[Tutorials]Deep Leaning Using Apache Spark's BigDL

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. Deep Leaning Tutorials on Apache Spark using BigDL
   
2. 
   
3. Step-by-step Deep Leaning Tutorials on Apache Spark using BigDL. The tutorials are inspired by Apache Spark examples, the Theano Tutorials and the Tensorflow tutorials.
   
4. 
   
5. Topics
   
6. 
   
7. RDD
   
8. DataFrame
   
9. SparkSQL
   
10. StructureStreaming
    
11. Forward and backward
    
12. Linear Regression
    
13. Introduction to MNIST
    
14. Logistic Regression
    
15. Feedforward Neural Network
    
16. Convolutional Neural Network
    
17. Recurrent Neural Network
    
18. LSTM
    
19. Bi-directional RNN
    
20. Auto-encoder
    
21. Environment
    
22. 
    
23. Python 2.7
    
24. JDK 8
    
25. Apache Spark 2.1.0
    
26. Jupyter Notebook 4.1
    
27. BigDL 0.2.0
    
28. Setup env on Mac OS / Setup env on Linux
    
29. Start Jupyter Server
    
30. 
    
31. Download BigDL 0.2.0(linux64, mac) and unzip file.
    
32. Run export BIGDL_HOME=where is your unzipped bigdl folder
    
33. Run export SPARK_HOME=where is your unpacked spark folder
    
34. Run ./start_notebook.sh
    
35. Run Demo
    
36. 
    
37. Open a browser - Suggest Chrome or Firefox or Safari
    
38. Access notebook client at address http://localhost:8888, open the example ipynb files and execute.

复制代码

本帖隐藏的内容

https://github.com/intel-analytics/BigDL-Tutorials

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1 回帖

关键词：Apache Spark Tutorials Tutorial Leaning apache

1. In [2]:
   
2. FEATURES_DIM = 2
   
3. data_len = 100
   
4. 
   
5. def gen_rand_sample():
   
6.     features = np.random.uniform(0, 1, (FEATURES_DIM))
   
7.     label = (2 * features).sum() + 0.4
   
8.     return Sample.from_ndarray(features, label)
   
9. 
   
10. rdd_train = sc.parallelize(range(0, data_len)).map( lambda i: gen_rand_sample() )

复制代码

1. In [3]:
   
2. # Parameters
   
3. learning_rate = 0.2
   
4. training_epochs = 5
   
5. batch_size = 4
   
6. n_input = FEATURES_DIM
   
7. n_output = 1
   
8. 
   
9. def linear_regression(n_input, n_output):
   
10.     # Initialize a sequential container
    
11.     model = Sequential()  
    
12.     # Add a linear layer
    
13.     model.add(Linear(n_input, n_output))
    
14. 
    
15.     return model
    
16. 
    
17. model = linear_regression(n_input, n_output)

复制代码

1. In [4]:
   
2. # Create an Optimizer
   
3. optimizer = Optimizer(
   
4.     model=model,
   
5.     training_rdd=rdd_train,
   
6.     criterion=MSECriterion(),
   
7.     optim_method=SGD(learningrate=learning_rate),
   
8.     end_trigger=MaxEpoch(training_epochs),
   
9.     batch_size=batch_size)

复制代码

1. In [5]:
   
2. # Start to train
   
3. trained_model = optimizer.optimize()

复制代码

1. In [6]:
   
2. # Print the first five predicted results of training data.
   
3. predict_result = trained_model.predict(rdd_train)
   
4. p = predict_result.take(5)
   
5. 
   
6. print("predict predict: \n")
   
7. for i in p:
   
8.     print(str(i) + "\n")

复制代码

1. In [7]:
   
2. def test_predict(trained_model):
   
3.     np.random.seed(100)
   
4.     total_length = 10
   
5.     features = np.random.uniform(0, 1, (total_length, 2))
   
6.     label = (features).sum() + 0.4
   
7.     predict_data = sc.parallelize(range(0, total_length)).map(
   
8.         lambda i: Sample.from_ndarray(features[i], label))
   
9.    
   
10.     predict_result = trained_model.predict(predict_data)
    
11.     p = predict_result.take(6)
    
12.     ground_label = np.array([[-0.47596836], [-0.37598032], [-0.00492062],
    
13.                                  [-0.5906958], [-0.12307882], [-0.77907401]], dtype="float32")
    
14.     mse = ((p - ground_label) ** 2).mean()
    
15.     print mse
    
16.    
    
