diff options
Diffstat (limited to 'docs/use_cases/kws.md')
| -rw-r--r-- | docs/use_cases/kws.md | 300 |
1 files changed, 164 insertions, 136 deletions
diff --git a/docs/use_cases/kws.md b/docs/use_cases/kws.md index dc0e1f5..9b0372c 100644 --- a/docs/use_cases/kws.md +++ b/docs/use_cases/kws.md @@ -17,50 +17,54 @@ ## Introduction -This document describes the process of setting up and running the Arm® Ethos™-U55 Keyword Spotting -example. +This document describes the process of setting up and running the Arm® *Ethos™-U55* Keyword Spotting example. -Use case code could be found in [source/use_case/kws](../../source/use_case/kws]) directory. +Use-case code could be found in the following directory: [source/use_case/kws](../../source/use_case/kws]). ### Preprocessing and feature extraction -The DS-CNN keyword spotting model that is supplied with the Code Samples expects audio data to be preprocessed in -a specific way before performing an inference. This section aims to provide an overview of the feature extraction -process used. +The `DS-CNN` keyword spotting model that is used with the Code Samples expects audio data to be preprocessed in a +specific way before performing an inference. -First the audio data is normalized to the range (-1, 1). +Therefore, this section aims to provide an overview of the feature extraction process used. -> **Note:** Mel-frequency cepstral coefficients (MFCCs) are a common feature extracted from audio data and can be used as ->input for machine learning tasks like keyword spotting and speech recognition. ->See source/application/main/include/Mfcc.hpp for implementation details. +First, the audio data is normalized to the range (`-1`, `1`). -Next, a window of 640 audio samples is taken from the start of the audio clip. From these 640 samples we calculate 10 +> **Note:** Mel-Frequency Cepstral Coefficients (MFCCs) are a common feature that is extracted from audio data and can +> be used as input for machine learning tasks. Such as keyword spotting and speech recognition. For implementation +> details, please refer to: `source/application/main/include/Mfcc.hpp` + +Next, a window of 640 audio samples is taken from the start of the audio clip. From these 640 samples, we calculate 10 MFCC features. The whole window is shifted to the right by 320 audio samples and 10 new MFCC features are calculated. This process of -shifting and calculating is repeated until the end of the 16000 audio samples needed to perform an inference is reached. -In total this will be 49 windows that each have 10 MFCC features calculated for them, giving an input tensor of shape -49x10. +shifting and calculating is repeated until the end of the 16000 audio samples required to perform an inference is +reached. + +In total, this is 49 windows that each have 10 MFCC features calculated for them, giving an input tensor of shape 49x10. -These extracted features are quantized, and an inference is performed. +These extracted features are quantized and an inference is performed.  -If the audio clip is longer than 16000 audio samples then the initial starting position is offset by 16000/2 = 8000 -audio samples. From this new starting point, MFCC features for the next 16000 audio samples are calculated and another -inference is performed (i.e. do an inference for samples 8000-24000). +If the audio clip is longer than 16000 audio samples, then the initial starting position is offset by `16000/2 = 8000` +audio samples. From this new starting point, MFCC features for the next `16000` audio samples are calculated and another +inference is performed. In other words, do an inference for samples `8000-24000`. -> **Note:** Parameters of the MFCC feature extraction such as window size, stride, number of features etc. all depend on ->what was used during model training. These values are specific to each model and if you try a different keyword spotting ->model that uses MFCC input then values are likely to need changing to match the new model. -In addition, MFCC feature extraction methods can vary slightly with different normalization methods or scaling etc. being used. +> **Note:** Parameters of the MFCC feature extraction all depend on what was used during model training. These values +> are specific to each model.