-
Notifications
You must be signed in to change notification settings - Fork 10
/
train_crypto.py
129 lines (109 loc) · 5.2 KB
/
train_crypto.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
from typing import Tuple
import os
import random
import pandas as pd
import numpy as np
from joblib import dump, load
import matplotlib.pyplot as plt
import tensorflow as tf
from tensorflow.keras.models import Sequential, load_model
from tensorflow.keras.callbacks import EarlyStopping
from tensorflow.keras.metrics import RootMeanSquaredError, MeanAbsoluteError
from preprocessing import CryptoPreprocessor
from SCINet import SCINet, StackedSCINet
from explore import lower_granularity
# Make model
def make_model(input_shape, output_shape):
inputs = tf.keras.Input(shape=(input_shape[1], input_shape[2]), name='inputs')
# x = SciNet(horizon, levels=L, h=h, kernel_size=kernel_size)(inputs)
# model = tf.keras.Model(inputs, x)
targets = tf.keras.Input(shape=(output_shape[1], output_shape[2]), name='targets')
predictions = StackedSCINet(horizon=horizon, features=input_shape[-1], stacks=K, levels=L, h=h,
kernel_size=kernel_size,
regularizer=(l1, l2))(inputs, targets)
model = tf.keras.Model(inputs=[inputs, targets], outputs=predictions)
model.summary()
tf.keras.utils.plot_model(model, to_file='modelDiagram.png', show_shapes=True)
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate), loss='mse',
metrics=['mean_squared_error', 'mean_absolute_error'])
return model
# Parametres
degree_of_differencing = 0
look_back_window, horizon = 184, 60
batch_size = 16
learning_rate = 9e-3
h, kernel_size, L, K = 4, 5, 3, 2
l1, l2 = 0.001, 0.1
# split_strides = look_back_window + horizon
split_strides = 1
if __name__ == '__main__':
# Load and preprocess data
data = pd.read_csv('crypto_data/USD-2021-06-17-2021-09-12.csv', index_col=0, header=[0, 1])
data.index = pd.to_datetime(data.index)
data = lower_granularity(data, pd.Timedelta(15, 'min'))
train_data = data[:int(0.6 * len(data))]
val_data = data[int(0.6 * len(data)):int(0.8 * len(data))]
test_data = data[int(0.8 * len(data)):]
# Train model
preprocessor = CryptoPreprocessor(look_back_window, horizon, split_strides, degree_of_differencing,
relative_diff=False, scaling='standard')
X_train, y_train = preprocessor.fit_transform(train_data)
X_val, y_val = preprocessor.transform(val_data)
print(f'Input shape: X{X_train.shape}, y{y_train.shape}')
model = make_model(X_train.shape, y_train.shape)
early_stopping = EarlyStopping(monitor='val_loss', patience=100, min_delta=0, verbose=1, restore_best_weights=True)
history = model.fit({'inputs': X_train, 'targets': y_train},
validation_data={'inputs': X_val, 'targets': y_val},
batch_size=batch_size, epochs=1600, callbacks=[early_stopping])
# Generate new id and create save directory
existing_ids = [int(name) for name in os.listdir('saved-models/') if name.isnumeric()]
run_id = random.choice(list(set(range(0, 1000)) - set(existing_ids)))
save_directory = f'saved-models/regressor/{run_id:03d}/'
os.makedirs(os.path.dirname(save_directory), exist_ok=True)
# Save model, preprocessor and training history
model.save(save_directory)
with open(save_directory + 'preprocessor', 'wb') as f:
dump(preprocessor, f, compress=3)
pd.DataFrame(history.history).to_csv(save_directory + 'train_history.csv')
# Plot accuracy
plt.plot(history.history['mean_absolute_error'])
plt.plot(history.history['val_mean_absolute_error'])
plt.title('model accuracy')
plt.ylabel('mean absolute error')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='upper right')
plt.savefig(save_directory + 'accuracy.png')
plt.clf()
# Plot loss
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('model loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='upper right')
plt.savefig(save_directory + 'loss.png')
# Evaluate
# run_id = 186
# model = load_model(f'saved-models/regressor/{run_id:03d}/')
# with open(f'saved-models/regressor/{run_id:03d}/preprocessor', 'rb') as f:
# preprocessor = load(f)
X_test, y_test = preprocessor.transform(test_data)
scores = model.evaluate({'inputs': X_test, 'targets': y_test})
# Save evaluation results
if not isinstance(scores, list):
scores = [scores]
row = [run_id] + scores + [pd.Timestamp.now(tz='Australia/Melbourne')]
try:
df_scores = pd.read_csv('saved-models/scores.csv')
df_scores.loc[len(df_scores)] = row
except (FileNotFoundError, ValueError):
df_scores = pd.DataFrame([row], columns=['id'] + list(model.metrics_names) + ['time'])
df_scores.to_csv('saved-models/scores.csv', index=False)
# Predict
# y_test is only used to calculate loss, how to get rid of it?
y_pred = model.predict({'inputs': X_test, 'targets': y_test})
y_pred = preprocessor.scaler.inverse_transform()
y_test = preprocessor.scaler.inverse_transform()
comparison = np.hstack([y_pred, y_test])
df = pd.DataFrame(comparison, columns=['Predicted', 'Actual'])
df.to_csv(f'saved-models/regressor/{run_id:03d}/comparison.csv')