17. test_predict(trained_model)

复制代码

谢谢楼主分享！

1. In [1]:
   
2. %pylab inline
   
3. import pandas
   
4. import datetime as dt
   
5. 
   
6. from bigdl.nn.layer import *
   
7. from bigdl.nn.criterion import *
   
8. from bigdl.optim.optimizer import *
   
9. from bigdl.util.common import *
   
10. from bigdl.dataset.transformer import *
    
11. from bigdl.dataset import mnist
    
12. from utils import get_mnist
    
13. 
    
14. init_engine()

复制代码

1. In [2]:
   
2. mnist_path = "datasets/mnist"
   
3. (train_data, test_data) = get_mnist(sc, mnist_path)
   
4. 
   
5. print train_data.count()
   
6. print test_data.count()

复制代码

1. In [3]:
   
2. # Parameters
   
3. learning_rate = 0.2
   
4. training_epochs = 15
   
5. batch_size = 2048
   
6. 
   
7. # Network Parameters
   
8. n_input = 784 # MNIST data input (img shape: 28*28)
   
9. n_classes = 10 # MNIST total classes (0-9 digits)

复制代码

1. In [4]:
   
2. # Define the logistic_regression model
   
3. def logistic_regression(n_input, n_classes):
   
4.     # Initialize a sequential container
   
5.     model = Sequential()
   
6. 
   
7.     model.add(Reshape([28*28]))
   
8.     model.add(Linear(n_input, n_classes))
   
9.     model.add(LogSoftMax())
   
10.    
    
11.     return model
    
12. 
    
13. model = logistic_regression(n_input, n_classes)

复制代码

1. # Create an Optimizer
   
2. 
   
3. optimizer = Optimizer(
   
4.     model=model,
   
5.     training_rdd=train_data,
   
6.     criterion=ClassNLLCriterion(),
   
7.     optim_method=SGD(learningrate=learning_rate),
   
8.     end_trigger=MaxEpoch(training_epochs),
   
9.     batch_size=batch_size)

复制代码

1. %%time
   
2. # Start to train
   
3. trained_model = optimizer.optimize()
   
4. print "Optimization Done."

复制代码

1. In [7]:
   
2. def map_predict_label(l):
   
3.     return l.argmax()
   
4. def map_groundtruth_label(l):
   
5.     return l[0] - 1

复制代码

1. In [8]:
   
2. # Prediction
   
3. predictions = trained_model.predict(test_data)
   
4. imshow(np.column_stack([np.array(s.features).reshape(28,28) for s in test_data.take(8)]),cmap='gray'); axis('off')
   
5. print 'Ground Truth labels:'
   
6. print ', '.join(str(map_groundtruth_label(s.label)) for s in test_data.take(8))
   
7. print 'Predicted labels:'
   
8. print ', '.join(str(map_predict_label(s)) for s in predictions.take(8))

复制代码

1. In [1]:
   
2. %pylab inline
   
3. import pandas
   
4. import datetime as dt
   
5. 
   
6. from bigdl.nn.layer import *
   
7. from bigdl.nn.criterion import *
   
8. from bigdl.optim.optimizer import *
   
9. from bigdl.util.common import *
   
10. from bigdl.dataset.transformer import *
    
11. from utils import get_mnist
    
12. 
    
13. init_engine()

复制代码

1. In [2]:
   
2. # Get and store MNIST into RDD of Sample, please edit the "mnist_path" accordingly.
   
3. mnist_path = "datasets/mnist"
   
4. (train_data, test_data) = get_mnist(sc, mnist_path)
   
5. 
   
6. print train_data.count()
   
7. print test_data.count()

复制代码

1. In [3]:
   
2. learning_rate = 0.2
   
3. training_epochs = 15
   
4. batch_size = 2048
   
5. display_step = 1
   
6. 
   
7. # Network Parameters
   
8. n_hidden_1 = 256 # 1st layer number of features
   
9. n_hidden_2 = 256 # 2nd layer number of features
   
10. n_input = 784 # MNIST data input (img shape: 28*28)
    
11. n_classes = 10 # MNIST total classes (0-9 digits)

复制代码

1. In [4]:
   
2. # Create model
   
3. 
   
