13 计算机视觉-代码详解

13.2 微调

为了防止在训练集上过拟合，有两种办法，第一种是扩大训练集数量，但是需要大量的成本；第二种就是应用迁移学习，将源数据学习到的知识迁移到目标数据集，即在把在源数据训练好的参数和模型（除去输出层）直接复制到目标数据集训练。

# IPython魔法函数，可以不用执行plt .show()
%matplotlib inline
import os
import torch
import torchvision
from torch import nn
from d2l import torch as d2l

13.2.1 获取数据集

#@save
d2l.DATA_HUB['hotdog'] = (d2l.DATA_URL + 'hotdog.zip','fba480ffa8aa7e0febbb511d181409f899b9baa5')data_dir = d2l.download_extract('hotdog')
train_imgs = torchvision.datasets.ImageFolder(os.path.join(data_dir, 'train'))
test_imgs = torchvision.datasets.ImageFolder(os.path.join(data_dir, 'test'))
hotdogs = [train_imgs[i][0] for i in range(8)]
not_hotdogs = [train_imgs[-i-1][0] for i in range(8)]
# 展示2行8列矩阵的图片，共16张
d2l.show_images(hotdogs+not_hotdogs,2,8,scale=1.5)

# 使用RGB通道的均值和标准差，以标准化每个通道
normalize = torchvision.transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
# 图像增广
train_augs = torchvision.transforms.Compose([torchvision.transforms.RandomResizedCrop(224),torchvision.transforms.RandomHorizontalFlip(),torchvision.transforms.ToTensor(),normalize])
test_augs = torchvision.transforms.Compose([torchvision.transforms.Resize([256, 256]),torchvision.transforms.CenterCrop(224),torchvision.transforms.ToTensor(),normalize])

 13.2.2 初始化模型

# 自动下载网上的训练模型
finetune_net = torchvision.models.resnet18(pretrained=True)
# 输入张量的形状还是源输入张量大小，输入张量大小改为2
finetune_net.fc = nn.Linear(finetune_net.fc.in_features, 2)
nn.init.xavier_uniform_(finetune_net.fc.weight);

13.2.3 微调模型

# 如果param_group=True，输出层中的模型参数将使用十倍的学习率
# 如果param_group=False，输出层中模型参数为随机值
# 训练模型
def train_fine_tuning(net, learning_rate, batch_size=128, num_epochs=5,param_group=True):train_iter = torch.utils.data.DataLoader(torchvision.datasets.ImageFolder(os.path.join(data_dir, 'train'), transform=train_augs),batch_size=batch_size, shuffle=True)test_iter = torch.utils.data.DataLoader(torchvision.datasets.ImageFolder(os.path.join(data_dir, 'test'), transform=test_augs),batch_size=batch_size)devices = d2l.try_all_gpus()loss = nn.CrossEntropyLoss(reduction="none")if param_group:params_1x = [param for name, param in net.named_parameters()if name not in ["fc.weight", "fc.bias"]]# params_1x的参数使用learning_rate学习率， net.fc.parameters()的参数使用0.001的学习率trainer = torch.optim.SGD([{'params': params_1x},{'params': net.fc.parameters(),'lr': learning_rate * 10}],lr=learning_rate, weight_decay=0.001)else:trainer = torch.optim.SGD(net.parameters(), lr=learning_rate,weight_decay=0.001)d2l.train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs,devices)
train_fine_tuning(finetune_net, 5e-5)

13.3 目标检测和边界框

有时候不仅要识别图像的类别，还需要识别图像的位置。在计算机视觉中叫做目标识别或者目标检测。这小节是介绍目标检测的深度学习方法。

%matplotlib inline
import torch
from d2l import torch as d2l
#@save
def box_corner_to_center(boxes):"""从（左上，右下）转换到（中间，宽度，高度）"""x1, y1, x2, y2 = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]# cx,xy,w,h的维度是ncx = (x1 + x2) / 2cy = (y1 + y2) / 2w = x2 - x1h = y2 - y1# torch.stack()沿着新维度对张量进行链接。boxes最开始维度是（n，4），axis=-1表示倒数第一个维度# torch.stack()将(cx, cy, w, h)的维度n将其沿着倒数第一个维度拼接在一起，又是（n，4）boxes = torch.stack((cx, cy, w, h), axis=-1)return boxes#@save
def box_center_to_corner(boxes):"""从（中间，宽度，高度）转换到（左上，右下）"""cx, cy, w, h = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]x1 = cx - 0.5 * wy1 = cy - 0.5 * hx2 = cx + 0.5 * wy2 = cy + 0.5 * hboxes = torch.stack((x1, y1, x2, y2), axis=-1)return boxes

