b401952caf
Add newly added vision transformer large/base 224x224 weights ported from JAX official repo
4 years ago
61200db0ab
in_chans=1 working w/ pretrained weights for vision_transformer
4 years ago
e90edce438
Support native silu activation (aka swish). An optimized ver is available in PyTorch 1.7.
4 years ago
da6cd2cc1f
Fix regression for pretrained classifier loading when using entrypt functions directly
4 years ago
f591e90b0d
Make sure num_features attr is present in vit models as with others
4 years ago
f944242cb0
Fix #262 , num_classes arg mixup. Make vision_transformers a bit closer to other models wrt get/reset classfier/forward_features. Fix torchscript for ViT.
4 years ago
736f209e7d
Update vision transformers to be compatible with official code. Port official ViT weights from jax impl.
4 years ago
27a93e9de7
Improve test crop for ViT models. Small now 77.85, added base weights at 79.35 top-1.
4 years ago
d4db9e7977
Add small vision transformer weights. 77.42 top-1.
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
e8e2d9cabf
Add DropPath (stochastic depth) to ReXNet and VoVNet. RegNet DropPath impl tweak and dedupe se args.
4 years ago
e8ca45854c
More models in sotabench, more control over sotabench run, dataset filename extraction consistency
4 years ago
9c406532bd
Add EfficientNet-EdgeTPU-M (efficientnet_em) model trained natively in PyTorch. More sotabench fiddling.
4 years ago
c40384f5bd
Add ResNet weights. 80.5 (top-1) ResNet-50-D, 77.1 ResNet-34-D, 72.7 ResNet-18-D.
4 years ago
33f8a1bf36
Updated README, add wide_resnet50_2 and seresnext50_32x4d weights
4 years ago
751b0bba98
Add global_pool (--gp) arg changes to allow passing 'fast' easily for train/validate to avoid channels_last issue with AdaptiveAvgPool
4 years ago
9c297ec67d
Cleanup Apex vs native AMP scaler state save/load. Cleanup CheckpointSaver a bit.
4 years ago
80c9d9cc72
Add 'fast' global pool option, remove redundant SEModule from tresnet, normal one is now 'fast'
4 years ago
90a01f47d1
hrnet features_only pretrained weight loading issue. Fix #232 .
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
470220b1f4
Fix MobileNetV3 crash with global_pool='', output consistent with other models but not equivalent due to efficient head.
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
0f5d9d8166
Add CSPResNet50 weights, 79.6 top-1 at 256x256
4 years ago
b1b6e7c361
Fix a few more issues related to #216 w/ TResNet (space2depth) and FP16 weights in wide resnets. Also don't completely dump pretrained weights in in_chans != 1 or 3 cases.
4 years ago
512b2dd645
Add new EfficientNet-B3 and RegNetY-3.2GF weights, both just over 82 top-1
4 years ago
6890300877
Add DropPath (stochastic depth) to RegNet
4 years ago
f6b56602f9
fix test_model_default_cfgs
4 years ago
d5145fa4d5
Change default_cfg names for senet to include the legacy and match model names
4 years ago
b1f1a54de9
More uniform treatment of classifiers across all models, reduce code duplication.
4 years ago
d72ddafe56
Fix some checkpoint / model str regressions
4 years ago
ac18adb9c3
Remove debug print from RexNet
4 years ago
ec4976fdba
Add EfficientNet-Lite0 weights trained with this code by @hal-314, 75.484 top-1
4 years ago
9ecd16bd7b
Add new seresnet50 (non-legacy) model weights, 80.274 top-1
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
a69c0e04f0
Fix pool size in cspnet
4 years ago
14ef7a0dd6
Rename csp.py -> cspnet.py
4 years ago
ec37008432
Add pretrained weight links to CSPNet for cspdarknet53, cspresnext50
4 years ago
08016e839d
Cleanup FeatureInfo getters, add TF models sourced Xception41/65/71 weights
4 years ago
7ba5a384d3
Add ReXNet w/ remapped weights, feature support
4 years ago
c9d54bc1c3
Add HRNet feature extraction, fix senet type, lower feature testing res to 96x96
4 years ago
2ac663f340
Add feature support to legacy senets, add 32x32 resnext models to exclude list for feature testing.
4 years ago
c146b54abc
Cleanup EfficientNet/MobileNetV3 feature extraction a bit, only two tap locations now, small mobilenetv3 models work
4 years ago
68fd8a267b
Merge branch 'master' into features
4 years ago
4e61c6a12d
Cleanup, refactoring of Feature extraction code, add tests, fix tests, non hook feature extraction working with torchscript
4 years ago
6eec3fb4a4
Move FeatureHooks into features.py, switch EfficientNet, MobileNetV3 to use build model helper
4 years ago
9eba134d79
More models supporting feature extraction, xception, gluon_xception, inception_v3, inception_v4, pnasnet, nasnet, dla. Fix DLA unused projection params.
