v1.34.0

• TensorRT 8.4.3 -> 8.4.0 backport. Fixes degrade the accuracy of FP16.
• TensorFlow v2.10.0-rc1 -> v2.10.0

What's Changed

• Cuda11.6.2 trt8.4.0 tf2.10 by @PINTO0309 in #126

Full Changelog: v1.33.1...v1.34.0

v1.33.1

• Model conversion time decreased from ~70s to ~25s for my particular model (MobileNetV3, 224x224x3), but time should relatively decrease more as model size increases.
• Added switch --dont_save_keras_model to avoid writing model to disk #120 @travisjayday
• Add --dont_convert_from_keras_model option

v1.33.0

• CUDA 11.7
• cuDNN 8.5
• TensorFlow v2.10.0-rc1
• PyTorch v1.12.1
• TorchVision 0.13.1
• TorchAudio 0.12.1
• OpenVINO 2022.1.0
• TensorRT 8.4.3+
• pycuda 2022.1
• onnxruntime-gpu 1.12.1 (CUDA, TensorRT, OpenVINO GPU/CPU)

v1.32.1

1. NCDHW input supported
2. Addition of --input_as_ncdhw (5D only)
3. AveragePooling1D, AveragePooling3D supported

v1.32.0

• Add tensor unit quantization option to tflite quantization options (experimental implementation) #114 @zye1996
  • --disable_per_channel

v1.31.6

• GroupConvolution bug fixes

v1.31.5

• Support for some SeparableConv2D situations. 4D only. #111 @taisuke-tomida

v1.31.4

• Interpolate bug fixes
• Support for U64, U32, and U16

v1.31.3

• Sqrt bug fixes

v1.31.2

• Replaces OP by specifying parameters directly in TensorFlow's Strided_Slice specification.
• begin, end, strides, begin_mask, end_mask, ellipsis_mask, new_axis_mask, shrink_axis_mask https://www.tensorflow.org/api_docs/python/tf/strided_slice

e.g.

{
    "layer_id": "927",
    "type": "StridedSlice",
    "replace_mode": "replace",
    "values": [
        [0,0,0,8],
        [2,7,11,16],
        [1,1,1,1],
        0,
        0,
        0,
        0,
        0
    ]
}