-
Notifications
You must be signed in to change notification settings - Fork 2
/
trainer.py
353 lines (301 loc) · 14.3 KB
/
trainer.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
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
"""
-*- coding: utf-8 -*-
@Author: Tenzing Dolmans
@Date: 2020-07-17 10:52:26
@Last Modified by: Tenzing Dolmans
@Last Modified time: 2020-09-08 14:41:34
@Description: Trainer class that is used to train models.
Also contains evaluation functions.
"""
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from tfrecord_reader import tfrecord_dataset
from Models.mlp_base import HeadClass as MlpClass # noqa
from Models.small_mlp_base import HeadClass as sMlpClass # noqa
from Models.literature_base import HeadClass as LitClass # noqa
from Models.small_literature_base import HeadClass as sLitClass # noqa
import tensorflow as tf
import os
import sys
sys.path.insert(1, os.path.join(sys.path[0], '..'))
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
class Trainer():
"""Initialise a Trainer class for classifying MWL. Methods included
are used to load TFRecord datasets, cycle through LR, and train."""
def __init__(self,
num_epochs=3,
batch_size=8,
dropout_rate=0.2,
min_lr=0.0007,
max_lr=0.1,
min_mom=0.85,
max_mom=0.95):
self.num_epochs = num_epochs
self.batch_size = batch_size
# Note that a standard model is defined here
self.model = sMlpClass(dropout_rate)
self.loss_object = tf.keras.losses.MeanSquaredError()
self.optimizer = tf.keras.optimizers.Adam()
self.min_lr = min_lr
self.max_lr = max_lr
self.min_mom = min_mom
self.max_mom = max_mom
def get_dataset(self):
"""Loads a TFRecord dataset from a file, performs train & test
splitting and shuffling. See Tensorflow Docs for more methods.
See "hpo_search.py" before deciding whether to return one dataset
or a split train/test set."""
file = 'ZebraDatasetV1.03.tfrecord'
raw_dataset = tf.data.TFRecordDataset(file)
dataset = tfrecord_dataset(self.batch_size, raw_dataset)
dataset = dataset.shuffle(buffer_size=4200,
reshuffle_each_iteration=False)
# Split Train (90%), Test (10%)
train_dataset = dataset.take(469)
test_dataset = dataset.skip(470)
train_dataset = train_dataset.shuffle(buffer_size=4200,
reshuffle_each_iteration=True)
test_dataset = test_dataset.shuffle(buffer_size=4200,
reshuffle_each_iteration=True)
return train_dataset, test_dataset
def get_step_size(self, train_dataset):
"""Returns step size based on the dataset and num_epochs,
which is used to calculate the learning rate."""
self.num_iterations = 0
for batch in train_dataset:
self.num_iterations += 1
step_size = (self.num_epochs * self.num_iterations / 2) * 0.85
return step_size
def get_learning_rate(self, iterations_seen):
"""Returns learning rate and momentum based on progession thus far.
https://github.com/nachiket273/One_Cycle_Policy/blob/master/OneCycle.py""" # noqa
cycle = np.floor(1 + iterations_seen / (2 * self.step_size))
x = np.abs(iterations_seen / self.step_size - 2 * cycle + 1)
x = np.maximum(0, 1 - x)
div = 100
if iterations_seen > 2 * self.step_size:
ratio = (iterations_seen - 2 * self.step_size) / (
self.num_iterations * self.num_epochs - 2 * self.step_size)
_lr = self.min_lr * (1 - ratio * (1 - 1 / div))
_mom = self.max_mom
else:
_lr = self.min_lr + x * (self.max_lr - self.min_lr)
_mom = self.max_mom - x * (self.max_mom - self.min_mom)
return _lr, _mom
def forward_pass(self, gsr, ppg, nirs, et, training):
"""Simple forward pass of all data through the model.
Is used in both the training and validation/testing stages."""
y_pred = self.model(gsr=gsr, ppg=ppg, nirs=nirs, et=et,
training=training)
return y_pred
@tf.function
def train_step(self, gsr, ppg, nirs, et, label):
"""Single training step, is done for each batch in every epoch."""
with tf.GradientTape() as tape:
y_pred = self.forward_pass(
gsr=gsr, ppg=ppg, nirs=nirs, et=et, training=True)
loss_value = self.loss_object(y_true=label, y_pred=y_pred)
grads = tape.gradient(loss_value, self.model.trainable_variables)
self.optimizer.apply_gradients(
zip(grads, self.model.trainable_variables))
return y_pred
def training_loop(self, train_dataset, test_dataset):
"""Loop that wraps train_step() in epochs and batches.
Resets iterations_seen on call, returns many metrics.
NOTE: Accuracy is simply the absolute difference between
the predicted vs true label. Hence, a lower accuracy is better."""
self.iterations_seen = 0
self.step_size = self.get_step_size(train_dataset)
train_loss_results = []
train_accuracy_results = []
test_loss_results = []
test_accuracy_results = []
test_labels = []
test_predictions = []
for epoch in range(self.num_epochs):
epoch_loss_avg = tf.keras.metrics.MeanSquaredError()
epoch_accuracy = []
test_loss_avg = tf.keras.metrics.MeanSquaredError()
test_accuracy = []
epoch_labels = []
epoch_predicts = []
# --- Training --- Optimise the model
for ppg, gsr, nirs, et, label, partno in train_dataset:
y_pred = self.train_step(
gsr=gsr, ppg=ppg, nirs=nirs, et=et, label=label)
_lr, _mom = self.get_learning_rate(self.iterations_seen)
self.optimizer.learning_rate = _lr
self.optimizer.momentum = _mom
self.iterations_seen += 1
epoch_loss_avg.update_state(label, y_pred)
epoch_accuracy.extend(abs(y_pred - label))
train_loss_results.append(epoch_loss_avg.result())
train_accuracy_results.append(np.mean(epoch_accuracy))
# --- Testing --- Only a forward pass, does not update gradients
for ppg, gsr, nirs, et, label, partno in test_dataset:
y_pred = self.forward_pass(
gsr=gsr, ppg=ppg, nirs=nirs, et=et, training=False)
test_loss_avg.update_state(label, y_pred)
test_accuracy.extend(abs(y_pred - label))
epoch_labels.extend(label)
epoch_predicts.extend(y_pred)
# Keep track of all new metrics
test_labels.append(epoch_labels)
test_predictions.append(epoch_predicts)
test_loss_results.append(test_loss_avg.result())
test_accuracy_results.append(np.mean(test_accuracy))
print('Epoch {:03d}: Loss {:.4f}, Accuracy {:.4f}',
'Test Loss: {:.4f}, Test Accuracy: {:.4f}'
.format(
epoch + 1, epoch_loss_avg.result(),
np.mean(epoch_accuracy), test_loss_avg.result(),
np.mean(test_accuracy)))
# Objective is what HPO will seek to optimise. Adjust as needed.
objective = test_accuracy_results[-1]
return (train_loss_results,
train_accuracy_results,
test_loss_results,
test_accuracy_results,
test_labels,
test_predictions, objective)
"""Evaluation Functions"""
def plot_loss(train_loss_results,
train_accuracy_results,
test_loss_results,
test_accuracy_results):
"""Plots the progression of traing and testing loss + accuracy."""
fig, axes = plt.subplots(4, sharex=True, figsize=(12, 8))
fig.suptitle('Training Metrics')
axes[0].set_ylabel("Train Loss", fontsize=14)
axes[0].plot(train_loss_results)
axes[1].set_ylabel("Train Accuracy", fontsize=14)
axes[1].plot(train_accuracy_results)
axes[2].set_ylabel("Test Loss", fontsize=14)
axes[2].plot(test_loss_results)
axes[3].set_xlabel("Epoch", fontsize=14)
axes[3].set_ylabel("Test Accuracy", fontsize=14)
axes[3].plot(test_accuracy_results)
plt.show()
def evaluate_performance(test_labels, test_predictions):
"""Select one epoch in test_labels and test_predictions to work with.
Then creates a dataframe that contains all labels, predictions,
and differences. DFs for (un)acceptable results are also made and returned.
These are later used for plotting."""
label_results = [l[0].numpy() for l in test_labels[-1]]
predicted_results = [l[0].numpy() for l in test_predictions[-1]]
d = {'labels': label_results, 'predicted': predicted_results}
df = pd.DataFrame(data=d)
df['difference'] = np.array(predicted_results) - np.array(label_results)
df = df.sort_values(by='labels')
# Prevalance refers to how common each label is
prevalence = df.groupby('labels')['predicted'].nunique()
# Acceptability bound is currently one label (1/6)
acceptable = df[abs(df.predicted - df.labels) <= 1/6]
unacceptable = df[abs(df.predicted - df.labels) > 1/6]
# Accuracy is the proportion of predictions that are within 1/6
acc = len(acceptable)/len(df.difference)
print('Mean difference between label and prediction: {}'
.format(np.mean(df.difference)))
print('Percentage of predictions that are within 1.5 levels: ', acc*100)
return df, prevalence, df.difference, acceptable, unacceptable
def level_accuracies(df, prevalence):
"""returns workable lists of results that are looped over
while plotting histograms."""
_all = [df[df.labels == prevalence.keys()[i]]
for i in range(len(prevalence))]
_accept = [acceptable[acceptable.labels == prevalence.keys()[i]]
for i in range(len(prevalence))]
_unaccept = [unacceptable[unacceptable.labels == prevalence.keys()[i]]
for i in range(len(prevalence))]
return _all, _accept, _unaccept
def plot_all_histograms(data, prevalence):
"""Plots as many histograms as there are lists in input data. The histograms
are separated by label and contain a vertical line of the label and bins of
predictions around said label. """
fig, ax = plt.subplots(len(data), figsize=(20, 20))
num_bins = 15
for i, current in enumerate(data):
mu = np.mean(current.predicted)
sigma = np.std(current.predicted) + 1e-4
x = current.predicted.values
print('Level: {} Mu: {:.3f} Sigma: {:.3f}'.format(i + 1, mu, sigma))
ax[i].set_xlim(left=-0.1, right=1.1)
ax[i].set_xlabel('Predicted Label')
ax[i].set_ylabel('Probability density')
ax[i].set_title(r'Histogram of predictions: $\mu={:.3f}$',
'$\\sigma={:.3f}$'.format(mu, sigma))
# the histogram of the data
n, bins, patches = ax[i].hist(x, num_bins, density=True)
# add a 'best fit' line
y = ((1 / (np.sqrt(2 * np.pi) * sigma)) * np.exp(
-0.5 * (1 / sigma * (bins - mu))**2))
ax[i].plot(bins, y, '--')
ax[i].axvline(prevalence.keys()[i])
fig.tight_layout() # Tweak spacing to prevent clipping of ylabel
plt.show()
def plot_histogram(difference, acceptable, unacceptable):
"""Plots a single histogram that contains all predictions'
relative offset. The slimmer and closer to zero the histogram,
the better the results."""
fig, ax = plt.subplots()
num_bins = 30
"""All differences"""
mu0 = np.mean(difference.predicted - difference.labels)
sigma0 = np.std(difference.predicted - difference.labels) + 1e-4
x0 = df.difference
ax.set_xlim(left=-1, right=1)
ax.set_xlabel('Predicted Label Difference')
ax.set_ylabel('Number of Predictions')
ax.set_title(r'Histogram of Predictions: $\mu={:.3f}$, $\sigma={:.3f}$'
.format(mu0, sigma0))
"""Differences within an acceptable range"""
mu1 = np.mean(acceptable.predicted - acceptable.labels)
sigma1 = np.std(acceptable.predicted - acceptable.labels) + 1e-4
x1 = acceptable.predicted - acceptable.labels
"""Differences outside an acceptable range"""
mu2 = np.mean(unacceptable.predicted - unacceptable.labels)
sigma2 = np.std(unacceptable.predicted - unacceptable.labels) + 1e-4
x2 = unacceptable.predicted - unacceptable.labels
"""Plot the Histogram(s)"""
n0, bins0, patches0 = ax.hist(x0, num_bins, density=False, color='blue')
n1, bins1, patches1 = ax.hist(x1, num_bins, density=False, color='green')
n2, bins2, patches2 = ax.hist(x2, num_bins, density=False, color='red')
"""Add 'best fit' line(s)"""
y0 = ((1 / (np.sqrt(2 * np.pi) * sigma0)) * np.exp(
-0.5 * (1 / sigma0 * (bins0 - mu0))**2))
ax.plot(bins0, y0, '-')
y1 = ((1 / (np.sqrt(2 * np.pi) * sigma1)) * np.exp(
-0.5 * (1 / sigma1 * (bins1 - mu1))**2))
ax.plot(bins1, y1, '--')
y2 = ((1 / (np.sqrt(2 * np.pi) * sigma2)) * np.exp(
-0.5 * (1 / sigma2 * (bins2 - mu2))**2))
ax.plot(bins2, y2, '--')
"""Vertical lines to indicate acceptable bounds."""
ax.axvline(-1/6)
ax.axvline(1/6)
fig.tight_layout() # Tweak spacing to prevent clipping of ylabel
plt.show()
if __name__ == "__main__":
# Select a GPU to train with
physical_devices = tf.config.experimental.list_physical_devices('GPU')
tf.config.experimental.set_memory_growth(physical_devices[0], True)
"""TRAINING"""
# Give the HPs that were found by HPO
trainer = Trainer(num_epochs=25,
batch_size=8,
dropout_rate=0.1789,
min_lr=0.000000796,
max_lr=0.0000341,
min_mom=0.7403,
max_mom=0.7985)
train_dataset, test_dataset = trainer.get_dataset()
(tr_loss, tr_acc, te_loss, te_acc,
te_labels, te_preds, objective) = trainer.training_loop(
train_dataset, test_dataset)
"""EVALUATION"""
(df, prevalence, difference,
acceptable, unacceptable) = evaluate_performance(te_labels, te_preds)
_all, _accept, _unaccept = level_accuracies(df, prevalence)
plot_all_histograms(_all, prevalence)
plot_histogram(difference, acceptable, unacceptable)