def main(args):
meta_info = {
"POLLUTION": [5, 50, 14],
"HR": [32, 50, 13],
"BATTERY": [20, 50, 3]
}
output_directory = "output/"
horizon = 10
output_dim = 1
dataset_name = args.dataset
save_model_file = args.save_model_file
load_model_file = args.load_model_file
lower_trial = args.lower_trial
upper_trial = args.upper_trial
learning_rate = args.learning_rate
meta_learning_rate = args.meta_learning_rate
adaptation_steps = args.adaptation_steps
batch_size = args.batch_size
model_name = args.model
is_test = args.is_test
patience_stopping = args.stopping_patience
epochs = args.epochs
noise_level = args.noise_level
noise_type = args.noise_type
assert model_name in ("FCN", "LSTM"), "Model was not correctly specified"
assert dataset_name in ("POLLUTION", "HR", "BATTERY")
window_size, task_size, input_dim = meta_info[dataset_name]
grid = [0., noise_level]
train_data = pickle.load(
open(
"../../Data/TRAIN-" + dataset_name + "-W" + str(window_size) +
"-T" + str(task_size) + "-NOML.pickle", "rb"))
train_data_ML = pickle.load(
open(
"../../Data/TRAIN-" + dataset_name + "-W" + str(window_size) +
"-T" + str(task_size) + "-ML.pickle", "rb"))
validation_data = pickle.load(
open(
"../../Data/VAL-" + dataset_name + "-W" + str(window_size) + "-T" +
str(task_size) + "-NOML.pickle", "rb"))
validation_data_ML = pickle.load(
open(
"../../Data/VAL-" + dataset_name + "-W" + str(window_size) + "-T" +
str(task_size) + "-ML.pickle", "rb"))
test_data = pickle.load(
open(
"../../Data/TEST-" + dataset_name + "-W" + str(window_size) +
"-T" + str(task_size) + "-NOML.pickle", "rb"))
test_data_ML = pickle.load(
open(
"../../Data/TEST-" + dataset_name + "-W" + str(window_size) +
"-T" + str(task_size) + "-ML.pickle", "rb"))
results_dict = {}
for trial in range(lower_trial, upper_trial):
output_directory = "../../Models/" + dataset_name + "_" + model_name + "_MAML/" + str(
trial) + "/"
save_model_file_ = output_directory + save_model_file
load_model_file_ = output_directory + load_model_file
try:
os.mkdir(output_directory)
except OSError as error:
print(error)
f = open(output_directory + "/results3.txt", "a+")
f.write("Learning rate :%f \n" % learning_rate)
f.write("Meta-learning rate: %f \n" % meta_learning_rate)
f.write("Adaptation steps: %f \n" % adaptation_steps)
f.write("\n")
f.close()
if model_name == "LSTM":
model = LSTMModel(batch_size=batch_size,
seq_len=window_size,
input_dim=input_dim,
n_layers=2,
hidden_dim=120,
output_dim=1)
elif model_name == "FCN":
kernels = [8, 5, 3] if dataset_name != "POLLUTION" else [4, 2, 1]
model = FCN(time_steps=window_size,
channels=[input_dim, 128, 128, 128],
kernels=kernels)
model.cuda()
maml = l2l.algorithms.MAML(model, lr=learning_rate, first_order=False)
opt = optim.Adam(maml.parameters(), lr=meta_learning_rate)
torch.backends.cudnn.enabled = False
total_num_tasks = train_data_ML.x.shape[0]
test(maml, model_name, dataset_name, test_data_ML, adaptation_steps,
learning_rate)
val_error = test(maml, model_name, dataset_name, validation_data_ML,
adaptation_steps, learning_rate)
early_stopping = EarlyStopping(patience=patience_stopping,
model_file=save_model_file_,
verbose=True)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(opt,
patience=200,
verbose=True)
#early_stopping(val_error, maml)
for iteration in range(epochs):
# Creates a clone of model
opt.zero_grad()
iteration_error = 0.0
print(iteration)
for task in range(batch_size):
learner = maml.clone()
task = np.random.randint(0, total_num_tasks - horizon)
#task_qry = np.random.randint(1,horizon+1)
x_spt, y_spt = train_data_ML[task]
#x_qry, y_qry = train_data_ML[(task+1):(task+1+horizon)]
x_qry, y_qry = train_data_ML[task + 1]
x_qry = x_qry.reshape(-1, window_size, input_dim)
y_qry = y_qry.reshape(-1, output_dim)
if model_name == "FCN":
x_qry = np.transpose(x_qry, [0, 2, 1])
x_spt = np.transpose(x_spt, [0, 2, 1])
