TF1 to TF2 migration

[ishaq-ansari]

Our current project depends on the TensorFlow 1.x implementation of Deeplab v3, but the code no longer runs on our new GPU hardware/drivers, which are compatible only with TensorFlow 2.x.

We need to migrate the model architecture and training pipeline from TF1 to TF2 while preserving compatibility with existing checkpoints and functionality.

[ishaq-ansari]

There are two options Dr. @anindya-paul,

Option 1: Still using TF-Slim but in TF2 environment
Steps:

1. Remove old "tf.contrib.layers" and replace them with TF Slim symbols and TF Addons for other tf.contrib symbols.
2. Rewrite the TF1.x model forward passes to run in TF2 with eager execution enabled.
3. Validate the accuracy and numerical correctness of the migrated code.
4. Upgrade the training, evaluation and model saving code to TF2 equivalents.

Option 2: Migrate the TF2-compatible tf.compat.v1 APIs including TF Slim usage to idiomatic TF2 APIs.

• This will may need significant amount of change.

[anindya-paul]

@ishaq-ansari I want to follow the same steps as performed by Anish in IFTA migration as we have a baseline of what worked. I think he followed tf.compat APIs in TF2 environment. Review his code in detail and let me know tomorrow.