Skip to content

Latest commit

 

History

History
executable file
·
70 lines (49 loc) · 2.83 KB

CHANGELOG.md

File metadata and controls

executable file
·
70 lines (49 loc) · 2.83 KB
Raw

Change Log

All notable changes to the maskrcnn with tensorflow 2 will be documented in this file.

[Unreleased] - 2021-09-25

Added

  • ResNet's backbones got resnet_leaky_relu option.
  • Mask and classifier heads also got mask_head_leaky_relu and cls_head_leaky_relu options.
  • Extended ResNets and EfficientNets backbones: added ResNet152, EfficientNet [B4, B5, B6, B7] backbones.
  • ResNeXt backbones [50, 101].
  • SE-ResNet backbones [18, 34, 50, 101, 152].
  • SE-ResNeXt backbones [50, 101].
  • SE-Net backbones [154].

Changed

  • New tensorboard log folder format: [maskrcnn_<BACKBONE_NAME>_<YYYY-MM-DD_hh-mm-ss>]
  • logs folder now is generated outside src
  • Updated README.md

Fixed

  • Normalization issue. Now normalization can be specified via model config in src/common/config.py.
  • ONNX ZeroPadding for TensorRT engines.

Known issues

  • There is a drop in performance after ONNX modification for TRT, also NaNs happens in TensorRT model output. The issue is under research.

  • TensorRT: ../rtSafe/cublas/cublasLtWrapper.cpp (279) - Assertion Error in getCublasLtHeuristic: 0 (cublasStatus == CUBLAS_STATUS_SUCCESS)

    • Possible cause: cuda and TensorRT versions mismatch;
    • Possible workaround if error still exists - remove cublasLt from tacticSources:
      • fp32: trtexec --onnx=<PATH_TO_ONNX_GRAPH> --saveEngine=<PATH_TO_TRT_ENGINE> --tacticSources=-cublasLt,+cublas --workspace=<WORKSPACE_SIZE> --verbose
      • fp16: trtexec --onnx=<PATH_TO_ONNX_GRAPH> --saveEngine=<PATH_TO_TRT_ENGINE> --tacticSources=-cublasLt,+cublas --fp16 --workspace=<WORKSPACE_SIZE> --verbose

  • Tensorlfow v2.5: AttributeError: module 'keras.utils' has no attribute 'get_file'

    • Possible cause: the influence of keras-nightly automatically suggested with tensorflow installation.

    • Possible workaround:

      Open __init__.py in <ANACONDA_PATH>/envs/tf2.5/lib/python3.8/site-packages/classification_models

      Modify file:

      import keras_applications as ka
from .__version__ import __version__
import tensorflow.keras.utils as utils
    
def get_submodules_from_kwargs(kwargs):
    backend = kwargs.get('backend', ka._KERAS_BACKEND)
    layers = kwargs.get('layers', ka._KERAS_LAYERS)
    models = kwargs.get('models', ka._KERAS_MODELS)
    return backend, layers, models, utils

  • senet154, efficientnetb5, efficientnetb6, efficientnetb7 backbones are not tested enough for now because of the high GPU memory consumption. Thus, the model with senet154 backbone may be not supported for ONNX graph modification.

[Unreleased] - 2021-12-17

Fixed

  • steps_per_epoch calculation in src/preprocess/preprocess.py. Related issue.