def predict(t_net: DaRnnNet, t_dat: TrainData, train_size: int, batch_size: int, T: int, on_train=False):
out_size = t_dat.targs.shape[1]
if on_train:
y_pred = np.zeros((train_size - T + 1, out_size))
else:
y_pred = np.zeros((t_dat.feats.shape[0] - train_size, out_size))
for y_i in range(0, len(y_pred), batch_size):
y_slc = slice(y_i, y_i + batch_size)
batch_idx = range(len(y_pred))[y_slc]
b_len = len(batch_idx)
X = np.zeros((b_len, T - 1, t_dat.feats.shape[1]))
y_history = np.zeros((b_len, T - 1, t_dat.targs.shape[1]))
for b_i, b_idx in enumerate(batch_idx):
if on_train:
idx = range(b_idx, b_idx + T - 1)
else:
idx = range(b_idx + train_size - T, b_idx + train_size - 1)
X[b_i, :, :] = t_dat.feats[idx, :]
y_history[b_i, :] = t_dat.targs[idx]
y_history = numpy_to_tvar(y_history)
_, input_encoded = t_net.encoder(numpy_to_tvar(X))
y_pred[y_slc] = t_net.decoder(input_encoded, y_history).cpu().data.numpy()
return y_pred
def train_iteration(t_net: DaRnnNet, loss_func: typing.Callable, X, y_history, y_target):
t_net.enc_opt.zero_grad()
t_net.dec_opt.zero_grad()
input_weighted, input_encoded = t_net.encoder(numpy_to_tvar(X))
y_pred = t_net.decoder(input_encoded, numpy_to_tvar(y_history))
y_true = numpy_to_tvar(y_target)
loss = loss_func(y_pred, y_true)
loss.backward()
t_net.enc_opt.step()
t_net.dec_opt.step()
return loss.item()
def da_rnn(train_data: TrainData, n_targs: int,
encoder_hidden_size=64, decoder_hidden_size=64,
T=10, learning_rate=0.01, batch_size=128):
train_cfg = TrainConfig(T, int(train_data.feats.shape[0] * 0.7), batch_size, nn.MSELoss())
logger.info(f"Training size: {train_cfg.train_size:d}.")
enc_kwargs = {"input_size": train_data.feats.shape[1], "hidden_size": encoder_hidden_size, "T": T}
encoder = Encoder(**enc_kwargs).to(device)
with open(os.path.join("data", "enc_kwargs.json"), "w") as f:
json.dump(enc_kwargs, f, indent=4)
dec_kwargs = {"encoder_hidden_size": encoder_hidden_size,
"decoder_hidden_size": decoder_hidden_size, "T": T, "out_feats": n_targs}
decoder = Decoder(**dec_kwargs).to(device)
with open(os.path.join("data", "dec_kwargs.json"), "w") as f:
json.dump(dec_kwargs, f, indent=4)
encoder_optimizer = optim.Adam(
params=[p for p in encoder.parameters() if p.requires_grad],
lr=learning_rate)
decoder_optimizer = optim.Adam(
params=[p for p in decoder.parameters() if p.requires_grad],
lr=learning_rate)
da_rnn_net = DaRnnNet(encoder, decoder, encoder_optimizer, decoder_optimizer)
return train_cfg, da_rnn_net
def predict(t_net: DaRnnNet,
t_dat: TrainData,
train_size: int,
batch_size: int,
T: int,
on_train=False,
eval=False):
out_size = t_dat.targs.shape[1]
if on_train:
y_pred = np.zeros((train_size - T + 1, out_size))
else:
y_pred = np.zeros((t_dat.feats.shape[0] - train_size, out_size))
#y_pred = np.zeros((VALI_SIZE, out_size))
if eval:
y_pred = np.zeros((VALI_SIZE, out_size))
for y_i in range(0, len(y_pred), batch_size):
y_slc = slice(y_i, y_i + batch_size)
batch_idx = range(len(y_pred))[y_slc]
b_len = len(batch_idx)
X = np.zeros((b_len, T - 1, t_dat.feats.shape[1]))
y_history = np.zeros((b_len, T - 1, t_dat.targs.shape[1]))
for b_i, b_idx in enumerate(batch_idx):
if on_train:
idx = range(b_idx, b_idx + T - 1)
else:
# ANDREA --> The validation set is chosen at random
# b_idx = np.random.randint(0, len(t_dat.feats)-train_size)
idx = range(b_idx + train_size - T, b_idx + train_size - 1)
X[b_i, :, :] = t_dat.feats[idx, :]
## Leave it zeros
# y_history[b_i, :] = t_dat.targs[idx]
y_history = numpy_to_tvar(y_history)
_, input_encoded = t_net.encoder(numpy_to_tvar(X))
