-
Notifications
You must be signed in to change notification settings - Fork 1
/
run_models.py
executable file
·349 lines (308 loc) · 12 KB
/
run_models.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
import numpy as np
import tensorflow as tf
from lib.neuralnetwork import Fair_MNN
from lib.zemel import ZemelFairRepresentations, k_nearest_neighbors_sp
from lib.general_functions import now
from lib.metrics import accuracy_np, discrimination_np, consistency_np
from lib.metrics import k_nearest_neighbors_sp, identify_monotonic_pairs
from lib.metrics import resentment_individual, resentment_pairwise
from lib.metrics import lipschitz_sample_estimate
from matplotlib import pyplot as plt
import pickle as pk
from lib.compas import load_data as load_data_compas
from lib.law_school import load_data as load_data_law_school
from lib.german import load_data as load_data_german
for data_name in ['compas', 'law_school', 'german']:
print now(), data_name
tf.reset_default_graph()
if data_name == 'compas':
load_data = load_data_compas
monotonicity = [1, 0, 1, 1, 1]
data = load_data(binarize=True)
if data_name == 'law_school':
load_data = load_data_law_school
monotonicity = [1, 1]
data = load_data(binarize=True)
if data_name == 'german':
load_data = load_data_german
data = load_data()
monotonicity = data['monotonicity']
use_minibatch = True
valid_frac = 0.2
n_valid = int(data['data_train'].shape[0] * valid_frac)
X_train = data['data_train'][:-n_valid, data['X_cols']].astype(np.number)
X_valid = data['data_train'][-n_valid:, data['X_cols']].astype(np.number)
X_test = data['data_test' ][:, data['X_cols']].astype(np.number)
Y_train = data['data_train'][:-n_valid, data['Y_col']:data['Y_col']+1].astype(np.number)
Y_valid = data['data_train'][-n_valid:, data['Y_col']:data['Y_col']+1].astype(np.number)
Y_test = data['data_test' ][:, data['Y_col']:data['Y_col']+1].astype(np.number)
A_train = data['data_train'][:-n_valid, data['A_cols']][:, -2:].astype(np.number)
A_valid = data['data_train'][-n_valid:, data['A_cols']][:, -2:].astype(np.number)
A_test = data['data_test' ][:, data['A_cols']][:, -2:].astype(np.number)
n, x_dim = X_train.shape
X_pretrain = X_train + 0.
Y_prevalid = X_valid + 0.
X_pretest = X_test + 0.
std = X_train.std(axis=0)
X_train = X_train / std
X_valid = X_valid / std
X_test = X_test / std
if use_minibatch:
batch_size = 128 if data_name == 'german' else 256
X_tf, Y_tf, A_tf = [tf.constant(d) for d in [X_train, Y_train, A_train]]
ds = tf.data.Dataset.from_tensor_slices({'X': X_tf, 'Y': Y_tf, 'A': A_tf})
ds = ds.repeat()
ds = ds.shuffle(buffer_size=batch_size * 3)
ds = ds.batch(batch_size=batch_size)
ds = ds.prefetch(batch_size)
iterator = ds.make_initializable_iterator()
next_element = iterator.get_next()
X_trn, Y_trn, A_trn = [next_element[key] for key in ['X', 'Y', 'A']]
else:
batch_size = X_train.shape[0]
X_trn, Y_trn, A_trn = [tf.constant(d) for d in [X_train, Y_train, A_train]]
batch_factor = n / batch_size
n, x_dim = X_train.shape
fnn = Fair_MNN(
x_dim = x_dim,
a_dim = 2,
var_scope = 'fair_mnn',
nn_params = {
'dtype': tf.float64,
'num_layers': 5,
'width': [[10], [10], [10], [10], [1]],
'output_dim': 1,
'activations': [tf.tanh],
'activate_last_layer': False,