\ +If you try a different keyword spotting model that uses MFCC input, then values check to see if the values need changing +to match the new model. + +In addition, MFCC feature extraction methods can vary slightly with different normalization methods or scaling being +used. ### Postprocessing -After an inference is complete the highest probability detected word is output to console, providing its probability is -larger than a threshold value (default 0.9). +After an inference is complete, the word with the highest detected probability is output to console. Providing that the +probability is larger than a threshold value. The default is set to `0.9`. -If multiple inferences are performed for an audio clip, then multiple results will be output. +If multiple inferences are performed for an audio clip, then multiple results are output. ### Prerequisites @@ -70,58 +74,67 @@ See [Prerequisites](../documentation.md#prerequisites) ### Build options -In addition to the already specified build option in the main documentation, keyword spotting use case adds: +In addition to the already specified build option in the main documentation, the Keyword Spotting use-case adds: -- `kws_MODEL_TFLITE_PATH` - Path to the NN model file in TFLite format. Model will be processed and included into the application axf file. The default value points to one of the delivered set of models. Note that the parameters `kws_LABELS_TXT_FILE`,`TARGET_PLATFORM` and `ETHOS_U55_ENABLED` should be aligned with the chosen model, i.e.: - - if `ETHOS_U55_ENABLED` is set to `On` or `1`, the NN model is assumed to be optimized. The model will naturally fall back to the Arm® Cortex®-M CPU if an unoptimized model is supplied. - - if `ETHOS_U55_ENABLED` is set to `Off` or `0`, the NN model is assumed to be unoptimized. Supplying an optimized model in this case will result in a runtime error. +- `kws_MODEL_TFLITE_PATH` - The path to the NN model file in `TFLite` format. The model is processed and then included + into the application `axf` file. The default value points to one of the delivered set of models. Note that the + parameters `kws_LABELS_TXT_FILE`,`TARGET_PLATFORM`, and `ETHOS_U55_ENABLED` must be aligned with the chosen model. In + other words: + - If `ETHOS_U55_ENABLED` is set to `On` or `1`, then the NN model is assumed to be optimized. The model naturally + falls back to the Arm® *Cortex®-M* CPU if an unoptimized model is supplied. + - If `ETHOS_U55_ENABLED` is set to `Off` or `0`, then the NN model is assumed to be unoptimized. Supplying an + optimized model in this case results in a runtime error. -- `kws_FILE_PATH`: Path to the directory containing audio files, or a path to single WAV file, to be used in the application. The default value points - to the resources/kws/samples folder containing the delivered set of audio clips. +- `kws_FILE_PATH`: The path to the directory containing audio files, or a path to single WAV file, to be used in the + application. The default value points to the `resources/kws/samples` folder that contains the delivered set of audio + clips. -- `kws_LABELS_TXT_FILE`: Path to the labels' text file. The file is used to map key word class index to the text - label. The default value points to the delivered labels.txt file inside the delivery package. +- `kws_LABELS_TXT_FILE`: Path to the text file of the label. The file is used to map letter class index to the text + label. The default value points to the delivered `labels.txt` file inside the delivery package. -- `kws_AUDIO_RATE`: Input data sampling rate. Each audio file from kws_FILE_PATH is preprocessed during the build to - match NN model input requirements. Default value is 16000. +- `kws_AUDIO_RATE`: The input data sampling rate. Each audio file from `kws_FILE_PATH` is preprocessed during the build + to match the NN model input requirements. The default value is `16000`. -- `kws_AUDIO_MONO`: If set to ON the audio data will be converted to mono. Default is ON. +- `kws_AUDIO_MONO`: If set to `ON`, then the audio data is converted to mono. The default value is `ON`. -- `kws_AUDIO_OFFSET`: Start loading audio data starting from this offset (in seconds). Default value is 0. +- `kws_AUDIO_OFFSET`: Begins loading audio data and starts from this specified offset, defined in seconds. the default + value is set to `0`. -- `kws_AUDIO_DURATION`: Length of the audio data to be used in the application in seconds. Default is 0 meaning the - whole audio file will be taken. +- `kws_AUDIO_DURATION`: The length of the audio data to be used in the application in seconds. The default is `0`, + meaning that the whole audio file is used. - `kws_AUDIO_MIN_SAMPLES`: Minimum number of samples required by the network model. If the audio clip is shorter than - this number, it is padded with zeros. Default value is 16000. + this number, then it is padded with zeros. The default value is `16000`. -- `kws_MODEL_SCORE_THRESHOLD`: Threshold value [0.0, 1.0] that must be applied to the inference results for a - label to be deemed valid. Default is 0.9 +- `kws_MODEL_SCORE_THRESHOLD`: Threshold value that must be applied to the inference