forked from sunprinceS/ML-Assignment3
-
Notifications
You must be signed in to change notification settings - Fork 2
/
p1.html
310 lines (253 loc) · 22.4 KB
/
p1.html
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
<!DOCTYPE html>
<html>
<link rel="shortcut icon" href="favicon.ico">
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/4.7.0/css/font-awesome.min.css">
<link rel="stylesheet" href="highlight.css">
<meta name="author" content="ntumlta" >
<meta property="og:image" content="joy.png"/>
<title>Machine Learning (2017, Fall)</title>
<xmp theme="cerulean" style="display:none;">
# Problem 1: Build Convolution Neural Network
Problem Description:
* tune performance
* record model structure
* record training procedure
## 範例
**[Note] 請不要直接使用助教的圖來當成作業交上來**
#### Model Structure
<center>Keras vis_utils</center>| <center> Tensorboard Graph</center>
:-------------------------:|:-------------------------:
<img src="http://i.imgur.com/XYkDQvY.png" alt="Drawing" style="width: 350px;"/> | <img src="http://i.imgur.com/Ucdf0ys.png" alt="Drawing" style="width: 1400px;"/>
#### Training Procedure
<img src="http://i.imgur.com/7w3WnTV.png" alt="Drawing" style="width: 700px;"/>
## TA hour
以下會以 **Keras** 為使用的套件來做介紹
<i class="fa fa-diamond"></i> Keywords: `keras.models`, `keras.layers`, `keras.callbacks`, `keras.utils.visualize_util.plot`, `Tensorboard`
<div id="doc" class="markdown-body container-fluid" style="position: relative;"><h1 id="hw3-手把手-q1"><a class="anchor hidden-xs" href="#hw3-手把手-q1" title="hw3-手把手-q1"><span class="octicon octicon-link"></span></a>HW3 手把手 Q1</h1><h2 id="建立模型"><a class="anchor hidden-xs" href="#建立模型" title="建立模型"><span class="octicon octicon-link"></span></a>建立模型</h2><pre><code class="python=3.6 hljs"><span class="hljs-comment">#!/usr/bin/env python</span>
<span class="hljs-comment"># -- coding: utf-8 --</span>
<span class="hljs-keyword">from</span> keras.models <span class="hljs-keyword">import</span> Model, load_model
<span class="hljs-keyword">from</span> keras.layers <span class="hljs-keyword">import</span> Input, Dense, Dropout, Flatten, Activation, Reshape
<span class="hljs-keyword">from</span> keras.layers.convolutional <span class="hljs-keyword">import</span> Conv2D, ZeroPadding2D
<span class="hljs-keyword">from</span> keras.layers.pooling <span class="hljs-keyword">import</span> MaxPooling2D, AveragePooling2D
<span class="hljs-keyword">from</span> keras.optimizers <span class="hljs-keyword">import</span> SGD, Adam, Adadelta
<span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">build_model</span><span class="hljs-params">()</span>:</span>
<span class="hljs-string">'''
#先定義好框架
#第一步從input吃起
'''</span>
input_img = Input(shape=(<span class="hljs-number">48</span>, <span class="hljs-number">48</span>, <span class="hljs-number">1</span>))
<span class="hljs-string">'''
先來看一下keras document 的Conv2D
keras.layers.Conv2D(filters, kernel_size, strides=(1, 1),
padding='valid', data_format=None, dilation_rate=(1, 1),
activation=None, use_bias=True, kernel_initializer='glorot_uniform',
bias_initializer='zeros', kernel_regularizer=None,
bias_regularizer=None, activity_regularizer=None,
kernel_constraint=None, bias_constraint=None)
'''</span>
block1 = Conv2D(<span class="hljs-number">64</span>, (<span class="hljs-number">5</span>, <span class="hljs-number">5</span>), padding=<span class="hljs-string">'valid'</span>, activation=<span class="hljs-string">'relu'</span>)(input_img)
block1 = ZeroPadding2D(padding=(<span class="hljs-number">2</span>, <span class="hljs-number">2</span>), data_format=<span class="hljs-string">'channels_last'</span>)(block1)
block1 = MaxPooling2D(pool_size=(<span class="hljs-number">5</span>, <span class="hljs-number">5</span>), strides=(<span class="hljs-number">2</span>, <span class="hljs-number">2</span>))(block1)
block1 = ZeroPadding2D(padding=(<span class="hljs-number">1</span>, <span class="hljs-number">1</span>), data_format=<span class="hljs-string">'channels_last'</span>)(block1)
block2 = Conv2D(<span class="hljs-number">64</span>, (<span class="hljs-number">3</span>, <span class="hljs-number">3</span>), activation=<span class="hljs-string">'relu'</span>)(block1)
block2 = ZeroPadding2D(padding=(<span class="hljs-number">1</span>, <span class="hljs-number">1</span>), data_format=<span class="hljs-string">'channels_last'</span>)(block2)
block3 = Conv2D(<span class="hljs-number">64</span>, (<span class="hljs-number">3</span>, <span class="hljs-number">3</span>), activation=<span class="hljs-string">'relu'</span>)(block2)
