3.2.0

This GitHub Release was done automatically by CircleCI

What's Changed

• Update yolox_loss.py by @eran-deci in #1265
• Fix documentation link! by @Louis-Dupont in #1300
• Addiing more logs to let user know when pretrained_weights is being used, and/or downloaded by @Louis-Dupont in #1298
• set HYDRA_FULL_ERROR by @Louis-Dupont in #1291
• remove arch_params from logs by @Louis-Dupont in #1297
• Make classification models inherit from BaseClassifier by @Louis-Dupont in #1314
• Make cache annotation optional by @Louis-Dupont in #1332
• Fix class_inclusion_list in DetectionDataset by @Louis-Dupont in #1327
• Support for segmentation extreme batch cases by @shaydeci in #1282
• Feature/sg 901 extreme batch visualization for object detection by @shaydeci in #1339
• Update OD docs with clarified output formats by @BloodAxe in #1348
• Fix model.get by @Louis-Dupont in #1287
• Feature/sg 1059 fix ci by @BloodAxe in #1350
• Fix datasetparams not showing by @Louis-Dupont in #1317
• Fix a bug in implementation of DetectionRGB2BGR.get_equivalent_preprocessing by @BloodAxe in #1352
• Add utilities to plot datasets to Weights & Biases + Add callback to log validation predictions to Weights & Biases by @soumik12345 in #1167
• DagsHub Logger: Fix unsupported metric formats for MLflow, Add example notebook by @nirbarazida in #915
• New Export API by @BloodAxe in #1318
• Feature/sg 000 fix predict in pose estimation by @BloodAxe in #1358
• Add model export tutorial to documentation website by @Louis-Dupont in #1362
• fix by @Louis-Dupont in #1367
• Feature/sg 000 propagate imagenet dataset params by @BloodAxe in #1368
• Doc changes by @pranoyr in #1253
• Summarize models, losses & metrics for segmentation by @BloodAxe in #1354
• Feature/sg 000 fix import of onnx graphsurgeon by @BloodAxe in #1359
• Feature/sg 1047 predict od with labels by @shaydeci in #1365
• Feature/sg 1033 fix yolox anchors by @BloodAxe in #1369
• version bumped by @shaydeci in #1374

New Contributors

• @nirbarazida made their first contribution in #915
• @pranoyr made their first contribution in #1253

Full Changelog: 3.1.3...3.2.0

3.1.3

New features
predict() support for pose estimation (PR: #1142) and classification (PR: #1220)
The possibility to log scalar values with explicit time units associated with it: log losses per step, use TimerCallback that measures and log batch time (PR: #1181)
Support torch compile (PR: #1172)
Class agnostic NMS (PR: #1232)
Allow to compute & log per-class mAP scores (PR: #1238)
Small features
Add batch_size option to predict (PR: #1273)
Allows to pass user-defined target of IoU thresholds without asking user to edit IoUThresholds enum in SG codebase (PR: #1223)
Add DetectionVerticalFlip for bounding box detection augmentation (PR: #1237)
Support the option for multiple ignored indices for segmentation metrics (PR: #1177)
Simpify torch schedulers integration (PR: #1230)
support single output in YoloX NMS (PR: #813)
Compute the best threshold for each class in an efficient manner without added loops (PR: #973)
Bug fix
Hotfix/alg 1470 drop boxes padding by (PR: #1107)
Fix evaluate from recipe by (PR: #1170)
fixed bug of whitespace by (PR: #1173)
Fix collections.Iterable -> typing.Iterable to fix crash in python 3.10 by @BloodAxe in #1178
Fix missing encoding (PR: #1185)
default quantization params set in qat from config (PR: #1192)
Fix mkdirs in checkpoint (PR: #1198)
Fix DEKR’s replace_head & improve repr for keypoints transforms (PR: #1227)
Getting rid of “module.” heritage (PR: #1184)
Fix LayerNorm have a bias parameter attribute but is not instance of torch primitive modules (PR: #1229)
moved exceptions from training (PR: #1260)
Added explicit casting of input in predict pipeline. (PR: #1281)
kd ema model unrwap model fix (PR: #1283)
New Contributors
@allankouidri made their first contribution in #1185
@djm93dev made their first contribution in #1199
@danielafrimi made their first contribution in #1220
@RanZilberstein made their first contribution in #1228
@itaylevy-deci made their first contribution in #973
@LukeAI made their first contribution in #1237
@jorgectf made their first contribution in #1240
@jacobmarks made their first contribution in #1278

