a16ea1e355
Merge remote-tracking branch 'origin/norm_norm_norm' into bits_and_tpu
3 years ago
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
3 years ago
066e490605
Merge branch 'norm_norm_norm' into bits_and_tpu
3 years ago
95cfc9b3e8
Merge remote-tracking branch 'origin/master' into norm_norm_norm
3 years ago
656757d26b
Fix MobileNetV2 head conv size for multiplier < 1.0. Add some missing modification copyrights, fix starting date of some old ones.
3 years ago
b27c21b09a
Update drop_path and drop_block (fast impl) to be symbolically traceable, slightly faster
3 years ago
214c84a235
Disable use of timm nn.Linear wrapper since AMP autocast + torchscript use appears fixed
3 years ago
7eb7e73216
File will not stay deleted
3 years ago
a52a614475
Remove layer experiment which should not have been added
3 years ago
d829858550
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
3 years ago
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
3 years ago
4d7a5544f7
Remove inplace sigmoid for consistency with other impl
3 years ago
88a5b54802
A few small evonorm tweaks for convergence comparisons
3 years ago
66daee4f31
Last change wasn't complete, missed adding full evo_norm changeset
3 years ago
7bbbd5ef1b
EvoNorm and GroupNormAct options for debugging TPU / XLA concerns
3 years ago
d04f2f1377
Update drop_path and drop_block (fast impl) to be symbolically traceable, slightly faster
3 years ago
834a9ec721
Disable use of timm nn.Linear wrapper since AMP autocast + torchscript use appears fixed
3 years ago
78912b6375
Updated EvoNorm implementations with some experimentation. Add FilterResponseNorm. Updated RegnetZ and ResNetV2 model defs for trials.
3 years ago
480c676ffa
Fix FX breaking assert in evonorm
3 years ago
93cc08fdc5
Make evonorm variables 1d to match other PyTorch norm layers, will break weight compat for any existing use (likely minimal, easy to fix).
3 years ago
af607b75cc
Prep a set of ResNetV2 models with GroupNorm, EvoNormB0, EvoNormS0 for BN free model experiments on TPU and IPU
3 years ago
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.
3 years ago
b25ff96768
wip - pre-rebase
3 years ago
e051dce354
Make all models FX traceable
3 years ago
0149ec30d7
wip - attempting to rebase
3 years ago
bc3d4eb403
wip -rebase
3 years ago
2ddef942b9
Better fix for #954 that doesn't break torchscript, pull torch._assert into timm namespace when it exists
3 years ago
4f0f9cb348
Fix #954 by bringing traceable _assert into timm to allow compat w/ PyTorch < 1.8
3 years ago
b745d30a3e
Fix formatting of last commit
3 years ago
3478f1d7f1
Traceability fix for vit models for some experiments
3 years ago
f658a72e72
Cleanup re-use of Dropout modules in Mlp modules after some twitter feedback :p
3 years ago
c02334d9fa
Add weights for regnetz_d and haloregnetz_c, update regnetz_c weights. Add commented PyTorch XLA code for halo attention
3 years ago
02daf2ab94
Add option to include relative pos embedding in the attention scaling as per references. See discussion #912
3 years ago
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
3 years ago
007bc39323
Some halo and bottleneck attn code cleanup, add halonet50ts weights, use optimal crop ratios
3 years ago
b1c2e3eb92
Match rel_pos_indices attr rename in conv branch
3 years ago
b49630a138
Add relative pos embed option to LambdaLayer, fix last transpose/reshape.
3 years ago
b81e79aae9
Fix bottleneck attn transpose typo, hopefully these train better now..
3 years ago
515121cca1
Use reshape instead of view in std_conv, causing issues in recent PyTorch in channels_last
3 years ago
5bd04714e4
Cleanup weight init for byob/byoanet and related
3 years ago
8642401e88
Swap botnet 26/50 weights/models after realizing a mistake in arch def, now figuring out why they were so low...
3 years ago
5f12de4875
Add initial AttentionPool2d that's being trialed. Fix comment and still trying to improve reliability of sgd test.
3 years ago
492c0a4e20
Update HaloAttn comment
3 years ago
3b9032ea48
Use Tensor.unfold().unfold() for HaloAttn, fast like as_strided but more clarity
3 years ago
8449ba210c
Improve performance of HaloAttn, change default dim calc. Some cleanup / fixes for byoanet. Rename resnet26ts to tfs to distinguish (extra fc).
3 years ago
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...
3 years ago
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.
3 years ago
8165cacd82
Realized LayerNorm2d won't work in all cases as is, fixed.
3 years ago
b9cfb64412
Support npz custom load for vision transformer hybrid models. Add posembed rescale for npz load.
3 years ago
8319e0c373
Add file docstring to std_conv.py
3 years ago
Ross Wightman