x_spt, y_spt = to_torch(x_spt), to_torch(y_spt)
x_qry = to_torch(x_qry)
y_qry = to_torch(y_qry)
#data augmentation
epsilon = grid[np.random.randint(0, len(grid))]
if noise_type == "additive":
y_spt = y_spt + epsilon
y_qry = y_qry + epsilon
else:
y_spt = y_spt * (1 + epsilon)
y_qry = y_qry * (1 + epsilon)
# Fast adapt
for step in range(adaptation_steps):
pred = learner(x_spt)
error = mae(pred, y_spt)
learner.adapt(
error) #, allow_unused=True)#, allow_nograd=True)
pred = learner(x_qry)
evaluation_error = mae(pred, y_qry)
iteration_error += evaluation_error
# Meta-update the model parameters
iteration_error /= batch_size
iteration_error.backward() #retain_graph = True)
print("loss iteration:", iteration_error.data)
opt.step()
if iteration % 1 == 0:
val_error = test(maml, model_name, dataset_name,
validation_data_ML, adaptation_steps,
learning_rate)
test_error = test(maml, model_name, dataset_name, test_data_ML,
adaptation_steps, learning_rate)
scheduler.step(val_error)
print("Val error:", val_error)
print("Test error:", test_error)
early_stopping(val_error, maml)
if early_stopping.early_stop:
print("Early stopping")
break
maml.load_state_dict(torch.load(save_model_file_))
validation_error = test(maml, model_name, dataset_name,
validation_data_ML, adaptation_steps,
learning_rate)
#validation_error2 = test(maml, model_name, dataset_name, validation_data_ML, adaptation_steps, learning_rate, with_early_stopping=True)
#validation_error3 = test(maml, model_name, dataset_name, validation_data_ML, 10, learning_rate, with_early_stopping=True)
#validation_error4 = test(maml, model_name, dataset_name, validation_data_ML, 10, learning_rate*0.1, with_early_stopping=True)
test_error = test(maml, model_name, dataset_name, test_data_ML,
adaptation_steps, learning_rate)
#test_error2 = test(maml, model_name, dataset_name, test_data_ML, adaptation_steps , learning_rate, with_early_stopping=True)
#test_error3 = test(maml, model_name, dataset_name, test_data_ML, 10 , learning_rate, with_early_stopping=True)
#test_error4 = test(maml, model_name, dataset_name, test_data_ML, 10, learning_rate*0.1, with_early_stopping=True)
f = open(output_directory + "/results3.txt", "a+")
f.write("Dataset :%s \n" % dataset_name)
f.write("Test error: %f \n" % test_error)
#f.write("Test error2: %f \n" % test_error2)
#f.write("Test error3: %f \n" % test_error3)
#f.write("Test error4: %f \n" % test_error4)
f.write("Validation error: %f \n" % validation_error)
#f.write("Validation error2: %f \n" %validation_error2)
#f.write("Validation error3: %f \n" %validation_error3)
#f.write("Validation error4: %f \n" %validation_error4)
f.write("\n")
f.close()
results_dict[str(trial) + "_val"] = validation_error
results_dict[str(trial) + "_test"] = test_error
np.save(
"npy_objects/run03_" + dataset_name + "_" + model_name + "_" +
noise_type + "_" + str(noise_level * 100000) + ".npy", results_dict)
def test2(maml,
model_name,
test_data_ML,
adaptation_steps,
learning_rate,
with_early_stopping=False,
horizon=10):
total_tasks_test = len(test_data_ML)
error_list = []
learner = maml.clone() # Creates a clone of model
learner.cuda()
accum_error = 0.0
count = 0
input_dim = test_data_ML.x.shape[-1]
window_size = test_data_ML.x.shape[-2]
output_dim = test_data_ML.y.shape[-1]
for task in range(0, (total_tasks_test - horizon - 1),
total_tasks_test // 100):
temp_file_idx = test_data_ML.file_idx[task:task + horizon + 1]
if (len(np.unique(temp_file_idx)) > 1):
continue
if model_name == "LSTM":
model2 = LSTMModel(batch_size=None,
seq_len=None,
input_dim=input_dim,
n_layers=2,
hidden_dim=120,
output_dim=1)
elif model_name == "FCN":
kernels = [8, 5, 3] if window_size != 5 else [4, 2, 1]
model2 = FCN(time_steps=window_size,
channels=[input_dim, 128, 128, 128],
kernels=kernels)
#model2.cuda()
#model2.load_state_dict(copy.deepcopy(maml.module.state_dict()))
