161 lines
4.9 KiB
Python
161 lines
4.9 KiB
Python
import numpy as np
| |
import pandas as pd
| |
| |
from sklearn.preprocessing import StandardScaler
| |
| |
import tensorflow as tf
| |
from tensorflow import keras
| |
from keras.optimizers import *
| |
from keras.models import Sequential
| |
from keras.layers import Dense, Dropout
| |
from keras.regularizers import *
| |
from keras.callbacks import *
| |
| |
| |
def train_predict_nn(df_X, df_y, df_X_test, model_config: dict):
| |
"""
| |
Train model with the specified hyper-parameters and return its predictions for the test data.
| |
"""
| |
model_pair = train_nn(df_X, df_y, model_config)
| |
y_test_hat = predict_nn(model_pair, df_X_test, model_config)
| |
return y_test_hat
| |
| |
def train_nn(df_X, df_y, model_config: dict):
| |
"""
| |
Train model with the specified hyper-parameters and return this model (and scaler if any).
| |
"""
| |
params = model_config.get("params", {})
| |
| |
is_scale = params.get("is_scale", True)
| |
is_regression = params.get("is_regression", False)
| |
| |
#
| |
# Scale
| |
#
| |
if is_scale:
| |
scaler = StandardScaler()
| |
scaler.fit(df_X)
| |
X_train = scaler.transform(df_X)
| |
else:
| |
scaler = None
| |
X_train = df_X.values
| |
| |
y_train = df_y.values
| |
| |
#
| |
# Create model
| |
#
| |
n_features = X_train.shape[1]
| |
layers = params.get("layers") # List of ints
| |
if not layers:
| |
layers = [n_features // 4] # Default
| |
if not isinstance(layers, list):
| |
layers = [layers]
| |
| |
# Topology
| |
model = Sequential()
| |
# sigmoid, relu, tanh, selu, elu, exponential
| |
# kernel_regularizer=l2(0.001)
| |
| |
reg_l2 = 0.001
| |
| |
train_conf = model_config.get("train", {})
| |
learning_rate = train_conf.get("learning_rate")
| |
n_epochs = train_conf.get("n_epochs")
| |
batch_size = train_conf.get("bs")
| |
| |
for i, out_features in enumerate(layers):
| |
in_features = n_features if i == 0 else layers[i-1]
| |
model.add(Dense(out_features, activation='sigmoid', input_dim=in_features)) # , kernel_regularizer=l2(reg_l2)
| |
#model.add(Dropout(rate=0.5))
| |
| |
if is_regression:
| |
model.add(Dense(units=1))
| |
| |
model.compile(
| |
loss='mean_squared_error',
| |
optimizer=Adam(learning_rate=learning_rate),
| |
metrics=[
| |
tf.keras.metrics.MeanAbsoluteError(name="mean_absolute_error"),
| |
tf.keras.metrics.MeanAbsolutePercentageError(name="mean_absolute_percentage_error"),
| |
tf.keras.metrics.R2Score(name="r2_score"),
| |
],
| |
)
| |
else:
| |
model.add(Dense(units=1, activation='sigmoid'))
| |
| |
model.compile(
| |
loss='binary_crossentropy',
| |
optimizer=Adam(learning_rate=learning_rate),
| |
metrics=[
| |
tf.keras.metrics.AUC(name="auc"),
| |
tf.keras.metrics.Precision(name="precision"),
| |
tf.keras.metrics.Recall(name="recall"),
| |
],
| |
)
| |
| |
#model.summary()
| |
| |
# Default arguments for early stopping
| |
es_args = dict(
| |
monitor = "loss", # val_loss loss
| |
min_delta = 0.00001, # Minimum change qualified as improvement
| |
patience = 5, # Number of epochs with no improvements
| |
verbose = 0,
| |
mode = 'auto',
| |
)
| |
es_args.update(train_conf.get("es", {})) # Overwrite default values with those explicitly specified in config
| |
| |
es = EarlyStopping(**es_args)
| |
| |
#
| |
# Train
| |
#
| |
model.fit(
| |
X_train,
| |
y_train,
| |
batch_size=batch_size,
| |
epochs=n_epochs,
| |
#validation_split=0.05,
| |
#validation_data=(X_validate, y_validate),
| |
#class_weight={0: 1, 1: 20},
| |
callbacks=[es],
| |
verbose=1,
| |
)
| |
| |
return (model, scaler)
| |
| |
def predict_nn(models: tuple, df_X_test, model_config: dict):
| |
"""
| |
Use the model(s) to make predictions for the test data.
| |
The first model is a prediction model and the second model (optional) is a scaler.
| |
"""
| |
#
| |
# Scale
| |
#
| |
scaler = models[1]
| |
is_scale = scaler is not None
| |
| |
input_index = df_X_test.index
| |
if is_scale:
| |
df_X_test = scaler.transform(df_X_test)
| |
df_X_test = pd.DataFrame(data=df_X_test, index=input_index)
| |
else:
| |
df_X_test = df_X_test
| |
| |
df_X_test_nonans = df_X_test.dropna() # Drop nans, possibly create gaps in index
| |
nonans_index = df_X_test_nonans.index
| |
| |
# Resets all (global) state generated by Keras
| |
# Important if prediction is executed in a loop to avoid memory leak
| |
tf.keras.backend.clear_session()
| |
| |
y_test_hat_nonans = models[0].predict_on_batch(df_X_test_nonans.values) # NN returns matrix with one column as prediction
| |
y_test_hat_nonans = y_test_hat_nonans[:, 0] # Or y_test_hat.flatten()
| |
y_test_hat_nonans = pd.Series(data=y_test_hat_nonans, index=nonans_index) # Attach indexes with gaps
| |
| |
df_ret = pd.DataFrame(index=input_index) # Create empty dataframe with original index
| |
df_ret["y_hat"] = y_test_hat_nonans # Join using indexes
| |
sr_ret = df_ret["y_hat"] # This series has all original input indexes but NaNs where input is NaN
| |
| |
return sr_ret
|