Skip to content

Latest commit

 

History

History
867 lines (745 loc) · 39.7 KB

AOT.md

File metadata and controls

867 lines (745 loc) · 39.7 KB
Raw

Creating standalone executable bundles

This document provides a short description about producing Ahead Of Time (AOT) compiled executable bundles. The motivation for this work is to remove the cost of compile time by allowing the users of Glow to compile a network model Ahead Of Time.

Overview

A bundle is a self-contained compiled network model that can be used to execute the model in a standalone mode. Glow has multiple backends which can run models but not all of them are capable of saving a compiled model.

The main backend used to generate bundles is the CPU backend where the bundle exists as a self-contained object file (library) containing all the necessary code to run a model. The bundle has a single entry function which performs inference for the model.

After following this document you will be able to compile models into bundles. You can view the CMake instructions (CMakeLists.txt) used to build the bundles and also the applications (main.cpp) which integrate the bundles for the ResNet50 and LeNetMnist models here:

When building Glow you can choose to build the ResNet50 and the LeNetMnist examples by enabling the GLOW_WITH_BUNDLES option as below:

mkdir build_Debug
cd build_Debug
cmake -G Ninja -DCMAKE_BUILD_TYPE=Debug -DGLOW_WITH_BUNDLES=ON ../glow
ninja all

After the build is complete, you can run the example applications using one of the following commands (the executables are found in the build_Debug/bundles folder):

./ResNet50Bundle cat_285.png
./QuantizedResNet50Bundle cat_285.png
./LeNetMnistBundle 0_1009.png
./QuantizedLeNetMnistBundle 0_1009.png

You can find the sample images used in the examples above in the following directories:

  • Sample images for the ResNet50 model from the IMAGENET database here
  • Sample image for the LeNetMnist model from the MNIST database here

Compile a bundle for a floating-point model

The model-compiler front-end tool is the main Glow tool used to compile ONNX, Caffe2 and TensorFlowLite models into bundles. The tool is generic in the sense that it can compile models with any number of inputs or outputs, without being limited to a particular application.

The command used to build a bundle using the CPU backend is the following:

model-compiler -backend=CPU -model=<model-path> -emit-bundle=<bundle-dir>
  • The option backend specifies the backend used to generated the bundle.
  • The option model specifies the path for the Caffe2 or ONNX model:
    • For Caffe2 models the <model-path> is expected to point to a directory which contains two model files: init_net.pb and predict_net.pb. Some Caffe2 model directories also contain a file named predict_net.pbtxt but that file is not required.
    • For ONNX models the <model-path> is expected to point to a file with the extension onnx or onnxtxt.
  • The option emit-bundle specifies the output directory where all the bundle artifacts will be generated. If the directory does not exist, it will be created.

There is a small difference when using this tool with ONNX/TensorFlowLite versus Caffe2 models:

  • For ONNX or TensorFlowLite models the tool can infer automatically the inputs of the model since the description of the input tensors is part of the model. Therefore the tool will be used in the form shown above: 
    model-compiler -backend=CPU -model=<onnx-model-path> -emit-bundle=<bundle-dir>
model-compiler -backend=CPU -model=<tflite-model-path> -emit-bundle=<bundle-dir>
  • For Caffe2 models the user must also explicitly provide the description of the input tensors which are not part of model. The option model-input will be used to specify the name, type and shape for each model input. The model input names can be deduced by inspecting the model with Netron. The option can be used multiple times to describe each model input separately: 
    model-compiler -backend=CPU -model=<caffe2-model-path> -emit-bundle=<bundle-dir>  \
  -model-input=<inputName1>,<inputType1>,<inputShape1>                            \
  -model-input=<inputName2>,<inputType2>,<inputShape2>                            \
  ....................................................
    For quantized types the format of the -model-input is slightly different since the scale and offset parameters should also be provided: 
    -model-input=<name>,<type>,<scale>,<offset>,<shape>
    For example we can can provide one or more inputs with: 
    -model-input=input_03_data,float,[1]
-model-input=data_bias,int32,[1,32,32]
-model-input=data,int8q,0.123,-13,[1,10]

After running the model-compiler tool, the following bundle artifacts will be generated in the output directory:

  • <network_name>.o - the bundle object file (code).
  • <network_name>.h - the bundle header file (API).
  • <network_name>.weights.bin - the model weights in binary format.
  • <network_name>.weights.txt - the model weights in text format as C text array.