'monotonicity': [0] * x_dim, # None = all positive
'positive_func': lambda w: 1. + tf.nn.elu(w - 1.),
'var_scope': 'ffnn',
},
fairness_loss_type = ['DP', 'EOdds', 'EOutcome', 'EOpportunity', 'ZemelDiscrimination'][4],
y_loss_type = ['CE', 'EC'][0],
use_minibatch = False,
batch_size = batch_size,
X_minibatch_tf = X_trn if use_minibatch else None,
Y_minibatch_tf = Y_trn if use_minibatch else None,
A_minibatch_tf = A_trn if use_minibatch else None,
)
fmnn = Fair_MNN(
x_dim = x_dim,
a_dim = 2,
var_scope = 'fair_mnn',
nn_params = {
'dtype': tf.float64,
'num_layers': 5,
'width': [[10], [10], [10], [10], [1]],
'output_dim': 1,
'activations': [tf.tanh],
'activate_last_layer': False,
'monotonicity': monotonicity, # None = all positive
'positive_func': lambda w: 1. + tf.nn.elu(w - 1.),
'var_scope': 'mffnn',
},
fairness_loss_type = ['DP', 'EOdds', 'EOutcome', 'EOpportunity', 'ZemelDiscrimination'][4],
y_loss_type = ['CE', 'EC'][0],
use_minibatch = False,
batch_size = batch_size,
X_minibatch_tf = X_trn if use_minibatch else None,
Y_minibatch_tf = Y_trn if use_minibatch else None,
A_minibatch_tf = A_trn if use_minibatch else None,
)
num_clusters = 10
zfr = ZemelFairRepresentations(
x_dim = x_dim,
num_clusters = num_clusters,
sigmas = X_train.var(axis=0),
var_scope = 'ZFR',
dtype=tf.float64,
optimizer=tf.train.AdamOptimizer,
)
sess = tf.Session()
fnn.sess=sess
fnn.sess=sess
zfr.sess=sess
if use_minibatch:
sess.run(iterator.initializer)
n_runs = 250
n_steps = 30
n_runs = 100
n_steps = 30
fnn_alphas = [0, 0.01, 0.1, 0.25, 0.5, 0.75, 0.9, 0.99, 0.999]
fnn_loss_by_step = np.zeros([n_runs, n_steps])
fnn_losses = {}
fnn_states = {}
print ''
print now(), 'fnn'
for run in range(n_runs):
if run < len(fnn_alphas):
alpha = fnn_alphas[run] + 0.
else:
alpha = np.random.beta(.5, .5, 1)[0]
fnn_losses[alpha, run] = np.inf
fnn_states[alpha, run] = []
fnn.initialize(sess=sess)
# Initialize the bias semi-intelligently
fnn.vars_tf[-1].load(np.array([0.]), sess)
for step in range(n_steps):
losses = fnn.fit(
X_np = None if use_minibatch else X_train,
Y_np = None if use_minibatch else Y_train,
A_np = None if use_minibatch else A_train,
A_0 = [0.4, 0.6],
N_0 = 0.5,
lambda_A = alpha,
lambda_N = 0.,
lambda_Y = 1. - alpha,
n_steps = int(10 * batch_factor),
learning_rate = (
1e-2 / batch_factor if step < 33 else
1e-3 / batch_factor # if step < 100 else
),
sess = sess,
)
losses = fnn.predict_and_losses(
X_np = X_valid,
Y_np = Y_valid,
A_np = A_valid,
A_0_np = [Y_train.mean()]*2,
N_0_np = Y_train.mean(),
lambda_A = alpha,
lambda_N = 0.,
lambda_Y = 1. - alpha,
)[1]
fnn_loss_by_step[run, step] = losses['loss'] + 0.
if losses['loss'] < fnn_losses[alpha, run]:
fnn_losses[alpha, run] = losses['loss'] + 0.
fnn_states[alpha, run] = [v + 0. for v in fnn.save_vars()]
print now(), run, alpha, losses['loss']
pk.dump(
{
'fnn_losses': fnn_losses,
'fnn_states': fnn_states,
'fnn_loss_by_step': fnn_loss_by_step,
},
open('data/{:}_fnn_var_state_3.pk'.format(data_name), 'wb'),
)
n_runs = 100
n_steps = 30
fmnn_alphas = [0, 0.01, 0.1, 0.25, 0.5, 0.75, 0.9, 0.99, 0.999]
fmnn_loss_by_step = np.zeros([n_runs, n_steps])
fmnn_losses = {}
fmnn_states = {}
print ''
print now(), 'fmnn'
for run in range(n_runs):
if run < len(fmnn_alphas):
alpha = fmnn_alphas[run] + 0.
