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
2020-06-09 18:28:48 +01:00
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.
2020-05-22 16:16:45 -07:00
a7f570c9b7
added MultiEpochsDataLoader
2020-05-05 14:47:02 +03:00
13cf68850b
Remove poorly named metrics from torch imagenet example origins. Use top1/top5 in csv output for consistency with existing validation results files, acc elsewhere. Fixes #111
2020-04-10 14:41:08 -07:00
27b3680d49
Revamp LR noise, move logic to scheduler base. Fixup PlateauLRScheduler and add it as an option.
2020-02-22 16:23:15 -08:00
514b0938c4
Experimenting with per-epoch learning rate noise w/ step scheduler
2020-02-21 11:51:05 -08:00
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.
2020-02-18 14:00:26 -08:00
b3cb5f3275
Working on CutMix impl as per #8 , integrating with Mixup, currently experimenting...
2020-02-16 20:09:11 -08:00
b72013def8
Added commandline argument validation-batch-size-multiplier with default set to 1.
2020-02-04 01:02:27 -08:00
5b7cc16ac9
Add warning about using sync-bn with zero initialized BN layers. Fixes #54
2020-01-31 11:44:24 -08:00
d9a6a9d0af
Merge pull request #74 from rwightman/augmix-jsd
...
AugMix, JSD loss, SplitBatchNorm (Auxiliary BN), and more
2020-01-11 12:04:29 -08:00
3eb4a96eda
Update AugMix, JSD, etc comments and references
2020-01-11 12:02:05 -08:00
7547119891
Add SplitBatchNorm. AugMix, Rand/AutoAugment, Split (Aux) BatchNorm, Jensen-Shannon Divergence, RandomErasing all working together
2020-01-05 19:58:59 -08:00
40fea63ebe
Add checkpoint averaging script. Add headers, shebangs, exec perms to all scripts
2020-01-03 14:57:46 -08:00
4666cc9aed
Add --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
2020-01-02 16:22:06 -08:00
232ab7fb12
Working on an implementation of AugMix with JensenShannonDivergence loss that's compatible with my AutoAugment and RandAugment impl
2019-12-20 23:04:11 -08:00
5719b493ad
Missed update dist-bn logic for EMA model
2019-12-19 23:03:04 -08:00
a435ea1327
Change reduce_bn to distribute_bn, add ability to choose between broadcast and reduce (mean). Add crop_pct arg to allow selecting validation crop while training.
2019-12-19 22:56:54 -08:00
3bff2b21dc
Add support for keeping running bn stats the same across distributed training nodes before eval/save
2019-12-05 22:35:40 -08:00
1f39d15f15
Allow float decay epochs arg for training, works out with step lr math
2019-11-22 13:28:58 -08:00
7b83e67f77
Pass drop connect arg through to EfficientNet models
2019-11-22 13:27:43 -08:00
4748c6dff2
Fix non-prefetch variant of Mixup. Fixes #50
2019-11-02 17:42:55 -07:00
187ecbafbe
Add support for loading args from yaml file (and saving them with each experiment)
2019-09-09 17:56:45 -07:00
b750b76f67
More AutoAugment work. Ready to roll...
2019-09-01 16:55:42 -07:00
3d9c8a6489
Add support for new AMP checkpointing support w/ amp.state_dict
2019-08-29 15:19:18 -07:00
fac58f609a
Add RAdam, NovoGrad, Lookahead, and AdamW optimizers, a few ResNet tweaks and scheduler factory tweak.
...
* Add some of the trendy new optimizers. Decent results but not clearly better than the standards.
* Can create a None scheduler for constant LR
* ResNet defaults to zero_init of last BN in residual
* add resnet50d config
2019-08-28 00:14:10 -07:00
66634d2200
Add support to split random erasing blocks into randomly selected number with --recount arg. Fix random selection of aspect ratios.
2019-08-12 16:01:58 -07:00
e7c8a37334
Make min-lr and cooldown-epochs cmdline args, change dash in color_jitter arg for consistency
2019-07-26 09:35:31 -07:00
c6b32cbe73
A number of tweaks to arguments, epoch handling, config
...
* reorganize train args
* allow resolve_data_config to be used with dict args, not just arparse
* stop incrementing epoch before save, more consistent naming vs csv, etc
* update resume and start epoch handling to match above
* stop auto-incrementing epoch in scheduler
2019-06-28 13:49:20 -07:00
b20bb58284
Distributed tweaks
...
* Support PyTorch native DDP as fallback if APEX not present
* Support SyncBN for both APEX and Torch native (if torch >= 1.1)
* EMA model does not appear to need DDP wrapper, no gradients, updated from wrapped model
2019-06-20 22:30:23 -07:00
6fc886acaf
Remove all prints, change most to logging calls, tweak alignment of batch logs, improve setup.py
2019-06-20 17:29:25 -07:00
aa4354f466
Big re-org, working towards making pip/module as 'timm'
2019-06-19 17:20:51 -07:00
7dab6d1ec7
Default to img_size in model default_cfg, defer output folder creation until later in the init sequence
2019-06-10 13:34:42 -07:00
9bcd65181b
Add exponential moving average for model weights + few other additions and cleanup
...
