數(shù)據(jù)增強(qiáng)

數(shù)據(jù)增強(qiáng)是一種通過使用現(xiàn)有圖像的不同變體創(chuàng)建新的訓(xùn)練圖像來更好地概括我們的模型的技術(shù)。我們當(dāng)前的訓(xùn)練集中只有 800 張圖像，因此數(shù)據(jù)增強(qiáng)對于確保我們的模型不會過擬合非常重要。

對于這個(gè)問題，我使用了翻轉(zhuǎn)、旋轉(zhuǎn)、中心裁剪和隨機(jī)裁剪。

這里唯一需要記住的是確保包圍盒也以與圖像相同的方式進(jìn)行轉(zhuǎn)換。

# modified from fast.ai

def crop(im, r, c, target_r, target_c): 

    return im[r:r+target_r, c:c+target_c]

# random crop to the original size

def random_crop(x, r_pix=8):

    """ Returns a random crop"""

    r, c,*_ = x.shape

    c_pix = round(r_pix*c/r)

    rand_r = random.uniform(0, 1)

    rand_c = random.uniform(0, 1)

    start_r = np.floor(2*rand_r*r_pix).astype(int)

    start_c = np.floor(2*rand_c*c_pix).astype(int)

    return crop(x, start_r, start_c, r-2*r_pix, c-2*c_pix)

def center_crop(x, r_pix=8):

    r, c,*_ = x.shape

    c_pix = round(r_pix*c/r)

    return crop(x, r_pix, c_pix, r-2*r_pix, c-2*c_pix)

def rotate_cv(im, deg, y=False, mode=cv2.BORDER_REFLECT, interpolation=cv2.INTER_AREA):

    """ Rotates an image by deg degrees"""

    r,c,*_ = im.shape

    M = cv2.getRotationMatrix2D((c/2,r/2),deg,1)

    if y:

        return cv2.warpAffine(im, M,(c,r), borderMode=cv2.BORDER_CONSTANT)

    return cv2.warpAffine(im,M,(c,r), borderMode=mode, flags=cv2.WARP_FILL_OUTLIERS+interpolation)

def random_cropXY(x, Y, r_pix=8):

    """ Returns a random crop"""

    r, c,*_ = x.shape

    c_pix = round(r_pix*c/r)

    rand_r = random.uniform(0, 1)

    rand_c = random.uniform(0, 1)

    start_r = np.floor(2*rand_r*r_pix).astype(int)

    start_c = np.floor(2*rand_c*c_pix).astype(int)

    xx = crop(x, start_r, start_c, r-2*r_pix, c-2*c_pix)

    YY = crop(Y, start_r, start_c, r-2*r_pix, c-2*c_pix)

    return xx, YY

def transformsXY(path, bb, transforms):

    x = cv2.imread(str(path)).astype(np.float32)

    x = cv2.cvtColor(x, cv2.COLOR_BGR2RGB)/255

    Y = create_mask(bb, x)

    if transforms:

        rdeg = (np.random.random()-.50)*20

        x = rotate_cv(x, rdeg)

        Y = rotate_cv(Y, rdeg, y=True)

        if np.random.random() > 0.5: 

            x = np.fliplr(x).copy()

            Y = np.fliplr(Y).copy()

        x, Y = random_cropXY(x, Y)

    else:

        x, Y = center_crop(x), center_crop(Y)

    return x, mask_to_bb(Y)

def create_corner_rect(bb, color='red'):

    bb = np.array(bb, dtype=np.float32)

    return plt.Rectangle((bb, bb), bb-bb, bb-bb, color=color,

                         fill=False, lw=3)

def show_corner_bb(im, bb):

    plt.imshow(im)

    plt.gca().add_patch(create_corner_rect(bb))

PyTorch 數(shù)據(jù)集

現(xiàn)在我們已經(jīng)有了數(shù)據(jù)增強(qiáng)，我們可以進(jìn)行訓(xùn)練驗(yàn)證拆分并創(chuàng)建我們的 PyTorch 數(shù)據(jù)集。我們使用 ImageNet 統(tǒng)計(jì)數(shù)據(jù)對圖像進(jìn)行標(biāo)準(zhǔn)化，因?yàn)槲覀兪褂玫氖穷A(yù)訓(xùn)練的 ResNet 模型并在訓(xùn)練時(shí)在我們的數(shù)據(jù)集中應(yīng)用數(shù)據(jù)增強(qiáng)。

X_train, X_val, y_train, y_val = train_test_split(X, Y, test_size=0.2, random_state=42)

def normalize(im):

    """Normalizes images with Imagenet stats."""

    imagenet_stats = np.array([[0.485, 0.456, 0.406], [0.229, 0.224, 0.225]])

    return (im - imagenet_stats)/imagenet_stats

class RoadDataset(Dataset):

    def __init__(self, paths, bb, y, transforms=False):

        self.transforms = transforms

        self.paths = paths.values

        self.bb = bb.values

        self.y = y.values

    def __len__(self):

        return len(self.paths)

    def __getitem__(self, idx):

        path = self.paths[idx]

        y_class = self.y[idx]

        x, y_bb = transformsXY(path, self.bb[idx], self.transforms)

        x = normalize(x)

        x = np.rollaxis(x, 2)

        return x, y_class, y_bb

train_ds = RoadDataset(X_train['new_path'],X_train['new_bb'] ,y_train, transforms=True)

valid_ds = RoadDataset(X_val['new_path'],X_val['new_bb'],y_val)

batch_size = 64

train_dl = DataLoader(train_ds, batch_size=batch_size, shuffle=True)

valid_dl = DataLoader(valid_ds, batch_size=batch_size)