13.4 锚框

目标检测算法通常会在图像中采集大量的样本，本小节介绍其中一个采样办法：以某个像素为中心，生成多个不同缩放比和宽高比的边界框。

13.4.1 生成多个锚框

%matplotlib inline
import torch
from d2l import torch as d2ltorch.set_printoptions(2)  # 精简输出精度，显示小数点后2位
"""
形成多个锚框
params:data：图像(批量大小，通道数，高，宽)sizes：缩放比尺寸集合ratios：宽高比集合"""
def multibox_prior(data, sizes, ratios):# 获取data后两位的值，也就是图像的高和宽in_height, in_width = data.shape[-2:]""" params:device：cpu或者gpunum_sizes：尺寸的个数nnum_ratios：宽高比个数m"""device, num_sizes, num_ratios = data.device, len(sizes), len(ratios)# 以同一像素为中心的锚框数量n+m-1boxes_per_pixel = (num_sizes + num_ratios - 1)size_tensor = torch.tensor(sizes, device=device)ratio_tensor = torch.tensor(ratios, device=device)# offset：为了将锚点移动到像素的中心，需要设置偏移量。# steps：归一化，将宽高规化到0-1之间，因为一个像素的高为1且宽为1，我们选择偏移我们的中心0.5offset_h, offset_w = 0.5, 0.5steps_h = 1.0 / in_height  # 在y轴上缩放步长steps_w = 1.0 / in_width  # 在x轴上缩放步长# 假设宽高512*216 那么torch.arange(in_height, device=device)=【0，1，2...511】,移动到中心就是[0.5,1.5...511.5]# 第一步：torch.arange(in_height, device=device) + offset_h代表移动到每个像素的中心，因为每个像素1*1大小.# 第二步：宽高进行归一化center_h = (torch.arange(in_height, device=device) + offset_h) * steps_hcenter_w = (torch.arange(in_width, device=device) + offset_w) * steps_w"""a = torch.tensor([1, 2, 3, 4])b = torch.tensor([4, 5, 6])x, y = torch.meshgrid(a, b,indexing='ij')print:tensor([[1, 1, 1],[2, 2, 2],[3, 3, 3],[4, 4, 4]])tensor([[4, 5, 6],[4, 5, 6],[4, 5, 6],[4, 5, 6]])x, y = torch.meshgrid(a, b,indexing='xy')print:tensor([[1, 2, 3, 4],[1, 2, 3, 4],[1, 2, 3, 4]])tensor([[4, 4, 4, 4],[5, 5, 5, 5],[6, 6, 6, 6]])"""# 对比上面例子，假设center_h=tensor([0.5,1.5...511.5])(实际上是0-1的值，这里为了简单理解写成这样)  # 则shift_y=tensor([0.5,0.5..],[1.5,1.5,...],...[511.5,511.5...])shift_y, shift_x = torch.meshgrid(center_h, center_w, indexing='ij')   # 将shift展平成一维序列，用上述的例子则shift_y为tensor([0.5,0.5...511.5,511.5])shift_y, shift_x = shift_y.reshape(-1), shift_x.reshape(-1)# 宽=h*s*sqrt(r)# 由于锚框只考虑s1和r1的组合，r1组合就是size_tensor * torch.sqrt(ratio_tensor[0])，s1组合就是sizes[0] * torch.sqrt(ratio_tensor[1:])# 此处要乘上in_height / in_width是因为，假设此时ratios宽高比为1，那么默认w=h，但是实际上ratios代表与原图宽高比一致，举个例子# 假设原图1000*10，那么当ratios为1时，此时w=h，而我们需要的是w/h = 1000/10，所以需要乘上in_height / in_width来与原尺寸保持一致w = torch.cat((size_tensor * torch.sqrt(ratio_tensor[0]),sizes[0] * torch.sqrt(ratio_tensor[1:])))\* in_height / in_width  # 处理矩形输入h = torch.cat((size_tensor / torch.sqrt(ratio_tensor[0]),sizes[0] / torch.sqrt(ratio_tensor[1:])))# 除以2来获得半高和半宽# 每一行(-w, -h, w, h)对应一个锚框一个锚框的左上角偏差和右下角偏差anchor_manipulations = torch.stack((-w, -h, w, h)).T.repeat(in_height * in_width, 1) / 2# 每个中心点都将有boxes_per_pixel=(n+m-1)个锚框，# 形状：（w*h*(n+m-1)， 4）out_grid = torch.stack([shift_x, shift_y, shift_x, shift_y],dim=1).repeat_interleave(boxes_per_pixel, dim=0)output = out_grid + anchor_manipulations# 添加一个维度return output.unsqueeze(0)img = d2l.plt.imread('../data/img/catdog.jpg')
h, w = img.shape[:2] # (1080, 1920)
X = torch.rand(size=(1, 3, h, w))
Y = multibox_prior(X, sizes=[0.75, 0.5, 0.25], ratios=[1, 2, 0.5])
print(Y.shape)

# 即将Y变成（高，宽，以同一像素点为中心的锚框数，4）
# 每个锚框有四个元素（锚框的左上角x，y坐标和锚框右下角的x，y坐标）
# n+m-1=3+3-1=5
boxes = Y.reshape(h, w, 5, 4)
# 访问以（250，250）为中心的第一个锚框
boxes[250, 250, 0, :]

# 显示以某个像素点为中心的所有锚框
"""
params:axes:图像坐标bboxes：某个像素点中心坐标labels：显示文本，例如s=0.2,r=1colors：锚框的颜色
"""
def show_bboxes(axes, bboxes, labels=None, colors=None):"""显示所有边界框"""def _make_list(obj, default_values=None):if obj is None:obj = default_valueselif not isinstance(obj, (list, tuple)):obj = [obj]return objlabels = _make_list(labels)colors = _make_list(colors, ['b', 'g', 'r', 'm', 'c'])for i, bbox in enumerate(bboxes):color = colors[i % len(colors)]# bbox_to_rect将边界框(左上x,左上y,右下x,右下y)格式转换成matplotlib格式：# ((左上x,左上y),宽,高)rect = d2l.bbox_to_rect(bbox.detach().numpy(), color)axes.add_patch(rect)if labels and len(labels) > i:text_color = 'k' if color == 'w' else 'w'axes.text(rect.xy[0], rect.xy[1], labels[i],va='center', ha='center', fontsize=9, color=text_color,bbox=dict(facecolor=color, lw=0))
d2l.set_figsize()
bbox_scale = torch.tensor((w, h, w, h))
fig = d2l.plt.imshow(img)
show_bboxes(fig.axes, boxes[750, 750, :, :] * bbox_scale,['s=0.75, r=1', 's=0.5, r=1', 's=0.25, r=1', 's=0.75, r=2','s=0.75, r=0.5'])

13.4.2 交并比

# 衡量锚框与真实框之间或者锚框与锚框之间的相似度，即A∩B/A∪B
def box_iou(boxes1, boxes2):"""计算两个锚框或边界框列表中成对的交并比"""box_area = lambda boxes: ((boxes[:, 2] - boxes[:, 0]) *(boxes[:, 3] - boxes[:, 1]))# boxes1,boxes2,areas1,areas2的形状:# boxes1：(boxes1的数量,4),# boxes2：(boxes2的数量,4),# areas1：(boxes1的数量,),# areas2：(boxes2的数量,)areas1 = box_area(boxes1)areas2 = box_area(boxes2)# inter_upperlefts,inter_lowerrights,inters的形状:# (boxes1的数量,boxes2的数量,2)inter_upperlefts = torch.max(boxes1[:, None, :2], boxes2[:, :2])inter_lowerrights = torch.min(boxes1[:, None, 2:], boxes2[:, 2:])inters = (inter_lowerrights - inter_upperlefts).clamp(min=0)# inter_areasandunion_areas的形状:(boxes1的数量,boxes2的数量)inter_areas = inters[:, :, 0] * inters[:, :, 1]union_areas = areas1[:, None] + areas2 - inter_areasreturn inter_areas / union_areas

