MobileNet for Microcontrollers

MobileNets are a family of neural networks architectures aimed at image classification. Their known to be small and fast yet still achieving a good classification accuracy with respect to more sophisticated architectures.

This document introduces MicroMobileNet, a Python package that gives you the possibility to train MobileNets in Python and then export them to plain C++, for use with Arduino, PlatformIO or ESP-IDF frameworks.

Install the dependencies

MicroMobileNet is based on Keras and Tensorflow.

pip install keras tensorflow Jinja2 sklearn Pillow cached_property
pip install micromobilenet

1 2

Prepare data

MicroMobileNet only works on grayscale, 96x96 images. If you have color images, you have to convert them to grayscale. If you have smaller/larger images, you have to resize/crop them.

The code sample below shows how to generate the training dataset assuming you have a folder structure that looks like the one depicted here.

DATA
|-- train
|---- apple
|------ ...list of images...
|---- orange
|------ ...list of images...
|---- banana
|------ ...list of images...
|-- validation
|---- apple
|------ ...list of images...
|---- orange
|------ ...list of images...
|---- banana
|------ ...list of images...
|-- test
|---- apple
|------ ...list of images...
|---- orange
|------ ...list of images...
|---- banana
|------ ...list of images...

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22

"""
Load images from folders
"""

import os
import numpy as np
from os import listdir
from glob import glob
from PIL import Image


def load_folder(folder: str):
    """
    Load images from folder as [0, 1] floats
    :param folder:
    :return:
    """
    for filename in sorted(glob(f"{folder}/*.jpg") + glob(f"{folder}/*.jpeg")):
        yield np.asarray(Image.open(filename).convert("L"), dtype=float) / 255.


def load_split(root: str, split_name: str):
    """
    Load images from train/val/test folder
    :param root:
    :param split_name:
    :return:
    """
    X = []
    Y = []
    folders = listdir(f"{root}/{split_name}")
    folders = [f"{root}/{split_name}/{f}" for f in folders if os.path.isdir(f"{root}/{split_name}/{f}")]

    for k, folder in enumerate(sorted(folders)):
        folder_x = list(load_folder(folder))
        X += folder_x
        Y += [k] * len(folder_x)

    # shuffle inputs
    shuffle_mask = np.random.permutation(len(X))
    X = np.asarray(X)[shuffle_mask]
    Y = np.asarray(Y)[shuffle_mask]

    return X, Y


if __name__ == '__main__':
    train_x, train_y = load_split("DATA", "train")
    val_x, val_y = load_split("DATA", "val")
    test_x, test_y = load_split("DATA", "test")
    print(train_x.shape)
    print(val_x.shape)
    print(test_x.shape)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53

Instantiate and Train

First of all, you start by instantiating the MobileNet variation you prefer (refer to the section Architectures for the list of available configurations and their size).

from micromobilenet import PicoMobileNet

# replace num_classes with the actual number of classes
net = PicoMobileNet(num_classes=10)

1 2 3 4

Then, you configure all the parameters required for training.

net.config.learning_rate = 0.01
net.config.batch_size = 32
net.config.verbosity = 1
net.config.loss = "categorical_crossentropy"
net.config.metrics = ["categorical_accuracy"]
net.config.checkpoint_path = "./checkpoints/pico"

1 2 3 4 5 6

Finally, it is time to train the network on your data.

net.build()
net.compile()
net.fit(train_x, train_y, val_x, val_y, epochs=30)

1 2 3

To evaluate the accuracy on the test set:

from sklearn.metrics import classification_report

predictions = net.predict(test_x)
print(classification_report(test_y.argmax(axis=1), predictions.argmax(axis=1)))

1 2 3 4

Convert to C++

After you have your model trained and tested, it is time to export it to C++.

with open("MobileNet.h", "w") as file:
    file.write(net.convert.to_cpp())

1 2

Generated code will look like the following.

/**
 * "Compiled" implementation of modified MobileNet
 */
class PicoMobileNet {
public:
    const uint16_t numInputs = 9216;
    const uint16_t numOutputs = 4;
    float outputs[4];
    float arena[6936];
    uint16_t output;
    float proba;

    /**
     *
     */
    MobileNet() : output(0), proba(0) {
        for (uint16_t i = 0; i < numOutputs; i++)
            outputs[i] = 0;
    }

    /**
     *
     * @param input
     */
    uint16_t predict(float *input) {
        float *ping = arena;
        float *pong = arena + 3468;

        // conv2d (0)
        for (int16_t d = 0; d < 3; d++)
            this->conv2d_3x3x1(input, ping + 32 * 32 * d, conv2d_0_weights[d], 96, 3);

        // padding (1)
        for (int16_t d = 0; d < 3; d++)
            this->pad(ping + 32 * 32 * d, pong + 34 * 34 * d, 32);

        memcpy(ping, pong, sizeof(float) * 34 * 34 * 3);

        // depthwise (1)
        for (int16_t d = 0; d < 3; d++)
            this->depthwise_conv(ping + 34 * 34 * d, pong + 16 * 16 * d, depthwise_1_weights[d], 34, 2);

        // pointwise (1)
        for (int16_t d = 0; d < 6; d++)
            this->pointwise_conv(pong, ping + 16 * 16 * d, pointwise_1_weights[d], 16, 3);

        // padding (2)
        for (int16_t d = 0; d < 6; d++)
            this->pad(ping + 16 * 16 * d, pong + 18 * 18 * d, 16);

        memcpy(ping, pong, sizeof(float) * 18 * 18 * 6);

        // depthwise (2)
        for (int16_t d = 0; d < 6; d++)
            this->depthwise_conv(ping + 18 * 18 * d, pong + 8 * 8 * d, depthwise_2_weights[d], 18, 2);

        // pointwise (2)
        for (int16_t d = 0; d < 12; d++)
            this->pointwise_conv(pong, ping + 8 * 8 * d, pointwise_2_weights[d], 8, 6);


        // padding (3)
        for (int16_t d = 0; d < 12; d++)
            this->pad(ping + 8 * 8 * d, pong + 10 * 10 * d, 8);

        memcpy(ping, pong, sizeof(float) * 10 * 10 * 12);


        // depthwise (3)
        for (int16_t d = 0; d < 12; d++)
            this->depthwise_conv(ping + 10 * 10 * d, pong + 4 * 4 * d, depthwise_3_weights[d], 10, 2);

