3

我在将自定义数据集加载到 pylearn2 时遇到问题。我正在尝试使用一个小的 XOR 数据集来训练一个简单的 MLP。我有一个与我的 yaml 文件在同一目录中命名的数据集，它与 pylearn2 的脚本xor.csv不在同一目录中。train.py

以下是 的全部内容xor.csv：

label,x,y
0,0,0
1,0,1
1,1,0
0,1,1

这是我的 YAML 文件的全部内容：

!obj:pylearn2.train.Train {
    dataset: &train !obj:pylearn2.datasets.csv_dataset.CSVDataset {
        path: 'xor.csv',
        task: 'classification'
    },
    model: !obj:pylearn2.models.mlp.MLP {
        layers: [
                 !obj:pylearn2.models.mlp.Sigmoid {
                     layer_name: 'h0',
                     dim: 10,
                     irange: 0.05,
                 },

                 !obj:pylearn2.models.mlp.Softmax {
                     layer_name: 'y',
                     n_classes: 1,
                     irange: 0.
                 }
                ],
        nvis: 2,
    },
    algorithm: !obj:pylearn2.training_algorithms.sgd.SGD {
        learning_rate: 1e-2,
        batch_size: 1,
        monitoring_dataset:
            {
                'train' : *train
            },
        termination_criterion: 
                !obj:pylearn2.termination_criteria.EpochCounter {
                    max_epochs: 10000
                },
    },
    extensions: [
        !obj:pylearn2.train_extensions.best_params.MonitorBasedSaveBest {
             channel_name: 'valid_y_misclass',
             save_path: "best.pkl"
        },
    ]
}

当我运行 pylearn2 的train.py脚本时，它在训练之前失败（大概是在编译 theano 函数时）。这是输出的全部内容：

[COMPUTER_NAME]:some_folder [MY_NAME]$ python [PATH_TO_PYLEARN2_SCRIPTS]/train.py example_mlp.yml 
/Users/[MY_NAME]/anaconda/lib/python2.7/site-packages/nose/plugins/manager.py:418: UserWarning: Module argparse was already imported from /Users/[MY_NAME]/anaconda/lib/python2.7/argparse.pyc, but /Users/[MY_NAME]/anaconda/lib/python2.7/site-packages is being added to sys.path
  import pkg_resources
Parameter and initial learning rate summary:
    h0_W: 0.01
    h0_b: 0.01
    softmax_b: 0.01
    softmax_W: 0.01
Compiling sgd_update...
Compiling sgd_update done. Time elapsed: 1.109511 seconds
compiling begin_record_entry...
compiling begin_record_entry done. Time elapsed: 0.090133 seconds
Monitored channels: 
    learning_rate
    total_seconds_last_epoch
    train_h0_col_norms_max
    train_h0_col_norms_mean
    train_h0_col_norms_min
    train_h0_max_x_max_u
    train_h0_max_x_mean_u
    train_h0_max_x_min_u
    train_h0_mean_x_max_u
    train_h0_mean_x_mean_u
    train_h0_mean_x_min_u
    train_h0_min_x_max_u
    train_h0_min_x_mean_u
    train_h0_min_x_min_u
    train_h0_range_x_max_u
    train_h0_range_x_mean_u
    train_h0_range_x_min_u
    train_h0_row_norms_max
    train_h0_row_norms_mean
    train_h0_row_norms_min
    train_objective
    train_y_col_norms_max
    train_y_col_norms_mean
    train_y_col_norms_min
    train_y_max_max_class
    train_y_mean_max_class
    train_y_min_max_class
    train_y_misclass
    train_y_nll
    train_y_row_norms_max
    train_y_row_norms_mean
    train_y_row_norms_min
    training_seconds_this_epoch
Compiling accum...
graph size: 115
Compiling accum done. Time elapsed: 1.647879 seconds
Traceback (most recent call last):
  File "/Users/[MY_NAME]/pylearn2/pylearn2/scripts/train.py", line 252, in <module>
    args.verbose_logging, args.debug)
  File "/Users/[MY_NAME]/pylearn2/pylearn2/scripts/train.py", line 242, in train
    train_obj.main_loop(time_budget=time_budget)
  File "/Users/[MY_NAME]/pylearn2/pylearn2/train.py", line 196, in main_loop
    self.run_callbacks_and_monitoring()
  File "/Users/[MY_NAME]/pylearn2/pylearn2/train.py", line 242, in run_callbacks_and_monitoring
    self.model.monitor()
  File "/Users/[MY_NAME]/pylearn2/pylearn2/monitor.py", line 254, in __call__
    for X in myiterator:
  File "/Users/[MY_NAME]/pylearn2/pylearn2/utils/iteration.py", line 859, in next
    for data, fn in safe_izip(self._raw_data, self._convert))
  File "/Users/[MY_NAME]/pylearn2/pylearn2/utils/iteration.py", line 859, in <genexpr>
    for data, fn in safe_izip(self._raw_data, self._convert))
  File "/Users/[MY_NAME]/pylearn2/pylearn2/utils/iteration.py", line 819, in fn
    return dspace.np_format_as(batch, sp)
  File "/Users/[MY_NAME]/pylearn2/pylearn2/space/__init__.py", line 458, in np_format_as
    space=space)
  File "/Users/[MY_NAME]/pylearn2/pylearn2/space/__init__.py", line 513, in _format_as
    self._validate(is_numeric, batch)
  File "/Users/[MY_NAME]/pylearn2/pylearn2/space/__init__.py", line 617, in _validate
    self._validate_impl(is_numeric, batch)
  File "/Users/[MY_NAME]/pylearn2/pylearn2/space/__init__.py", line 984, in _validate_impl
    super(IndexSpace, self)._validate_impl(is_numeric, batch)
  File "/Users/[MY_NAME]/pylearn2/pylearn2/space/__init__.py", line 796, in _validate_impl
    (batch.dtype, self.dtype))
TypeError: Cannot safely cast batch dtype float64 to space's dtype int64.

这到底是什么意思？我查看了 的代码CSVDataset，它在 using 中加载了数据np.loadtxt，这应该将它们作为浮点数引入。如果我编辑xor.csv为看起来像浮动（1 -> 1.0例如），则没有任何变化。

4

1 回答 1

6

这是因为 CSVDataset 的 y 属性类型设置为 float64。
我已经修复了 csv_dataset.py 的 __init__() 如下，它可以工作。
我不知道这是否是 pylearn2 的问题。

    if self.task == 'regression':
        super(CSVDataset, self).__init__(X=X, y=y)
    else:
        super(CSVDataset, self).__init__(X=X, y=y.astype(int),
                                         y_labels=np.max(y) + 1)

顺便说一句，你应该修复你的 yaml

  • Softmax 层的 n_classes 应该是 2
  • “channel_name: 'valid_y_misclass'” 会导致错误，因为您没有设置 monitoring_dataset 的“valid”属性。
    尝试将监控数据集设置为“有效”，或改用“train_y_misclass”。
于 2014-12-23T13:43:04.897 回答