fa38f24967Update SK network configs, add weights for skresnet8 and skresnext50
Ross Wightman
2020-02-14 15:37:00 -0800
f098fda2caAdd map_location='cpu' to ModelEma resume, should improve #72
Ross Wightman
2020-02-12 13:23:56 -0800
b9496995d8Add L2-475 PyTorch preprocessing result, update sotabench for new weights
Ross Wightman
2020-02-12 13:20:37 -0800
ba15ca47e8Add ported EfficientNet-L2, B0-B7 NoisyStudent weights from TF TPU
Ross Wightman
2020-02-12 11:26:38 -0800
5d2df0746erefactor cecabam with cbam
Chris Ha
2020-02-11 22:55:07 +0900
1448a19d67Update .gitignore
Chris Ha
2020-02-11 22:27:06 +0900
26922eb2ceMerge branch 'attention' into EcaBam
Chris Ha
2020-02-11 22:26:53 +0900
d5a577bb84Merge remote-tracking branch 'upstream/master'
Chris Ha
2020-02-11 22:25:57 +0900
04a29656dcUpdate .gitignore
Chris Ha
2020-02-11 22:25:41 +0900
a4b1993c84Restore Bias
Chris Ha
2020-02-11 19:35:06 +0900
5e6dbbaf30Add CBAM for experimentation
Ross Wightman
2020-02-10 16:23:09 -0800
d725991870Remove debug print from ECA module
Ross Wightman
2020-02-10 16:21:33 -0800
2a7d256fd5Re-enable mem-efficient/jit activations after torchscript tests
Ross Wightman
2020-02-10 11:59:36 -0800
f902bcd54cLayer 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
Ross Wightman
2020-02-10 11:55:03 -0800
a99ec4e7d1A bunch more layer reorg, splitting many layers into own files. Improve torchscript compatibility.
Ross Wightman
2020-02-09 14:46:28 -0800
13746a33fcBig move, layer modules and fn to timm/models/layers
Ross Wightman
2020-02-09 13:13:08 -0800
f54612f648Merge branch 'select_kernel' into attention
Ross Wightman
2020-02-09 12:59:24 -0800
4defbbbaa8Fix module name mistake, start layers sub-package
Ross Wightman
2020-02-09 12:44:26 -0800
7011cd0902A little bit of ECA cleanup
Ross Wightman
2020-02-09 12:41:18 -0800
46471df7b2Merge pull request #82 from VRandme/eca
Ross Wightman
2020-02-09 12:31:05 -0800
d0eb59ef46Remove unused default_init for EfficientNets, experimenting with fanout calc for #84
Ross Wightman
2020-02-09 11:32:05 -0800
cade829105Add EfficientNet-ES to sotabench
Ross Wightman
2020-02-09 11:04:48 -0800
dbdb3ce963Add CecaBam
Chris Ha
2020-02-09 23:09:13 +0900
e6a762346aImplement Adaptive Kernel selection
#82
Chris Ha
2020-02-09 11:58:03 +0900
13e8da2b46SelectKernel 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.
Ross Wightman
2020-02-07 22:42:04 -0800
99142b378eMerge remote-tracking branch 'upstream/master'
Chris Ha
2020-02-07 19:36:53 +0900
6db087a1ffMerge remote-tracking branch 'upstream/master' into eca
Chris Ha
2020-02-07 19:36:35 +0900
904c618040Update EcaModule.py
Chris Ha
2020-02-07 19:36:18 +0900
db91ba053bEcaModule(CamelCase)
Chris Ha
2020-02-07 19:28:07 +0900
Merge pull request #83 from andravin/validation-batch-size-multiplier
Ross Wightman
2020-02-04 22:49:42 -0800
f87fcd7e88Implement Eca modules
Chris Ha
2020-02-04 23:15:29 +0900
697e05cb3eMerge remote-tracking branch 'upstream/master'
Chris Ha
2020-02-04 21:02:55 +0900
b72013def8Added commandline argument validation-batch-size-multiplier with default set to 1.
#83
Andrew Lavin
2020-02-04 01:02:27 -0800
4808b3c32fBump version for PyPi update, fix few out of date README items/mistakes, add README updates for TF EfficientNet-B8 (RandAugment)
Ross Wightman
2020-02-03 11:43:14 -0800
Add results/README.md
Ross Wightman
2020-02-02 19:35:44 -0800
89a32454e1Merge remote-tracking branch 'upstream/master'
Chris Ha
2020-02-02 23:05:49 +0900
7d07ebb660Adding some configs to sknet, incl ResNet50 variants from 'Compounding ... Assembled Techniques' paper and original SKNet50
Ross Wightman
2020-02-01 23:28:48 -0800
a9d2424fd1Add separate zero_init_last_bn function to support more block variety without a mess
Ross Wightman
2020-01-30 16:51:49 -0800
355aa152d5Just leave it float for now, will look at fp16 later. Remove unused reference code.
