-
Notifications
You must be signed in to change notification settings - Fork 0
/
SMS_MXNet_script.py
191 lines (146 loc) · 6.54 KB
/
SMS_MXNet_script.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
from __future__ import print_function
import logging
import mxnet as mx
from mxnet import gluon, autograd
from mxnet.gluon import nn
import numpy as np
import json
import time
import pip
try:
from pip import main as pipmain
except:
from pip._internal import main as pipmain
pipmain(['install', 'pandas'])
import pandas
logging.basicConfig(level=logging.DEBUG)
# ------------------------------------------------------------ #
# Training methods #
# ------------------------------------------------------------ #
def train(hyperparameters, input_data_config, channel_input_dirs, output_data_dir,
num_gpus, num_cpus, hosts, current_host, **kwargs):
# SageMaker passes num_cpus, num_gpus and other args we can use to tailor training to
# the current container environment, but here we just use simple cpu context.
ctx = mx.cpu()
# retrieve the hyperparameters we set in notebook (with some defaults)
batch_size = hyperparameters.get('batch_size', 200)
epochs = hyperparameters.get('epochs', 10)
learning_rate = hyperparameters.get('learning_rate', 0.01)
momentum = hyperparameters.get('momentum', 0.9)
log_interval = hyperparameters.get('log_interval', 200)
train_data_path = channel_input_dirs['train']
val_data_path = channel_input_dirs['val']
train_data = get_train_data(train_data_path, batch_size)
val_data = get_val_data(val_data_path, batch_size)
# define the network
net = define_network()
# Collect all parameters from net and its children, then initialize them.
net.initialize(mx.init.Normal(sigma=1.), ctx=ctx)
# Trainer is for updating parameters with gradient.
if len(hosts) == 1:
kvstore = 'device' if num_gpus > 0 else 'local'
else:
kvstore = 'dist_device_sync' if num_gpus > 0 else 'dist_sync'
trainer = gluon.Trainer(net.collect_params(), 'sgd',
{'learning_rate': learning_rate, 'momentum': momentum},
kvstore=kvstore)
metric = mx.metric.Accuracy()
loss = gluon.loss.SoftmaxCrossEntropyLoss()
for epoch in range(epochs):
# reset data iterator and metric at begining of epoch.
metric.reset()
btic = time.time()
for i, (data, label) in enumerate(train_data):
# Copy data to ctx if necessary
data = data.as_in_context(ctx)
label = label.as_in_context(ctx)
# Start recording computation graph with record() section.
# Recorded graphs can then be differentiated with backward.
with autograd.record():
output = net(data)
L = loss(output, label)
L.backward()
# take a gradient step with batch_size equal to data.shape[0]
trainer.step(data.shape[0])
# update metric at last.
metric.update([label], [output])
if i % log_interval == 0 and i > 0:
name, acc = metric.get()
print('[Epoch %d Batch %d] Training: %s=%f, %f samples/s' %
(epoch, i, name, acc, batch_size / (time.time() - btic)))
btic = time.time()
name, acc = metric.get()
print('[Epoch %d] Training: %s=%f' % (epoch, name, acc))
name, val_acc = test(ctx, net, val_data)
print('[Epoch %d] Validation: %s=%f' % (epoch, name, val_acc))
return net
def save(net, model_dir):
# save the model
y = net(mx.sym.var('data'))
y.save('%s/model.json' % model_dir)
net.collect_params().save('%s/model.params' % model_dir)
def define_network():
net = nn.Sequential()
with net.name_scope():
net.add(nn.Dense(16, activation='relu'))
net.add(nn.Dense(16, activation='relu'))
net.add(nn.Dense(2))
#net.add(nn.Dense(1))
return net
def get_train_data(data_path, batch_size):
print('Train data path: ' + data_path)
df = pandas.read_csv(data_path + '/sms_train_set.gz')
features = df[df.columns[1:]].values.astype(dtype=np.float32)
labels = df[df.columns[0]].values.reshape((-1, 1)).astype(dtype=np.float32)
return gluon.data.DataLoader(gluon.data.ArrayDataset(features, labels), batch_size=batch_size, shuffle=True)
def get_val_data(data_path, batch_size):
print('Validation data path: ' + data_path)
df = pandas.read_csv(data_path + '/sms_test_set.gz')
features = df[df.columns[1:]].values.astype(dtype=np.float32)
labels = df[df.columns[0]].values.reshape((-1, 1)).astype(dtype=np.float32)
return gluon.data.DataLoader(gluon.data.ArrayDataset(features, labels), batch_size=batch_size, shuffle=False)
def test(ctx, net, val_data):
metric = mx.metric.Accuracy()
for data, label in val_data:
data = data.as_in_context(ctx)
label = label.as_in_context(ctx)
output = net(data)
metric.update([label], [output])
return metric.get()
# ------------------------------------------------------------ #
# Hosting methods #
# ------------------------------------------------------------ #
def model_fn(model_dir):
"""
Load the gluon model. Called once when hosting service starts.
:param: model_dir The directory where model files are stored.
:return: a model (in this case a Gluon network)
"""
symbol = mx.sym.load('%s/model.json' % model_dir)
outputs = mx.symbol.softmax(data=symbol, name='softmax_label')
inputs = mx.sym.var('data')
param_dict = gluon.ParameterDict('model_')
net = gluon.SymbolBlock(outputs, inputs, param_dict)
net.load_params('%s/model.params' % model_dir, ctx=mx.cpu())
return net
def transform_fn(net, data, input_content_type, output_content_type):
"""
Transform a request using the Gluon model. Called once per request.
:param net: The Gluon model.
:param data: The request payload.
:param input_content_type: The request content type.
:param output_content_type: The (desired) response content type.
:return: response payload and content type.
"""
# we can use content types to vary input/output handling, but
# here we just assume json for both
try:
parsed = json.loads(data)
nda = mx.nd.array(parsed)
output = net(nda)
#prediction = mx.nd.argmax(output, axis=1)
response_body = json.dumps(output.asnumpy().tolist())
return response_body, output_content_type
except Exception as ex:
response_body = '{error: }' + str(ex)
return response_body, output_content_type