13.4.3 在训练数据中标注锚框

%matplotlib inline
import torch
from d2l import torch as d2ltorch.set_printoptions(2)  # 精简输出精度，显示小数点后2位
"""
形成多个锚框
params:data：图像(批量大小，通道数，高，宽)sizes：缩放比尺寸集合ratios：宽高比集合"""
def multibox_prior(data, sizes, ratios):# 获取data后两位的值，也就是图像的高和宽in_height, in_width = data.shape[-2:]""" params:device：cpu或者gpunum_sizes：尺寸的个数nnum_ratios：宽高比个数m"""device, num_sizes, num_ratios = data.device, len(sizes), len(ratios)# 以同一像素为中心的锚框数量n+m-1boxes_per_pixel = (num_sizes + num_ratios - 1)size_tensor = torch.tensor(sizes, device=device)ratio_tensor = torch.tensor(ratios, device=device)# offset：为了将锚点移动到像素的中心，需要设置偏移量。# steps：归一化，将宽高规化到0-1之间，因为一个像素的高为1且宽为1，我们选择偏移我们的中心0.5offset_h, offset_w = 0.5, 0.5steps_h = 1.0 / in_height  # 在y轴上缩放步长steps_w = 1.0 / in_width  # 在x轴上缩放步长# 假设宽高512*216 那么torch.arange(in_height, device=device)=【0，1，2...511】,移动到中心就是[0.5,1.5...511.5]# 第一步：torch.arange(in_height, device=device) + offset_h代表移动到每个像素的中心，因为每个像素1*1大小.# 第二步：宽高进行归一化center_h = (torch.arange(in_height, device=device) + offset_h) * steps_hcenter_w = (torch.arange(in_width, device=device) + offset_w) * steps_w"""a = torch.tensor([1, 2, 3, 4])b = torch.tensor([4, 5, 6])x, y = torch.meshgrid(a, b,indexing='ij')print:tensor([[1, 1, 1],[2, 2, 2],[3, 3, 3],[4, 4, 4]])tensor([[4, 5, 6],[4, 5, 6],[4, 5, 6],[4, 5, 6]])x, y = torch.meshgrid(a, b,indexing='xy')print:tensor([[1, 2, 3, 4],[1, 2, 3, 4],[1, 2, 3, 4]])tensor([[4, 4, 4, 4],[5, 5, 5, 5],[6, 6, 6, 6]])"""# 对比上面例子，假设center_h=tensor([0.5,1.5...511.5])(实际上是0-1的值，这里为了简单理解写成这样)  # 则shift_y=tensor([0.5,0.5..],[1.5,1.5,...],...[511.5,511.5...])shift_y, shift_x = torch.meshgrid(center_h, center_w, indexing='ij')   # 将shift展平成一维序列，用上述的例子则shift_y为tensor([0.5,0.5...511.5,511.5])shift_y, shift_x = shift_y.reshape(-1), shift_x.reshape(-1)# 宽=h*s*sqrt(r)# 由于锚框只考虑s1和r1的组合，r1组合就是size_tensor * torch.sqrt(ratio_tensor[0])，s1组合就是sizes[0] * torch.sqrt(ratio_tensor[1:])# 此处要乘上in_height / in_width是因为，假设此时ratios宽高比为1，那么默认w=h，但是实际上ratios代表与原图宽高比一致，举个例子# 假设原图1000*10，那么当ratios为1时，此时w=h，而我们需要的是w/h = 1000/10，所以需要乘上in_height / in_width来与原尺寸保持一致w = torch.cat((size_tensor * torch.sqrt(ratio_tensor[0]),sizes[0] * torch.sqrt(ratio_tensor[1:])))\* in_height / in_width  # 处理矩形输入h = torch.cat((size_tensor / torch.sqrt(ratio_tensor[0]),sizes[0] / torch.sqrt(ratio_tensor[1:])))# 除以2来获得半高和半宽# 每一行(-w, -h, w, h)对应一个锚框一个锚框的左上角偏差和右下角偏差anchor_manipulations = torch.stack((-w, -h, w, h)).T.repeat(in_height * in_width, 1) / 2# 每个中心点都将有boxes_per_pixel=(n+m-1)个锚框，# 形状：（w*h*(n+m-1)， 4）out_grid = torch.stack([shift_x, shift_y, shift_x, shift_y],dim=1).repeat_interleave(boxes_per_pixel, dim=0)output = out_grid + anchor_manipulations# 添加一个维度return output.unsqueeze(0)img = d2l.plt.imread('../data/img/catdog.jpg')
h, w = img.shape[:2] # (1080, 1920)
X = torch.rand(size=(1, 3, h, w))
Y = multibox_prior(X, sizes=[0.75, 0.5, 0.25], ratios=[1, 2, 0.5])
print(Y.shape)

