f88c606fcf
fixing channels_last on cond_conv2d; update nvfuser debug env variable
2022-04-25 12:41:46 -07:00
f670d98cb8
Make a few more layers symbolically traceable (remove from FX leaf modules)
...
* remove dtype kwarg from .to() calls in EvoNorm as it messed up script + trace combo
* BatchNormAct2d always uses custom forward (cut & paste from original) instead of super().forward. Fixes #1176
* BlurPool groups==channels, no need to use input.dim[1]
2022-03-24 21:43:56 -07:00
b049a5c5c6
Merge remote-tracking branch 'origin/master' into norm_norm_norm
2022-03-21 13:41:43 -07:00
9440a50c95
Merge branch 'mrT23-master'
2022-03-21 12:30:02 -07:00
372ad5fa0d
Significant model refactor and additions:
...
* All models updated with revised foward_features / forward_head interface
* Vision transformer and MLP based models consistently output sequence from forward_features (pooling or token selection considered part of 'head')
* WIP param grouping interface to allow consistent grouping of parameters for layer-wise decay across all model types
* Add gradient checkpointing support to a significant % of models, especially popular architectures
* Formatting and interface consistency improvements across models
* layer-wise LR decay impl part of optimizer factory w/ scale support in scheduler
* Poolformer and Volo architectures added
2022-02-28 13:56:23 -08:00
95cfc9b3e8
Merge remote-tracking branch 'origin/master' into norm_norm_norm
2022-01-25 22:20:45 -08:00
656757d26b
Fix MobileNetV2 head conv size for multiplier < 1.0. Add some missing modification copyrights, fix starting date of some old ones.
2022-01-14 16:28:27 -08:00
b27c21b09a
Update drop_path and drop_block (fast impl) to be symbolically traceable, slightly faster
2022-01-06 16:04:58 -08:00
214c84a235
Disable use of timm nn.Linear wrapper since AMP autocast + torchscript use appears fixed
2022-01-06 16:01:51 -08:00
a52a614475
Remove layer experiment which should not have been added
2021-12-14 14:29:32 -08:00
ab49d275de
Significant norm update
...
* ConvBnAct layer renamed -> ConvNormAct and ConvNormActAa for anti-aliased
* Significant update to EfficientNet and MobileNetV3 arch to support NormAct layers and grouped conv (as alternative to depthwise)
* Update RegNet to add Z variant
* Add Pre variant of XceptionAligned that works with NormAct layers
* EvoNorm matches bits_and_tpu branch for merge
2021-12-14 13:48:30 -08:00
d04f2f1377
Update drop_path and drop_block (fast impl) to be symbolically traceable, slightly faster
2021-12-05 15:36:56 -08:00
834a9ec721
Disable use of timm nn.Linear wrapper since AMP autocast + torchscript use appears fixed
2021-12-01 14:58:09 -08:00
78912b6375
Updated EvoNorm implementations with some experimentation. Add FilterResponseNorm. Updated RegnetZ and ResNetV2 model defs for trials.
2021-12-01 12:09:01 -08:00
c11f4c3218
support CNNs
2021-11-30 08:48:08 +02:00
d6701d8a81
Merge branch 'rwightman:master' into master
2021-11-30 08:07:44 +02:00
480c676ffa
Fix FX breaking assert in evonorm
2021-11-24 09:24:47 -08:00
41559247e9
use_ml_decoder_head
2021-11-22 17:50:39 +02:00
93cc08fdc5
Make evonorm variables 1d to match other PyTorch norm layers, will break weight compat for any existing use (likely minimal, easy to fix).
2021-11-20 15:50:51 -08:00
af607b75cc
Prep a set of ResNetV2 models with GroupNorm, EvoNormB0, EvoNormS0 for BN free model experiments on TPU and IPU
2021-11-19 17:37:00 -08:00
c976a410d9
Add ResNet-50 w/ GN (resnet50_gn) and SEBotNet-33-TS (sebotnet33ts_256) model defs and weights. Update halonet50ts weights w/ slightly better variant in1k val, more robust to test sets.