block3 = AveragePooling2D(pool_size=(<span class="hljs-number">3</span>, <span class="hljs-number">3</span>), strides=(<span class="hljs-number">2</span>, <span class="hljs-number">2</span>))(block3)
block3 = ZeroPadding2D(padding=(<span class="hljs-number">1</span>, <span class="hljs-number">1</span>), data_format=<span class="hljs-string">'channels_last'</span>)(block3)
block4 = Conv2D(<span class="hljs-number">128</span>, (<span class="hljs-number">3</span>, <span class="hljs-number">3</span>), activation=<span class="hljs-string">'relu'</span>)(block3)
block4 = ZeroPadding2D(padding=(<span class="hljs-number">1</span>, <span class="hljs-number">1</span>), data_format=<span class="hljs-string">'channels_last'</span>)(block4)
block5 = Conv2D(<span class="hljs-number">128</span>, (<span class="hljs-number">3</span>, <span class="hljs-number">3</span>), activation=<span class="hljs-string">'relu'</span>)(block4)
block5 = ZeroPadding2D(padding=(<span class="hljs-number">1</span>, <span class="hljs-number">1</span>), data_format=<span class="hljs-string">'channels_last'</span>)(block5)
block5 = AveragePooling2D(pool_size=(<span class="hljs-number">3</span>, <span class="hljs-number">3</span>), strides=(<span class="hljs-number">2</span>, <span class="hljs-number">2</span>))(block5)
block5 = Flatten()(block5)
fc1 = Dense(<span class="hljs-number">1024</span>, activation=<span class="hljs-string">'relu'</span>)(block5)
fc1 = Dropout(<span class="hljs-number">0.5</span>)(fc1)
fc2 = Dense(<span class="hljs-number">1024</span>, activation=<span class="hljs-string">'relu'</span>)(fc1)
fc2 = Dropout(<span class="hljs-number">0.5</span>)(fc2)
predict = Dense(<span class="hljs-number">7</span>)(fc2)
predict = Activation(<span class="hljs-string">'softmax'</span>)(predict)
model = Model(inputs=input_img, outputs=predict)
<span class="hljs-comment"># opt = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)</span>
<span class="hljs-comment"># opt = Adam(lr=1e-3)</span>
opt = Adadelta(lr=<span class="hljs-number">0.1</span>, rho=<span class="hljs-number">0.95</span>, epsilon=<span class="hljs-number">1e-08</span>)
model.compile(loss=<span class="hljs-string">'categorical_crossentropy'</span>, optimizer=opt, metrics=[<span class="hljs-string">'accuracy'</span>])
model.summary()
<span class="hljs-keyword">return</span> model
</code></pre><hr><h2 id="training-file"><a class="anchor hidden-xs" href="#training-file" title="training-file"><span class="octicon octicon-link"></span></a>Training File</h2><h5 id="不同於平常網路上的keras教學，我們使用比較多manual的loop-code來讓同學比較能瞭解epoch-“batch”-…的概念是什麼"><a class="anchor hidden-xs" href="#不同於平常網路上的keras教學，我們使用比較多manual的loop-code來讓同學比較能瞭解epoch-“batch”-…的概念是什麼" title="不同於平常網路上的keras教學，我們使用比較多manual的loop-code來讓同學比較能瞭解epoch-“batch”-…的概念是什麼"><span class="octicon octicon-link"></span></a>不同於平常網路上的keras教學，我們使用比較多manual的loop code來讓同學比較能瞭解"epoch" “batch” …的概念是什麼</h5><pre><code class="python=3.6 hljs"><span class="hljs-comment"># This code is using tensorflow backend</span>
<span class="hljs-comment">#!/usr/bin/env python</span>
<span class="hljs-comment"># -- coding: utf-8 --</span>
<span class="hljs-keyword">from</span> cnn_simple <span class="hljs-keyword">import</span> *
<span class="hljs-keyword">from</span> utils <span class="hljs-keyword">import</span> *
<span class="hljs-keyword">import</span> os
<span class="hljs-keyword">import</span> numpy <span class="hljs-keyword">as</span> np
<span class="hljs-keyword">import</span> argparse
<span class="hljs-keyword">import</span> time
os.system(<span class="hljs-string">'echo $CUDA_VISIBLE_DEVICES'</span>)
PATIENCE = <span class="hljs-number">5</span> <span class="hljs-comment"># The parameter is used for early stopping</span>
<span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">main</span><span class="hljs-params">()</span>:</span>
parser = argparse.ArgumentParser(prog=<span class="hljs-string">'train.py'</span>)
parser.add_argument(<span class="hljs-string">'--epoch'</span>, type=int, default=<span class="hljs-number">1</span>)
parser.add_argument(<span class="hljs-string">'--batch'</span>, type=int, default=<span class="hljs-number">64</span>)
parser.add_argument(<span class="hljs-string">'--pretrain'</span>, type=bool, default=<span class="hljs-keyword">False</span>)
parser.add_argument(<span class="hljs-string">'--save_every'</span>, type=int, default=<span class="hljs-number">1</span>)
parser.add_argument(<span class="hljs-string">'--model_name'</span>, type=str, default=<span class="hljs-string">'model/model-1'</span>)
args = parser.parse_args()
<span class="hljs-string">'''
To begin with, you should first read your csv training file and
cut them into training set and validation set.
