尝试使用Cloud TPU还原更新的BERT模型检查点时出现InfeedEnqueueTuple问题

时间:2018-11-16 08:07:54

标签: python tensorflow google-colaboratory google-cloud-tpu

感谢您对以下内容的任何帮助,在此先感谢您。我制作了Google Bert's notebook on fine-tuning的副本,并使用Cloud TPU和Bucket在其上训练了SQUAD数据集。关于开发集的预测是可以的,因此我在本地下载了检查点,model.ckpt.meta,model.ckpt.index和model.ckpt.data文件,并尝试使用代码进行还原:

sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
saver = tf.train.import_meta_graph(META_FILE) # META_FILE being path to .meta
saver.restore(sess, 'model.ckpt')

但是,我得到了错误:

    op_def = op_dict[node.op]
KeyError: 'InfeedEnqueueTuple'

我认为它是Cloud TPU Tools的一部分,我应该继续使用Cloud TPU,所以我尝试了以下(reference):

# code from cells before includes
...
tf.contrib.cloud.configure_gcs(session, credentials=auth_info)
...
tpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver(TPU_ADDRESS)
run_config = tf.contrib.tpu.RunConfig(
    cluster=tpu_cluster_resolver,
    model_dir=OUTPUT_DIR,
    save_checkpoints_steps=SAVE_CHECKPOINTS_STEPS,
    tpu_config=tf.contrib.tpu.TPUConfig(
        iterations_per_loop=ITERATIONS_PER_LOOP,
        num_shards=NUM_TPU_CORES,
        per_host_input_for_training=tf.contrib.tpu.InputPipelineConfig.PER_HOST_V2))
...

问题单元格:

"""
# not valid checkpoint error. <bucket> placeholder for cloud bucket name
sess = tf.Session()
META_FILE = "gs://<bucket>/bert/models/bertsquad/model.ckpt-10949.meta"
CKPT_FILE = "gs://<bucket>/bert/models/bertsquad/model.ckpt"
saver = tf.train.import_meta_graph(META_FILE)
saver.restore(sess, CKPT_FILE)
"""

from google.cloud import storage
from tensorflow import MetaGraphDef

client = storage.Client(project="agent-helper-4a014")
bucket = client.get_bucket(<bucket>)
metafile = "bert/models/bertsquad/model.ckpt-10949.meta"
# using full path gs://<bucket>/bert/models/bertsquad doesn't work

blob = bucket.get_blob(metafile)
#blob = bucket.blob(metafile)
#model_graph = blob.download_to_filename("model.ckpt")
model_graph = blob.download_as_string()

mgd = MetaGraphDef()
mgd.ParseFromString(model_graph)

with tf.Session() as sess:
    saver = tf.train.import_meta_graph(mgd, clear_devices=True)
    init_checkpoint = saver.restore(sess, 'model.ckpt')

这又导致了以下错误:

InvalidArgumentError (see above for traceback): Restoring from checkpoint failed. This is most likely due to a mismatch between the current graph and the graph from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint. Original error:

No OpKernel was registered to support Op 'InfeedEnqueueTuple' with these attrs.  Registered devices: [CPU,XLA_CPU], Registered kernels:
  <no registered kernels>

     [[node input_pipeline_task0/while/InfeedQueue/enqueue/0 (defined at <ipython-input-67-e4b52b7b5944>:21)  = InfeedEnqueueTuple[_class=["loc:@input_pipeline_task0/while/IteratorGetNext"], device_ordinal=0, dtypes=[DT_INT32, DT_INT32, DT_INT32, DT_INT32, DT_INT32, DT_INT32], shapes=[[2], [2,384], [2,384], [2,384], [2], [2]], _device="/job:worker/task:0/device:CPU:0"](input_pipeline_task0/while/IteratorGetNext, input_pipeline_task0/while/IteratorGetNext:1, input_pipeline_task0/while/IteratorGetNext:2, input_pipeline_task0/while/IteratorGetNext:3, input_pipeline_task0/while/IteratorGetNext:4, input_pipeline_task0/while/IteratorGetNext:5)]]

1 个答案:

答案 0 :(得分:0)

如果您的动机是预测,则只需指定model_dir位置(必须是GCS存储桶)即可,其中保存了检查点和元文件。该代码将不再进行训练(因为检查点已保存为训练步骤数,并且模型图中没有任何更改)。它将直接跳至预测。

但是,如果您的用例确实要保存检查点,并且仅出于推理而将其还原,则请执行以下步骤:

  • 与原始模型一样,为每个图层手动创建模型网络,或者使用保存的.meta文件通过tf.train.import()函数来重新创建网络,如下所示:

saver = tf.train.import_meta_graph('<filename>.meta')

  • 现在,使用以下方法恢复检查点:saver.restore(sess, 'model.ckpt')

注意:要将检查点还原到的模型图,应与保存这些检查点的原始图完全相同。

希望这可以解决您的问题。