ruby-dnn 1.1.4 → 1.2.2
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/.gitignore +1 -0
- data/.travis.yml +2 -1
- data/README.md +39 -22
- data/examples/api-examples/early_stopping_example.rb +6 -6
- data/examples/api-examples/initializer_example.rb +6 -6
- data/examples/api-examples/regularizer_example.rb +6 -6
- data/examples/api-examples/save_example.rb +6 -6
- data/examples/dcgan/dcgan.rb +27 -27
- data/examples/judge-number/README.md +29 -0
- data/examples/judge-number/capture.PNG +0 -0
- data/examples/judge-number/convnet8.rb +70 -0
- data/examples/judge-number/make_weights.rb +5 -0
- data/examples/judge-number/mnist_predict.rb +20 -0
- data/examples/judge-number/mnist_train.rb +19 -0
- data/examples/judge-number/public/httpRequest.js +44 -0
- data/examples/judge-number/public/judgeNumber.js +61 -0
- data/examples/judge-number/server.rb +19 -0
- data/examples/judge-number/trained_mnist_params.marshal +0 -0
- data/examples/judge-number/views/index.erb +7 -0
- data/examples/mnist_conv2d_example.rb +3 -3
- data/examples/mnist_define_by_run.rb +7 -7
- data/examples/mnist_gpu.rb +47 -0
- data/examples/mnist_lstm_example.rb +1 -1
- data/examples/pix2pix/dcgan.rb +54 -66
- data/examples/pix2pix/train.rb +2 -2
- data/examples/vae.rb +13 -13
- data/img/cart-pole.gif +0 -0
- data/img/cycle-gan.PNG +0 -0
- data/img/facade-pix2pix.png +0 -0
- data/lib/dnn.rb +24 -3
- data/lib/dnn/core/callbacks.rb +6 -4
- data/lib/dnn/core/layers/basic_layers.rb +40 -22
- data/lib/dnn/core/layers/cnn_layers.rb +33 -5
- data/lib/dnn/core/layers/math_layers.rb +17 -9
- data/lib/dnn/core/layers/merge_layers.rb +2 -26
- data/lib/dnn/core/layers/split_layers.rb +39 -0
- data/lib/dnn/core/link.rb +14 -33
- data/lib/dnn/core/losses.rb +6 -12
- data/lib/dnn/core/models.rb +77 -10
- data/lib/dnn/core/optimizers.rb +8 -1
- data/lib/dnn/core/utils.rb +23 -0
- data/lib/dnn/image.rb +48 -0
- data/lib/dnn/version.rb +1 -1
- data/ruby-dnn.gemspec +2 -15
- metadata +40 -20
- data/bin/console +0 -14
- data/bin/setup +0 -8
checksums.yaml
CHANGED
@@ -1,7 +1,7 @@
|
|
1
1
|
---
|
2
2
|
SHA256:
|
3
|
-
metadata.gz:
|
4
|
-
data.tar.gz:
|
3
|
+
metadata.gz: 2aa12b717ef532b8afe44de7cb388c7d87cb271bf38a25adaad2c335c8817d4b
|
4
|
+
data.tar.gz: e1322f86b06c11ac3728e948e18469ccc6e454eeefb180108a0a9bfc2dbd8143
|
5
5
|
SHA512:
|
6
|
-
metadata.gz:
|
7
|
-
data.tar.gz:
|
6
|
+
metadata.gz: 6c9c53ca73a5ab7fc53935f37e0804d761e36735aa06972f63e151ce76a44f01878a2c01dfe0622bf39baa5c98b16ee8d06db212356fcc329a9dd40ae2d78f1c
|
7
|
+
data.tar.gz: 5a195c0afd677127afad2433df2fda3a09c711464c46cb174713d00d695026d2b86faa64f82cc96f2b417b468fc372c04efee22761800c900e2f6b0ca05ac57d
|
data/.gitignore
CHANGED
data/.travis.yml
CHANGED
data/README.md
CHANGED
@@ -1,6 +1,7 @@
|
|
1
1
|
# ruby-dnn
|
2
|
-
[![Gem Version](https://badge.fury.io/rb/ruby-dnn.svg)](https://badge.fury.io/rb/ruby-dnn)
|
2
|
+
[![Gem Version](https://badge.fury.io/rb/ruby-dnn.svg)](https://badge.fury.io/rb/ruby-dnn)
|
3
3
|
[![Build Status](https://travis-ci.org/unagiootoro/ruby-dnn.svg?branch=master)](https://travis-ci.org/unagiootoro/ruby-dnn)
|
4
|
+
[![Docs Latest](https://img.shields.io/badge/docs-latest-blue.svg)](https://rubydoc.info/gems/ruby-dnn)
|
4
5
|
|
5
6
|
ruby-dnn is a ruby deep learning library. This library supports full connected neural network and convolution neural network
|
6
7
|
and recurrent neural network.