results for a label to be deemed + valid. Goes from 0.00 to 1.0. The default is `0.9`. -- `kws_ACTIVATION_BUF_SZ`: The intermediate/activation buffer size reserved for the NN model. By default, it is set - to 1MiB and should be enough for most models. +- `kws_ACTIVATION_BUF_SZ`: The intermediate, or activation, buffer size reserved for the NN model. By default, it is set + to 2MiB and is enough for most models -In order to build **ONLY** keyword spotting example application add to the `cmake` command line specified in [Building](../documentation.md#Building) `-DUSE_CASE_BUILD=kws`. +To **ONLY** build the automatic speech recognition example application, add `-DUSE_CASE_BUILD=kws` to the `cmake` +command line, as specified in: [Building](../documentation.md#Building). ### Build process -> **Note:** This section describes the process for configuring the build for `MPS3: SSE-300` for different target platform see [Building](../documentation.md#Building) section. +> **Note:** This section describes the process for configuring the build for the *MPS3: SSE-300*. To build for a +> different target platform, please refer to: [Building](../documentation.md#Building). -In order to build **only** the keyword spotting example, create a build directory and -navigate inside, for example: +To build **only** the keyword spotting example, create a build directory and navigate inside, like so: ```commandline mkdir build_kws && cd build_kws ``` -On Linux, execute the following command to build Keyword Spotting application to run on the Ethos-U55 Fast Model when providing only the mandatory arguments for CMake configuration: +On Linux, when providing only the mandatory arguments for CMake configuration, execute the following command to build +**only** the Keyword Spotting application to run on the *Ethos-U55* Fast Model: ```commandline cmake ../ -DUSE_CASE_BUILD=kws ``` -To configure a build that can be debugged using Arm-DS, we can just specify -the build type as `Debug` and use the `Arm Compiler` toolchain file: +To configure a build that can be debugged using Arm DS specify the build type as `Debug` and then use the `Arm Compiler` +toolchain file: ```commandline cmake .. \ @@ -130,24 +143,25 @@ cmake .. \ -DUSE_CASE_BUILD=kws ``` -Also see: +For further information, please refer to: - [Configuring with custom TPIP dependencies](../sections/building.md#configuring-with-custom-tpip-dependencies) - [Using Arm Compiler](../sections/building.md#using-arm-compiler) - [Configuring the build for simple_platform](../sections/building.md#configuring-the-build-for-simple_platform) -- [Working with model debugger from Arm FastModel Tools](../sections/building.md#working-with-model-debugger-from-arm-fastmodel-tools) +- [Working with model debugger from Arm Fast Model Tools](../sections/building.md#working-with-model-debugger-from-arm-fastmodel-tools) -> **Note:** If re-building with changed parameters values, it is highly advised to clean the build directory and re-run the CMake command. +> **Note:** If re-building with changed parameters values, we recommend that you clean the build directory and re-run +> the CMake command. -If the CMake command succeeded, build the application as follows: +If the CMake command succeeds, build the application as follows: ```commandline make -j4 ``` -Add VERBOSE=1 to see compilation and link details. +To see compilation and link details, add `VERBOSE=1`. -Results of the build will be placed under `build/bin` folder: +Results of the build are placed under the `build/bin` folder, like so: ```tree bin @@ -157,29 +171,31 @@ bin └── sectors ├── images.txt └── kws - ├── dram.bin + ├── ddr.bin └── itcm.bin ``` -Where: +The `bin` folder contains the following files: -- `ethos-u-kws.axf`: The built application binary for the Keyword Spotting use case. +- `ethos-u-kws.axf`: The built application binary for the Keyword Spotting use-case. -- `ethos-u-kws.map`: Information from building the application (e.g. libraries used, what was optimized, location of - objects) +- `ethos-u-kws.map`: Information from building the application. For example: The libraries used, what was optimized, and + the location of objects. - `ethos-u-kws.htm`: Human readable file containing the call graph of application functions. -- `sectors/kws`: Folder containing the built application, split into files for loading into different FPGA memory regions. +- `sectors/kws`: Folder containing the built application. It is split into files for loading into different FPGA memory + regions. -- `sectors/images.txt`: Tells the FPGA which memory regions to use for loading the binaries in sectors/\*\* folder. +- `sectors/images.txt`: Tells the FPGA which memory regions to use for loading the binaries in the `sectors/..