4. def multilayer_perceptron(n_hidden_1, n_hidden_2, n_input, n_classes):
   
5.     # Initialize a sequential container
   
6.     model = Sequential()
   
7.     # Hidden layer with ReLu activation
   
8.     model.add(Reshape([28*28]))
   
9.     model.add(Linear(n_input, n_hidden_1).set_name('mlp_fc1'))
   
10.     model.add(ReLU())
    
11.     # Hidden layer with ReLu activation
    
12.     model.add(Linear(n_hidden_1, n_hidden_2).set_name('mlp_fc2'))
    
13.     model.add(ReLU())
    
14.     # output layer
    
15.     model.add(Linear(n_hidden_2, n_classes).set_name('mlp_fc3'))
    
16.     model.add(LogSoftMax())
    
17.     return model
    
18. 
    
19. model = multilayer_perceptron(n_hidden_1, n_hidden_2, n_input, n_classes)

复制代码

1. In [5]:
   
2. # Create an Optimizer
   
3. optimizer = Optimizer(
   
4.     model=model,
   
5.     training_rdd=train_data,
   
6.     criterion=ClassNLLCriterion(),
   
7.     optim_method=SGD(learningrate=learning_rate),
   
8.     end_trigger=MaxEpoch(training_epochs),
   
9.     batch_size=batch_size)
   
10. 
    
11. # Set the validation logic
    
12. optimizer.set_validation(
    
13.     batch_size=batch_size,
    
14.     val_rdd=test_data,
    
15.     trigger=EveryEpoch(),
    
16.     val_method=[Top1Accuracy()]
    
17. )
    
18. 
    
19. app_name='multilayer_perceptron-'+dt.datetime.now().strftime("%Y%m%d-%H%M%S")
    
20. train_summary = TrainSummary(log_dir='/tmp/bigdl_summaries',
    
21.                                      app_name=app_name)
    
22. train_summary.set_summary_trigger("Parameters", SeveralIteration(50))
    
23. val_summary = ValidationSummary(log_dir='/tmp/bigdl_summaries',
    
24.                                         app_name=app_name)
    
25. optimizer.set_train_summary(train_summary)
    
26. optimizer.set_val_summary(val_summary)
    
27. print "saving logs to ",app_name

复制代码

1. In [6]:
   
2. %%time
   
3. # Boot training process
   
4. trained_model = optimizer.optimize()
   
5. print "Optimization Done."

复制代码

1. In [7]:
   
2. def map_predict_label(l):
   
3.     return np.array(l).argmax()
   
4. def map_groundtruth_label(l):
   
5.     return l[0] - 1

复制代码

1. In [8]:
   
2. %%time
   
3. predictions = trained_model.predict(test_data)
   
4. imshow(np.column_stack([np.array(s.features).reshape(28,28) for s in test_data.take(8)]),cmap='gray'); axis('off')
   
5. print 'Ground Truth labels:'
   
6. print ', '.join(str(map_groundtruth_label(s.label)) for s in test_data.take(8))
   
7. print 'Predicted labels:'
   
8. print ', '.join(str(map_predict_label(s)) for s in predictions.take(8))

复制代码

1. In [3]:
   
2. # Create a LeNet model
   
3. def build_model(class_num):
   
4.     model = Sequential()
   
5.     model.add(Reshape([1, 28, 28]))
   
6.     model.add(SpatialConvolution(1, 6, 5, 5).set_name('conv1'))
   
7.     model.add(Tanh())
   
8.     model.add(SpatialMaxPooling(2, 2, 2, 2).set_name('pool1'))
   
9.     model.add(Tanh())
   
10.     model.add(SpatialConvolution(6, 12, 5, 5).set_name('conv2'))
    
11.     model.add(SpatialMaxPooling(2, 2, 2, 2).set_name('pool2'))
    
12.     model.add(Reshape([12 * 4 * 4]))
    
13.     model.add(Linear(12 * 4 * 4, 100).set_name('fc1'))
    
14.     model.add(Tanh())
    
15.     model.add(Linear(100, class_num).set_name('score'))
    
16.     model.add(LogSoftMax())
    
17.     return model
    
18. lenet_model = build_model(10)

复制代码

1. In [4]:
   
2. # Create an Optimizer
   
3. 
   
4. optimizer = Optimizer(
   
5.     model=lenet_model,
   
6.     training_rdd=train_data,
   
7.     criterion=ClassNLLCriterion(),
   
8.     optim_method=SGD(learningrate=0.4, learningrate_decay=0.0002),
   
9.     end_trigger=MaxEpoch(20),
   
10.     batch_size=2048)
    
11. 
    
12. # Set the validation logic
    
13. optimizer.set_validation(
    
14.     batch_size=2048,
    
15.     val_rdd=test_data,
    
16.     trigger=EveryEpoch(),
    
17.     val_method=[Top1Accuracy()]
    
18. )
    
19. 
    
