imt-ring 1.3.6__tar.gz → 1.3.8__tar.gz

{imt_ring-1.3.6 → imt_ring-1.3.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: imt-ring
-Version: 1.3.6
+Version: 1.3.8
 Summary: RING: Recurrent Inertial Graph-based Estimator
 Author-email: Simon Bachhuber <simon.bachhuber@fau.de>
 Project-URL: Homepage, https://github.com/SimiPixel/ring

{imt_ring-1.3.6 → imt_ring-1.3.8}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "imt-ring"
-version = "1.3.6"
+version = "1.3.8"
 authors = [
   { name="Simon Bachhuber", email="simon.bachhuber@fau.de" },
 ]

{imt_ring-1.3.6 → imt_ring-1.3.8}/src/imt_ring.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: imt-ring
-Version: 1.3.6
+Version: 1.3.8
 Summary: RING: Recurrent Inertial Graph-based Estimator
 Author-email: Simon Bachhuber <simon.bachhuber@fau.de>
 Project-URL: Homepage, https://github.com/SimiPixel/ring

{imt_ring-1.3.6 → imt_ring-1.3.8}/src/imt_ring.egg-info/SOURCES.txt

@@ -64,6 +64,7 @@ src/ring/ml/callbacks.py
 src/ring/ml/ml_utils.py
 src/ring/ml/optimizer.py
 src/ring/ml/ringnet.py
+src/ring/ml/rnno_v1.py
 src/ring/ml/train.py
 src/ring/ml/training_loop.py
 src/ring/ml/params/0x13e3518065c21cd8.pickle

{imt_ring-1.3.6 → imt_ring-1.3.8}/src/ring/algorithms/generator/batch.py

@@ -147,7 +147,7 @@ def _data_fn_from_paths(
     paths = [utils.parse_path(p, mkdir=False) for p in paths]
 
     extensions = list(set([Path(p).suffix for p in paths]))
-    assert len(extensions) == 1
+    assert len(extensions) == 1, f"{extensions}"
 
     if extensions[0] == ".h5":
         N = sum([utils.hdf5_load_length(p) for p in paths])

{imt_ring-1.3.6 → imt_ring-1.3.8}/src/ring/algorithms/generator/motion_artifacts.py

@@ -180,9 +180,13 @@ def setup_fn_randomize_damping_stiffness_factory(
             link_spring_stiffness = link_spring_stiffness.at[slice].set(stif)
             link_damping = link_damping.at[slice].set(damp)
 
-        assert len(imus_surely_rigid) == len(triggered_surely_rigid)
+        assert len(imus_surely_rigid) == len(
+            triggered_surely_rigid
+        ), f"{imus_surely_rigid}, {triggered_surely_rigid}"
         for imu_surely_rigid in imus_surely_rigid:
-            assert imu_surely_rigid in triggered_surely_rigid
+            assert (
+                imu_surely_rigid in triggered_surely_rigid
+            ), f"{imus_surely_rigid} not in {triggered_surely_rigid}"
 
         return sys.replace(
             link_damping=link_damping, link_spring_stiffness=link_spring_stiffness

{imt_ring-1.3.6 → imt_ring-1.3.8}/src/ring/ml/__init__.py

@@ -3,6 +3,7 @@ from . import callbacks
 from . import ml_utils
 from . import optimizer
 from . import ringnet
+from . import rnno_v1
 from . import train
 from . import training_loop
 from .base import AbstractFilter
@@ -42,17 +43,28 @@ def RNNO(
     params=None,
     eval: bool = True,
     samp_freq: float | None = None,
+    v1: bool = False,
     **kwargs,
 ):
     assert "message_dim" not in kwargs
     assert "link_output_normalize" not in kwargs
     assert "link_output_dim" not in kwargs
 
+    if v1:
+        kwargs.update(
+            dict(forward_factory=rnno_v1.rnno_v1_forward_factory, output_dim=output_dim)
+        )
+    else:
+        kwargs.update(
+            dict(
+                message_dim=0,
+                link_output_normalize=False,
+                link_output_dim=output_dim,
+            )
+        )
+
     ringnet = RING(  # noqa: F811
         params=params,
-        message_dim=0,
-        link_output_normalize=False,
-        link_output_dim=output_dim,
         **kwargs,
     )
     ringnet = base.NoGraph_FilterWrapper(ringnet, quat_normalize=return_quats)

{imt_ring-1.3.6 → imt_ring-1.3.8}/src/ring/ml/callbacks.py

@@ -245,7 +245,8 @@ class SaveParamsTrainingLoopCallback(training_loop.TrainingLoopCallback):
             else:
                 value = "{:.2f}".format(ele.value).replace(".", ",")
                 filename = parse_path(
-                    self.path_to_file + f"_episode={ele.episode}_value={value}",
+                    str(Path(self.path_to_file).with_suffix(""))
+                    + f"_episode={ele.episode}_value={value}",
                     extension="pickle",
                 )
 
