Models: ResNet50V1 classification checkpoint not loaded when training Faster-RCNN from scratch on custom dataset

[ujjwal-ai]

I am trying to use this config file for training on a different dataset.

My tensorflow version is 2.3 (stable).

In order to use the classification checkpoint of ResNet50V1, I did the following:

import tensorflow as tf
model = tf.keras.applications.ResNet50(weights='imagenet')
checkpoint = tf.train.Checkpoint(model=model)
checkpoint.save(<CKPT_PATH>)

Then I provide the path to this checkpoint file in the config file. On running the training using model_main_tf2.py, I receive the same message as has been previously referenced in #8892 .

How can this issue be resolved ? Am I making a mistake here ?

[ujjwal-ai]

I would normally have posted this issue on StackOverFlow, but following are reasons for this being here:

1. There is no clarity about which classification checkpoints to use.
2. There is no proper example or documentation about it.
3. Multiple issues (#8892 for example and there are others) which report the same problem and have not yet been resolved.
4. This backward compatibility issue is a big mess because too much of legacy code is mixed in with the new code. It would have been better to have two branches ( one for TF1 and one for TF2). Me as a developer would have loved to contribute to achieve that.

I am against tagging this issue as support. Documentation is a quintessential part of any project. If the documentation does not mention anything and this has been released, it looks like a bug to me.

[ujjwal-ai]

What else have I tried ?
a) Tried on TF 2.2 ( DOES NOT WORK )
b) Tried using pretrained models from tfhub ( DOES NOT WORK)
c) Tried using checkpoints on google cloud ( DOES NOT WORK)

[aabbas90]

What else have I tried ?
a) Tried on TF 2.2 ( DOES NOT WORK )
b) Tried using pretrained models from tfhub ( DOES NOT WORK)
c) Tried using checkpoints on google cloud ( DOES NOT WORK)

Perhaps some discussion from this issue is helpful: https://github.com/tensorflow/models/issues/8967