|
@@ -54,18 +55,18 @@ When create a model with 'define by run' style:
|
|
54
55
|
class MLP < Model
|
55
56
|
def initialize
|
56
57
|
super
|
57
|
-
@
|
58
|
-
@
|
59
|
-
@
|
58
|
+
@d1 = Dense.new(256)
|
59
|
+
@d2 = Dense.new(256)
|
60
|
+
@d3 = Dense.new(10)
|
60
61
|
end
|
61
62
|
|
62
63
|
def forward(x)
|
63
64
|
x = InputLayer.new(784).(x)
|
64
|
-
x = @
|
65
|
+
x = @d1.(x)
|
65
66
|
x = ReLU.(x)
|
66
|
-
x = @
|
67
|
+
x = @d2.(x)
|
67
68
|
x = ReLU.(x)
|
68
|
-
x = @
|
69
|
+
x = @d3.(x)
|
69
70
|
x
|
70
71
|
end
|
71
72
|
end
|
@@ -84,6 +85,23 @@ puts "loss: #{loss}"
|
|
84
85
|
Please refer to examples for basic usage.
|
85
86
|
If you want to know more detailed information, please refer to the source code.
|
86
87
|
|
88
|
+
## Sample
|
89
|
+
|
90
|
+
* Pix2pix
|
91
|
+
Convert an abstract image into a building image.
|
92
|
+
[https://github.com/unagiootoro/facade-pix2pix](https://github.com/unagiootoro/facade-pix2pix)
|
93
|
+
![facade-pix2pix.png](img/facade-pix2pix.png)
|
94
|
+
|
95
|
+
* Cycle-GAN
|
96
|
+
Convert apples to oranges and oranges to apples.
|
97
|
+
[https://github.com/unagiootoro/apple2orange-cyclegan](https://github.com/unagiootoro/apple2orange-cyclegan)
|
98
|
+
![cycle-gan.PNG](img/cycle-gan.PNG)
|
99
|
+
|
100
|
+
* DQN
|
101
|
+
Learn the game so that the pole on the cart does not fall.
|
102
|
+
[https://github.com/unagiootoro/ruby-rl](https://github.com/unagiootoro/ruby-rl)
|
103
|
+
![cart-pole.gif](img/cart-pole.gif)
|
104
|
+
|
87
105
|
## Implemented
|
88
106
|
|| Implemented classes |
|
89
107
|
|:-----------|------------:|
|
@@ -95,28 +113,27 @@ If you want to know more detailed information, please refer to the source code.
|
|
95
113
|
| Losses | MeanSquaredError, MeanAbsoluteError, Hinge, HuberLoss, SoftmaxCrossEntropy, SigmoidCrossEntropy |
|
96
114
|
|
97
115
|
## Datasets
|
98
|
-
By setting the environment variable
|
116
|
+
By setting the environment variable `RUBY_DNN_DOWNLOADS_PATH`, you can specify the path to read dataset.
|
99
117
|
|
100
|
-
|
101
|
-
|
102
|
-
|
103
|
-
|
104
|
-
|
105
|
-
|
118
|
+
* Iris
|
119
|
+
* MNIST
|
120
|
+
* Fashion-MNIST
|
121
|
+
* CIFAR-10
|
122
|
+
* CIFAR-100
|
123
|
+
* STL-10
|
106
124
|
|
107
|
-
##
|
108
|
-
|
109
|
-
|
110
|
-
● Pix2pix
|
125
|
+
## Use GPU
|
126
|
+
If you do `require "cumo/narray"` before `require "dnn"`, you can run it on GPU.
