0b64117592
Take `no_emb_class` into account when calling `resize_pos_embed`
2 years ago
56c3a84db3
Update type hint for `register_notrace_module`
...
register_notrace_module is used to decorate types (i.e. subclasses of nn.Module).
It is not called on module instances.
2 years ago
1b278136c3
Change models with mean 0,0,0 std 1,1,1 from int to float for consistency as mentioned in #1355
2 years ago
909705e7ff
Remove some redundant requires_grad=True from nn.Parameter in third party code
2 years ago
c5e0d1c700
Add dilation support to convnext, allows output_stride=8 and 16 use. Fix #1341
2 years ago
dc376e3676
Ensure all model entrypoint fn default to `pretrained=False` (a few didn't)
2 years ago
23b102064a
Add cs3sedarknet_x weights w/ 82.65 @ 288 top1. Add 2 cs3 edgenet models (w/ 3x3-1x1 block), remove aa from cspnet blocks (not needed)
2 years ago
0dbd9352ce
Add bulk_runner script and updates to benchmark.py and validate.py for better error handling in bulk runs (used for benchmark and validation result runs). Improved batch size decay stepping on retry...
2 years ago
92b91af3bb
version 0.6.6
2 years ago
05313940e2
Add cs3darknet_x, cs3sedarknet_l, and darknetaa53 weights from TPU sessions. Move SE btwn conv1 & conv2 in DarkBlock. Improve SE/attn handling in Csp/DarkNet. Fix leaky_relu bug on older csp models.
2 years ago
51cca82aa1
👽 use hf_hub_download instead of cached_download
2 years ago
324a4e58b6
disable nvfuser for jit te/legacy modes (for PT 1.12+)
2 years ago
2898cf6e41
version 0.6.5 for pypi release
2 years ago
a45b4bce9a
x and xx small edgenext models do benefit from larger test input size
2 years ago
a8e34051c1
Unbreak gamma remap impacting beit checkpoint load, version bump to 0.6.4
2 years ago
1c5cb819f9
bump version to 0.6.3 before merge
2 years ago
a1cb25066e
Add edgnext_small_rw weights trained with swin like recipe. Better than original 'small' but not the recent 'USI' distilled weights.
2 years ago
7c7ecd2492
Add --use-train-size flag to force use of train input_size (over test input size) for validation. Default test-time pooling to use train input size (fixes issues).
2 years ago
ce65a7b29f
Update vit_relpos w/ some additional weights, some cleanup to match recent vit updates, more MLP log coord experiments.
2 years ago
58621723bd
Add CrossStage3 DarkNet (cs3) weights
2 years ago
9be0c84715
Change set -> dict w/ None keys for dataset split synonym search, so always consistent if more than 1 exists. Fix #1224
2 years ago
db0cee9910
Refactor cspnet configuration using dataclasses, update feature extraction for new cs3 variants.
2 years ago
eca09b8642
Add MobileVitV2 support. Fix #1332 . Move GroupNorm1 to common layers (used in poolformer + mobilevitv2). Keep ol custom ConvNeXt LayerNorm2d impl as LayerNormExp2d for reference.
2 years ago
06307b8b41
Remove experimental downsample in block support in ConvNeXt. Experiment further before keeping it in.
2 years ago
bfc0dccb0e
Improve image extension handling, add methods to modify / get defaults. Fix #1335 fix #1274 .
2 years ago
7d4b3807d5
Support DeiT-3 (Revenge of the ViT) checkpoints. Add non-overlapping (w/ class token) pos-embed support to vit.
2 years ago
d0c5bd5722
Rename cs2->cs3 for darknets. Fix features_only for cs3 darknets.
2 years ago
d765305821
Remove first_conv for resnetaa50 def
2 years ago
dd9b8f57c4
Add feature_info to edgenext for features_only support, hopefully fix some fx / test errors
2 years ago
377e9bfa21
Add TPU trained darknet53 weights. Add mising pretrain_cfg for some csp/darknet models.
2 years ago
c170ba3173
Add weights for resnet10t, resnet14t, and resnetaa50 models. Fix #1314
2 years ago
188c194b0f
Left some experiment stem code in convnext by mistake
2 years ago
70d6d2c484
support test_crop_size in data config resolve
2 years ago
6064d16a2d
Add initial EdgeNeXt import. Significant cleanup / reorg (like ConvNeXt). Fix #1320
...
* edgenext refactored for torchscript compat, stage base organization
* slight refactor of ConvNeXt to match some EdgeNeXt additions
* remove use of funky LayerNorm layer in ConvNeXt and just use nn.LayerNorm and LayerNorm2d (permute)
2 years ago
7a9c6811c9
Add eps arg to LayerNorm2d, add 'tf' (tensorflow) variant of trunc_normal_ that applies scale/shift after sampling (instead of needing to move a/b)
2 years ago
82c311d082
Add more experimental darknet and 'cs2' darknet variants (different cross stage setup, closer to newer YOLO backbones) for train trials.
