cloudml reconining inception - 收到有效范围之外的标签值

Question

我正在关注花朵tutorial，以便在google cloud ml上重新训练。我可以运行教程，训练，预测，就好了。

然后我用花数据集代替了我自己的测试数据集。图像数字的光学字符识别。

训练模型时，我收到错误：

Invalid argument: Received a label value of 13 which is outside the valid range of [0, 6). Label values: 6 3 2 7 3 7 6 6 12 6 5 2 3 6 8 8 8 8 4 6 5 13 7 4 8 12 5 2 4 12 12 8 8 8 12 6 4 2 12 4 3 8 2 6 8 12 2 8 4 6 2 4 12 5 5 7 6 2 2 3 2 8 2 5 2 8 2 7 4 12 8 4 2 4 8 2 2 8 2 8 7 6 8 3 5 5 5 8 8 2 5 3 9 8 5 8 3 2 5 4

训练和评估数据集的格式如下所示：

root@e925cd9502c0:~/MeerkatReader/cloudML# head training_dataGCS.csv
gs://api-project-773889352370-ml/TrainingData/0_2.jpg,H
gs://api-project-773889352370-ml/TrainingData/0_4.jpg,One
gs://api-project-773889352370-ml/TrainingData/0_5.jpg,Five

dict文件看起来像这样

$ cat cloudML/dict.txt
Eight
F
Five
Forward_slash
Four
H
Nine
One
Seven
Six
Three
Two
Zero

我最初有1,2,3,4和/的标签，但我将它们更改为字符串，以防它们是特殊字符（尤其是/）。我可以看到一个类似的消息here，但这与基于0的索引有关。

该消息的奇怪之处在于确实有13种标签类型。不知何故，tensorflow只寻找7（0-6）。我的问题是什么样的格式错误可能使张量流认为有更少的标签然后有。我可以确认80-20分割的测试和训练数据都具有所有标签类别（尽管频率不同）。

我正在使用google提供的最近的docker build。

`docker run -it -p "127.0.0.1:8080:8080" --entrypoint=/bin/bash  gcr.io/cloud-datalab/datalab:local-20161227

我使用

提交培训工作

  # Submit training job.
gcloud beta ml jobs submit training "$JOB_ID" \
  --module-name trainer.task \
  --package-path trainer \
  --staging-bucket "$BUCKET" \
  --region us-central1 \
  -- \
  --output_path "${GCS_PATH}/training" \
  --eval_data_paths "${GCS_PATH}/preproc/eval*" \
  --train_data_paths "${GCS_PATH}/preproc/train*"