|
127
|
+
Or, set the environment variable `RUBY_DNN_USE_CUMO` to `ENABLE` to force the GPU to be used.
|
111
128
|
|
112
129
|
## TODO
|
113
|
-
|
114
|
-
|
115
|
-
|
130
|
+
* Write a test.
|
131
|
+
* Write a document.
|
132
|
+
* Improve performance when using GPU.
|
116
133
|
|
117
134
|
## Development
|
118
135
|
|
119
|
-
After checking out the repo, run `bin/setup` to install dependencies. Then, run `rake "
|
136
|
+
After checking out the repo, run `bin/setup` to install dependencies. Then, run `rake "test"` to run the tests. You can also run `bin/console` for an interactive prompt that will allow you to experiment.
|
120
137
|
|
121
138
|
To install this gem onto your local machine, run `bundle exec rake install`. To release a new version, update the version number in `version.rb`, and then run `bundle exec rake release`, which will create a git tag for the version, push git commits and tags, and push the `.gem` file to [rubygems.org](https://rubygems.org).
|
122
139
|
|
@@ -27,22 +27,22 @@ y_test = DNN::Utils.to_categorical(y_test, 10, Numo::SFloat)
|
|
27
27
|
class MLP < Model
|
28
28
|
def initialize
|
29
29
|
super
|
30
|
-
@
|
31
|
-
@
|
32
|
-
@
|
30
|
+
@d1 = Dense.new(256)
|
31
|
+
@d2 = Dense.new(256)
|
32
|
+
@d3 = Dense.new(10)
|
33
33
|
@bn1 = BatchNormalization.new
|
34
34
|
@bn2 = BatchNormalization.new
|
35
35
|
end
|
36
36
|
|
37
37
|
def forward(x)
|
38
38
|
x = InputLayer.new(784).(x)
|
39
|
-
x = @
|
39
|
+
x = @d1.(x)
|
40
40
|
x = @bn1.(x)
|
41
41
|
x = ReLU.(x)
|
42
|
-
x = @
|
42
|
+
x = @d2.(x)
|
43
43
|
x = @bn2.(x)
|
44
44
|
x = ReLU.(x)
|
45
|
-
x = @
|
45
|
+
x = @d3.(x)
|
46
46
|
x
|
47
47
|
end
|
48
48
|
end
|
@@ -28,22 +28,22 @@ class MLP < Model
|
|
28
28
|
def initialize
|
29
29
|
super
|
30
30
|
# Set the initial values of weight and bias to the initial values of He.
|
31
|
-
@
|
32
|
-
@
|
33
|
-
@
|
31
|
+
@d1 = Dense.new(256, weight_initializer: He.new, bias_initializer: He.new)
|
32
|
+
@d2 = Dense.new(256, weight_initializer: He.new, bias_initializer: He.new)
|
33
|
+
@d3 = Dense.new(10, weight_initializer: He.new, bias_initializer: He.new)
|
34
34
|
@bn1 = BatchNormalization.new
|
35
35
|
@bn2 = BatchNormalization.new
|
36
36
|
end
|
37
37
|
|
38
38
|
def forward(x)
|
39
39
|
x = InputLayer.new(784).(x)
|
40
|
-
x = @
|
40
|
+
x = @d1.(x)
|
41
41
|
x = @bn1.(x)
|
42
42
|
x = ReLU.(x)
|
43
|
-
x = @
|
43
|
+
x = @d2.(x)
|
44
44
|
x = @bn2.(x)
|
45
45
|
x = ReLU.(x)
|
46
|
-
x = @
|
46
|
+
x = @d3.(x)
