-
-
Save kongqi404/336a8031901a3cebc05585b78fa650ef to your computer and use it in GitHub Desktop.
probabilistic forcasting
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
class GaussianDistribution(pl.LightningModule): | |
def __init__(self, model: torch.nn.Module, seq_len: int, pred_len: int, | |
data_path: str, | |
quantile: list[float] = [0.1, 0.25, 0.5, 0.75, 0.9], | |
lr: float = 0.00001, | |
batch_size: int = 32, | |
pdf_path: str = "./") -> None: | |
super().__init__() | |
self.save_hyperparameters(ignore=["model"]) | |
self.model = model | |
self.data_pl = ElectricDataModule( | |
batch_size=batch_size, data_path=data_path) | |
self.lr = lr | |
self.pdf_str = pdf_path+self.model.__class__.__name__ + \ | |
str(time.strftime("%Y-%m-%d-%H%M%S", time.localtime()))+".pdf" | |
self.quantile = torch.as_tensor(quantile) | |
self.seq_len = seq_len | |
self.pred_len = pred_len | |
def forward(self, x): | |
self.model(x) | |
def configure_optimizers(self) -> torch.optim.Optimizer: | |
return torch.optim.Adam(self.model.parameters(), lr=self.lr) | |
def training_step(self, train_batch, batch_idx): | |
batch_x, batch_y, batch_x_mark, batch_y_mark = train_batch | |
mu, sigma = self.model(batch_x) | |
f_dim = 0 | |
mu_new = mu.detach() | |
distritution = torch.distributions.Normal( | |
mu_new[:, -self.pred_len:, f_dim], | |
sigma[:, -self.pred_len:, f_dim]) | |
mse_loss = F.mse_loss(mu[:, -self.pred_len:, f_dim], | |
batch_y[:, -self.pred_len:, f_dim]) | |
loss = - \ | |
distritution.log_prob( | |
batch_y[:, -self.pred_len:, f_dim]).mean() + mse_loss | |
self.log("train_loss", loss) | |
return loss | |
def validation_step(self, valid_batch, batch_idx): | |
batch_x, batch_y, batch_x_mark, batch_y_mark = valid_batch | |
mu, sigma = self.model(batch_x) | |
f_dim = 0 | |
distritution = torch.distributions.Normal( | |
mu[:, -self.pred_len:, f_dim], sigma[:, -self.pred_len:, f_dim]) | |
mse_loss = F.mse_loss(mu[:, -self.pred_len:, f_dim], | |
batch_y[:, -self.pred_len:, f_dim]) | |
loss = - \ | |
distritution.log_prob( | |
batch_y[:, -self.pred_len:, f_dim]).mean() + mse_loss | |
self.log("val_loss", loss) | |
def on_test_epoch_start(self): | |
self.pdf = PdfPages(self.pdf_str) | |
def test_step(self, test_batch, batch_idx): | |
batch_x, batch_y, batch_x_mark, batch_y_mark = test_batch | |
mu, sigma = self.model(batch_x) | |
distritution = torch.distributions.Normal( | |
mu[:, -self.pred_len:, :], sigma[:, -self.pred_len:, :]) | |
mse_loss = F.mse_loss(mu[:, -self.pred_len:, :], | |
batch_y[:, -self.pred_len:, :]) | |
loss = - \ | |
distritution.log_prob( | |
batch_y[:, -self.pred_len:, :]).mean() + mse_loss | |
self.log("test_loss", loss) | |
sample = distritution.sample((100,)) | |
self.quantile = self.quantile.to(batch_x.device) | |
sample_quantile = torch.quantile(sample, self.quantile, dim=0) | |
lower_ninety = sample_quantile[0] | |
lower_quarter = sample_quantile[1] | |
mean = distritution.mean | |
upper_quarter = sample_quantile[3] | |
upper_ninety = sample_quantile[4] | |
outputs_true = self.data_pl.inverse_transform( | |
data=mean.reshape(-1, mean.size(-1)).cpu().detach() | |
.numpy()).reshape(mean.size()) | |
lower_ninety = self.data_pl.inverse_transform( | |
data=lower_ninety.reshape(-1, lower_ninety.size(-1) | |
).cpu().detach() | |
.numpy()).reshape(lower_ninety.size()) | |
lower_quarter = self.data_pl.inverse_transform( | |
data=lower_quarter.reshape(-1, lower_quarter.size(-1) | |
).cpu().detach() | |
.numpy()).reshape(lower_quarter.size()) | |
upper_quarter = self.data_pl.inverse_transform( | |
data=upper_quarter.reshape(-1, upper_quarter.size(-1) | |
).cpu().detach() | |
.numpy()).reshape(upper_quarter.size()) | |
upper_ninety = self.data_pl.inverse_transform( | |
data=upper_ninety.reshape(-1, upper_ninety.size(-1) | |
).cpu().detach() | |
.numpy()).reshape(upper_ninety.size()) | |
batch_y_true = self.data_pl.inverse_transform( | |
data=batch_y.reshape(-1, batch_y.size(-1)).cpu().detach() | |
.numpy()).reshape(batch_y.size()) | |
batch_x_true = self.data_pl.inverse_transform( | |
data=batch_x.reshape(-1, batch_y.size(-1)).cpu().detach() | |
.numpy()).reshape(batch_x.size()) | |
outputs_true = np.maximum(outputs_true, 0) | |
outputs_true = outputs_true[:, -self.pred_len:, :] | |
batch_y_true = batch_y_true[:, -self.pred_len:, :] | |
lower_ninety = lower_ninety[:, -self.pred_len:, :] | |
lower_quarter = lower_quarter[:, -self.pred_len:, :] | |
upper_quarter = upper_quarter[:, -self.pred_len:, :] | |
upper_ninety = upper_ninety[:, -self.pred_len:, :] | |
self.log("test_acc", accuracy( | |
outputs_true[:, :, 0], batch_y_true[:, :, 0])) | |
self.log("test_rmse", rmse( | |
outputs_true[:, :, 0], batch_y_true[:, :, 0])) | |
self.log("test_mae", mae(outputs_true[:, :, 0], batch_y_true[:, :, 0])) | |
self.log("test_mape", mape( | |
outputs_true[:, :, 0], batch_y_true[:, :, 0])) | |
self.log("test_mse", mse(outputs_true[:, :, 0], batch_y_true[:, :, 0])) | |
x_axis = list(range(self.seq_len+self.pred_len)) | |
y_all = np.concatenate( | |
(batch_x_true, batch_y_true), axis=1) | |
for idx in np.random.choice(outputs_true.shape[0], | |
int(outputs_true.shape[0]*0.1), | |
replace=False): | |
plt.figure() | |
plt.plot(x_axis, y_all[idx, :, 0], label="true", | |
linewidth=0.75, alpha=0.75) | |
plt.plot(x_axis[-self.pred_len:], outputs_true[idx, :, 0], | |
label="pred", linewidth=0.75) | |
plt.fill_between(x=x_axis[-self.pred_len:], | |
y1=lower_ninety[idx, :, 0], | |
y2=upper_ninety[idx, :, 0], | |
alpha=0.25, label="90%", | |
color="green", linewidth=0.1) | |
plt.fill_between(x=x_axis[-self.pred_len:], | |
y1=lower_quarter[idx, :, 0], | |
y2=upper_quarter[idx, :, 0], | |
alpha=0.5, label="75%", | |
color="green", linewidth=0.1) | |
plt.legend(["true", "pred", "90%", "75%"]) | |
self.pdf.savefig() | |
plt.close() | |
def test_epoch_end(self, outputs): | |
self.pdf.close() |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment