-
Notifications
You must be signed in to change notification settings - Fork 0
/
data.py
135 lines (108 loc) · 5.59 KB
/
data.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
#!/usr/bin/python3
import tensorflow as tf
import numpy as np
import pandas as pd
class Data(object):
@staticmethod
def load_data(filename, evaluate=False, adversary=False, parquet=False, tune=False,
pivots=['_pivot'], auxillary=None, adv_n_classes=8):
if parquet:
import pyarrow.parquet as pq
dataset = pq.ParquetDataset(filename)
df = dataset.read(nthreads=4).to_pandas()
else:
df = pd.read_hdf(filename, key='df')
if not evaluate:
df = df.sample(frac=1).reset_index(drop=True)
if auxillary is None:
# Cleanup + omit variables prefixed with an underscore from training
auxillary = [col for col in df.columns if col.startswith('_')]
auxillary = list(set(auxillary))
df_features = df.drop(auxillary, axis=1)
print('Data shape:', df_features.shape)
print('Features', df_features.columns.tolist())
pivot_df = df[pivots]
pivot_features = pivot_df[pivots]
if adversary:
# Bin variable -> discrete classification problem
# Each protected variable must be binned separately
if len(pivots) == 1:
pivot = pivots[0]
pivot_df = pivot_df.assign(pivot_labels=pd.qcut(df[pivot], q=adv_n_classes, labels=False))
pivot_labels = pivot_df.pivot_labels
marginal_df = pivot_features[pivots]
marginal_df = marginal_df.sample(frac=1)
marginal_df = marginal_df.reset_index(drop=True)
pivot_features[[pivot + '_marginal' for pivot in pivots]] = marginal_df
if evaluate:
if tune:
from ray.tune.util import pin_in_object_store as pin
return pin(np.nan_to_num(df_features.values)), pin(df._label.values.astype(np.int32)), \
pin(pivot_features.values.astype(np.float32)), pin(df[auxillary])
else:
return df, np.nan_to_num(df_features.values), df._label.values.astype(np.int32), \
pivot_features.values.astype(np.float32)
else:
if adversary:
return np.nan_to_num(df_features.values), df._label.values.astype(np.int32), \
pivot_features.values.astype(np.float32), pivot_labels.values.astype(np.int32)
else:
if tune:
from ray.tune.util import pin_in_object_store as pin
return pin(np.nan_to_num(df_features.values)), pin(df._label.values.astype(np.int32)), \
pin(pivot_features.values.astype(np.float32))
else:
return np.nan_to_num(df_features.values), df._label.values.astype(np.int32), \
pivot_features.values.astype(np.float32)
@staticmethod
def load_dataset(features_placeholder, labels_placeholder, pivots_placeholder, batch_size, test=False,
evaluate=False, sequential=False, prefetch_size=2, adversary=False, pivot_labels_placeholder=None):
if adversary:
dataset = tf.data.Dataset.from_tensor_slices((features_placeholder, labels_placeholder,
pivots_placeholder, pivot_labels_placeholder))
padded_shapes = (tf.TensorShape([None]), tf.TensorShape([]), tf.TensorShape([]), tf.TensorShape([]))
padding_values = (0.,0,0.,0)
else:
dataset = tf.data.Dataset.from_tensor_slices((features_placeholder, labels_placeholder,
pivots_placeholder))
padded_shapes = (tf.TensorShape([None]), tf.TensorShape([]), tf.TensorShape([]))
padding_values = (0.,0,0.)
# Retain order if evaluate=True
if evaluate is False:
dataset = dataset.shuffle(buffer_size=10**5)
if sequential:
dataset = dataset.padded_batch(
batch_size,
padded_shapes=padded_shapes,
padding_values=padding_values,
drop_remainder=True)
else:
# Don't bottleneck GPU with data loading
dataset = dataset.batch(batch_size, drop_remainder=True)
dataset = dataset.prefetch(prefetch_size)
if test is True:
dataset = dataset.repeat()
return dataset
@staticmethod
def load_tfr_dataset(filenames_placeholder, n_features, batch_size, test=False, evaluate=False):
# TODO: Parallelize tfrecord decoding, fuse map/batch
num_parallel_calls = 4
prefetch_size = 4
def _parse_function(example_proto):
features_desc = {"data": tf.FixedLenFeature([n_features], tf.float32),
"labels": tf.FixedLenFeature((), tf.float32),
"meta": tf.FixedLenFeature((), tf.float32)}
parsed_features = tf.parse_single_example(example_proto, features_desc)
return parsed_features["data"], parsed_features["labels"], parsed_features["meta"]
dataset = tf.data.TFRecordDataset(filenames_placeholder)
# Parse the record into tensors.
dataset = dataset.map(_parse_function, num_parallel_calls=num_parallel_calls)
if test: # Repeat input indefinitely
dataset = dataset.apply(tf.data.experimental.shuffle_and_repeat(buffer_size=int(1e6), count=None))
else:
if evaluate is False: # Retain order for evaluation
dataset = dataset.shuffle(buffer_size=int(1e6))
dataset = dataset.batch(batch_size, drop_remainder=True)
# Enqueue batches on CPU
dataset = dataset.prefetch(prefetch_size)
return dataset