Irregularly sample

.pre-commit-config.yaml

@@ -3,23 +3,30 @@ repos:
     rev: 24.3.0
     hooks:
       - id: black
+        exclude: '\.ipynb$'   # Ignore notebooks
 
   - repo: https://github.com/nbQA-dev/nbQA
     rev: 1.7.0
     hooks:
       - id: nbqa-black
         additional_dependencies: [black, setuptools]
+        exclude: '\.ipynb$'
       - id: nbqa-isort
         additional_dependencies: [isort, setuptools]
+        exclude: '\.ipynb$'
       - id: nbqa-flake8
         additional_dependencies: [flake8, setuptools]
+        exclude: '\.ipynb$'
 
   - repo: https://github.com/PyCQA/isort
     rev: 5.13.2
     hooks:
       - id: isort
+        exclude: '\.ipynb$'
 
   - repo: https://github.com/pycqa/flake8
     rev: 6.1.0
     hooks:
-      - id: flake8
+      - id: flake8
+        exclude: '\.ipynb$'
+

data_dir/datasets.py

@@ -140,60 +140,61 @@ def dataset_generator(
         val_data, val_labels = data[idxs[1]], labels[idxs[1]]
         test_data, test_labels = None, None
 
+    if include_time:
+        ts_train = train_data[:, :, 0]
+        ts_val = val_data[:, :, 0]
+        ts_test = test_data[:, :, 0]
+    else:
+        ts_train = (T / train_data.shape[1]) * jnp.repeat(
+            jnp.arange(train_data.shape[1])[None, :], train_data.shape[0], axis=0
+        )
+        ts_val = (T / val_data.shape[1]) * jnp.repeat(
+            jnp.arange(val_data.shape[1])[None, :], val_data.shape[0], axis=0
+        )
+        ts_test = (T / test_data.shape[1]) * jnp.repeat(
+            jnp.arange(test_data.shape[1])[None, :], test_data.shape[0], axis=0
+        )
+
     train_paths = batch_calc_paths(train_data, stepsize, depth)
     val_paths = batch_calc_paths(val_data, stepsize, depth)
     test_paths = batch_calc_paths(test_data, stepsize, depth)
-    intervals = jnp.arange(0, train_data.shape[1], stepsize)
-    intervals = jnp.concatenate((intervals, jnp.array([train_data.shape[1]])))
-    intervals = intervals * (T / train_data.shape[1])
+    indexes = np.unique(np.r_[0 : train_data.shape[1] : stepsize])
+    intervals = ts_train[0, indexes]
+    intervals = jnp.concatenate((intervals, jnp.array([T])))
 
     train_coeffs = calc_coeffs(train_data, include_time, T)
     val_coeffs = calc_coeffs(val_data, include_time, T)
     test_coeffs = calc_coeffs(test_data, include_time, T)
     train_coeff_data = (
-        (T / train_data.shape[1])
-        * jnp.repeat(
-            jnp.arange(train_data.shape[1])[None, :], train_data.shape[0], axis=0
-        ),
+        ts_train,
         train_coeffs,
         train_data[:, 0, :],
     )
     val_coeff_data = (
-        (T / val_data.shape[1])
-        * jnp.repeat(jnp.arange(val_data.shape[1])[None, :], val_data.shape[0], axis=0),
+        ts_val,
         val_coeffs,
         val_data[:, 0, :],
     )
     if idxs is None:
         test_coeff_data = (
-            (T / test_data.shape[1])
-            * jnp.repeat(
-                jnp.arange(test_data.shape[1])[None, :], test_data.shape[0], axis=0
-            ),
+            ts_test,
             test_coeffs,
             test_data[:, 0, :],
         )
 
     train_path_data = (
-        (T / train_data.shape[1])
-        * jnp.repeat(
-            jnp.arange(train_data.shape[1])[None, :], train_data.shape[0], axis=0
-        ),
+        ts_train,
         train_paths,
         train_data[:, 0, :],
     )
     val_path_data = (
-        (T / val_data.shape[1])
-        * jnp.repeat(jnp.arange(val_data.shape[1])[None, :], val_data.shape[0], axis=0),
+        ts_val,
         val_paths,
         val_data[:, 0, :],
     )
     if idxs is None:
         test_path_data = (
-            (T / test_data.shape[1])
-            * jnp.repeat(
-                jnp.arange(test_data.shape[1])[None, :], test_data.shape[0], axis=0
-            ),
+            ts_test,
             test_paths,
             test_data[:, 0, :],
         )
@@ -234,7 +235,7 @@ def dataset_generator(
 
 
 def _scale_to_minus_one_one(x, data_min, data_max, eps=1e-8):
-    """Affine‑maps x from [data_min,data_max] → [‑1,1] with broadcasting."""
+    """Affine-maps x from [data_min,data_max] → [-1,1] with broadcasting."""
     return 2.0 * (x - data_min) / (data_max - data_min + eps) - 1.0
 