Ross Wightman
2020-01-29 14:51:34 -0800
ef457555d3BlockDrop working on GPU
Ross Wightman
2020-01-29 14:34:45 -0800
3ff19079f9Missed nn_ops.py from last commit
Ross Wightman
2020-01-29 13:11:38 -0800
9f11b4e8a2Add ConvBnAct layer to parallel integrated SelectKernelConv, add support for DropPath and DropBlock to ResNet base and SK blocks
Ross Wightman
2020-01-29 13:01:35 -0800
cefc9b7761Move SelectKernelConv to conv2d_layers and more * always apply attention in SelectKernelConv, leave MixedConv for no attention alternative * make MixedConv torchscript compatible * refactor first/previous dilation name to make more sense in ResNet* networks
Ross Wightman
2020-01-27 21:48:28 -0800
9abe610931Used wrong channel var for split
Ross Wightman
2020-01-26 11:33:31 -0800
58e28dc7e7Move Selective Kernel blocks/convs to their own sknet.py file
Ross Wightman
2020-01-26 11:23:39 -0800
a93bae6dc5A SelectiveKernelBasicBlock for more experiments
Ross Wightman
2020-01-25 18:31:08 -0800
ad087b4b17Missed bias=False in selection conv
Ross Wightman
2020-01-24 19:54:37 -0800
c8b3d6b81aInitial impl of Selective Kernel Networks. Very much a WIP.
Ross Wightman
2020-01-24 19:45:05 -0800
fd98fb33c5Update sotabench with tf_efficientnet_b8 model
Ross Wightman
2020-02-01 18:12:25 -0800
7c88356682Add update results on ImageNet validation, ImageNetV2, ImageNet-A, and ImageNet-Sketch for all models
Ross Wightman
2020-02-01 18:10:13 -0800
1daa303744Add support to Dataset for class id mapping file, clean up a bit of old logic. Add results file arg for validation and update script.
Ross Wightman
2020-02-01 18:07:32 -0800
Update README.md
Ross Wightman
2020-01-13 17:49:49 -0800
7a17ee9c3bUpdate README.md
Ross Wightman
2020-01-12 18:16:36 -0800
e42503e1d0Update sotabench
Ross Wightman
2020-01-12 18:00:12 -0800
2f41905ba5Update ResNet50 weights to AuxMix trained 78.994 top-1. A few commentes re 'tiered_narrow' tn variant.
Ross Wightman
2020-01-12 17:55:58 -0800
cc0b1f4130Add support to clean_checkpoint.py to remove aux_bn weights/biases from SplitBatchNorm
Ross Wightman
2020-01-12 17:52:19 -0800
Update README.md
Ross Wightman
2020-01-07 08:52:12 -0800
833066b540A few minor things in SplitBN
Ross Wightman
2020-01-05 20:07:03 -0800
7547119891Add SplitBatchNorm. AugMix, Rand/AutoAugment, Split (Aux) BatchNorm, Jensen-Shannon Divergence, RandomErasing all working together
Ross Wightman
2020-01-05 19:58:59 -0800
2e955cfd0cUpdate RandomErasing with some improved arg names, tweak to aspect range
Ross Wightman
2020-01-05 14:31:48 -0800
3cc0f91e23Fix augmix variable name scope overlap, default non-blended mode
Ross Wightman
2020-01-05 14:27:27 -0800
Update README.md
Ross Wightman
2020-01-04 11:51:38 -0800
ec0dd4053aAdd updated RandAugment trained EfficientNet-B0 trained weights from @michaelklachko
Ross Wightman
2020-01-03 17:18:46 -0800
8662454bd7Plural for averaging script.
Ross Wightman
2020-01-03 15:00:20 -0800
40fea63ebeAdd checkpoint averaging script. Add headers, shebangs, exec perms to all scripts
Ross Wightman
2020-01-03 14:56:51 -0800
4666cc9aedAdd --pin-mem arg to enable dataloader pin_memory (showing more benefit in some scenarios now), also add --torchscript arg to validate.py for testing models with jit.script
Ross Wightman
2020-01-02 16:22:06 -0800
53001dd292ResNet / Res2Net additions: * ResNet torchscript compat * output_stride arg supported to limit network stride via dilations (support for dilation added to Res2Net) * allow activation layer to be changed via act_layer arg
Ross Wightman
2020-01-01 17:15:56 -0800
f96b3e5e92InceptionResNetV2 torchscript compatible
Ross Wightman
2020-01-01 17:13:37 -0800
19d93fe454Add selecsls60 weights
Ross Wightman
2019-12-31 16:49:04 -0800
Merge pull request #66 from rwightman/selecsls_updates
Ross Wightman
2019-12-30 16:07:20 -0800
84ca3d1f4dAdd SelecSLS to sotabench list
#66
Ross Wightman
2019-12-30 16:04:39 -0800
0062c15fb0Update checkpoint url with modelzoo compatible ones.
Ross Wightman
2019-12-30 15:59:19 -0800
b5315e66b5Streamline SelecSLS model without breaking checkpoint compat. Move cfg handling out of model class. Update feature/pooling behaviour to match current.
Ross Wightman
2019-12-30 15:44:47 -0800
d59a756c16Run PyCharm autoformat on selecsls and change mix cap variables and model names to all lower
Ross Wightman
2019-12-30 14:30:46 -0800