Question

我有一个网络，该网络具有两个头部和两个不同的输出，我想分别查看其度量/精度。 CNTK的培训师课程有可能吗？现在，我将损失和指标定义为：

policyLoss  = cntk.cross_entropy_with_softmax(net.outputs[0], policyVar)
valueLoss   = cntk.cross_entropy_with_softmax(net.outputs[1], valueVar)
loss        = policyLoss + valueLoss

policyError = cntk.element_not(cntk.classification_error(net.outputs[0], policyVar))
valueError  = cntk.element_not(cntk.classification_error(net.outputs[1], valueVar))
#error       = (valueError + policyError) / 2
 error       = valueError # Right now I make due with only viewing valueError for the training data

虽然这很适合训练，但是能够单独查看错误，以便将它们与我使用网络输出手动计算的验证集错误进行比较，将是非常不错的。

我目前正在使用其内置的培训师来像这样训练批次：

progressPrinter = cntk.logging.ProgressPrinter(tag='Training', num_epochs=maxEpochs)   
trainer         = cntk.Trainer(net, (loss, error), learner, progressPrinter)

for epoch in range(maxEpochs):

    miniBatches = 0
    while miniBatches < gen.stepsPerEpoch:
        X, Y, W = next(g)
        miniBatches += 1 # TODO: NEED to make sure this doesn't go over minibatchSize so we're not inputting more than we're saying we are
        trainer.train_minibatch({net.arguments[0] : X, policyVar : Y, valueVar : W}) 

    trainer.summarize_training_progress()

是否有办法同时获取policyVar的度量和valueVar的度量以分别显示？我现在还手动计算验证准确性，因为我想同时查看valueVar和policyVar。是否可以用CNTK做到这一点？

当前输出：

Finished Epoch[1 of 100]: [Training] loss = 3.010285 * 9984, metric = 59.91% * 9984 29.532s (338.1 samples/s);

所需的输出（或类似的内容）：

Finished Epoch[1 of 100]: [Training] loss = 3.010285 * 9984, (policy)metric = 59.91% * 9984 29.532s (338.1 samples/s);
Finished Epoch[1 of 100]: [Training] loss = 3.010285 * 9984, (value )metric = 63.91% * 9984 29.532s (338.1 samples/s);

显示两个指标CNTK网络？

0 个答案: