1.首先需要了解一下Pytorch的DataLoader, DataSet, Sampler之间的关联
其中DataSet类来自from torch.utils.data import Dataset， 指针对输入数据提供读取方式，数据增强等处理手段。直接继承DataSet实现就可以，主要需要实现以下三个方法，构造函数定义数据处理需要的各种参数，def getitem(self, index)允许通过 dataset[index]的方式访问第index个数据，而对数据处理和增强的主要代码都写在这里，def len(self)允许通过len(dataset)返回dataset对象的长度，即样本个数。而这里的主要参数index就来自于DataLoader中的Sampler。
class Dataset(object):
def __init__(self):
...
def __getitem__(self, index):
return ...
def __len__(self):
return ...
DataLoader的定义如下，pytorch中的数据加载模块 Dataloader，使用Sampler或者batch_sampler（都由生成器实现）来返回数据的索引，使用迭代器 来返回需要的张量数据（ 其中输入为dataset和sampler获取的索引 indices，例如：batch = self.collate_fn([self.dataset[i] for i in indices])），可以在大量数据情况下，实现小批量循环迭代式的读取，避免了内存不足问题
关于迭代器，生成器等概念以及其在Dataloader中的使用可以参考这里
class DataLoader(object):
def __init__(self, dataset, batch_size=1, shuffle=False, sampler=None,
batch_sampler=None, num_workers=0, collate_fn=default_collate,
pin_memory=False, drop_last=False, timeout=0,
worker_init_fn=None)
关于Pytorch的DataLoader, DataSet, Sampler之间的关联，更详细的可以参考别人的博客：
点击这里，已经有人讲的很详细，就不多赘述，下面直接分析代码。
创建dataloader 的代码如下：
def create_dataloader(path, img_size, batch_size, stride, opt, hyp=None, augment=False, cache=False, pad=0.0, rect=False, local_rank=-1, world_size=-1):
"""
path：Txt file containing picture path or folder path containing picture
img_size：Network input picture size
batch_size: batch size
stride：Maximum total stride size of down sampling
opt：input parameter when use train.py in terminal
hyp：hyper parameter, Here we mainly use some coefficients about data augmentation (rotation, translation, etc.)
augment：data augmentation or not
cache：Whether to cache pictures to memory in advance to speed up training
pad：padding when setting shape of rectangle training
rect：Rectangular validation or not
local_rank: DDP mode or not
world_size: Number of pc， set in Multi pc and multi GPU
"""
with torch_distributed_zero_first(local_rank):
dataset = LoadImagesAndLabels(path, img_size, batch_size,
augment=augment, # augment images
hyp=hyp, # augmentation hyperparameters
rect=rect, # rectangular validation
cache_images=cache,
single_cls=opt.single_cls,
stride=int(stride),
pad=pad)
batch_size = min(batch_size, len(dataset))
num_workers = min([os.cpu_count() // world_size, batch_size if batch_size > 1 else 0, 8])
train_sampler = torch.utils.data.distributed.DistributedSampler(dataset) if local_rank != -1 else None
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
num_workers=num_workers,
sampler=train_sampler,
pin_memory=True,
collate_fn=LoadImagesAndLabels.collate_fn)
return dataloader, dataset
（1）.先了解上下文管理器的相关机制
关于@contextmanager 装饰器的主要机理可参考这里
其中local_rank=-1表示不是DDP模式，local_rank ！=-1表示DDP模式，其中local_rank=0表示主进程，其他为从进程。
而在这里代码运行的流程如下：
a. 进入 torch_distributed_zero_first内部 -> b. 运行yield之前的代码 -> c. 返回with中，执行LoadImagesAndLabels() -> d. 进入torch_distributed_zero_first内部，接着运行yield后面的代码
with torch_distributed_zero_first(local_rank):
dataset = LoadImagesAndLabels(path, img_size, batch_size,
augment=augment, # augment images
hyp=hyp, # augmentation hyperparameters
rect=rect, # rectangular validation
cache_images=cache,
single_cls=opt.single_cls,
stride=int(stride),
pad=pad)
@contextmanager
def torch_distributed_zero_first(local_rank: int):
"""
Decorator to make all processes in distributed training wait for each local_master to do something.