* ModelEma class added to track an EMA set of weights for the model being trained
* EMA handling added to train, validation and clean_checkpoint scripts
* Add multi checkpoint or multi-model validation support to validate.py
* Add syncbn option (APEX) to train script for experimentation
* Cleanup interface of CheckpointSaver while adding ema functionality
2019-06-07 15:39:36 -07:00
e6c14427c0
More appropriate/correct loss name
2019-05-28 21:41:10 -07:00
127487369f
Fix bug for prefetcher
...
Set input and target to Cuda when without using prefetcher.
2019-05-26 15:37:19 +10:00
4d2056722a
Mixup and prefetcher improvements
...
* Do mixup in custom collate fn if prefetcher enabled, reduces performance impact
* Move mixup code to own file
* Add arg to disable prefetcher
* Fix no cuda transfer when prefetcher off
* Random erasing when prefetcher off wasn't changed to match new args, fixed
* Default random erasing to off (prob = 0.) for train
2019-05-18 22:17:51 -07:00
780c0a96a4
Change args for RandomErasing so only one required for pixel/color mode
2019-05-18 12:29:30 -07:00
76539d905e
Some transform/data/loader refactoring, hopefully didn't break things
...
* factor out data related constants to own file
* move data related config helpers to own file
* add a variant of RandomResizeCrop that randomizes interpolation method
* remove old Numpy version of RandomErasing
* cleanup torch version of RandomErasing and use it in either GPU loader batch mode or single image cpu Transform
2019-05-16 22:52:17 -07:00
fee607edf6
Mixup implemention in progress
...
* initial impl w/ label smoothing converging, but needs more testing
2019-05-13 19:05:40 -07:00
8fbd62a169
Exclude batchnorm and bias params from weight_decay by default
2019-04-22 17:33:22 -07:00
bc264269c9
Morph mnasnet impl into a generic mobilenet that covers Mnasnet, MobileNetV1/V2, ChamNet, FBNet, and related
...
* add an alternate RMSprop opt that applies eps like TF
* add bn params for passing through alternates and changing defaults to TF style
2019-04-21 15:54:28 -07:00
e9c7961efc
Fix pooling in mnasnet, more sensible default for AMP opt level
2019-04-17 18:06:37 -07:00
0562b91c38
Add per model crop pct, interpolation defaults, tie it all together
...
* create one resolve fn to pull together model defaults + cmd line args
* update attribution comments in some models
* test update train/validation/inference scripts
2019-04-12 22:55:24 -07:00
c328b155e9
Random erasing crash fix and args pass through
2019-04-11 22:06:43 -07:00
9c3859fb9c
Uniform pretrained model handling.
...
* All models have 'default_cfgs' dict
* load/resume/pretrained helpers factored out
* pretrained load operates on state_dict based on default_cfg
* test all models in validate
* schedule, optim factor factored out
* test time pool wrapper applied based on default_cfg
2019-04-11 21:32:16 -07:00
f1cd1a5ce3
Cleanup CheckpointSaver, add support for increasing or decreasing metric, switch to prec1 metric in train loop
2019-04-07 10:22:55 -07:00
5180f94c7e
Distributed (multi-process) train, multi-gpu single process train, and NVIDIA AMP support
2019-04-05 10:53:04 -07:00
45cde6f0c7
Improve creation of data pipeline with prefetch enabled vs disabled, fixup inception_res_v2 and dpn models
2019-03-11 22:17:42 -07:00
2295cf56c2
Add some Nvidia performance enhancements (prefetch loader, fast collate), and refactor some of training and model fact/transforms
2019-03-10 14:23:16 -07:00
9d927a389a
Add adabound, random erasing
2019-03-01 22:03:42 -08:00
1577c52976
Resnext added, changes to bring it and seresnet in line with rest of models
2019-03-01 15:44:04 -08:00
31055466fc
Fixup validate/inference script args, fix senet init for better test accuracy
2019-02-22 14:07:50 -08:00
b1a5a71151
Update schedulers
2019-02-17 12:50:15 -08:00
b5255960d9
Tweaking tanh scheduler, senet weight init (for BN), transform defaults
2019-02-13 23:11:09 -08:00
a336e5bff3
Minor updates
2019-02-08 20:56:24 -08:00
cf0c280e1b
Cleanup tranforms, add custom schedulers, tweak senet34 model
2019-02-06 20:19:11 -08:00
c57717d325
Fix tta train bug, improve logging
2019-02-02 10:17:04 -08:00
72b4d162a2
Increase training performance
2019-02-01 22:48:31 -08:00
5855b07ae0
Initial commit, puting some ol pieces together
2019-02-01 22:07:34 -08:00
Antoine Broyelle
Ross Wightman
AFLALO, Jonathan Isaac
Andrew Lavin
Ross Wightman
Zhun Zhong