add constrained SVD model

exploratory.ipynb

@@ -455,7 +455,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.6.9"
+   "version": "3.8.5"
   }
  },
  "nbformat": 4,

fnn/models.py

@@ -510,6 +510,20 @@ def transform(self, X, y=None):
 #             **kwargs
 #         )
 
+class ETDConstrainedEmbedding(NeuralNetworkEmbedding):
+    def __init__(
+        self,
+        *args,
+        **kwargs
+    ):
+        super().__init__(*args, **kwargs)
+        kwargs.pop("time_window")
+        self.model = ETDConstrained(
+            self.n_latent,
+            self.time_window,
+            **kwargs
+        )
+
 class MLPEmbedding(NeuralNetworkEmbedding):
     def __init__(
         self,

fnn/networks.py

@@ -547,4 +547,47 @@ def __init__(
     def call(self, inputs, training=False):
         outputs = self.decoder(self.encoder(inputs))
         return outputs
-
+
+
+class ETDConstrained(tf.keras.Model):
+    """
+    A batchwise fully-connected eigen-time-delay embedding (linear autoencoder)
+    """
+    def __init__(
+        self,
+        n_latent,
+        time_window,
+        n_features=1,
+        latent_regularizer=None,
+        rnn_opts=dict(),
+        random_state=None,
+        **kwargs
+    ):
+        super(ETDConstrained, self).__init__()
+        self.n_latent = n_latent
+        self.time_window = time_window
+        self.n_features = n_features
+
+        # Initialize state
+        tf.random.set_seed(random_state)
+
+        # Encoder
+        self.encoder = tf.keras.Sequential()
+        self.encoder.add(tf.keras.layers.InputLayer(input_shape=(time_window, n_features)))
+        self.encoder.add(tf.keras.layers.Flatten())
+        self.encoder.add(tf.keras.layers.Dense(n_latent, activation=None, **rnn_opts))
+        self.encoder.add(
+            tf.keras.layers.Reshape(
+                (n_latent,),  
+                activity_regularizer=latent_regularizer
+            )
+        )
+
+        self.decoder = tf.keras.Sequential()
+        self.decoder.add(tf.keras.layers.Flatten())
+        self.decoder.add(tf.keras.layers.Dense(time_window*n_features, activation=None, **rnn_opts))
+        self.decoder.add(tf.keras.layers.Reshape((time_window, n_features)))
+
+    def call(self, inputs, training=False):
+        outputs = self.decoder(self.encoder(inputs))
+        return outputs

fnn/utils.py

@@ -1,12 +1,7 @@
 import numpy as np
 import warnings
 
-import scipy
-from scipy.linalg import hankel, orthogonal_procrustes
-from scipy.signal import periodogram, argrelextrema, savgol_filter
-from scipy.spatial.distance import pdist, squareform, directed_hausdorff
-from scipy.spatial import procrustes
-
+from scipy.linalg import hankel
 import matplotlib.pyplot as plt
 
 ###------------------------------------###