#opt2 = optim.Adam(model2.parameters(), lr=learning_rate)
learner = maml.clone()
x_spt, y_spt = test_data_ML[task]
x_qry = test_data_ML.x[(task + 1):(task + 1 + horizon)].reshape(
-1, window_size, input_dim)
y_qry = test_data_ML.y[(task + 1):(task + 1 + horizon)].reshape(
-1, output_dim)
if model_name == "FCN":
x_qry = np.transpose(x_qry, [0, 2, 1])
x_spt = np.transpose(x_spt, [0, 2, 1])
x_spt, y_spt = to_torch(x_spt), to_torch(y_spt)
x_qry = to_torch(x_qry)
y_qry = to_torch(y_qry)
early_stopping = EarlyStopping(patience=2,
model_file="temp/temp_file.pt",
verbose=True)
#learner.module.train()
#model2.eval()
for step in range(adaptation_steps):
#model2.train()
pred = learner(x_spt)
error = mae(pred, y_spt)
#opt2.zero_grad()
#error.backward()
learner.adapt(error)
#opt2.step()
if with_early_stopping:
with torch.no_grad():
model2.load_state_dict(
copy.deepcopy(learner.module.state_dict()))
#model2.eval()
pred = model2(x_qry)
error = mae(pred, y_qry)
early_stopping(error, model2)
if early_stopping.early_stop:
print("Early stopping")
break
if with_early_stopping:
model2.load_state_dict(torch.load("temp/temp_file.pt"))
#model2.eval()
#learner.module.eval()
pred = learner(x_qry)
error = mae(pred, y_qry)
accum_error += error.data
count += 1
error = accum_error / count
return error
def main(args):
meta_info = {
"POLLUTION": [5, 50, 14],
"HR": [32, 50, 13],
"BATTERY": [20, 50, 3]
}
output_directory = "output/"
verbose = True
batch_size = 64
freeze_model_flag = True
params = {'batch_size': batch_size, 'shuffle': True, 'num_workers': 0}
dataset_name = args.dataset
mode = args.mode
save_model_file = args.save_model_file
load_model_file = args.load_model_file
lower_trial = args.lower_trial
upper_trial = args.upper_trial
learning_rate = args.learning_rate
regularization_penalty = args.regularization_penalty
model_name = args.model
is_test = args.is_test
patience_stopping = args.patience_stopping
epochs = args.epochs
freeze_model_flag = args.freeze_model
assert mode in ("WFT", "WOFT", "50"), "Mode was not correctly specified"
assert model_name in ("FCN", "LSTM"), "Model was not correctly specified"
assert dataset_name in ("POLLUTION", "HR", "BATTERY")
window_size, task_size, input_dim = meta_info[dataset_name]
task_size = args.task_size
task_size = args.task_size
train_data = pickle.load(
open(
"../Data/TRAIN-" + dataset_name + "-W" + str(window_size) + "-T" +
str(task_size) + "-NOML.pickle", "rb"))
validation_data = pickle.load(
open(
"../Data/VAL-" + dataset_name + "-W" + str(window_size) + "-T" +
str(task_size) + "-NOML.pickle", "rb"))
validation_data_ML = pickle.load(
open(
"../Data/VAL-" + dataset_name + "-W" + str(window_size) + "-T" +
str(task_size) + "-ML.pickle", "rb"))
test_data = pickle.load(
open(
"../Data/TEST-" + dataset_name + "-W" + str(window_size) + "-T" +
str(task_size) + "-NOML.pickle", "rb"))
test_data_ML = pickle.load(
open(
"../Data/TEST-" + dataset_name + "-W" + str(window_size) + "-T" +
str(task_size) + "-ML.pickle", "rb"))
#device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
for trial in range(lower_trial, upper_trial):
output_directory = "../Models/" + dataset_name + "_" + model_name + "/" + str(
trial) + "/"
save_model_file_ = output_directory + save_model_file
load_model_file_ = output_directory + load_model_file
try:
os.mkdir(output_directory)
except OSError as error:
print(error)
if mode == "WOFT":
f = open(output_directory + "/results.txt", "a+")
f.write("Dataset :%s \n" % dataset_name)
f.write("Learning rate:%f \n" % learning_rate)
f.write("Epochs:%f \n" % epochs)
f.write("Model name:%s \n" % model_name)
f.close()
if model_name == "FCN":
kernels = [8, 5, 3
] if dataset_name != "POLLUTION" else [4, 2, 1]
train_data.x = np.transpose(train_data.x, [0, 2, 1])
test_data.x = np.transpose(test_data.x, [0, 2, 1])
validation_data.x = np.transpose(validation_data.x, [0, 2, 1])