PyTorch 模型

對于這個(gè)模型，我使用了一個(gè)非常簡單的預(yù)先訓(xùn)練的 resNet-34模型。由于我們有兩個(gè)任務(wù)要完成，這里有兩個(gè)最后的層: 包圍盒回歸器和圖像分類器。

class BB_model(nn.Module):

    def __init__(self):

        super(BB_model, self).__init__()

        resnet = models.resnet34(pretrained=True)

        layers = list(resnet.children())[:8]

        self.features1 = nn.Sequential(*layers[:6])

        self.features2 = nn.Sequential(*layers[6:])

        self.classifier = nn.Sequential(nn.BatchNorm1d(512), nn.Linear(512, 4))

        self.bb = nn.Sequential(nn.BatchNorm1d(512), nn.Linear(512, 4))

    def forward(self, x):

        x = self.features1(x)

        x = self.features2(x)

        x = F.relu(x)

        x = nn.AdaptiveAvgPool2d((1,1))(x)

        x = x.view(x.shape, -1)

        return self.classifier(x), self.bb(x)

訓(xùn)練

對于損失，我們需要同時(shí)考慮分類損失和邊界框回歸損失，因此我們使用交叉熵和 L1 損失（真實(shí)值和預(yù)測坐標(biāo)之間的所有絕對差之和）的組合。我已經(jīng)將 L1 損失縮放了 1000 倍，因?yàn)榉诸惡突貧w損失都在相似的范圍內(nèi)。除此之外，它是一個(gè)標(biāo)準(zhǔn)的 PyTorch 訓(xùn)練循環(huán)（使用 GPU）：

def update_optimizer(optimizer, lr):

    for i, param_group in enumerate(optimizer.param_groups):

        param_group["lr"] = lr

def train_epocs(model, optimizer, train_dl, val_dl, epochs=10,C=1000):

    idx = 0

    for i in range(epochs):

        model.train()

        total = 0

        sum_loss = 0

        for x, y_class, y_bb in train_dl:

            batch = y_class.shape

            x = x.cuda().float()

            y_class = y_class.cuda()

            y_bb = y_bb.cuda().float()

            out_class, out_bb = model(x)

            loss_class = F.cross_entropy(out_class, y_class, reduction="sum")

            loss_bb = F.l1_loss(out_bb, y_bb, reduction="none").sum(1)

            loss_bb = loss_bb.sum()

            loss = loss_class + loss_bb/C

            optimizer.zero_grad()

            loss.backward()

            optimizer.step()

            idx += 1

            total += batch

            sum_loss += loss.item()

        train_loss = sum_loss/total

        val_loss, val_acc = val_metrics(model, valid_dl, C)

        print("train_loss %.3f val_loss %.3f val_acc %.3f" % (train_loss, val_loss, val_acc))

    return sum_loss/total

def val_metrics(model, valid_dl, C=1000):

    model.eval()

    total = 0

    sum_loss = 0

    correct = 0 

    for x, y_class, y_bb in valid_dl:

        batch = y_class.shape

        x = x.cuda().float()

        y_class = y_class.cuda()

        y_bb = y_bb.cuda().float()

        out_class, out_bb = model(x)

        loss_class = F.cross_entropy(out_class, y_class, reduction="sum")

        loss_bb = F.l1_loss(out_bb, y_bb, reduction="none").sum(1)

        loss_bb = loss_bb.sum()

        loss = loss_class + loss_bb/C

        _, pred = torch.max(out_class, 1)

        correct += pred.eq(y_class).sum().item()

        sum_loss += loss.item()

        total += batch

    return sum_loss/total, correct/total

model = BB_model().cuda()

parameters = filter(lambda p: p.requires_grad, model.parameters())

optimizer = torch.optim.Adam(parameters, lr=0.006)

train_epocs(model, optimizer, train_dl, valid_dl, epochs=15)

測試

現(xiàn)在我們已經(jīng)完成了訓(xùn)練，我們可以選擇一個(gè)隨機(jī)圖像并在上面測試我們的模型。盡管我們只有相當(dāng)少量的訓(xùn)練圖像，但是我們最終在測試圖像上得到了一個(gè)相當(dāng)不錯(cuò)的預(yù)測。

使用手機(jī)拍攝真實(shí)照片并測試模型將是一項(xiàng)有趣的練習(xí)。另一個(gè)有趣的實(shí)驗(yàn)是不執(zhí)行任何數(shù)據(jù)增強(qiáng)并訓(xùn)練模型并比較兩個(gè)模型。

# resizing test image

im = read_image('./road_signs/images_resized/road789.png')

im = cv2.resize(im, (int(1.49*300), 300))

cv2.imwrite('./road_signs/road_signs_test/road789.jpg', cv2.cvtColor(im, cv2.COLOR_RGB2BGR))

# test Dataset

test_ds = RoadDataset(pd.DataFrame([{'path':'./road_signs/road_signs_test/road789.jpg'}])['path'],pd.DataFrame([{'bb':np.array([0,0,0,0])}])['bb'],pd.DataFrame([{'y':}])['y'])

x, y_class, y_bb = test_ds

xx = torch.FloatTensor(x[None,])

xx.shape

# prediction

out_class, out_bb = model(xx.cuda())

out_class, out_bb

總結(jié)

現(xiàn)在我們已經(jīng)介紹了目標(biāo)檢測的基本原理，并從頭開始實(shí)現(xiàn)它，您可以將這些想法擴(kuò)展到多對象情況，并嘗試更復(fù)雜的模型，如 RCNN 和 YOLO！