Deepdetect Versions Save

Features

• ml: data augmentation for object detection with torch backend (95942b9)
• ml: Visformer architecture with torch backend (40ec03f)
• torch: add batch size > 1 for detection models (91bde66)
• torch: image data augmentation with random geometric perspectives (d163fd8)
• api: introduce predict output parameter (c9ee71a)
• api: use DTO for NCNN init parameters (2ee11f0)

Bug Fixes

• build: docker builds with tcmalloc (6b8411a)
• doc: api traced models list (342b909)
• graph: loading weights from previous model does not fail (5e7c8f6)
• torch: fix faster rcnn model export for training (cbbbd99)
• torch: retinanet now trains correctly (351d6c6)

Docker images:

• CPU version: docker pull jolibrain/deepdetect_cpu:v0.17.0
• GPU (CUDA only): docker pull jolibrain/deepdetect_gpu:v0.17.0
• GPU (CUDA and Tensorrt) :docker pull jolibrain/deepdetect_cpu_tensorrt:v0.17.0
• GPU with torch backend: docker pull jolibrain/deepdetect_gpu_torch:v0.17.0
• All images available on https://hub.docker.com/u/jolibrain

Features

• torch: add confidence threshold for classification (0e75d88)
• torch: add more backbones to traced detection models (f4d05e1)
• torch: allow FP16 inference on GPU (705d3d7)
• torch: madgrad optimizer (0657d82)
• torch: training of detection models on backend torch (b920999)

Bug Fixes

• torch: default gradient clipping to true when using madgrad (5979019)
• remove dirty git flag on builds (6daa4f5)
• services names were not always case insentitive (bee3183)
• chains: cloning of image crops in chains (2e62b7e)
• ml: refinedet image dimensions configuration via API (20d56e4)
• TensorRT: fix some memory allocation weirdness in trt backend (4f952c3)
• timeseries: throw if no data found (a95e7f9)
• torch: allow partial or mismatching weights loading only if finetuning (23666ea)
• torch: Fix underflow in CSVTS::serialize_bounds (c8b11b6)
• torch: fix very long ETA with iter_size != 1 (0c716a6)
• torch: parameters are added only once to solver during traced model training (86cbcf5)

Docker images:

• CPU version: docker pull jolibrain/deepdetect_cpu:v0.16.0
• GPU (CUDA only): docker pull jolibrain/deepdetect_gpu:v0.16.0
• GPU (CUDA and Tensorrt) :docker pull jolibrain/deepdetect_cpu_tensorrt:v0.16.0
• GPU with torch backend: docker pull jolibrain/deepdetect_gpu_torch:v0.16.0
• All images available on https://hub.docker.com/u/jolibrain

Features

• nbeats: default backcast loss coeff to zero, allows very short forecast length to learn smoothly (db17a41)
• timeseries: add MAE and MSE metrics (847830d)
• timeseries: do not output per serie metrics as a default, add prefix _all for displaying all metrics (5b6bc4e)
• torch: model publishing with the platform (da14d33)
• torch: save last model at training service interruption (b346923)
• torch: SWA for RANGER/torch (https://arxiv.org/abs/1803.05407) (74cf54c)
• torch/csvts: create db incrementally (4336e89)

Bug Fixes

• caffe/detection: fix rare spurious detection decoding, see bug 1190 (94935b5)
• chore: add opencv imgcodecs explicit link (8ff5851)
• compile flags typo (8f0c947)
• docker cpu link in readme (1541dcc)
• tensorrt tests on Jetson nano (25b12f5)
• nbeats: make seasonality block work (d035c79)
• torch: display msg if resume fails, also fails if not best_model.txt file (d8c5418)

Docker images:

• CPU version: docker pull jolibrain/deepdetect_cpu:v0.15.0
• GPU (CUDA only): docker pull jolibrain/deepdetect_gpu:v0.15.0
• GPU (CUDA and Tensorrt) :docker pull jolibrain/deepdetect_cpu_tensorrt:v0.15.0
• GPU with torch backend: docker pull jolibrain/deepdetect_gpu_torch:v0.15.0
• All images available on https://hub.docker.com/u/jolibrain

