-
Notifications
You must be signed in to change notification settings - Fork 0
/
run_sgd.py
292 lines (206 loc) · 10.4 KB
/
run_sgd.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
import numpy as np
import tensorflow as tf
import sys
import pickle as pkl
from models import Dense_2Layer, Big_Dense, Big_Conv_net, Small_Conv_net
import random
import os
from sklearn.metrics import balanced_accuracy_score, f1_score, roc_auc_score, accuracy_score
gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
try:
# Currently, memory growth needs to be the same across GPUs
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
logical_gpus = tf.config.experimental.list_logical_devices('GPU')
print(len(gpus), "Physical GPUs,", len(logical_gpus), "Logical GPUs")
except RuntimeError as e:
# Memory growth must be set before GPUs have been initialized
print(e)
# Reading Arguments
rand_seed = 1
tf.random.set_seed(rand_seed)
dataset = sys.argv[1]
name = sys.argv[2]
part = sys.argv[3]
swapped = sys.argv[4]
re_enter = sys.argv[5]
noisy = sys.argv[6]
save_logs = bool(int(sys.argv[7]))
per_sw = swapped
print(dataset, name, part, swapped, re_enter, noisy, save_logs)
# Loading Dataset
if dataset == "Ozone":
x_train = np.load(name + "/" + dataset + "/x_train_part_" + part + ".npy")
y_train = np.load(name + "/" + dataset + "/y_train_part_" + part + ".npy")
x_test = np.load(name + "/" + dataset + "/x_test_part_" + part + ".npy")
y_test = np.load(name + "/" + dataset + "/y_test_part_" + part + ".npy")
logs = "./logs/" + name + "/" + dataset + "/part_" + part + "/"
x_test = (x_test - x_train.mean(axis=0)) / (x_train.std(axis=0) + np.finfo(np.float32).eps)
x_train = (x_train - x_train.mean(axis=0)) / (x_train.std(axis=0) + np.finfo(np.float32).eps)
elif dataset == "adult":
x_train = np.load(name + "/" + dataset + "/log_x_train.npy")
y_train = np.load(name + "/" + dataset + "/log_y_train.npy")
x_test = np.load(name + "/" + dataset + "/log_x_test.npy")
y_test = np.load(name + "/" + dataset + "/log_y_test.npy")
logs = "./logs/" + name + "/" + dataset + "/"
elif dataset == "credit":
x_train = np.load(name + "/" + dataset + "/x_train_part_" + part + ".npy")
y_train = np.load(name + "/" + dataset + "/y_train_part_" + part + ".npy")
x_test = np.load(name + "/" + dataset + "/x_test_part_" + part + ".npy")
y_test = np.load(name + "/" + dataset + "/y_test_part_" + part + ".npy")
logs = "./logs/" + name + "/" + dataset + "/part_" + part + "/"
x_test = (x_test - x_train.mean(axis=0)) / (x_train.std(axis=0) + np.finfo(np.float32).eps)
x_train = (x_train - x_train.mean(axis=0)) / (x_train.std(axis=0) + np.finfo(np.float32).eps)
elif dataset == "mnist":
x_train = np.load(name + "/" + dataset + "/1_minor_0.1/x_train_" + part + ".npy")
y_train = np.load(name + "/" + dataset + "/1_minor_0.1/y_train_" + part + ".npy")
x_test = np.load(name + "/" + dataset + "/1_minor_0.1/x_test.npy")
y_test = np.load(name + "/" + dataset + "/1_minor_0.1/y_test.npy")
logs = "./logs/" + name + "/" + dataset + "/1_minor_0.1/" + part + "/"
x_test = x_test.reshape((x_test.shape[0], x_test.shape[1], x_test.shape[2], 1))
x_test = x_test / 255.0
x_train = x_train / 255.0
elif dataset == "cifar10":
x_train = np.load(name + "/" + dataset + "/1_minor_0.1/x_train_" + part + ".npy")
