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
4d96165989
Merge branch 'master' into cleanup_xla_model_fixes
3 years ago
8880f696b6
Refactoring, cleanup, improved test coverage.
...
* Add eca_nfnet_l2 weights, 84.7 @ 384x384
* All 'non-std' (ie transformer / mlp) models have classifier / default_cfg test added
* Fix #694 reset_classifer / num_features / forward_features / num_classes=0 consistency for transformer / mlp models
* Add direct loading of npz to vision transformer (pure transformer so far, hybrid to come)
* Rename vit_deit* to deit_*
* Remove some deprecated vit hybrid model defs
* Clean up classifier flatten for conv classifiers and unusual cases (mobilenetv3/ghostnet)
* Remove explicit model fns for levit conv, just pass in arg
3 years ago
ba2ca4b464
One codepath for stdconv, switch layernorm to batchnorm so gain included. Tweak epsilon values for nfnet, resnetv2, vit hybrid.
3 years ago
b7a568f065
Fix torchscript issue in bat
3 years ago
8e4ac3549f
All ScaledStdConv and StdConv uses default to using F.layernorm so that they work with PyTorch XLA. eps value tweaking is a WIP.
3 years ago
bda8ab015a
Remove min channels for SelectiveKernel, divisor should cover cases well enough.
4 years ago
a27f4aec4a
Missed args for skresnext w/ refactoring.
4 years ago
307a935b79
Add non-local and BAT attention. Merge attn and self-attn factories into one. Add attention references to README. Add mlp 'mode' to ECA.
4 years ago
8bf63b6c6c
Able to use other attn layer in EfficientNet now. Create test ECA + GC B0 configs. Make ECA more configurable.
4 years ago
9611458e19
Throw in some FBNetV3 code I had lying around, some refactoring of SE reduction channel calcs for all EffNet archs.
4 years ago
f615474be3
Fix broken test, repvgg block doesn't have attn_last attr.
4 years ago
742c2d5247
Add Gather-Excite and Global Context attn modules. Refactor existing SE-like attn for consistency and refactor byob/byoanet for less redundancy.
4 years ago
9c78de8c02
Fix #661 , move hardswish out of default args for LeViT. Enable native torch support for hardswish, hardsigmoid, mish if present.
4 years ago
f45de37690
Merge branch 'master' into levit_visformer_rednet
4 years ago
d5af752117
Add preliminary gMLP and ResMLP impl to Mlp-Mixer
4 years ago
3bffc701f1
Merge branch 'master' into levit_visformer_rednet
4 years ago
ecc7552c5c
Add levit, levit_c, and visformer model defs. Largely untested and not finished cleanup.
4 years ago
165fb354b2
Add initial RedNet model / Involution layer impl for testing
4 years ago
c4f482a08b
EfficientNetV2 official impl w/ weights ported from TF. Cleanup/refactor of related EfficientNet classes and models.
4 years ago
715519a5ef
Rethink name of patch embed grid info
4 years ago
b2c305c2aa
Move Mlp and PatchEmbed modules into layers. Being used in lots of models now...
4 years ago
0721559511
Improved (hopefully) init for SA/SA-like layers used in ByoaNets
4 years ago
0d87650fea
Remove filter hack from BlurPool w/ non-persistent buffer. Use BlurPool2d instead of AntiAliasing.. for TResNet. Breaks PyTorch < 1.6.
4 years ago
9cc7dda6e5
Fixup byoanet configs to pass unit tests. Add swin_attn and swinnet26t model for testing.
4 years ago
e15c3886ba
Defaul lambda r=7. Define '26t' stage 4/5 256x256 variants for all of bot/halo/lambda nets for experiment. Add resnet50t for exp. Fix a few comments.
4 years ago
4e4b863b15
Missed norm.py
4 years ago
ce62f96d4d
ByoaNet with bottleneck transformer, lambda resnet, and halo net experiments
4 years ago
a5310a3451
Merge remote-tracking branch 'origin/benchmark-fixes-vit_hybrids' into pit_and_vit_update
4 years ago
cf5fec5047
Cleanup experimental vit weight init a bit
4 years ago
740f32c96a
Add ECA-NFNet-L0 weights and update model name. Update README and bump version to 0.4.6
4 years ago
f57db99101
Update README, fix iabn pip version print.
4 years ago
8563609b28
Update notes in ScaledStdConv impl
4 years ago
678ba4e0a2
Add NFNet-F model weights ported from DeepMind Haiku impl and new set of models w/ compatible config.
4 years ago
d8e69206be
Merge pull request #419 from rwightman/byob_vgg_models
...
More models, GPU-Efficient Nets, RepVGG, classic VGG, and flexible Byob backbone.
4 years ago
94ca140b67
update collections.abc import
4 years ago
1bcc69e0ad
Use in_channels for depthwise groups, allows using `out_channels=N * in_channels` (does not impact existing models). Fix #354 .
4 years ago
9811e229f7
Fix regression in models with 1001 class pretrained weights. Improve batchnorm arg and BatchNormAct layer handling in several models.
4 years ago
a39c3ee216
Merge branch 'master' into eca-weights
4 years ago
68a4144882
Add new weights for ecaresnet26t/50t/269d models. Remove distinction between 't' and 'tn' (tiered models), tn is now t. Add test time img size spec to default cfg.
4 years ago
b9843f954b
Merge pull request #282 from tigert1998/patch-1
...
Add symbolic for SwishJitAutoFn to support onnx
4 years ago
7a4be5c035
add operator HardSwishJitAutoFn export to onnx
4 years ago
90980de4a9
Fix up a few details in NFResNet models, managed stable training. Add support for gamma gain to be applied in activation or ScaleStdConv. Some tweaks to ScaledStdConv.