        // pointwise (3)
        for (int16_t d = 0; d < 24; d++)
            this->pointwise_conv(pong, ping + 4 * 4 * d, pointwise_3_weights[d], 4, 12);

        this->maxpool(ping, pong, 4, 24);

        for (uint16_t d = 0; d < numOutputs; d++)
            this->dot(pong, ping + d, conv2d_last_weights[d], conv2d_last_bias[d], 24);

        this->softmax(ping, outputs, numOutputs);

        return this->argmax();
    }

    /**
 * Get index of max output
 */
    uint16_t argmax() {
        this->output = 0;
        this->proba = outputs[0];

        for (uint16_t i = 1; i < numOutputs; i++) {
            if (outputs[i] > this->proba) {
                this->proba = outputs[i];
                this->output = i;
            }
        }

        return this->output;
    }

protected:
    const float conv2d_0_weights[3][9] = {{-0.41666251421, -0.11826507002, -0.23049902916, 0.17382600904, -0.55161094666, 0.10363399237, -0.42671826482, -0.12638387084, -0.52779608965},
                                          {0.31084173918, -0.04917881638, 0.40228804946, -0.18794929981, -0.02449743077, 0.03094954416, 0.44947320223, 0.16917343438, -0.30685693026},
                                          {-0.08407333493, 0.57998371124, 0.34822252393, 0.56157398224, 0.24396187067, 0.32406872511, 0.18443229795, 0.53056180477, 0.13178956509}};

    const float depthwise_1_weights[3][9] = {{-0.35425487161, -0.23616252840, -0.22997750342, 0.24456115067, -0.20624010265, -0.00013742072, 0.26879012585, -0.26807692647, 0.01581084728},
                                             {0.07073932886, 0.48199057579, -0.10555473715, -0.19673484564, -0.22651126981, -0.01001045667, 0.37793800235, -0.19346579909, -0.32536745071},
                                             {-0.28408280015, -0.47497844696, -0.11035950482, 0.36142480373, 0.21333804727, 0.31504249573, 0.41812920570, -0.31300476193, 0.04843502119}};
    const float pointwise_1_weights[6][3] = {{-0.09117162228, 0.82249736786, -0.20695370436},
                                             {-0.85948276520, 0.19161477685, 1.10251903534},
                                             {-0.71232938766, -0.03556078672, -0.08283454180},
                                             {0.46851244569, 0.76725500822, 0.50938290358},
                                             {-0.25427860022, -0.72349339724, -0.79621297121},
                                             {-0.47162520885, 0.25396701694, -0.49285510182}};

    const float depthwise_2_weights[6][9] = {{0.35895720124, 0.23784333467, -0.07534114271, -0.24524940550, -0.04549394548, -0.26059886813, 0.28477019072, -0.00712140510, -0.27584657073},
                                             {0.51091092825, 0.62111812830, -0.05044540763, 0.52927201986, 0.25487384200, 0.33370712399, 0.18966604769, 0.11532534659, 0.16275504231},
                                             {0.29592916369, -0.15394289792, 0.12004908919, 0.07607766986, 0.24132254720, -0.20747616887, -0.04176491499, -0.27908378839, 0.13809236884},
                                             {0.02603448555, 0.25732424855, -0.36077418923, 0.07963499427, -0.01887336187, -0.33278352022, 0.24708244205, -0.19360905886, 0.13639056683},
                                             {0.21229699254, 0.00662785769, -0.19093543291, -0.09667769074, -0.05552545190, 0.01491189003, -0.20547568798, 0.27846303582, -0.03822237253},
                                             {-0.06207026541, 0.15863336623, -0.16388191283, 0.15762025118, -0.19919943810, -0.07231600583, -0.13571833074, -0.02249273658, -0.21008148789}};
    const float pointwise_2_weights[12][6] = {{0.00063503534, 0.48058623075, -0.36563238502, 0.51302105188, 0.21660172939, -0.32831248641},
                                              {-0.50620383024, -0.35271933675, 0.48594498634, -0.50641745329, 0.35882031918, 0.22168761492},
                                              {-0.43700292706, -0.38995116949, -0.53454947472, -0.06846392155, -0.13144075871, 0.07794147730},
                                              {0.37599769235, -0.43717575073, 0.38039886951, -0.55730378628, 0.38651198149, -0.11370623112},
                                              {-0.02578501403, -0.49355089664, 0.18120104074, 0.10755521804, -0.52842414379, 0.23777391016},
                                              {0.32847705483, -0.22344671190, 0.04166835546, 0.46865811944, 0.56128060818, -0.33481889963},
                                              {-0.31596237421, -0.04046857357, 0.37702673674, -0.46384748816, 0.48273956776, -0.26779732108},
                                              {0.49115952849, -0.01721261069, 0.42422366142, -0.06754394621, 0.39422857761, 0.13803933561},
                                              {-0.10960361362, 0.57471114397, -0.27886447310, 0.14179545641, -0.08440622687, -0.20258137584},
                                              {0.35399246216, 0.33130452037, 0.31853711605, 0.43343657255, -0.34315919876, 0.11230981350},
                                              {0.31758838892, 0.80079740286, -0.42914772034, 0.01409940235, -0.23797056079, -0.39399114251},
                                              {-0.22277000546, -0.39749902487, 0.18530112505, -0.02303826809, -0.17399805784, -0.52532655001}};

