I run kaggle store forecast for training
I have the following
model_lstm = Sequential()
#it use to be 190
model_lstm.add(LSTM(50, activation=‘tanh’, return_sequences=True,input_shape=(X_train_series.shape[1], X_train_series.shape[2]), kernel_regularizer=regularizers.l2(0.001))) # Use n_steps_in and n_features here
#model_lstm.add(Dropout(0.2))
model_lstm.add(LSTM(250, activation=‘tanh’))
#model_lstm.add(RepeatVector(6))
model_lstm.add(Dense(250, activation=‘relu’)) # Additional Dense layer
model_lstm.add(Dense(50, activation=‘relu’)) # Additional Dense layer
model_lstm.add(Dense(1))
#model_lstm.add(TimeDistributed(Dense(1)))
#it used to be 0.0001
model_lstm.compile(loss=‘mse’, optimizer=optimizers.Adam(learning_rate=0.001, clipnorm=1.0))
model_lstm.summary()
and i found out by running fit inside a for loop performs better
for i in range(1,140):
lstm_history = model_lstm.fit(X_train_series, Y_train,
validation_data=(X_valid_series, Y_valid),
epochs=1,
batch_size=512,
#callbacks=[lr_scheduler],
verbose=2)
I need guidance on what should i do Store Item demand forecasting | Kaggle to reduce the underfitting in particular cases