O mercado de rastreamento ocular deverá crescer e crescer: de US $ 560 milhões em 2020 para US $ 1.786 bilhões em 2025 . Então, qual é a alternativa para dispositivos relativamente caros? Claro, uma webcam simples! Como outras, essa abordagem encontra muitas dificuldades, sejam elas: uma ampla variedade de dispositivos (portanto, é difícil escolher configurações que funcionem igualmente em todas as câmeras), forte variabilidade de parâmetros (da iluminação à inclinação da câmera e sua posição em relação ao rosto), computação decente poder (vários núcleos de cuda e Xeon - é isso) ...

Embora espere um minuto, é realmente necessário gastar dinheiro em hardware topo de linha e até comprar uma placa de vídeo? Talvez haja uma maneira de ajustar todos os cálculos na CPU e não perder velocidade ao fazê-lo?

(Bem, se não houvesse esse caminho, não haveria um artigo sobre como treinar um neurônio no PyTorch)

Dados

Como sempre na ciência de dados, a questão mais importante. Após algum tempo de pesquisa, encontrei o conjunto de dados MPIIGaze . Os autores do artigo sugeriram várias maneiras legais de processá-lo (por exemplo, normalizando a posição da cabeça), mas seguiremos o caminho simples.

Então, inicie o Colab , carregue o laptop e inicie:

#  
import os

import numpy as np
import pandas as pd
import scipy
import scipy.io

from PIL import Image
import cv2

import seaborn as sns
import matplotlib
import matplotlib.pyplot as plt

No Colab, você pode usar os utilitários de sistema diretamente do seu laptop, semear, baixar e descompactar o conjunto de dados:

!wget https://datasets.d2.mpi-inf.mpg.de/MPIIGaze/MPIIGaze.tar.gz
!tar xvzf MPIIGaze.tar.gz MPIIGaze

Data/Original . 15, . Annotation Subset , — . header', , .

database_path = "/content/MPIIGaze"

#      
def load_image_data(patient_name):
    global database_path

    annotation_path = os.path.join(database_path, "Annotation Subset", patient_name + ".txt")
    data_folder = os.path.join(database_path, "Data", "Original", patient_name)

    annotation = pd.read_csv(annotation_path, sep=" ", header=None)

    points = np.array(annotation.loc[:, list(range(1, 17))])

    filenames = np.array(annotation.loc[:, [0]]).reshape(-1)
    images = [np.array(Image.open(os.path.join(data_folder, filename))) for filename in filenames]

    return images, points

images, points = load_image_data("p00")

plt.imshow(images[0])
colors = ["r", "g", "b", "magenta", "y", "cyan", "brown", "lightcoral"]
for i in range(0, len(points[0]), 2):
    x, y = points[0, i:i+2] #     ,      2,    X, Y
    plt.scatter([x], [y], c=colors[i//2])

:

, : , , .

, . , : , (, , ), , 2:1. 2 1 , .

#     
def distance(x1, y1, x2, y2):
    return int(((x1 - x2) ** 2 + (y1 - y2) ** 2) ** 0.5)

image_shape = (16, 32)

#           
#   
def handle_eye(image, p1, p2, pupil):
    global image_shape

    line_len = distance(*p1, *p2)
    # x, y -> y, x
    p1 = p1[::-1]
    p2 = p2[::-1]
    pupil = pupil[::-1]

    corner1 = p1 - np.array([line_len//4, 0])
    corner2 = p2 + np.array([line_len//4, 0])

    sub_image = image[corner1[0]:corner2[0]+1, corner1[1]:corner2[1]+1]

    pupil_new = pupil - corner1
    pupil_new = pupil_new / sub_image.shape[:2]

    sub_image = cv2.resize(sub_image, image_shape[::-1], interpolation=cv2.INTER_AREA)
    sub_image = cv2.cvtColor(sub_image, cv2.COLOR_RGB2GRAY)

    return sub_image, pupil_new

2 , — :

def image_to_train_data(image, points):
    eye_right_p1 = points[0:2]
    eye_right_p2 = points[2:4]
    eye_right_pupil = points[12:14]

    right_image, right_pupil = handle_eye(image, eye_right_p1, eye_right_p2, eye_right_pupil)

    eye_left_p1 = points[4:6]
    eye_left_p2 = points[6:8]
    eye_left_pupil = points[14:16]

    left_image, left_pupil = handle_eye(image, eye_left_p1, eye_left_p2, eye_left_pupil)

    return right_image, right_pupil, left_image, left_pupil

( ):

#   
right_image, right_pupil, left_image, left_pupil = image_to_train_data(images[10], points[10])

plt.imshow(right_image, cmap="gray")

r_p_x = int(right_pupil[1] * image_shape[1])
r_p_y = int(right_pupil[0] * image_shape[0])
plt.scatter([r_p_x], [r_p_y], c="red")

