82cc53237e
Working on support for siglip (w/ attn pool) vit backbone, and adding registers (reg tokens)
2023-09-30 16:03:01 -07:00
6bae514656
Add pretrained patch embed resizing to swin
2023-09-27 10:27:28 -07:00
5c504b4ded
flip these two
2023-09-27 10:24:12 -07:00
8ba2038e6b
fast_vit: propagate act_layer argument
2023-09-27 10:24:12 -07:00
95ba90157f
Update tiny_vit.py to fix bug
2023-09-23 10:05:52 -07:00
245ad4f41a
Added missing RuntimeError to builder functions of models that do not currently support feature extraction
2023-09-19 08:19:14 -07:00
d4c21b95f4
Update repghost.py
2023-09-15 11:41:56 -07:00
7eb7d13845
Fix in_features for linear layer in reset_classifier.
2023-09-13 09:29:38 -07:00
0d124ffd4f
Update README. Fine-grained layer-wise lr decay working for tiny_vit and both efficientvits. Minor fixes.
2023-09-01 15:05:29 -07:00
2f0fbb59b3
TinyViT weights on HF hub
2023-09-01 11:05:56 -07:00
507cb08acf
TinyVitBlock needs adding as leaf for FX now, tweak a few dim names
2023-09-01 11:05:56 -07:00
9caf32b93f
Move levit style pos bias resize with other rel pos bias utils
2023-09-01 11:05:56 -07:00
63417b438f
TinyViT adjustments
...
* keep most of net in BCHW layout, performance appears same, can remove static resolution attribs and features easier to use
* add F.sdpa, decent gains in pt 2.1
* tweak crop pct based on eval
2023-09-01 11:05:56 -07:00
39aa44b192
Fixing tinyvit trace issue
2023-09-01 11:05:56 -07:00
aea3b9c854
Fixing tinyvit input_size issue
2023-09-01 11:05:56 -07:00
fabc4e5bcd
Fixing tinyvit torchscript issue
2023-09-01 11:05:56 -07:00
bae949f830
fix attention_bias_cache in tinyvit
2023-09-01 11:05:56 -07:00
170a5b6e27
add tinyvit
2023-09-01 11:05:56 -07:00
983310d6a2
Fix #1935 , torch.amp.autocast -> torch.autocast (namespace issue for 1.10 pt compat)
2023-08-30 15:03:28 -07:00
c8b2f28096
Fix a few typos, fix fastvit proj_drop, add code link
2023-08-28 21:26:29 -07:00
fc5d705b83
dynamic_size -> dynamic_img_size, add dynamic_img_pad for padding option
2023-08-27 15:58:35 -07:00
1f4512fca3
Support dynamic_resize in eva.py models
2023-08-27 15:58:35 -07:00
ea3519a5f0
Fix dynamic_resize for deit models (distilled or no_embed_cls) and vit w/o class tokens
2023-08-27 15:58:35 -07:00
4d8ecde6cc
Fix torchscript for vit-hybrid dynamic_resize
2023-08-27 15:58:35 -07:00
fdd8c7c2da
Initial impl of dynamic resize for existing vit models (incl vit-resnet hybrids)
2023-08-27 15:58:35 -07:00
5d599a6a10
RepViT weights on HF hub
2023-08-25 10:39:02 -07:00
56c285445c
Wrong pool size for 384x384 inception_next_base
2023-08-24 18:31:44 -07:00
af9f56f3bf
inception_next dilation support, weights on hf hub, classifier reset / global pool / no head fixes
2023-08-24 18:31:44 -07:00
2d33b9df6c
Add features_only support to inception_next
2023-08-24 18:31:44 -07:00
3d8d7450ad
InceptionNeXt using timm builder, more cleanup
2023-08-24 18:31:44 -07:00
f4cf9775c3
Adding InceptionNeXt
2023-08-24 18:31:44 -07:00
d6c348765a
Fix first_conv for mobileone and fastvit
2023-08-23 22:50:37 -07:00
16334e4bec
Fix two fastvit issues
2023-08-23 22:50:37 -07:00
5242ba6edc
MobileOne and FastViT weights on HF hub, more code cleanup and tweaks, features_only working. Add reparam flag to validate and benchmark, support reparm of all models with fuse(), reparameterize() or switch_to_deploy() methods on modules
2023-08-23 22:50:37 -07:00
40dbaafef5
Stagify FastViT /w downsample to top of stage
2023-08-23 22:50:37 -07:00
8470eb1cb5
More fastvit & mobileone updates, ready for weight upload
2023-08-23 22:50:37 -07:00
8474508d07
More work on FastViT, use own impl of MobileOne, validation working with remapped weight, more refactor TODO
2023-08-23 22:50:37 -07:00
c7a20cec13
Begin adding FastViT
2023-08-23 22:50:37 -07:00
7fd3674d0d
Add mobileone and update repvgg
2023-08-23 22:50:37 -07:00
3055411c1b
Fix samvit bug, add F.sdpa support and ROPE option ( #1920 )
...
* Fix a bug I introduced in samvit, add F.sdpa support and ROPE option to samvit, neck is LayerNorm if not used and standard classifier used
* Add attn dropout to F.sdpa
* Fix fx trace for sam vit
* Fixing torchscript issues in samvit
* Another torchscript fix
* samvit head fc name fix
2023-08-20 21:22:59 -07:00
300f54a96f
Another effcientvit (mit) tweak, fix torchscript/fx conflict with autocast disable
2023-08-20 15:07:25 -07:00
dc18cda2e7
efficientvit (mit) msa attention q/k/v ops need to be in float32 to train w/o NaN
2023-08-20 11:49:36 -07:00
be4e0d8f76
Update attrib comment to include v2
2023-08-19 23:39:09 -07:00
126a58e563
Combine ghostnetv2 with ghostnet, reduec redundancy, add weights to hf hub.
2023-08-19 23:33:43 -07:00
3f320a9e57
Merge branch 'Add-GhostNetV2' of github.com:yehuitang/pytorch-image-models into yehuitang-Add-GhostNetV2
2023-08-19 22:07:54 -07:00
7c2728c6fe
Merge pull request #1919 from ChengpengChen/main
...
Add RepGhost models and weights
2023-08-19 16:26:45 -07:00
69e0ca2e36
Weights on hf hub, bicubic yields slightly better eval
2023-08-19 16:25:45 -07:00
b8011565bd
Merge pull request #1894 from seefun/master
...
add two different EfficientViT models
2023-08-19 09:24:14 -07:00
7d7589e8da
Fixing efficient_vit torchscript, fx, default_cfg issues
2023-08-18 23:23:11 -07:00
58ea1c02c4
Add fixed_input_size flag to msra efficient_vit
2023-08-18 16:48:17 -07:00
Ross Wightman
Yassine
Nguyen Nhat Hoang
belfner
Thorsten Hempel
方曦
Ross Wightman