-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathshakespeare_fedavg_main.py
205 lines (155 loc) · 7.81 KB
/
shakespeare_fedavg_main.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
import collections
import functools
from absl import app
from absl import flags
import numpy as np
import tensorflow as tf
import tensorflow_federated as tff
import pickle
import random
import os
from tensorflow.compat.v1.keras import backend as Kbackbend
import simple_fedavg_tf
import simple_fedavg_tff
from tensorflow_federated.python.simulation import hdf5_client_data
# Training hyperparameters
flags.DEFINE_integer('total_rounds', 151, 'Number of total training rounds.')
flags.DEFINE_integer('rounds_per_eval', 5, 'How often to evaluate')
flags.DEFINE_integer('train_clients_per_round', 32,
'How many clients to sample per round.')
flags.DEFINE_integer('expected_clients_per_round', 2,
'How many clients to communicate per round.')
flags.DEFINE_integer('j_max_iter_greedy_alg', 4,
'Maximum number of iteration of greedy algorithm.')
flags.DEFINE_integer('client_epochs_per_round', 1,
'Number of epochs in the client to take per round.')
flags.DEFINE_integer('batch_size', 8, 'Batch size used on the client and also for test data.')
flags.DEFINE_bool('importance_sampling', True, 'Importance sampling is used.')
flags.DEFINE_string('name', 'shakespeare', 'Name of the experiment.')
flags.DEFINE_integer('random_seed', 1, 'Random seed that should be used for client sampling.')
flags.DEFINE_integer('seq_len', 5, 'Sequence length.')
# Optimizer configuration (this defines one or more flags per optimizer).
flags.DEFINE_float('server_learning_rate', 1.0, 'Server learning rate.')
flags.DEFINE_float('client_learning_rate', 0.1, 'Client learning rate.')
FLAGS = flags.FLAGS
# A fixed vocabularly of ASCII chars that occur in the works of Shakespeare and Dickens:
vocab = list('dhlptx@DHLPTX $(,048cgkoswCGKOSW[_#\'/37;?bfjnrvzBFJNRVZ"&*.26:\naeimquyAEIMQUY]!%)-159\r')
# Creating a mapping from unique characters to indices
char2idx = {u:i for i, u in enumerate(vocab)}
idx2char = np.array(vocab)
class FlattenedCategoricalAccuracy(tf.keras.metrics.SparseCategoricalAccuracy):
def __init__(self, name='accuracy', dtype=tf.float32):
super().__init__(name, dtype=dtype)
def update_state(self, y_true, y_pred, sample_weight=None):
y_true = tf.reshape(y_true, [-1, 1])
y_pred = tf.reshape(y_pred, [-1, len(vocab), 1])
return super().update_state(y_true, y_pred, sample_weight)
def get_shakespeare_dataset():
"""Loads and preprocesses the Shakespeare dataset.
Returns:
A `(ss_train, ss_test)` tuple where `ss_train` is a
`tff.simulation.ClientData` object representing the training data and
`ss_test` is a single `tf.data.Dataset` representing the test data of
all clients.
"""
table = tf.lookup.StaticHashTable(
tf.lookup.KeyValueTensorInitializer(
keys=vocab, values=tf.constant(list(range(len(vocab))), dtype=tf.int64)
),
default_value=0)
def to_ids(x):
s = tf.reshape(x['snippets'], shape=[1])
chars = tf.strings.bytes_split(s).values
ids = table.lookup(chars)
return ids
def split_input_target(chunk):
input_text = tf.map_fn(lambda x: x[:-1], chunk)
target_text = tf.map_fn(lambda x: x[1:], chunk)
return collections.OrderedDict(
x=input_text, y=target_text)
def preprocess_train_dataset(dataset):
return dataset.map(to_ids).unbatch().batch(FLAGS.seq_len+1, drop_remainder=True).shuffle(buffer_size=100).repeat(
count=FLAGS.client_epochs_per_round).batch(
FLAGS.batch_size, drop_remainder=True).map(split_input_target)
def preprocess_test_dataset(dataset):
return dataset.map(to_ids).unbatch().batch(FLAGS.seq_len+1, drop_remainder=True).batch(
FLAGS.batch_size, drop_remainder=True).map(split_input_target)
ss_train, ss_test = tff.simulation.datasets.shakespeare.load_data()
ss_train = ss_train.preprocess(preprocess_train_dataset)
ss_test = preprocess_test_dataset(
ss_test.create_tf_dataset_from_all_clients())
return ss_train, ss_test
def load_model(batch_size):
model = tf.keras.models.Sequential()
model.add(tf.keras.layers.Embedding(input_dim=86, output_dim=256))
model.add(tf.keras.layers.GRU(units=256, return_sequences=True, batch_input_shape=(batch_size, None, 256)))
model.add(tf.keras.layers.GRU(units=256, return_sequences=True))
model.add(tf.keras.layers.Dense(86))
return model
def server_optimizer_fn():
return tf.keras.optimizers.SGD(learning_rate=FLAGS.server_learning_rate)
def client_optimizer_fn():
return tf.keras.optimizers.SGD(learning_rate=FLAGS.client_learning_rate)
def main(argv):
if len(argv) > 1:
raise app.UsageError('Too many command-line arguments.')