, - . :

images_left_conc = []
images_right_conc = []
pupils_left_conc = []
pupils_right_conc = []

patients_path = os.path.join(database_path, "Data", "Original")
for patient in os.listdir(patients_path):
    print(patient)
    images, points = load_image_data(patient)
    for i in range(len(images)):
        signle_image_data = image_to_train_data(images[i], points[i])

        if any(stuff is None for stuff in signle_image_data):
            continue

        right_image, right_pupil, left_image, left_pupil = signle_image_data

        if any(right_pupil < 0) or any(left_pupil < 0):
            continue

        images_right_conc.append(right_image)
        images_left_conc.append(left_image)
        pupils_right_conc.append(right_pupil)
        pupils_left_conc.append(left_pupil)

images_left_conc = np.array(images_left_conc)
images_right_conc = np.array(images_right_conc)
pupils_left_conc = np.array(pupils_left_conc)
pupils_right_conc = np.array(pupils_right_conc)

:

images_left_conc = images_left_conc / 255
images_right_conc = images_right_conc / 255

, : :

pupils_conc = np.zeros_like(pupils_left_conc)
for i in range(2):
    pupils_conc[:, i] = (pupils_left_conc[:, i] + pupils_right_conc[:, i]) / 2

:

viz_pupils = np.zeros(image_shape)
for y, x in pupils_conc:
    y = int(y * image_shape[0])
    x = int(x * image_shape[1])
    viz_pupils[y, x] += 1
max_val = viz_pupils.max()
viz_pupils = viz_pupils / max_val

plt.imshow(viz_pupils, cmap="hot")

, .

#    
from sklearn.model_selection import train_test_split

import torch
from torch.utils.data import DataLoader, TensorDataset

# ,      --   
def make_2eyes_datasets(images_left, images_right, pupils, train_size=0.8):
    n, height, width = images_left.shape

    images_left = images_left.reshape(n, 1, height, width)
    images_right = images_right.reshape(n, 1, height, width)

    images_left_train, images_left_val, images_right_train, images_right_val, pupils_train, pupils_val = train_test_split(
        images_left, images_right, pupils, train_size=train_size
    )

    def make_dataset(im_left, im_right, pups):
        return TensorDataset(
            torch.from_numpy(im_left.astype(np.float32)), torch.from_numpy(im_right.astype(np.float32)), torch.from_numpy(pups.astype(np.float32))
        )

    train_dataset = make_dataset(images_left_train, images_right_train, pupils_train)
    val_dataset = make_dataset(images_left_val, images_right_val, pupils_val)

    return train_dataset, val_dataset

#    
def make_dataloaders(train_dataset, val_dataset, batch_size=256):
    train_dataloader = DataLoader(train_dataset, batch_size=batch_size)
    val_dataloader = DataLoader(val_dataset, batch_size=batch_size)

    return train_dataloader, val_dataloader

batch_size = 256

eyes_datasets = make_2eyes_datasets(images_left_conc, images_right_conc, pupils_conc)
eyes_train_loader, eyes_val_loader = make_dataloaders(*eyes_datasets, batch_size=batch_size)

import torch
import torch.nn as nn
import torch.nn.functional as F

# ,  `keras.layers.Reshape`
class Reshaper(nn.Module):
    def __init__(self, target_shape):
        super(Reshaper, self).__init__()
        self.target_shape = target_shape

    def forward(self, input):
        return torch.reshape(input, (-1, *self.target_shape))

#  
class EyesNet(nn.Module):
    def __init__(self):
        super(EyesNet, self).__init__()

        #  feature-extractor'   
        self.features_left = nn.Sequential(
            nn.Conv2d(in_channels=1, out_channels=32, kernel_size=5, stride=2, padding=2),
            nn.LeakyReLU(),
            nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3, stride=2, padding=1),
            nn.LeakyReLU(),
            nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=2, padding=1),
            nn.LeakyReLU(),
            nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=2, padding=1),
            nn.LeakyReLU(),
            nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=2, padding=1),
            nn.LeakyReLU(),
            Reshaper([64])
        )
        self.features_right = nn.Sequential(
            nn.Conv2d(in_channels=1, out_channels=32, kernel_size=5, stride=2, padding=2),
            nn.LeakyReLU(),
            nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3, stride=2, padding=1),
            nn.LeakyReLU(),
            nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=2, padding=1),
            nn.LeakyReLU(),
            nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=2, padding=1),
            nn.LeakyReLU(),
            nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=2, padding=1),
            nn.LeakyReLU(),
            Reshaper([64])
        )
        self.fc = nn.Sequential(
            nn.Linear(128, 64),
            nn.LeakyReLU(),
            nn.Linear(64, 16),
            nn.LeakyReLU(),
            nn.Linear(16, 2),
            nn.Sigmoid()
        )

    def forward(self, x_left, x_right):
        #      ,    
        x_left = self.features_left(x_left)
        x_right = self.features_right(x_right)
        x = torch.cat((x_left, x_right), 1)
        x = self.fc(x)

        return x

, GPU ( CPU ≈ ), 8 .

