时间戳

首页 > 解决方案 > ValueError: input_fn 返回的模型特征必须具有静态形状，使用 TPU Estimator API

python - ValueError: input_fn 返回的模型特征必须具有静态形状，使用 TPU Estimator API

问题描述

我正在尝试使用 Cloud TPU 上的 TPU Estimator API 训练模型。错误日志和读取我的输入数据的代码附在下面。我尝试使用 python 调试器来确定遇到错误的位置。traing_input_fn在遇到错误之前，控件不会退出功能。所以，我相信我的数据管道是问题的根源。有人可以帮我解决这个问题吗？如有必要，我很乐意提供更多信息。谢谢

INFO:tensorflow:Error recorded from training_loop: The features to the model returned by input_fn must have static shape. Tensor: Tensor("Inf[25/1805]
dequeue:0", shape=(16, ?, 50, 1024), dtype=float32, device=/device:TPU_REPLICATED_CORE:0)
INFO:tensorflow:training_loop marked as finished
WARNING:tensorflow:Reraising captured error
Traceback (most recent call last):
  File "estimator_task.py", line 303, in <module>
    main(**arguments)
  File "estimator_task.py", line 261, in main
    estimator.train(input_fn=train_input_fn, max_steps=train_steps, hooks=hooks)
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/contrib/tpu/python/tpu/tpu_estimator.py", line 2457, in train
    rendezvous.raise_errors()
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/contrib/tpu/python/tpu/error_handling.py", line 128, in raise_error$
    six.reraise(typ, value, traceback)
  File "/home/abi/.local/lib/python3.5/site-packages/six.py", line 693, in reraise
    raise value
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/contrib/tpu/python/tpu/tpu_estimator.py", line 2452, in train
    saving_listeners=saving_listeners)
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow_estimator/python/estimator/estimator.py", line 358, in train
    loss = self._train_model(input_fn, hooks, saving_listeners)
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow_estimator/python/estimator/estimator.py", line 1124, in _train_mode$
    return self._train_model_default(input_fn, hooks, saving_listeners)
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow_estimator/python/estimator/estimator.py", line 1154, in _train_mode$
_default
    features, labels, model_fn_lib.ModeKeys.TRAIN, self.config)
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/contrib/tpu/python/tpu/tpu_estimator.py", line 2251, in _call_model$
fn
    config)
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow_estimator/python/estimator/estimator.py", line 1112, in _call_model$
fn
    model_fn_results = self._model_fn(features=features, **kwargs)
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/contrib/tpu/python/tpu/tpu_estimator.py", line 2558, in _model_fn
    _train_on_tpu_system(ctx, model_fn_wrapper, dequeue_fn))
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/contrib/tpu/python/tpu/tpu_estimator.py", line 2893, in _train_on_t$
u_system
    device_assignment=ctx.device_assignment)
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/contrib/tpu/python/tpu/tpu.py", line 890, in split_compile_and_shar$
    name=name)
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/contrib/tpu/python/tpu/tpu.py", line 689, in split_compile_and_repl$
cate
    outputs = computation(*computation_inputs)
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/contrib/tpu/python/tpu/tpu_estimator.py", line 2886, in multi_tpu_t$
ain_steps_on_single_shard
    [_INITIAL_LOSS])
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/contrib/tpu/python/tpu/training_loop.py", line 208, in repeat
    cond, body_wrapper, inputs=inputs, infeed_queue=infeed_queue, name=name)
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/contrib/tpu/python/tpu/training_loop.py", line 170, in while_loop
    condition_wrapper, body_wrapper, inputs, name="", parallel_iterations=1)
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/python/ops/control_flow_ops.py", line 3556, in while_loop
    return_same_structure)
File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/python/ops/control_flow_ops.py", line 3087, in BuildLoop
    pred, body, original_loop_vars, loop_vars, shape_invariants)
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/python/ops/control_flow_ops.py", line 3022, in _BuildLoop
    body_result = body(*packed_vars_for_body)
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/contrib/tpu/python/tpu/training_loop.py", line 121, in body_wrapper
    outputs = body(*(inputs + dequeue_ops))
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/contrib/tpu/python/tpu/training_loop.py", line 204, in body_wrapper
    return [i + 1] + _convert_to_list(body(*args))
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/contrib/tpu/python/tpu/tpu_estimator.py", line 1359, in train_step
    self._call_model_fn(features, labels))
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/contrib/tpu/python/tpu/tpu_estimator.py", line 1552, in _call_model_
fn
    self._validate_model_features_and_labels(features, labels, is_export_mode)
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/contrib/tpu/python/tpu/tpu_estimator.py", line 1546, in _validate_mo
del_features_and_labels
    validate(features, 'features')
  File "/home/abi/anaconda3/envs/myenv_3_5/lib/python3.5/site-packages/tensorflow/contrib/tpu/python/tpu/tpu_estimator.py", line 1538, in validate
    ' Tensor: {}'.format(obj_name, obj))
ValueError: The features to the model returned by input_fn must have static shape. Tensor: Tensor("InfeedQueue/dequeue:0", shape=(16, ?, 50, 1024), dt
ype=float32, device=/device:TPU_REPLICATED_CORE:0)

这是我的训练数据管道

def train_input_fn(params):


    def decode_example(example_proto, t=50, dim=1024):

        features = tf.parse_single_example(
          example_proto,
          features = {
            'X': tf.FixedLenSequenceFeature([], tf.float32, allow_missing=True),
            'Y': tf.FixedLenSequenceFeature([], tf.int64, allow_missing=True),
          }
        )

        feat = features['X']
        feat = tf.squeeze(feat)
        feat.set_shape([t, dim])
        labels = features['Y']
        labels = tf.cast(labels, dtype=tf.int32)

        return feat, labels

    train_files = params["train_filenames"]
    batch_size = params['batch_size']

    dataset = tf.data.TFRecordDataset(train_files, num_parallel_reads=8)
    dataset = dataset.apply(
            tf.contrib.data.shuffle_and_repeat(buffer_size=100))
    dataset = dataset.apply(
            tf.contrib.data.map_and_batch(decode_example, batch_size, drop_remainder=False))

    dataset = dataset.prefetch(1)
    dataset = dataset.apply(tf.contrib.data.batch_and_drop_remainder(batch_size))
    return dataset

标签： pythontensorflowmachine-learningdeep-learning

解决方案

我有一个非常相似的问题，我通过使用显式值设置形状尺寸来解决它，在你的情况下：

feat.set_shape([50, 1024])

不是很方便，但它对我有用。

推荐阅读