os.environ['PYTHONHASHSEED'] = str(FLAGS.random_seed)
np.random.seed(FLAGS.random_seed)
random.seed(FLAGS.random_seed)
tf.random.set_seed(FLAGS.random_seed)
session_conf = tf.compat.v1.ConfigProto(intra_op_parallelism_threads=1, inter_op_parallelism_threads=1)
sess = tf.compat.v1.Session(graph=tf.compat.v1.get_default_graph(), config=session_conf)
Kbackbend.set_session(sess)
train_data, test_data = get_shakespeare_dataset()
def tff_model_fn():
"""Constructs a fully initialized model for use in federated averaging."""
keras_model = load_model(FLAGS.batch_size)
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
return simple_fedavg_tf.KerasModelWrapper(keras_model,
test_data.element_spec, loss)
federated_averaging = simple_fedavg_tff.FedAvg(
tff_model_fn, FLAGS.expected_clients_per_round, FLAGS.train_clients_per_round,
FLAGS.j_max_iter_greedy_alg, FLAGS.importance_sampling, server_optimizer_fn, client_optimizer_fn)
server_state = federated_averaging.initialize()
metric_acc = FlattenedCategoricalAccuracy(name='test_accuracy')
model = tff_model_fn()
train_loss = []
val_acc = []
probs = []
clients = train_data.client_ids
print('preprocessing...')
exclude_clients = []
for i, client in enumerate(clients):
client_dataset = train_data.create_tf_dataset_for_client(client)
len_dataset = len(list(client_dataset))
if len_dataset == 0:
exclude_clients.append(i) # exclude empty clients
np.random.seed(FLAGS.random_seed)
sampled_clients_list = [np.random.choice(
[i for i in range(len(clients)) if i not in exclude_clients],
size=FLAGS.train_clients_per_round,
replace=False) for _ in range(FLAGS.total_rounds)]
for round_num in range(FLAGS.total_rounds):
sampled_clients = [clients[i] for i in sampled_clients_list[round_num]]
sampled_train_data = []
for client in sampled_clients:
client_dataset = train_data.create_tf_dataset_for_client(client)
sampled_train_data.append(client_dataset)
server_state, train_metrics, prob = federated_averaging.next(
server_state, sampled_train_data)
print(f'Round {round_num} training loss: {train_metrics}')
train_loss.append(train_metrics)
probs.append(prob)
if round_num % FLAGS.rounds_per_eval == 0:
model.from_weights(server_state.model_weights)
accuracy = simple_fedavg_tf.keras_evaluate(model.keras_model, test_data, metric_acc)
print(f'Round {round_num} validation accuracy: {accuracy * 100.0}')
val_acc.append(accuracy)
with open(f'./tff_save/{FLAGS.name}_valacc.pk', 'wb') as f:
pickle.dump(val_acc, f)
with open(f'./tff_save/{FLAGS.name}_trainloss.pk', 'wb') as f:
pickle.dump(train_loss, f)
with open(f'./tff_save/{FLAGS.name}_probs.pk', 'wb') as f:
pickle.dump(probs, f)
if __name__ == '__main__':
app.run(main)