[Microsoft Cognitive Toolkit]MultiLayer Perceptron using CNTK

是 否 +2 论坛币

k人 参与回答

经管之家送您一份

应届毕业生专属福利!

求职就业群

赵安豆老师微信：zhaoandou666

经管之家联合CDA

送您一个全额奖学金名额~ !

立即领取

感谢您参与论坛问题回答

经管之家送您两个论坛币！

+2 论坛币

1. In this tutorial, we train a multi-layer perceptron on MNIST data. This notebook provides the recipe using Python APIs. If you are looking for this example in BrainScript, please look here
   
2. Introduction
   
3. Problem As in CNTK 103B, we will continue to work on the same problem of recognizing digits in MNIST data. The MNIST data comprises hand-written digits with little background noise.

复制代码

本帖隐藏的内容

MultiLayer Perceptron using CNTK.pdf (100.5 KB)

2017-9-17 02:44:28 上传

扫码加我 拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

分享0 收藏0 回帖

关键词：Microsoft Cognitive Toolkit Micro soft

1. # Copyright (c) Microsoft. All rights reserved.
   
2. 
   
3. # Licensed under the MIT license. See LICENSE.md file in the project root
   
4. # for full license information.
   
5. # ==============================================================================
   
6. 
   
7. import argparse
   
8. import numpy as np
   
9. import sys
   
10. import os
    
11. import cntk as C
    
12. from cntk.train import Trainer, minibatch_size_schedule
    
13. from cntk.io import MinibatchSource, CTFDeserializer, StreamDef, StreamDefs, INFINITELY_REPEAT
    
14. from cntk.device import cpu, try_set_default_device
    
15. from cntk.learners import adadelta, learning_rate_schedule, UnitType
    
16. from cntk.ops import relu, element_times, constant
    
17. from cntk.layers import Dense, Sequential, For
    
18. from cntk.losses import cross_entropy_with_softmax
    
19. from cntk.metrics import classification_error
    
20. from cntk.train.training_session import *
    
21. from cntk.logging import ProgressPrinter, TensorBoardProgressWriter
    
22. 
    
23. abs_path = os.path.dirname(os.path.abspath(__file__))
    
24. 
    
25. def check_path(path):
    
26.     if not os.path.exists(path):
    
27.         readme_file = os.path.normpath(os.path.join(
    
28.             os.path.dirname(path), "..", "README.md"))
    
29.         raise RuntimeError(
    
30.             "File '%s' does not exist. Please follow the instructions at %s to download and prepare it." % (path, readme_file))
    
31. 
    
32. def create_reader(path, is_training, input_dim, label_dim):
    
33.     return MinibatchSource(CTFDeserializer(path, StreamDefs(
    
34.         features  = StreamDef(field='features', shape=input_dim, is_sparse=False),
    
35.         labels    = StreamDef(field='labels',   shape=label_dim, is_sparse=False)
    
36.     )), randomize=is_training, max_sweeps = INFINITELY_REPEAT if is_training else 1)
    
37. 
    
38. 
    
39. # Creates and trains a feedforward classification model for MNIST images
    
40. 
    
41. def simple_mnist(tensorboard_logdir=None):
    
42.     input_dim = 784
    
43.     num_output_classes = 10
    
44.     num_hidden_layers = 1
    
45.     hidden_layers_dim = 200
    
46. 
    
47.     # Input variables denoting the features and label data
    
48.     feature = C.input_variable(input_dim, np.float32)
    
49.     label = C.input_variable(num_output_classes, np.float32)
    
50. 
    
51.     # Instantiate the feedforward classification model
    
52.     scaled_input = element_times(constant(0.00390625), feature)
    
53. 
    
54.     z = Sequential([For(range(num_hidden_layers), lambda i: Dense(hidden_layers_dim, activation=relu)),
    
55.                     Dense(num_output_classes)])(scaled_input)
    
56. 
    
57.     ce = cross_entropy_with_softmax(z, label)
    
58.     pe = classification_error(z, label)
    
59. 
    
60.     data_dir = os.path.join(abs_path, "..", "..", "..", "DataSets", "MNIST")
    
61. 
    
62.     path = os.path.normpath(os.path.join(data_dir, "Train-28x28_cntk_text.txt"))
    
63.     check_path(path)
    
64. 
    
65.     reader_train = create_reader(path, True, input_dim, num_output_classes)
    
66. 
    
67.     input_map = {
    
68.         feature  : reader_train.streams.features,
    
69.         label  : reader_train.streams.labels
    
70.     }
    
71. 
    
72.     # Training config
    
73.     minibatch_size = 64
    
74.     num_samples_per_sweep = 60000
    
75.     num_sweeps_to_train_with = 10
    
76. 
    
77.     # Instantiate progress writers.
    
78.     #training_progress_output_freq = 100
    
79.     progress_writers = [ProgressPrinter(
    
80.         #freq=training_progress_output_freq,
    
81.         tag='Training',
    
82.         num_epochs=num_sweeps_to_train_with)]
    
83. 
    
84.     if tensorboard_logdir is not None:
    
85.         progress_writers.append(TensorBoardProgressWriter(freq=10, log_dir=tensorboard_logdir, model=z))
    
86. 
    
87.     # Instantiate the trainer object to drive the model training
    
88.     lr = learning_rate_schedule(1, UnitType.sample)
    
89.     trainer = Trainer(z, (ce, pe), adadelta(z.parameters, lr), progress_writers)
    
90. 
    
91.     training_session(
    
92.         trainer=trainer,
    
93.         mb_source = reader_train,
    
94.         mb_size = minibatch_size,
    
95.         model_inputs_to_streams = input_map,
    
96.         max_samples = num_samples_per_sweep * num_sweeps_to_train_with,
    
97.         progress_frequency=num_samples_per_sweep
    
98.     ).train()
    
99.    
    
100.     # Load test data
     
101.     path = os.path.normpath(os.path.join(data_dir, "Test-28x28_cntk_text.txt"))
     
102.     check_path(path)
     
103. 
     
104.     reader_test = create_reader(path, False, input_dim, num_output_classes)
     
105. 
     
106.     input_map = {
     
107.         feature  : reader_test.streams.features,
     
108.         label  : reader_test.streams.labels
     
109.     }
     
110. 
     
111.     # Test data for trained model
     
112.     test_minibatch_size = 1024
     
113.     num_samples = 10000
     
114.     num_minibatches_to_test = num_samples / test_minibatch_size
     
115.     test_result = 0.0
     
116.     for i in range(0, int(num_minibatches_to_test)):
     
117.         mb = reader_test.next_minibatch(test_minibatch_size, input_map=input_map)
     
118.         eval_error = trainer.test_minibatch(mb)
     
119.         test_result = test_result + eval_error
     
120. 
     
121.     # Average of evaluation errors of all test minibatches
     
122.     return test_result / num_minibatches_to_test
     
123. 
     
124. 
     
125. if __name__=='__main__':
     
126.     # Specify the target device to be used for computing, if you do not want to
     
127.     # use the best available one, e.g.
     
128.     # try_set_default_device(cpu())
     
129. 
     
130.     parser = argparse.ArgumentParser()
     
131.     parser.add_argument('-tensorboard_logdir', '--tensorboard_logdir',
     
132.                         help='Directory where TensorBoard logs should be created', required=False, default=None)
     
133.     args = vars(parser.parse_args())
     
134. 
     
135.     error = simple_mnist(args['tensorboard_logdir'])
     
136.     print("Error: %f" % error)

复制代码

谢谢楼主分享！