Skip to content

TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting.

Python

Add finetuning support

misha-chertushkin65700643df89
diff --git a/notebooks/finetuning_torch.py b/notebooks/finetuning_torch.py
new file mode 100644
index 0000000..50a7f4d
--- /dev/null
+++ b/notebooks/finetuning_torch.py
@@ -0,0 +1,274 @@
+# Filename: tutorial_timesfm.py
+
+import yfinance as yf
+import numpy as np
+import pandas as pd
+import torch
+from torch.utils.data import Dataset, DataLoader
+import torch.optim as optim
+import timesfm
+from os import path
+from typing import Any, Sequence
+
+import numpy as np
+import torch
+from huggingface_hub import snapshot_download
+
+
+from timesfm.pytorch_patched_decoder import TimesFMConfig, PatchedTimeSeriesDecoder
+
+import torch
+import matplotlib.pyplot as plt
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+
+# --------------------------------------------------
+# 1. Download stock data via yfinance
+# --------------------------------------------------
+def download_yfinance_data(ticker="AAPL", start="2020-01-01", end="2022-01-01"):
+ """
+ Download daily stock data for a given ticker from Yahoo Finance.
+ Returns a pandas DataFrame with columns like 'Open', 'High', 'Low', 'Close', 'Volume'.
+ """
+ df = yf.download(ticker, start=start, end=end)
+ df = df.dropna()
+ return df["Close"].reset_index(drop=True)
+
+
+# --------------------------------------------------
+# 2. Create a dataset class for TimesFM
+# --------------------------------------------------
+class FinancialDataset(Dataset):
+ def __init__(
+ self,
+ series: pd.Series,
+ config: TimesFMConfig,
+ context_length=128, # how many past timesteps as input
+ horizon_length=32, # how many future steps to predict
+ ):
+ super().__init__()
+
+ self.series = series.values.astype(np.float32)
+ self.context_length = context_length
+ self.horizon_length = horizon_length
+ self.config = config
+
+ self.samples = []
+ # We want to ensure we have at least context_length + horizon_length points.
+ for start_idx in range(0, len(self.series) - (context_length + horizon_length)):
+ end_idx = start_idx + context_length
+ # context slice
+ x_context = self.series[start_idx:end_idx]
+ # future/horizon slice
+ x_future = self.series[end_idx : end_idx + horizon_length]
+ self.samples.append((x_context, x_future))
+
+ def __len__(self):
+ return len(self.samples)
+
+ def __getitem__(self, index):
+ x_context, x_future = self.samples[index]
+ # Convert to torch
+ x_context = torch.tensor(x_context, dtype=torch.float32)
+ x_future = torch.tensor(x_future, dtype=torch.float32)
+
+ input_padding = torch.zeros_like(x_context)
+
+ freq = torch.zeros(1, dtype=torch.long)
+
+ return x_context, input_padding, freq, x_future
+
+
+def collate_fn(batch):
+ xs_context = [item[0] for item in batch]
+ xs_padding = [item[1] for item in batch]
+ freqs = [item[2] for item in batch]
+ xs_future = [item[3] for item in batch]
+
+ x_context = torch.stack(xs_context, dim=0)
+ input_pad = torch.stack(xs_padding, dim=0)
+ freq = torch.stack(freqs, dim=0) # shape [B, 1]
+ x_future = torch.stack(xs_future, dim=0)
+
+ return x_context, input_pad, freq, x_future
+
+
+def get_model(*, load_weights: bool = False):
+ # standard model hack
+ repo_id = "google/timesfm-2.0-500m-pytorch"
+ tfm = timesfm.TimesFm(
+ hparams=timesfm.TimesFmHparams(
+ backend="cuda",
+ per_core_batch_size=32,
+ horizon_len=128,
+ num_layers=50,
+ use_positional_embedding=False,
+ context_len=192,
+ ),
+ checkpoint=timesfm.TimesFmCheckpoint(huggingface_repo_id=repo_id),
+ )
+
+ model = PatchedTimeSeriesDecoder(tfm._model_config)
+
+ if load_weights:
+ checkpoint_path = path.join(snapshot_download(repo_id), "torch_model.ckpt")
+ print(model.state_dict()["input_ff_layer.hidden_layer.0.weight"])
+ loaded_checkpoint = torch.load(checkpoint_path, weights_only=True)
+ model.load_state_dict(loaded_checkpoint)
+ print("After loading:")
+ print(model.state_dict()["input_ff_layer.hidden_layer.0.weight"])
+ model = model.to(device)
+
+ # import sys
+ # sys.exit(-1)
+ # repo_id = "google/timesfm-1.0-200m"
+ return model, tfm._model_config
+
+
+def train_model(
+ ticker="AAPL", start="2015-01-01", end="2022-01-01", train_split=0.8, batch_size=8, num_epochs=20, pretrained=False
+):
+ df_close = download_yfinance_data(ticker, start=start, end=end)