20. app_name='lenet-'+dt.datetime.now().strftime("%Y%m%d-%H%M%S")
    
21. train_summary = TrainSummary(log_dir='/tmp/bigdl_summaries',
    
22.                                      app_name=app_name)
    
23. train_summary.set_summary_trigger("Parameters", SeveralIteration(50))
    
24. val_summary = ValidationSummary(log_dir='/tmp/bigdl_summaries',
    
25.                                         app_name=app_name)
    
26. optimizer.set_train_summary(train_summary)
    
27. optimizer.set_val_summary(val_summary)
    
28. print "saving logs to ",app_name

复制代码

1. In [5]:
   
2. %%time
   
3. # Boot training process
   
4. trained_model = optimizer.optimize()
   
5. print "Optimization Done."

复制代码

1. In [6]:
   
2. def map_predict_label(l):
   
3.     return np.array(l).argmax()
   
4. def map_groundtruth_label(l):
   
5.     return l[0] - 1

复制代码

1. In [7]:
   
2. # label-1 to restore the original label.
   
3. print "Ground Truth labels:"
   
4. print ', '.join([str(map_groundtruth_label(s.label)) for s in train_data.take(8)])
   
5. imshow(np.column_stack([np.array(s.features).reshape(28,28) for s in train_data.take(8)]),cmap='gray'); axis('off')

复制代码

1. In [9]:
   
2. params = trained_model.parameters()
   
3. 
   
4. #batch num, output_dim, input_dim, spacial_dim
   
5. for layer_name, param in params.iteritems():
   
6.     print layer_name,param['weight'].shape,param['bias'].shape

复制代码

1. In [10]:
   
2. #vis_square is borrowed from caffe example
   
3. def vis_square(data):
   
4.     """Take an array of shape (n, height, width) or (n, height, width, 3)
   
5.        and visualize each (height, width) thing in a grid of size approx. sqrt(n) by sqrt(n)"""
   
6.    
   
7.     # normalize data for display
   
8.     data = (data - data.min()) / (data.max() - data.min())
   
9.     # force the number of filters to be square
   
10.     n = int(np.ceil(np.sqrt(data.shape[0])))
    
11.     padding = (((0, n ** 2 - data.shape[0]),
    
12.                (0, 1), (0, 1))                 # add some space between filters
    
13.                + ((0, 0),) * (data.ndim - 3))  # don't pad the last dimension (if there is one)
    
14.     data = np.pad(data, padding, mode='constant', constant_values=1)  # pad with ones (white)
    
15.    
    
16.     # tile the filters into an image
    
17.     data = data.reshape((n, n) + data.shape[1:]).transpose((0, 2, 1, 3) + tuple(range(4, data.ndim + 1)))
    
18.     data = data.reshape((n * data.shape[1], n * data.shape[3]) + data.shape[4:])
    
19.   
    
20.     plt.imshow(data,cmap='gray'); plt.axis('off')

复制代码

1. In [11]:
   
2. filters_conv1 = params['conv1']['weight']
   
3. 
   
4. filters_conv1[0,0,0]
   
5. 
   
6. vis_square(np.squeeze(filters_conv1, axis=(0,)).reshape(1*6,5,5))

复制代码

1. In [12]:
   
2. # the parameters are a list of [weights, biases]
   
3. filters_conv2 = params['conv2']['weight']
   
4. 
   
5. vis_square(np.squeeze(filters_conv2, axis=(0,)).reshape(12*6,5,5))

复制代码

1. In [13]:
   
2. loss = np.array(train_summary.read_scalar("Loss"))
   
3. top1 = np.array(val_summary.read_scalar("Top1Accuracy"))
   
4. 
   
5. plt.figure(figsize = (12,12))
   
6. plt.subplot(2,1,1)
   
7. plt.plot(loss[:,0],loss[:,1],label='loss')
   
8. plt.xlim(0,loss.shape[0]+10)
   
9. plt.grid(True)
   
10. plt.title("loss")
    
11. plt.subplot(2,1,2)
    
12. plt.plot(top1[:,0],top1[:,1],label='top1')
    
13. plt.xlim(0,loss.shape[0]+10)
    
14. plt.title("top1 accuracy")
    
15. plt.grid(True)

复制代码

1. In [2]:
   
2. # Get and store MNIST into RDD of Sample, please edit the "mnist_path" accordingly.
   