4 years ago
298fba09ac
Back out some activation hacks trialing upcoming pytorch changes
5 years ago
3b9004bef9
Lots of changes to model creation helpers, close to finalizing feature extraction / interfaces
5 years ago
e2cc481310
Update CSP ResNets for cross expansion without activation. Fix VovNet IABN compatibility with fixed activation arg.
5 years ago
3b6cce4c95
Add initial impl of CrossStagePartial networks, yet to be trained, not quite the same as darknet cfgs.
5 years ago
3aebc2f06c
Switch DPN to use BnAct layer, train a new DPN 68b model with RA to 79.21
5 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
5 years ago
a66df5fb91
More model feature extraction support, start to deprecate senet.py, dilations added to regnet, add proper aligned xception
5 years ago
7729f40dca
Fix another bug, update all gluon resnet models to use new creation method (feature support)
5 years ago
d72ac0db25
Fix #173 , lr cycle default 0 vs 1. Fix #177 , mirror resnest weights for future stability.
5 years ago
d0113f9cdb
Fix a few issues that came up in tests
5 years ago
d23a2697d0
Working on feature extraction, interfaces refined, a number of models working, some in progress.
5 years ago
328339a086
Add ESE-VoVNet-19-DW weights
5 years ago
2d83752bc2
Fix default interpolation/crop of largest 2 ResNeSt models
5 years ago
39f27c1add
Almost ready to merge, vovnet typo, version bump, readme addition
5 years ago
151679c2f1
Add custom grad tests, fix cut & paste error with hard_mish ME, add a few more pytorch act fns to factory
5 years ago
6c7932fe75
Update sotabench.py, tweak VovNet cfg
5 years ago
0aca08384f
Update regnet cfg keys to match model names so registry works properly
5 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
5 years ago
13c0a6290e
Add ESE-VovNet39 weights, 79.32 top-1
5 years ago
4ddde1d3a4
Fix two regressions
5 years ago
a7e8cadd15
Remove pointless densenet configs, add an iabn version of 264 as it makes more sense to try someday...
5 years ago
e78daf586a
better densenet121 and densenetblur121d weights
5 years ago
7be299504f
Add missing feature_info() on MobileNetV3, make hook feature output order/type consistent with bottleneck (list, decreasing fmap size)
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
6441e9cc1b
Fix memory_efficient mode for DenseNets. Add AntiAliasing (Blur) support for DenseNets and create one test model. Add lr cycle/mul params to train args.
5 years ago
afb6bd0669
Add backward and default_cfg tests and fix a few issues found. Fix #153
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
c4ca016656
Merge pull request #145 from rwightman/resnest
...
ResNeSt
5 years ago
208e7912f7
Missed one of the abalation model entrypoints, update README
5 years ago
9cc289f18c
Exclude EfficientNet-L2 models from test
5 years ago
2f884a0ce5
Add resnest14, resnest26, and two of the abalation grouped resnest50 models
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
022ed001f3
Update DenseNet to latest in Torchvision (torchscript compat, checkpointing, proper init). Start adding ehanced configurability, stem options...
5 years ago
a7ebe09029
Replace all None by nn.Identity() in all models reset_classifier when False-values num_classes is given.
...
Make small code refactoring
5 years ago
f0eb021620
Replace all None by nn.Identity() in HRNet modules
5 years ago
8d8677e03b
Fix #139 . Broken SKResNets after BlurPool addition, as a plus, SKResNets support AA now too.
5 years ago
9c15d57505
added efficientnet pruned weights
5 years ago
ea300709f0
Fix model create fn not passing num_classes through. Fix #135
5 years ago
2c438c4e0d
Merge branch 'master' of github.com:rwightman/pytorch-models
5 years ago
2681a8d618
Final blurpool2d cleanup and add resnetblur50 weights, match tresnet Downsample arg order to BlurPool2d for interop
5 years ago
9590f301a9
Merge branch 'blur' of https://github.com/VRandme/pytorch-image-models into VRandme-blur
5 years ago
a5220ad3aa
Merge pull request #125 from Separius/patch-1
...
fix typo in eca
5 years ago
0834fbc01c
Move pruned model adapt strings to separate txt files. A few minor formatting alignment tweaks.
5 years ago
ba793f5c1c
Merge branch 'adding_ECA_resnet' of https://github.com/yoniaflalo/pytorch-image-models into yoniaflalo-adding_ECA_resnet
5 years ago
07f19dd699
added eca resnet
5 years ago
be7c784d21
Small TResNet simplification, just use SelectAdaptivePool, don't notice any perf difference
5 years ago
e3a98171b2
Remove some CRLF that made it into TResNet merge
5 years ago
0004f37d25
Some TResNet cleanup.
...
* allow use of global pool arg, test-time-pooling
* clean checkpoints to just contain state dict, add 448 res checkpoints
* support DataParallel via lazy filter creation for JIT Downsample
* some minor formatting (mostly alignment) preferences
5 years ago
20290b5b56
fix typo in eca
5 years ago
ebf82b84ac
Merge pull request #122 from mrT23/master
...
TResNet models
5 years ago
bdb165a8a4
Merge changes in feature extraction interface to MobileNetV3
...