"""
if local_rank not in [-1, 0]:
torch.distributed.barrier()
yield
if local_rank == 0:
torch.distributed.barrier()
在上面的代码示例中，如果执行create_dataloader()函数的进程不是主进程，即rank不等于0或者-1，上下文管理器会执行相应的torch.distributed.barrier()，设置一个阻塞栅栏，让此进程处于等待状态，等待所有进程到达栅栏处（包括主进程数据处理完毕）；如果执行create_dataloader()函数的进程是主进程，其会直接去读取数据并处理，然后其处理结束之后会接着遇到torch.distributed.barrier()，此时，所有进程都到达了当前的栅栏处，这样所有进程就达到了同步，并同时得到释放，这里讲解来自视觉弘毅
（2）关于torch.utils.data.DataLoade的参数
num_workers即处理输入数据的进程数，一般根据cpu的性能来设置
train_sampler 如果是DDP模式需要设置DistributedSampler，否则不设置
pin_memory=True，这里可以简单的理解为加速。
collate_fn主要是用来将一个batch中的数据打包在一起。 batch = self.collate_fn([self.dataset[i] for i in indices])
这里关于参数需要注意的是DataLoader的部分初始化参数之间存在互斥关系：
如果你自定义了batch_sampler,那么这些参数都必须使用默认值：batch_size, shuffle,sampler,drop_last.
如果你自定义了sampler，那么shuffle需要设置为False
如果sampler和batch_sampler都为None,那么batch_sampler使用Pytorch已经实现好的BatchSampler,而sampler分两种情况：
若shuffle=True,则sampler=RandomSampler(dataset)
若shuffle=False,则sampler=SequentialSampler(dataset)
batch_size = min(batch_size, len(dataset))
num_workers = min([os.cpu_count() // world_size, batch_size if batch_size > 1 else 0, 8])
train_sampler = torch.utils.data.distributed.DistributedSampler(dataset) if local_rank != -1 else None
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
num_workers=num_workers,
sampler=train_sampler,
pin_memory=True,
collate_fn=LoadImagesAndLabels.collate_fn)
- class LoadImagesAndLabels(Dataset)
（1） 构造函数，定义参数和部分预处理
def __init__(self, path, img_size=640, batch_size=16, augment=False, hyp=None, rect=False, image_weights=False,
cache_images=False, single_cls=False, stride=32, pad=0.0):
# --------------------------------------------------------------------------------------------------------------
# define img_files(images path list) 保存图像路径列表
try:
files = [] # images path list
for p in path if isinstance(path, list) else [path]:
parent_dir = str(Path(p).parent) + os.sep
dir = str(Path(p))
if os.path.isfile(dir): # file
with open(dir, 'r') as f:
img_dir_list = f.read().splitlines()
files += [i.replace('./', parent_dir) if i.startswith('./') else i for i in img_dir_list]
elif os.path.isdir(dir): # folder
files += glob.iglob(dir + os.sep + '*.*') # [] + generater 将遍历迭代器中所有元素并添加到[]中
else:
raise Exception('Not exist file dir: %s.' % dir)
self.img_files = sorted([i.replace('/', os.sep)
for i in files if os.path.splitext(i)[-1].lower() in img_formats])
except Exception as e:
raise Exception('Error loading data from %s: %s\nSee %s' % (path, e, help_url))
num_files = len(self.img_files)
assert num_files > 0, 'No images found in %s. See %s' % (path, help_url)
# --------------------------------------------------------------------------------------------------------------
# define label_files(labels path list) 保存标签路径列表
self.label_files = [i.replace('images', 'labels').replace(os.path.splitext(i)[-1], 'txt')
for i in self.img_files]
# --------------------------------------------------------------------------------------------------------------
# define batch index, number of batches 后续Rectangular train/val/test/detection 会用
batch_index = np.array([i // batch_size for i in range(num_files)], dtype=np.int)
num_batches = batch_index[-1] + 1
# --------------------------------------------------------------------------------------------------------------
# parameters
self.num_files = num_files
self.batch = batch_index # batch index of image
self.img_size = img_size
self.augment = augment
self.hyp = hyp # 超参数，这里主要是图像空间增强（平移缩放旋转错切）等参数和色域增强（hsv）