early_stopping = EarlyStopping(patience=patience_stopping,
model_file=save_model_file_,
verbose=verbose)
if model_name == "LSTM":
model = LSTMModel(batch_size=batch_size,
seq_len=window_size,
input_dim=input_dim,
n_layers=2,
hidden_dim=120,
output_dim=1)
elif model_name == "FCN":
model = FCN(time_steps=window_size,
channels=[input_dim, 128, 128, 128],
kernels=kernels)
model.cuda()
train_loader = DataLoader(train_data, **params)
val_loader = DataLoader(validation_data, **params)
if is_test:
test_loader = DataLoader(test_data, **params)
else:
test_loader = DataLoader(validation_data, **params)
train(model, train_loader, val_loader, early_stopping,
learning_rate, epochs)
test(model, test_loader, output_directory, save_model_file_)
elif mode == "WFT":
f = open(output_directory + "/results.txt", "a+")
f.write("Dataset :%s \n" % dataset_name)
f.write("Learning rate:%f \n" % learning_rate)
f.write("Epochs:%f \n" % epochs)
f.write("Model name:%s \n" % model_name)
f.write("Regularization:%s \n" % regularization_penalty)
f.close()
save_model_file_ = output_directory + save_model_file
load_model_file_ = output_directory + load_model_file
assert save_model_file != load_model_file, "Files cannot be the same"
n_tasks, task_size, dim, channels = test_data_ML.x.shape if is_test else validation_data_ML.x.shape
horizon = 10
test_loss_list1 = []
test_loss_list2 = []
initial_test_loss_list1 = []
initial_test_loss_list2 = []
if task_size == 50:
step = n_tasks // 100
else:
step = 1
for task_id in range(0, (n_tasks - horizon - 1), step):
#check that all files blong to the same domain
temp_file_idx = test_data_ML.file_idx[
task_id:task_id + horizon +
1] if is_test else validation_data_ML.file_idx[
task_id:task_id + horizon + 1]
if (len(np.unique(temp_file_idx)) > 1):
continue
if is_test:
temp_x_train = test_data_ML.x[task_id]
temp_y_train = test_data_ML.y[task_id]
temp_x_test1 = test_data_ML.x[(task_id +
1):(task_id + horizon +
1)].reshape(
-1, dim, channels)
temp_y_test1 = test_data_ML.y[(task_id +
1):(task_id + horizon +
1)].reshape(-1, 1)
temp_x_test2 = test_data_ML.x[(task_id + 1)].reshape(
-1, dim, channels)
temp_y_test2 = test_data_ML.y[(task_id + 1)].reshape(-1, 1)
else:
temp_x_train = validation_data_ML.x[task_id]
temp_y_train = validation_data_ML.y[task_id]
temp_x_test1 = validation_data_ML.x[(task_id +
1):(task_id +
horizon +
1)].reshape(
-1, dim,
channels)
temp_y_test1 = validation_data_ML.y[(task_id +
1):(task_id +
horizon +
1)].reshape(
-1, 1)
temp_x_test2 = validation_data_ML.x[(task_id + 1)].reshape(
-1, dim, channels)
temp_y_test2 = validation_data_ML.y[(task_id + 1)].reshape(
-1, 1)
if model_name == "FCN":
kernels = [8, 5, 3
] if dataset_name != "POLLUTION" else [4, 2, 1]
temp_x_train = np.transpose(temp_x_train, [0, 2, 1])
temp_x_test1 = np.transpose(temp_x_test1, [0, 2, 1])
temp_x_test2 = np.transpose(temp_x_test2, [0, 2, 1])
#temp_x_val = np.transpose(temp_x_val, [0,2,1])
early_stopping = EarlyStopping(patience=patience_stopping,
model_file=save_model_file_,
verbose=verbose)
if model_name == "LSTM":
model = LSTMModel(batch_size=batch_size,
seq_len=window_size,
input_dim=input_dim,
n_layers=2,
hidden_dim=120,
output_dim=1)
elif model_name == "FCN":
model = FCN(time_steps=window_size,
channels=[input_dim, 128, 128, 128],
kernels=kernels)
model.load_state_dict(torch.load(load_model_file_))
train_loader = DataLoader(
SimpleDataset(x=temp_x_train, y=temp_y_train), **params)
test_loader1 = DataLoader(
SimpleDataset(x=temp_x_test1, y=temp_y_test1), **params)
test_loader2 = DataLoader(
SimpleDataset(x=temp_x_test2, y=temp_y_test2), **params)
verbose = False
model.cuda()
initial_loss1 = test(model, test_loader1, output_directory,
load_model_file_, verbose)
initial_loss2 = test(model, test_loader2, output_directory,
load_model_file_, verbose)
if freeze_model_flag:
freeze_model(model)
#early_stopping(initial_loss, model)