Features

• bench: Add parameters for torch image backend (5d24f3d)
• ml: ViT support for Realformer from https://arxiv.org/abs/2012.11747v2 (5312de7)
• nbeats: add parameter coefficient to backcast loss (35b3c31)
• torch: add inference for torch detection models (516eeb6)
• torch: Sharpness Aware Minimization (2010.01412) (45a8408)
• torch: support for multiple test sets (c0dcec9)
• torch: temporal transformers (encoder only) (non autoreg) (3538eb7)
• CSV parser support for quotes and string labels (efa4c79)
• new cropping action parameters in chains (6597b53)
• running custom methods from jit models (73d1eef)
• torch/txt: display msg if vocab not found (31837ec)
• SSD MAP-x threshold control (acd252a)
• use oatpp::DTO to parse img-input-connector APIData (33aee72)

Bug Fixes

• build: pytorch with custom spdlog (1fb19a0)
• caffe/cudnn: force default engine option in case of cudnn not compiled in (b6dec4e)
• chore: typo when trying to use syslog (374e6c4)
• client: Change python package name to dd_client (b96b0fa)
• csvts: read from memory (6d1dba8)
• csvts: throw proper error when a csv file is passed at training time (90aab20)
• docker: ensure pip3 is working on all images (a374a58)
• ncnn: update innerproduct so that it does not pack data (9d88187)
• torch: add error message when repository contains multiple models (a08285f)
• -Werror=deprecated-copy gcc 9.3 (0371cfa)
• action cv macros with opencv >= 3 (37d2926)
• caffe build spdlog dependency (62e781a)
• docker /opt/models permissions (82e2695)
• prevent softmax after layer extraction (cbee659)
• tag syntax for github releases (4de3807)
• torch backend CPU build and tests (44343f6)
• typo in oatpp chain HTTP endpoint (955b178)
• torch: gather torchscript model parameters correctly (99e4dbe)
• torch: set seed of torchdataset during training (d02404a)
• torch/ranger: allow not to use lookahead (d428d08)
• torch/timeseries: in case of db, correctly finalize db (aabedbd)

Docker images:

• CPU version: docker pull jolibrain/deepdetect_cpu:v0.14.0
• GPU (CUDA only): docker pull jolibrain/deepdetect_gpu:v0.14.0
• GPU (CUDA and Tensorrt) :docker pull jolibrain/deepdetect_cpu_tensorrt:v0.14.0
• GPU with torch backend: docker pull jolibrain/deepdetect_gpu_torch:v0.14.0
• All images available on https://hub.docker.com/u/jolibrain

Features

• support for batches for NCNN image models (b85d79e)
• ml: retain_graph control via API for torch autograd (d109558)
• ml: torch image basic data augmentation (b9f8525)
• ncnn: use master from tencent/ncnn (044e181)
• upgrade oatpp to pre-1.2.5 (596f6f4)

Bug Fixes

• torch: csvts forecast mode needs sequence of length backcast during predict (4c89a1c)
• add missing spdlog patch (4d0a4fa)
• caffe linkage with our spdlog (967fdef)
• copy .git in docker image builder (570323d)
• deactivate the csvts NCNN test when caffe is not built (5a5c8f1)
• missing support for parent_id in chains with Python client (a5fad50)
• NCNN chain with images and actions (38b1d07)
• throw if hard image read error in caffe classification input (f1c0d09)
• doc: similarity search_nn number of results API (5eaf343)
• torch: remove potential segfault in csvts connector (ba96b4e)

Docker images:

• CPU version: docker pull jolibrain/deepdetect_cpu:v0.13.0
• GPU (CUDA only): docker pull jolibrain/deepdetect_gpu:v0.13.0
• GPU (CUDA and Tensorrt) :docker pull jolibrain/deepdetect_cpu_tensorrt:v0.13.0
• GPU with torch backend: docker pull jolibrain/deepdetect_gpu_torch:v0.13.0
• All images available on https://hub.docker.com/u/jolibrain

