b81e79aae9
Fix bottleneck attn transpose typo, hopefully these train better now..
3 years ago
80075b0b8a
Add worker_seeding arg to allow selecting old vs updated data loader worker seed for (old) experiment repeatability
3 years ago
6478bcd02c
Fix regnetz_d conv layer name, use inception mean/std
3 years ago
0387e6057e
Update binary cross ent impl to use thresholding as an option (convert soft targets from mixup/cutmix to 0, 1)
3 years ago
f8a63a3b71
Add worker_init_fn to loader for numpy seed per worker
3 years ago
515121cca1
Use reshape instead of view in std_conv, causing issues in recent PyTorch in channels_last
3 years ago
da06cc61d4
ResNetV2 seems to work best without zero_init residual
3 years ago
8e11da0ce3
Add experimental RegNetZ(ish) models for training / perf trials.
3 years ago
6bbc50beb4
make it possible to provide norm_layer via create_model
3 years ago
cf5ac2800c
BotNet models were still off, remove weights for bad configs. Add good SE-HaloNet33-TS weights.
3 years ago
24720abe3b
Merge branch 'master' into attn_update
3 years ago
1c9284c640
Add BeiT 'finetuned' 1k weights and pretrained 22k weights, pretraining specific (masked) model excluded for now
3 years ago
f8a215cfe6
A few more crossvit tweaks, fix training w/ no_weight_decay names, add crop option for scaling, adjust default crop_pct for large img size to 1.0 for better results
3 years ago
7ab2491ab7
Better handling of crossvit for tests / forward_features, fix torchscript regression in my changes
3 years ago
f1808e0970
Post crossvit merge cleanup, change model names to reflect input size, cleanup img size vs scale handling, fix tests
3 years ago
4027412757
Add resnet33ts weights, update resnext26ts baseline weights
3 years ago
9fe5798bee
fix bug for reset classifier and fix for validating the dimension
3 years ago
3718c5a5bd
fix loading pretrained model
3 years ago
bb50b69a57
fix for torch script
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
76881d207b
Add baseline resnet26t @ 256x256 weights. Add 33ts variant of halonet with at least one halo in stage 2,3,4
3 years ago
484e61648d
Adding the attn series weights, tweaking model names, comments...
3 years ago
fb94350896
Update training script and loader factory to allow use of scheduler updates, repeat augment, and bce loss
3 years ago
f262137ff2
Add RepeatAugSampler as per DeiT RASampler impl, showing promise for current (distributed) training experiments.
3 years ago
ba9c1108a1
Add a BCE loss impl that converts dense targets to sparse /w smoothing as an alternate to CE w/ smoothing. For training experiments.
3 years ago
29a37e23ee
LR scheduler update:
...
* add polynomial decay 'poly'
* cleanup cycle specific args for cosine, poly, and tanh sched, t_mul -> cycle_mul, decay -> cycle_decay, default cycle_limit to 1 in each opt
* add k-decay for cosine and poly sched as per https://arxiv.org/abs/2004.05909
* change default tanh ub/lb to push inflection to later epochs
3 years ago
492c0a4e20
Update HaloAttn comment
3 years ago
7ab9d4555c
add crossvit
3 years ago
3b9032ea48
Use Tensor.unfold().unfold() for HaloAttn, fast like as_strided but more clarity
3 years ago
78933122c9
Fix silly typo
3 years ago
2568ffc5ef
Merge branch 'master' into attn_update
3 years ago
708d87a813
Fix ViT SAM weight compat as weights at URL changed to not use repr layer. Fix #825 . Tweak optim test.
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
a8b65695f1
Add resnet26ts and resnext26ts models for non-attn baselines
3 years ago
a5a542f17d
Fix typo
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
d667351eac
Tweak accuracy topk safety. Fix #807
3 years ago
35c9740826
Fix accuracy when topk > num_classes
3 years ago
a16a753852
Add lamb/lars to optim init imports, remove stray comment
3 years ago
c207e02782
MOAR optimizer changes. Woo!
3 years ago
a426511c95
More optimizer cleanup. Change all to no longer use .data. Improve (b)float16 use with adabelief. Add XLA compatible Lars.
3 years ago
9541f4963b
One more scalar -> tensor fix for lamb optimizer
3 years ago
8f68193c91
Update lamp.py comment
3 years ago
4d284017b8
Merge pull request #813 from rwightman/opt_cleanup
...
Optimizer cleanup and additions
3 years ago
a6af48be64
add madgradw optimizer
3 years ago
55fb5eedf6
Remove experiment from lamb impl
3 years ago
8a9eca5157
A few optimizer comments, dead import, missing import
3 years ago
ac469b50da
Optimizer improvements, additions, cleanup
...
* Add MADGRAD code
* Fix Lamb (non-fused variant) to work w/ PyTorch XLA
* Tweak optimizer factory args (lr/learning_rate and opt/optimizer_name), may break compat
* Use newer fn signatures for all add,addcdiv, addcmul in optimizers
* Use upcoming PyTorch native Nadam if it's available
* Cleanup lookahead opt
* Add optimizer tests
* Remove novograd.py impl as it was messy, keep nvnovograd
* Make AdamP/SGDP work in channels_last layout
* Add rectified adablief mode (radabelief)
* Support a few more PyTorch optim, adamax, adagrad
3 years ago
Ross Wightman