train(model, train_loader, test_loader1, early_stopping,
learning_rate, epochs, regularization_penalty)
early_stopping(0.0, model)
loss1 = test(model, test_loader1, output_directory,
save_model_file_, verbose, True)
loss2 = test(model, test_loader2, output_directory,
save_model_file_, verbose, True)
print(loss1)
test_loss_list1.append(loss1)
initial_test_loss_list1.append(initial_loss1)
test_loss_list2.append(loss2)
initial_test_loss_list2.append(initial_loss2)
f = open(output_directory + "/results.txt", "a+")
#f.write("Initial Test error1 :%f \n"% np.mean(initial_test_loss_list1))
f.write("Test error1: %f \n" % np.mean(test_loss_list1))
f.write("Standard deviation1: %f \n" % np.std(test_loss_list1))
#f.write("Initial Test error2 :%f \n"% np.mean(initial_test_loss_list2))
f.write("Test error2: %f \n" % np.mean(test_loss_list2))
f.write("Standard deviation2: %f \n" % np.std(test_loss_list2))
f.write("\n")
f.close()
#np.save("test_loss_wtf_"+model_name+"_"+dataset_name+"_"+str(trial)+".npy", test_loss_list)
elif mode == "50":
assert save_model_file != load_model_file, "Files cannot be the same"
with open(output_directory + "/results.txt", "a+") as f:
f.write("Dataset :%s" % dataset_name)
f.write("\n")
save_model_file_ = output_directory + save_model_file
load_model_file_ = output_directory + load_model_file
if is_test == 0:
test_data = validation_data
train_idx, val_idx, test_idx = split_idx_50_50(
test_data.file_idx) if is_test else split_idx_50_50(
validation_data.file_idx)
n_domains_in_test = np.max(test_data.file_idx) + 1
test_loss_list = []
initial_test_loss_list = []
for domain in range(n_domains_in_test):
print("Domain:", domain)
f = open(output_directory + "/results.txt", "a+")
f.write("Mode :%d" % domain)
f.write("\n")
f.close()
x_test = test_data.x
y_test = test_data.y
temp_train_data = SimpleDataset(x=x_test[train_idx[domain]],
y=y_test[train_idx[domain]])
temp_val_data = SimpleDataset(x=x_test[val_idx[domain]],
y=y_test[val_idx[domain]])
temp_test_data = SimpleDataset(x=x_test[test_idx[domain]],
y=y_test[test_idx[domain]])
early_stopping = EarlyStopping(patience=patience_stopping,
model_file=save_model_file_,
verbose=True)
if model_name == "LSTM":
model = LSTMModel(batch_size=batch_size,
seq_len=window_size,
input_dim=input_dim,
n_layers=2,
hidden_dim=120,
output_dim=1)
elif model_name == "FCN":
kernels = [8, 5, 3
] if dataset_name != "POLLUTION" else [4, 2, 1]
model = FCN(time_steps=window_size,
channels=[input_dim, 128, 128, 128],
kernels=kernels)
temp_train_data.x = np.transpose(temp_train_data.x,
[0, 2, 1])
temp_test_data.x = np.transpose(temp_test_data.x,
[0, 2, 1])
temp_val_data.x = np.transpose(temp_val_data.x, [0, 2, 1])
if freeze_model_flag:
freeze_model(model)
model.load_state_dict(torch.load(load_model_file_))
model.cuda()
temp_train_loader = DataLoader(temp_train_data, **params)
temp_val_loader = DataLoader(temp_val_data, **params)
temp_test_loader = DataLoader(temp_test_data, **params)
initial_loss = test(model, temp_test_loader, output_directory,
load_model_file_, False)
train(model, temp_train_loader, temp_val_loader,
early_stopping, learning_rate, epochs,
regularization_penalty)
loss = test(model, temp_test_loader, output_directory,
save_model_file_, True, True)
initial_test_loss_list.append(initial_loss)
test_loss_list.append(loss)
total_loss = np.mean(test_loss_list)
initial_loss = np.mean(initial_test_loss_list)
f = open(output_directory + "/results.txt", "a+")
f.write("Total error :%f \n" % total_loss)
f.write("Learning rate: %f \n" % learning_rate)
f.write("Initial Total error :%f \n" % initial_loss)
f.write("Std: %f\n" % np.std(test_loss_list))
f.write("\n")
f.close()
def test2(maml, model, model_name, dataset_name, test_data_ML, adaptation_steps, learning_rate, noise_level, noise_type, is_test = True, horizon = 10):
total_tasks_test = len(test_data_ML)
error_list = []