y_pred[y_slc] = t_net.decoder(input_encoded,
y_history).cpu().data.numpy()
return y_pred
def train_iteration(t_net: DaRnnNet, loss_func: typing.Callable, X, y_history,
y_target):
t_net.enc_opt.zero_grad()
t_net.dec_opt.zero_grad()
X = Variable(torch.from_numpy(X).type(torch.FloatTensor).to(device))
y_history = Variable(
torch.from_numpy(y_history).type(torch.FloatTensor).to(device))
y_target = Variable(
torch.from_numpy(y_target).type(torch.FloatTensor).to(device))
input_weighted, input_encoded = t_net.encoder(X)
y_pred = t_net.decoder(input_encoded, y_history)
y_true = y_target
loss = loss_func(y_pred, y_true)
loss.backward()
t_net.enc_opt.step()
t_net.dec_opt.step()
return loss.item()
def set_params(train_data, device, **da_rnn_kwargs):
train_configs = TrainConfig(da_rnn_kwargs["time_step"],
int(train_data.shape[0] * 0.95),
da_rnn_kwargs["batch_size"],
nn.MSELoss())
enc_kwargs = {
"input_size": train_data.shape[1],
"hidden_size": da_rnn_kwargs["en_hidden_size"],
"time_step":
int(da_rnn_kwargs["time_step"] / self.predict_size)
}
dec_kwargs = {
"encoder_hidden_size": da_rnn_kwargs["en_hidden_size"],
"decoder_hidden_size": da_rnn_kwargs["de_hidden_size"],
"time_step":
int(da_rnn_kwargs["time_step"] / self.predict_size),
"out_feats": da_rnn_kwargs["target_cols"]
}
encoder = Encoder(**enc_kwargs).to(device)
decoder = Decoder(**dec_kwargs).to(device)
encoder_optimizer = optim.Adam(
params=[p for p in encoder.parameters() if p.requires_grad],
lr=da_rnn_kwargs["learning_rate"],
betas=(0.9, 0.999),
eps=1e-08)
decoder_optimizer = optim.Adam(
params=[p for p in decoder.parameters() if p.requires_grad],
lr=da_rnn_kwargs["learning_rate"],
betas=(0.9, 0.999),
eps=1e-08)
da_rnn_net = DaRnnNet(encoder, decoder, encoder_optimizer,
decoder_optimizer)
return train_configs, da_rnn_net
def da_rnn(train_data: TrainData,
n_targs: int,
learning_rate=0.01,
encoder_hidden_size=64,
decoder_hidden_size=64,
T=10,
batch_size=128):
# passed arguments are data, n_targs=len(targ_cols), learning_rate=.001, **da_rnn_kwargs
#here n_args : int means that this argument takes only an integer as its value
#train_data = TrainData means that this train_data argument takes only the datatype TrainData that we have defined as its value
training_data_size_out_of_total = train_data.feats.shape[0] * 0.7
training_configuration = TrainConfig(T,
int(training_data_size_out_of_total),
batch_size, nn.MSELoss())
'''
class TrainConfig(typing.NamedTuple):
T: int
train_size: int
batch_size: int
loss_func: typing.Callable
'''
logger.info(f"Training size: {training_configuration.train_size:d}.")
encoder_kwargs = {
"input_size": train_data.feats.shape[1],
"hidden_size": encoder_hidden_size,
"T": T
}
encoder = Encoder(**encoder_kwargs).to(device)
with open(os.path.join("data", "enc_kwargs.json"), "w") as fi:
json.dump(encoder_kwargs, fi, indent=4)
decoder_kwargs = {
"encoder_hidden_size": encoder_hidden_size,
"decoder_hidden_size": decoder_hidden_size,
"T": T,
"out_feats": n_targs
}
decoder = Decoder(**decoder_kwargs).to(device)
with open(os.path.join("data", "dec_kwargs.json"), "w") as fi:
json.dump(decoder_kwargs, fi, indent=4)
encoder_optimizer = optim.Adam(
params=[p for p in encoder.parameters() if p.requires_grad],
lr=learning_rate)
decoder_optimizer = optim.Adam(
params=[p for p in decoder.parameters() if p.requires_grad],
lr=learning_rate)
da_rnn_net = DaRnnNet(
encoder, decoder, encoder_optimizer, decoder_optimizer
) #-------------------------------return the DA-RNN network
return training_configuration, da_rnn_net