eca_halonext26ts
- 79.5 @ 256resnet50_gn
(new) - 80.1 @ 224, 81.3 @ 288resnet50
- 80.7 @ 224, 80.9 @ 288 (trained at 176, not replacing current a1 weights as default since these don't scale as well to higher res, weights)resnext50_32x4d
- 81.1 @ 224, 82.0 @ 288sebotnet33ts_256
(new) - 81.2 @ 224lamhalobotnet50ts_256
- 81.5 @ 256halonet50ts
- 81.7 @ 256halo2botnet50ts_256
- 82.0 @ 256resnet101
- 82.0 @ 224, 82.8 @ 288resnetv2_101
(new) - 82.1 @ 224, 83.0 @ 288resnet152
- 82.8 @ 224, 83.5 @ 288regnetz_d8
(new) - 83.5 @ 256, 84.0 @ 320regnetz_e8
(new) - 84.5 @ 256, 85.0 @ 320vit_base_patch8_224
(85.8 top-1) & in21k
variant weights added thanks Martins Bruveristimm bits
branch).data
, a bit more consistency, unit tests for all!efficientnetv2_rw_t
weights, a custom 'tiny' 13.6M param variant that is a bit better than (non NoisyStudent) B3 models. Both faster and better accuracy (at same or lower res)
vit_base_patch16_sam_224
) and B/32 (vit_base_patch32_sam_224
) models.jx_nest_base
- 83.534, jx_nest_small
- 83.120, jx_nest_tiny
- 81.426gmlp_s16_224
trained to 79.6 top-1, matching paper. Hparams for this and other recent MLP training herevit_large_patch16_384
(87.1 top-1), vit_large_r50_s32_384
(86.2 top-1), vit_base_patch16_384
(86.0 top-1)vit_deit_*
renamed to just deit_*
gmixer_24_224
MLP /w GLU, 78.1 top-1 w/ 25M params.eca_nfnet_l2
weights from my 'lightweight' series. 84.7 top-1 at 384x384.efficientnetv2_rw_m
model and weights (started training before official code). 84.8 top-1, 53M params.tf_efficientnetv2_s/m/l
tf_efficientnetv2_s/m/l_in21k
tf_efficientnetv2_s/m/l_in21ft1k
tf_efficientnetv2_b0
through b3
efficientnet_v2s
-> efficientnetv2_rw_s
efficientnetv2_*
models in-place for future native PyTorch trainingswinnet
benchmark.py
script for bulk timm
model benchmarking of train and/or inferencetimm
cleanup/style tweaks and weights have hub download supportnfnet_l0b
->nfnet_l0
) weights 82.75 top-1 @ 288x288dm_
. They require SAME padding conv, skipinit enabled, and activation gains applied in act fn.s
variants.dm_nfnet_f6
- 86.352dm_nfnet_f5
- 86.100dm_nfnet_f4
- 85.834dm_nfnet_f3
- 85.676dm_nfnet_f2
- 85.178dm_nfnet_f1
- 84.696dm_nfnet_f0
- 83.464--clip-grad .01 --clip-mode agc
--clip-grad 1.0
--clip-grad 10. --clip-mode value
byobnet.py
byobnet.py
vgg
)--channels-last
and --torchscript
model training, APEX does not.ecaresnet26t
- 79.88 top-1 @ 320x320, 79.08 @ 256x256ecaresnet50t
- 82.35 top-1 @ 320x320, 81.52 @ 256x256ecaresnet269d
- 84.93 top-1 @ 352x352, 84.87 @ 320x320t
) vs tiered_narrow (tn
) ResNet model defs, all tn
changed to t
and t
models removed (seresnext26t_32x4d
only model w/ weights that was removed).test_input_size
and remove extra _320
suffix ResNet model defs that were just for test.train.py /data/tfds --dataset tfds/oxford_iiit_pet --val-split test --model resnet50 -b 256 --amp --num-classes 37 --opt adamw --lr 3e-4 --weight-decay .001 --pretrained -j 2
validate.py /data/fall11_whole.tar --model resnetv2_50x1_bitm_in21k --amp
efficientnet_em
) model trained in PyTorch, 79.3 top-1--channels-last
, --native-amp
vs --apex-amp
)in_chans
!= 3 on several models.Universal feature extraction, new models, new weights, new test sets.
features_only=True
argument for create_model
call to return a network that extracts feature maps from the deepest layer at each stride.results/README.md
Bunch of changes:
fix_group_fanout=False
in _init_weight_goog
fn if you need to reproducte past behaviourlayers
subfolder/module of models
and organize in a more granular fashion.se
position for all ResNetsseresnext26tn_32x4d
- 77.99 top-1, 93.75 top-5 added to tiered experiment, higher img/s than 't' and 'd'avg_checkpoints.py
script for post training weight averaging and update all scripts with header docstrings and shebangs.efficientnet_b3
- 81.5 top-1, 95.7 top-5 at default res/crop, 81.9, 95.8 at 320x320 1.0 crop-pct
seresnext26d_32x4d
- 77.6 top-1, 93.6 top-5
seresnext26t_32x4d
- 78.0 top-1, 93.7 top-5
--dist-bn
argument added to train.py, will distribute BN stats between nodes after each train epoch, before evalreset_classifer
, and forward_features
across models
forward_features
always returns unpooled feature maps nowdrop-connect
cmd line arg finally added to train.py
, no need to hack model fns. Works for efficientnet/mobilenetv3 based models, ignored otherwise.Dear OpenI User
Thank you for your continuous support to the Openl Qizhi Community AI Collaboration Platform. In order to protect your usage rights and ensure network security, we updated the Openl Qizhi Community AI Collaboration Platform Usage Agreement in January 2024. The updated agreement specifies that users are prohibited from using intranet penetration tools. After you click "Agree and continue", you can continue to use our services. Thank you for your cooperation and understanding.
For more agreement content, please refer to the《Openl Qizhi Community AI Collaboration Platform Usage Agreement》