# 即将Y变成（高，宽，以同一像素点为中心的锚框数，4）
# 每个锚框有四个元素（锚框的左上角x，y坐标和锚框右下角的x，y坐标）
# n+m-1=3+3-1=5
boxes = Y.reshape(h, w, 5, 4)
# 访问以（250，250）为中心的第一个锚框
boxes[250, 250, 0, :]

# 显示以某个像素点为中心的所有锚框
"""
params:axes:图像坐标bboxes：某个像素点中心坐标labels：显示文本，例如s=0.2,r=1colors：锚框的颜色
"""
def show_bboxes(axes, bboxes, labels=None, colors=None):"""显示所有边界框"""def _make_list(obj, default_values=None):if obj is None:obj = default_valueselif not isinstance(obj, (list, tuple)):obj = [obj]return objlabels = _make_list(labels)colors = _make_list(colors, ['b', 'g', 'r', 'm', 'c'])for i, bbox in enumerate(bboxes):color = colors[i % len(colors)]# bbox_to_rect将边界框(左上x,左上y,右下x,右下y)格式转换成matplotlib格式：# ((左上x,左上y),宽,高)rect = d2l.bbox_to_rect(bbox.detach().numpy(), color)axes.add_patch(rect)if labels and len(labels) > i:text_color = 'k' if color == 'w' else 'w'axes.text(rect.xy[0], rect.xy[1], labels[i],va='center', ha='center', fontsize=9, color=text_color,bbox=dict(facecolor=color, lw=0))
d2l.set_figsize()
bbox_scale = torch.tensor((w, h, w, h))
fig = d2l.plt.imshow(img)
show_bboxes(fig.axes, boxes[750, 750, :, :] * bbox_scale,['s=0.75, r=1', 's=0.5, r=1', 's=0.25, r=1', 's=0.75, r=2','s=0.75, r=0.5'])

