Issue creating TFrecords in TAO EfficientDet

[DarshnaZE]

I want to train efficientDet on custom dataset. The issue i am facing is with running create_coco_tf_record.py file

and with the error:

[Morganh]

Did you ever run efficientdet jupyter noteboook successfully?
https://docs.nvidia.com/tao/tao-toolkit/text/tao_toolkit_quick_start_guide.html#download-jupyter-notebook

[DarshnaZE]

Resolved the first issue in first image. but getting this now.

[DarshnaZE]

yeah. notebook run was successful.

[Morganh]

So, no other issues now, right?

[DarshnaZE]

issue is with custom dataset. the image i uploaded last.

[Morganh]

How did you run into the error? Can you share full command and full log?

[DarshnaZE]

Here is the specs

The command that i use to create TF records

the error log

[DarshnaZE]

This is what generated with create_coco_tf_record.py

[Morganh]

So, you have generated tfrecord files successfully with your own dataset?
At the top of this topic, you share below issue. Is the issue gone?

[DarshnaZE]

yeah…this issue is solved. now its the params shape issue.the error log image

[DarshnaZE]

for tfrecords i am not sure if it is the same tfrecords required by efficientdet. as its outputted with different name than in example one. i.e. here its starts with coco_train.record… and in example one it is named as train-*.tfrecord

[Morganh]

Please check if below topics help you.
https://forums.developer.nvidia.com/search?expanded=true&q=%22params%20is%20empty%22

[DarshnaZE]

My issue has been resoved. thanks for the help

[Morganh]

Could you share the experience how did you solve the error? Thanks.

[DarshnaZE]

The issue was with the Tfrecords generation. earlier i was using the create_tf_record.py from tf-models repo. changing that with the create_tf_record.py file that comes with efficientDet resolved the issue. I think the tfrecords generated from these two different files might be different. so the issue came up while parsing those tfrecords during training.

[Morganh]

Thanks for sharing!