Experimental feature extraction interface seems to be changed
a little bit with the most up to date version apparently found
in EfficientNet class. Here these changes are added to
MobileNetV3 class to make it support it and work again, too.
5 years ago
8a63c1add8
finalizing
5 years ago
6209146738
TResNet models
5 years ago
1a8f5900ab
Update EfficientNet feature extraction for EfficientDet. Add needed MaxPoolSame as well.
5 years ago
06a50a94a8
Fix minor typos in create_attn.py and resnet.py
...
'eca'->'ceca'
and
doest not-> does not
5 years ago
1a9ab07307
One too many changes at a time, fix missing C
5 years ago
f17b42bc33
Blur filter no longer a buffer
5 years ago
6cdeca24a3
Some cleanup and fixes for initial BlurPool impl. Still some testing and tweaks to go...
5 years ago
acd1b6cccd
Implement Functional Blur on resnet.py
...
1. add ResNet argument blur=''
2. implement blur for maxpool and strided convs in downsampling blocks
5 years ago
ce3d82b58b
Update blurpool.py
...
clean up code for PR
5 years ago
3a287a6e76
Create blurpool.py
...
Initial implementation of blur layer.
currently tests as correct against Downsample of original github
5 years ago
c99a5abed4
Merge pull request #115 from rwightman/mobilenetv2-experiment
...
MobileNet-V2 experiments
5 years ago
e34074b4da
Add final weights for MobileNet-V2 experiments
5 years ago
a6a5565de7
Fiddling...
5 years ago
5a16c533ff
Add better resnext50_32x4d weights trained by andravin
5 years ago
bc998cad91
Experimenting with some MobileNetV2 variations to compare against EfficientNet-Lite
5 years ago
3406e582cf
Add EfficientNet-Lite results, update README
5 years ago
bd05258f7b
EfficientNet-Lite model added w/ converted checkpoints, validation in progress...
5 years ago
56e2ac3a6d
Merge pull request #94 from rwightman/lr_noise
...
Learning rate noise, MobileNetV3 weights, and activate MobileNetV3/EfficientNet weight init change
5 years ago
c16f25ced2
Add MobileNetV3 Large weights, results, update README and sotabench for merge
5 years ago
c60069c1eb
Annotate types on drop fns to avoid torchscript error
5 years ago
9fee316752
Enable fixed fanout calc in EfficientNet/MobileNetV3 weight init by default. Fix #84
5 years ago
43225d110c
Unify drop connect vs drop path under 'drop path' name, switch all EfficientNet/MobilenetV3 refs to 'drop_path'. Update factory to handle new drop args.
5 years ago
f1d5f8a6c4
Update comments for Selective Kernel and DropBlock/Path impl, add skresnet34 weights
5 years ago
569419b38d
Tweak some comments, add SKNet models with weights to sotabench, remove an unused branch
5 years ago
53c47479c4
Batch validation batch size adjustment, tweak L2 crop pct
5 years ago
08553e16b3
Merge branch 'master' into attention
5 years ago
fa38f24967
Update SK network configs, add weights for skresnet8 and skresnext50
5 years ago
ba15ca47e8
Add ported EfficientNet-L2, B0-B7 NoisyStudent weights from TF TPU
5 years ago
5e6dbbaf30
Add CBAM for experimentation
5 years ago
d725991870
Remove debug print from ECA module
5 years ago
2a7d256fd5
Re-enable mem-efficient/jit activations after torchscript tests
5 years ago
f902bcd54c
Layer refactoring continues, ResNet downsample rewrite for proper dilation in 3x3 and avg_pool cases
...
* select_conv2d -> create_conv2d
* added create_attn to create attention module from string/bool/module
* factor padding helpers into own file, use in both conv2d_same and avg_pool2d_same
* add some more test eca resnet variants
* minor tweaks, naming, comments, consistency
5 years ago
a99ec4e7d1
A bunch more layer reorg, splitting many layers into own files. Improve torchscript compatibility.
5 years ago
13746a33fc
Big move, layer modules and fn to timm/models/layers
5 years ago
f54612f648
Merge branch 'select_kernel' into attention
5 years ago
4defbbbaa8
Fix module name mistake, start layers sub-package
5 years ago
7011cd0902
A little bit of ECA cleanup
5 years ago
46471df7b2
Merge pull request #82 from VRandme/eca
...
ECA-Net Efficient Channel Attention
5 years ago
d0eb59ef46
Remove unused default_init for EfficientNets, experimenting with fanout calc for #84
5 years ago
e6a762346a
Implement Adaptive Kernel selection
...
When channel size is given,
calculate adaptive kernel size according to original paper.
Otherwise use the given kernel size(k_size), which defaults to 3
5 years ago
13e8da2b46
SelectKernel split_input works best when input channels split like grouped conv, but output is full width. Disable zero_init for SK nets, seems a bad combo.
5 years ago
6db087a1ff
Merge remote-tracking branch 'upstream/master' into eca
5 years ago
Ross Wightman