` folder. ### Add custom input -The application performs inference on audio data found in the folder, or an individual file, set by the CMake parameter `kws_FILE_PATH`. +The application anomaly detection is set up to perform inferences on data found in the folder, or an individual file, +that is pointed to by the parameter `kws_FILE_PATH`. -To run the application with your own audio clips first create a folder to hold them and then copy the custom audio files -into this folder, for example: +To run the application with your own audio clips, first create a folder to hold them and then copy the custom clips into +the following folder: ```commandline mkdir /tmp/custom_wavs @@ -189,7 +205,7 @@ cp my_clip.wav /tmp/custom_wavs/ > **Note:** Clean the build directory before re-running the CMake command. -Next set `kws_FILE_PATH` to the location of this folder when building: +Next, when building, set `kws_FILE_PATH` to the location of the following folder: ```commandline cmake .. \ @@ -197,10 +213,11 @@ cmake .. \ -DUSE_CASE_BUILD=kws ``` -The audio clips found in the `kws_FILE_PATH` folder will be picked up and automatically converted to C++ files during the -CMake configuration stage and then compiled into the application during the build phase for performing inference with. +The audio flies found in the `kws_FILE_PATH` folder are picked up and automatically converted to C++ files during the +CMake configuration stage. They are then compiled into the application during the build phase for performing inference +with. -The log from the configuration stage should tell you what audio clip directory path has been used: +The log from the configuration stage tells you what audio directory path has been used: ```log -- User option kws_FILE_PATH is set to /tmp/custom_wavs @@ -212,25 +229,29 @@ The log from the configuration stage should tell you what audio clip directory p -- kws_FILE_PATH=/tmp/custom_wavs ``` -After compiling, your custom inputs will have now replaced the default ones in the application. +After compiling, your custom inputs have now replaced the default ones in the application. -> **Note:** The CMake parameter `kws_AUDIO_MIN_SAMPLES` determine the minimum number of input sample. When building the application, -if the size of the audio clips is less then `kws_AUDIO_MIN_SAMPLES` then it will be padded so that it does. +> **Note:** The CMake parameter `kws_AUDIO_MIN_SAMPLES` determines the minimum number of input samples. When building +> the application, if the size of the audio clips is less then `kws_AUDIO_MIN_SAMPLES`, then it is padded until it +> matches. ### Add custom model The application performs inference using the model pointed to by the CMake parameter `kws_MODEL_TFLITE_PATH`. -> **Note:** If you want to run the model using Ethos-U55, ensure your custom model has been run through the Vela compiler successfully before continuing. See [Optimize model with Vela compiler](../sections/building.md#Optimize-custom-model-with-Vela-compiler). +> **Note:** If you want to run the model using an *Ethos-U55*, ensure that your custom model has been successfully run +> through the Vela compiler *before* continuing. + +For further information: [Optimize model with Vela compiler](../sections/building.md#Optimize-custom-model-with-Vela-compiler). -To run the application with a custom model you will need to provide a labels_<model_name>.txt file of labels -associated with the model. Each line of the file should correspond to one of the outputs in your model. See the provided -ds_cnn_labels.txt file for an example. +To run the application with a custom model, you must provide a `labels_<model_name>.txt` file of labels that are +associated with the model. Each line of the file must correspond to one of the outputs in your model. Refer to the +provided `ds_cnn_labels.txt` file for an example. -Then, you must set kws_MODEL_TFLITE_PATH to the location of the Vela processed model file and kws_LABELS_TXT_FILE -to the location of the associated labels file. +Then, you must set `kws_MODEL_TFLITE_PATH` to the location of the Vela processed model file and `kws_LABELS_TXT_FILE`to +the location of the associated labels file. -An example: +For example: ```commandline cmake .. \ @@ -241,11 +262,11 @@ cmake .. \ > **Note:** Clean the build directory before re-running the CMake command. -The `.tflite` model file pointed to by `kws_MODEL_TFLITE_PATH` and labels text file pointed to by `kws_LABELS_TXT_FILE` will -be converted to C++ files during the CMake configuration stage and then compiled into the application for performing -inference with. +The `.tflite` model file pointed to by `kws_MODEL_TFLITE_PATH` and labels text file pointed to by `kws_LABELS_TXT_FILE` +are converted to C++ files during the CMake configuration stage. They are then compiled into the application for +performing inference with. -The log from the configuration stage should tell you what model path and labels file have been used: +The log from the configuration stage tells you what model path and labels file have been used, for example: ```log -- User option kws_MODEL_TFLITE_PATH is set to <path/to/custom_model_after_vela.tflite> @@ -260,38 +281,43 @@ custom_model_after_vela.tflite.cc ... ``` -After compiling, your custom model will have now replaced the default one in the application. +After compiling, your custom model has now replaced the default one in the application. ## Setting up and running Ethos-U55 code sample ### Setting up the Ethos-U55 Fast Model -The FVP is available publicly from [Arm Ecosystem FVP downloads](https://developer.arm.com/tools-and-software/open-source-software/arm-platforms-software/arm-ecosystem-fvps). +The FVP is available publicly from +[Arm Ecosystem FVP downloads](https://developer.arm.com/tools-and-software/open-source-software/arm-platforms-software/arm-ecosystem-fvps). -For Ethos-U55 evaluation, please download the MPS3 version of the Arm® Corstone™-300 model that contains Ethos-U55 and -Cortex-M55. The model is currently only supported on Linux based machines. To install the FVP: +For the *Ethos-U55* evaluation, please download the MPS3 version of the Arm® *Corstone™-300* model that contains both +the *Ethos-U55* and *Cortex-M55*. The model is only supported on Linux-based machines. -- Unpack the archive +To install the FVP: -- Run the install script in the extracted package +- Unpack the archive. + +- Run the install script in the extracted package: ```commandline ./FVP_Corstone_SSE-300_Ethos-U55.sh ``` -- Follow the instructions to install the FVP to your desired location +- Follow the instructions to install the FVP to the required location. ### Starting Fast Model simulation -Once completed the building step, application binary ethos-u-kws.axf can be found in the `build/bin` folder. -Assuming the install location of the FVP was set to ~/FVP_install_location, the simulation can be started by: +Once the building has been completed, the application binary `ethos-u-kws.axf` can be found in the `build/bin` folder. + +Assuming that the install location of the FVP was set to `~/FVP_install_location`, then the simulation can be started by +using: ```commandline ~/FVP_install_location/models/Linux64_GCC-6.4/FVP_Corstone_SSE-300_Ethos-U55 ./bin/mps3-sse-300/ethos-u-kws.axf ``` -A log output should appear on the terminal: +A log output appears on the terminal: ```log telnetterminal0: Listening for serial connection on port 5000 @@ -300,12 +326,15 @@ telnetterminal2: Listening for serial connection on port 5002 telnetterminal5: Listening for serial connection on port 5003 ``` -This will also launch a telnet window with the sample application's standard output and error log entries containing -information about the pre-built application version, TensorFlow Lite Micro library version used, data type as well as -the input and output tensor sizes of the model compiled into the executable binary. +This also launches a telnet window with the standard output of the sample application. It also includes error log +entries containing information about the pre-built application version, TensorFlow Lite Micro library version used, and +data types. The log also includes the input and output tensor sizes of the model compiled into the executable binary. -After the application has started if `kws_FILE_PATH` pointed to a single file (or a folder containing a single input file) -the inference starts immediately. In case of multiple inputs choice, it outputs a menu and waits for the user input from telnet terminal: +After the application has started, if `kws_FILE_PATH` points to a single file, or even a folder that contains a single +input file, then the inference starts immediately. If there are multiple inputs, it outputs a menu and then waits for +input from the user. + +For example: ```log User input required @@ -321,49 +350,46 @@ Choice: ``` -1. “Classify next audio clip” menu option will run inference on the next in line voice clip from the collection of the - compiled audio. +What the preceding choices do: - > **Note:** Note that if the clip is over a certain length, the application will invoke multiple inference runs to cover the entire file. +1. Classify next audio clip: Runs a single inference on the next in line. -2. “Classify audio clip at chosen index” menu option will run inference on the chosen audio clip. +2. Classify