For example, to compile the ResNet50 model provided by Glow in Caffe2 format, you can use the following command. The name of the input tensor is gpu_0/data, the type is float and the shape is 1 x 3 x 224 x 224 corresponding to a RGB image with NCHW layout:

model-compiler -backend=CPU -model=resnet50 -emit-bundle=build -model-input=gpu_0/data,float,[1,3,224,224]

Similarly, to compile the LeNetMnist model provided by Glow in Caffe2 format, you can use the following command. The name of the input tensor is data, the type is float and the shape is 1 x 1 x 28 x 28 corresponding to a grayscale image with NCHW layout:

model-compiler -backend=CPU -model=lenet_mnist -emit-bundle=build -model-input=data,float,[1,1,28,28]

For more information about the options of the model-compiler tool use:

model-compiler -help

Compile a bundle for a quantized model

Glow support for out-of-the-box quantized models is a work in progress mainly because the quantized tensors and operators are not well standardized in formats like Caffe2 and ONNX.

The way Glow produces a quantized bundle is by taking a floating-point model and converting it to a quantized model using it's own internal representation before compiling to a bundle.

The procedure used for quantizing the model is called profile-guided quantization (more details here). Before the model is quantized and compiled with the model-compiler tool, the quantization profile must be acquired.

It is important to note that the profiling phase is independent on the quantization parameters so there is no need to specify the quantization schema, precision or other parameters.

In order to compute the quantization profile, one option is to use the model-profiler tool. This application is generic and can be used with any model and requires a set of files (in either text or binary format) corresponding to the model input tensors in order to feed the model with a dataset and get the profile. The command has the following format:

model-profiler -model=<model-path> -dump-profile=profile.yaml \
    -input-dataset=<name1,format1,source1,opts1>              \
    -input-dataset=<name2,format2,source2,opts2>              \
    ............................................
  • The command above is for ONNX or TensorFlowLite models. For Caffe2 models the user must also provide the information about the model inputs by using the option model-input similar to the model-compiler tool.
  • The option dump-profile specifies the file name used to dump the profile in YAML format.
  • The option input-dataset specifies the dataset used to feed each of the model inputs. The option has the following comma separated fields:
    • <name> the name of the model input placeholder (tensor) where the dataset files will be loaded during run-time.
    • <format> the format of all the files from the given dataset:
      • rawbin: raw binary format. Each binary file corresponds to a tensor and contains the data serialized as a binary blob without extra meta information (tensor data type or shape) because the tensor is statically configured before loading the data. The data is expected to be serialized with the correct size and layout as the tensor in which it will be loaded. For example, for a float32 tensor with shape [2,3], the binary file is expected to have the size 2 x 3 x 4 (float32) = 24 bytes.
      • rawtxt: raw text format. Each text file corresponds to a tensor and contains data serialized as a linear list of comma separated values in text format without extra meta information (tensor data type or shape) because the tensor is statically configured before loading the data. The data is expected to be serialized with the correct size and layout as the tensor in which it will be loaded. For example, for a float32 tensor with shape [2,3], the text file is expected to contain a list of 6 values separated by comma like this (extra spaces and newlines are allowed): 
        1.0, 2.0, 3.0, 4.0, 5.0, 6.0,
    • <source> specifies the dataset source:
      • file: the dataset is specified as a text file which contains the relative or absolute paths of all the files in the dataset, listed one per line, separated by comma or not. The path of the dataset file is given as the first argument in the <opts> list. If a second argument is given in the <opts> list (optional), that will be concatenated (prepended) to all the paths from the file. The dataset file must contain only ONE PATH PER LINE. After the first comma or space character, the rest of the line is ignored. All the examples below are valid: 
        data0.bin
data1.bin,
data2.bin 'cat'
data3.bin,dog
data4.bin ,2
data5.bin,1
        Do NOT use file paths which contain spaces.
      • dir: the dataset is specified as all the files from a given directory listed alphabetically. The directory path is specified with the first argument in the <opts> list. Make sure the directory does not contain other items than the dataset files (folders, symlinks, etc).
    • <opts> extra options dependent on the <source> field.
    • This option will be used for each of the model inputs.
    • Example 1: -input-dataset=input1,rawbin,file,dataset.csv The dataset paths for the 'input1' model input are read from the 'dataset.csv' file which could have the following content: 
      /data_folder/data0.dat,
/data_folder/data1.dat,
.......................
      All the files listed are assumed to be in raw binary format.
    • Example 2: -input-dataset=input2,rawbin,file,dataset.csv,/data_folder The dataset files for the 'input2' model input are read from the 'dataset.csv' file which could have the following content: 
      data0.dat,
data1.dat,
..........
      All the file paths listed will be concatenated (prepended) with the '/data_folder' base directory path when loading. All the files listed are assumed to be in raw binary format.
    • Example 3: -input-dataset=input3,rawtxt,dir,/data_folder The dataset files for the 'input3' model input are all the files from the '/data_folder' directory listed alphabetically. The files are assumed to be in raw text format.