Highlights

• Vision Transformer (ViT) image classification models support with libtorch
• Support for native Torch vision classification models
• Improved N-BEATS for multivariate time-series
• OATPP new webserver interface

Features

• switch back to cppnet-lib by default (ebe3b15)
• torch: native models can load weights from any jit file (69af7f4)
• torch: update libtorch to 1.7.1 (41d5375)
• add access log for oat++ (4291bf8)
• add cudnn cmake find package (5983ffd)
• add some more error messages to log (e4ec772)
• enable backtrace on segfault for oatpp (96b2184)
• enhance cppnetlib req timing logs (6fc3e76)
• gives also per target error and not only global eucl when selecting measure:eucll (dd2fc79)
• introduce aotpp interface for deepdetect (04b79f4)
• print stacktrace on segfault (11ab359)
• provide predict/transform duration in ms (0197991)
• service stats provide predict and transform duration total (9a24125)
• track oatpp request timing (68749d3)
• ml: image regression model training with libtorch (968c551)
• tools: trace_torchvision can trace models for loading weights with dd native models (c11b551)
• torch: Add multigpu support for native models training (33cd1df)
• torch: Add native resnet support (0a01e57)
• torch: add wide resnet and resnext to the vision models (aba6efb)
• use jolibrain fork of faiss (8eb6e53)
• use oatpp by default (c1d6620)
• vision transformer (ViT) multi-gpu with torch (88b65c2)
• graph: correct data size computation if different ouputs of an op have different sizes (288dd5b)
• ml: added vision tranformer (ViT) as torch native template (72c0269)
• ml: torch db stores encoded images instead of tensors (e7f3c19)
• ml: torch regression and classification models training from list of files without db (e049caa)
• torch: clip gradient options for all optimizers (c2ddee5)
• torch: implement resume mllib option for torchlib: if true, reuse previous solver state (02e3177)
• torch/nbeats: allow different sizes for backcast and forecast , also implements minimal change in csvtstorchinputconn in order to do forecast of signals instead of label predicting (d4e27f3)
• torch/timeseries: add (p)ReLU layers in recurrent template, allowing to compute mlp-like embeddings before LSTM layers (930bee2)
• torch/timeseries: log much more info on data (17d9a49)
• allow to disable warning on build (75b2928)
• one protobuf to rule them all (37a0867)