2021-11-19 14:24:43 -08:00
b25ff96768
wip - pre-rebase
2021-11-12 20:45:05 +00:00
e051dce354
Make all models FX traceable
2021-11-12 20:45:05 +00:00
0149ec30d7
wip - attempting to rebase
2021-11-12 20:45:05 +00:00
bc3d4eb403
wip -rebase
2021-11-12 20:45:05 +00:00
2ddef942b9
Better fix for #954 that doesn't break torchscript, pull torch._assert into timm namespace when it exists
2021-11-02 11:22:33 -07:00
4f0f9cb348
Fix #954 by bringing traceable _assert into timm to allow compat w/ PyTorch < 1.8
2021-11-02 09:21:40 -07:00
b745d30a3e
Fix formatting of last commit
2021-10-25 15:15:14 -07:00
3478f1d7f1
Traceability fix for vit models for some experiments
2021-10-25 15:13:08 -07:00
f658a72e72
Cleanup re-use of Dropout modules in Mlp modules after some twitter feedback :p
2021-10-25 00:40:59 -07:00
c02334d9fa
Add weights for regnetz_d and haloregnetz_c, update regnetz_c weights. Add commented PyTorch XLA code for halo attention
2021-10-19 12:32:09 -07:00
02daf2ab94
Add option to include relative pos embedding in the attention scaling as per references. See discussion #912
2021-10-12 15:37:01 -07:00
e2b8d44ff0
Halo, bottleneck attn, lambda layer additions and cleanup along w/ experimental model defs
...
* align interfaces of halo, bottleneck attn and lambda layer
* add qk_ratio to all of above, control q/k dim relative to output dim
* add experimental haloregnetz, and trionet (lambda + halo + bottle) models
2021-10-06 16:32:48 -07:00
007bc39323
Some halo and bottleneck attn code cleanup, add halonet50ts weights, use optimal crop ratios
2021-10-02 15:51:42 -07:00
b1c2e3eb92
Match rel_pos_indices attr rename in conv branch
2021-09-30 23:19:05 -07:00
b49630a138
Add relative pos embed option to LambdaLayer, fix last transpose/reshape.
2021-09-30 22:45:09 -07:00
b81e79aae9
Fix bottleneck attn transpose typo, hopefully these train better now..
2021-09-28 16:38:41 -07:00
515121cca1
Use reshape instead of view in std_conv, causing issues in recent PyTorch in channels_last
2021-09-23 15:43:48 -07:00
5bd04714e4
Cleanup weight init for byob/byoanet and related
2021-09-05 15:34:05 -07:00
8642401e88
Swap botnet 26/50 weights/models after realizing a mistake in arch def, now figuring out why they were so low...
2021-09-05 15:17:19 -07:00
5f12de4875
Add initial AttentionPool2d that's being trialed. Fix comment and still trying to improve reliability of sgd test.
2021-09-05 12:41:14 -07:00
492c0a4e20
Update HaloAttn comment
2021-09-01 17:14:31 -07:00
3b9032ea48
Use Tensor.unfold().unfold() for HaloAttn, fast like as_strided but more clarity
2021-08-27 12:45:53 -07:00
8449ba210c
Improve performance of HaloAttn, change default dim calc. Some cleanup / fixes for byoanet. Rename resnet26ts to tfs to distinguish (extra fc).
2021-08-26 21:56:44 -07:00
925e102982
Update attention / self-attn based models from a series of experiments:
...
* remove dud attention, involution + my swin attention adaptation don't seem worth keeping
* add or update several new 26/50 layer ResNe(X)t variants that were used in experiments
* remove models associated with dead-end or uninteresting experiment results
* weights coming soon...
2021-08-20 16:13:11 -07:00
01cb46a9a5
Add gc_efficientnetv2_rw_t weights (global context instead of SE attn). Add TF XL weights even though the fine-tuned ones don't validate that well. Change default arg for GlobalContext to use scal (mul) mode.
2021-08-07 16:45:29 -07:00
8165cacd82
Realized LayerNorm2d won't work in all cases as is, fixed.
2021-07-05 18:21:34 -07:00
b9cfb64412
Support npz custom load for vision transformer hybrid models. Add posembed rescale for npz load.
2021-06-14 12:31:44 -07:00
8319e0c373
Add file docstring to std_conv.py
2021-06-13 12:31:06 -07:00
4d96165989
Merge branch 'master' into cleanup_xla_model_fixes
2021-06-12 23:19:25 -07:00
jjsjann123
Ross Wightman
talrid
mrT23
Alexander Soare