3. mnist_path = "datasets/mnist"
   
4. (train_data, test_data) = get_mnist(sc, mnist_path)
   
5. 
   
6. train_data = train_data.map(lambda s: Sample.from_ndarray(np.resize(s.features, (28, 28)), s.label))
   
7. test_data = test_data.map(lambda s: Sample.from_ndarray(np.resize(s.features, (28, 28)), s.label))
   
8. print train_data.count()
   
9. print test_data.count()

复制代码

1. In [3]:
   
2. # Parameters
   
3. batch_size = 64
   
4. 
   
5. # Network Parameters
   
6. n_input = 28 # MNIST data input (img shape: 28*28)
   
7. n_hidden = 128 # hidden layer num of features
   
8. n_classes = 10 # MNIST total classes (0-9 digits)

复制代码

1. In [4]:
   
2. def build_model(input_size, hidden_size, output_size):
   
3.     model = Sequential()
   
4.     recurrent = Recurrent()
   
5.     recurrent.add(RnnCell(input_size, hidden_size, Tanh()))
   
6.     model.add(InferReshape([-1, input_size], True))
   
7.     model.add(recurrent)
   
8.     model.add(Select(2, -1))
   
9.     model.add(Linear(hidden_size, output_size))
   
10.     return model
    
11. rnn_model = build_model(n_input, n_hidden, n_classes)

复制代码

1. In [5]:
   
2. # Create an Optimizer
   
3. 
   
4. #criterion = TimeDistributedCriterion(CrossEntropyCriterion())
   
5. criterion = CrossEntropyCriterion()
   
6. optimizer = Optimizer(
   
7.     model=rnn_model,
   
8.     training_rdd=train_data,
   
9.     criterion=criterion,
   
10.     optim_method= Adam(),
    
11.     end_trigger=MaxEpoch(5),
    
12.     batch_size=batch_size)
    
13. 
    
14. # Set the validation logic
    
15. optimizer.set_validation(
    
16.     batch_size=batch_size,
    
17.     val_rdd=test_data,
    
18.     trigger=EveryEpoch(),
    
19.     val_method=[Top1Accuracy()]
    
20. )
    
21. 
    
22. app_name='rnn-'+dt.datetime.now().strftime("%Y%m%d-%H%M%S")
    
23. train_summary = TrainSummary(log_dir='/tmp/bigdl_summaries',
    
24.                                      app_name=app_name)
    
25. train_summary.set_summary_trigger("Parameters", SeveralIteration(50))
    
26. val_summary = ValidationSummary(log_dir='/tmp/bigdl_summaries',
    
27.                                         app_name=app_name)
    
28. optimizer.set_train_summary(train_summary)
    
29. optimizer.set_val_summary(val_summary)
    
30. print "saving logs to ",app_name

复制代码

1. In [6]:
   
2. %%time
   
3. # Boot training process
   
4. trained_model = optimizer.optimize()
   
5. print "Optimization Done."

复制代码

1. In [7]:
   
2. def map_predict_label(l):
   
3.     return np.array(l).argmax()
   
4. def map_groundtruth_label(l):
   
5.     return l[0] - 1

复制代码

1. In [8]:
   
2. %%time
   
3. predictions = trained_model.predict(test_data)
   
4. imshow(np.column_stack([np.array(s.features).reshape(28,28) for s in test_data.take(8)]),cmap='gray'); axis('off')
   
5. print 'Ground Truth labels:'
   
6. print ', '.join(str(map_groundtruth_label(s.label)) for s in test_data.take(8))
   
7. print 'Predicted labels:'
   
8. print ', '.join(str(map_predict_label(s)) for s in predictions.take(8))

复制代码

1. In [9]:
   
2. loss = np.array(train_summary.read_scalar("Loss"))
   
3. top1 = np.array(val_summary.read_scalar("Top1Accuracy"))
   
4. 
   
5. plt.figure(figsize = (12,12))
   
6. plt.subplot(2,1,1)
   
7. plt.plot(loss[:,0],loss[:,1],label='loss')
   
8. plt.xlim(0,loss.shape[0]+10)
   
9. plt.grid(True)
   
10. plt.title("loss")
    
11. plt.subplot(2,1,2)
    
12. plt.plot(top1[:,0],top1[:,1],label='top1')
    
13. plt.xlim(0,loss.shape[0]+10)
    
14. plt.title("top1 accuracy")
    
15. plt.grid(True)

复制代码

1. In [2]:
   
2. # Get and store MNIST into RDD of Sample, please edit the "mnist_path" accordingly.
   
