Mobilenet v2 pretrained weights with lowered depth multipliers

[Patrick_Conway]

I’m training ssd mobilenet v2 on custom data

As per this blog post

Custom object detection in the browser using TensorFlow.js

I want to use a lower depth multiplier to decrease the model size

When I change the depth multiplier in the pipeline config, the pretrained weigths are not compatible.

Are there any pretrained weigths available for tf2 with different depth and input size
Or is it possible to load the checkpoints provided here

tensorflow/models/blob/master/research/slim/nets/mobilenet/README.md

and export them in a format which can be used to train with model_main_tf2.py

[8bitmp3]

@Patrick_Conway

Are there any pretrained weigths available for tf2 with different depth and input size

Hi @Patrick_Conway

How about SSD MobileNet V2 for object detection : TensorFlow Hub

trained on the COCO2017 dataset. For example:

• https://tfhub.dev/tensorflow/ssd_mobilenet_v2
• https://tfhub.dev/tensorflow/ssd_mobilenet_v2/fpnlite_320x320
• https://tfhub.dev/tensorflow/ssd_mobilenet_v2/fpnlite_640x640

---

Alternatively, here are the MobileNet V2s for TF2 (various tasks) with some with different depth multipliers: https://tfhub.dev/s?network-architecture=mobilenet-v2&tf-version=tf2

[Patrick_Conway]

Thanks for the reply, I am trying to recreate coco ssd from tensorflow js on a custom dataset.
When training https://tfhub.dev/tensorflow/ssd_mobilenet_v2 I end up with a model that is around 10Mb , but the coco ssd pretrained model when using lite_mobilenet_v2 is around 1Mb

tfjs-models/coco-ssd at master · tensorflow/tfjs-models

I got a lot of browser crashes and ‘aw snap something went wrong’ errors when running my model trained with ssd_mobilenet_v2 and tf2.

So my basic goal is to reduce the model size and apply the same optimizations to improve the performance for browser execution which have been applied in the coco ssd implementation.

I believe the model size can be reduced by reducing the depth multiplier and applying uint8 quantization.

[8bitmp3]

@Patrick_Conway

So my basic goal is to reduce the model size and apply the same optimizations to improve the performance for browser execution which have been applied in the coco ssd implementation.

Awesome. Not an expert here but off the top of my head:

@8bitmp3

have you considered the following by any chance to improve generalization?

• Post-training quantization (Post-training quantization | TensorFlow Lite ):
  • For example, float16 quantization can lead to a “minimal loss in accuracy”.
  • Another example: integer-only “16x8” quantization (experimental), which “can improve accuracy significantly, but only slightly increase model size”.
• Some basic things to try to reduce overfitting and improve generalization (Overfit and underfit | TensorFlow Core):
  • More training data; “quality” data/data quality, etc.
  • Weight regularization, dropout, and various other techniques, which are active areas in ML research.
  • In “imbalanced datasets” …" overfitting is a significant concern from the lack of training data" (Classification on imbalanced data | TensorFlow Core ).

As well as:

• Model optimization  |  TensorFlow Lite
• Post-training quantization  |  TensorFlow Lite
• Post-training dynamic range quantization  |  TensorFlow Lite
• Post-training integer quantization  |  TensorFlow Lite
• Post-training float16 quantization  |  TensorFlow Lite
• Post-training integer quantization with int16 activations  |  TensorFlow Lite
• TensorFlow Lite 8-bit quantization specification
• Inspecting Quantization Errors with Quantization Debugger  |  TensorFlow Lite

This post may be a bit dated but useful:

• TensorFlow Model Optimization Toolkit — float16 quantization halves model size — The TensorFlow Blog

[8bitmp3]

@Patrick_Conway

to recreate coco ssd from tensorflow js on a custom dataset.

Another doc you may find useful - retraining with a custom dataset (scroll down to the Download the headless model section of the tutorial): Transfer learning with TensorFlow Hub  |  TensorFlow Core

[Patrick_Conway]

thanks, I will give it a try