In order for the profiling phase to be correct, make sure the data used to feed the network is pre-processed in the same way as it would be in the case of inference. For example, for an image classification model make sure the input raw data:

  • has the correct data layout (NHWC or NCHW)
  • has the correct channel order (RGB or BGR)
  • is scaled properly: the values are in the range [0,1], [-1,1], [-127,128], [0,255.0] etc.

For more information about the options of the model-profiler tool use:

model-profiler -help

Another tool used to compute the quantization profile is the image-classifier tool which is specialized for image classification models only and requires a set of images to do the inference and compute the profile. This application has the benefit that it provides a mechanism to load directly PNG images and also to pre-process them according to the model needs (layout conversion, channel ordering, scaling):

image-classifier <images> <image-opts> -model=<model-path> -model-input-name=<name> \
    -dump-profile=profile.yaml
  • The image paths are specified one after the other, separated by space. Only images in PNG format are supported (RGB or grayscale). If many images are to be used then it is more convenient to provide a directory of images using the option input-image-dir. This option can be used multiple times for multiple directories and can also be combined with individual images listed in a positional fashion: image-classifier <images> -input-image-dir=<dir1> -input-image-dir=<dir2> ...
  • model specifies the path for the Caffe2, ONNX or TensorFlowLite model.
  • model-input-name specifies the name of the model input.
  • The option dump-profile specifies the file name used to dump the profile in YAML format.

Extra options are available to specify how the images are pre-processed before they are fed to the model during inference:

  • The option minibatch can be used to specify the batch size of the model used to perform the inference during profiling. The default value is 1. Note that the number of images provided during profiling must be divisible by minibatch.
  • The option image-mode can be used to specify how the images are normalized:
    • neg1to1 for the range [-1, 1].
    • 0to1 for the range [0, 1].
    • 0to255 for the range [0, 255] (Default).
    • neg128to127 for the range [-128,127].
  • The option image-channel-order can be used to specify the order of the channels:
    • RGB for R,G,B channel ordering.
    • BGR for B,G,R channel ordering (Default).
  • The option image-layout can be used to specify the tensor dimension ordering (layout) which must match the layout of the model input:
    • NCHW for NCHW tensor layout (Default).
    • NHWC for NHWC tensor layout.

For more information about the options of the image-classifier tool use:

image-classifier -help

After the quantization profile profile.yaml has been generated, we can use the model-compiler tool to compile the model into a bundle by loading the previously generated profile:

model-compiler ... -load-profile=profile.yaml -quantization-schema=<schema>

When compiling a quantized bundle with the model-compiler some quantization parameters can be specified:

  • quantization-schema specifies the quantization schema:
    • asymmetric for Asymmetric quantization schema (Default).
    • symmetric for Symmetric quantization schema.
    • symmetric_with_uint8 for SymmetricWithUint8 quantization schema.
    • symmetric_with_power2_scale for SymmetricWithPower2Scale quantization schema.
  • quantization-precision specifies the precision used to quantized the nodes:
    • Int8 for int8 quantization (Default).
    • Int16 for int16 quantization.
  • quantization-precision-bias specifies the precision used to quantize the bias operand of some of the nodes (e.g. FullyConnected, Convolution):
    • Int8 for int8 quantization.
    • Int32 for int32 quantization (Default). More details about the quantization parameters can be found here.