|
47
47
|
x
|
48
48
|
end
|
49
49
|
end
|
@@ -29,22 +29,22 @@ class MLP < Model
|
|
29
29
|
def initialize
|
30
30
|
super
|
31
31
|
# Set L2 regularizer(weight decay) for weight and bias.
|
32
|
-
@
|
33
|
-
@
|
34
|
-
@
|
32
|
+
@d1 = Dense.new(256, weight_regularizer: L2.new(L2_LAMBDA), bias_regularizer: L2.new(L2_LAMBDA))
|
33
|
+
@d2 = Dense.new(256, weight_regularizer: L2.new(L2_LAMBDA), bias_regularizer: L2.new(L2_LAMBDA))
|
34
|
+
@d3 = Dense.new(10, weight_regularizer: L2.new(L2_LAMBDA), bias_regularizer: L2.new(L2_LAMBDA))
|
35
35
|
@bn1 = BatchNormalization.new
|
36
36
|
@bn2 = BatchNormalization.new
|
37
37
|
end
|
38
38
|
|
39
39
|
def forward(x)
|
40
40
|
x = InputLayer.new(784).(x)
|
41
|
-
x = @
|
41
|
+
x = @d1.(x)
|
42
42
|
x = @bn1.(x)
|
43
43
|
x = ReLU.(x)
|
44
|
-
x = @
|
44
|
+
x = @d2.(x)
|
45
45
|
x = @bn2.(x)
|
46
46
|
x = ReLU.(x)
|
47
|
-
x = @
|
47
|
+
x = @d3.(x)
|
48
48
|
x
|
49
49
|
end
|
50
50
|
end
|
@@ -38,22 +38,22 @@ y_test = DNN::Utils.to_categorical(y_test, 10, Numo::SFloat)
|
|
38
38
|
class MLP < Model
|
39
39
|
def initialize
|
40
40
|
super
|
41
|
-
@
|
42
|
-
@
|
43
|
-
@
|
41
|
+
@d1 = Dense.new(256)
|
42
|
+
@d2 = Dense.new(256)
|
43
|
+
@d3 = Dense.new(10)
|
44
44
|
@bn1 = BatchNormalization.new
|
45
45
|
@bn2 = BatchNormalization.new
|
46
46
|
end
|
47
47
|
|
48
48
|
def forward(x)
|
49
49
|
x = InputLayer.new(784).(x)
|
50
|
-
x = @
|
50
|
+
x = @d1.(x)
|
51
51
|
x = @bn1.(x)
|
52
52
|
x = ReLU.(x)
|
53
|
-
x = @
|
53
|
+
x = @d2.(x)
|
54
54
|
x = @bn2.(x)
|
55
55
|
x = ReLU.(x)
|
56
|
-
x = @
|
56
|
+
x = @d3.(x)
|
57
57
|
x
|
58
58
|
end
|
59
59
|
end
|
data/examples/dcgan/dcgan.rb
CHANGED
@@ -4,13 +4,13 @@ include DNN::Layers
|
|
4
4
|
class Generator < Model
|
5
5
|
def initialize
|
6
6
|
super
|
7
|
-
@
|
8
|
-
@
|
9
|
-
@
|
10
|
-
@
|
11
|
-
@
|
12
|
-
@
|
13
|
-
@
|
7
|
+
@d1 = Dense.new(1024)
|
8
|
+
@d2 = Dense.new(7 * 7 * 64)
|
9
|
+
@cv1 = Conv2D.new(64, 4, padding: true)
|
10
|
+
@cvt1 = Conv2DTranspose.new(64, 4, strides: 2, padding: true)
|
11
|
+
@cvt2 = Conv2DTranspose.new(32, 4, strides: 2, padding: true)
|
12
|
+
@cv2 = Conv2D.new(32, 4, padding: true)
|
13
|
+
@cv3 = Conv2D.new(1, 4, padding: true)
|
14
14
|
@bn1 = BatchNormalization.new
|
15
15
|
@bn2 = BatchNormalization.new
|
16
16
|
@bn3 = BatchNormalization.new
|
@@ -21,32 +21,32 @@ class Generator < Model
|
|
21
21
|
|
22
22
|
def forward(x)
|
23
23
|
x = InputLayer.new(20).(x)
|
24
|
-
x = @
|
24
|
+
x = @d1.(x)
|
25
25
|
x = @bn1.(x)
|
26
26
|
x = ReLU.(x)
|
27
27
|
|
28
|
-
x = @