Bug Fixes

• allows mean+std image transform with NCNN inference (c038f47)
• benchmark tool to pass input size on every predict call (997023a)
• bounds on two ml tests with non deterministic outputs (eeb783a)
• broken API doc formatting (8b0ab32)
• caffe backend internal exception if bbox rescaling fails (47d589f)
• copy oatpp static files in docker images (d6568d8)
• copy service name between input and output rapidjson document (46456dd)
• ddimg logger ptr (4e9e871)
• do not display all euclidean metrics for autoencoders (8e09e48)
• ensure redownloaded test archives are extracted (d09eb2a)
• fix compilation w/o caffe (0e693c5)
• forward our cuda to xgboost (021b5a8)
• init XXX_total_duration_ms to 0ms (292d891)
• missing libboost-stacktrace-dev dep in docs (37e6008)
• models flops and number of parameters in API (b856534)
• NCNN backend using the common protobuf (a8dc531)
• NCNN bbox loop (8d029c9)
• NCNN best parameter for classification models (a7ac187)
• ONNX tensorrt engine with correct enqueueV2 (1aede85)
• pass CUDA_ARCH correctly to caffe (3b9f5a1)
• raise error is jsonapi is invalid (0e0a892)
• refinedet vovnet deploy parameter setup (d7ff1e6)
• reraise same signal on abort (391568d)
• scale image input with NCNN (49cddfe)
• setuptools drop support of python27 (71cd789)
• simsearch build with annoy (39a9cda)
• some HTTP return codes (703553f)
• start-dede.sh permissions (66b49c4)
• tensorrt input size for caffe source models (5488c99)
• tensorrt max workspace size overflow (0358c4a)
• tentative torch faster tests (cbdefa7)
• torch image input connector mean and std scaling (13374dd)
• wrong NCNN bbox output scaling (113c7d2)
• wrong template_params with torch from DD API (bee39c4)
• build: allow cmake 3.10 (Ubuntu 18.04) for builds with -DUSE_CPU_ONLY=ON (d6fff8d)
• build: fix tensorrt-oss include (bdf4ec2)
• caffe: check model destination permissions (913489e)
• csv: remove trailing \n from header if there are any, robustify in case of data passed from mem (400b53a)
• csvts: clarify error msg on seq lenths (ad99049)
• csvts: fix typo in logged msg (a81b291)
• nbeats: fix for multi gpu (627c8db)
• nbeats: fix loss for backcast part (986728c)
• oatpp: use parent _logger object (88fc246)
• timeseries: throw exception if recurrent template w/o layers specification (efc49de)
• timeseries: throw if smthg wrong with scaling values (bf316b9)
• torch: also remove not best models in case of native model (45819df)
• torch: correctly implements range state comparison (c58a0df)
• torch: Display correct remaining time and iteration duration (ec7c031)
• torch: do not add linear head if native model (f81b997)
• torch: loosen ut test (30659c8)
• torch: loosen ut test 2 (ad6da05)
• torch: when resume, override solver parameters after solver reload (d08695d)
• torch/timeseries: restore previous behavior wrt timesteps API param : timestep not used at predict time (a343d4b)
• ViT blocks modification for importing pre-trained models (f7f66a4)
• torch/timeseries: check if number of labels is not larger than number of columns (294a5ed)
• ViT: typo that changed computations a lot (95acf05)
• tensorrt refinedet CI and linking to nvcaffeparsers with tensorrt OSS (846d6f5)
• torch/native: do not fail if trying to load weights before (late) allocation of native model (e0e1d9b)

Docker images:

• CPU version: docker pull jolibrain/deepdetect_cpu:v0.12.0
• GPU (CUDA only): docker pull jolibrain/deepdetect_gpu:v0.12.0
• GPU (CUDA and Tensorrt) :docker pull jolibrain/deepdetect_cpu_tensorrt:v0.12.0
• GPU with torch backend: docker pull jolibrain/deepdetect_gpu_torch:v0.12.0
• All images available on https://hub.docker.com/u/jolibrain

Features

• bench: support for regression model benchmarking (a385292)
• make python client an install package (ec2f5e2)
• one protobuf to rule them all (77912fe)
• api: add versions and compile flags to /info (67b1d99), closes #897
• caffe: add new optimizers flavors to API (d534a16)
• ml: tensorrt support for regression models (77a016b)
• tensorrt: Add support for onnx image classification models (a8b81f2)
• torch: ranger optimizer (ie rectified ADAM + lookahead) + \ (a3004f0)

Bug Fixes

• torch: best model was never saved on last iteration (6d1aa4d)
• torch: clip gradient in rectified adam as stated in annex B of original paper (1561269)
• torch: Raise an exception if gpu is not available (1f0887a)
• add pytorch fatbin patch (43a698c)
• add tool to generate debian buster image with the workaround (5570db4)
• building documentation up to date for 18.04, tensorrt and tests (18ba916)
• docker adds missing pytorch deps (314160c)
• docker build readme link from doc (c6682bf)
• handle int64 in conversion from json to APIData (863e697)
• ignore JSON conversion throw in partial chains output (742c1c7)
• missing main in bench.py (8b8b196)
• proper cleanup of tensorrt models and services (d6749d0)
• put useful informations in case of unexpected exception (5ab90c7)
• readme table of backends, models and data formats (f606aa8)
• regression benchmark tool parameter (3840218)
• tensorrt output layer lookup now throws when layer does not exist (ba7c839)
• csvts/torch: allow to read csv timeserie directly from query (76023db)
• doc: update to neural network templates and output connector (2916daf)
• docker: don't share apt cache between arch build (75dc9e9)
• graph: correctly discard dropout (16409a6)
• stats: measure of inference count (b517910)
• timeseries: do not segfault if no valid files in train/test dir (1977bba)
• torch: add missing header needed in case of building w/o caffe backend (2563b74)
• torch: load weights only once (0052a03)
• torch: reload solver params on API device (30fa16f)
• tensorrt fp16 and int8 selector (36c7488)
• torch/native: prevent loading weights before instanciating native model (b15d767)
• torch/timeseries: do not double read query data (d54f60d)