For example, in order to profile, quantize and compile the ResNet50 model you can use the commands:

image-classifier cat_285.png ... -image-mode=0to1 -model=resnet50 -model-input-name=gpu_0/data -dump-profile=profile.yml
model-compiler -backend=CPU -model=resnet50 -emit-bundle=build -model-input=gpu_0/data,float,[1,3,224,224] -load-profile=profile.yml

Similarly, for the LeNetMnist model:

image-classifier 0_1009.png ... -image-mode=0to1 -model=lenet_mnist -model-input-name=data -dump-profile=profile.yml
model-compiler -backend=CPU -model=lenet_mnist -emit-bundle=build -model-input=data,float,[1,1,28,28] -load-profile=profile.yml

It is important to note that by default the quantization of a model is performed only for the intermediate nodes of the graph, without affecting the data types of the model inputs and outputs. In the examples above, the data type for the model input (image tensor) and the model output remains float even though the intermediate operators and tensors use int8 data type. If you want to convert also the model input and output placeholders you can use the option convert-placeholders.

Note that in order for the model quantization to take place properly, the batch size of the model used during profiling (e.g. using the minibatch option for the image-classifier) must be the same as the batch size of the model during quantization (e.g. using the model-input option for the model-compiler). Here is the generic example for quantizing the LeNetMnist model:

image-classifier -minibatch=<N> ...
model-compiler -model-input=data,float,[<N>,1,28,28] ...

In the above example, if different batch sizes N are used during profiling/quantization then the model-compiler will throw an error which signals that different graphs were used during profiling and quantization.

When compiling a quantized bundle you can choose to disable the quantization of some of the graph operators which might be more susceptible to quantization by using the option keep-original-precision-for-nodes. For example in order to disable the quantization for the operators Add,Sub,Mul,Div,Exp and SoftMax you can use:

model-compiler ... -keep-original-precision-for-nodes=Add,Sub,Mul,Div,Exp,SoftMax

Cross-compile a bundle for a specific architecture

Since the CPU backend is based on LLVM, the Glow tools can be used to cross-compile bundles for different target architectures. To specify the target architecture you must use the -target and -mcpu flags (if no target flags are provided the bundle will be generated by default for the native architecture - the one which is running Glow). Below you have a table with examples for some of the common architecures:

Architecture Option
x86 - 32 bit -target=i386
x86 - 64 bit -target=x86_64
ARM Cortex M0 -target=arm -mcpu=cortex-m0
ARM Cortex M4 -target=arm -mcpu=cortex-m4
ARM Cortex M7 -target=arm -mcpu=cortex-m7
ARM Cortex M33 -target=arm -mcpu=cortex-m33
ARM Cortex A53 -target=aarch64 -mcpu=cortex-a53
ARM Cortex A72 -target=aarch64 -mcpu=cortex-a72

The bundle can be cross-compiled for any target architecture supported by LLVM. For the complete list of LLVM target architectures you can type llc -version command in Linux (assuming you have LLVM installed). For example the LLVM 8.0.1 has the following supported architectures:

LLVM (http://llvm.org/):
  LLVM version 8.0.1

  Optimized build.
  Default target: x86_64-pc-linux-gnu
  Host CPU: skylake

  Registered Targets:
    aarch64    - AArch64 (little endian)
    aarch64_be - AArch64 (big endian)
    amdgcn     - AMD GCN GPUs
    arm        - ARM
    arm64      - ARM64 (little endian)
    armeb      - ARM (big endian)
    avr        - Atmel AVR Microcontroller
    bpf        - BPF (host endian)
    bpfeb      - BPF (big endian)
    bpfel      - BPF (little endian)
    hexagon    - Hexagon
    lanai      - Lanai
    mips       - MIPS (32-bit big endian)
    mips64     - MIPS (64-bit big endian)
    mips64el   - MIPS (64-bit little endian)
    mipsel     - MIPS (32-bit little endian)
    msp430     - MSP430 [experimental]
    nvptx      - NVIDIA PTX 32-bit
    nvptx64    - NVIDIA PTX 64-bit
    ppc32      - PowerPC 32
    ppc64      - PowerPC 64
    ppc64le    - PowerPC 64 LE
    r600       - AMD GPUs HD2XXX-HD6XXX
    sparc      - Sparc
    sparcel    - Sparc LE
    sparcv9    - Sparc V9
    systemz    - SystemZ
    thumb      - Thumb
    thumbeb    - Thumb (big endian)
    wasm32     - WebAssembly 32-bit
    wasm64     - WebAssembly 64-bit
    x86        - 32-bit X86: Pentium-Pro and above
    x86-64     - 64-bit X86: EM64T and AMD64
    xcore      - XCore

Extra options

  • If you want to change the base name of the bundle artifacts you can use the option:

    -network-name=<network-name>

    For example by using the option -network-name=mb2 the generated bundle artifacts will be named mb2.o, mb2.h, mb2.weights.bin and mb2.weights.txt.