    const float depthwise_3_weights[12][9] = {{0.43468713760, 0.27347934246, 0.26273813844, -0.04167792201, 0.16433015466, -0.12557785213, 0.50365906954, 0.19630661607, 0.48740977049},
                                              {0.04266124964, -0.16427403688, 0.08567957580, 0.17073456943, 0.21095471084, -0.03015229106, -0.01741993427, -0.11063526571, 0.06490472704},
                                              {0.18609096110, 0.04467250407, -0.16391515732, -0.18312613666, 0.01121036988, 0.10224457085, 0.13521514833, -0.10507968068, 0.08436971158},
                                              {0.15235303342, -0.04396350682, 0.13825593889, -0.05687613785, 0.11330153048, -0.20187327266, -0.00074075162, 0.04703579471, -0.15467861295},
                                              {0.00457271701, 0.24002927542, 0.04937255010, 0.10910945386, 0.19974878430, 0.10104069859, 0.06970567256, 0.12173694372, -0.26547023654},
                                              {-0.07644123584, -0.21463918686, 0.02911583148, -0.18811574578, 0.09112719446, 0.11265213788, -0.01146453898, -0.00454099616, -0.14186237752},
                                              {0.06569774449, -0.02157643437, 0.19389937818, 0.19328580797, 0.14999692142, 0.15664963424, -0.02148352563, 0.22635035217, -0.17734022439},
                                              {0.10683796555, 0.05706703290, 0.21315959096, -0.05024305359, 0.19207747281, 0.00748820370, 0.01584845968, 0.21641007066, 0.21460226178},
                                              {-0.19390374422, -0.14212438464, -0.23052616417, -0.07321915030, 0.27654990554, 0.06163945794, -0.16758939624, 0.33695048094, 0.07404905558},
                                              {-0.22942769527, -0.17088742554, 0.02193632536, 0.35620847344, 0.20238873363, 0.31436139345, -0.34888097644, -0.39686203003, -0.25719842315},
                                              {0.32601067424, 0.25143274665, 0.26007786393, 0.03177319467, 0.19906532764, -0.17288939655, 0.37424448133, 0.12854650617, 0.31418201327},
                                              {-0.15625628829, 0.01465466619, 0.05003134906, 0.09593714774, 0.12668652833, -0.22036601603, 0.09534128010, -0.10707643628, 0.18672065437}};
    const float pointwise_3_weights[24][12] = {{-0.33242326975, 0.32112824917, -0.34323650599, -0.19911736250, 0.16705293953, 0.01022295561, -0.30409169197, 0.01617346518, -0.12629058957, 0.02393108793, 0.16568218172, 0.10778385401},
                                               {0.45878762007, -0.34416934848, -0.18845045567, -0.29416739941, 0.03930538893, 0.00911496207, -0.26870962977, 0.33837494254, -0.22767841816, -0.06902970374, 0.72774738073, -0.07464709878},
                                               {-0.15224465728, 0.25462841988, 0.16669172049, -0.11165502667, -0.40522453189, -0.29562169313, 0.07739520073, -0.31246533990, -0.39020511508, -0.39069032669, -0.28850054741, 0.08244985342},
                                               {0.36490887403, -0.08350038528, 0.36102640629, -0.15604324639, -0.29153943062, 0.00717086066, -0.10952049494, -0.08671616763, -0.32154598832, 0.00325861014, 0.62585133314, -0.19151093066},
                                               {0.33634936810, 0.09808677435, 0.00602692366, 0.03131084517, 0.28147563338, 0.38172629476, -0.10785871744, 0.14068019390, -0.25693386793, -0.10714749992, 0.56367319822, 0.39001709223},
                                               {0.61772930622, 0.13471519947, 0.04854518548, -0.00608583074, -0.01802019961, -0.37180465460, -0.00401133299, 0.33786326647, -0.12424690276, -0.28323772550, 0.39886355400, -0.29021489620},
                                               {0.04374134541, -0.37529510260, 0.06371120363, 0.04935376719, 0.02217261866, 0.30530115962, -0.23669055104, 0.06503948569, 0.01679839753, 0.48886525631, -0.27314275503, -0.34578859806},
                                               {-0.01946349815, -0.22470764816, 0.09743384272, -0.02405748889, 0.29618066549, -0.25876462460, -0.34475395083, 0.21627385914, -0.10402329266, 0.19197782874, -0.42406490445, 0.25275290012},
                                               {0.63850814104, -0.23805379868, 0.27066165209, 0.25643539429, 0.37469926476, -0.11610201001, -0.29154029489, -0.02091231011, -0.19745121896, -0.16873377562, 0.24683740735, 0.06828689575},
                                               {-0.13936965168, -0.23779028654, -0.39054390788, -0.16376510262, -0.08882397413, 0.41016262770, 0.18596971035, -0.25503417850, 0.36027094722, -0.28058910370, -0.23572920263, -0.37204989791},
                                               {0.40478792787, 0.18747282028, 0.32068753242, 0.29423406720, 0.36582773924, 0.20912985504, -0.40184441209, 0.28225576878, -0.12299652398, -0.08685125411, 0.47822484374, -0.04130911827},
                                               {0.03607089445, 0.18253159523, 0.33984237909, 0.10524179041, -0.24410746992, 0.09393396229, -0.19550697505, -0.38074532151, 0.29552531242, 0.45382964611, -0.38232478499, 0.19928234816},
                                               {-0.36115184426, -0.03652039170, 0.13133662939, 0.12104801089, 0.02864059806, 0.13996993005, 0.27846491337, -0.33686831594, 0.00807717629, 0.44147452712, 0.10711596906, 0.24524736404},
                                               {-0.47024169564, 0.03156501055, -0.07844874263, -0.39789190888, 0.19768652320, 0.18511816859, 0.37028717995, 0.14206841588, 0.42253586650, 0.48465278745, 0.15217860043, 0.29450023174},
                                               {0.25448688865, 0.04511585832, 0.28480070829, -0.01936831698, 0.14015029371, 0.14111196995, -0.12783417106, 0.19556348026, 0.01343316026, -0.15595127642, 0.42841988802, 0.18192327023},
                                               {-0.14106854796, 0.30786058307, 0.01031529903, 0.26337873936, 0.15846376121, 0.28183579445, 0.10774451494, -0.32656732202, 0.35945361853, -0.24068254232, -0.35250753164, 0.21529477835},
                                               {0.61793971062, -0.02668470144, 0.32367497683, -0.38035601377, -0.07861428708, 0.39642265439, -0.29068851471, 0.15355134010, 0.11719731987, -0.24253894389, 0.65259689093, -0.14480277896},
                                               {0.32112109661, -0.17774070799, 0.35220271349, -0.19650655985, 0.04403254017, -0.16583316028, 0.01203846931, 0.21680326760, 0.06325855851, -0.06204054505, -0.15532730520, -0.16143210232},
                                               {0.18882560730, -0.23018650711, 0.23278686404, -0.09548929334, 0.36104604602, -0.22795256972, -0.32385107875, -0.24018083513, -0.30034396052, -0.00216405559, -0.14133879542, 0.02028653026},
                                               {0.37537524104, -0.29309490323, -0.00938728452, -0.29328131676, 0.34983760118, -0.09490379691, 0.35412961245, -0.01121471357, 0.32010972500, -0.30494660139, 0.42134764791, -0.16008812189},
                                               {-0.64491266012, 0.26568508148, -0.12225946784, 0.18564531207, 0.28982475400, -0.35740262270, 0.25998938084, 0.23804895580, 0.10513759404, 0.42896071076, 0.20906865597, -0.34997045994},
                                               {-0.20663002133, -0.14476191998, 0.06242525578, -0.25408589840, -0.16947199404, -0.19733618200, -0.00828516483, -0.14818486571, -0.18266969919, -0.00224149227, -0.38471829891, 0.28299516439},
                                               {0.13125912845, -0.39106506109, -0.03080397844, 0.24858620763, 0.02611339465, 0.11944877356, -0.20792202652, -0.11530396342, 0.22887417674, -0.09155285358, -0.18884035945, 0.32079821825},
                                               {0.15044319630, -0.14401194453, -0.18125914037, -0.28057807684, 0.08085042238, -0.18810100853, -0.27874565125, -0.07419086248, 0.22713482380, -0.21203750372, 0.15332724154, 0.09756356478}};

    const float conv2d_last_weights[4][24] = {{0.18572680652, 0.10660002381, 0.34570229053, -0.13073261082, 0.04635358602, 0.40537646413, -0.42621853948, -0.19123505056, -0.13323357701, -0.08411549777, -0.28852567077, -0.40656703711, -0.04649058729, -0.26070696115, 0.37145623565, -0.20005775988, 0.16340047121, -0.51885360479, -0.43397870660, -0.09329127520, -0.35178303719, -0.18354952335, -0.06395643204, -0.22809614241},
                                              {0.31888130307, 0.52352815866, 0.04355543852, 0.25176772475, 0.62861680984, 0.51203912497, -0.19495809078, -0.17149986327, 0.43246814609, 0.20671278238, 0.51400816441, 0.10310800374, -0.10653696954, -0.14129965007, -0.20243576169, -0.25948125124, 0.46354496479, 0.20531697571, -0.29653131962, 0.35552129149, -0.37847694755, -0.20610806346, -0.23213934898, 0.28872734308},
                                              {-0.10687508434, -0.63240849972, 0.11593413353, -0.77483934164, -0.24154321849, -0.78893947601, 0.58301407099, 0.12191224843, -0.38278028369, -0.57541137934, -0.58911901712, 0.58145040274, 0.60270369053, 0.19607372582, -0.78608351946, -0.53846079111, -0.15513655543, -0.04214842990, -0.14708326757, -0.82148188353, 0.37131208181, 0.31574213505, -0.44903215766, -0.35200369358},
                                              {-0.44866114855, 0.28538855910, -0.26954653859, 0.23931820691, -0.21181413531, -0.25706809759, -0.52325546741, 0.09330182523, 0.05173591897, 0.17546258867, 0.26352429390, -0.04025872424, -0.25341644883, -0.26888474822, 0.21390053630, -0.05746645853, 0.48053106666, 0.40644150972, 0.51892417669, 0.22470143437, -0.40383216739, 0.06014847755, 0.10987602919, 0.12020519376}};
    const float conv2d_last_bias[4] = {0.03802097589, -0.09491597861, 0.07504304498, -0.01769104972};

    /**
 * Multiply 3x3 kernel on single 3x3 image patch
 *
 * @param inputs
 * @param kernel
 * @param width
 */
    inline float mult3x3(float *inputs, const float kernel[9], const uint16_t width) {
        const float *i1 = inputs;
        const float *i2 = inputs + width;
        const float *i3 = inputs + width + width;

        return i1[0] * kernel[0] +
               i1[1] * kernel[1] +
               i1[2] * kernel[2] +
               i2[0] * kernel[3] +
               i2[1] * kernel[4] +
               i2[2] * kernel[5] +
               i3[0] * kernel[6] +
               i3[1] * kernel[7] +
               i3[2] * kernel[8];
    }
    /**
 * Zero padding 2D
 *
 * @param inputs
 * @param outputs
 * @param width
 */
    void pad(float *inputs, float *outputs, uint16_t width) {
        const uint16_t paddedWidth = width + 2;
        uint16_t i = 0;
        uint16_t o = 0;

        // first row of zeros
        for (uint16_t x = 0; x < paddedWidth; x++)
            outputs[o++] = 0;

        for (uint16_t y = 0; y < width; y++) {
            outputs[o++] = 0;

            for (uint16_t x = 0; x < width; x++)
                outputs[o++] = inputs[i++];

            outputs[o++] = 0;
        }

        // last row of zeros
        for (uint16_t x = 0; x < paddedWidth; x++)
            outputs[o++] = 0;
    }
    /**
 * Depthwise 3x3 convolution without ReLU
 *
 * @param input
 * @param output
 * @param kernel
 * @param width
 * @param stride
 */
    void conv2d_3x3x1(float *input, float *output, const float *kernel, const uint16_t width, uint8_t stride) {
        uint16_t o = 0;

        for (uint16_t y = 0; y <= width - 3; y += stride) {
            const uint16_t offset = y * width;
            float *i = input + offset;

            for (uint16_t x = 0; x <= width - 3; x += stride) {
                output[o++] = this->mult3x3(i + x, kernel, width);
            }
        }
    }
    /**
 * Depthwise 3x3 convolution with ReLU
 *
 * @param inputs
 * @param outputs
 * @param kernel
 * @param width
 * @param stride
 */
    void depthwise_conv(float *inputs, float *outputs, const float *kernel, const uint16_t width, uint8_t stride) {
        uint16_t o = 0;

        for (uint16_t y = 0; y <= width - 3; y += stride) {
            const uint16_t offset = y * width;
            float *i = inputs + offset;

            for (uint16_t x = 0; x <= width - 3; x += stride) {
                float val = this->mult3x3(i + x, kernel, width);

                if (val < 0) val = 0;
                else if (val > 6) val = 6;

                outputs[o++] = val;
            }
        }
    }
    /**
 * Pointwise 1x1 convolution with ReLU
 *
 * @param inputs
 * @param outputs
 * @param kernel
 * @param width
 * @param channels
 */
    void pointwise_conv(float *inputs, float *outputs, const float *kernel, const uint16_t width, const uint16_t channels) {
        const uint16_t size = width * width;
        uint16_t o = 0;

