mmpretrain/tools/model_converters/vig_to_mmpretrain.py

99 lines
3.3 KiB
Python

# Copyright (c) OpenMMLab. All rights reserved.
import argparse
import os.path as osp
import re
from collections import OrderedDict
import mmengine
import torch
from mmengine.runner import CheckpointLoader
def convert_vig(ckpt):
new_ckpt = OrderedDict()
for k, v in ckpt.items():
new_key = k
new_value = v
if 'pos_embed' in new_key:
new_key = new_key.replace('pos_embed', 'backbone.pos_embed')
elif 'stem' in new_key:
new_key = new_key.replace('stem.convs', 'backbone.stem')
elif 'backbone' in new_key:
new_key = new_key.replace('backbone', 'backbone.blocks')
elif 'prediction.0' in new_key:
new_key = new_key.replace('prediction.0', 'head.fc1')
new_value = v.squeeze(-1).squeeze(-1)
elif 'prediction.1' in new_key:
new_key = new_key.replace('prediction.1', 'head.bn')
elif 'prediction.4' in new_key:
new_key = new_key.replace('prediction.4', 'head.fc2')
new_value = v.squeeze(-1).squeeze(-1)
new_ckpt[new_key] = new_value
return new_ckpt
def convert_pvig(ckpt):
new_ckpt = OrderedDict()
stage_idx = 0
stage_blocks = 0
for k, v in ckpt.items():
new_key: str = k
new_value = v
if 'pos_embed' in new_key:
new_key = new_key.replace('pos_embed', 'backbone.pos_embed')
elif 'stem' in new_key:
new_key = new_key.replace('stem.convs', 'backbone.stem')
elif re.match(r'^backbone\.\d+\.conv', new_key) is not None:
if new_key.endswith('0.weight'):
stage_idx += 1
stage_blocks = int(new_key.split('.')[1])
other = new_key.split('.', maxsplit=3)[-1]
new_key = f'backbone.stages.{stage_idx}.0.' + other
elif 'backbone' in new_key:
block_idx = int(new_key.split('.')[1]) - stage_blocks
other = new_key.split('.', maxsplit=2)[-1]
new_key = f'backbone.stages.{stage_idx}.{block_idx}.' + other
elif 'prediction.0' in new_key:
new_key = new_key.replace('prediction.0', 'head.fc1')
new_value = v.squeeze(-1).squeeze(-1)
elif 'prediction.1' in new_key:
new_key = new_key.replace('prediction.1', 'head.bn')
elif 'prediction.4' in new_key:
new_key = new_key.replace('prediction.4', 'head.fc2')
new_value = v.squeeze(-1).squeeze(-1)
new_ckpt[new_key] = new_value
return new_ckpt
def main():
parser = argparse.ArgumentParser(
description='Convert keys in pretrained vig models to '
'mmpretrain style.')
parser.add_argument('src', help='src model path or url')
# The dst path must be a full path of the new checkpoint.
parser.add_argument('dst', help='save path')
args = parser.parse_args()
checkpoint = CheckpointLoader.load_checkpoint(args.src, map_location='cpu')
if 'model' in checkpoint:
state_dict = checkpoint['model']
else:
state_dict = checkpoint
if 'backbone.2.conv.0.weight' in state_dict:
weight = convert_pvig(state_dict)
else:
weight = convert_vig(state_dict)
mmengine.mkdir_or_exist(osp.dirname(args.dst))
torch.save(weight, args.dst)
print('Done!!')
if __name__ == '__main__':
main()