  • The generated bundle object file <network_name>.o is non-relocatable by default. It is possible to control the relocation model with the command line option -relocation-model=<mode>. This option supports two modes:

    • static: (Default) Produce non-relocatable code.
    • pic: Produce position independent code.

  • The ONNX format allows some of the tensor dimensions to be undefined (that is, marked as symbols and not as actual sizes) within the model. In the ONNX proto description the symbolic sizes are marked with dim_param while the actual sizes are marked with dim_value. For example, when inspecting an image classification model with Netron one might see that the input tensor size is defined as None x 3 x 224 x 224 where None is the undefined size (symbol) associated with the batch size of the model. Glow cannot compile a model with undefined sizes, therefore the user must assign manually the actual values for all the symbols. The user will be prompted with an error if there is an undefined symbol (for example ONNX model symbol 'None' is undefined.). In the above example, the user can define the symbol None to have the value 1 by using the following option:

    -onnx-define-symbol=None,1

    Multiple symbols can be defined using this option, for example:

    -onnx-define-symbol=<symbol_name1>,<symbol_value1>
-onnx-define-symbol=<symbol_name2>,<symbol_value2>
..................................................

  • When cross-compiling bundles for some target architectures you might be interested in generating a bundle compatible with a given float ABI (Application Binary Interface) type (soft or hard). The LLVM backend can be instructed to generate an object file using a specific float ABI by using the option -float-abi=hard or -float-abi=soft.

  • When compiling the bundle it is useful to view the final form of the graph after all the transformations and optimizations performed by Glow (which might differ from the initial model). You can generate the graph visual representation in .dot format by using the -dump-graph-DAG option like in this:

    -dump-graph-DAG=graph.dot

    Additionally, you can convert the .dot file to .pdf format using the dot utility available on Linux like this:

    dot -Tpdf graph.dot -o graph.pdf

  • For debugging purposes you can choose to compile the model in instrumentation mode with the option -instrument-debug such that the model will display during run-time the content of all the tensors associated to all the operands of all the IR instructions. Additionally, you can choose the format of the dumped information with the option -instrument-debug-format=<format>:

    • console (Default) All the operands (tensors) are displayed in text format in the console. Large tensors are only partially displayed.
    • bin The operands (tensors) are dumped in binary format in separate files (one tensor per file). Each file will contain the tensor type and tensor data.
    • txt The operands (tensors) are dumped in text format in separate files (one tensor per file). Each file will contain the tensor type and tensor data. One such file might look like this: 
      float<1 x 2 x 3>
1.1, 2.2, 3.3, 4.4, 5.5, 6.6,
    • rawbin The operands (tensors) are dumped in binary format in separate files (one tensor per file). Each file will contain ONLY the tensor data.
    • rawtxt The operands (tensors) are dumped in text format in separate files (one tensor per file). Each file will contain ONLY the tensor data. One such file might look like this: 
      1.1, 2.2, 3.3, 4.4, 5.5, 6.6,

    The names of the dump files have a simple format data[idx].bin or data[idx].txt but a separate meta file debug.info is dumped at compile-time which makes the association between each binary file and the operand of the IR instruction to which it belongs. The debug.info meta file might look like this:

    Format: bin

Type: quantize
Name: Conv_MobilenetV1_MobilenetV1_Conv2d_0_Conv2D__2_quantize
[0] Dest:      data0000.bin    i8[S:0.0156 O:0][-2.000,1.984]<1 x 224 x 224 x 3>
[1] Src:       data0001.bin    float<1 x 224 x 224 x 3>

Type: convolution
Name: Conv_MobilenetV1_MobilenetV1_Conv2d_0_Conv2D__2
[0] Dest:      data0002.bin    i8[S:0.2500 O:0][-32.000,31.750]<1 x 112 x 112 x 32>
[1] Src:       data0003.bin    i8[S:0.0156 O:0][-2.000,1.984]<1 x 224 x 224 x 3>
[2] Filter:    data0004.bin    i8[S:0.0312 O:0][-4.000,3.969]<32 x 3 x 3 x 3>
[3] Bias:      data0005.bin    i32[S:0.0005 O:0][-1048576.000,1048576.000]<32>

    All the dump files and the meta file debug.info are dumped in a folder debug relative to the current directory at compile-time (if it does not exist it is created). You can choose a different directory with the option -instrument-debug-dir=<dir>.

