From 3a0d97879bd3fd5d04367e383336b09b5dcf79b5 Mon Sep 17 00:00:00 2001
From: Jan <jan.mandel@gmail.com>
Date: Sat, 2 Jul 2022 15:50:28 -0600
Subject: [PATCH] adding variabke number of dense layers and the simplest
 example consisting of one cell only

---
 fmda_kf_rnn.ipynb | 34 +++++++++++++++++++++++++++++-----
 1 file changed, 29 insertions(+), 5 deletions(-)

diff --git a/fmda_kf_rnn.ipynb b/fmda_kf_rnn.ipynb
index 679665b..c418b5d 100644
--- a/fmda_kf_rnn.ipynb
+++ b/fmda_kf_rnn.ipynb
@@ -1413,7 +1413,8 @@
       },
       "outputs": [],
       "source": [
-        "def create_RNN_2(hidden_units, dense_units, activation, stateful=False, batch_shape=None, input_shape=None):\n",
+        "def create_RNN_2(hidden_units, dense_units, activation, stateful=False, \n",
+        "                 batch_shape=None, input_shape=None, dense_layers=1):\n",
         "    if stateful:\n",
         "      inputs = tf.keras.Input(batch_shape=batch_shape)\n",
         "    else:\n",
@@ -1423,27 +1424,50 @@
         "    x = inputs\n",
         "    x = tf.keras.layers.SimpleRNN(hidden_units,activation=activation[0],stateful=stateful)(x)\n",
         "    # x = tf.keras.layers.Dense(hidden_units, activation=activation[1])(x)\n",
-        "    x = tf.keras.layers.Dense(dense_units, activation=activation[1])(x)\n",
+        "    for i in range(dense_layers):\n",
+        "      x = tf.keras.layers.Dense(dense_units, activation=activation[1])(x)\n",
         "    model = tf.keras.Model(inputs=inputs, outputs=x)\n",
         "    model.compile(loss='mean_squared_error', optimizer='adam')\n",
         "    return model\n",
         "def create_fit_predict_RNN(hidden_units, dense_units, \n",
-        "                             samples, timesteps, features, activation):\n",
+        "                             samples, timesteps, features, dense_layers=1, activation=['tanh', 'tanh']):\n",
         "    # statefull model version with with fixed number of batches\n",
         "    model_fit=create_RNN_2(hidden_units=hidden_units, dense_units=dense_units, \n",
         "                        batch_shape=(samples, timesteps, features),stateful = True,\n",
-        "                        activation=activation)\n",
+        "                        activation=activation,dense_layers=dense_layers)\n",
         "    print(model_fit.summary())\n",
         "    # same model for prediction on the entire dataset\n",
         "    model_predict=create_RNN_2(hidden_units=hidden_units, dense_units=dense_units,  \n",
         "                        input_shape=(None,features),stateful = False,\n",
-        "                        activation=activation)\n",
+        "                        activation=activation,dense_layers=dense_layers)\n",
         "    print(model_predict.summary())\n",
         "    return model_fit, model_predict"
       ]
     },
     {
       "cell_type": "code",
+      "source": [
+        "# the simplest model possible\n",
+        "fmda_model, fmda_model_eval = create_fit_predict_RNN(hidden_units=1, dense_units=1, \n",
+        "      samples=samples, timesteps=timesteps, features=1, dense_layers=0,\n",
+        "      activation=['linear'])\n",
+        "fmda_model.fit(x_train, y_train, epochs=40, verbose=2,batch_size=samples)\n",
+        "# Same model as stateless for prediction:\n",
+        "w=fmda_model.get_weights()\n",
+        "fmda_model_eval.set_weights(w)\n",
+        "# prediction on the entire dataset from zero state\n",
+        "mt = fmda_model_eval.predict(Et)\n",
+        "m = scalery.inverse_transform(mt)\n",
+        "plot_m(m,title='RNN prediction')"
+      ],
+      "metadata": {
+        "id": "PZw7DNQD4Inr"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
       "execution_count": null,
       "metadata": {
         "id": "R2jkoZlAIaSb"
-- 
2.11.4.GIT