y_train = np.load(name + "/" + dataset + "/1_minor_0.1/y_train_" + part + ".npy")
x_test = np.load(name + "/" + dataset + "/1_minor_0.1/x_test.npy")
y_test = np.load(name + "/" + dataset + "/1_minor_0.1/y_test.npy")
logs = "./logs/" + name + "/" + dataset + "/1_minor_0.1/" + part + "/"
x_test = x_test / 255.0
x_train = x_train / 255.0
numcl = len(np.unique(y_train))
y_train = tf.keras.utils.to_categorical(y_train, num_classes=numcl)
y_test = tf.keras.utils.to_categorical(y_test, num_classes=numcl)
classes = y_train.shape[1]
# Initializing models
opt = tf.keras.optimizers.Adam()
if (dataset == "Ozone") or (dataset == "adult"):
epochs = 10
model = Dense_2Layer(x_train.shape[1], numcl)
model.compile(loss="categorical_crossentropy", optimizer=opt, metrics=["acc"])
elif (dataset == "credit") :
epochs = 10
model = Big_Dense(x_train.shape[1], numcl)
model.compile(loss="categorical_crossentropy", optimizer=opt, metrics=["acc"])
elif dataset == "mnist":
epochs = 15
model = Small_Conv_net()
model.compile(loss="categorical_crossentropy", optimizer=opt, metrics=["acc"])
elif dataset == "cifar10":
epochs = 35
model = Big_Conv_net()
model.compile(loss="categorical_crossentropy", optimizer=opt, metrics=["acc"])
N = x_train.shape[0]
indices = np.arange(N)
np.random.shuffle(indices)
if N > 64:
batch_size = 64
else:
batch_size = int(N / 2)
num_batches = x_train.shape[0] // batch_size
swapped = round(float(swapped) * N)
ma_param = 0.6
full_batch_losses = np.zeros(N)
EMA_batch_losses = np.zeros(N)
train_results = []
test_results = []
re_enter = (int(re_enter) == 1)
noisy = (int(noisy) == 1)
if re_enter:
sw_times = np.zeros(N)
if noisy:
var_noise = 0.2
# Start training
for ep in range(epochs):
print("Epoch: ", ep)
if (noisy & (ep > 0)):
noise_mask = np.round(np.random.uniform(0, 1, x_train.shape))
mask = np.in1d(indices, batch_low_swap).astype(int)
noise = np.random.normal(0, var_noise, x_train.shape)
ax = list(range(len(x_train.shape)))[1:]
x_train_temp = x_train[indices] + np.expand_dims(mask, axis=ax) * x_train[indices] * noise_mask * noise
y_train_temp = y_train[indices]
else:
x_train_temp = x_train[indices]
y_train_temp = y_train[indices]
for b in range(num_batches+1):
if b == num_batches:
batch_x = x_train_temp[b * batch_size :]
batch_y = y_train_temp[b * batch_size :]
else:
batch_x = x_train_temp[b * batch_size : (b+1) * batch_size]
batch_y = y_train_temp[b * batch_size : (b+1) * batch_size]
with tf.GradientTape() as tape:
preds = model(batch_x, training=True)
losses = tf.keras.losses.categorical_crossentropy(batch_y, preds)
# print(losses.numpy().mean())
grads = tape.gradient(losses, model.trainable_variables)
opt.apply_gradients(zip(grads, model.trainable_variables))
if b == num_batches:
full_batch_losses[indices[b * batch_size : ]] = losses
EMA_batch_losses[indices[b * batch_size : ]] = ma_param * EMA_batch_losses[indices[b * batch_size : ]] + (1 - ma_param) * losses
else:
full_batch_losses[indices[b * batch_size : (b+1) * batch_size]] = losses
EMA_batch_losses[indices[b * batch_size : (b+1) * batch_size]] = ma_param * EMA_batch_losses[indices[b * batch_size : (b+1) * batch_size]] + (1 - ma_param) * losses
# --------------------------------TRAIN SCORES---------------------------------------------
tr_preds = model.predict(x_train)