Bundle memory layout

The memory of a bundle is organized in three separate memory regions which must be allocated by the user application code and provided through the bundle interface:

  • constantWeight - contains the model constant weights. The user application must:

    • allocate this memory region (statically or dynamically)
    • initialize this memory region with the content of the generated weights file in one of two possible formats:
      • binary format (<network_name>.weights.bin) used to initialize this memory region (allocated statically or dynamically) by loading the binary file dynamically at run-time using standard C function like fopen.
      • text format (<network_name>.weights.txt) used to initialize this memory region (only if statically allocated) by including the text file statically at compile-time as a C array using the #include pre-processor directive. This format is suitable for target architectures which do not have file systems (for example microcontrollers).
    • provide the base address of this memory region to the inference function

  • mutableWeight - contains all the model inputs and outputs (graph placeholders). The tensors corresponding to different inputs and outputs are identified using offsets relative to the base address of this memory region. The user application must:

    • allocate this memory region (statically or dynamically)
    • initialize the model input tensors from this memory region with the desired input data before running the inference
    • provide the base address of this memory region to the inference function
    • read the model output tensors from this memory region after running the inference

  • activations - this memory region is a scratch memory required for the bundle code to store the intermediate results of the graph computation (activations). The user application must:

    • allocate this memory region (statically or dynamically)
    • provide the base address of this memory region to the inference function
    • this memory region is NOT required to be initialized

The required sizes for all the memory regions described above are provided in the bundle interface. Also all the memory regions must be allocated with a minimum alignment which is also provided in the interface (typically 64 bytes). For example, for aligning a statically allocated buffer one can use the following C syntax:

__attribute__((aligned(64)))
uint8_t aligned_buffer[BUFFER_SIZE];

Static bundle API

This is the default bundle API obtained by generating the bundle with the option -bundle-api=static. Below is an example of how the auto-generated header file looks like for the Lenet Mnist model:

// Placeholder address offsets within mutable buffer (bytes)
#define LENET_MNIST_data        0
#define LENET_MNIST_softmax__1  3136

// Memory sizes (bytes)
#define LENET_MNIST_CONSTANT_MEM_SIZE     1724672
#define LENET_MNIST_MUTABLE_MEM_SIZE      3200
#define LENET_MNIST_ACTIVATIONS_MEM_SIZE  57600

// Memory alignment (bytes)
#define LENET_MNIST_MEM_ALIGN  64

// Bundle entry point (inference function)
void lenet_mnist(uint8_t *constantWeight, uint8_t *mutableWeight, uint8_t *activations);

The header file contains all the information required to run the bundle, defined in a static manner using macro defines:

  • the offsets of all the placeholders (graph inputs/outputs) within the mutableWeight memory
  • the sizes for all the memory regions
  • the alignment required for allocating the memory regions
  • the inference function prototype

All the definitions names (the macros and the inference function) are prefixed with the model name, in this example with lenet_mnist. If you want to change the model name you can use the command line option -network-name, for example -network-name=my_bundle.

The auto-generated header file file also contains some extra defines to help with writing the user application code:

// Memory alignment definition with given alignment size
// for static allocation of memory.
#define GLOW_MEM_ALIGN(size)  __attribute__((aligned(size)))

// Macro function to get the absolute address of a
// placeholder using the base address of the mutable
// weight buffer and placeholder offset definition.
#define GLOW_GET_ADDR(mutableBaseAddr, placeholderOff)  (((uint8_t*)(mutableBaseAddr)) + placeholderOff)

For example, in order to allocate and initialize all the memory regions, you need to write the following in the user application (lenet_mnist.weights.txt is the file containing the model weights serialized as text):

