features_train = y_train[:, :, :, 1:]
targets_val = y_val[:, :, :, [0]]
features_val = y_val[:, :, :, 1:]
targets_test = y_test[:, :, :, [0]]
features_test = y_test[:, :, :, 1:]
lstm = LSTM(input_size, hidden_size, output_size, n_layers, dropout)
if os.path.isfile(checkpoint_file):
print("Loading checkpoint...")
lstm.load_state_dict(torch.load(checkpoint_file))
if use_cuda:
lstm.cuda()
# optimizer = optim.Adam(lstm.parameters(), lr=lr)
#
# best_val_loss = 1000
# train_loss = 0
# for epoch in range(n_epochs):
# n_batches = x_train.shape[0]
# for i in range(n_batches):
# lstm.hidden = None
# input_batches = x_train[i]
# target_batches = targets_train[i]
# train_loss = train(input_batches, target_batches, lstm, optimizer, use_cuda)
#
# epoch_train_loss = evaluate(x_train, targets_train, lstm, use_cuda)
# epoch_val_loss = evaluate(x_val, targets_val, lstm, use_cuda)
def setup(config):
if config.task.name == 'copy':
task = CopyTask(
batch_size=config.task.batch_size,
min_len=config.task.min_len,
max_len=config.task.max_len,
bit_width=config.task.bit_width,
seed=config.task.seed,
)
elif config.task.name == 'repeat':
task = RepeatCopyTask(
batch_size=config.task.batch_size,
bit_width=config.task.bit_width,
min_len=config.task.min_len,
max_len=config.task.max_len,
min_rep=config.task.min_rep,
max_rep=config.task.max_rep,
norm_max=config.task.norm_max,
seed=config.task.seed,
)
elif config.task.name == 'recall':
task = AssociativeRecallTask(
batch_size=config.task.batch_size,
bit_width=config.task.bit_width,
item_len=config.task.item_len,
min_cnt=config.task.min_cnt,
max_cnt=config.task.max_cnt,
seed=config.task.seed,
)
else:
logging.info('Unknown task')
exit(0)
torch.manual_seed(config.model.seed)
if config.model.name == 'lstm':
model = LSTM(
n_inputs=task.full_input_width,
n_outputs=task.full_output_width,
n_hidden=config.model.n_hidden,
n_layers=config.model.n_layers,
)
elif config.model.name == 'ntm':
model = NTM(
input_size=task.full_input_width,
output_size=task.full_output_width,
mem_word_length=config.model.mem_word_length,
mem_cells_count=config.model.mem_cells_count,
n_writes=config.model.n_writes,
n_reads=config.model.n_reads,
controller_n_hidden=config.model.controller_n_hidden,
controller_n_layers=config.model.controller_n_layers,
clip_value=config.model.clip_value,
)
elif config.model.name == 'dnc':
model = DNC(
input_size=task.full_input_width,
output_size=task.full_output_width,
cell_width=config.model.cell_width,
n_cells=config.model.n_cells,
n_reads=config.model.n_reads,
controller_n_hidden=config.model.controller_n_hidden,
controller_n_layers=config.model.controller_n_layers,
clip_value=config.model.clip_value,
masking=config.model.masking,
mask_min=config.model.mask_min,
dealloc=config.model.dealloc,
diff_alloc=config.model.diff_alloc,
links=config.model.links,
links_sharpening=config.model.links_sharpening,
normalization=config.model.normalization,
dropout=config.model.dropout,
)
else:
logging.info('Unknown model')
exit(0)
if config.gpu and torch.cuda.is_available():
model = model.cuda()
# Setup optimizer
if config.optimizer == 'sgd':
optimizer = torch.optim.SGD(model.parameters(),
lr=config.learning_rate,
momentum=config.momentum)
if config.optimizer == 'rmsprop':
optimizer = torch.optim.RMSprop(
model.parameters(),
lr=config.learning_rate,
momentum=config.momentum,
)
if config.optimizer == 'adam':
optimizer = torch.optim.Adam(model.parameters(),
lr=config.learning_rate)
step = 0
if config.load:
logging.info('Restoring model from checkpoint')
model, optimizer, task, step = utils.load_checkpoint(
model,
optimizer,
task,
config.load,
)
return model, optimizer, task, step
dset,
batch_size=batch_size,
shuffle=False,
num_workers=4,
pin_memory=True # CUDA only
)
# Network Definition + Optimizer + Scheduler
model = LSTM(hidden_size=n_hidden1,
hidden_size2=n_hidden2,
num_securities=n_stocks,
dropout=0.2,
n_layers=2,
T=T)
if use_cuda:
model.cuda()
optimizer = optim.RMSprop(model.parameters(),
lr=learning_rate,
weight_decay=0.0) # n
scheduler_model = lr_scheduler.StepLR(optimizer, step_size=1, gamma=1.0)
# loss function
criterion = nn.MSELoss(size_average=True).cuda()
# Store successive losses
losses = []
it = 0
for i in range(max_epochs):
loss_ = 0.