# [los, s] = model.evaluate(x_train, y_train)
los = np.mean(tf.keras.losses.categorical_crossentropy(y_train, tr_preds))
s = accuracy_score(y_train.argmax(axis=1), tr_preds.argmax(axis=1))
f1_macro = f1_score(y_train.argmax(axis=1), tr_preds.argmax(axis=1), average='macro')
f1_micro = f1_score(y_train.argmax(axis=1), tr_preds.argmax(axis=1), average='micro')
balacc = balanced_accuracy_score(y_train.argmax(axis=1), tr_preds.argmax(axis=1))
rocauc_macro = roc_auc_score(y_train, tr_preds, average='macro', multi_class="ovr")
rocauc_micro = roc_auc_score(y_train, tr_preds, average='micro', multi_class="ovr")
print(" Train Loss: ", los)
print(" Train Accuracy: ", s)
print(" Train F1: ", f1_macro, ", ", f1_micro)
print(" Train Balanced Acc: ", balacc)
print(" Train ROC AUC: ", rocauc_macro, ", ", rocauc_micro)
train_results.append([los, s, f1_macro, f1_micro, balacc, rocauc_macro, rocauc_micro])
# ---------------------------------TEST SCORES-----------------------------------------------
tst_preds = model.predict(x_test)
# [los, s] = model.evaluate(x_test, y_test)
los = np.mean(tf.keras.losses.categorical_crossentropy(y_test, tst_preds))
s = accuracy_score(y_test.argmax(axis=1), tst_preds.argmax(axis=1))
f1_macro2 = f1_score(y_test.argmax(axis=1), tst_preds.argmax(axis=1), average='macro')
f1_micro2 = f1_score(y_test.argmax(axis=1), tst_preds.argmax(axis=1), average='micro')
balacc2 = balanced_accuracy_score(y_test.argmax(axis=1), tst_preds.argmax(axis=1))
rocauc_macro2 = roc_auc_score(y_test, tst_preds, average='macro', multi_class="ovr")
rocauc_micro2 = roc_auc_score(y_test, tst_preds, average='micro', multi_class="ovr")
print(" Test Loss: ", los)
print(" Test Accuracy: ", s)
print(" Test F1: ", f1_macro2, ", ", f1_micro2)
print(" Test Balanced Acc: ", balacc2)
print(" Test ROC AUC: ", rocauc_macro2, ", ", rocauc_micro2)
test_results.append([los, s, f1_macro2, f1_micro2, balacc2, rocauc_macro2, rocauc_micro2])
# ------------------------------------------------------------------------------------
#------------------Swapping--------------------------------------
if swapped != 0:
print("Swapping ", swapped, " samples.")
if re_enter:
# sw_thr = % of epochs
re_inds = sw_times >= int(0.2 * epochs)
mean_loss = full_batch_losses.mean()
full_batch_losses[re_inds] = mean_loss
ind_batch_low = np.argpartition(full_batch_losses, swapped)
# ind_ma_low = np.argpartition(EMA_batch_losses, swapped)
# ind_batch_high = np.argpartition(full_batch_losses, N - swapped)
ind_ma_high = np.argpartition(EMA_batch_losses, N - swapped)
batch_low_swap = ind_batch_low[swapped:]
ma_high_swap = ind_ma_high[-swapped:]
indices = np.concatenate((batch_low_swap, ma_high_swap))
if re_enter:
swapped_ind = ind_batch_low[:swapped]
sw_times[indices] = 0
sw_times[swapped_ind] += 1
np.random.shuffle(indices)
train_results = np.array(train_results)
test_results = np.array(test_results)
# Saving results
if save_logs:
if not os.path.exists(logs):
os.makedirs(logs)
np.save(logs + "train_results_part" + "_" + str(per_sw) + "_" + str(re_enter) + "_" + str(noisy), train_results)
np.save(logs + "test_results__part" + "_" + str(per_sw) + "_" + str(re_enter) + "_" + str(noisy), test_results)