initial commit

nuclearspy · Aug 29, 2017 · 2c4be38 · 2c4be38
1 parent a8a1d32
commit 2c4be38
Show file tree

Hide file tree

Showing 5 changed files with 937 additions and 10 deletions.
diff --git a/README.md b/README.md
@@ -0,0 +1,76 @@
+# Time Series Predictions
+
+> Play with time
+
+## 1. Shampoo Sales Prediction
+
+> `ShampooSales.ipynb`
+
+sales goes like this, need to predict according to history.
+
+<img src='./assets/sales.png'>
+
+A wonderful tutorial to convert time series prediction to supervised problem: [Time Series Forecasting as Supervised Learning](https://machinelearningmastery.com/time-series-forecasting-supervised-learning/)
+
+### Result
+
+best fit before overfitting:
+
+<img src='./assets/train.png'>
+
+### Stateful LSTM
+
+Core code
+
+```
+model = Sequential()
+model.add(LSTM(4, batch_input_shape=(BATCH_SIZE, X.shape[1], X.shape[2]), stateful=True))
+model.add(Dropout(0.5))
+model.add(Dense(1, activation='linear'))
+
+model.compile(loss='mse', optimizer='adadelta')
+
+# way 1
+for i in range(EPOCHS):
+    model.fit(X, y, epochs=1, shuffle=False, batch_size=BATCH_SIZE)
+    model.reset_states()
+
+# way 2
+class StatusResetCallback(Callback):
+    def on_batch_begin(self, batch, logs={}):
+        self.model.reset_states()
+
+model.fit(X, y, epochs=EPOCHS, batch_size=BATCH_SIZE,
+		shuffle=False, callbacks=[StatusResetCallback()])
+```
+
+## 2. Stateful LSTM in Keras
+
+> `StatefulLSTM.ipynb`
+
+Learning from [Stateful LSTM in Keras](http://philipperemy.github.io/keras-stateful-lstm/) by Philippe Remy, which is a wonderful and simple tutorial. The composed dataset is simple and clean:
+
+```
+     X           y
+1 0 0 ... 0      1
+0 0 0 ... 0      0
+0 0 0 ... 0      0
+1 0 0 ... 0      1
+1 0 0 ... 0      1
+...
+```
+
+Obviously, if the first of X seq is 1, y = 1, else 0. We will see if the 1 status will pass along to predict the result.
+
+### Stateless LSTM Can't Converge
+
+```
+model = Sequential()
+model.add(LSTM(LSTM_UNITS, input_shape=X_train.shape[1:], return_sequences=False, stateful=False))
+model.add(Dense(1, activation='sigmoid'))
+model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
+```
+
+### Stateful LSTM
+
+it works. Talk is cheap, see the code.
diff --git a/TimeSeries.ipynb → ShampooSales.ipynb b/TimeSeries.ipynb → ShampooSales.ipynb
@@ -3,7 +3,9 @@
   {
    "cell_type": "code",
    "execution_count": 27,
-   "metadata": {},
+   "metadata": {
+    "collapsed": true
+   },
    "outputs": [],
    "source": [
     "import numpy as np\n",
@@ -243,8 +245,8 @@
     }
    ],
    "source": [
-    "BATCH_SIEE = 4\n",
-    "train_len = len(train_scaled) // BATCH_SIEE * BATCH_SIEE\n",
+    "BATCH_SIZE = 4\n",
+    "train_len = len(train_scaled) // BATCH_SIZE * BATCH_SIZE\n",
     "train_scaled = train_scaled[:train_len]\n",
     "train_scaled.shape"
    ]
@@ -281,10 +283,10 @@
    "source": [
     "def get_model():\n",
     "    model = Sequential()\n",
-    "    model.add(LSTM(4, batch_input_shape=(BATCH_SIEE, X.shape[1], X.shape[2]), stateful=True))\n",
+    "    model.add(LSTM(4, batch_input_shape=(BATCH_SIZE, X.shape[1], X.shape[2]), stateful=True))\n",
     "\n",
-    "#     model.add(LSTM(16, batch_input_shape=(BATCH_SIEE, X.shape[1], X.shape[2]), stateful=True, return_sequences=True))\n",
-    "#     model.add(LSTM(16, batch_input_shape=(BATCH_SIEE, X.shape[1], X.shape[2]), stateful=True, return_sequences=False))\n",
+    "#     model.add(LSTM(16, batch_input_shape=(BATCH_SIZE, X.shape[1], X.shape[2]), stateful=True, return_sequences=True))\n",
+    "#     model.add(LSTM(16, batch_input_shape=(BATCH_SIZE, X.shape[1], X.shape[2]), stateful=True, return_sequences=False))\n",
     "\n",
     "    model.add(Dropout(0.5))\n",
     "    model.add(Dense(1, activation='linear'))\n",
@@ -303,13 +305,13 @@
    "outputs": [],
    "source": [
     "def make_prediction(model, X):\n",
-    "    gap = BATCH_SIEE - X.shape[0] % BATCH_SIEE\n",
+    "    gap = BATCH_SIZE - X.shape[0] % BATCH_SIZE\n",
     "    if gap != 0:\n",
     "        X = np.concatenate([X, X[:gap]], axis=0)\n",
     "        \n",
     "    X_reshape = X.reshape((X.shape[0], 1, X.shape[1]))\n",
     "    \n",
-    "    y_pred = model.predict(X_reshape, batch_size=BATCH_SIEE)\n",
+    "    y_pred = model.predict(X_reshape, batch_size=BATCH_SIZE)\n",
     "    \n",
     "    data_scaled = np.concatenate([X, y_pred], axis=1)\n",
     "    data = scaler.inverse_transform(data_scaled)\n",
@@ -606,7 +608,7 @@
    ],
    "source": [
     "for i in range(EPOCHS):\n",
-    "    model.fit(X, y, epochs=1, shuffle=False, verbose=0, batch_size=BATCH_SIEE)\n",
+    "    model.fit(X, y, epochs=1, shuffle=False, verbose=0, batch_size=BATCH_SIZE)\n",
     "    model.reset_states()\n",
     "    \n",
     "    if (i+1) % 500 == 0:\n",
@@ -671,7 +673,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.5.2"
+   "version": "3.6.1"
   }
  },
  "nbformat": 4,