learner = maml.clone() # Creates a clone of model
learner.cuda()
accum_error = 0.0
accum_std = 0.0
count = 0.0
grid = [0., noise_level]
input_dim = test_data_ML.x.shape[-1]
window_size = test_data_ML.x.shape[-2]
output_dim = test_data_ML.y.shape[-1]
if is_test:
step = total_tasks_test//100
else:
step = 1
step = 1 if step == 0 else step
for task in range(0, (total_tasks_test-horizon-1), step):
temp_file_idx = test_data_ML.file_idx[task:task+horizon+1]
if(len(np.unique(temp_file_idx))>1):
continue
if model_name == "LSTM":
model2 = LSTMModel( batch_size=None, seq_len = None, input_dim = input_dim, n_layers = 2, hidden_dim = 120, output_dim =1)
elif model_name == "LSTM_MRA":
model2 = LSTMModel_MRA( batch_size=None, seq_len = window_size, input_dim = input_dim, n_layers = 2, hidden_dim = 120, output_dim =1)
elif model_name == "FCN":
kernels = [8,5,3] if window_size != 5 else [4,2,1]
model2 = FCN(time_steps = window_size, channels=[input_dim, 128, 128, 128] , kernels=kernels)
#model2.cuda()
#model2.load_state_dict(copy.deepcopy(maml.module.state_dict()))
#opt2 = optim.Adam(model2.parameters(), lr=learning_rate)
learner = maml.clone()
x_spt, y_spt = test_data_ML[task]
x_qry = test_data_ML.x[(task+1):(task+1+horizon)].reshape(-1, window_size, input_dim)
y_qry = test_data_ML.y[(task+1):(task+1+horizon)].reshape(-1, output_dim)
#x_qry = test_data_ML.x[(task+1)].reshape(-1, window_size, input_dim)
#y_qry = test_data_ML.y[(task+1)].reshape(-1, output_dim)
if model_name == "FCN":
x_qry = np.transpose(x_qry, [0,2,1])
x_spt = np.transpose(x_spt, [0,2,1])
x_spt, y_spt = to_torch(x_spt), to_torch(y_spt)
x_qry = to_torch(x_qry)
y_qry = to_torch(y_qry)
epsilon = grid[np.random.randint(0,len(grid))]
if noise_type == "additive":
y_spt = y_spt+epsilon
y_qry = y_qry+epsilon
else:
y_spt = y_spt*(1+epsilon)
y_qry = y_qry*(1+epsilon)
#learner.module.train()
#model2.eval()
for step in range(adaptation_steps):
#model2.train()
pred = learner(model.encoder(x_spt))
error = mae(pred, y_spt)
#opt2.zero_grad()
#error.backward()
learner.adapt(error)
#opt2.step()
#model2.eval()
#learner.module.eval()
y_pred = learner(model.encoder(x_qry))
y_pred = torch.clamp(y_pred, 0, 1)
error = mae(y_pred, y_qry)
accum_error += error.data
accum_std += error.data**2
count += 1
error = accum_error/count
print("std:", accum_std/count)
return error
def main(args):
meta_info = {
"POLLUTION": [5, 50, 14],
"HR": [32, 50, 13],
"BATTERY": [20, 50, 3]
}
output_directory = "output/"
horizon = 10
output_dim = 1
dataset_name = args.dataset
save_model_file = args.save_model_file
load_model_file = args.load_model_file
lower_trial = args.lower_trial
upper_trial = args.upper_trial
learning_rate = args.learning_rate
meta_learning_rate = args.meta_learning_rate
adaptation_steps = args.adaptation_steps
batch_size = args.batch_size
model_name = args.model
is_test = args.is_test
patience_stopping = args.stopping_patience
epochs = args.epochs
noise_level = args.noise_level
noise_type = args.noise_type
ml_horizon = args.ml_horizon
experiment_id = args.experiment_id
window_size, task_size, input_dim = meta_info[dataset_name]
task_size = args.task_size
assert model_name in ("FCN", "LSTM"), "Model was not correctly specified"
assert dataset_name in ("POLLUTION", "HR", "BATTERY")
grid = [0., noise_level]
train_data = pickle.load(
open(
"../../Data/TRAIN-" + dataset_name + "-W" + str(window_size) +
"-T" + str(task_size) + "-NOML.pickle", "rb"))
train_data_ML = pickle.load(
open(
"../../Data/TRAIN-" + dataset_name + "-W" + str(window_size) +
"-T" + str(task_size) + "-ML.pickle", "rb"))
validation_data = pickle.load(
open(
"../../Data/VAL-" + dataset_name + "-W" + str(window_size) + "-T" +
str(task_size) + "-NOML.pickle", "rb"))
validation_data_ML = pickle.load(
open(
"../../Data/VAL-" + dataset_name + "-W" + str(window_size) + "-T" +
str(task_size) + "-ML.pickle", "rb"))