        for (uint16_t y = 0; y < width; y += 1) {
            const uint16_t offset = y * width;
            for (uint16_t x = 0; x < width; x += 1) {
                float val = 0;

                for (uint16_t c = 0; c < channels; c++)
                    val += inputs[(offset + x) + size * c] * kernel[c];

                if (val < 0) val = 0;
                else if (val > 6) val = 6;

                outputs[o++] = val;
            }
        }
    }
    /**
 * (Global) MaxPooling
 *
 * @param inputs
 * @param outputs
 * @param width
 * @param channels
 */
    void maxpool(float *inputs, float *outputs, const uint16_t width, const uint16_t channels) {
        const uint16_t size = width * width;

        for (uint16_t c = 0; c < channels; c++) {
            const uint16_t offset = size * c;
            float *in = inputs + offset;
            float greatest = in[0];

            for (uint16_t j = 1; j < size; j++)
                if (in[j] > greatest)
                    greatest = in[j];

            outputs[c] = greatest;
        }
    }
    /**
 * Dot product with ReLU
 * @param inputs
 * @param outputs
 * @param kernel
 * @param bias
 * @param length
 */
    void dot(float *inputs, float *outputs, const float *weights, const float bias, const uint16_t length) {
        float sum = 0;

        for (uint16_t i = 0; i < length; i++)
            sum += inputs[i] * weights[i];

        outputs[0] = sum + bias;
    }
    /**
 * Softmax activation
 *
 * @param inputs
 * @param outputs
 * @param numOutputs
 */
    void softmax(float *inputs, float *outputs, uint16_t numOutputs) {
        float sum = 0;

        for (uint16_t i = 0; i < numOutputs; i++) {
            const float e = exp(inputs[i]);
            outputs[i] = e;
            sum += e;
        }

        for (uint16_t i = 0; i < numOutputs; i++)
            outputs[i] /= sum;
    }
};

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365

This is a fully self-contained, statically allocated class that implements the MobileNet variation of choice. It doesn't required external runtimes to run, doesn't require a TENSOR_ARENA_SIZE to be defined beforehand, doesn't throw cryptic errors during compilation nor execution.

It is written as plain C++ and doesn't contain any vendor-specific optimizations at the moment (e.g. CMSIS for ARM Cortex chipsets). They will be added in future versions, if demand supports the effort.

Use with Arduino

To use the network with Arduino (or any other framework, actually), you need to just include the generated class and call predict() on the instance.

// sample image is a float[96 * 96] array
#include "sample_image.h"
#include "MobileNet.h"

MobileNet net;

void setup() {
  Serial.begin(115200);
  Serial.println("MobileNet demo");
  
  // no complicated setup! 
}

void loop() {
  size_t start = micros();
  net.predict(sample_image);

  Serial.print("Predicted output = ");
  Serial.println(net.output);
  Serial.print("It took ");
  Serial.print(micros() - start);
  Serial.println(" us to run MobileNet");
  delay(2000);
}

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Benchmarks

How small are these MobileNet variations? How fast do they run? Making an extensive test would take a lot of time, so I run most of the experiments on an ESP32S3 board and a couple more on an Arduino Nano 33 BLE Sense (ARM Cortex M4) and Arduino Portenta H7 (ARM Cortex M7).

Benchmarks for ESP32S3
┏━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Architecture       ┃ Program space (kb) ┃  RAM (kb)  ┃ Execution time (us) ┃
┡━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩━━━━━━━━━━━━━━━━━━━━━┩
│ Pico               │              4,52  │    30,55   │        2832         |
├────────────────────┼────────────────────┼────────────┤─────────────────────┤
│ Nano               │              8,54  │    64,23   │        6543         |
├────────────────────┼────────────────────┼────────────┤─────────────────────┤
│ Micro              │             19,75  │   132,36   │       31987         |
├────────────────────┼────────────────────┼────────────┤─────────────────────┤
│ Milli              │             49,70  │   162,12   │       37641         |
├────────────────────┼────────────────────┼────────────┤─────────────────────┤
│ Base               │            123,60  │   235,47   │       53944         |
└────────────────────┴────────────────────┴────────────┴─────────────────────┘

1 2 3 4 5 6 7 8 9 10 11 12 13 14

Benchmarks for Arduino Nano 33 BLE Sense (Cortex M4)
┏━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Architecture       ┃ Program space (kb) ┃  RAM (kb)  ┃ Execution time (us) ┃
┡━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩━━━━━━━━━━━━━━━━━━━━━┩
│ Pico               │              5,42  │    31,17   │       10969         |
├────────────────────┼────────────────────┼────────────┤─────────────────────┤
│ Nano               │              9,27  │    64,22   │       25169         |
├────────────────────┼────────────────────┼────────────┤─────────────────────┤
│ Micro              │             19,75  │   132,36   │      123669         |
└────────────────────┴────────────────────┴────────────┴─────────────────────┘

1 2 3 4 5 6 7 8 9 10

Benchmarks for Portenta H7 (Cortex M7)
┏━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┓
┃ Architecture       ┃  Execution time (us) ┃
┡━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━┩
│ Nano               │         3000         |
├────────────────────┼──────────────────────┤
│ Base               │         25357        │
└────────────────────┴──────────────────────┘

1 2 3 4 5 6 7 8

Regarding the execution speed, you can see that for the widely available ESP32S3 chip, you can get the middle Micro network to run in 32ms. This should allow realtime camera frames processing in most of vision-based projects.

The Arduino Portenta H7 runs roughly twice as fast as the ESP32S3. The Arduino Nano BLE Sense is much slower at every scale.

Accuracy

Ok, these variations of MobileNet are small and fast. But are they accurate enough? It depends on you dataset. These networks are meant to be trained from scratch on your data and the results may vary. I was able to achieve >90% accuracy on a 10 textures classification project and 75% accuracy on a subset of Fashion MNIST training PicoMobileNet for 20 epochs, so good results are totally achievable.

If you find that on your dataset you can't get satisfactory results, then you should explore alternatives (e.g. EdgeImpulse, that uses the full version of MobileNetV1 with transfer learning).

Entire Python script

import os.path
from collections import Counter

import numpy as np
from os import listdir
from glob import glob
from PIL import Image
from keras.utils import to_categorical
from sklearn.metrics import classification_report
from micromobilenet import PicoMobileNet


def load_folder(folder: str):
    """
    Load images from folder as [0, 1] floats
    :param folder:
    :return:
    """
    for filename in sorted(glob(f"{folder}/*.jpg") + glob(f"{folder}/*.jpeg")):
        yield np.asarray(Image.open(filename).convert("L"), dtype=float) / 255.


def load_split(root: str, split_name: str):
    """
    Load images from train/val/test folder
    :param root:
    :param split_name:
    :return:
    """
    X = []
    Y = []
    folders = listdir(f"{root}/{split_name}")
    folders = [f"{root}/{split_name}/{f}" for f in folders if os.path.isdir(f"{root}/{split_name}/{f}")]

    for k, folder in enumerate(sorted(folders)):
        folder_x = list(load_folder(folder))
        X += folder_x
        Y += [k] * len(folder_x)

    # shuffle inputs
    shuffle_mask = np.random.permutation(len(X))
    X = np.asarray(X)[shuffle_mask]
    Y = np.asarray(Y)[shuffle_mask]

    return X, Y


def make_data():
    train_x, train_y = load_split("DATA", "train")
    val_x, val_y = load_split("DATA", "val")
    test_x, test_y = load_split("DATA", "test")

    print("train counts", Counter(train_y))
    print("val counts", Counter(val_y))
    print("test counts", Counter(test_y))

    return train_x, to_categorical(train_y), val_x, to_categorical(val_y), test_x, to_categorical(test_y)


def make_network(num_classes: int):
    net = PicoMobileNet(num_classes=num_classes)
    net.config.learning_rate = 0.01
    net.config.batch_size = 32
    net.config.verbosity = 1
    net.config.loss = "categorical_crossentropy"
    net.config.metrics = ["categorical_accuracy"]
    net.config.checkpoint_path = "./checkpoints/pico"

    net.build()
    net.compile()

    return net


if __name__ == '__main__':
    train_x, train_y, val_x, val_y, test_x, test_y = make_data()
    net = make_network(num_classes=10)

    net.fit(train_x, train_y, val_x, val_y, epochs=30)
    predictions = net.predict(test_x)
    print(classification_report(test_y.argmax(axis=1), predictions.argmax(axis=1)))
    print(net.convert.to_cpp())

Architectures

There are different architectures available in the micromobilenet package. They differ in the number of layers, ranging from the "canonical" MobileNet down to an extremely lightweight version with only N layers.

MobileNet

:noscroll:
Total params: 30,040 (117.34 KB)
Trainable params: 30,040 (117.34 KB)
Non-trainable params: 0 (0.00 B)
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ conv2d_0 (Conv2D)               │ (None, 47, 47, 3)      │            27 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__padding               │ (None, 49, 49, 3)      │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__dw (DepthwiseConv2D)  │ (None, 47, 47, 3)      │            27 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__relu_1 (ReLU)         │ (None, 47, 47, 3)      │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__pw (Conv2D)           │ (None, 47, 47, 6)      │            18 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__relu_2 (ReLU)         │ (None, 47, 47, 6)      │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__padding               │ (None, 49, 49, 6)      │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__dw (DepthwiseConv2D)  │ (None, 24, 24, 6)      │            54 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__relu_1 (ReLU)         │ (None, 24, 24, 6)      │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__pw (Conv2D)           │ (None, 24, 24, 12)     │            72 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__relu_2 (ReLU)         │ (None, 24, 24, 12)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__padding               │ (None, 26, 26, 12)     │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__dw (DepthwiseConv2D)  │ (None, 24, 24, 12)     │           108 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__relu_1 (ReLU)         │ (None, 24, 24, 12)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__pw (Conv2D)           │ (None, 24, 24, 12)     │           144 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__relu_2 (ReLU)         │ (None, 24, 24, 12)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__padding               │ (None, 26, 26, 12)     │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__dw (DepthwiseConv2D)  │ (None, 12, 12, 12)     │           108 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__relu_1 (ReLU)         │ (None, 12, 12, 12)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__pw (Conv2D)           │ (None, 12, 12, 24)     │           288 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__relu_2 (ReLU)         │ (None, 12, 12, 24)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_5__padding               │ (None, 14, 14, 24)     │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_5__dw (DepthwiseConv2D)  │ (None, 12, 12, 24)     │           216 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_5__relu_1 (ReLU)         │ (None, 12, 12, 24)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_5__pw (Conv2D)           │ (None, 12, 12, 24)     │           576 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_5__relu_2 (ReLU)         │ (None, 12, 12, 24)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_6__padding               │ (None, 14, 14, 24)     │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_6__dw (DepthwiseConv2D)  │ (None, 6, 6, 24)       │           216 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_6__relu_1 (ReLU)         │ (None, 6, 6, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_6__pw (Conv2D)           │ (None, 6, 6, 48)       │         1,152 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_6__relu_2 (ReLU)         │ (None, 6, 6, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_7__padding               │ (None, 8, 8, 48)       │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_7__dw (DepthwiseConv2D)  │ (None, 6, 6, 48)       │           432 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_7__relu_1 (ReLU)         │ (None, 6, 6, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_7__pw (Conv2D)           │ (None, 6, 6, 48)       │         2,304 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_7__relu_2 (ReLU)         │ (None, 6, 6, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_8__padding               │ (None, 8, 8, 48)       │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_8__dw (DepthwiseConv2D)  │ (None, 6, 6, 48)       │           432 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_8__relu_1 (ReLU)         │ (None, 6, 6, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_8__pw (Conv2D)           │ (None, 6, 6, 48)       │         2,304 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_8__relu_2 (ReLU)         │ (None, 6, 6, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_9__padding               │ (None, 8, 8, 48)       │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_9__dw (DepthwiseConv2D)  │ (None, 6, 6, 48)       │           432 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_9__relu_1 (ReLU)         │ (None, 6, 6, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_9__pw (Conv2D)           │ (None, 6, 6, 48)       │         2,304 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_9__relu_2 (ReLU)         │ (None, 6, 6, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_10__padding              │ (None, 8, 8, 48)       │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_10__dw (DepthwiseConv2D) │ (None, 6, 6, 48)       │           432 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_10__relu_1 (ReLU)        │ (None, 6, 6, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_10__pw (Conv2D)          │ (None, 6, 6, 48)       │         2,304 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_10__relu_2 (ReLU)        │ (None, 6, 6, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_11__padding              │ (None, 8, 8, 48)       │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_11__dw (DepthwiseConv2D) │ (None, 3, 3, 48)       │           432 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_11__relu_1 (ReLU)        │ (None, 3, 3, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_11__pw (Conv2D)          │ (None, 3, 3, 96)       │         4,608 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_11__relu_2 (ReLU)        │ (None, 3, 3, 96)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_12__padding              │ (None, 5, 5, 96)       │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_12__dw (DepthwiseConv2D) │ (None, 3, 3, 96)       │           864 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_12__relu_1 (ReLU)        │ (None, 3, 3, 96)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_12__pw (Conv2D)          │ (None, 3, 3, 96)       │         9,216 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_12__relu_2 (ReLU)        │ (None, 3, 3, 96)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ maxpool_last (MaxPool2D)        │ (None, 1, 1, 96)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ dropout (Dropout)               │ (None, 1, 1, 96)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ conv2d_last (Conv2D)            │ (None, 1, 1, 10)       │           970 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ reshape (Reshape)               │ (None, 10)             │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ softmax (Softmax)               │ (None, 10)             │             0 │
└─────────────────────────────────┴────────────────────────┴───────────────┘

MilliImageNet

:noscroll:
Total params: 11,704 (45.72 KB)
Trainable params: 11,704 (45.72 KB)
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ conv2d_0 (Conv2D)               │ (None, 47, 47, 3)      │            27 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__padding               │ (None, 49, 49, 3)      │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__dw (DepthwiseConv2D)  │ (None, 47, 47, 3)      │            27 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__relu_1 (ReLU)         │ (None, 47, 47, 3)      │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__pw (Conv2D)           │ (None, 47, 47, 6)      │            18 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__relu_2 (ReLU)         │ (None, 47, 47, 6)      │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__padding               │ (None, 49, 49, 6)      │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__dw (DepthwiseConv2D)  │ (None, 24, 24, 6)      │            54 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__relu_1 (ReLU)         │ (None, 24, 24, 6)      │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__pw (Conv2D)           │ (None, 24, 24, 12)     │            72 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__relu_2 (ReLU)         │ (None, 24, 24, 12)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__padding               │ (None, 26, 26, 12)     │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__dw (DepthwiseConv2D)  │ (None, 24, 24, 12)     │           108 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__relu_1 (ReLU)         │ (None, 24, 24, 12)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__pw (Conv2D)           │ (None, 24, 24, 12)     │           144 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__relu_2 (ReLU)         │ (None, 24, 24, 12)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__padding               │ (None, 26, 26, 12)     │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__dw (DepthwiseConv2D)  │ (None, 12, 12, 12)     │           108 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__relu_1 (ReLU)         │ (None, 12, 12, 12)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__pw (Conv2D)           │ (None, 12, 12, 24)     │           288 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__relu_2 (ReLU)         │ (None, 12, 12, 24)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_5__padding               │ (None, 14, 14, 24)     │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_5__dw (DepthwiseConv2D)  │ (None, 12, 12, 24)     │           216 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_5__relu_1 (ReLU)         │ (None, 12, 12, 24)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_5__pw (Conv2D)           │ (None, 12, 12, 24)     │           576 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_5__relu_2 (ReLU)         │ (None, 12, 12, 24)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_6__padding               │ (None, 14, 14, 24)     │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_6__dw (DepthwiseConv2D)  │ (None, 6, 6, 24)       │           216 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_6__relu_1 (ReLU)         │ (None, 6, 6, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_6__pw (Conv2D)           │ (None, 6, 6, 48)       │         1,152 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_6__relu_2 (ReLU)         │ (None, 6, 6, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_7__padding               │ (None, 8, 8, 48)       │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_7__dw (DepthwiseConv2D)  │ (None, 6, 6, 48)       │           432 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_7__relu_1 (ReLU)         │ (None, 6, 6, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_7__pw (Conv2D)           │ (None, 6, 6, 48)       │         2,304 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_7__relu_2 (ReLU)         │ (None, 6, 6, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_8__padding               │ (None, 8, 8, 48)       │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_8__dw (DepthwiseConv2D)  │ (None, 3, 3, 48)       │           432 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_8__relu_1 (ReLU)         │ (None, 3, 3, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_8__pw (Conv2D)           │ (None, 3, 3, 48)       │         2,304 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_8__relu_2 (ReLU)         │ (None, 3, 3, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_9__padding               │ (None, 5, 5, 48)       │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_9__dw (DepthwiseConv2D)  │ (None, 3, 3, 48)       │           432 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_9__relu_1 (ReLU)         │ (None, 3, 3, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_9__pw (Conv2D)           │ (None, 3, 3, 48)       │         2,304 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_9__relu_2 (ReLU)         │ (None, 3, 3, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ maxpool_last (MaxPool2D)        │ (None, 1, 1, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ dropout (Dropout)               │ (None, 1, 1, 48)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ conv2d_last (Conv2D)            │ (None, 1, 1, 10)       │           490 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ reshape (Reshape)               │ (None, 10)             │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ softmax (Softmax)               │ (None, 10)             │             0 │
└─────────────────────────────────┴────────────────────────┴───────────────┘

