我從CNTK的培訓(xùn)師/進(jìn)步作家那里得到了輸出，告訴我，我的準(zhǔn)確度> 99％，而實際上是0.5％左右。根據(jù)這個指標(biāo)的確意味著損失，但是得知我不正確地使用了CNTK的訓(xùn)練器/損失功能并不奇怪，這并不會讓我感到驚訝。這是下面示例的示例輸出（與我的模型不同，但產(chǎn)生了相似的效果）：-------------------------------------------------------------------Finished Epoch[1 of 20]: [Training] loss = 2.302585 * 100, metric = 48.10% * 100 0.802s (124.7 samples/s);Accuracy % 11.0Finished Epoch[2 of 20]: [Training] loss = 2.302514 * 100, metric = 49.82% * 100 0.043s (2325.6 samples/s);Accuracy % 15.0這是一個最小的工作示例，演示了真實準(zhǔn)確度與度量標(biāo)準(zhǔn)所報告的真實度之間的差異。我編寫了一個小的精度函數(shù)來對其進(jìn)行測試，我敢肯定它可以正確實現(xiàn)。import cntk as Cimport numpy as npfrom cntk.ops import relufrom cntk.layers import Dense, Convolution2DminibatchSize = 100def printAccuracy(net, X, Y):    outs = net(X)    pred = np.argmax(Y, 1)    indx = np.argmax(outs, 1)    same = pred == indx    print("Accuracy %", np.sum(same)/minibatchSize*100)outputs = 10input_var = C.input_variable((7, 19, 19), name='features')label_var = C.input_variable((outputs))epochs = 20cc = C.layers.Convolution2D((3,3), 64, activation=relu)(input_var)net = C.layers.Dense(outputs)(cc)loss = C.cross_entropy_with_softmax(net, label_var)pe = C.classification_error(net, label_var)    learner = C.adam(net.parameters, 0.0018, 0.9, minibatch_size=minibatchSize)progressPrinter = C.logging.ProgressPrinter(tag='Training', num_epochs=epochs)trainer = C.Trainer(net, (loss, pe), learner, progressPrinter)    for i in range(epochs):    X = np.zeros((minibatchSize, 7, 19, 19), dtype=np.float32)    Y = np.random.rand(minibatchSize, outputs)    trainer.train_minibatch({input_var : X, label_var : Y})        trainer.summarize_training_progress()    printAccuracy(net, X, Y)