audio clip at chosen index: Runs inference on the chosen audio clip. - > **Note:** Please make sure to select audio clip index in the range of supplied audio clips during application build. - By default, pre-built application has 4 files, indexes from 0 to 3. + > **Note:** Please make sure to select audio clip index within the range of supplied audio clips during application + > build. By default, a pre-built application has four files, with indexes from `0` to `3`. -3. “Run classification on all audio clips” menu option triggers sequential inference executions on all built-in voice - samples. +3. Run ... on all: Triggers sequential inference executions on all built-in applications. -4. “Show NN model info” menu option prints information about model data type, input and output tensor sizes: +4. Show NN model info: Prints information about the model data type, input, and output, tensor sizes: ```log INFO - uTFL version: 2.5.0 INFO - Model info: INFO - Model INPUT tensors: - INFO - tensor type is INT8 - INFO - tensor occupies 490 bytes with dimensions - INFO - 0: 1 - INFO - 1: 1 - INFO - 2: 49 - INFO - 3: 10 + INFO - tensor type is INT8 + INFO - tensor occupies 490 bytes with dimensions + INFO - 0: 1 + INFO - 1: 1 + INFO - 2: 49 + INFO - 3: 10 INFO - Quant dimension: 0 INFO - Scale[0] = 1.107164 INFO - ZeroPoint[0] = 95 INFO - Model OUTPUT tensors: - INFO - tensor type is INT8 - INFO - tensor occupies 12 bytes with dimensions - INFO - 0: 1 - INFO - 1: 12 + INFO - tensor type is INT8 + INFO - tensor occupies 12 bytes with dimensions + INFO - 0: 1 + INFO - 1: 12 INFO - Quant dimension: 0 INFO - Scale[0] = 0.003906 INFO - ZeroPoint[0] = -128 INFO - Activation buffer (a.k.a tensor arena) size used: 72848 INFO - Number of operators: 1 - INFO - Operator 0: ethos-u + INFO - Operator 0: ethos-u ``` -5. “List audio clips” menu option prints a list of pair audio indexes - the original filenames embedded in the - application: +5. List audio clips: Prints a list of pair ... indexes. The original filenames are embedded in the application, like so: ```log [INFO] List of Files: @@ -375,9 +401,9 @@ Choice: ### Running Keyword Spotting -Selecting the first option will run inference on the first file. +Please select the first menu option to execute inference on the first file. -The following example illustrates application output for classification: +The following example illustrates the output for classification: ```logINFO - Running inference on audio clip 0 => down.wav INFO - Inference 1/1 @@ -393,26 +419,28 @@ INFO - NPU IDLE cycles: 561 INFO - NPU total cycles: 681172 ``` -Each inference should take less than 30 seconds on most systems running Fast Model. +On most systems running Fast Model, each inference takes under 30 seconds. + The profiling section of the log shows that for this inference: -- Ethos-U55's PMU report: +- *Ethos-U55* PMU report: - - 681,172 total cycle: The number of NPU cycles + - 681,172 total cycle: The number of NPU cycles. - - 680,611 active cycles: The number of NPU cycles that were used for computation + - 680,611 active cycles: The number of NPU cycles that were used for computation. - - 561 idle cycles: number of cycles for which the NPU was idle + - 561 idle cycles: The number of cycles for which the NPU was idle. - - 217,385 AXI0 read beats: The number of AXI beats with read transactions from AXI0 bus. - AXI0 is the bus where Ethos-U55 NPU reads and writes to the computation buffers (activation buf/tensor arenas). + - 217,385 AXI0 read beats: The number of AXI beats with read transactions from the AXI0 bus. AXI0 is the bus where the + *Ethos-U55* NPU reads and writes to the computation buffers, activation buf, or tensor arenas. - 82,607 write cycles: The number of AXI beats with write transactions to AXI0 bus. - - 59,608 AXI1 read beats: The number of AXI beats with read transactions from AXI1 bus. - AXI1 is the bus where Ethos-U55 NPU reads the model (read only) + - 59,608 AXI1 read beats: The number of AXI beats with read transactions from the AXI1 bus. AXI1 is the bus where the + *Ethos-U55* NPU reads the model. So, read-only. -- For FPGA platforms, CPU cycle count can also be enabled. For FVP, however, CPU cycle counters should not be used as - the CPU model is not cycle-approximate or cycle-accurate. +- For FPGA platforms, a CPU cycle count can also be enabled. However, do not use cycle counters for FVP, as the CPU + model is not cycle-approximate or cycle-accurate. -The application prints the highest confidence score and the associated label from ds_cnn_labels.txt file. +> **Note:** The application prints the highest confidence score and the associated label from the `ds_cnn_labels.txt` +> file. |