self.image_weights = image_weights
self.rect = False if image_weights else rect # Rectangular train 和 mosaic train只能二选一
self.mosaic = self.augment and not self.rect # load 4 images at a time into a mosaic (only during training)
self.mosaic_border = [-img_size // 2, -img_size // 2] # mosaic增强时使用
self.stride = stride
# --------------------------------------------------------------------------------------------------------------
# Check cache cache[img]=[label, shape] 首次读入label则缓存label到 label.cache中
cache_dir = str(Path(self.label_files[0]).parent) + '.cache' # cached labels
if os.path.isfile(cache_dir):
cache = torch.load(cache_dir)
if cache['hash'] != get_hash(self.img_files + self.label_files): # dataset changed
cache = self.cache_labels(cache_dir) # re-cache
else:
cache = self.cache_labels(cache_dir) # cache
# --------------------------------------------------------------------------------------------------------------
# Get labels
labels, shapes = zip(*[cache[img] for img in self.img_files])
self.shapes = np.array(shapes, dtype=np.float64)
self.labels = labels
# --------------------------------------------------------------------------------------------------------------
# Rectangular Training https://github.com/ultralytics/yolov3/issues/232
if self.rect:
s = self.shapes # 这里是用Image获取的shape，为w,h，cv2读入的shape顺序为h,w
aspect_ratio = s[:, 1] / s[:, 0] # h/w # 将所有数据按照图像aspect ratio从小到大排列
aspect_ratio_index = np.argsort(aspect_ratio)
shape_batch = np.array([[1, 1]] * num_batches)
self.labels = self.labels[aspect_ratio_index]
self.img_files = self.img_files[aspect_ratio_index]
self.label_files = self.label_files[aspect_ratio_index]
aspect_ratio = aspect_ratio[aspect_ratio_index]
for batch in range(num_batches):
aspect_ratio_batch = aspect_ratio[batch_index == batch] # 获取一个batch的aspect_ratio
ar_min = aspect_ratio_batch.min()
ar_max = aspect_ratio_batch.max()
# 计算batch中ar的最小值和最大值，如果最大值比1还小，即所有图像的h都小于w，则将整个batch的shape设置为h_batch,w_batch=（h/w, 1）
# 如果最小值最1还大，即所有图形的h都大于w，则将整个batch的shape设置为h_batch,w_batch=（1, w/h）
# 如果不符合以上条件则h_batch,w_batch=（1, 1）
if ar_max < 1:
shape_batch[batch] = [ar_max, 1]
elif ar_min > 1:
shape_batch[batch] = [1, 1 / ar_min]
# 使用输入的img_size获取每个batch的图像尺寸，作为 Rectangular Training需要的shape，确保为stride的倍数
self.batch_shapes = [np.ceil(shape_batch[i] * img_size / stride).astype(np.int) * stride
for i in range(num_batches)]
# --------------------------------------------------------------------------------------------------------------
# Cache labels
num_missing, num_found, num_empty, num_duplicate = 0, 0, 0, 0 # number missing, found, empty, duplicate
pbar = tqdm(self.label_files)
for i, file in enumerate(pbar):
l = self.shapes[i]
if l.shape[0]:
assert l.shape[1] == 5, '> 5 label columns: %s' % file
assert (l[:, 1:] <= 1).all(), 'non-normalized or out of bounds coordinate labels: %s' % file
assert (l >= 0).all(), 'negative labels: %s' % file
if np.unique(l, axis=0).shape[0] < l.shape[0]: # duplicate rows
num_duplicate += 1
if single_cls:
l[:, 0] = 0 # force dataset into single-class mode
self.shapes[i] = l
num_found += 1
else:
num_empty += 1
pbar.desc = 'Scanning labels %s (%g found, %g missing, %g empty, %g duplicate, for %g images)' % (
cache_dir, num_found, num_missing, num_empty, num_duplicate, num_files)
if num_found == 0:
s = 'WARNING: No labels found in %s. See %s' % (os.path.dirname(cache_dir) + os.sep, help_url)
print(s)
assert not augment, '%s. Can not train without labels.' % s