2 years ago
a050fde5cd
Add resnet10t (basic block) and resnet14t (bottleneck) with 1,1,1,1 repeats
2 years ago
e6d7df40ec
no longer a point using kwargs for pretrain_cfg resolve, just pass explicit arg
2 years ago
07d0c4ae96
Improve repr for DropPath module
2 years ago
e27c16b8a0
Remove unecessary code for synbn guard
2 years ago
0da3c9ebbf
Remove SiLU layer in default args that breaks import on old old PyTorch
2 years ago
7d657d2ef4
Improve resolve_pretrained_cfg behaviour when no cfg exists, warn instead of crash. Improve usability ex #1311
2 years ago
879df47c0a
Support BatchNormAct2d for sync-bn use. Fix #1254
2 years ago
7cedc8d474
Follow up to #1256 , fix interpolation warning in auto_autoaugment as well
2 years ago
db64393c0d
use `Image.Resampling` namespace for PIL mapping ( #1256 )
...
* use `Image.Resampling` namespace for PIL mapping
PIL shows a deprecation warning when accessing resampling constants via the `Image` namespace. The suggested namespace is `Image.Resampling`. This commit updates `_pil_interpolation_to_str` to use the `Image.Resampling` namespace.
```
/tmp/ipykernel_11959/698124036.py:2: DeprecationWarning: NEAREST is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.NEAREST or Dither.NONE instead.
Image.NEAREST: 'nearest',
/tmp/ipykernel_11959/698124036.py:3: DeprecationWarning: BILINEAR is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.BILINEAR instead.
Image.BILINEAR: 'bilinear',
/tmp/ipykernel_11959/698124036.py:4: DeprecationWarning: BICUBIC is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.BICUBIC instead.
Image.BICUBIC: 'bicubic',
/tmp/ipykernel_11959/698124036.py:5: DeprecationWarning: BOX is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.BOX instead.
Image.BOX: 'box',
/tmp/ipykernel_11959/698124036.py:6: DeprecationWarning: HAMMING is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.HAMMING instead.
Image.HAMMING: 'hamming',
/tmp/ipykernel_11959/698124036.py:7: DeprecationWarning: LANCZOS is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.LANCZOS instead.
Image.LANCZOS: 'lanczos',
```
* use new pillow resampling enum only if it exists
2 years ago
20a1fa63f8
Make dev version 0.6.2.dev0 for pypi pre
3 years ago
347308faad
Update README.md, version to 0.6.2
3 years ago
4b30bae67b
Add updated vit_relpos weights, and impl w/ support for official swin-v2 differences for relpos. Add bias control support for MLP layers
3 years ago
d4c0588012
Remove persistent buffers from Swin-V2. Change SwinV2Cr cos attn + tau/logit_scale to match official, add ckpt convert, init_value zeros resid LN weight by default
3 years ago
27c42f0830
Fix torchscript use for offician Swin-V2, add support for non-square window/shift to WindowAttn/Block
3 years ago
2f2b22d8c7
Disable nvfuser fma / opt level overrides per #1244
3 years ago
c0211b0bf7
Swin-V2 test fixes, typo
3 years ago
9a86b900fa
Official SwinV2 models
3 years ago
d07d015173
Merge pull request #1249 from okojoalg/sequencer
...
Add Sequencer
3 years ago
d30685c283
Merge pull request #1251 from hankyul2/fix-multistep-scheduler
...
fix: multistep lr decay epoch bugs
3 years ago
a16171335b
fix: change milestones to decay-milestones
...
- change argparser option `milestone` to `decay-milestone`
3 years ago
39b725e1c9
Fix tests for rank-4 output where feature channels dim is -1 (3) and not 1
3 years ago
78a32655fa
Fix poolformer group_matcher to merge proj downsample with previous block, support coarse
3 years ago
d79f3d9d1e
Fix torchscript use for sequencer, add group_matcher, forward_head support, minor formatting
3 years ago
37b6920df3
Fix group_matcher regex for regnet.py
3 years ago
93a79a3dd9
Fix num_features in Sequencer
3 years ago
57a988df30
fix: multistep lr decay epoch bugs
...
- add milestones arguments
- change decay_epochs to milestones variable
3 years ago
578d52e752
Add Sequencer
3 years ago
f5ca4141f7
Adjust arg order for recent vit model args, add a few comments
3 years ago
41dc49a337
Vision Transformer refactoring and Rel Pos impl
3 years ago
b7cb8d0337
Add Swin-V2 Small-NS weights (83.5 @ 224). Add layer scale like 'init_values' via post-norm LN weight scaling
3 years ago
f88c606fcf
fixing channels_last on cond_conv2d; update nvfuser debug env variable
3 years ago
09e9f3defb
migrate azure blob for beit checkpoints
...