# 衡量锚框与真实框之间或者锚框与锚框之间的相似度，即A∩B/A∪B
def box_iou(boxes1, boxes2):"""计算两个锚框或边界框列表中成对的交并比"""box_area = lambda boxes: ((boxes[:, 2] - boxes[:, 0]) *(boxes[:, 3] - boxes[:, 1]))# boxes1,boxes2,areas1,areas2的形状:# boxes1：(boxes1的数量,4),# boxes2：(boxes2的数量,4),# areas1：(boxes1的数量,),# areas2：(boxes2的数量,)areas1 = box_area(boxes1)areas2 = box_area(boxes2)# inter_upperlefts,inter_lowerrights,inters的形状:# (boxes1的数量,boxes2的数量,2)inter_upperlefts = torch.max(boxes1[:, None, :2], boxes2[:, :2])inter_lowerrights = torch.min(boxes1[:, None, 2:], boxes2[:, 2:])inters = (inter_lowerrights - inter_upperlefts).clamp(min=0)# inter_areasandunion_areas的形状:(boxes1的数量,boxes2的数量)inter_areas = inters[:, :, 0] * inters[:, :, 1]union_areas = areas1[:, None] + areas2 - inter_areasreturn inter_areas / union_areas
# 将最接近的真实边界框分配给锚框
# iou_threshold：阈值
def assign_anchor_to_bbox(ground_truth, anchors, device, iou_threshold=0.5):# num_anchors=na num_gt_boxes=nbnum_anchors, num_gt_boxes = anchors.shape[0], ground_truth.shape[0]# 位于第i行和第j列的元素x_ij是锚框i和真实边界框j的IoUjaccard = box_iou(anchors, ground_truth)# 对于每个锚框，分配的真实边界框的张量，初始值为-1anchors_bbox_map = torch.full((num_anchors,), -1, dtype=torch.long,device=device)# 找到每一行中最大交并比的ground_truth和anchors索引号max_ious, indices = torch.max(jaccard, dim=1)# 找到剩余交并比大于阈值的索引号anc_i = torch.nonzero(max_ious >= iou_threshold).reshape(-1)box_j = indices[max_ious >= iou_threshold]anchors_bbox_map[anc_i] = box_j# 删去这些索引行和列col_discard = torch.full((num_anchors,), -1)row_discard = torch.full((num_gt_boxes,), -1)for _ in range(num_gt_boxes):max_idx = torch.argmax(jaccard)box_idx = (max_idx % num_gt_boxes).long()anc_idx = (max_idx / num_gt_boxes).long()anchors_bbox_map[anc_idx] = box_idxjaccard[:, box_idx] = col_discardjaccard[anc_idx, :] = row_discardreturn anchors_bbox_map
#@save
def offset_boxes(anchors, assigned_bb, eps=1e-6):"""对锚框偏移量的转换"""c_anc = d2l.box_corner_to_center(anchors)c_assigned_bb = d2l.box_corner_to_center(assigned_bb)offset_xy = 10 * (c_assigned_bb[:, :2] - c_anc[:, :2]) / c_anc[:, 2:]offset_wh = 5 * torch.log(eps + c_assigned_bb[:, 2:] / c_anc[:, 2:])offset = torch.cat([offset_xy, offset_wh], axis=1)return offset
#@save
def multibox_target(anchors, labels):"""使用真实边界框标记锚框"""batch_size, anchors = labels.shape[0], anchors.squeeze(0)batch_offset, batch_mask, batch_class_labels = [], [], []device, num_anchors = anchors.device, anchors.shape[0]for i in range(batch_size):label = labels[i, :, :]anchors_bbox_map = assign_anchor_to_bbox(label[:, 1:], anchors, device)bbox_mask = ((anchors_bbox_map >= 0).float().unsqueeze(-1)).repeat(1, 4)# 将类标签和分配的边界框坐标初始化为零class_labels = torch.zeros(num_anchors, dtype=torch.long,device=device)assigned_bb = torch.zeros((num_anchors, 4), dtype=torch.float32,device=device)# 使用真实边界框来标记锚框的类别。# 如果一个锚框没有被分配，标记其为背景（值为零）indices_true = torch.nonzero(anchors_bbox_map >= 0)bb_idx = anchors_bbox_map[indices_true]class_labels[indices_true] = label[bb_idx, 0].long() + 1assigned_bb[indices_true] = label[bb_idx, 1:]# 偏移量转换offset = offset_boxes(anchors, assigned_bb) * bbox_maskbatch_offset.append(offset.reshape(-1))batch_mask.append(bbox_mask.reshape(-1))batch_class_labels.append(class_labels)bbox_offset = torch.stack(batch_offset)bbox_mask = torch.stack(batch_mask)class_labels = torch.stack(batch_class_labels)return (bbox_offset, bbox_mask, class_labels)ground_truth = torch.tensor([[0, 0.1, 0.08, 0.52, 0.92],[1, 0.55, 0.2, 0.9, 0.88]])
anchors = torch.tensor([[0, 0.1, 0.2, 0.3], [0.15, 0.2, 0.4, 0.4],[0.63, 0.05, 0.88, 0.98], [0.66, 0.45, 0.8, 0.8],[0.57, 0.3, 0.92, 0.9]])