test_data = pickle.load(
open(
"../../Data/TEST-" + dataset_name + "-W" + str(window_size) +
"-T" + str(task_size) + "-NOML.pickle", "rb"))
test_data_ML = pickle.load(
open(
"../../Data/TEST-" + dataset_name + "-W" + str(window_size) +
"-T" + str(task_size) + "-ML.pickle", "rb"))
loss_fn = mae
results_list = []
results_dict = {}
results_dict["Experiment_id"] = experiment_id
results_dict["Model"] = model_name
results_dict["Dataset"] = dataset_name
results_dict["Learning rate"] = learning_rate
results_dict["Noise level"] = noise_level
results_dict["Task size"] = task_size
results_dict["Evaluation loss"] = "MAE Test"
results_dict["Vrae weight"] = "-"
results_dict["Training"] = "MAML"
results_dict["ML-Horizon"] = ml_horizon
results_dict["Meta-learning rate"] = meta_learning_rate
#loss_fn = nn.SmoothL1Loss()
for trial in range(lower_trial, upper_trial):
output_directory = "../../Models/" + dataset_name + "_" + model_name + "_MAML/" + str(
trial) + "/"
save_model_file_ = output_directory + experiment_id + "_" + "encoder_" + save_model_file
save_model_file_2 = output_directory + experiment_id + "_" + save_model_file
load_model_file_ = output_directory + load_model_file
try:
os.mkdir(output_directory)
except OSError as error:
print(error)
with open(output_directory + "/results3.txt", "a+") as f:
f.write("Learning rate :%f \n" % learning_rate)
f.write("Meta-learning rate: %f \n" % meta_learning_rate)
f.write("Adaptation steps: %f \n" % adaptation_steps)
f.write("Noise level: %f \n" % noise_level)
f.write("\n")
if model_name == "LSTM":
model = LSTMModel(batch_size=batch_size,
seq_len=window_size,
input_dim=input_dim,
n_layers=2,
hidden_dim=120,
output_dim=1)
model2 = nn.Linear(120, 1)
elif model_name == "FCN":
kernels = [8, 5, 3] if dataset_name != "POLLUTION" else [4, 2, 1]
model = FCN(time_steps=window_size,
channels=[input_dim, 128, 128, 128],
kernels=kernels)
model2 = nn.Linear(128, 1)
model.cuda()
model2.cuda()
maml = l2l.algorithms.MAML(model2, lr=learning_rate, first_order=False)
opt = optim.Adam(list(maml.parameters()) + list(model.parameters()),
lr=meta_learning_rate)
#torch.backends.cudnn.enabled = False
total_num_tasks = train_data_ML.x.shape[0]
#test2(maml, model_name, dataset_name, test_data_ML, adaptation_steps, learning_rate)
#val_error = test(maml, model_name, dataset_name, validation_data_ML, adaptation_steps, learning_rate)
early_stopping = EarlyStopping(patience=patience_stopping,
model_file=save_model_file_,
verbose=True)
early_stopping2 = EarlyStopping(patience=patience_stopping,
model_file=save_model_file_2,
verbose=True)
#scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(opt, patience =200, verbose=True)
#early_stopping(val_error, maml)
for iteration in range(epochs):
# Creates a clone of model
opt.zero_grad()
iteration_error = 0.0
print(iteration)
for task in range(batch_size):
learner = maml.clone()
task = np.random.randint(0, total_num_tasks - horizon)
if train_data_ML.file_idx[task +
1] != train_data_ML.file_idx[task]:
continue
#task_qry = np.random.randint(1,horizon+1)
x_spt, y_spt = train_data_ML[task]
#x_qry, y_qry = train_data_ML[(task+1):(task+1+horizon)]
x_qry, y_qry = train_data_ML[task + ml_horizon]
#x_qry, y_qry = train_data_ML[task_qry]
x_qry = x_qry.reshape(-1, window_size, input_dim)
y_qry = y_qry.reshape(-1, output_dim)
if model_name == "FCN":
x_qry = np.transpose(x_qry, [0, 2, 1])
x_spt = np.transpose(x_spt, [0, 2, 1])
x_spt, y_spt = to_torch(x_spt), to_torch(y_spt)
x_qry = to_torch(x_qry)
y_qry = to_torch(y_qry)
#data augmentation
epsilon = grid[np.random.randint(0, len(grid))]
if noise_type == "additive":
y_spt = y_spt + epsilon
y_qry = y_qry + epsilon
else:
y_spt = y_spt * (1 + epsilon)
y_qry = y_qry * (1 + epsilon)
# Fast adapt
for _ in range(adaptation_steps):
pred = learner(model.encoder(x_spt))
error = loss_fn(pred, y_spt)
learner.adapt(