# Store current predictions
predicted = []
gt = []
def setup_model(config):
# Load data
if config.task.name == 'arithmetic':
train_data = Arithmetic(
batch_size=config.task.batch_size,
min_len=config.task.min_len,
max_len=config.task.max_len,
task=config.task.task,
seed=config.seed,
)
np.random.seed(config.seed)
params = [20, 30, 40, 60]
validation_data = []
for length in params:
example = train_data.gen_batch(batch_size=50,
min_len=length,
max_len=length,
distribution=np.array([
1,
]))
validation_data.append((example, length))
loss = Arithmetic.loss
else:
logging.info('Unknown task')
exit(0)
# Setup model
torch.manual_seed(config.seed)
if config.model.name == 'lstm':
model = LSTM(
n_inputs=train_data.symbols_amount,
n_outputs=train_data.symbols_amount,
n_hidden=config.model.n_hidden,
n_layers=config.model.n_layers,
)
elif config.model.name == 'ntm':
model = NTM(input_size=train_data.symbols_amount,
output_size=train_data.symbols_amount,
mem_word_length=config.model.mem_word_length,
mem_cells_count=config.model.mem_cells_count,
n_writes=config.model.n_writes,
n_reads=config.model.n_reads,
controller_n_hidden=config.model.controller_n_hidden,
controller_n_layers=config.model.controller_n_layers,
controller=config.model.controller,
layer_sizes=config.model.layer_sizes,
controller_output=config.model.controller_output,
clip_value=config.model.clip_value,
dropout=config.model.dropout)
elif config.model.name == 'dnc':
model = DNC(
input_size=train_data.symbols_amount,
output_size=train_data.symbols_amount,
n_cells=config.model.n_cells,
cell_width=config.model.cell_width,
n_reads=config.model.n_reads,
controller_n_hidden=config.model.controller_n_hidden,
controller_n_layers=config.model.controller_n_layers,
clip_value=config.model.clip_value,
)
else:
logging.info('Unknown model')
exit(0)
if config.gpu and torch.cuda.is_available():
model = model.cuda()
logging.info('Loaded model')
logging.info('Total number of parameters %d', model.calculate_num_params())
# Setup optimizer
if config.optimizer == 'sgd':
optimizer = torch.optim.SGD(model.parameters(),
lr=config.learning_rate,
momentum=config.momentum)
if config.optimizer == 'rmsprop':
optimizer = torch.optim.RMSprop(
model.parameters(),
lr=config.learning_rate,
momentum=config.momentum,
)
if config.optimizer == 'adam':
optimizer = torch.optim.Adam(model.parameters(),
lr=config.learning_rate)
if config.scheduler is not None:
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer,
mode='min',
factor=config.scheduler.factor,
patience=config.scheduler.patience,
verbose=config.scheduler.verbose,
threshold=config.scheduler.threshold,
)
optimizer = (optimizer, scheduler)
if config.load:
model, optimizer, train_data, step = utils.load_checkpoint(
model,
optimizer,
train_data,
config.load,
)
return model, optimizer, loss, train_data, validation_data