Such as:
with open(csvFile, 'r') as f:
f.readline()
for i, line in enumerate(f):
data = line.split(',')
label = data[0]
pixel = data[1]
...
...
In addition, we maintain it in array structure and save it in pickle
'''</span>
<span class="hljs-comment"># training data</span>
train_pixels = load_pickle(<span class="hljs-string">'../train_pixels.pkl'</span>)
train_labels = load_pickle(<span class="hljs-string">'../train_labels.pkl'</span>)
<span class="hljs-keyword">print</span> (<span class="hljs-string">'# of training instances: '</span> + str(len(train_labels)))
<span class="hljs-comment"># validation data</span>
valid_pixels = load_pickle(<span class="hljs-string">'../valid_pixels.pkl'</span>)
valid_labels = load_pickle(<span class="hljs-string">'../valid_labels.pkl'</span>)
<span class="hljs-keyword">print</span> (<span class="hljs-string">'# of validation instances: '</span> + str(len(valid_labels)))
<span class="hljs-string">'''
Modify the answer format so as to correspond with the output of keras model
We can also do this to training data here,
but we choose to do it in "train" function
'''</span>
<span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> range(len(valid_labels)):
valid_pixels[i] = np.fromstring(valid_pixels[i], dtype=float, sep=<span class="hljs-string">' '</span>).reshape((<span class="hljs-number">48</span>, <span class="hljs-number">48</span>, <span class="hljs-number">1</span>))
onehot = np.zeros((<span class="hljs-number">7</span>, ), dtype=np.float)
onehot[int(valid_labels[i])] = <span class="hljs-number">1.</span>
valid_labels[i] = onehot
<span class="hljs-comment"># start training</span>
train(args.batch, args.epoch, args.pretrain, args.save_every,
train_pixels, train_labels,
np.asarray(valid_pixels), np.asarray(valid_labels),
args.model_name)
<span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">train</span><span class="hljs-params">(batch_size, num_epoch, pretrain, save_every, train_pixels, train_labels, val_pixels, val_labels, model_name=None)</span>:</span>
<span class="hljs-keyword">if</span> pretrain == <span class="hljs-keyword">False</span>:
model = build_model()
<span class="hljs-keyword">else</span>:
model = load_model(model_name)
<span class="hljs-string">'''
"1 Epoch" means you have been looked all of the training data once already.
Batch size B means you look B instances at once when updating your parameter.
Thus, given 320 instances, batch size 32, you need 10 iterations in 1 epoch.