GLOW_MEM_ALIGN(LENET_MNIST_MEM_ALIGN)
uint8_t constantWeight[LENET_MNIST_CONSTANT_MEM_SIZE] = {
#include "lenet_mnist.weights.txt"
};

GLOW_MEM_ALIGN(LENET_MNIST_MEM_ALIGN)
uint8_t mutableWeight[LENET_MNIST_MUTABLE_MEM_SIZE];

GLOW_MEM_ALIGN(LENET_MNIST_MEM_ALIGN)
uint8_t activations[LENET_MNIST_ACTIVATIONS_MEM_SIZE];

In order to obtain the absolute addresses of the model inputs/outputs you need to write the following in the user application:

uint8_t *inputAddr  = GLOW_GET_ADDR(mutableWeight, LENET_MNIST_data);
uint8_t *outputAddr = GLOW_GET_ADDR(mutableWeight, LENET_MNIST_softmax__1);

Dynamic bundle API

This is the bundle API obtained by generating the bundle with the option -bundle-api=dynamic. Below is an example of how the auto-generated header file looks like for the Resnet50 model:

// Bundle memory configuration (memory layout)
extern BundleConfig resnet50_config;

// Bundle entry point (inference function)
void resnet50(uint8_t *constantWeight, uint8_t *mutableWeight, uint8_t *activations);

This API has all the information about the memory configuration encapsulated in a structure named <network_name>_config. The layout of this structure is defined by the type BundleConfig which is also included in the generated header file:

// Type describing the config of a generated bundle.
struct BundleConfig {
  // Size of the constant weight variables memory area.
  uint64_t constantWeightVarsMemSize;
  // Size of the mutable weight variables memory area.
  uint64_t mutableWeightVarsMemSize;
  // Size of the activations memory area.
  uint64_t activationsMemSize;
  // Alignment to be used for weights and activations.
  uint64_t alignment;
  // Number of symbols in the symbol table.
  uint64_t numSymbols;
  // Symbol table.
  const SymbolTableEntry *symbolTable;
};

Similar to the static API, this structure contains:

  • the sizes for all the memory regions
  • the alignment required for allocating all the memory regions
  • the number of symbols
  • the descriptions of all the symbols as an array of symbol entries

In this case the notion of symbol might include not only the model placeholders but also the model constant weights. Each symbol is described according to the SymbolTableEntry structure definition (included also in the header file):

// Type describing a symbol table entry of a generated bundle.
struct SymbolTableEntry {
  // Name of a variable.
  const char *name;
  // Offset of the variable inside the memory area.
  uint64_t offset;
  // The number of elements inside this variable.
  uint64_t size;
  // Variable kind: 1 if it is a mutable variable, 0 otherwise.
  char kind;
};

For each symbol the following information is registered:

  • the symbol name
  • the symbol kind: whether is mutable (placeholder) or not (constant)
  • the size in bytes
  • the offset: if the symbol is mutable this is the offset of the variable within the mutableWeight buffer, otherwise this is the offset of the variable within the constantWeight buffer

The user has to look up the symbol entries to find the model variables (placeholders or constants) at run-time (dynamically).

How to use the bundle

This section describes how to use the bundle produced with the CPU backend. Other backends may have different interfaces. To integrate the bundle into your project, you need to do the following:

  1. Link the object file <network_name>.o to your application.
  2. Include the header file <network_name>.h in your application.
  3. Allocate memory for all the required buffers: constantWeight, mutableWeight, activations:
    • For the static API the buffer sizes and alignments are given by the macro defines from the header file <network_name>.h and are available both at compile-time and run-time.
    • For the dynamic API the buffer sizes and alignments are provided in the <network_name>_config structure and are only available at run-time.
  4. Initialize the content of the constantWeight buffer with the model weights using either the <network_name>.weights.bin file or the <network_name>.weights.txt file.
  5. Initialize the model input tensors from the mutableWeight buffer (e.g. image data).
  6. Invoke the main entry function <network_name> by providing the base addresses of the memory regions previously allocated.
  7. After the main entry function has returned, you can find the results in the corresponding model output tensors from the mutableWeight buffer.

Extra utilities

Visualize models

A very popular tool for visualizing the original model before compiling with Glow is Netron which has:

  • an online browser version here (drag & drop a model into the browser window).
  • an offline standalone version here (drag & drop a model into the GUI window).