# --------------------------------------------------------------------------------------------------------------
# Cache images into memory for faster training (WARNING: large datasets may exceed system RAM)
self.imgs = [None] * num_files
if cache_images:
gb = 0 # Gigabytes of cached images
pbar = tqdm(range(len(self.img_files)), desc='Caching images')
self.img_hw0, self.img_hw = [None] * num_files, [None] * num_files
for i in pbar:
self.imgs[i], self.img_hw0[i], self.img_hw[i] = self.load_image(i)
gb += self.imgs[i].nbytes
pbar.desc = 'Caching images (%.1fGB)' % (gb / 1E9)
# --------------------------------------------------------------------------------------------------------------
（2）def getitem(self, index)， 图像增强
def __len__(self):
return len(self.img_files)
def __getitem__(self, index):
# --------------------------------------------------------------------------------------------------------------
# data augmentation,
# if train, use Mosaic or Rectangular(when self.rect=true),
# if val or test, use Rectangular, letterbox(but not minimum rectangle)
# if detection, use Rectangular, letterbox(minimum rectangle) 只有推理时才使用最小矩形框，加速推理
if self.mosaic:
img, label = self.load_mosaic(index)
shapes = None
else:
img, (h0, w0), (h, w) = self.load_image(index)
shape = self.batch_shapes[self.batch[index]] if self.rect else self.img_size # final letterboxed shape
img, ratio, (dw, dh) = letterbox(img, shape, auto=False, scaleup=self.augment)
shapes = (h0, w0), ((h / h0, w / w0), (dw, dh)) # for COCO mAP rescaling ??????
# according to letterbox change corresponding label
x = self.labels[index]
label = x.copy()
if x.size > 0:
# Normalized xywh to xywh ratio * w， ratio * h表示rect后的矩形框尺寸,dw和dh表示图像四周填充的边框
label[:, 1] = ratio * w * x[:, 1] + dw # pad width
label[:, 2] = ratio * h * x[:, 2] + dh # pad height
label[:, 3] = ratio * w * x[:, 3]
label[:, 4] = ratio * h * x[:, 4]
# --------------------------------------------------------------------------------------------------------------
# image augmentation, image space 缩放，平移，旋转，错切
if self.augment:
if not self.mosaic:
img, label = random_affine(img, label, self.hyp['degrees'], self.hyp['translate'],
self.hyp['scale'], self.hyp['shear'])
# --------------------------------------------------------------------------------------------------------------
# image augmentation, color space 色域转换
augment_hsv(img, self.hyp['hsv_h'], self.hyp['hsv_s'], self.hyp['hsv_v'])
# --------------------------------------------------------------------------------------------------------------
# label normalization
if label.size > 0:
label[:, 1] = label[:, 1] / img.shape[1] # x
label[:, 2] = label[:, 2] / img.shape[0] # y
label[:, 3] = label[:, 3] / img.shape[1] # w
label[:, 4] = label[:, 4] / img.shape[0] # h
# --------------------------------------------------------------------------------------------------------------
# image augmentation, flip
if self.augment:
# random left-right flip
lr_flip = True
if lr_flip and random.random() < 0.5:
img = np.fliplr(img)
if label.size > 0:
label[:, 1] = 1 - label[:, 1]
# random up-down flip
ud_flip = False
if ud_flip and random.random() < 0.5:
img = np.flipud(img)
if label.size > 0:
label[:, 2] = 1 - label[:, 2]
# --------------------------------------------------------------------------------------------------------------
# output labels新增一列用来存储在每个batch中每个图像的索引
num_label = len(label)
labels = torch.zeros((num_label, 6))
if num_label:
