Created
February 8, 2018 19:23
-
-
Save jamartinh/a38d8f0322818eeca9b6c5fcae5bc165 to your computer and use it in GitHub Desktop.
A Deep Statefull LSTM
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
import os | |
import keras | |
from keras import Input | |
from keras.layers import BatchNormalization, Dense, LSTM, concatenate, Dropout, TimeDistributed | |
from keras.optimizers import RMSprop | |
from keras.utils import plot_model | |
from metrics import in_top5, top5_accuracy | |
def gen_nn_model(hyperparameters, target_symbols): | |
# - Input definitions: | |
batch_size = hyperparameters['batch_size'] | |
n_cols = hyperparameters['n_cols'] | |
n_history = hyperparameters['n_history'] | |
# -- Daily Input | |
daily_input = Input(shape = (n_history, n_cols), batch_shape = (batch_size, n_history, n_cols), dtype = 'float32', | |
name = 'daily_input') | |
normalized_daily_input = BatchNormalization()(daily_input) | |
# normalized_daily_input = TimeDistributed(Dense(n_cols, activation='linear'))(normalized_daily_input) | |
normalized_daily_input = TimeDistributed(Dense(n_cols, activation = 'relu'))(normalized_daily_input) | |
normalized_daily_input = TimeDistributed(Dense(n_cols, activation = 'relu'))(normalized_daily_input) | |
normalized_daily_input = TimeDistributed(Dense(n_cols, activation = 'relu'))(normalized_daily_input) | |
normalized_daily_input = Dropout(0.2)(normalized_daily_input) | |
# Deep LSTM Layers | |
x = LSTM(n_cols // 7, | |
return_state = False, | |
return_sequences = True, | |
stateful = True)(normalized_daily_input) | |
x = LSTM(n_cols // 7, | |
stateful = False)(x) | |
r = Dense(n_cols - 1, activation = 'tanh')(x) | |
r = Dense(n_cols - 1, activation = 'tanh')(r) | |
r = Dense(len(target_symbols), activation = 'linear', name = 'r')(r) | |
joint_flow = concatenate([x, r]) | |
p = Dense(n_cols - 1, activation = 'sigmoid')(joint_flow) | |
p = Dense(n_cols - 1, activation = 'sigmoid')(p) | |
# p = keras.layers.core.Lambda(lambda x: x *5)(p) | |
p = Dense(len(target_symbols), activation = 'softmax', name = 'p')(p) | |
# l = keras.layers.core.Lambda(lambda x: negative(x))(p) | |
l = Dense(n_cols - 1, activation = 'sigmoid')(joint_flow) | |
l = Dense(n_cols - 1, activation = 'sigmoid')(l) | |
# l = keras.layers.core.Lambda(lambda x: x / 2)(l) | |
l = Dense(len(target_symbols), activation = 'softmax', name = 'l')(l) | |
# -- Model creation | |
inputs = [daily_input] | |
outputs = [r, p, l] | |
model = keras.models.Model(inputs = inputs, outputs = outputs) | |
return batch_size, model | |
def compile_model(model): | |
lr = 0.1 | |
loss = { 'r': 'mean_absolute_error', | |
'p': 'kullback_leibler_divergence', | |
'l': 'kullback_leibler_divergence' | |
} | |
loss_weights = { 'r': .20, | |
'p': .40, | |
'l': .40, | |
} | |
opt = RMSprop(lr = lr) | |
# opt = 'adam' | |
metrics = { 'p': [in_top5, top5_accuracy], | |
'l': [in_top5, top5_accuracy] } | |
model.compile(optimizer = opt, | |
loss = loss, | |
loss_weights = loss_weights, | |
metrics = metrics | |
) | |
return model | |
def show_model_details(EXP_NAME, EXP_PATH, model): | |
print(model.summary()) | |
str_image_file_name = os.path.join(EXP_PATH, '{0}.png'.format(EXP_NAME)) | |
plot_model(model, to_file = str_image_file_name, | |
show_shapes = True, | |
show_layer_names = True, | |
rankdir = 'TD') |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment