我在尝试使用Tensorflow中的tf.contrib.training.stratified_sample来平衡类别。我做了一个简单的示例来测试它，从两个不平衡的类别中以平衡的方式抽取样本并进行验证，但却遇到了错误。

import tensorflow as tffrom tensorflow.python.framework import opsfrom tensorflow.python.framework import dtypesbatch_size = 10data = ['a']*9990+['b']*10labels = [1]*9990+[0]*10data_tensor = ops.convert_to_tensor(data, dtype=dtypes.string)label_tensor = ops.convert_to_tensor(labels)target_probs = [0.5,0.5]data_batch, label_batch = tf.contrib.training.stratified_sample(    data_tensor, label_tensor, target_probs, batch_size,    queue_capacity=2*batch_size)with tf.Session() as sess:    d,l = sess.run(data_batch,label_batch)print('percentage "a" = %.3f' % (np.sum(l)/len(l)))

我得到的错误是：

Traceback (most recent call last):   File "/home/jason/code/scrap.py", line 56, in <module>    test_stratified_sample()   File "/home/jason/code/scrap.py", line 47, in test_stratified_sample    queue_capacity=2*batch_size)   File "/usr/local/lib/python3.4/dist-packages/tensorflow/contrib/training/python/training/sampling_ops.py", line 191, in stratified_sample    with ops.name_scope(name, 'stratified_sample', tensors + [labels]):   File "/usr/local/lib/python3.4/dist-packages/tensorflow/python/ops/math_ops.py", line 829, in binary_op_wrapper    y = ops.convert_to_tensor(y, dtype=x.dtype.base_dtype, name="y")   File "/usr/local/lib/python3.4/dist-packages/tensorflow/python/framework/ops.py", line 676, in convert_to_tensor    as_ref=False)   File "/usr/local/lib/python3.4/dist-packages/tensorflow/python/framework/ops.py", line 741, in internal_convert_to_tensor    ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)   File "/usr/local/lib/python3.4/dist-packages/tensorflow/python/framework/constant_op.py", line 113, in _constant_tensor_conversion_function    return constant(v, dtype=dtype, name=name)   File "/usr/local/lib/python3.4/dist-packages/tensorflow/python/framework/constant_op.py", line 102, in constant    tensor_util.make_tensor_proto(value, dtype=dtype, shape=shape, verify_shape=verify_shape))   File "/usr/local/lib/python3.4/dist-packages/tensorflow/python/framework/tensor_util.py", line 374, in make_tensor_proto    _AssertCompatible(values, dtype)   File "/usr/local/lib/python3.4/dist-packages/tensorflow/python/framework/tensor_util.py", line 302, in _AssertCompatible    (dtype.name, repr(mismatch), type(mismatch).__name__)) TypeError: Expected string, got list containing Tensors of type '_Message' instead.

错误信息并没有解释我做错了什么。我还尝试过直接输入原始数据和标签（不转换为张量），以及使用tf.train.slice_input_producer来创建数据和标签张量的初始队列。

有没有人成功使用过stratified_sample？我没能找到任何示例。

回答：

我已经修改了代码，使其对我来说可以工作。修改的总结如下：

• 使用enqueue_many=True来一次性加入一批具有不同标签的示例。否则，它期望一个单一的标量标签张量（这在由队列运行器评估时可能是随机的）。
• 第一个参数期望是一个张量列表。它应该有更好的错误信息（我想这就是你遇到的问题）。请提交一个拉取请求或在Github上打开一个问题来改进错误信息。
• 启动队列运行器。否则使用队列的代码会死锁。或者使用Estimator或MonitoredSession，这样你就不需要担心这个问题了。
• （基于评论的编辑）stratified_sample不打乱数据，它只是接受/拒绝！所以如果你的数据没有随机化，考虑在抽样前通过slice_input_producer（enqueue_many=False）或shuffle_batch（enqueue_many=True）来随机化数据，如果你希望它以随机顺序输出的话。

修改后的代码（基于@人名的评论改进）：

import numpyimport tensorflow as tffrom tensorflow.python.framework import opsfrom tensorflow.python.framework import dtypeswith tf.Graph().as_default():  batch_size = 100  data = ['a']*9000+['b']*1000  labels = [1]*9000+[0]*1000  data_tensor = ops.convert_to_tensor(data, dtype=dtypes.string)  label_tensor = ops.convert_to_tensor(labels, dtype=dtypes.int32)  shuffled_data, shuffled_labels = tf.train.slice_input_producer(      [data_tensor, label_tensor], shuffle=True, capacity=3*batch_size)  target_probs = numpy.array([0.5,0.5])  data_batch, label_batch = tf.contrib.training.stratified_sample(      [shuffled_data], shuffled_labels, target_probs, batch_size,      queue_capacity=2*batch_size)  with tf.Session() as session:    tf.local_variables_initializer().run()    tf.global_variables_initializer().run()    coordinator = tf.train.Coordinator()    tf.train.start_queue_runners(session, coord=coordinator)    num_iter = 10    sum_ones = 0.    for _ in range(num_iter):      d, l = session.run([data_batch, label_batch])      count_ones = l.sum()      sum_ones += float(count_ones)      print('percentage "a" = %.3f' % (float(count_ones) / len(l)))    print('Overall: {}'.format(sum_ones / (num_iter * batch_size)))    coordinator.request_stop()    coordinator.join()

输出结果：

percentage "a" = 0.480percentage "a" = 0.440percentage "a" = 0.580percentage "a" = 0.570percentage "a" = 0.580percentage "a" = 0.520percentage "a" = 0.480percentage "a" = 0.460percentage "a" = 0.390percentage "a" = 0.530Overall: 0.503