MicroMobileNet

:noscroll:
Total params: 4,264 (16.66 KB)
Trainable params: 4,264 (16.66 KB)
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ conv2d_0 (Conv2D)               │ (None, 47, 47, 3)      │            27 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__padding               │ (None, 49, 49, 3)      │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__dw (DepthwiseConv2D)  │ (None, 47, 47, 3)      │            27 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__relu_1 (ReLU)         │ (None, 47, 47, 3)      │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__pw (Conv2D)           │ (None, 47, 47, 6)      │            18 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__relu_2 (ReLU)         │ (None, 47, 47, 6)      │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__padding               │ (None, 49, 49, 6)      │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__dw (DepthwiseConv2D)  │ (None, 24, 24, 6)      │            54 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__relu_1 (ReLU)         │ (None, 24, 24, 6)      │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__pw (Conv2D)           │ (None, 24, 24, 12)     │            72 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__relu_2 (ReLU)         │ (None, 24, 24, 12)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__padding               │ (None, 26, 26, 12)     │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__dw (DepthwiseConv2D)  │ (None, 24, 24, 12)     │           108 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__relu_1 (ReLU)         │ (None, 24, 24, 12)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__pw (Conv2D)           │ (None, 24, 24, 12)     │           144 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__relu_2 (ReLU)         │ (None, 24, 24, 12)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__padding               │ (None, 26, 26, 12)     │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__dw (DepthwiseConv2D)  │ (None, 12, 12, 12)     │           108 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__relu_1 (ReLU)         │ (None, 12, 12, 12)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__pw (Conv2D)           │ (None, 12, 12, 24)     │           288 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__relu_2 (ReLU)         │ (None, 12, 12, 24)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_5__padding               │ (None, 14, 14, 24)     │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_5__dw (DepthwiseConv2D)  │ (None, 12, 12, 24)     │           216 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_5__relu_1 (ReLU)         │ (None, 12, 12, 24)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_5__pw (Conv2D)           │ (None, 12, 12, 24)     │           576 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_5__relu_2 (ReLU)         │ (None, 12, 12, 24)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_6__padding               │ (None, 14, 14, 24)     │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_6__dw (DepthwiseConv2D)  │ (None, 6, 6, 24)       │           216 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_6__relu_1 (ReLU)         │ (None, 6, 6, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_6__pw (Conv2D)           │ (None, 6, 6, 24)       │           576 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_6__relu_2 (ReLU)         │ (None, 6, 6, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_7__padding               │ (None, 8, 8, 24)       │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_7__dw (DepthwiseConv2D)  │ (None, 6, 6, 24)       │           216 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_7__relu_1 (ReLU)         │ (None, 6, 6, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_7__pw (Conv2D)           │ (None, 6, 6, 24)       │           576 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_7__relu_2 (ReLU)         │ (None, 6, 6, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_8__padding               │ (None, 8, 8, 24)       │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_8__dw (DepthwiseConv2D)  │ (None, 3, 3, 24)       │           216 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_8__relu_1 (ReLU)         │ (None, 3, 3, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_8__pw (Conv2D)           │ (None, 3, 3, 24)       │           576 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_8__relu_2 (ReLU)         │ (None, 3, 3, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ maxpool_last (MaxPool2D)        │ (None, 1, 1, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ dropout (Dropout)               │ (None, 1, 1, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ conv2d_last (Conv2D)            │ (None, 1, 1, 10)       │           250 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ reshape (Reshape)               │ (None, 10)             │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ softmax (Softmax)               │ (None, 10)             │             0 │
└─────────────────────────────────┴────────────────────────┴───────────────┘

NanoMobileNet

:noscroll:
Total params: 1,636 (6.39 KB)
Trainable params: 1,636 (6.39 KB)
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ conv2d_0 (Conv2D)               │ (None, 47, 47, 3)      │            27 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__padding               │ (None, 49, 49, 3)      │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__dw (DepthwiseConv2D)  │ (None, 24, 24, 3)      │            27 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__relu_1 (ReLU)         │ (None, 24, 24, 3)      │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__pw (Conv2D)           │ (None, 24, 24, 6)      │            18 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__relu_2 (ReLU)         │ (None, 24, 24, 6)      │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__padding               │ (None, 26, 26, 6)      │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__dw (DepthwiseConv2D)  │ (None, 12, 12, 6)      │            54 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__relu_1 (ReLU)         │ (None, 12, 12, 6)      │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__pw (Conv2D)           │ (None, 12, 12, 12)     │            72 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__relu_2 (ReLU)         │ (None, 12, 12, 12)     │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__padding               │ (None, 14, 14, 12)     │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__dw (DepthwiseConv2D)  │ (None, 6, 6, 12)       │           108 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__relu_1 (ReLU)         │ (None, 6, 6, 12)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__pw (Conv2D)           │ (None, 6, 6, 24)       │           288 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__relu_2 (ReLU)         │ (None, 6, 6, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__padding               │ (None, 8, 8, 24)       │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__dw (DepthwiseConv2D)  │ (None, 3, 3, 24)       │           216 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__relu_1 (ReLU)         │ (None, 3, 3, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__pw (Conv2D)           │ (None, 3, 3, 24)       │           576 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_4__relu_2 (ReLU)         │ (None, 3, 3, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ maxpool_last (MaxPool2D)        │ (None, 1, 1, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ dropout (Dropout)               │ (None, 1, 1, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ conv2d_last (Conv2D)            │ (None, 1, 1, 10)       │           250 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ reshape (Reshape)               │ (None, 10)             │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ softmax (Softmax)               │ (None, 10)             │             0 │
└─────────────────────────────────┴────────────────────────┴───────────────┘

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62

PicoMobileNet

:noscroll:
Total params: 844 (3.30 KB)
Trainable params: 844 (3.30 KB)
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ conv2d_0 (Conv2D)               │ (None, 32, 32, 3)      │            27 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__padding               │ (None, 34, 34, 3)      │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__dw (DepthwiseConv2D)  │ (None, 16, 16, 3)      │            27 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__relu_1 (ReLU)         │ (None, 16, 16, 3)      │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__pw (Conv2D)           │ (None, 16, 16, 6)      │            18 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_1__relu_2 (ReLU)         │ (None, 16, 16, 6)      │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__padding               │ (None, 18, 18, 6)      │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__dw (DepthwiseConv2D)  │ (None, 8, 8, 6)        │            54 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__relu_1 (ReLU)         │ (None, 8, 8, 6)        │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__pw (Conv2D)           │ (None, 8, 8, 12)       │            72 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_2__relu_2 (ReLU)         │ (None, 8, 8, 12)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__padding               │ (None, 10, 10, 12)     │             0 │
│ (ZeroPadding2D)                 │                        │               │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__dw (DepthwiseConv2D)  │ (None, 4, 4, 12)       │           108 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__relu_1 (ReLU)         │ (None, 4, 4, 12)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__pw (Conv2D)           │ (None, 4, 4, 24)       │           288 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ hidden_3__relu_2 (ReLU)         │ (None, 4, 4, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ maxpool_last (MaxPool2D)        │ (None, 1, 1, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ dropout (Dropout)               │ (None, 1, 1, 24)       │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ conv2d_last (Conv2D)            │ (None, 1, 1, 10)       │           250 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ reshape (Reshape)               │ (None, 10)             │             0 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ softmax (Softmax)               │ (None, 10)             │             0 │
└─────────────────────────────────┴────────────────────────┴───────────────┘

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51

MobileNet for Microcontrollers MobileNet for Microcontrollers