DeeplabV3 MobileNetV2 Quantization for USB Coral TPU with custom input size ((100, 100) rather than (513, 513))

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): No
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): MacOS
• Mobile device (e.g., Pixel 4, Samsung Galaxy 10) if the issue happens on mobile device: Coral TPU
• TensorFlow installed from (source or binary): pip install tensorflow==1.15.0
• TensorFlow version (use command below):1.15.0
• Python version:3.7
• Bazel version (if compiling from source):
• GCC/Compiler version (if compiling from source):
• CUDA/cuDNN version: N/A
• GPU model and memory: N/A

Please provide the entire URL of the model you are using?

https://github.com/tensorflow/models/blob/394baa9f21424d3522ccfbdcee8acd3840891ff6/research/deeplab/g3doc/quantize.md

Describe the current behavior
Good Evening,

I have been following the steps in https://github.com/tensorflow/models/blob/394baa9f21424d3522ccfbdcee8acd3840891ff6/research/deeplab/g3doc/quantize.md to quantize a DeeplabV3 MobileNetV2 model for the USB Coral TPU but I cannot get it to work.

Even when I start from the models provided at the bottom (not the 8_bits, the raw models), I cannot get it to run. And I even get different problem for both models.

For the deeplabv3_mnv2_pascal_train_aug:

• The deeplab/train.py script for quantizing runs successfully.
• The deeplab/export_model.py scripts step works.
• The tflite_convert works too.
• EdgeTPU compiler gets cryptic error message: Internal compiler error. Aborting!

For the deeplabv3_mnv2_dm05_pascal_trainaug:

• The deeplab/train.py script for quantizing runs gets:
  ValueError: Total size of new array must be unchanged for MobilenetV2/Conv/weights lh_shape: [(3, 3, 3, 16)], rh_shape: [(3, 3, 3, 32)]
  I tried this solution unsuccessfully.

• Using the .pb file directly, the deeplab/train.py with quantization aware training fails with python crashing.

• Doing a post-quantization in python directly works.

• edgetpu_compiler works.

• But once ran on the TPU, we get 0.87 TPS, most likely because the edgetpu-converter cannot convert all to edgtpu.

Any guidance to get it working would be great. I am also open to solutions not following the guide from tensorflow.

Describe the expected behavior

It should not crash.

Code to reproduce the issue

Take the 2 models at the bottom of the https://github.com/tensorflow/models/blob/394baa9f21424d3522ccfbdcee8acd3840891ff6/research/deeplab/g3doc/quantize.md page (not the 8_bits one) and try to quantize them following the guide and run them on an USB Coral Edge TPU.

Other info / logs

See above