- shift+enter: run cell
- a: insert cell above
- b: insert cell below
- dd: delete cell
- Use an exclamation mark in the command to use linux commands. eg.
!mkdir datawill make a data directory.
Before you jump in to copying and pasting the model below answer these questions:
- Is it a regression/binary classifiction/ multi class classification problem? see activation and loss function
- Is the X data real (floating point) numbers? Then normalise it so that
\begin{align}
X_{new} = \frac{X - mean(X)}{stddev(X)}
\end{align}.
Same with
y. - Is the X data categorical? (eg. words, colours, car type). Then create a dictionary that will map categories to numbers. Then use the
Embeddinglayer to input the data.
from keras.layers import Dense, Activation
# Must provide either input_dim or input_shape for the **first layer only**
model.add(Dense(units=64, input_dim=100, activation='relu'))
model.add(Dense(units=10))
model.add(Activation('softmax'))
model.compile(loss='sparse_categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
Whatever the layer that you need, you simply add it (eg. model.add(LSTM(...)), model.add(Dropout(...))).
Training the model:
model.fit(X_train, y_train, epochs=5, batch_size=256)
Predicting with model:
model.predict(X_train, batch_size=256)
- Regression - None or 'linear' (no activation implies linear)
- Classifiction (two classes) - 'sigmoid'
- Classifiction (multiple classes) - 'softmax'
- Regression - 'mean_squared_error'
- Classifiction (two classes) - 'binary_crossentropy'
- Classifiction (multiple classes) - 'sparse_categorical_crossentropy' It is sparse since with the other option 'categorical_crossentropy' you need to provide a one hot encoded version of the classes.