search

Bisonai / mobilenetv3-tensorflow

Unofficial implementation of MobileNetV3 architecture described in paper Searching for MobileNetV3.
Apache License 2.0
225 stars 81 forks source link

Do you know how to use NAS and NetAdapt to optimize mobilenet block and layer to adapt to other task? #6

Closed alexliyang closed 5 years ago

alexliyang commented 5 years ago

when I use paper parameters to train my task , the result is not good enough.

how to use NAS and NetAdapt to optimize mobilenet block and layer to adapt to other task?

martinkersner commented 5 years ago

Hi @alexliyang!

MobileNetV3 and its predecessors should be suitable for many different tasks. What task are you training for?

Cheers, Martin

alexliyang commented 5 years ago

Hi, @martinkersner :+1: I use MobilenetV3 to train image classes for mobilephone, but on my dataset, the result is not good , so I try to optimize the parameters again. Maybe it is a big task

martinkersner commented 5 years ago

@alexliyang What does your data look like? How many classes do you train for? The problem could be due to the hyperparameters you use. It is better to investigate it before moving on NAS or NetAdapt.

alexliyang commented 5 years ago

@martinkersner my data just like flowers, pets, foods and others' image , about 7 classes, I resized the images to 224,224 or 128,128 , I trained mobilenetv3 on the dataset about 100 epoches, the acc is about 0.95, but on the test data is very poor

martinkersner commented 5 years ago

@alexliyang Is your test data from the same distribution as train data? Do you use any preprocessing/augmentation? Did you try to train with CIFAR10 or MNIST before?