Home
MobileNet on TensorFlow with ability to fine-tune and incorporate center or triplet loss
A tensorflow implementation of Google's MobileNets for re-training/fine-tuning on your own custom dataset with the addition of (optional) center loss or triplet loss. Additionally, this repo can be used to re-train Inception network as well with the above added benefits.
Currently this repo is compatible with Tensorflow 1.3.0.
News| Date | Update |
|---|---|
| 2017-10-25 | Currently working on triplet loss |
| 2017-10-25 | Added code to support center loss |
Pre-trained ModelInception_v3 is the most accurate model, but also the slowest. For faster or smaller models, choose a MobileNet with the form mobilenet_<parameter_size>_<input_size>_[(optional)quantized]. For example,'mobilenet_1.0_224' will pick a model that is 17 MB in size and takes 224
pixel input images, while 'mobilenet_0.25_128_quantized' will choose a much
less accurate, but smaller and faster network that's 920 KB on disk and
takes 128x128 images.
These models are automatically downloaded for you.
InstallationDetails on how to install and re-train on your own dataset can be found on the wiki page. Different parameters that can be tweaked are also given there.
InspirationThe code is heavily inspired by the Tensorflow's Retrain Script and FaceNet.