#   
def train(model, train_loader, test_loader, epochs, lr, folder="gazenet"):
    os.makedirs(folder, exist_ok=True)

    optimizer = torch.optim.Adam(model.parameters(), lr=lr)
    mse = nn.MSELoss()

    for epoch in range(epochs):
        running_loss = 0
        error_mean = []
        error_std = []
        for i, (*xs_batch, y_batch) in enumerate(train_loader):
            xs_batch = [x_batch.cuda() for x_batch in xs_batch]
            y_batch = y_batch.cuda()

            optimizer.zero_grad()

            y_batch_pred = model(*xs_batch)
            loss = mse(y_batch_pred, y_batch)

            loss.backward()
            optimizer.step()

            running_loss += loss.item()

            difference = (y_batch - y_batch_pred).detach().cpu().numpy().reshape(-1)
            error_mean.append(np.mean(difference))
            error_std.append(np.std(difference))

        error_mean = np.mean(error_mean)
        error_std = np.mean(error_std)

        print(f"Epoch {epoch+1}/{epochs}, train loss: {running_loss}, error mean: {error_mean}, error std: {error_std}")

        running_loss = 0
        error_mean = []
        error_std = []
        for i, (*xs_batch, y_batch) in enumerate(train_loader):
            xs_batch = [x_batch.cuda() for x_batch in xs_batch]
            y_batch = y_batch.cuda()

            y_batch_pred = model(*xs_batch)
            loss = mse(y_batch_pred, y_batch)

            loss.backward()
            running_loss += loss.item()

            difference = (y_batch - y_batch_pred).detach().cpu().numpy().reshape(-1)
            error_mean.append(np.mean(difference))
            error_std.append(np.std(difference))

        error_mean = np.mean(error_mean)
        error_std = np.mean(error_std)

        print(f"Epoch {epoch+1}/{epochs}, val loss: {running_loss}, error mean: {error_mean}, error std: {error_std}")

        epoch_path = os.path.join(folder, f"epoch_{epoch+1}.pth")
        torch.save(model.state_dict(), epoch_path)

eyesnet = EyesNet().cuda()
#      *eyes_net*
train(eyesnet, eyes_train_loader, eyes_val_loader, 300, 1e-3, "eyes_net")

, 300 ( , ):

Epoch 1/300, train loss: 0.3125856015831232, error mean: -0.019309822469949722, error std: 0.08668763190507889
Epoch 1/300, val loss: 0.18365296721458435, error mean: -0.008721884340047836, error std: 0.07283741235733032
Epoch 2/300, train loss: 0.1700970521196723, error mean: 0.0001489206333644688, error std: 0.07033108174800873
Epoch 2/300, val loss: 0.1475073655601591, error mean: -0.001808341359719634, error std: 0.06572529673576355
...
Epoch 299/300, train loss: 0.003378463063199888, error mean: -8.133996743708849e-05, error std: 0.009488753043115139
Epoch 299/300, val loss: 0.004163481352406961, error mean: -0.001996406354010105, error std: 0.010547727346420288
Epoch 300/300, train loss: 0.003569353237253381, error mean: -9.1125002654735e-05, error std: 0.00977678969502449
Epoch 300/300, val loss: 0.004456713928448153, error mean: 0.0008482271223329008, error std: 0.010923181660473347

299 , .

:

import random

#            
def show_output(model, data_loader, batch_num=0, samples=5, grid_shape=(5, 1), figsize=(10, 10)):
    for i, (*xs, y) in enumerate(data_loader):
        if i == batch_num:
            break
    xs = [x.cuda() for x in xs]
    y_pred = model(*xs).detach().cpu().numpy().reshape(-1, 2)

    xs = [x.detach().cpu().numpy().reshape(-1, 16, 32) for x in xs]
    imgs_conc = np.hstack(xs)
    y = y.cpu().numpy().reshape(-1, 2)

    indices = random.sample(range(len(y_pred)), samples)
    fig, axes = plt.subplots(*grid_shape, figsize=figsize)
    for i, index in enumerate(indices):
        row = i // grid_shape[1]
        column = i % grid_shape[1]

        axes[row, column].imshow(imgs_conc[index])
        axes[row, column].scatter([y_pred[index, 1]*32, y_pred[index, 1]*32], [y_pred[index, 0]*16, (y_pred[index, 0]+1)*16], c="r")
        axes[row, column].scatter([y[index, 1]*32, y[index, 1]*32], [y[index, 0]*16, (y[index, 0]+1)*16], c="g")

#  299 
eyesnet.load_state_dict(torch.load("eyes_net/epoch_299.pth"))

show_output(eyesnet, eyes_val_loader, 103, 16, (4, 4))

, , "" , . — -, - , -, . , .