labels[:, 1:] = torch.from_numpy(label)
img = img[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, (h, w, 3) to (3, h, w)
img = np.ascontiguousarray(img)
#此处的返回值为单张图像数据，一个打包成batch的操作在collate_fn中完成
return torch.from_numpy(img), labels, self.img_files[index], shapes
# --------------------------------------------------------------------------------------------------------------
（3） def collate_fn(batch)， torch.utils.data.DataLoade中负责打包一个batch数据的函数，当_getitem__返回的数据不止image,label时，需要重写。
@staticmethod
def collate_fn(batch):
# merge a batch images and labels
batch_imgs, batch_labels, batch_paths, batch_shapes = zip(*batch)
for i, label in enumerate(batch_labels):
if label.size[0]:
label[:, 0] = i
return torch.stack(batch_imgs, dim=0), torch.cat(batch_labels, dim=0), batch_paths, batch_shapes
（4）mosaic
def load_mosaic(self, index):
# mosaic for image , mosaic means combining four images to a new image
# --------------------------------------------------------------------------------------------------------------
label_mosaic = [] # label after mosaic
indices = [index] + [random.randint(0, self.num_files-1) for _ in range(3)] # current image and other 3 images
# --------------------------------------------------------------------------------------------------------------
# (x_cross, y_cross) is cross point of four images, 四张图像合并的交点
# Range(0.5*img_mosaic_w ~ 1.5*img_mosaic_w, 0.5*img_mosaic_h ~ 1.5*img_mosaic_h)
y_cross, x_cross = [int(random.uniform(-x, 2 * self.img_size + x)) for x in self.mosaic_border]
# --------------------------------------------------------------------------------------------------------------
# put four small pictures into large mosaic picture
for i, index in enumerate(indices):
# ----------------------------------------------------------------------------------------------------------
img, _, (h, w) = self.load_image(index) # current image
# (x1_mosaic, y1_mosaic, x2_mosaic, y2_mosaic) means (x_min,y_min,x_max,y_max) in mosaic image
# (x1, y1, x2, y2) means (x_min,y_min,x_max,y_max) in one of four small images
if i == 0: # upper left image
# ------------------------------------------------------------------------------------------------------
# define mosaic image
img_mosaic = np.full((2 * self.img_size, 2 * self.img_size, img.shape[2]), 114, dtype=np.uint8)
# ------------------------------------------------------------------------------------------------------
x1_mosaic, y1_mosaic, x2_mosaic, y2_mosaic = max(x_cross-w, 0), max(y_cross-h, 0), x_cross, y_cross
x1, y1, x2, y2 = w - (x2_mosaic-x1_mosaic), h - (y2_mosaic-y1_mosaic), w, h
if i == 1: # upper right image
x1_mosaic, y1_mosaic, x2_mosaic, y2_mosaic = x_cross, max(y_cross-h, 0), \
min(x_cross+w, 2*self.img_size), y_cross
# 这里原来的代码中 x2 = min(w, x2_mosaic-x1_mosaic), 我认为x2_mosaic-x1_mosaic永远不会大于w,所以
#没有必要这样设置，直接x2 =x2_mosaic-x1_mosaic就好，后面有好几处也是这个道理
x1, y1, x2, y2 = 0, h - (y2_mosaic-y1_mosaic), x2_mosaic-x1_mosaic, h
if i == 2: # bottom left image
x1_mosaic, y1_mosaic, x2_mosaic, y2_mosaic = max(x_cross-w, 0), y_cross, \
x_cross, min(y_cross+h, 2*self.img_size)
x1, y1, x2, y2 = w-(x2_mosaic-x1_mosaic), 0, w, y2_mosaic-y1_mosaic
if i == 3: # bottom right image
x1_mosaic, y1_mosaic, x2_mosaic, y2_mosaic = x_cross, y_cross, min(x_cross+w, 2*self.img_size), \
min(y_cross+h, 2*self.img_size)
x1, y1, x2, y2 = 0, 0, x2_mosaic-x1_mosaic, y2_mosaic-y1_mosaic
img_mosaic[y1_mosaic:y2_mosaic, x1_mosaic:x2_mosaic] = img[y1:y2, x1:x2]