Convert models

The Glow compiler currently has support only for Caffe2, ONNX and TensorFlowLite model formats. Since a lot of popular models are available in other formats, for example TensorFlow, it is useful to have tools to convert models between different formats. The most used tools for format conversion are:

Here we demonstrate how to convert a TensorFlow model to ONNX. We will convert a MobileNet V1 image classification model which operates on 128 x 128 RGB images and 1001 classes. You can download the model archive in TensorFlow format from here.

In order to convert the model to ONNX format using the MMdnn tool:

  1. Install MMdnn from here.
  2. Download the MobileNet V1 model archive and unzip it.
  3. Run the following command to convert the TensorFlow frozen file mobilenet_v1_0.25_128_frozen.pb to the ONNX model file mobilenet_v1_0.25_128.onnx: 
    mmconvert                                              \
  --srcFramework tensorflow                            \
  --inputWeight mobilenet_v1_0.25_128_frozen.pb        \
  --inNodeName input                                   \
  --inputShape 128,128,3                               \
  --dstNodeName MobilenetV1/Predictions/Softmax        \
  --dstFramework onnx                                  \
  --outputModel mobilenet_v1_0.25_128.onnx

In order to convert the model to ONNX format using the tf2onnx tool:

  1. Install tf2onnx from here.
  2. Download the MobileNet V1 model archive and unzip it.
  3. Run the following command to convert the TensorFlow frozen file mobilenet_v1_0.25_128_frozen.pb to the ONNX model file mobilenet_v1_0.25_128.onnx: 
    python -m tf2onnx.convert                      \
  --input mobilenet_v1_0.25_128_frozen.pb      \
  --inputs input:0                             \
  --outputs MobilenetV1/Predictions/Softmax:0  \
  --output mobilenet_v1_0.25_128.onnx

After converting the model to ONNX you can build a bundle using the Glow model-compiler tool with the command:

model-compiler -backend=CPU -model=mobilenet_v1_0.25_128.onnx -emit-bundle=bundle

When converting the model with tf2onnx some undefined symbols might end up in the ONNX model. For example, the input batch size might be encoded as a symbol unk__286 which can be defined to a value of 1 by using the extra option:

-onnx-define-symbol=unk__286,1

Edit protocol buffers

There are times when it is useful to manually edit the protocol buffers which are used for model formats like ONNX and TensorFlow. One such example is when one needs to modify an ONNX model in order to circumvent a limitation of the Glow compiler (for example when an attribute is not relevant but causes compiler issues, etc).

In order modify a protocol buffer you need:

  • the proto compiler application protoc which can be found here.
  • the proto definition:
    • for ONNX format the proto definition onnx.proto can be found here.
    • for TensorFlow format the proto definition saved_model.proto can be found here.

When you need to manually edit an ONNX model you will need to:

  • decode the model proto to text format.
  • edit the model in text format.
  • encode the model back to proto format.

After you install the protoc application and you download the proto definitions, you can perform the actions defined below:

  • Decode an ONNX model from proto model.onnx to text file model.onnxtxt:

    protoc onnx.proto --decode=onnx.ModelProto < model.onnx > model.onnxtxt

  • Encode an ONNX model from text file model.onnxtxt to proto model.onnx:

    protoc onnx.proto --encode=onnx.ModelProto < model.onnxtxt > model.onnx

  • Decode an ONNX tensor from proto tensor.pb to text file tensor.pbtxt:

    protoc onnx.proto --decode=onnx.TensorProto < tensor.pb > tensor.pbtxt

  • Encode an ONNX tensor from text file tensor.pbtxt to proto tensor.pb:

    protoc onnx.proto --encode=onnx.TensorProto < tensor.pbtxt > tensor.pb

  • Decode/Encode TensorFlow saved model to/from text file:

    protoc saved_model.proto --decode tensorflow.SavedModel < model.pb > model.pbtxt
protoc saved_model.proto --encode tensorflow.SavedModel < model.pbtxt < model.pb

  • Decode/Encode TensorFlow frozen graph to/from text file:

    protoc saved_model.proto --decode tensorflow.GraphDef < model.pb > model.pbtxt
protoc saved_model.proto --encode tensorflow.GraphDef < model.pbtxt > model.pb