3.1.2

What's Changed

• Hotfix/sg 000 fix doc typo rf100 by @Louis-Dupont in #897
• Fixed false-positive warning message by @BloodAxe in #886
• Update deprecate -> deprecated by @BloodAxe in #906
• Fix not-closed tqdm by @BloodAxe in #909
• Update of PoseEstimation model by @BloodAxe in #872
• Update the notebooks to install super_gradients==3.1 by @BloodAxe in #924
• Fix the bug with YOLONAS of not supporting overriding in_channels by @BloodAxe in #923
• Feature/sg 000 no explicit pil by @BloodAxe in #908
• Force predictions as list by @Louis-Dupont in #945
• Feature/sg 850 support torch 2.0 by @shaydeci in #944
• Bug/sg 764 wrong ckpt when resuming with external ckpt root dir by @shaydeci in #948
• Fix typo in welcome.html by @eltociear in #941
• Add detection prediction tutorial by @Louis-Dupont in #943
• Update YoloNASQuickstart.md by @avideci in #970
• Fix preprocessing by @Louis-Dupont in #972
• Fix train_loader not initalized properly by @Louis-Dupont in #981
• Fix empty mask in MaskAttentionLoss by @lkdci in #982
• changed requirement by @shaydeci in #1002
• Fix convert_to_onnx to correctly handle case when input_shape is None by @BloodAxe in #989
• Feature/sg 757 resume for spots by @shaydeci in #870
• Cityscapes AutoLabelling dataset by @lkdci in #1000
• Predict on fused model by @Louis-Dupont in #998
• Proposal of issue template by @Louis-Dupont in #1018
• Fix doc by @Louis-Dupont in #1019
• Fixed missing encoding (Issue #999) by @T0T4R4 in #1044
• Cache model from platform locally by @BloodAxe in #1009
• Feature/infra 000 nightly by @shaydeci in #1051
• Added CoreML export notebook & method for Apple devices. by @avideci in #1068
• Bug/sg 861 decouple qat from train from config by @shaydeci in #1001
• updated contributing.md by @shaydeci in #965
• Draw most likely bbox on top (#1061) by @laszlovandenhoek in #1063
• replace documentation 'onnx' with 'coreml' by @shanibenbaruch in #1074
• Fix deprecated usage of pycocotools by @fcakyon in #1052
• YoloNAS benchmarking tutorial by @shaydeci in #1076
• Fix arch_param deprecated by @Louis-Dupont in #1084
• fix deprecated naming for stdc by @ofrimasad in #1094
• fix: Plot boxes by @yurkovak in #1097
• Added link to new QAT fine-tuning notebook (YoloNAS) + removed github notebooks by @shaydeci in #1100
• Hotfix/sg 000 fix breaking changes in some imports by @BloodAxe in #1101
• Hotfix/sg 000 dekr weights hotfix by @BloodAxe in #1102
• Fixed resuming training with DDP when average_best_models=true by @shaydeci in #1111
• Update deci-lab-client usage by @roikoren755 in #1047
• Feature/sg 815 fix override dataset params by @BloodAxe in #1092
• Bug/sg 896 add deprecation for previous breaking changes by @shaydeci in #1121
• Actualize pose estimatino zoo. by @BloodAxe in #1114
• Typo fix in modify_params_for_qat by @shaydeci in #1129
• Bug/sg 897 support for min samples with ddp by @shaydeci in #1128
• strict_load fixed in recipe default checkpoint_params by @shaydeci in #1130
• bump version by @shaydeci in #1132
• Update Weights & Biases Integration + Add Prediction Visualization for Object Detection by @soumik12345 in #1010
• Bug/sg 000 fix tests by @shaydeci in #1131
• Allow checkpoint_path to be an url (Required for transfer learning) by @BloodAxe in #1138
• Fix incorrect use of collections.OrderedDict in type annotation by @BloodAxe in #1141
• Bug/sg 000 merge failure for datasetparams by @shaydeci in #1140