'''</span>
num_instances = len(train_labels)
iter_per_epoch = int(num_instances / batch_size) + <span class="hljs-number">1</span>
batch_cutoff = [<span class="hljs-number">0</span>]
<span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> range(iter_per_epoch - <span class="hljs-number">1</span>):
batch_cutoff.append(batch_size * (i+<span class="hljs-number">1</span>))
batch_cutoff.append(num_instances)
total_start_t = time.time()
best_metrics = <span class="hljs-number">0.0</span>
early_stop_counter = <span class="hljs-number">0</span>
<span class="hljs-keyword">for</span> e <span class="hljs-keyword">in</span> range(num_epoch):
<span class="hljs-comment">#shuffle data in every epoch</span>
rand_idxs = np.random.permutation(num_instances)
<span class="hljs-keyword">print</span> (<span class="hljs-string">'#######'</span>)
<span class="hljs-keyword">print</span> (<span class="hljs-string">'Epoch '</span> + str(e+<span class="hljs-number">1</span>))
<span class="hljs-keyword">print</span> (<span class="hljs-string">'#######'</span>)
start_t = time.time()
<span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> range(iter_per_epoch):
<span class="hljs-keyword">if</span> i % <span class="hljs-number">50</span> == <span class="hljs-number">0</span>:
<span class="hljs-keyword">print</span> (<span class="hljs-string">'Iteration '</span> + str(i+<span class="hljs-number">1</span>))
X_batch = []
Y_batch = []
<span class="hljs-string">''' fill data into each batch '''</span>
<span class="hljs-keyword">for</span> n <span class="hljs-keyword">in</span> range(batch_cutoff[i], batch_cutoff[i+<span class="hljs-number">1</span>]):
X_batch.append(train_pixels[rand_idxs[n]])
Y_batch.append(np.zeros((<span class="hljs-number">7</span>, ), dtype=np.float))
X_batch[<span class="hljs-number">-1</span>] = np.fromstring(X_batch[<span class="hljs-number">-1</span>], dtype=float, sep=<span class="hljs-string">' '</span>).reshape((<span class="hljs-number">48</span>, <span class="hljs-number">48</span>, <span class="hljs-number">1</span>))
Y_batch[<span class="hljs-number">-1</span>][int(train_labels[rand_idxs[n]])] = <span class="hljs-number">1.</span>
<span class="hljs-string">''' use these batch data to train your model '''</span>
model.train_on_batch(np.asarray(X_batch),np.asarray(Y_batch))
<span class="hljs-string">'''
The above process is one epoch, and then we can check the performance now.
'''</span>
loss_and_metrics = model.evaluate(val_pixels, val_labels, batch_size)
<span class="hljs-keyword">print</span> (<span class="hljs-string">'\nloss & metrics:'</span>)
<span class="hljs-keyword">print</span> (loss_and_metrics)
<span class="hljs-string">'''
early stop is a mechanism to prevent your model from overfitting
'''</span>
<span class="hljs-keyword">if</span> loss_and_metrics[<span class="hljs-number">1</span>] >= best_metrics:
best_metrics = loss_and_metrics[<span class="hljs-number">1</span>]
<span class="hljs-keyword">print</span> (<span class="hljs-string">"save best score!! "</span>+str(loss_and_metrics[<span class="hljs-number">1</span>]))
early_stop_counter = <span class="hljs-number">0</span>
<span class="hljs-keyword">else</span>:
early_stop_counter += <span class="hljs-number">1</span>
<span class="hljs-string">'''
Sample code to write result :
if e == e:
val_proba = model.predict(val_pixels)
val_classes = val_proba.argmax(axis=-1)
with open('result/simple_%s.csv' % str(e), 'w') as f:
f.write('acc = %s\n' % str(loss_and_metrics[1]))
f.write('id,label')
for i in range(len(val_classes)):
f.write('\n' + str(i) + ',' + str(val_classes[i]))
'''</span>
<span class="hljs-keyword">print</span> (<span class="hljs-string">'Elapsed time in epoch '</span> + str(e+<span class="hljs-number">1</span>) + <span class="hljs-string">': '</span> + str(time.time() - start_t))
<span class="hljs-keyword">if</span> (e+<span class="hljs-number">1</span>) % save_every == <span class="hljs-number">0</span>:
model.save(<span class="hljs-string">'model/model-%d.h5'</span> %(e+<span class="hljs-number">1</span>))
<span class="hljs-keyword">print</span> (<span class="hljs-string">'Saved model %s!'</span> %str(e+<span class="hljs-number">1</span>))
<span class="hljs-keyword">if</span> early_stop_counter >= PATIENCE:
<span class="hljs-keyword">print</span> (<span class="hljs-string">'Stop by early stopping'</span>)