+ model, config = get_model(load_weights=pretrained)
+
+ total_len = len(df_close)
+ train_size = int(total_len * train_split)
+ val_size = total_len - train_size
+
+ train_series = df_close.iloc[:train_size].reset_index(drop=True)
+ val_series = df_close.iloc[train_size:].reset_index(drop=True)
+
+ train_dataset = FinancialDataset(
+ series=train_series, config=config, context_length=128, horizon_length=config.horizon_len
+ )
+ val_dataset = FinancialDataset(
+ series=val_series, config=config, context_length=128, horizon_length=config.horizon_len
+ )
+ print("Train samples:", len(train_dataset))
+ print("Val samples:", len(val_dataset))
+ train_dataloader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, collate_fn=collate_fn)
+ val_dataloader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False, collate_fn=collate_fn)
+
+ optimizer = optim.Adam(model.parameters(), lr=1e-4)
+
+ for epoch in range(num_epochs):
+ model.train()
+ total_train_loss = 0.0
+
+ for x_context, x_padding, freq, x_future in train_dataloader:
+ x_context, x_padding, freq, x_future = (
+ x_context.to(device),
+ x_padding.to(device),
+ freq.to(device),
+ x_future.to(device),
+ )
+ predictions = model(x_context, x_padding.float(), freq)
+ # predictions shape => [B, N, horizon_len, (1 + #quantiles)]
+ predictions_mean = predictions[..., 0] # => [B, N, horizon_len]
+ last_patch_pred = predictions_mean[:, -1, :] # => [B, horizon_len]
+
+ # x_future => [B, horizon_len]
+ loss = torch.mean((last_patch_pred - x_future.squeeze(-1)) ** 2)
+
+ optimizer.zero_grad()
+ loss.backward()
+ optimizer.step()
+
+ total_train_loss += loss.item()
+
+ avg_train_loss = total_train_loss / len(train_dataloader)
+
+ # -------- Compute validation loss --------
+ model.eval()
+ total_val_loss = 0.0
+ with torch.no_grad():
+ for x_context, x_padding, freq, x_future in val_dataloader:
+ x_context, x_padding, freq, x_future = (
+ x_context.to(device),
+ x_padding.to(device),
+ freq.to(device),
+ x_future.to(device),
+ )
+ predictions = model(x_context, x_padding.float(), freq)
+ predictions_mean = predictions[..., 0]
+ last_patch_pred = predictions_mean[:, -1, :]
+ val_loss = torch.mean((last_patch_pred - x_future.squeeze(-1)) ** 2)
+ total_val_loss += val_loss.item()
+
+ avg_val_loss = total_val_loss / max(len(val_dataloader), 1)
+
+ print(f"[Epoch {epoch+1}] Train Loss: {avg_train_loss:.4f} | Val Loss: {avg_val_loss:.4f}")
+
+ torch.save(model.state_dict(), "timesfm_finetuned.pth")
+ return model, train_dataloader, val_dataloader
+
+
+def plot_predictions(model, dataloader):
+ model.eval()
+ with torch.no_grad():
+ x_context, x_padding, freq, x_future = next(iter(dataloader))
+ x_context, x_padding, freq, x_future = (
+ x_context.to(device),
+ x_padding.to(device),
+ freq.to(device),
+ x_future.to(device),
+ )
+ # Forward pass
+ predictions = model(x_context, x_padding.float(), freq)
+ # => [B, N, horizon_len, (1 + #quantiles)]
+ predictions_mean = predictions[..., 0] # => [B, N, horizon_len]
+ last_patch_prediction = predictions_mean[:, -1, :] # => [B, horizon_len]
+
+ # We'll plot only the first sample in the batch
+ i = 0
+ pred_vals = last_patch_prediction[i].cpu().numpy() # [horizon_len]
+ context_vals = x_context[i].cpu().numpy() # [context_len]
+ future_vals = x_future[i].cpu().numpy() # [horizon_len]
+
+ horizon_len = future_vals.shape[0]
+ context_len = context_vals.shape[0]
+
+ plt.figure(figsize=(10, 5))
+
+ # Plot context
+ plt.plot(range(context_len), context_vals, label="Context (History)", color="blue")
+
+ # Plot predicted future
+ plt.plot(
+ range(context_len, context_len + horizon_len),
+ pred_vals,
+ label="Predicted Future",
+ color="orange",
+ )
+
+ # Plot ground truth future
+ plt.plot(
+ range(context_len, context_len + horizon_len),
+ future_vals,
+ label="Ground Truth Future",
+ color="green",
+ linestyle="--",
+ )
+
+ plt.xlabel("Time")
+ plt.ylabel("Value")
+ plt.title("Model Forecast vs. Ground Truth")
+ plt.legend()
+ plt.show()
+ plt.savefig("pic_predictions.png")
+
+
+if __name__ == "__main__":
+ # Example usage
+ model, train_dl, val_dl = train_model(
+ ticker="AAPL",
+ start="2012-01-01",
+ end="2019-01-01",
+ train_split=0.8,
+ batch_size=256,
+ num_epochs=50,
+ pretrained=True,
+ )
+
+ plot_predictions(model, val_dl)