/-archived-VaporPlusBenchmark

Primary LanguagePython

VaporPlusBenchmark

Packages:

  1. ONNX pip install keras2onnx pip install tf2onnx

  2. TRT

Problems:

  1. Tensorflow 1.1 ONNX 1.1.1 VaporPlusGraphs EFFDET0 conversion errs: UnicodeDecodeError: 'utf-8' codec can't decode byte 0xff in position 0: invalid start byte 1.1.2 AUTOML EFFDET EFFDET conversion errs: UnicodeDecodeError: 'utf-8' codec can't decode byte 0xff in position 0: invalid start byte Requires tf24 environment to run keras model 1.1.3 PYTORCH_EFFDET Could not convert to ONNX: RuntimeError: Sizes of tensors must match except in dimension 2. Got 135 and 136 (The offending index is 0) 1.1.4 PYTORCH_HUB: RuntimeError: Sizes of tensors must match except in dimension 2. Got 135 and 136 (The offending index is 0)

1.2 TRT 1.2.1 Keras applications Could not convert any default keras applycations models into TRT with tensorflow-gpu==2.1.0 1.2.2 AutoML Could not convert any effdet to TRT ValueError: NodeDef mentions attr 'explicit_paddings' not in Op<name=MaxPool; signature=input:T -> output:T; attr=T:type,default=DT_FLOAT,allowed=[DT_HALF, DT_BFLOAT16, DT_FLOAT, DT_DOUBLE, DT_INT32, DT_INT64, DT_UINT8, DT_INT16, DT_INT8, DT_UINT16, DT_QINT8]; attr=ksize:list(int),min=4; attr=strides:list(int),min=4; attr=padding:string,allowed=["SAME", "VALID"]; attr=data_format:string,default="NHWC",allowed=["NHWC", "NCHW", "NCHW_VECT_C"]>; NodeDef: {{node resample_p6/max_pooling2d/MaxPool}}. (Check whether your GraphDef-interpreting binary is up to date with your GraphDef-generating binary.). The conversion was made with native automl python conversion script into UBUNTU 18.04 When infering: google/automl#440 1.2.3 PYTORCH_EFFDET Could not convert to TRT: RuntimeError: Sizes of tensors must match except in dimension 2. Got 135 and 136 (The offending index is 0) 1.2.4 PYTORCH_HUB - YOLOS v3/v5 RuntimeError: Sizes of tensors must match except in dimension 2. Got 135 and 136 (The offending index is 0) 1.2.5 VaporPlusGraphs Conversion 'freezes', maybe needs more time (?). Needs investigation.

ENVIRONMENTS USED:

  1. tf23py36: conda create -n tf23py36 python=3.6 anaconda opencv tensorflow-gpu=2.1.0 pip install tensorflow-gpu==2.3.0 conda install pytorch torchvision cudatoolkit=10.1 -c pytorch

cd third_party/pytorch/torch2trt python setup.py install

BENCHMARKS: PYTORCH_TRT

  1. tf23: ---will have python=3.7.9 conda create -n tf23 anaconda tensorflow-gpu=2.1.0 pip install tensorflow-gpu==2.3.0 conda install pytorch torchvision cudatoolkit=10.1 -c pytorch conda install -c conda-forge onnx pip install onnxruntime-gpu conda install -c conda-forge timm import onnxruntime

BENCHMARKS: PYTORCH PYTORCH_ONNX PYTORCH_EFFDET PYTORCH_HUB PYTORCH_HUB_ONNX AUTOML_EFFDET_PB AUTOML_EFFDET_TRT KERAS_ONNX KERAS_PB KERAS_TRT KERAS VAPOR_GRAPHS_PB VAPOR_GRAPHS VAPOR_GRAPHS_ONNX VAPOR_GRAPHS_TRT

  1. tf24 conda create -n tf24 anaconda conda install pytorch torchvision cudatoolkit=10.1 -c pytorch pip install tensorflow-gpu==2.4.1 cd third_party/lumm_tensorflow pip install -r automl_effdet_requirements.txt

BENCHMARKS: AUTOML_EFFDET

Streaming restore not supported from name-based checkpoints when graph building. File a feature request if this limitation bothers you. As a workaround, consider either using tf.train.Checkpoint to load name-based checkpoints or enabling eager execution.

Investigate !!!!! Why AUTOML_EFFDET_TRT, KERAS_TRT works on tf23 but PYTORCH_TRT need tf23py36