3. mnist_path = "datasets/mnist"
   
4. (train_data, test_data) = get_mnist(sc, mnist_path)
   
5. 
   
6. train_data = train_data.map(lambda s: Sample.from_ndarray(np.resize(s.features, (28, 28)), s.label))
   
7. test_data = test_data.map(lambda s: Sample.from_ndarray(np.resize(s.features, (28, 28)), s.label))
   
8. print train_data.count()
   
9. print test_data.count()

复制代码

1. In [3]:
   
2. # Parameters
   
3. batch_size = 64
   
4. 
   
5. # Network Parameters
   
6. n_input = 28 # MNIST data input (img shape: 28*28)
   
7. n_hidden = 128 # hidden layer num of features
   
8. n_classes = 10 # MNIST total classes (0-9 digits)

复制代码

1. In [4]:
   
2. def build_model(input_size, hidden_size, output_size):
   
3.     model = Sequential()
   
4.     recurrent = Recurrent()
   
5.     recurrent.add(LSTM(input_size, hidden_size))
   
6.     model.add(InferReshape([-1, input_size], True))
   
7.     model.add(recurrent)
   
8.     model.add(Select(2, -1))
   
9.     model.add(Linear(hidden_size, output_size))
   
10.     return model
    
11. rnn_model = build_model(n_input, n_hidden, n_classes)

复制代码

1. In [5]:
   
2. # Create an Optimizer
   
3. 
   
4. #criterion = TimeDistributedCriterion(CrossEntropyCriterion())
   
5. criterion = CrossEntropyCriterion()
   
6. optimizer = Optimizer(
   
7.     model=rnn_model,
   
8.     training_rdd=train_data,
   
9.     criterion=criterion,
   
10.     optim_method=Adam(),
    
11.     end_trigger=MaxEpoch(5),
    
12.     batch_size=batch_size)
    
13. 
    
14. # Set the validation logic
    
15. optimizer.set_validation(
    
16.     batch_size=batch_size,
    
17.     val_rdd=test_data,
    
18.     trigger=EveryEpoch(),
    
19.     val_method=[Top1Accuracy()]
    
20. )
    
21. 
    
22. app_name='rnn-'+dt.datetime.now().strftime("%Y%m%d-%H%M%S")
    
23. train_summary = TrainSummary(log_dir='/tmp/bigdl_summaries',
    
24.                                      app_name=app_name)
    
25. train_summary.set_summary_trigger("Parameters", SeveralIteration(50))
    
26. val_summary = ValidationSummary(log_dir='/tmp/bigdl_summaries',
    
27.                                         app_name=app_name)
    
28. optimizer.set_train_summary(train_summary)
    
29. optimizer.set_val_summary(val_summary)
    
30. print "saving logs to ",app_name

复制代码

1. In [6]:
   
2. %%time
   
3. # Boot training process
   
4. trained_model = optimizer.optimize()
   
5. print "Optimization Done."

复制代码

1. In [7]:
   
2. def map_predict_label(l):
   
3.     return np.array(l).argmax()
   
4. def map_groundtruth_label(l):
   
5.     return l[0] - 1

复制代码

1. In [8]:
   
2. %%time
   
3. predictions = trained_model.predict(test_data)
   
4. imshow(np.column_stack([np.array(s.features).reshape(28,28) for s in test_data.take(8)]),cmap='gray'); axis('off')
   
5. print 'Ground Truth labels:'
   
6. print ', '.join(str(map_groundtruth_label(s.label)) for s in test_data.take(8))
   
7. print 'Predicted labels:'
   
8. print ', '.join(str(map_predict_label(s)) for s in predictions.take(8))

复制代码

1. In [9]:
   
2. loss = np.array(train_summary.read_scalar("Loss"))
   
3. top1 = np.array(val_summary.read_scalar("Top1Accuracy"))
   
4. 
   
5. plt.figure(figsize = (12,12))
   
6. plt.subplot(2,1,1)
   
7. plt.plot(loss[:,0],loss[:,1],label='loss')
   
8. plt.xlim(0,loss.shape[0]+10)
   
9. plt.grid(True)
   
10. plt.title("loss")
    
11. plt.subplot(2,1,2)
    
12. plt.plot(top1[:,0],top1[:,1],label='top1')
    
13. plt.xlim(0,loss.shape[0]+10)
    
14. plt.title("top1 accuracy")
    
15. plt.grid(True)

复制代码

谢谢分享