Skip to content

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

Python

(1) Make config hashable. (2) Flax

siriuz4225bb4f20e8f6
diff --git a/pyproject.toml b/pyproject.toml
index 3cb0346..c5344be 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -24,13 +24,12 @@ optax = { version = ">=0.2.2", optional = true }
einshape = { version = ">=0.8.0", optional = true }
orbax-checkpoint = { version = ">=0.5.15", optional = true }
jaxtyping = { version = ">=0.2.29", optional = true }
-typeguard = { version = ">=4.3.0", optional = true }
jax = { version = ">=0.4.26", optional = true }
[tool.poetry.extras]
torch = ["torch"]
-flax = ["flax", "optax", "einshape", "orbax-checkpoint", "jaxtyping", "typeguard", "jax"]
+flax = ["flax", "optax", "einshape", "orbax-checkpoint", "jaxtyping", "jax"]
[tool.ruff]
line-length = 88
diff --git a/src/timesfm/configs.py b/src/timesfm/configs.py
index 104ea42..88a63af 100644
--- a/src/timesfm/configs.py
+++ b/src/timesfm/configs.py
@@ -18,7 +18,7 @@ import dataclasses
from typing import Literal
[email protected](frozen=False)
[email protected](frozen=True)
class ForecastConfig:
"""Options for forecasting.
diff --git a/src/timesfm/flax/dense.py b/src/timesfm/flax/dense.py
index 7728d36..7de49b9 100644
--- a/src/timesfm/flax/dense.py
+++ b/src/timesfm/flax/dense.py
@@ -18,7 +18,6 @@ from flax import nnx
import jax
import jax.numpy as jnp
import jaxtyping
-import typeguard
from .. import configs
@@ -64,7 +63,6 @@ class ResidualBlock(nnx.Module):
else:
raise ValueError(f"Activation: {config.activation} not supported.")
- @jaxtyping.jaxtyped(typechecker=typeguard.typechecked)
def __call__(self, x: Float[Array, "b ... i"]) -> Float[Array, "b ... o"]:
return self.output_layer(
self.activation(self.hidden_layer(x))
@@ -99,7 +97,6 @@ class RandomFourierFeatures(nnx.Module):
rngs=rngs,
)
- @jaxtyping.jaxtyped(typechecker=typeguard.typechecked)
def __call__(self, x: Float[Array, "b ... i"]) -> Float[Array, "b ... o"]:
projected = self.projection_layer(x)
cos_features = jnp.cos(projected)
diff --git a/src/timesfm/flax/normalization.py b/src/timesfm/flax/normalization.py
index 6f6e0be..07349c8 100644
--- a/src/timesfm/flax/normalization.py
+++ b/src/timesfm/flax/normalization.py
@@ -18,7 +18,6 @@ from flax import nnx
import jax
import jax.numpy as jnp
import jaxtyping
-import typeguard
Array = jaxtyping.Array
Bool = jaxtyping.Bool
@@ -44,7 +43,6 @@ class RMSNorm(nnx.Module):
self.num_features = num_features
self.epsilon = epsilon
- @jaxtyping.jaxtyped(typechecker=typeguard.typechecked)
def __call__(
self, inputs: Float[Array, "b ... d"]
) -> Float[Array, "b ... d"]:
@@ -69,7 +67,6 @@ class LayerNorm(nnx.Module):
self.num_features = num_features
self.epsilon = epsilon
- @jaxtyping.jaxtyped(typechecker=typeguard.typechecked)
def __call__(
self, inputs: Float[Array, "b ... d"]
) -> Float[Array, "b ... d"]:
diff --git a/src/timesfm/flax/transformer.py b/src/timesfm/flax/transformer.py
index fb56d4b..add1d85 100644
--- a/src/timesfm/flax/transformer.py
+++ b/src/timesfm/flax/transformer.py
@@ -23,7 +23,6 @@ import jax
from jax import lax
import jax.numpy as jnp
import jaxtyping
-import typeguard
from .. import configs
from . import normalization, util
@@ -127,7 +126,6 @@ class PerDimScale(nnx.Module):
self.num_dims = num_dims