## Motivation
We are going to use a new blob account to store the checkpoints.
## Modification
Modify the azure blob storage URLs for BEiT checkpoints.
3 years ago
52ac881402
Missed first_conv in latest seresnext 'D' default_cfgs
3 years ago
7629d8264d
Add two new SE-ResNeXt101-D 32x8d weights, one anti-aliased and one not. Reshuffle default_cfgs vs model entrypoints for resnet.py so they are better aligned.
3 years ago
8f0bc0591e
fix convnext args
3 years ago
c5a8e929fb
Add initial swinv2 tiny / small weights
3 years ago
f670d98cb8
Make a few more layers symbolically traceable (remove from FX leaf modules)
...
* remove dtype kwarg from .to() calls in EvoNorm as it messed up script + trace combo
* BatchNormAct2d always uses custom forward (cut & paste from original) instead of super().forward. Fixes #1176
* BlurPool groups==channels, no need to use input.dim[1]
3 years ago
ec4e9aa5a0
Add ConvNeXt tiny and small pretrain in22k
...
Add ConvNeXt tiny and small pretrain in22k from ConvNeXt repo:
06f7b05f92
3 years ago
575924ed60
Update test crop for new RegNet-V weights to match Y
3 years ago
1618527098
Add layer scale and parallel blocks to vision_transformer
3 years ago
c42be74621
Add attrib / comments about Swin-S3 (AutoFormerV2) weights
3 years ago
474ac906a2
Add 'head norm first' convnext_tiny_hnf weights
3 years ago
dc51334cdc
Fix pruned adapt for EfficientNet models that are now using BatchNormAct layers
3 years ago
024fc4d9ab
version 0.6.1 for master
3 years ago
e1e037ba52
Fix bad tuple typing fix that was on XLA branch bust missed on master merge
3 years ago
341b464a5a
Remove redundant noise attr from Plateau scheduler (use parent)
3 years ago
fe457c1996
Update SwinTransformerV2Cr post-merge, update with grad checkpointing / grad matcher
...
* weight compat break, activate norm3 for final block of final stage (equivalent to pre-head norm, but while still in BLC shape)
* remove fold/unfold for TPU compat, add commented out roll code for TPU
* add option for end of stage norm in all stages
* allow weight_init to be selected between pytorch default inits and xavier / moco style vit variant
3 years ago
b049a5c5c6
Merge remote-tracking branch 'origin/master' into norm_norm_norm
3 years ago
7cdd164d77
Fix #1184 , scheduler noise bug during merge madness
3 years ago
9440a50c95
Merge branch 'mrT23-master'
3 years ago
d98aa47d12
Revert ml-decoder changes to model factory and train script
3 years ago
b20665d379
Merge pull request #1007 from qwertyforce/patch-1
...
update arxiv link
3 years ago
7a0994f581
Merge pull request #1150 from ChristophReich1996/master
...
Swin Transformer V2
3 years ago
61d3493f87
Fix hf-hub handling when hf-hub is config source
3 years ago
5f47518f27
Fix pit implementation to be clsoer to deit/levit re distillation head handling
3 years ago
0862e6ebae
Fix correctness of some group matching regex (no impact on result), some formatting, missed forward_head for resnet
3 years ago
94bcdebd73
Add latest weights trained on TPU-v3 VM instances
3 years ago
0557c8257d
Fix bug introduced in non layer_decay weight_decay application. Remove debug print, fix arg desc.
3 years ago
372ad5fa0d
Significant model refactor and additions:
...
* All models updated with revised foward_features / forward_head interface
* Vision transformer and MLP based models consistently output sequence from forward_features (pooling or token selection considered part of 'head')
* WIP param grouping interface to allow consistent grouping of parameters for layer-wise decay across all model types
* Add gradient checkpointing support to a significant % of models, especially popular architectures
* Formatting and interface consistency improvements across models
* layer-wise LR decay impl part of optimizer factory w/ scale support in scheduler
* Poolformer and Volo architectures added
3 years ago
1420c118df
Missed comitting outstanding changes to default_cfg keys and test exclusions for swin v2
3 years ago
c6e4b7895a
Swin V2 CR impl refactor.
...
* reformat and change some naming so closer to existing timm vision transformers
* remove typing that wasn't adding clarity (or causing torchscript issues)
* support non-square windows
* auto window size adjust from image size
* post-norm + main-branch no
3 years ago
67d140446b
Fix bug in classification head
3 years ago
29add820ac
Refactor (back to relative imports)
3 years ago
74a04e0016
Add parameter to change normalization type
3 years ago
Ceshine Lee