如何从minibatch获得标签?

时间:2017-01-05 09:31:01

标签: cntk

我正在编写本教程:

https://github.com/Microsoft/CNTK/blob/master/Tutorials/CNTK_201B_CIFAR-10_ImageHandsOn.ipynb

测试/训练数据文件是简单的制表符分隔文本文件,包含图像文件名和正确的标签,如下所示:

...\data\CIFAR-10\test\00000.png    3
...\data\CIFAR-10\test\00001.png    8
...\data\CIFAR-10\test\00002.png    8

如何从小批量中提取原始标签?

我尝试过这段代码:

reader_test = MinibatchSource(ImageDeserializer('test_map.txt', StreamDefs(
    features = StreamDef(field='image', transforms=transforms), # first column in map file is referred to as 'image'
    labels   = StreamDef(field='label', shape=num_classes)      # and second as 'label'
)))

test_minibatch = reader_test.next_minibatch(10)
labels_stream_info = reader_test['labels']
orig_label = test_minibatch[labels_stream_info].value
print(orig_label)

<cntk.cntk_py.Value; proxy of <Swig Object of type 'CNTK::ValuePtr *' at 0x0000000007A32C00> >

但是,如上所示,结果不是带有标签的数组。

获取标签的正确代码是什么?

此代码有效,但它使用的是不同的文件格式,而不是ImageDeserializer。

文件格式:

|labels 0 0 1 0 0 0 |features 0
|labels 1 0 0 0 0 0 |features 457

工作代码:

mb_source = text_format_minibatch_source('test_map2.txt', [
    StreamConfiguration('features', 1),
    StreamConfiguration('labels', num_classes)])

test_minibatch = mb_source.next_minibatch(2)

labels_stream_info = mb_source['labels']
orig_label = test_minibatch[labels_stream_info].value
print(orig_label)

[[[ 0.  0.  1.  0.  0.  0.]]
 [[ 1.  0.  0.  0.  0.  0.]]]

使用ImageDeserializer时,如何获取输入中的标签?

2 个答案:

答案 0 :(得分:2)

您可以尝试使用:

orig_label = test_minibatch[labels_stream_info].value

答案 1 :(得分:1)

我只是试图重复 - 我认为这里潜伏着一些奇怪的错误。我的预感是,实际上labels对象不会作为有效的numpy数组返回。我将以下调试输出插入到教程train_and_evaluate中的CNTK_201B函数中:

for epoch in range(max_epochs):       # loop over epochs
    sample_count = 0
    while sample_count < epoch_size:  # loop over minibatches in the epoch
        data = reader_train.next_minibatch(min(minibatch_size, epoch_size - sample_count), input_map=input_map) # fetch minibatch.
        print("Features:")
        print(data[input_var].shape)
        print(data[input_var].value.shape)
        print("Labels:")
        print(data[label_var].shape)
        print(data[label_var].value.shape)

输出:

Training 116906 parameters in 10 parameter tensors.
Features:
(64, 1, 3, 32, 32)
(64, 1, 3, 32, 32)
Labels:
(64, 1, 10)
()

标签显示为numpy.ndarray,但没有有效的shape

我称之为一个错误。