TF-TRT convert to INT8 error

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

HI:
I followed TF-TRT docunment:Accelerating Inference In TF-TRT User Guide :: NVIDIA Deep Learning Frameworks Documentation, succeed convert model to INT8 with feed_dict_fn to send calibrate data, but when i use input_map_fn i get error:

Here is my platform:
Centos 7
P4
CUDA 10
Python 2.7
Tensorflow 1.15

Here is my code:

dataset = tf.data.TFRecordDataset(tf_calib_data_files)
iterator = dataset.make_one_shot_iterator()
features = iterator.get_next()
def input_map_fn():
    return {'input:0': features}

...

converted_graph_def = converter.calibrate(
    fetch_names=['detection_boxes_l1:0','detection_scores_l1:0','detection_classes_l1:0','detection_boxes_l2:0','detection_scores_l2:0','detection_classes_l2:0'],
    num_runs=10,
    input_map_fn=input_map_fn)

Then i get error:
ValueError: node ‘IteratorGetNext’ in input_map does not exist in graph (input_map entry: input:0->IteratorGetNext:0)

This error come from import_graph_def(), but why?

2

Hi,

Could you please share the model file to reproduce the issue so we can better help?

Thanks

3

Thanks for your response，
the model is just the retrained ResNet V1 50 model provided by tensorflow slim，and i converted it to pb file，download from
https://github.com/tensorflow/models/tree/r1.13.0/research/slim#pre-trained-models
it seems the model file is too large to update, i am sorry

4

I also have this issue. I cannot solve it. Maybe it is a bug for the latest version (TensorFlow container from 19.11-19.12). Waiting for the reply of the solution.

5

Please refer below support matrix:

Will recommend you to use below latest NGC container image and let us know in case issue persist.

Thanks

6

Any solution?