New Contributors

• @eltociear made their first contribution in #941
• @T0T4R4 made their first contribution in #1044
• @laszlovandenhoek made their first contribution in #1063
• @fcakyon made their first contribution in #1052
• @soumik12345 made their first contribution in #1010

Full Changelog: 3.1.1...3.1.2

3.1.1

This GitHub Release was done automatically by CircleCI

3.1.0

This GitHub Release was done automatically by CircleCI

3.0.9

Release notes 3.0.9
Features

• Object Detection (Yolox, PPyoloE) predict function on images, videos, GIF, and folders paths. The predict function works out of the box, no need to define pre and post-processing. Both pre and post-processing are taken from the training. Examples here. PR: #815, #804, #807, #829, #845

• Support yolov5 format detection dataset by for YoloX, PPyoloE #847

• Base recipes for PPYoloE and Yolox of RF100 dataset.

• DetectionOutputAdapter tutorial. The DetectionOutputAdapter is a class that converts the output of a detection model into a user-appropriate format.

• Segformer model and recipe

• Predict function on images and videos - taking the pre+post processing from the training recipe.

• QAT&PTQ notebooks

• Added student_adapter feature to KDModule, which lets the user pass a manipulated version of the input to the student model. PR: #820

• Introduce min_samples param to dataloader_params, which repeats the dataset in case its size is smaller than the value of this parameter. PR: #838

Bug fixes:

• Bug Fix wandb logger on resume by @Louis-Dupont in #766

3.0.8

Bug fixes:
Fix the compatibility of ONNX conversion with torch 1.11+
Fix the docker publication
Fix Pre-Prediciton callback override bug
Fix dataloader multiprocessing (for Mac OS and Windows)
Fix strict_load not being passed in evaluate_from_recipe
Resolve many potential circular imports
Correct albumentations import log
Fix incorrect tensor type & device when entire batch has no targets
Fix documentation format in many files

Features:
New documentation for many features (EMA, AMP, QAT, Knowledge-Distillation, Dataloaders, Optimizer, Schedulers…)
Added Lion optimizer
Added registration for SG Logger (to allow custom-defined loggers to be used from YAML)
DDRNet 39 pre-trained segmentation model
Mapillary dataset
RoboFlow100 benchmark and all 100 datasets
QATTrainer for training with QAT
QAT from recipe script
PP-YoloE pre-trained models
Pose Estimation models loss functions
Added Auxiliary heads to Unet
Channel wise distillation
ONNX Simplifier on ONNX conversion
Unet and Loss function for binary segmentation
New Transforms: DetectionPadToSize, DetectionImagePermute
Allow activation factory type to take resolved type as inp
New strict_load mode: key_matching.

