darts model pre-trained


Keywords
automl, darts, eeea-nets, image-classification, mobilenet, models, nas, nasnet, pcdarts, pdarts, pytorch
License
Apache-2.0
Install
pip install darmo==0.1.10

Documentation

DARMO

Darts family models pre-trained

PyPI version PyPI Release DOI

What's New

Sep 26, 2021

  • Add ResT ImageNet-1k models, Thank from ResT

Sep 20, 2021

  • Add OFA-595 ImageNet-1k and ImageNet-21k, Thank from 21k-weight
  • Add TResNet-M ImageNet-21k, Thank from 21k-weight

Sep 5, 2021

  • Add ResNet50 ImageNet-1k and ImageNet-21k, Thank from 21k-weight

Aug 17, 2021

  • Add EEEA-Net-C1 model

Aug 3, 2021

  • Add EEEA-Net-C2 model

July 20, 2021

  • Add test imagenet sctipt and result

April 5, 2021

  • Add NASNet model
  • Set params auxiliary

Oct 27, 2020

  • Add DARTSv2, PDART, RelativeNAS models

Supported Models

Install

The library can be installed with pip:

pip install darmo

Create models

import darmo

# just change -> "dartsv2", "pdarts", "relative_nas", "nasnet", "eeea_c2", "eeea_c1"
model = darmo.create_model("dartsv2", num_classes=1000, pretrained=True, auxiliary=True)

Supported Transfer learning

# create model with ImageNet pretrained
model = darmo.create_model("dartsv2", num_classes=1000, pretrained=True, auxiliary=True)

# Reset classifier layer with add dropout before classifier layer
model.reset_classifier(num_classes=100, dropout=0.2)

Test model on ImageNet

git clone https://github.com/jitdee-ai/darmo/
cd darmo
python test_imagenet.py --arch relative_nas --data [path of imagenet include val folder]

Results of ImageNet

Model Top-1 Acc Top-5 Acc
dartsv2 73.59 91.40
pdarts 75.94 92.74
eeea_c1 74.19 91.49
eeea_c2 76.17 92.66

Citations this source code

@software{chakkrit_termritthikun_2020_4139755,
author       = {Chakkrit Termritthikun},
title        = {jitdee-ai/darmo: pre-trained models for darts},
month        = oct,
year         = 2020,
publisher    = {Zenodo},
version      = {0.0.4},
doi          = {10.5281/zenodo.4139755},
url          = {https://doi.org/10.5281/zenodo.4139755}
}