Docker images:

• CPU version: docker pull jolibrain/deepdetect_cpu:v0.11.0
• GPU (CUDA only): docker pull jolibrain/deepdetect_gpu:v0.11.0
• GPU (CUDA and Tensorrt) :docker pull jolibrain/deepdetect_cpu_tensorrt:v0.11.0
• GPU with torch backend: docker pull jolibrain/deepdetect_gpu_torch:v0.11.0
• All images available on https://hub.docker.com/u/jolibrain

Features

• timeseries: MAPE, sMAPE, MASE, OWA metrics (c1f4ef9)
• automatically push image build for master (19e9674)
• build: add script to create cppnet-lib debian package (28247b4)
• build: allow to change CUDA_ARCH (67ad43e)
• dede: Training for image classification with torch (6e81915)
• docker: publish image as soon as ready (957e07c)
• docker: publish image as soon as ready (5f7013d)
• docker: rework Dockerfile (8bc9ddf)
• docker: use prebuild cppnet-lib (c929773)
• graph: lstm autoencoder (038a74c)
• nbeats: expose hidden size param in API (d7e5515)
• add auto release tools (98b41b0)
• imginputfile: histogram equalization of input image (2f0061c), closes #778
• imginputfile: histogram equalization of input image (576f2d8), closes #778
• stats: added service statistics mechanism (1839e4a)
• torch: in case of timeseries, warn if file do not contain enough timesteps (1a5f905)
• torch: nbeats (f288665)
• torch: upgrade to torch 1.6 (f8f7dbb)
• torch,native: extract_layer (d37e182)
• add json output to dd_bench.py (874fc01)
• added bw image input support to dd_bench (6e558d6)
• trains-status: add tflops to body.measures (af31c8b), closes #785
• Docker images optimization (fba637a)
• format the code with clang-format (07d6bdc)
• LSTM over torch , preliminary internal graph representation (25faa8b)
• update all docker images to ubuntu 18.04 (eaf0421)