3.0.7

What's Changed

• Ci fix context and fail on SG by @ranrubin in #615
• Feature/sg 000 break inner train loop by @shaydeci in #617
• Black on factories and data_interface by @Louis-Dupont in #620
• Apply black on utils by @Louis-Dupont in #619
• Added QARepVGGBlock by @spsancti in #562
• fix default value of max_train_batches, max_valid_batches by @Louis-Dupont in #623
• Hotfix/sg 000 transform backward compatibility by @Louis-Dupont in #622
• Added TestTrainingUtils suite to unit tests by @BloodAxe in #621
• Apply black by @BloodAxe in #624
• New callbacks API by @BloodAxe in #607
• Register crashtips automatically by @Louis-Dupont in #626
• Improve env sanity check to be more robust by @Louis-Dupont in #612
• Add dataset instruction by @Louis-Dupont in #602
• Interpollation crash tip by @Louis-Dupont in #616
• Group env variables by @Louis-Dupont in #618
• Remove unintended print (micro PR) by @Louis-Dupont in #630
• [BLACK PR] apply black on tests by @Louis-Dupont in #633
• Skip CI on .md files - SG side by @shaydeci in #635
• Feature/sg 624 remove avg from model plots by @Louis-Dupont in #629
• Add metadata to average_model by @Louis-Dupont in #625
• Bug/sg 000 update kd train from config by @shaydeci in #638
• Support 1.13 by @Louis-Dupont in #603
• Refactored scheduler callbacks (epoch-based/step-based warmup) by @BloodAxe in #568
• Feature/sg 541 auto batch selection by @shaydeci in #628
• Hotfix/sg 000 reduce import loops risk by @Louis-Dupont in #636
• Feature/sg 627 add dev option for DeciClient by @Louis-Dupont in #627
• Fixed id_tensor registry, so reparametrization works when .cuda() is called by @spsancti in #632
• Added new detection transforms that are used in PPYoloE by @BloodAxe in #641
• Feature/sg 493 modelnames instead of strings by @shaydeci in #614
• Fix registry exception tip by @Louis-Dupont in #644
• Feature/sg 568 single epoch sanity test for all recipes by @shaydeci in #645
• Feature/sg 634 improve deciclient errors by @Louis-Dupont in #648
• UNet registers by @lkdci in #646
• Mini change in DeciClient use of SG version by @Louis-Dupont in #651
• Feature/sg 573 pose estimation by @BloodAxe in #634
• Fix grad clipping stage + upscale gradients before clipping by @shaydeci in #653
• classification regresion tests split into 2 by @shaydeci in #656
• Added AdamW optimizer by @BloodAxe in #652
• Losses.md docs by @shaydeci in #658
• fixed num gpus missing arg for ddp compatibility by @shaydeci in #660
• documentation on using configuration files by @ofrimasad in #661
• multi gpu mode explicitly mentioned by @shaydeci in #665
• Hotfix/sg 000 fix regression tests by @shaydeci in #666
• Feature/sg 573 Integrate new EMA decay schedules by @BloodAxe in #647
• Hotfix/sg 645 regression tests essential fixes by @shaydeci in #669
• Feature/sg 646 rename logs by @Louis-Dupont in #667
• Feature/sg 644 upload big files properly on end by @Louis-Dupont in #671
• Hotfix/sg 000 temporary mute upload print by @Louis-Dupont in #672
• Bug/sg 000 fix lr function scheduling and add deprecation by @shaydeci in #675
• Feature/sg 600 metrics tutorial by @shaydeci in #673
• apply black formatter to all of "training" directory 2 by @ofrimasad in #678
• apply black formatter to all of "common" directory by @ofrimasad in #676
• apply black formatter to all of "training" directory by @ofrimasad in #677
• Segmentation Readme by @Louis-Dupont in #657
• Add log readme by @Louis-Dupont in #664
• fix by @shaydeci in #681

Full Changelog: 3.0.6...3.0.7

3.0.6

What's Changed

Bugfixes:

• Make WandB logger not ignore tag by @shaydeci in #596
• Add a version variable by @ofrimasad in #552
• Save code renamed to _save_code_lines() in wandb logger by @shaydeci in #601
• Fix master installation by @ofrimasad in #604
• Adding what happens when no rc tag exists by @strelok899 in #610
• Version and readme fix by @shaydeci in #606
• Bugfix/sg 000 release version verify test fix by @ranrubin in #613

Features:

• Feature/sg 521 gpu tests by @shaydeci in #587
• Add a decorator that saves the code to file by @ofrimasad in #589
• Feature/sg 492 fuzzy logic for get param and factories by @shaydeci in #595
• Add model names to Readme.md by @Louis-Dupont in #592
• Feature/sg 456 centralize ddp setup by @Louis-Dupont in #544
• Feature/sg 132 models convert by @shaydeci in #598
• Feature/infra 1481 call integration tests by @ranrubin in #549
• Feature/sg 193 extend output formator by @Louis-Dupont in #561
• Feature/infra 1571 support new orb by @ranrubin in #605
• Feature-OPS-1419_integration-tests-and-docker by @strelok899 in #339
• Export message after success + suffix replacement fix by @shaydeci in #611

New Contributors

• @strelok899 made their first contribution in #591

Full Changelog: 3.0.5...3.0.6

3.0.5

What's Changed

• Added warning for drop_last when train_loader is not divisible by batch_size by @shaydeci in #586
• fix compute_detection_metrics_per_cls return value when no detection … by @ofrimasad in #583
• Quantization infra mods for different calibrators and learnable amax by @spsancti in #537
• PLFM-3331 Register experiments with model name by @roikoren755 in #585
• fix ignore index for DiceCEEdgeLoss by @lkdci in #588

Full Changelog: 3.0.4...3.0.5