<span class="hljs-keyword">print</span> (<span class="hljs-string">'Best score: '</span>+str(best_metrics))
<span class="hljs-keyword">break</span>
<span class="hljs-keyword">print</span> (<span class="hljs-string">'Elapsed time in total: '</span> + str(time.time() - total_start_t))
<span class="hljs-keyword">if</span> __name__==<span class="hljs-string">'__main__'</span>:
main()
</code></pre><hr><h2 id="plot-model"><a class="anchor hidden-xs" href="#plot-model" title="plot-model"><span class="octicon octicon-link"></span></a>Plot model</h2><pre><code class="python hljs"><span class="hljs-comment">#!/usr/bin/env python</span>
<span class="hljs-comment"># -- coding: utf-8 --</span>
<span class="hljs-keyword">import</span> argparse
<span class="hljs-keyword">from</span> keras.utils.vis_utils <span class="hljs-keyword">import</span> plot_model
<span class="hljs-keyword">from</span> keras.models <span class="hljs-keyword">import</span> load_model
<span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">main</span><span class="hljs-params">()</span>:</span>
parser = argparse.ArgumentParser(prog=<span class="hljs-string">'plot_model.py'</span>,
description=<span class="hljs-string">'Plot the model.'</span>)
parser.add_argument(<span class="hljs-string">'--model'</span>,type=str,default=<span class="hljs-string">'model/model-7.h5'</span>)
args = parser.parse_args()
emotion_classifier = load_model(args.model)
emotion_classifier.summary()
plot_model(emotion_classifier,to_file=<span class="hljs-string">'model.png'</span>)
<span class="hljs-keyword">if</span> __name__==<span class="hljs-string">'__main__'</span>:
main()
</code></pre></div>
<div class="ui-toc dropup unselectable hidden-print" style="display:none;">
<div class="pull-right dropdown">
<a id="tocLabel" class="ui-toc-label btn btn-default" data-toggle="dropdown" href="#" role="button" aria-haspopup="true" aria-expanded="false" title="Table of content">
<i class="fa fa-bars"></i>
</a>
<ul id="ui-toc" class="ui-toc-dropdown dropdown-menu" aria-labelledby="tocLabel">
<div class="toc"><ul class="nav"><li class=""><a href="#hw3-手把手-q1">HW3 手把手 Q1</a><ul class="nav"><li><a href="#建立模型">建立模型</a></li><li><a href="#training-file">Training File</a><ul class="nav"><li><a href="#不同於平常網路上的keras教學，我們使用比較多manual的loop-code來讓同學比較能瞭解epoch-“batch”-…的概念是什麼">不同於平常網路上的keras教學，我們使用比較多manual的loop code來讓同學比較能瞭解"epoch" “batch” …的概念是什麼</a></li></ul></li><li><a href="#plot-model">Plot model</a></li></ul></li></ul></div><div class="toc-menu"><a class="expand-toggle" href="#">Expand all</a><a class="back-to-top" href="#">Back to top</a><a class="go-to-bottom" href="#">Go to bottom</a></div>
</ul>
</div>
</div>
<div id="ui-toc-affix" class="ui-affix-toc ui-toc-dropdown unselectable hidden-print" data-spy="affix" style="top:17px;display:none;" >
<div class="toc"><ul class="nav"><li class=""><a href="#hw3-手把手-q1">HW3 手把手 Q1</a><ul class="nav"><li><a href="#建立模型">建立模型</a></li><li><a href="#training-file">Training File</a><ul class="nav"><li><a href="#不同於平常網路上的keras教學，我們使用比較多manual的loop-code來讓同學比較能瞭解epoch-“batch”-…的概念是什麼">不同於平常網路上的keras教學，我們使用比較多manual的loop code來讓同學比較能瞭解"epoch" “batch” …的概念是什麼</a></li></ul></li><li><a href="#plot-model">Plot model</a></li></ul></li></ul></div><div class="toc-menu"><a class="expand-toggle" href="#">Expand all</a><a class="back-to-top" href="#">Back to top</a><a class="go-to-bottom" href="#">Go to bottom</a></div>
</div>
</xmp>
</xmp> <script src="strapdown.js"></script> </html>
<footer>
<center><a href="./index.html"><i class="fa fa-home"></i></a></center>
<center><i class="fa fa-github"></i></a> Posted by: <a href="https://github.com/ntumlta/" target="_blank">ntumlta</a> </center>
<center><i class="fa fa-envelope"></i> Contact information: <a href="mailto:"> ntu.mlta@gmail.com </a>.</center>
<center><i class="fa fa-mortar-board"></i> Course information: <a href="http://speech.ee.ntu.edu.tw/~tlkagk/courses_ML17_2.html", target="_blank">Machine Learning (2017, Fall) @ National Taiwan University</a>.</center>
</footer>
<script>
(function(i,s,o,g,r,a,m){i['GoogleAnalyticsObject']=r;i[r]=i[r]||function(){
(i[r].q=i[r].q||[]).push(arguments)},i[r].l=1*new Date();a=s.createElement(o),
m=s.getElementsByTagName(o)[0];a.async=1;a.src=g;m.parentNode.insertBefore(a,m)
})(window,document,'script','https://www.google-analytics.com/analytics.js','ga');
ga('create', 'UA-59748795-2', 'auto');
ga('send', 'pageview');
</script>
</html>