error) #, allow_unused=True)#, allow_nograd=True)
pred = learner(model.encoder(x_qry))
evaluation_error = loss_fn(pred, y_qry)
iteration_error += evaluation_error
#evaluation_error.backward()
# Meta-update the model parameters
#for p in maml.parameters():
#p.grad.data.mul_(1.0 / batch_size)
iteration_error /= batch_size
iteration_error.backward() #retain_graph = True)
#print("loss iteration:",iteration_error)
opt.step()
if iteration % 1 == 0:
val_error = test2(loss_fn,
maml,
model,
model_name,
dataset_name,
validation_data_ML,
adaptation_steps,
learning_rate,
noise_level,
noise_type,
horizon=10)
test_error = test2(loss_fn,
maml,
model,
model_name,
dataset_name,
test_data_ML,
adaptation_steps,
learning_rate,
0,
noise_type,
horizon=10)
#scheduler.step(val_error)
print("Val error:", val_error)
print("Test error:", test_error)
if iteration > 10:
early_stopping(val_error, model)
early_stopping2(val_error, maml)
if early_stopping.early_stop:
print("Early stopping")
break
model.load_state_dict(torch.load(save_model_file_))
maml.load_state_dict(torch.load(save_model_file_2))
validation_error = test2(loss_fn, maml, model, model_name,
dataset_name, validation_data_ML,
adaptation_steps, learning_rate, 0,
noise_type)
initial_val_error = test2(loss_fn, maml, model, model_name,
dataset_name, validation_data_ML, 0,
learning_rate, 0, noise_type)
test_error = test2(loss_fn, maml, model, model_name, dataset_name,
test_data_ML, adaptation_steps, learning_rate, 0,
noise_type)
initial_test_error = test2(loss_fn, maml, model, model_name,
dataset_name, test_data_ML, 0,
learning_rate, 0, noise_type)
test_error2 = test2(loss_fn,
maml,
model,
model_name,
dataset_name,
test_data_ML,
adaptation_steps,
learning_rate,
0,
noise_type,
horizon=1)
initial_test_error2 = test2(loss_fn,
maml,
model,
model_name,
dataset_name,
test_data_ML,
0,
learning_rate,
0,
noise_type,
horizon=1)
with open(output_directory + "/results3.txt", "a+") as f:
f.write("Dataset :%s \n" % dataset_name)
f.write("Test error: %f \n" % test_error)
f.write("Test error2: %f \n" % test_error2)
f.write("Initial Test error: %f \n" % initial_test_error)
f.write("Initial Test error2: %f \n" % initial_test_error2)
f.write("Validation error: %f \n" % validation_error)
f.write("Initial validation error: %f \n" % initial_val_error)
f.write("\n")
print("Adaptation_steps:", adaptation_steps)
temp_results_dict = copy.copy(results_dict)
temp_results_dict["Trial"] = trial
temp_results_dict["Adaptation steps"] = adaptation_steps
temp_results_dict["Horizon"] = 10
temp_results_dict["Value"] = float(test_error)
temp_results_dict["Val error"] = float(validation_error)
temp_results_dict["Final_epoch"] = iteration
results_list.append(temp_results_dict)
temp_results_dict = copy.copy(results_dict)
temp_results_dict["Trial"] = trial
temp_results_dict["Adaptation steps"] = 0
temp_results_dict["Horizon"] = 10
temp_results_dict["Value"] = float(initial_test_error)
temp_results_dict["Val error"] = float(initial_val_error)
temp_results_dict["Final_epoch"] = iteration
results_list.append(temp_results_dict)
temp_results_dict = copy.copy(results_dict)
temp_results_dict["Trial"] = trial
temp_results_dict["Adaptation steps"] = adaptation_steps
temp_results_dict["Horizon"] = 1
temp_results_dict["Value"] = float(test_error2)
temp_results_dict["Final_epoch"] = iteration
results_list.append(temp_results_dict)
temp_results_dict = copy.copy(results_dict)
temp_results_dict["Trial"] = trial
temp_results_dict["Adaptation steps"] = 0
temp_results_dict["Horizon"] = 1
temp_results_dict["Value"] = float(initial_test_error2)
temp_results_dict["Final_epoch"] = iteration
results_list.append(temp_results_dict)
try:
os.mkdir("../../Results/json_files/")
except OSError as error:
print(error)
with open("../../Results/json_files/" + experiment_id + ".json",
'w') as outfile:
json.dump(results_list, outfile)