AttributeError: '_ThreadPoolDataset' object has no attribute 'make_initializable_iterator'

[dmyershov]

TensorFlow - v1.13.1, benchmarks - cnn_tf_v1.13_compatible.
Tried to evaluate resnet50 model by running tf_cnn_benchmarks.py in following configuration:

TensorFlow: 1.13
Model: resnet50
Dataset: imagenet
Mode: evaluation
SingleSess: False
Batch size: 1 global
1 per device
Num batches: 100
Num epochs: 0.00
Devices: ['/gpu:0']
NUMA bind: False
Data format: NCHW
Optimizer: sgd
Variables: parameter_server

Got following error (last line from call stack):
File ".../benchmarks/scripts/tf_cnn_benchmarks/preprocessing.py", line 547, in create_iterator
ds_iterator = ds.make_initializable_iterator()
AttributeError: '_ThreadPoolDataset' object has no attribute 'make_initializable_iterator'

Was able to fix it and continue evaluation by using:
ds_iterator = tf.compat.v1.data.make_initializable_iterator(ds)

[reedwm]

Can you please give the command line to reproduce?

[dmyershov]

This is from my running shell script:

export ROOT_DIR=pwd/..
export MY_PYTHON=python3
export DATA_NAME=imagenet
export DATA_DIR=/media/tools/TF_ImageNet_latest
export MODEL_NAME=${ROOT_DIR}/benchmarks/scripts/tf_cnn_benchmarks/tf_cnn_benchmarks.py
export MODEL=resnet50
export TRAIN_DIR=${ROOT_DIR}/resnet50/train_dir
export EVAL_DIR=${ROOT_DIR}/resnet50
export DATA_FORMAT=NCHW
export BATCH_SIZE=1

${MY_PYTHON}
${MODEL_NAME} --eval
--model=${MODEL}
--num_inter_threads=2
--batch_size=${BATCH_SIZE}
--data_name=${DATA_NAME}
--data_dir=${DATA_DIR}
--data_format=${DATA_FORMAT}
--eval_dir=${EVAL_DIR}
--train_dir=${TRAIN_DIR}

[reedwm]

Thank you for finding this! Do you want to send a PR to fix this by using tf.compat.v1.data.make_initializable_iterator(ds)? If not I'm happy to fix it.

[reedwm]

Also note, this is broken at HEAD as well as 1.13.

[dmyershov]

Please proceed with a fix.

Also, could you please take into account following?:
I was slightly confused by "As a last resort" clause from (TensorFlow repository) tensorflow/python/data/ops/dataset_ops.py:

@deprecation.deprecated(
None, "Use for ... in dataset: to iterate over a dataset. If using "
"tf.estimator, return the Dataset object directly from your input "
"function. As a last resort, you can use "
"tf.compat.v1.data.make_initializable_iterator(dataset).")
def make_initializable_iterator(self, shared_name=None):

Therefore, it might be not appropriate solution to use tf.compat.v1.data.make_initializable_iterator.