|
28
|
+
x = @d2.(x)
|
29
29
|
x = @bn2.(x)
|
30
30
|
x = ReLU.(x)
|
31
31
|
|
32
32
|
x = Reshape.(x, [7, 7, 64])
|
33
|
-
x = @
|
33
|
+
x = @cvt1.(x)
|
34
34
|
x = @bn3.(x)
|
35
35
|
x = ReLU.(x)
|
36
36
|
|
37
|
-
x = @
|
37
|
+
x = @cv1.(x)
|
38
38
|
x = @bn4.(x)
|
39
39
|
x = ReLU.(x)
|
40
40
|
|
41
|
-
x = @
|
41
|
+
x = @cvt2.(x)
|
42
42
|
x = @bn5.(x)
|
43
43
|
x = ReLU.(x)
|
44
44
|
|
45
|
-
x = @
|
45
|
+
x = @cv2.(x)
|
46
46
|
x = @bn6.(x)
|
47
47
|
x = ReLU.(x)
|
48
48
|
|
49
|
-
x = @
|
49
|
+
x = @cv3.(x)
|
50
50
|
x = Tanh.(x)
|
51
51
|
x
|
52
52
|
end
|
@@ -55,12 +55,12 @@ end
|
|
55
55
|
class Discriminator < Model
|
56
56
|
def initialize
|
57
57
|
super
|
58
|
-
@
|
59
|
-
@
|
60
|
-
@
|
61
|
-
@
|
62
|
-
@
|
63
|
-
@
|
58
|
+
@cv1 = Conv2D.new(32, 4, strides: 2, padding: true)
|
59
|
+
@cv2 = Conv2D.new(32, 4, padding: true)
|
60
|
+
@cv3 = Conv2D.new(64, 4, strides: 2, padding: true)
|
61
|
+
@cv4 = Conv2D.new(64, 4, padding: true)
|
62
|
+
@d1 = Dense.new(1024)
|
63
|
+
@d2 = Dense.new(1)
|
64
64
|
@bn1 = BatchNormalization.new
|
65
65
|
@bn2 = BatchNormalization.new
|
66
66
|
@bn3 = BatchNormalization.new
|
@@ -68,26 +68,26 @@ class Discriminator < Model
|
|
68
68
|
|
69
69
|
def forward(x)
|
70
70
|
x = InputLayer.new([28, 28, 1]).(x)
|
71
|
-
x = @
|
71
|
+
x = @cv1.(x)
|
72
72
|
x = LeakyReLU.(x, 0.2)
|
73
73
|
|
74
|
-
x = @
|
74
|
+
x = @cv2.(x)
|
75
75
|
x = @bn1.(x)
|
76
76
|
x = LeakyReLU.(x, 0.2)
|
77
77
|
|
78
|
-
x = @
|
78
|
+
x = @cv3.(x)
|
79
79
|
x = @bn2.(x)
|
80
80
|
x = LeakyReLU.(x, 0.2)
|
81
81
|
|
82
|
-
x = @
|
82
|
+
x = @cv4.(x)
|
83
83
|
x = @bn3.(x)
|
84
84
|
x = LeakyReLU.(x, 0.2)
|
85
85
|
|
86
86
|
x = Flatten.(x)
|
87
|
-
x = @
|
87
|
+
x = @d1.(x)
|
88
88
|
x = LeakyReLU.(x, 0.2)
|
89
89
|
|
90
|
-
x = @
|
90
|
+
x = @d2.(x)
|
91
91
|
x
|
92
92
|
end
|
93
93
|
|
@@ -129,7 +129,7 @@ class DCGAN < Model
|
|
129
129
|
y_fake = Numo::SFloat.zeros(batch_size, 1)
|
130
130
|
@dis.enable_training
|
131
131
|
dis_loss = @dis.train_on_batch(x_batch, y_real)
|
132
|
-
dis_loss
|
132
|
+
dis_loss += @dis.train_on_batch(images, y_fake)
|
133
133
|
|
134
134
|
noise = Numo::SFloat.new(batch_size, 20).rand(-1, 1)
|
135
135
|
label = Numo::SFloat.cast([1] * batch_size).reshape(batch_size, 1)
|
@@ -0,0 +1,29 @@
|
|
1
|
+
# Prepare
|
2
|
+
This example use to sinatra.
|
3
|
+
|
4
|
+
```
|
5
|
+
$ gem install sinatra
|
6
|
+
$ gem install sinatra-contrib
|
7
|
+
```
|
8
|
+
|
9
|
+
# Let's try
|
10
|
+
This example prepared weights that have already been trained.
