5

我正在尝试为我的模型制作一个评估器。到目前为止，所有其他组件都很好，但是当我尝试这个配置时：

eval_config = tfma.EvalConfig(
    model_specs=[
        tfma.ModelSpec(label_key='Category'),
    ],
    metrics_specs=tfma.metrics.default_multi_class_classification_specs(),
    slicing_specs=[
        tfma.SlicingSpec(),
        tfma.SlicingSpec(feature_keys=['Category'])
    ])

使这个评估器：

model_resolver = ResolverNode(
      instance_name='latest_blessed_model_resolver',
      resolver_class=latest_blessed_model_resolver.LatestBlessedModelResolver,
      model=Channel(type=Model),
      model_blessing=Channel(type=ModelBlessing))
context.run(model_resolver)

evaluator = Evaluator(
    examples=example_gen.outputs['examples'],
    model=trainer.outputs['model'],
    baseline_model=model_resolver.outputs['model'],
    eval_config=eval_config)
context.run(evaluator)

我明白了：

[...]
IndexError                                Traceback (most recent call last)
/opt/miniconda3/envs/archiving/lib/python3.7/site-packages/apache_beam/runners/common.cpython-37m-darwin.so in apache_beam.runners.common.DoFnRunner.process()

/opt/miniconda3/envs/archiving/lib/python3.7/site-packages/apache_beam/runners/common.cpython-37m-darwin.so in apache_beam.runners.common.PerWindowInvoker.invoke_process()

/opt/miniconda3/envs/archiving/lib/python3.7/site-packages/apache_beam/runners/common.cpython-37m-darwin.so in apache_beam.runners.common.PerWindowInvoker._invoke_process_per_window()

/opt/miniconda3/envs/archiving/lib/python3.7/site-packages/apache_beam/runners/common.cpython-37m-darwin.so in apache_beam.runners.common._OutputProcessor.process_outputs()

/opt/miniconda3/envs/archiving/lib/python3.7/site-packages/apache_beam/runners/worker/operations.cpython-37m-darwin.so in apache_beam.runners.worker.operations.SingletonConsumerSet.receive()

/opt/miniconda3/envs/archiving/lib/python3.7/site-packages/apache_beam/runners/worker/operations.cpython-37m-darwin.so in apache_beam.runners.worker.operations.PGBKCVOperation.process()

/opt/miniconda3/envs/archiving/lib/python3.7/site-packages/apache_beam/runners/worker/operations.cpython-37m-darwin.so in apache_beam.runners.worker.operations.PGBKCVOperation.process()

/opt/miniconda3/envs/archiving/lib/python3.7/site-packages/tensorflow_model_analysis/evaluators/metrics_and_plots_evaluator_v2.py in add_input(self, accumulator, element)
    355     for i, (c, a) in enumerate(zip(self._combiners, accumulator)):
--> 356       result = c.add_input(a, get_combiner_input(elements[0], i))
    357       for e in elements[1:]:

/opt/miniconda3/envs/archiving/lib/python3.7/site-packages/tensorflow_model_analysis/metrics/calibration_histogram.py in add_input(self, accumulator, element)
    141             flatten=True,
--> 142             class_weights=self._class_weights)):
    143       example_weight = float(example_weight)

/opt/miniconda3/envs/archiving/lib/python3.7/site-packages/tensorflow_model_analysis/metrics/metric_util.py in to_label_prediction_example_weight(inputs, eval_config, model_name, output_name, sub_key, class_weights, flatten, squeeze, allow_none)
    283     elif sub_key.top_k is not None:
--> 284       label, prediction = select_top_k(sub_key.top_k, label, prediction)
    285 

/opt/miniconda3/envs/archiving/lib/python3.7/site-packages/tensorflow_model_analysis/metrics/metric_util.py in select_top_k(top_k, labels, predictions, scores)
    621   if not labels.shape or labels.shape[-1] == 1:
--> 622     labels = one_hot(labels, predictions)
    623 

/opt/miniconda3/envs/archiving/lib/python3.7/site-packages/tensorflow_model_analysis/metrics/metric_util.py in one_hot(tensor, target)
    671   # indexing the -1 and then removing it after.
--> 672   tensor = np.delete(np.eye(target.shape[-1] + 1)[tensor], -1, axis=-1)
    673   return tensor.reshape(target.shape)

IndexError: arrays used as indices must be of integer (or boolean) type

During handling of the above exception, another exception occurred:
[...]

IndexError: arrays used as indices must be of integer (or boolean) type [while running 'ExtractEvaluateAndWriteResults/ExtractAndEvaluate/EvaluateMetricsAndPlots/ComputeMetricsAndPlots()/ComputePerSlice/ComputeUnsampledMetrics/CombinePerSliceKey/WindowIntoDiscarding']

我以为这是我的配置，但我不明白这有什么问题。

我正在使用这个数据集Kaggle - BBC News Classification。我关注了这个笔记本：TFX - Chicago Taxi，以便为我的模型提供 Tensorflow Serving。

注意：我使用的模型如下所示：

def _build_keras_model(vectorize_layer: TextVectorization) -> tf.keras.Model: 

  input_layer = tf.keras.layers.Input(shape=(1,), dtype=tf.string)

  deep = vectorize_layer(input_layer)
  deep = layers.Embedding(_max_features + 1, _embedding_dim)(deep)
  deep = layers.Dropout(0.5)(deep)
  deep = layers.GlobalAveragePooling1D()(deep)
  deep = layers.Dropout(0.5)(deep)

  output = layers.Dense(5, activation=tf.nn.softmax)(deep)

  model = tf.keras.Model(input_layer, output)
  model.compile(
      loss=losses.SparseCategoricalCrossentropy(from_logits=True),
      optimizer='adam', 
      metrics=['accuracy'])
  model.summary(print_fn=absl.logging.info)  
  return model
4

1 回答 1

2

我让它工作。我的问题是，在数据集中，标签（文档类别）是字符串格式（例如：“sport”、“business”、...）。因此，为了将其编码为整数，我使用了 Transform 组件对其进行预处理。

但是，在构建评估器组件时，我传递了未对数据进行任何处理的 ExampleGen 组件。因此，评估者试图从 ExampleGen 中转换字符串以匹配模型的整数输出。

所以，为了解决这个问题，我只是这样做了：

model_resolver = ResolverNode(
      instance_name='latest_blessed_model_resolver',
      resolver_class=latest_blessed_model_resolver.LatestBlessedModelResolver,
      model=Channel(type=Model),
      model_blessing=Channel(type=ModelBlessing))
context.run(model_resolver)

evaluator = Evaluator(
    examples=transform.outputs['transformed_examples'],
    model=trainer.outputs['model'],
    baseline_model=model_resolver.outputs['model'],
    eval_config=eval_config)
context.run(evaluator)

我使用了转换组件中的示例。当然，我也更改了配置中的标签键以匹配转换组件的标签名称。

我不知道是否有一种“更清洁”的方式来执行此操作（或者如果我做错了，请纠正我！）

于 2020-08-17T12:35:44.293 回答