 
@@ -266,18 +267,13 @@ def create_uea_dataset(
             test_data = pickle.load(f)
         with open(data_dir + f"/processed/UEA/{name}/y_test.pkl", "rb") as f:
             test_labels = pickle.load(f)
+        t = (T / train_data.shape[1]) * jnp.arange(train_data.shape[1])[None, :]
         if include_time:
-            ts = (T / train_data.shape[1]) * jnp.repeat(
-                jnp.arange(train_data.shape[1])[None, :], train_data.shape[0], axis=0
-            )
+            ts = jnp.repeat(t, train_data.shape[0], axis=0)
             train_data = jnp.concatenate([ts[:, :, None], train_data], axis=2)
-            ts = (T / val_data.shape[1]) * jnp.repeat(
-                jnp.arange(val_data.shape[1])[None, :], val_data.shape[0], axis=0
-            )
+            ts = jnp.repeat(t, val_data.shape[0], axis=0)
             val_data = jnp.concatenate([ts[:, :, None], val_data], axis=2)
-            ts = (T / test_data.shape[1]) * jnp.repeat(
-                jnp.arange(test_data.shape[1])[None, :], test_data.shape[0], axis=0
-            )
+            ts = jnp.repeat(t, test_data.shape[0], axis=0)
             test_data = jnp.concatenate([ts[:, :, None], test_data], axis=2)
         data = (train_data, val_data, test_data)
         onehot_labels = (train_labels, val_labels, test_labels)
@@ -286,6 +282,7 @@ def create_uea_dataset(
             data = pickle.load(f)
         with open(data_dir + f"/processed/UEA/{name}/labels.pkl", "rb") as f:
             labels = pickle.load(f)
+        t = (T / data.shape[1]) * jnp.arange(data.shape[1])[None, :]
         onehot_labels = jnp.zeros((len(labels), len(jnp.unique(labels))))
         onehot_labels = onehot_labels.at[jnp.arange(len(labels)), labels].set(1)
         if use_idxs:
@@ -295,9 +292,7 @@ def create_uea_dataset(
             idxs = None
 
         if include_time:
-            ts = (T / data.shape[1]) * jnp.repeat(
-                jnp.arange(data.shape[1])[None, :], data.shape[0], axis=0
-            )
+            ts = jnp.repeat(t, data.shape[0], axis=0)
             data = jnp.concatenate([ts[:, :, None], data], axis=2)
 
     if scale:

experiment_configs/repeats/bd_linear_ncde/EigenWorms.json

@@ -26,5 +26,6 @@
     "lambd": 0.001,
     "block_size": 4,
     "stepsize": 12,
-    "depth": 2
+    "depth": 2,
+    "parallel_steps": 128
 }

experiment_configs/repeats/bd_linear_ncde/EthanolConcentration.json

@@ -26,5 +26,6 @@
     "block_size": 4,
     "depth": 1,
     "stepsize": 1,
-    "lambd": 0.000001
+    "lambd": 0.000001,
+    "parallel_steps": 128
 }

experiment_configs/repeats/bd_linear_ncde/Heartbeat.json

@@ -26,5 +26,6 @@
     "block_size": 4,
     "depth": 2,
     "stepsize": 2,
-    "lambd": 0.000001
+    "lambd": 0.000001,
+    "parallel_steps": 128
 }

experiment_configs/repeats/bd_linear_ncde/MotorImagery.json

@@ -26,5 +26,6 @@
     "block_size": 4,
     "depth": 2,
     "stepsize": 16,
-    "lambd": 0.001
+    "lambd": 0.001,
+    "parallel_steps": 128
 }

experiment_configs/repeats/bd_linear_ncde/SelfRegulationSCP1.json

@@ -26,5 +26,6 @@
     "block_size": 4,
     "stepsize": 16,
     "depth": 2,
-    "lambd": 0.0
+    "lambd": 0.0,
+    "parallel_steps": 128
 }

experiment_configs/repeats/bd_linear_ncde/SelfRegulationSCP2.json

@@ -26,5 +26,6 @@
     "block_size": 4,
     "stepsize": 4,
     "depth": 2,
-    "lambd": 0.001
+    "lambd": 0.001,
+    "parallel_steps": 128
 }

experiment_configs/repeats/dense_linear_ncde/EigenWorms.json

@@ -12,7 +12,7 @@
     "num_steps": 100000,
     "print_steps": 1000,
     "early_stopping_steps": 10,
-    "batch_size": 16,
+    "batch_size": 32,
     "model_name": "linear_ncde",
     "metric": "accuracy",
     "classification": true,
@@ -26,5 +26,6 @@
     "lambd": 0.001,
     "block_size": 128,
     "stepsize": 12,
-    "depth": 2
+    "depth": 2,
+    "parallel_steps": 128
 }

