motion-coremltools

motion-coremltools (coremotiontools) is the wrapper tool for converting neural networks trained with motion sensor data.

Usage

The usage is the same as coremltools' Unified Conversation API. Currently, only tensorflow.keras.Model is supported. Also, only 1-demensional CNNs are supported.

import coremltools as ct
import tensorflow as tf
from tensorflow.keras.layers import Input, Conv1D, MaxPooling1D, Flatten, Dense
from tensorflow.keras.models import Model

from coremotiontools import convert

# Build model
inputs = Input(shape=(256*3, 1))
x = Conv1D(16, kernel_size=3, strides=1, padding='same', activation='relu', kernel_initalizer='he_normal')(inputs)
x = MaxPooling1D(pool_size=2, padding='same')(x)

x = Flatten()(x)
x = Dense(1024, activation='relu', kernel_initalizer='he_normal')(x)
outputs = Dense(6, activation='softmax')(x)

model = Model(inputs=inputs, outputs=outputs)

# Convert to Core ML
classifier_config = ct.ClassifierConfig(class_labels=["stay", "walk", "jog", "skip", "stUp", "stDown"])
mlmodel = convert(model, classifier_config=classifier_config)
mlmodel.save("ActivityClassifier.mlmodel")

See here for more detailed usages.

In Core ML

When the converted model is used in Core ML, the input type is MLMultiArray.

let input: [Double] = [0.1, 0.2, 0.3, ...] // Order of x-axis, y-axis, z-axis, x, y, z, ...
let mlArray = try! MLMultiArray.fromDouble(input) // MLMultiArray.fromDouble() is extension

// Predict
let model = ActivityClassifier()
let output = try! model.prediction(input: ActivityClassifierInput(input: mlArray))

See here for more detailed usages.

Install

Supported pip install

pip install git+https://github.com/Shakshi3104/motion-coremltools.git

Requirements

• coremltools 4.1
• tensorflow >=2.1