C# Redis幫助類
由於一些原因,視訊錄製要告一段落了。再寫一篇關於cntk的文章分享出來吧。我也很想將這個事情進行下去。以後如果條件允許還會接著做。
cntk2.0框架生成的模型才可以支援python。1.0不支援。
python可以匯入cntk.exe生成的框架,也可以匯入python呼叫cntk生成的框架。舉兩個例子:
1 、匯入cntk.exe生成的框架。
from cntk.ops.functions import load_model from PIL import Image import numpy as np from sklearn.utils import shuffle np.random.seed(0) def generate(N, mean, cov, diff): #import ipdb;ipdb.set_trace() samples_per_class = int(N/2) X0 = np.random.multivariate_normal(mean, cov, samples_per_class) Y0 = np.zeros(samples_per_class) for ci, d in enumerate(diff): X1 = np.random.multivariate_normal(mean+d, cov, samples_per_class) Y1 = (ci+1)*np.ones(samples_per_class) X0 = np.concatenate((X0,X1)) Y0 = np.concatenate((Y0,Y1)) X, Y = shuffle(X0, Y0) return X,Y mean = np.random.randn(2) cov = np.eye(2) features, labels = generate(6, mean, cov, [[3.0], [3.0, 0.0]]) features= features.astype(np.float32) labels= labels.astype(np.int) print(features) print(labels) z = load_model("MC.dnn") print(z.parameters[0].value) print(z.parameters[0]) print(z) print(z.uid) #print(z.signature) #print(z.layers[0].E.shape) #print(z.layers[2].b.value) for index in range(len(z.inputs)): print("Index {} for input: {}.".format(index, z.inputs[index])) for index in range(len(z.outputs)): print("Index {} for output: {}.".format(index, z.outputs[index].name)) import cntk as ct z_out = ct.combine([z.outputs[2].owner]) predictions = np.squeeze(z_out.eval({z_out.arguments[0]:[features]})) ret = list() for t in predictions: ret.append(np.argmax(t)) top_class = np.argmax(predictions) print(ret) print("predictions{}.top_class{}".format(predictions,top_class))
上述的程式碼生成一個.py檔案。放到3分類例子中,跟模型一個資料夾下(需要預先用cntk.exe生成模型)。CNTK-2.0.beta15.0\CNTK-2.0.beta15.0\Tutorials\HelloWorld-
LogisticRegression\Models
2 、python生成模型和使用自己的模型:
程式碼如下:
# -*- coding: utf-8 -*- """ Created on Mon Apr 10 04:59:27 2017 @author: Administrator """ from __future__ import print_function import matplotlib.pyplot as plt import numpy as np from matplotlib.colors import colorConverter, ListedColormap from cntk.learners import sgd, learning_rate_schedule, UnitType #old in learner from cntk.ops.functions import load_model from cntk.ops import * #softmax from cntk.io import CTFDeserializer, MinibatchSource, StreamDef, StreamDefs from cntk import * from cntk.layers import Dense, Sequential from cntk.logging import ProgressPrinter def generate_random_data(sample_size, feature_dim, num_classes): # Create synthetic data using NumPy. Y = np.random.randint(size=(sample_size, 1), low=0, high=num_classes) # Make sure that the data is separable X = (np.random.randn(sample_size, feature_dim) + 3) * (Y + 1) X = X.astype(np.float32) # converting class 0 into the vector "1 0 0", # class 1 into vector "0 1 0", ... class_ind = [Y == class_number for class_number in range(num_classes)] Y = np.asarray(np.hstack(class_ind), dtype=np.float32) return X, Y # Read a CTF formatted text (as mentioned above) using the CTF deserializer from a file def create_reader(path, is_training, input_dim, num_label_classes): return MinibatchSource(CTFDeserializer(path, StreamDefs( labels = StreamDef(field='labels', shape=num_label_classes, is_sparse=False), features = StreamDef(field='features', shape=input_dim, is_sparse=False) )), randomize = is_training, epoch_size = INFINITELY_REPEAT if is_training else FULL_DATA_SWEEP) def ffnet(): inputs = 2 outputs = 2 layers = 2 hidden_dimension = 50 # input variables denoting the features and label data features = input((inputs), np.float32) label = input((outputs), np.float32) # Instantiate the feedforward classification model my_model = Sequential ([ Dense(hidden_dimension, activation=sigmoid,name='d1'), Dense(outputs)]) z = my_model(features) ce = cross_entropy_with_softmax(z, label) pe = classification_error(z, label) # Instantiate the trainer object to drive the model training lr_per_minibatch = learning_rate_schedule(0.125, UnitType.minibatch) # Initialize the parameters for the reader input_dim=2 num_output_classes=2 num_samples_per_sweep = 6000 # Get minibatches of training data and perform model training minibatch_size = 25 num_minibatches_to_train = 1024 num_sweeps_to_train_with = 2#10 num_minibatches_to_train = (num_samples_per_sweep * num_sweeps_to_train_with) / minibatch_size # progress_printer = ProgressPrinter(0) progress_printer = ProgressPrinter(tag='Training',num_epochs=num_sweeps_to_train_with) trainer = Trainer(z, (ce, pe), [sgd(z.parameters, lr=lr_per_minibatch)], [progress_printer]) #trainer = Trainer(z, (ce, pe), [sgd(z.parameters, lr=lr_per_minibatch)]) train_file = "Train2-noLiner_cntk_text.txt" # Create the reader to training data set reader_train = create_reader(train_file, True, input_dim, num_output_classes) # Map the data streams to the input and labels. input_map = { label : reader_train.streams.labels, features : reader_train.streams.features } print(reader_train.streams.keys()) aggregate_loss = 0.0 #for i in range(num_minibatches_to_train): for i in range(0, int(num_minibatches_to_train)): #train_features, labels = generate_random_data(minibatch_size, inputs, outputs) # Specify the mapping of input variables in the model to actual minibatch data to be trained with #trainer.train_minibatch({features : train_features, label : labels}) # Read a mini batch from the training data file data = reader_train.next_minibatch(minibatch_size, input_map = input_map) trainer.train_minibatch(data) sample_count = trainer.previous_minibatch_sample_count aggregate_loss += trainer.previous_minibatch_loss_average * sample_count # last_avg_error = aggregate_loss / trainer.total_number_of_samples_seen trainer.summarize_training_progress() z.save_model("myfirstmod.dnn") print(z) print(z.parameters) print(z.d1) print(z.d1.signature) print(z.d1.root_function) print(z.d1.placeholders) print(z.d1.parameters) print(z.d1.op_name) print(z.d1.type) print(z.d1.output) print(z.outputs) test_features, test_labels = generate_random_data(minibatch_size, inputs, outputs) avg_error = trainer.test_minibatch({features : test_features, label : test_labels}) print(' error rate on an unseen minibatch: {}'.format(avg_error)) return last_avg_error, avg_error np.random.seed(98052) ffnet() print("-------------分割-----------------") inputs = 2 outputs = 2 minibatch_size = 5 features = input((inputs), np.float32) label = input((outputs), np.float32) test_features, test_labels = generate_random_data(minibatch_size, inputs, outputs) print('fea={}'.format(test_features)) z = load_model("myfirstmod.dnn") ce = cross_entropy_with_softmax(z, label) pe = classification_error(z, label) lr_per_minibatch = learning_rate_schedule(0.125, UnitType.minibatch) progress_printer = ProgressPrinter(0) trainer = Trainer(z, (ce, pe), [sgd(z.parameters, lr=lr_per_minibatch)], [progress_printer]) avg_error = trainer.test_minibatch({z.arguments[0] : test_features, label : test_labels}) print(' error rate on an unseen minibatch: {}'.format(avg_error)) result1 = z.eval({z.arguments[0] : test_features}) #print("r={} ".format(result1)) out = softmax(z) result = out.eval({z.arguments[0] : test_features}) print(result) print("Label :", [np.argmax(label) for label in test_labels]) print("Predicted :", [np.argmax(label) for label in result]) #print("Predicted:", [np.argmax(result[i,:,:]) for i in range(result.shape[0])]) type1_x=[] type1_y=[] type2_x=[] type2_y=[] for i in range(len(test_labels)): #for i in range(6): if np.argmax(test_labels[i]) == 0: type1_x.append( test_features[i][0] ) type1_y.append( test_features[i][1] ) if np.argmax(test_labels[i]) == 1: type2_x.append( test_features[i][0] ) type2_y.append( test_features[i][1] ) type1 = plt.scatter(type1_x, type1_y,s=40, c='red',marker='+' ) type2 = plt.scatter(type2_x, type2_y, s=40, c='green',marker='+') nb_of_xs = 100 xs1 = np.linspace(2, 8, num=nb_of_xs) xs2 = np.linspace(2, 8, num=nb_of_xs) xx, yy = np.meshgrid(xs1, xs2) # create the grid featureLine = np.vstack((np.array(xx).reshape(1,nb_of_xs*nb_of_xs),np.array(yy).reshape(1,yy.size))) print(featureLine.T) r = out.eval({z.arguments[0] : featureLine.T}) print(r) # Initialize and fill the classification plane classification_plane = np.zeros((nb_of_xs, nb_of_xs)) for i in range(nb_of_xs): for j in range(nb_of_xs): #classification_plane[i,j] = nn_predict(xx[i,j], yy[i,j]) #r = out.eval({z.arguments[0] : [xx[i,j], yy[i,j]]}) classification_plane[i,j] = np.argmax(r[i*nb_of_xs+j] ) print(classification_plane) # Create a color map to show the classification colors of each grid point cmap = ListedColormap([ colorConverter.to_rgba('r', alpha=0.30), colorConverter.to_rgba('b', alpha=0.30)]) # Plot the classification plane with decision boundary and input samples plt.contourf(xx, yy, classification_plane, cmap=cmap) plt.xlabel('x1') plt.ylabel('x2') #axes.legend((type1, type2,type3), ('0', '1','2'),loc=1) plt.show()
程式碼內容:
1先生成模型。並打印出模型裡面的引數
2呼叫模型,測試下模型錯誤率
3呼叫模型,輸出結果
4將資料視覺化
輸出:dict_keys([‘features', ‘labels'])
Finished Epoch[1 of 2]: [Training] loss = 0.485836 * 12000, metric = 20.36% *
12000 0.377s (31830.2 samples/s);
Composite(Dense): Input(‘Input456', [#], [2]) -> Output(‘Block577_Output_0',
[#], [2])
(Parameter(‘W', [], [50 x 2]), Parameter(‘b', [], [2]), Parameter(‘W', [], [2
x 50]), Parameter(‘b', [], [50]))
Dense: Input(‘Input456', [#], [2]) -> Output(‘d1', [#], [50])
(Input(‘Input456', [#], [2]),)
Dense: Input(‘Input456', [#], [2]) -> Output(‘d1', [#], [50])
()
(Parameter(‘W', [], [2 x 50]), Parameter(‘b', [], [50]))
Dense
Tensor[50]
Output(‘d1', [#], [50])
(Output(‘Block577_Output_0', [#], [2]),)
error rate on an unseen minibatch: 0.6