Bug Fixes

• fix split_data in csvts connector (8f554b5)
• build: CUDA_ARCH not escaped correctly (696087f)
• build: ensure all xgboost submodules are checkouted (12aaa1a)
• clang-format: signed/unsigned comparaison (af8e144)
• clang-format: signed/unsigned comparaison (0ccabb6)
• clang-format: typo in dataset tarball command (04ddad7)
• csvts: correctly store and print test file names (12d4639)
• dede: Remove unnecessary caffe include that prevent build with torch only (a471b82)
• dede: support all version of spdlog while building with syslog (81f47c9)
• docker: add missing .so at runtime (4cc24ce)
• docker: add missing gpu_tensorrt.Dockerfile (97ff2b3)
• docker: add some missing runtime deps (0883a33)
• docker: add some missing runtime deps (a91f35f)
• docker: fixup base runtime image (6238dd4)
• docker: install rapidjson-dev package (30fb2ca)
• native: do not raise exception if no template_param is given (d0705ab)
• nbeats: correctly setup trend and seasonality models (implement paper version and not code version) (75accc6)
• nbeats: much lower memory use in case of large dim signals (639e222)
• tests: inc iteration of torchapi.service_train_image test (4c93ace)
• torch: Fix conditions to add classification head. (f46a710)
• torch/timeseries: unscale prediction output if needed (aa30e88)
• /api/ alias when deployed on deepdetect.com (4736893)
• add support and automated processing of categorical variables in timeseries data (1a9af3e)
• allow serialization/deserializationt of Inf/-Inf/NaN (976c892)
• allows to specify size and color/bw with segmentation models (58ecb4a)
• build with -DUSE_TENSORRT_OSS=ON (39bd675)
• convolution layer initialization of SE-ResNeXt network templates (69ff0fb)
• in tensorrt builds, remove forced cuda version and unused lib output + force-select tensorrt when tensorrt_oss is selected (9430fb4)
• input image transforms in API doc (f513f17)
• install cmake version 3.10 (10666b8)
• missing variant package in docker files (dcf738b)
• race condition in xgboost|dede build (fd32eae)
• remove unecessary limit setting call to protobuf codedstream (ae26f59)
• replace db":true by db":false in json files when copying models (06ac6df)
• set caffe smooth l1 loss threshold to 1 (0e329f0)
• ssd_300_res_128 deploy file is missing a quote (4e52a0e)
• svm prediction with alll db combinations (d8e2961)
• svm with db training (6e925f2)
• tensorrt does not support blank_label (7916500)
• typo in docker image builder (cb5ae19)
• unusual builds (ie w/o torch or with tsne only lead to build errors (241bf6b)
• update caffe cudnn engine without template (ca58c51)
• torch: handle case where sequence data is < wanted timestep (b6d394a)
• TRT: refinedet (b6152b6)

Docker images:

• CPU version: docker pull jolibrain/deepdetect_cpu:v0.10.1
• GPU (CUDA only): docker pull jolibrain/deepdetect_gpu:v0.10.1
• GPU (CUDA and Tensorrt) :docker pull jolibrain/deepdetect_cpu_tensorrt:v0.10.1
• All images available on https://hub.docker.com/u/jolibrain

Features & updates

• BERT + GPT2 training / finetuning / inference with Torch C++ API #673
• TensorRT 5, 6 & 7 OSS #683
• lr dropout #668 as an implementation of https://arxiv.org/abs/1912.00144 (also see https://github.com/jolibrain/caffe/pull/58)
• confidence on OCR/CTC output prediction #694
• support for logger inside chain actions #693
• added support for OCR models (CRNN) Squeeze-Excitation ResNet-50 and ResNeXt-50 #685

API changes

• mllib.lr_dropout#668
• image rotation action in chains #677

Bug fixes

• inverted ymin and ymax in detection model output, #690
• chain action crop failing when bbox is not within the image #693
• fixed bug in raw measure data output for model calibration #679
• fixed error message on wrong TensorRT model filename #681
• fixed potentially empty chain output #692

This release updates to C++ torch 1.4, improves speed and accuracy when training object detection models, and fixes various issues.

Features & Updates

• Support for Torch 1.4 with BERT and image classification models update #698
• New CUDNN convolution backend for Caffe saves a lot of memory for ResNext architectures and grouped convolutions, see https://github.com/jolibrain/caffe/pull/65
• Geometry transforms for object detection training, #702 and https://github.com/jolibrain/caffe/pull/67
• Much faster training of object detectors, see https://github.com/jolibrain/caffe/pull/70
• Added unit tests for TensorRT backend #697
• Segmentation benchmark script #711

Models

• added RefineDet VoVNet39 512x512 architecture for SotA object detection #700

API changes

• fine-grained CUDNN engine selection #696

Bug fixes

• Fixed Dlib logger in chains #699
• Fixes to torch API unit tests #704
• Fixed TensorRT classification best API keyword behavior #720
• Fixed logging from within output connectors #705
• Fixed rare case in metrics #708
• Fixed CUDA arch for FAISS builds #710
• Fix to rotate actions in chains #713
• Fix of Torch backend build along with Caffe #714
• Fixed error handling on model parser failure #712