train_auc, train_accuracy))
valid_loss, valid_accuracy, valid_auc = test(model, params, optimizer, valid_q_data, valid_qa_data)
print('Epoch %d/%d, valid auc : %3.5f, valid accuracy : %3.5f' % (idx + 1, params.max_iter, valid_auc, valid_accuracy))
all_train_auc[idx + 1] = train_auc
all_train_accuracy[idx + 1] = train_accuracy
all_train_loss[idx + 1] = train_loss
all_valid_loss[idx + 1] = valid_loss
all_valid_accuracy[idx + 1] = valid_accuracy
all_valid_auc[idx + 1] = valid_auc
#
# output the epoch with the best validation auc
if valid_auc > best_valid_auc:
print('%3.4f to %3.4f' % (best_valid_auc, valid_auc))
best_valid_auc = valid_auc
def train_one_dataset(params, file_name, train_q_data, train_qa_data,
valid_q_data, valid_qa_data):
### ================================== model initialization ==================================
g_model = MODEL(n_question=params.n_question,
seqlen=params.seqlen,
batch_size=params.batch_size,
q_embed_dim=params.q_embed_dim,
qa_embed_dim=params.qa_embed_dim,
memory_size=params.memory_size,
memory_key_state_dim=params.memory_key_state_dim,
memory_value_state_dim=params.memory_value_state_dim,
final_fc_dim=params.final_fc_dim)
# create a module by given a Symbol
net = mx.mod.Module(symbol=g_model.sym_gen(),
data_names=['q_data', 'qa_data'],
label_names=['target'],
context=params.ctx)
# create memory by given input shapes
net.bind(data_shapes=[
mx.io.DataDesc(name='q_data',
shape=(params.seqlen, params.batch_size),
layout='SN'),
mx.io.DataDesc(name='qa_data',
shape=(params.seqlen, params.batch_size),
layout='SN')
],
label_shapes=[
mx.io.DataDesc(name='target',
shape=(params.seqlen, params.batch_size),
layout='SN')
])
# initial parameters with the default random initializer
net.init_params(initializer=mx.init.Normal(sigma=params.init_std))
# decay learning rate in the lr_scheduler
lr_scheduler = mx.lr_scheduler.FactorScheduler(
step=20 * (train_q_data.shape[0] / params.batch_size),
factor=0.667,
stop_factor_lr=1e-5)
net.init_optimizer(optimizer='sgd',
optimizer_params={
'learning_rate': params.lr,
'momentum': params.momentum,
'lr_scheduler': lr_scheduler
})
for parameters in net.get_params()[0]:
print(parameters, net.get_params()[0][parameters].asnumpy().shape)
print("\n")
### ================================== start training ==================================
all_train_loss = {}
all_train_accuracy = {}
all_train_auc = {}
all_valid_loss = {}
all_valid_accuracy = {}
all_valid_auc = {}
best_valid_auc = 0
for idx in range(params.max_iter):
train_loss, train_accuracy, train_auc = train(net,
params,
train_q_data,
train_qa_data,
label='Train')
valid_loss, valid_accuracy, valid_auc = test(net,
params,
valid_q_data,
valid_qa_data,
label='Valid')
print('epoch', idx + 1)
print("valid_auc\t", valid_auc, "\ttrain_auc\t", train_auc)
print("valid_accuracy\t", valid_accuracy, "\ttrain_accuracy\t",
train_accuracy)
print("valid_loss\t", valid_loss, "\ttrain_loss\t", train_loss)
if not os.path.isdir('model'):
os.makedirs('model')
if not os.path.isdir(os.path.join('model', params.save)):
os.makedirs(os.path.join('model', params.save))
all_valid_auc[idx + 1] = valid_auc
all_train_auc[idx + 1] = train_auc
all_valid_loss[idx + 1] = valid_loss
all_train_loss[idx + 1] = train_loss
all_valid_accuracy[idx + 1] = valid_accuracy
all_train_accuracy[idx + 1] = train_accuracy
# output the epoch with the best validation auc
if valid_auc > best_valid_auc:
best_valid_auc = valid_auc
best_epoch = idx + 1
# here the epoch is default, set to be 100
# we only save the model in the epoch with the better results
net.save_checkpoint(prefix=os.path.join('model', params.save,
file_name),
epoch=100)
if not os.path.isdir('result'):
os.makedirs('result')
if not os.path.isdir(os.path.join('result', params.save)):
os.makedirs(os.path.join('result', params.save))
f_save_log = open(os.path.join('result', params.save, file_name), 'w')
f_save_log.write("valid_auc:\n" + str(all_valid_auc) + "\n\n")
f_save_log.write("train_auc:\n" + str(all_train_auc) + "\n\n")
f_save_log.write("valid_loss:\n" + str(all_valid_loss) + "\n\n")
f_save_log.write("train_loss:\n" + str(all_train_loss) + "\n\n")
f_save_log.write("valid_accuracy:\n" + str(all_valid_accuracy) + "\n\n")
f_save_log.write("train_accuracy:\n" + str(all_train_accuracy) + "\n\n")
f_save_log.close()
return best_epoch
def main():
