This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

TDA4VM: Can user specify the quantization configuration for each layer instead of automatically quantization?

Other Parts Discussed in Thread: TDA4VM

TIDL automatically quantize the network. If I want to specify the quantizaiton method for each layer, is there any example?

  • Please specify what processor you are using, and which software SDK, resp. version.

  • I use TDA4VM and want to do inference using C71. Is there any way to import my own quantization configuration, e.g., import the quantization params using prototxt?

  • I find caffe-jacinto has quantization params. If I trained model in caffe-jacinto, and the trained prototxt has QuantizationParameter for each layer.  Will the TIDL import the QuantizationParameter into the .bin file automatically?

    message QuantizationParameter {
    optional QuantizationParameter_Precision precision = 1 [default = QuantizationParameter_Precision_DYNAMIC_FIXED_POINT];
    optional QuantizationParameter_Rounding rounding_scheme = 2 [default = QuantizationParameter_Rounding_NEAREST];
    optional bool power2_scale_weights = 3 [default = false];
    optional bool power2_scale_activations = 4 [default = false];

    // Dynamic fixed point params
    message QParams {
    optional bool quantize = 1 [default = false];
    optional uint32 bitwidth = 2 [default = 8];
    optional int32 fracbits = 3 [default = 0];
    optional bool unsigned_data = 4 [ default = false];
    optional bool unsigned_quant = 5 [ default = false];
    optional float scale_target = 6 [default = 1.0];
    optional float scale_applied = 7 [default = 1.0];
    optional int32 shiftbits = 8 [default = 0];
    optional float offset = 9 [default = 0];
    optional float min = 10 [default = 0];
    optional float max = 11 [default = 0];
    }
    repeated QParams qparam_in = 5;
    repeated QParams qparam_w = 6;
    repeated QParams qparam_out = 7;

    optional int32 quantized_infer_count = 8 [ default = 0];
    }

  • tidl_tensor_range_update.c
    Fullscreen
    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    17
    18
    19
    /*
    *
    * Copyright (c) {2015 - 2017} Texas Instruments Incorporated
    *
    * All rights reserved not granted herein.
    *
    * Limited License.
    *
    * Texas Instruments Incorporated grants a world-wide, royalty-free, non-exclusive
    * license under copyrights and patents it now or hereafter owns or controls to make,
    * have made, use, import, offer to sell and sell ("Utilize") this software subject to the
    * terms herein. With respect to the foregoing patent license, such license is granted
    * solely to the extent that any such patent is necessary to Utilize the software alone.
    * The patent license shall not apply to any combinations which include this software,
    * other than combinations with devices manufactured by or for TI ("TI Devices").
    * No hardware patent is licensed hereunder.
    *
    * Redistributions must preserve existing copyright notices and reproduce this license
    * (including the above copyright notice and the disclaimer and (if applicable) source
    XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX

    If you are referring to the feature vector range. After importing the model, you can use the attached code to update the feature vector range.

    For Parameter quantization, you can update Quantization logic in the import tool are use QAT. 

    Refer below documentation for more details

    http://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/latest/exports/docs/tidl_j7_01_01_00_10/ti_dl/docs/user_guide_html/md_tidl_fsg_quantization.html