# ----------------------------------------------------------------------------------------------------------
label = self.labels[index]
if label.size > 0:
# normalized xywh to xyxy
x = label.copy()
label[:, 1] = (x[:, 1] - x[:, 3] / 2) * w
label[:, 2] = (x[:, 2] - x[:, 4] / 2) * h
label[:, 3] = (x[:, 1] + x[:, 3] / 2) * w
label[:, 4] = (x[:, 2] + x[:, 4] / 2) * h
# 利用（x1,y1）和 （x1_mosaic，y1_mosaic ）是相同的顶点来转换坐标系
# distance between label and (x1, y1)
delta_x1 = label[:, 1] - x1
delta_y1 = label[:, 2] - y1
delta_x2 = label[:, 3] - x1
delta_y2 = label[:, 4] - y1
# convert label coordinate system from img to img_mosaic
label[:, 1] = delta_x1 + x1_mosaic
label[:, 2] = delta_y1 + y1_mosaic
label[:, 3] = delta_x2 + x1_mosaic
label[:, 4] = delta_y2 + y1_mosaic
label_mosaic.append(label)
# --------------------------------------------------------------------------------------------------------------
if len(label_mosaic):
label_mosaic = np.concatenate(label_mosaic, axis=0)
# clip label_mosaic
np.clip(label_mosaic[:, 1:], 0, 2 * self.img_size, out=label_mosaic[:, 1:])
# ----------------------------------------------------------------------------------------------------------
# Augment
img_mosaic, label_mosaic = random_affine(img_mosaic, label_mosaic,
degrees=self.hyp['degrees'],
translate=self.hyp['translate'],
scale=self.hyp['scale'],
shear=self.hyp['shear'],
border=self.mosaic_border) # border to remove
return img_mosaic, label_mosaic
（5）load image
def load_image(self, index):
# loads 1 image from dataset, returns img, original hw, resized hw
img = self.imgs[index]
if img is None: # not cached
path = self.img_files[index]
img0 = cv2.imread(path) # BGR
assert img0 is not None, 'Image Not Found ' + path
h0, w0 = img0.shape[0:2] # orig hw
ratio = self.img_size / max(h0, w0) # resize image to img_size
if ratio != 1: # always resize down, only resize up if training with augmentation
interp = cv2.INTER_AREA if ratio < 1 and not self.augment else cv2.INTER_LINEAR
h, w = int(h0 * ratio), int(w0 * ratio)
img = cv2.resize(img0, (w, h), interpolation=interp)
else:
h, w = h0, w0
img = img0
return img, (h0, w0), (h, w) # img, hw_original, hw_resized
else:
return img, self.img_hw0[index], self.img_hw[index] # img, hw_original, hw_resized
（6） letterbox， yolo v5特有的矩形训练/测试/推理，其中推理时采用最小矩形框，可减少计算量，增加速度。
关于矩形推理的原理可以参考江大白在yolo v5讲解中的说明
def letterbox(img, new_shape=(640, 640), color=(114, 114, 114), auto=False, scaleFill=False, scaleup=True):
shape = img.shape[:2] # h,w
if isinstance(new_shape, int):
new_shape = (new_shape, new_shape) # h,w
# Scale ratio (new / old)
ratio = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
if not scaleup: # only scale down for val/test/detection
ratio = min(ratio, 1.0)
# pad: image padding in w and h
new_shape_pad = int(round(shape[1] * ratio)), int(round(shape[0] * ratio)) # w,h
pad = new_shape[1] - new_shape_pad[0], new_shape[0] - new_shape_pad[1] # w,h
if auto: # minimum rectangle
pad = np.mod(pad[0], 32), np.mod(pad[1], 32)
pad /= 2 # divide padding into 2 sides
if shape[::-1] != new_shape_pad:
img = cv2.resize(img, new_shape_pad, interpolation=cv2.INTER_LINEAR)
top, bottom = int(round(pad[1] - 0.1)), int(round(pad[1] + 0.1))
left, right = int(round(pad[0] - 0.1)), int(round(pad[0] + 0.1))
img = cv2.copyMakeBorder(img, top, bottom, left, right, borderType=cv2.BORDER_CONSTANT, value=color)
return img, ratio, (pad[0], pad[1])
1万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