# Utility object for reading, writing parameters, etc.
utils = UTILS()
# Reading parameters from para.dat file
parameters = utils.read_parameter_file()
# Command line specified parameters overide parameter file values
utils.read_command_line_arg(parameters, sys.argv)
# Printing parameters for user
utils.print_parameters(parameters)
# Defining Hamiltonian
H = HAMILTONIAN(**parameters)
# Defines the model, and precomputes evolution matrices given set of states
model = MODEL(H, parameters)
#n_step = parameters['n_step']
#X,y=sample_m0(10000,n_step,model)
#print(y[0:10])
#plt.hist(y,bins=20)
#plt.show()
rob_vs_T = {}
n_eval = {}
fid = {}
res = {}
visit = {}
T_list = np.arange(0.05, 10.001, 0.05)
n_step_list = [40, 50, 60, 70, 80, 90, 100, 110]
for T in T_list:
for n_step in n_step_list: #[40,50,60,70,80,90,100,110,120]:
##for T in np.arange(0.025,10.001,0.025):
# for n_step in [100,200,400] :
parameters['T'] = T
parameters['n_step'] = n_step
parameters['dt'] = T / n_step
file = utils.make_file_name(parameters, root='data/')
res = parse_data(file)
n_eval[(n_step, hash(T))] = res['n_fid']
fid[(n_step, hash(T))] = res['F']
visit[(n_step, hash(T))] = res['n_visit']
''' with open(file,'rb') as f:
_, data = pickle.load(f)
n_elem = len(data)
n_eval[(n_step,hash(T))]=[]
n_fid[(n_step,hash(T))]=[]
for elem in data:
n_eval[(n_step,hash(T))].append(elem[0])
n_fid[(n_step,hash(T))].append(elem[1])'''
#print(n_eval)
#exit()
n_eval_mean = {}
fid_mean = {}
visit_mean = {}
#print(visit[(40,115292150460684704)])
#exit()
for n_step in n_step_list:
n_eval_mean[n_step] = []
fid_mean[n_step] = []
visit_mean[n_step] = []
for T in T_list:
hT = hash(T)
n_eval_mean[n_step].append(
[T, np.mean(n_eval[(n_step, hT)]) / (n_step * n_step)])
fid_mean[n_step].append([T, np.mean(fid[(n_step, hT)])])
visit_mean[n_step].append(
[T, np.mean(visit[(n_step, hT)]) / (n_step)])
c_list = [
'#d53e4f', '#f46d43', '#fdae61', '#fee08b', '#e6f598', '#abdda4',
'#66c2a5', '#3288bd'
]
for i, n_step in enumerate(n_step_list):
x = np.array(n_eval_mean[n_step])
plt.plot(x[:, 0], x[:, 1], c='black', zorder=0)
plt.scatter(x[:, 0],
x[:, 1],
c=c_list[i],
marker='o',
s=5,
label='$N=%i$' % n_step,
zorder=1)
plt.title('Number of fidelity evaluations vs. ramp time \n for 2 flip')
plt.ylabel('$N_{eval}/N^2$')
plt.xlabel('$T$')
plt.legend(loc='best')
plt.tight_layout()
plt.show()
for i, n_step in enumerate(n_step_list):
x = np.array(visit_mean[n_step])
plt.plot(x[:, 0], x[:, 1], c='black', zorder=0)
plt.scatter(x[:, 0],
x[:, 1],
c=c_list[i],
marker='o',
s=5,
label='$N=%i$' % n_step,
zorder=1)
plt.title('Number of visited states vs. ramp time \n for 2 flip')
plt.ylabel('$N_{visit}/N$')
plt.xlabel('$T$')
plt.legend(loc='best')
plt.tight_layout()
plt.show()
'''