完整错误：

Error reported to Coordinator: <class 'tensorflow.python.framework.errors_impl.InvalidArgumentError'>, Received a label value of 13 which is outside the valid range of [0, 6). Label values: 6 3 2 7 3 7 6 6 12 6 5 2 3 6 8 8 8 8 4 6 5 13 7 4 8 12 5 2 4 12 12 8 8 8 12 6 4 2 12 4 3 8 2 6 8 12 2 8 4 6 2 4 12 5 5 7 6 2 2 3 2 8 2 5 2 8 2 7 4 12 8 4 2 4 8 2 2 8 2 8 7 6 8 3 5 5 5 8 8 2 5 3 9 8 5 8 3 2 5 4 [[Node: evaluate/xentropy/xentropy = SparseSoftmaxCrossEntropyWithLogits[T=DT_FLOAT, Tlabels=DT_INT64, _device="/job:master/replica:0/task:0/cpu:0"](final_ops/input/Wx_plus_b/fully_connected_1/BiasAdd, inputs/Squeeze)]] Caused by op u'evaluate/xentropy/xentropy', defined at: File "/usr/lib/python2.7/runpy.py", line 162, in _run_module_as_main "__main__", fname, loader, pkg_name) File "/usr/lib/python2.7/runpy.py", line 72, in _run_code exec code in run_globals File "/root/.local/lib/python2.7/site-packages/trainer/task.py", line 545, in <module> tf.app.run() File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/platform/app.py", line 43, in run sys.exit(main(sys.argv[:1] + flags_passthrough)) File "/root/.local/lib/python2.7/site-packages/trainer/task.py", line 308, in main run(model, argv) File "/root/.local/lib/python2.7/site-packages/trainer/task.py", line 439, in run dispatch(args, model, cluster, task) File "/root/.local/lib/python2.7/site-packages/trainer/task.py", line 480, in dispatch Trainer(args, model, cluster, task).run_training() File "/root/.local/lib/python2.7/site-packages/trainer/task.py", line 187, in run_training self.args.batch_size) File "/root/.local/lib/python2.7/site-packages/trainer/model.py", line 278, in build_train_graph return self.build_graph(data_paths, batch_size, GraphMod.TRAIN) File "/root/.local/lib/python2.7/site-packages/trainer/model.py", line 256, in build_graph loss_value = loss(logits, labels) File "/root/.local/lib/python2.7/site-packages/trainer/model.py", line 396, in loss logits, labels, name='xentropy') File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/nn_ops.py", line 1544, in sparse_softmax_cross_entropy_with_logits precise_logits, labels, name=name) File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/gen_nn_ops.py", line 2376, in _sparse_softmax_cross_entropy_with_logits features=features, labels=labels, name=name) File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/op_def_library.py", line 759, in apply_op op_def=op_def) File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 2238, in create_op original_op=self._default_original_op, op_def=op_def) File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 1130, in __init__ self._traceback = _extract_stack() InvalidArgumentError (see above for traceback): Received a label value of 13 which is outside the valid range of [0, 6). Label values: 6 3 2 7 3 7 6 6 12 6 5 2 3 6 8 8 8 8 4 6 5 13 7 4 8 12 5 2 4 12 12 8 8 8 12 6 4 2 12 4 3 8 2 6 8 12 2 8 4 6 2 4 12 5 5 7 6 2 2 3 2 8 2 5 2 8 2 7 4 12 8 4 2 4 8 2 2 8 2 8 7 6 8 3 5 5 5 8 8 2 5 3 9 8 5 8 3 2 5 4 [[Node: evaluate/xentropy/xentropy = SparseSoftmaxCrossEntropyWithLogits[T=DT_FLOAT, Tlabels=DT_INT64, _device="/job:master/replica:0/task:0/cpu:0"](final_ops/input/Wx_plus_b/fully_connected_1/BiasAdd, inputs/Squeeze)]]

我的桶里的一切看起来还不错

保存我的日志事件。

Answer 1

我认为您在提交培训工作时需要指定--label_count 13。该标志应该在--之后的第二组标志之后，因为它需要传递给您正在执行的代码，而不是传递给gcloud / Cloud ML。

问题在于TensorFlow培训代码需要知道在开始逐步执​​行数据之前要进行多少输出记录;所以它无法检查预处理步骤中的中间文件。

如果有帮助，请告诉我。

Answer 2

特定于 - 模型之后的标志，并在model.py文件中指定：

def create_model():
  """Factory method that creates model to be used by generic task.py."""
  parser = argparse.ArgumentParser()
  # Label count needs to correspond to nubmer of labels in dictionary used
  # during preprocessing.
  parser.add_argument('--label_count', type=int, default=5)
  parser.add_argument('--dropout', type=float, default=0.5)
  parser.add_argument(
      '--inception_checkpoint_file',
      type=str,
      default=DEFAULT_INCEPTION_CHECKPOINT)
  args, task_args = parser.parse_known_args()
  override_if_not_in_args('--max_steps', '1000', task_args)
  override_if_not_in_args('--batch_size', '100', task_args)
  override_if_not_in_args('--eval_set_size', '370', task_args)
  override_if_not_in_args('--eval_interval_secs', '2', task_args)
  override_if_not_in_args('--log_interval_secs', '2', task_args)
  override_if_not_in_args('--min_train_eval_rate', '2', task_args)
  return Model(args.label_count, args.dropout,
               args.inception_checkpoint_file), task_args

请注意，您可以更改label_count，dropout，max_steps等内容，以影响模型培训。

HTH。