mirror of
https://github.com/open-mmlab/mmsegmentation.git
synced 2025-06-03 22:03:48 +08:00
b6c7c77a08
* add Swin Transformer * add Swin Transformer * fixed import * Add some swin training settings. * Fix some filename error. * Fix attribute name: pretrain -> pretrained * Upload mmcls implementation of swin transformer. * Refactor Swin Transformer to follow mmcls style. * Refactor init_weigths of swin_transformer.py * Fix lint * Match inference precision * Add some comments * Add swin_convert to load official style ckpt * Remove arg: auto_pad * 1. Complete comments for each block; 2. Correct weight convert function; 3. Fix the pad of Patch Merging; * Clean function args. * Fix vit unit test. * 1. Add swin transformer unit tests; 2. Fix some pad bug; 3. Modify config to adapt new swin implementation; * Modify config arg * Update readme.md of swin * Fix config arg error and Add some swin benchmark msg. * Add MeM and ms test content for readme.md of swin transformer. * Fix doc string of swin module * 1. Register swin transformer to model list; 2. Modify pth url which keep meta attribute; * Update swin.py * Merge config settings. * Modify config style. * Update README.md Add ViT link * Modify main readme.md Co-authored-by: Jiarui XU <xvjiarui0826@gmail.com> Co-authored-by: sennnnn <201730271412@mail.scut.edu.cn> Co-authored-by: Junjun2016 <hejunjun@sjtu.edu.cn>
91 lines
2.8 KiB
Python
91 lines
2.8 KiB
Python
from collections import OrderedDict
|
|
|
|
|
|
def swin_convert(ckpt):
|
|
new_ckpt = OrderedDict()
|
|
|
|
def correct_unfold_reduction_order(x):
|
|
out_channel, in_channel = x.shape
|
|
x = x.reshape(out_channel, 4, in_channel // 4)
|
|
x = x[:, [0, 2, 1, 3], :].transpose(1,
|
|
2).reshape(out_channel, in_channel)
|
|
return x
|
|
|
|
def correct_unfold_norm_order(x):
|
|
in_channel = x.shape[0]
|
|
x = x.reshape(4, in_channel // 4)
|
|
x = x[[0, 2, 1, 3], :].transpose(0, 1).reshape(in_channel)
|
|
return x
|
|
|
|
for k, v in ckpt.items():
|
|
if k.startswith('head'):
|
|
continue
|
|
elif k.startswith('layers'):
|
|
new_v = v
|
|
if 'attn.' in k:
|
|
new_k = k.replace('attn.', 'attn.w_msa.')
|
|
elif 'mlp.' in k:
|
|
if 'mlp.fc1.' in k:
|
|
new_k = k.replace('mlp.fc1.', 'ffn.layers.0.0.')
|
|
elif 'mlp.fc2.' in k:
|
|
new_k = k.replace('mlp.fc2.', 'ffn.layers.1.')
|
|
else:
|
|
new_k = k.replace('mlp.', 'ffn.')
|
|
elif 'downsample' in k:
|
|
new_k = k
|
|
if 'reduction.' in k:
|
|
new_v = correct_unfold_reduction_order(v)
|
|
elif 'norm.' in k:
|
|
new_v = correct_unfold_norm_order(v)
|
|
else:
|
|
new_k = k
|
|
new_k = new_k.replace('layers', 'stages', 1)
|
|
elif k.startswith('patch_embed'):
|
|
new_v = v
|
|
if 'proj' in k:
|
|
new_k = k.replace('proj', 'projection')
|
|
else:
|
|
new_k = k
|
|
else:
|
|
new_v = v
|
|
new_k = k
|
|
|
|
new_ckpt[new_k] = new_v
|
|
|
|
return new_ckpt
|
|
|
|
|
|
def vit_convert(ckpt):
|
|
|
|
new_ckpt = OrderedDict()
|
|
|
|
for k, v in ckpt.items():
|
|
if k.startswith('head'):
|
|
continue
|
|
if k.startswith('norm'):
|
|
new_k = k.replace('norm.', 'ln1.')
|
|
elif k.startswith('patch_embed'):
|
|
if 'proj' in k:
|
|
new_k = k.replace('proj', 'projection')
|
|
else:
|
|
new_k = k
|
|
elif k.startswith('blocks'):
|
|
if 'norm' in k:
|
|
new_k = k.replace('norm', 'ln')
|
|
elif 'mlp.fc1' in k:
|
|
new_k = k.replace('mlp.fc1', 'ffn.layers.0.0')
|
|
elif 'mlp.fc2' in k:
|
|
new_k = k.replace('mlp.fc2', 'ffn.layers.1')
|
|
elif 'attn.qkv' in k:
|
|
new_k = k.replace('attn.qkv.', 'attn.attn.in_proj_')
|
|
elif 'attn.proj' in k:
|
|
new_k = k.replace('attn.proj', 'attn.attn.out_proj')
|
|
else:
|
|
new_k = k
|
|
new_k = new_k.replace('blocks.', 'layers.')
|
|
else:
|
|
new_k = k
|
|
new_ckpt[new_k] = v
|
|
|
|
return new_ckpt
|