@@ -404,7 +405,7 @@ class CheckpointCallback(training_loop.TrainingLoopCallback):
         # only checkpoint if run has been killed
         if training_loop.recv_kill_run_signal():
             path = parse_path(
-                "~/.xxy_checkpoints", ml_utils.unique_id(), extension="pickle"
+                "~/.ring_checkpoints", ml_utils.unique_id(), extension="pickle"
             )
             data = {"params": self.params, "opt_state": self.opt_state}
             pickle_save(

{imt_ring-1.3.6 → imt_ring-1.3.8}/src/ring/ml/ringnet.py

@@ -191,8 +191,16 @@ class LSTM(hk.RNNCore):
 
 
 class RING(ml_base.AbstractFilter):
-    def __init__(self, params=None, lam=None, jit: bool = True, name=None, **kwargs):
-        self.forward_lam_factory = partial(make_ring, **kwargs)
+    def __init__(
+        self,
+        params=None,
+        lam=None,
+        jit: bool = True,
+        name=None,
+        forward_factory=make_ring,
+        **kwargs,
+    ):
+        self.forward_lam_factory = partial(forward_factory, **kwargs)
         self.params = self._load_params(params)
         self.lam = lam
         self._name = name

imt_ring-1.3.8/src/ring/ml/rnno_v1.py

@@ -0,0 +1,37 @@
+from typing import Sequence
+
+import haiku as hk
+import jax
+import jax.numpy as jnp
+
+
+def rnno_v1_forward_factory(
+    output_dim: int,
+    rnn_layers: Sequence[int] = (400, 300),
+    linear_layers: Sequence[int] = (200, 100, 50, 50, 25, 25),
+    layernorm: bool = True,
+    act_fn_linear=jax.nn.relu,
+    act_fn_rnn=jax.nn.elu,
+):
+    @hk.without_apply_rng
+    @hk.transform_with_state
+    def forward_fn(X):
+        assert X.shape[-2] == 1
+
+        for i, n_units in enumerate(rnn_layers):
+            state = hk.get_state(f"rnn_{i}", shape=[n_units], init=jnp.zeros)
+            X, state = hk.dynamic_unroll(hk.GRU(n_units), X, state)
+            hk.set_state(f"rnn_{i}", state)
+
+            if layernorm:
+                X = hk.LayerNorm(axis=-1, create_scale=False, create_offset=False)(X)
+            X = act_fn_rnn(X)
+
+        for n_units in linear_layers:
+            X = hk.Linear(n_units)(X)
+            X = act_fn_linear(X)
+
+        y = hk.Linear(output_dim)(X)
+        return y[..., None, :]
+
+    return forward_fn

{imt_ring-1.3.6 → imt_ring-1.3.8}/src/ring/ml/train.py

@@ -5,6 +5,8 @@ from typing import Callable, Optional, Tuple
 import jax
 import jax.numpy as jnp
 import optax
+import tree_utils
+
 from ring import maths
 from ring.algorithms.generator import types
 from ring.ml import base as ml_base
@@ -13,10 +15,7 @@ from ring.ml import ml_utils
 from ring.ml import training_loop
 from ring.utils import distribute_batchsize
 from ring.utils import expand_batchsize
-from ring.utils import parse_path
 from ring.utils import pickle_load
-import tree_utils
-
 import wandb
 
 # (T, N, F) -> Scalar
@@ -142,15 +141,17 @@ def train_fn(
         Wether or not the training run was killed by a callback.
     """
 
+    filter = filter.nojit()
+
     if checkpoint is not None:
         checkpoint = Path(checkpoint).with_suffix(".pickle")
         recv_checkpoint: dict = pickle_load(checkpoint)
-        filter.params = recv_checkpoint["params"]
+        filter_params = recv_checkpoint["params"]
         opt_state = recv_checkpoint["opt_state"]
+        del recv_checkpoint
+    else:
+        filter_params = filter.search_attr("params")
 
-    filter = filter.nojit()
-
-    filter_params = filter.search_attr("params")
     if filter_params is None:
         X, _ = generator(jax.random.PRNGKey(1))
         filter_params, _ = filter.init(X=X, seed=seed_network)
@@ -215,7 +216,7 @@ def train_fn(
 
         callbacks_all.append(
             ml_callbacks.SaveParamsTrainingLoopCallback(
-                path_to_file=parse_path(callback_save_params, extension=""),
+                path_to_file=callback_save_params,
                 last_n_params=3,
                 track_metrices=callback_save_params_track_metrices,
                 cleanup=False,