else:
alpha = np.random.beta(.5, .5, 1)[0]
fmnn_losses[alpha, run] = np.inf
fmnn_states[alpha, run] = []
fmnn.initialize(sess=sess)
# Initialize the bias semi-intelligently
fmnn.vars_tf[-1].load(np.array([0.]), sess)
for step in range(n_steps):
losses = fmnn.fit(
X_np = None if use_minibatch else X_train,
Y_np = None if use_minibatch else Y_train,
A_np = None if use_minibatch else A_train,
A_0 = [0.4, 0.6],
N_0 = 0.5,
lambda_A = alpha,
lambda_N = 0.,
lambda_Y = 1. - alpha,
n_steps = int(10 * batch_factor),
learning_rate = (
1e-2 / batch_factor if step < 33 else
1e-3 / batch_factor # if step < 100 else
),
sess = sess,
)
losses = fmnn.predict_and_losses(
X_np = X_valid,
Y_np = Y_valid,
A_np = A_valid,
A_0_np = [Y_train.mean()]*2,
N_0_np = Y_train.mean(),
lambda_A = alpha,
lambda_N = 0.,
lambda_Y = 1. - alpha,
)[1]
fmnn_loss_by_step[run, step] = losses['loss']
if losses['loss'] < fmnn_losses[alpha, run]:
fmnn_losses[alpha, run] = losses['loss'] + 0.
fmnn_states[alpha, run] = [v + 0. for v in fmnn.save_vars()]
print now(), run, alpha, losses['loss']
pk.dump(
{
'fmnn_losses': fmnn_losses,
'fmnn_states': fmnn_states,
'fmnn_loss_by_step': fmnn_loss_by_step,
},
open('data/{:}_fmnn_var_state_3.pk'.format(data_name), 'wb'),
)
n_runs = 100
n_steps = 40
zfr_alphas = [0, 0.01, 0.1, 0.25, 0.5, 0.75, 0.9, 0.99, 0.999]
zfr_lambdas_X = 10. ** np.array([-1., -.5, 0., .5, 1.])
zfr_loss_by_step = np.zeros([n_runs, n_steps])
zfr_losses = {}
zfr_states = {}
print ''
print now(), 'zfr'
for run in range(n_runs):
if run < len(zfr_alphas) * len(zfr_lambdas_X):
alpha = zfr_alphas[run % len(zfr_alphas)] + 0.
lambda_X = zfr_lambdas_X[run // len(zfr_alphas)] + 0.
else:
alpha = np.random.beta(.5, .5, 1)[0]
lambda_X = 10 ** (2. * np.random.beta(1., 1., 1)[0] - 1.)
zfr_losses[alpha, lambda_X, run] = np.inf
zfr_states[alpha, lambda_X, run] = []
zfr.initialize(sess=sess)
# Do a smart(er) initialization
X_sample = X_train[np.random.choice(X_train.shape[0], size=num_clusters), :]
while np.array([X_sample[i, :] == X_sample[j, :]
for i in range(num_clusters-1)
for j in range(i+1, num_clusters)]
).min(1).max():
X_sample = X_train[np.random.choice(X_train.shape[0], size=num_clusters), :]
zfr.set_weights(
V_np = X_sample.transpose(),
W_np = np.random.uniform(size=[num_clusters, 1]),
)
for step in range(n_steps):
zfr.fit(
X_np = X_train,
Y_np = Y_train,
A_np = A_train,
alpha_Z = alpha,
alpha_X = lambda_X,
alpha_Y = 1. - alpha,
n_steps = n_steps,
learning_rate = 1e-2 if step < 30 else 1e-3,
sess=sess,
)
loss = zfr.predict_and_losses(
X_np = X_valid,
Y_np = Y_valid,
A_np = A_valid,
alpha_Z = alpha,
alpha_X = 0.,
alpha_Y = 1. - alpha,
)[1][0]
zfr_loss_by_step[run, step] = loss + 0.
zfr_losses[alpha, lambda_X, run] = loss + 0.
zfr_states[alpha, lambda_X, run] = [zfr.W_np + 0., zfr.V_np + 0.]
print now(), run, alpha, lambda_X, loss
pk.dump(
{
'zfr_losses': zfr_losses,
'zfr_states': zfr_states,
'zfr_loss_by_step': zfr_loss_by_step,
},
open('data/{:}_zfr_var_state_3.pk'.format(data_name), 'wb'),
)
print ''
sess.close()