mmsegmentation/tools/model_converters/mit_convert.py

77 lines
2.7 KiB
Python

import argparse
from collections import OrderedDict
import torch
def mit_convert(ckpt):
new_ckpt = OrderedDict()
# Process the concat between q linear weights and kv linear weights
for k, v in ckpt.items():
if k.startswith('head'):
continue
# patch embedding convertion
elif k.startswith('patch_embed'):
stage_i = int(k.split('.')[0].replace('patch_embed', ''))
new_k = k.replace(f'patch_embed{stage_i}', f'layers.{stage_i-1}.0')
new_v = v
if 'proj.' in new_k:
new_k = new_k.replace('proj.', 'projection.')
# transformer encoder layer convertion
elif k.startswith('block'):
stage_i = int(k.split('.')[0].replace('block', ''))
new_k = k.replace(f'block{stage_i}', f'layers.{stage_i-1}.1')
new_v = v
if 'attn.q.' in new_k:
sub_item_k = k.replace('q.', 'kv.')
new_k = new_k.replace('q.', 'attn.in_proj_')
new_v = torch.cat([v, ckpt[sub_item_k]], dim=0)
elif 'attn.kv.' in new_k:
continue
elif 'attn.proj.' in new_k:
new_k = new_k.replace('proj.', 'attn.out_proj.')
elif 'attn.sr.' in new_k:
new_k = new_k.replace('sr.', 'sr.')
elif 'mlp.' in new_k:
string = f'{new_k}-'
new_k = new_k.replace('mlp.', 'ffn.layers.')
if 'fc1.weight' in new_k or 'fc2.weight' in new_k:
new_v = v.reshape((*v.shape, 1, 1))
new_k = new_k.replace('fc1.', '0.')
new_k = new_k.replace('dwconv.dwconv.', '1.')
new_k = new_k.replace('fc2.', '4.')
string += f'{new_k} {v.shape}-{new_v.shape}'
# norm layer convertion
elif k.startswith('norm'):
stage_i = int(k.split('.')[0].replace('norm', ''))
new_k = k.replace(f'norm{stage_i}', f'layers.{stage_i-1}.2')
new_v = v
else:
new_k = k
new_v = v
new_ckpt[new_k] = new_v
return new_ckpt
def parse_args():
parser = argparse.ArgumentParser(
'Convert official segformer backbone weights to mmseg style.')
parser.add_argument(
'src', help='Source path of official segformer backbone weights.')
parser.add_argument(
'dst',
help='Destination path of converted segformer backbone weights.')
return parser.parse_args()
if __name__ == '__main__':
args = parse_args()
src_path = args.src
dst_path = args.dst
ckpt = torch.load(src_path, map_location='cpu')
ckpt = mit_convert(ckpt)
torch.save(ckpt, dst_path)