experiment_configs/repeats/dense_linear_ncde/EthanolConcentration.json

@@ -26,5 +26,6 @@
     "block_size": 64,
     "depth": 1,
     "stepsize": 1,
-    "lambd": 0.000001
+    "lambd": 0.000001,
+    "parallel_steps": 128
 }

experiment_configs/repeats/dense_linear_ncde/Heartbeat.json

@@ -26,5 +26,6 @@
     "block_size": 16,
     "depth": 2,
     "stepsize": 2,
-    "lambd": 0.000001
+    "lambd": 0.000001,
+    "parallel_steps": 128
 }

experiment_configs/repeats/dense_linear_ncde/MotorImagery.json

@@ -26,5 +26,6 @@
     "block_size": 16,
     "depth": 2,
     "stepsize": 16,
-    "lambd": 0.001
+    "lambd": 0.001,
+    "parallel_steps": 128
 }

experiment_configs/repeats/dense_linear_ncde/SelfRegulationSCP1.json

@@ -26,5 +26,6 @@
     "block_size": 64,
     "stepsize": 16,
     "depth": 2,
-    "lambd": 0.0
+    "lambd": 0.0,
+    "parallel_steps": 128
 }

experiment_configs/repeats/dense_linear_ncde/SelfRegulationSCP2.json

@@ -26,5 +26,6 @@
     "block_size": 128,
     "stepsize": 4,
     "depth": 2,
-    "lambd": 0.001
+    "lambd": 0.001,
+    "parallel_steps": 128
 }

experiment_configs/repeats/diagonal_dense_linear_ncde/EigenWorms.json

@@ -0,0 +1,35 @@
+{
+    "seeds": [
+        2345,
+        3456,
+        4567,
+        5678,
+        6789
+    ],
+    "data_dir": "data_dir",
+    "output_parent_dir": "",
+    "lr_scheduler": "lambda lr: lr",
+    "num_steps": 100000,
+    "print_steps": 1000,
+    "early_stopping_steps": 10,
+    "batch_size": 32,
+    "model_name": "linear_ncde",
+    "metric": "accuracy",
+    "classification": true,
+    "dataset_name": "EigenWorms",
+    "use_presplit": false,
+    "T": 1,
+    "scale": 1,
+    "time": "True",
+    "lr": "0.001",
+    "hidden_dim": "128",
+    "lambd": 0.001,
+    "block_size": 16,
+    "stepsize": 12,
+    "depth": 2,
+    "parallel_steps": 128,
+    "walsh_hadamard": false,
+    "diagonal_dense": true,
+    "sparsity": 1.0,
+    "rank": 0
+}

experiment_configs/repeats/diagonal_dense_linear_ncde/EthanolConcentration.json

@@ -0,0 +1,35 @@
+{
+    "seeds": [
+        2345,
+        3456,
+        4567,
+        5678,
+        6789
+    ],
+    "data_dir": "data_dir",
+    "output_parent_dir": "",
+    "lr_scheduler": "lambda lr: lr",
+    "num_steps": 100000,
+    "print_steps": 1000,
+    "early_stopping_steps": 10,
+    "batch_size": 32,
+    "model_name": "linear_ncde",
+    "metric": "accuracy",
+    "classification": true,
+    "dataset_name": "EthanolConcentration",
+    "use_presplit": false,
+    "T": 1,
+    "scale": 1,
+    "time": "True",
+    "lr": "0.0001",
+    "hidden_dim": "64",
+    "block_size": 16,
+    "depth": 1,
+    "stepsize": 1,
+    "lambd": 0.000001,
+    "parallel_steps": 128,
+    "walsh_hadamard": false,
+    "diagonal_dense": true,
+    "sparsity": 1.0,
+    "rank": 0
+}

experiment_configs/repeats/diagonal_dense_linear_ncde/Heartbeat.json

@@ -0,0 +1,35 @@
+{
+    "seeds": [
+        2345,
+        3456,
+        4567,
+        5678,
+        6789
+    ],
+    "data_dir": "data_dir",
+    "output_parent_dir": "",
+    "lr_scheduler": "lambda lr: lr",
+    "num_steps": 100000,
+    "print_steps": 1000,
+    "early_stopping_steps": 10,
+    "batch_size": 32,
+    "model_name": "linear_ncde",
+    "metric": "accuracy",
+    "classification": true,
+    "dataset_name": "Heartbeat",
+    "use_presplit": false,
+    "T": 1,
+    "scale": 1,
+    "time": "True",
+    "lr": "0.001",
+    "hidden_dim": "16",
+    "block_size": 16,
+    "depth": 2,
+    "stepsize": 2,
+    "lambd": 0.000001,
+    "parallel_steps": 128,
+    "walsh_hadamard": false,
+    "diagonal_dense": false,
+    "sparsity": 1.0,
+    "rank": 0
+}