self.per_dim_scale = nnx.Param(jnp.zeros(shape=(num_dims,)))
- @jaxtyping.jaxtyped(typechecker=typeguard.typechecked)
def __call__(self, x: Float[Array, "b ... d"]) -> Float[Array, "b ... d"]:
return x * (
1.442695041
@@ -344,7 +342,6 @@ class Transformer(nnx.Module):
else:
raise ValueError(f"Activation: {config.ff_activation} not supported.")
- @jaxtyping.jaxtyped(typechecker=typeguard.typechecked)
def __call__(
self,
input_embeddings: Float[Array, "b n d"],
diff --git a/src/timesfm/flax/util.py b/src/timesfm/flax/util.py
index d98ef5c..de4d3c7 100644
--- a/src/timesfm/flax/util.py
+++ b/src/timesfm/flax/util.py
@@ -40,7 +40,6 @@ class DecodeCache:
value: Float[Array, "b n h d"]
[email protected](typechecker=typeguard.typechecked)
@jax.jit
def update_running_stats(
n: Float[Array, "b"],
diff --git a/src/timesfm/timesfm_2p5/timesfm_2p5_flax.py b/src/timesfm/timesfm_2p5/timesfm_2p5_flax.py
index 87c09d9..48a61ad 100644
--- a/src/timesfm/timesfm_2p5/timesfm_2p5_flax.py
+++ b/src/timesfm/timesfm_2p5/timesfm_2p5_flax.py
@@ -46,6 +46,8 @@ Array = jaxtyping.Array
def try_gc():
for d in jax.local_devices():
stats = d.memory_stats()
+ if stats is None:
+ return
if stats["bytes_in_use"] / stats["bytes_limit"] > 0.75:
gc.collect()
break
@@ -458,9 +460,10 @@ class TimesFM_2p5_200M_flax(timesfm_2p5_base.TimesFM_2p5):
"""Flax implementation of TimesFM 2.5 with 200M parameters."""
model: nnx.Module = TimesFM_2p5_200M_flax_module()
-
+
+ @classmethod
def from_pretrained(
- self,
+ cls,
*,
path: str | None = None,
hf_repo_id: str | None = "google/timesfm-2.5-200m-flax",
@@ -476,11 +479,13 @@ class TimesFM_2p5_200M_flax(timesfm_2p5_base.TimesFM_2p5):
logging.info("Loading checkpoint from: %s", path)
else:
raise ValueError("Either path or hf_repo_id must be provided.")
-
+
+ instance = cls()
checkpointer = ocp.StandardCheckpointer()
- graph, state = nnx.split(self.model)
+ graph, state = nnx.split(instance.model)
state = checkpointer.restore(path, state)
- self.model = nnx.merge(graph, state)
+ instance.model = nnx.merge(graph, state)
+ return instance
def compile(self, forecast_config: configs.ForecastConfig, **kwargs):
diff --git a/src/timesfm/timesfm_2p5/timesfm_2p5_torch.py b/src/timesfm/timesfm_2p5/timesfm_2p5_torch.py
index 897dbe7..260cbd3 100644
--- a/src/timesfm/timesfm_2p5/timesfm_2p5_torch.py
+++ b/src/timesfm/timesfm_2p5/timesfm_2p5_torch.py
@@ -13,6 +13,7 @@
# limitations under the License.
"""TimesFM models."""
+import dataclasses
import logging
import math
import os
@@ -349,7 +350,7 @@ class TimesFM_2p5_200M_torch(timesfm_2p5_base.TimesFM_2p5, ModelHubMixin):
self.model.p,
new_context := math.ceil(fc.max_context / self.model.p) * self.model.p,
)
- fc.max_context = new_context
+ fc = dataclasses.replace(fc, max_context=new_context)
if fc.max_horizon % self.model.o != 0:
logging.info(
"When compiling, max horizon needs to be multiple of the output patch"
@@ -357,7 +358,7 @@ class TimesFM_2p5_200M_torch(timesfm_2p5_base.TimesFM_2p5, ModelHubMixin):
self.model.o,
new_horizon := math.ceil(fc.max_horizon / self.model.o) * self.model.o,
)
- fc.max_horizon = new_horizon
+ forecast_config = dataclasses.replace(fc, max_horizon=new_horizon)
if fc.max_context + fc.max_horizon > self.model.config.context_limit:
raise ValueError(
"Context + horizon must be less than the context limit."