4 years ago
5a8e1e643e
Initial Normalizer-Free Reg/ResNet impl. A bit of related layer refactoring.
4 years ago
231d04e91a
ResNetV2 pre-act and non-preact model, w/ BiT pretrained weights and support for ViT R50 model. Tweaks for in21k num_classes passing. More to do... tests failing.
4 years ago
2ed8f24715
A few more changes for 0.3.2 maint release. Linear layer change for mobilenetv3 and inception_v3, support no bias for linear wrapper.
4 years ago
460eba7f24
Work around casting issue with combination of native torch AMP and torchscript for Linear layers
4 years ago
5f4b6076d8
Fix inplace arg compat for GELU and PreLU via activation factory
4 years ago
fd962c4b4a
Native SiLU (Swish) op doesn't export to ONNX
4 years ago
43f2500c26
Add symbolic for SwishJitAutoFn to support onnx
4 years ago
e90edce438
Support native silu activation (aka swish). An optimized ver is available in PyTorch 1.7.
4 years ago
f31933cb37
Initial Vision Transformer impl w/ patch and hybrid variants. Refactor tuple helpers.
4 years ago
fcb6258877
Add missing leaky_relu layer factory defn, update Apex/Native loss scaler interfaces to support unscaled grad clipping. Bump ver to 0.2.2 for pending release.
4 years ago
e8ca45854c
More models in sotabench, more control over sotabench run, dataset filename extraction consistency
4 years ago
80c9d9cc72
Add 'fast' global pool option, remove redundant SEModule from tresnet, normal one is now 'fast'
4 years ago
110a7c4982
AdaptiveAvgPool2d -> mean((2,3)) for all SE/attn layers to avoid NaN with AMP + channels_last layout. See https://github.com/pytorch/pytorch/issues/43992
4 years ago
fc8b8afb6f
Fix a silly bug in Sample version of EvoNorm missing x* part of swish, update EvoNormBatch to accumulated unbiased variance.
4 years ago
b1f1a54de9
More uniform treatment of classifiers across all models, reduce code duplication.
4 years ago
7995295968
Merge branch 'logger' into features. Change 'logger' to '_logger'.
4 years ago
1998bd3180
Merge branch 'feature/AB/logger' of https://github.com/antoinebrl/pytorch-image-models into logger
4 years ago
6c17d57a2c
Fix some attributions, add copyrights to some file docstrings
4 years ago
7ba5a384d3
Add ReXNet w/ remapped weights, feature support
4 years ago
298fba09ac
Back out some activation hacks trialing upcoming pytorch changes
4 years ago
3b9004bef9
Lots of changes to model creation helpers, close to finalizing feature extraction / interfaces
4 years ago
3b6cce4c95
Add initial impl of CrossStagePartial networks, yet to be trained, not quite the same as darknet cfgs.
4 years ago
3aebc2f06c
Switch DPN to use BnAct layer, train a new DPN 68b model with RA to 79.21
4 years ago
f122f0274b
Significant ResNet refactor:
...
* stage creation + make_layer moved to separate fn with more sensible dilation/output_stride calc
* drop path rate decay easy to impl with refactored block creation loops
* fix dilation + blur pool combo
4 years ago
d0113f9cdb
Fix a few issues that came up in tests
4 years ago
151679c2f1
Add custom grad tests, fix cut & paste error with hard_mish ME, add a few more pytorch act fns to factory
4 years ago
78fa0772cc
Leverage python hierachical logger
...
with this update one can tune the kind of logs generated by timm but
training and inference traces are unchanged
4 years ago
4ddde1d3a4
Fix two regressions
5 years ago
88129b2569
Add set_layer_config contextmgr to adjust all layer configs at once, use in create_module with new args. Remove a few old warning causing constant annotations for jit.
5 years ago
f28170df3f
Fix an an untested change, remove a debug print
5 years ago
eb7653614f
Monster commit, activation refactor, VoVNet, norm_act improvements, more
...
* refactor activations into basic PyTorch, jit scripted, and memory efficient custom auto
* implement hard-mish, better grad for hard-swish
* add initial VovNet V1/V2 impl, fix #151
* VovNet and DenseNet first models to use NormAct layers (support BatchNormAct2d, EvoNorm, InplaceIABN)
* Wrap IABN for any models that use it
* make more models torchscript compatible (DPN, PNasNet, Res2Net, SelecSLS) and add tests
5 years ago
0ea53cecc3
Merge branch 'master' into densenet_update_and_more
5 years ago
50658b9a67
Add RegNet models and weights
5 years ago
7df83258c9
Merge branch 'master' into densenet_update_and_more
5 years ago
1904ed8fec
Improve dropblock impl, add fast variant, and better AMP speed, inplace, batchwise... few ResNeSt cleanups
5 years ago
17270c69b9
Remove annoying InceptionV3 dependency on scipy and insanely slow trunc_norm init. Bring InceptionV3 code into this codebase and use upcoming torch trunch_norm_ init.
5 years ago
208e7912f7
Missed one of the abalation model entrypoints, update README
5 years ago
f4cdc2ac31
Add ResNeSt models
5 years ago
780860d140
Add norm_act factory method, move JIT of norm layers to factory
5 years ago
14edacdf9a
DenseNet converted to support ABN (norm + act) modules. Experimenting with EvoNorm, IABN
5 years ago
8d8677e03b
Fix #139 . Broken SKResNets after BlurPool addition, as a plus, SKResNets support AA now too.
5 years ago
Ross Wightman