|
11
|
+
If you want to try it immediately, skip steps (1) and (2).
|
12
|
+
|
13
|
+
### (1) Training MNIST
|
14
|
+
```
|
15
|
+
$ ruby mnist_train.rb
|
16
|
+
```
|
17
|
+
|
18
|
+
### (2) Make weights
|
19
|
+
```
|
20
|
+
$ ruby make_weights.rb
|
21
|
+
```
|
22
|
+
|
23
|
+
### (3) Launch sinatra server
|
24
|
+
```
|
25
|
+
$ ruby server.rb
|
26
|
+
```
|
27
|
+
|
28
|
+
### (4) Access 127.0.0.1:4567 with your browser
|
29
|
+
![](capture.PNG)
|
Binary file
|
@@ -0,0 +1,70 @@
|
|
1
|
+
require "dnn"
|
2
|
+
require "numo/linalg/autoloader"
|
3
|
+
|
4
|
+
include DNN::Models
|
5
|
+
include DNN::Layers
|
6
|
+
include DNN::Optimizers
|
7
|
+
include DNN::Losses
|
8
|
+
|
9
|
+
class ConvNet < Model
|
10
|
+
def self.create(input_shape)
|
11
|
+
convnet = ConvNet.new(input_shape, 32)
|
12
|
+
convnet.setup(Adam.new, SoftmaxCrossEntropy.new)
|
13
|
+
convnet
|
14
|
+
end
|
15
|
+
|
16
|
+
def initialize(input_shape, base_filter_size)
|
17
|
+
super()
|
18
|
+
@input_shape = input_shape
|
19
|
+
@cv1 = Conv2D.new(base_filter_size, 3, padding: true)
|
20
|
+
@cv2 = Conv2D.new(base_filter_size, 3, padding: true)
|
21
|
+
@cv3 = Conv2D.new(base_filter_size * 2, 3, padding: true)
|
22
|
+
@cv4 = Conv2D.new(base_filter_size * 2, 3, padding: true)
|
23
|
+
@cv5 = Conv2D.new(base_filter_size * 4, 3, padding: true)
|
24
|
+
@cv6 = Conv2D.new(base_filter_size * 4, 3, padding: true)
|
25
|
+
@bn1 = BatchNormalization.new
|
26
|
+
@bn2 = BatchNormalization.new
|
27
|
+
@bn3 = BatchNormalization.new
|
28
|
+
@bn4 = BatchNormalization.new
|
29
|
+
@d1 = Dense.new(512)
|
30
|
+
@d2 = Dense.new(10)
|
31
|
+
end
|
32
|
+
|
33
|
+
def forward(x)
|
34
|
+
x = InputLayer.new(@input_shape).(x)
|
35
|
+
|
36
|
+
x = @cv1.(x)
|
37
|
+
x = ReLU.(x)
|
38
|
+
x = Dropout.(x, 0.25)
|
39
|
+
|
40
|
+
x = @cv2.(x)
|
41
|
+
x = @bn1.(x)
|
42
|
+
x = ReLU.(x)
|
43
|
+
x = MaxPool2D.(x, 2)
|
44
|
+
|
45
|
+
x = @cv3.(x)
|
46
|
+
x = ReLU.(x)
|
47
|
+
x = Dropout.(x, 0.25)
|
48
|
+
|
49
|
+
x = @cv4.(x)
|
50
|
+
x = @bn2.(x)
|
51
|
+
x = ReLU.(x)
|
52
|
+
x = MaxPool2D.(x, 2)
|
53
|
+
|
54
|
+
x = @cv5.(x)
|
55
|
+
x = ReLU.(x)
|
56
|
+
x = Dropout.(x, 0.25)
|
57
|
+
|
58
|
+
x = @cv6.(x)
|
59
|
+
x = @bn3.(x)
|
60
|
+
x = ReLU.(x)
|
61
|
+
x = MaxPool2D.(x, 2)
|
62
|
+
|
63
|
+
x = Flatten.(x)
|
64
|
+
x = @d1.(x)
|
65
|
+
x = @bn4.(x)
|
66
|
+
x = ReLU.(x)
|
67
|
+
x = @d2.(x)
|
68
|
+
x
|
69
|
+
end
|
70
|
+
end
|