( X Y), :

def error_distribution(model, data_loader, image_shape=(16, 32), bins=32, digits=2, figsize=(10,10)):
    ys_true = []
    ys_pred = []
    for *xs, y in data_loader:
        xs = [x.cuda() for x in xs]
        y_pred = model(*xs)

        ys_true.append(y.detach().cpu().numpy())
        ys_pred.append(y_pred.detach().cpu().numpy())
    ys_true = np.concatenate(ys_true)
    ys_pred = np.concatenate(ys_pred)
    indices = np.arange(len(ys_true))

    fig, axes = plt.subplots(2, figsize=figsize)
    for ax_num in range(2):
        ys_true_subset = ys_true[:, ax_num]
        ys_pred_subset = ys_pred[:, ax_num]
        counts, ranges = np.histogram(ys_true_subset, bins=bins)

        errors = []
        labels = []
        for i in range(len(counts)):
            begin, end = ranges[i], ranges[i + 1]
            range_indices = indices[(ys_true_subset >= begin) & (ys_true_subset <= end)]

            diffs = np.abs(ys_pred_subset[range_indices] - ys_true_subset[range_indices])
            label = (begin + end) / 2
            if image_shape:
                diffs = diffs * image_shape[ax_num]
                label = label * image_shape[ax_num]
            else:
                label = round(label, digits)
            errors.append(diffs)
            labels.append(str(label)[:2+digits])

        axes[ax_num].boxplot(errors, labels=labels)

        if image_shape:
            y_label = "difference, px"
            x_label = "true position, px"
        else:
            y_label = "difference"
            x_label = "true position"
        axes[ax_num].set_ylabel(y_label)
        axes[ax_num].set_xlabel(x_label)

        if ax_num == 0:
            title = "Y"
        else:
            title = "X"
        axes[ax_num].set_title(title)

error_distribution(eyesnet, eyes_val_loader, figsize=(20, 10))

,

-, . , . :

import time

def measure_time(model, data_loader, n_batches=5):
    begin_time = time.time()

    batch_num = 0
    n_samples = 0

    predicted = []
    for *xs, y in data_loader:
        xs = [x.cpu() for x in xs]

        y_pred = model(*xs)
        predicted.append(y_pred.detach().cpu().numpy().reshape(-1))

        batch_num += 1
        n_samples += len(y)

        if batch_num >= n_batches:
            break

    end_time = time.time()

    time_per_sample = (end_time - begin_time) / n_samples
    return time_per_sample

eyesnet_cpu = EyesNet().cpu()
eyesnet_cpu.load_state_dict(torch.load("eyes_net/epoch_299.pth", map_location="cpu"))

#  dataloader,    ,     realtime
_, eyes_val_loader_single = make_dataloaders(*eyes_datasets, batch_size=1)

tps = measure_time(eyesnet_cpu, eyes_val_loader_single)
print(f"{tps} seconds per sample")
>>> 0.003347921371459961 seconds per sample

, VGG16 ( , ):

import torchvision.models as models

class VGG16Based(nn.Module):
    def __init__(self):
        super(VGG16Based, self).__init__()

        self.vgg = models.vgg16(pretrained=False)
        self.vgg.classifier = nn.Sequential(
            nn.Linear(25088, 256),
            nn.LeakyReLU(),
            nn.Linear(256, 2),
            nn.Sigmoid()
        )

    def forward(self, x_left, x_right):
        x_mid = (x_left + x_right) / 2
        x = torch.cat((x_left, x_mid, x_right), dim=1)

        #  ,  VGG16   
        x_pad = torch.zeros((x.shape[0], 3, 32, 32))
        x_pad[:, :, :16, :] = x

        x = self.vgg(x_pad)

        return x

vgg16 = VGG16Based()
vgg16_tps = measure_time(vgg16, eyes_val_loader_single)
print(f"{vgg16_tps} seconds per sample")
>>> 0.023713159561157226 seconds per sample

, (AMD A10-4600M APU, 1500 MHz):

python benchmark.py 
0.003980588912963867 seconds per sample, EyesNet
0.12246298789978027 seconds per sample, VGG16-based

, , , ( VGG16 80 , EyesNet — 1 ; , , ). , , . , :

1. ( ).
2. . , float8 float32 ( , , ).
3. PyTorch Mobile — PyTorch . .
4. . — GazeCapture. , , — :
5. TFLite — TensorFlow . !

, . Data science ( — *^*) . — FARADAY Lab. — , .

c:

:

• MPIIGaze
• PyTorch mobile
• Tensorflow lite