Release v1.0.1-2021.12.26

Release model weight fine-tuned on CIFAR

This release includes fine-tuned model weights.

you can use model weights using below command

import torch

model_name = 'efficientnet_v2_s'
weight_path = 'efficientnet_v2_s_cifar100.pth'
model = torch.hub.load('hankyul2/EfficientNetV2-pytorch', model_name, nclass=100, skip_validation=True)
model.load_state_dict(torch.load(weight_path))

v1.0.0-2021.11.13

EfficientNetV2 Numpy weights (Converted from official tensorflow weights)

This release contains efficientnetv2 numpy weights converted from official tensorflow checkpoint. We just download checkpoint from official repo and convert trainable parameters to numpy array without changing any parameters.

How do we convert to numpy format

import os
import shutil
import subprocess
import tensorflow as tf

def download_from_url_and_convert_to_numpy(url, file_name):
    subprocess.run(["wget", "-r", "-nc", '-O', file_name, url])
    shutil.unpack_archive(file_name)
    ckpt_path = os.path.splitext(file_name)[0]
    pretrained_ckpt = tf.train.latest_checkpoint(ckpt_path)
    np.save(f"{ckpt_path}.npy",
            {'/'.join(name.split('/')[1:]): np.array(tf.train.load_variable(ckpt_path, name)) for name, shape in
             tf.train.list_variables(pretrained_ckpt)})

url = "https://storage.googleapis.com/cloud-tpu-checkpoints/efficientnet/v2/efficientnetv2-s.tgz"
file_name = "efficientnet_v2.tgz"
download_from_url_and_convert_to_numpy(url, file_name)

How to see variable name and shape

import numpy as np
weight = np.load('efficientnetv2-s.npy', allow_pickle=True).item()
for k, v in weight.items():
    print(k, v.shape)

Below links are official tensorflow ckpt download link (got from here)

• efficientnet_v2_s
• efficientnet_v2_m
• efficientnet_v2_l
• efficientnet_v2_s_in21k
• efficientnet_v2_m_in21k
• efficientnet_v2_l_in21k
• efficientnet_v2_xl_in21k