开源
自从yolov5开源以来，(不管因为啥原因)深受瞩目，我最近用tensorflow实现了其主要部分。可能是第一个纯正的tensorfow2版本，欢迎try and star：
github.com/LongxingTan…
之前在工作中接触过yolov3(跑过demo应该就算接触过了)，效果惊艳。我在视觉领域只是个新人(悲伤的是我一个中年人却在哪儿哪儿都TM是新人)，能力有限，疏漏难免。从头开始实现，对我来说是一次不错的经历，遇到和解决了一些细节问题。
如readme中所指出的，主要特点如下：
纯tensorflow2的实现
用yaml文件来配置模型和控制模型大小
支持自定义数据训练
马赛克数据增强
通过iou或长宽比匹配anchor
相邻正样本增强
支持多GPU训练
相对详细的代码注释
缺点多，提升空间巨大
原理
结合代码简单回顾一下其主要原理和改进。知乎上已经有不少非常不错的解析文章可以参考，尤其是下面几篇。如有可能，直接读代码应该是更清晰、细节。
江大白：深入浅出Yolo系列之Yolov5核心基础知识完整讲解
深度眸：进击的后浪yolov5深度可视化解析
目标检测：Yolov5集百家之长
来自 @江大白 的模型图
通过长宽比或iou匹配anchor
这里的实现了v3中根据iou分配anchor，以及v4/v5中根据宽高比来分配anchor，新的匹配用来解决物体靠近网格时的敏感问题，并可以涨点。
def assign_criterion_wh(self, gt_wh, anchors, anchor_threshold): # return: please note that the v5 default anchor_threshold is 4.0, related to the positive sample augment gt_wh = tf.expand_dims(gt_wh, 0) # => 1 * n_gt * 2 anchors = tf.expand_dims(anchors, 1) # => n_anchor * 1 * 2 ratio = gt_wh / anchors # => n_anchor * n_gt * 2 matched_matrix = tf.reduce_max(tf.math.maximum(ratio, 1 / ratio), axis=2) < anchor_threshold # => n_anchor * n_gt return matched_matrixdef assign_criterion_iou(self, gt_wh, anchors, anchor_threshold): # by IOU, anchor_threshold < 1 box_wh = tf.expand_dims(gt_wh, 0) # => 1 * n_gt * 2 box_area = box_wh[..., 0] * box_wh[..., 1] # => 1 * n_gt anchors = tf.cast(anchors, tf.float32) # => n_anchor * 2 anchors = tf.expand_dims(anchors, 1) # => n_anchor * 1 * 2 anchors_area = anchors[..., 0] * anchors[..., 1] # => n_anchor * 1 inter = tf.math.minimum(anchors[..., 0], box_wh[..., 0]) * tf.math.minimum(anchors[..., 1], box_wh[..., 1]) # n_gt * n_anchor iou = inter / (anchors_area + box_area - inter + 1e-9) iou = iou > anchor_threshold return iou复制代码
正样本增强
正负样本的均衡一直是目标检测领域要解决的问题。v5中根据长宽比匹配到anchor之后，进一步将匹配到的网格的上下或左右最接近的邻居也增强为正样本。匹配的anchor以及坐标与初始匹配到的网格相同
def enrich_pos_by_position(self, assigned_label, assigned_anchor, gain, matched_matrix, rect_style='rect4'): # using offset to extend more postive result, if x assigned_xy = assigned_label[..., 0:2] # n_matched * 2 offset = tf.constant([[0, 0], [1, 0], [0, 1], [-1, 0], [0, -1]], tf.float32) grid_offset = tf.zeros_like(assigned_xy) if rect_style == 'rect2': g = 0.2 # offset elif rect_style == 'rect4': g = 0.5 # matched = (assigned_xy % 1. < g) & (assigned_xy > 1.) matched_left = matched[:, 0] matched_up = matched[:, 1] matched = (assigned_xy % 1. > (1 - g)) & (assigned_xy < tf.expand_dims(gain[0:2], 0) - 1.) matched_right = matched[:, 0] matched_down = matched[:, 1] assigned_anchor = tf.concat([assigned_anchor, assigned_anchor[matched_left], assigned_anchor[matched_up], assigned_anchor[matched_right], assigned_anchor[matched_down]], axis=0) assigned_label = tf.concat([assigned_label, assigned_label[matched_left], assigned_label[matched_up], assigned_label[matched_right], assigned_label[matched_down]], axis=0) grid_offset = g * tf.concat( [grid_offset, grid_offset[matched_left] + offset[1], grid_offset[matched_up] + offset[2], grid_offset[matched_right] + offset[3], grid_offset[matched_down] + offset[4]], axis=0) return assigned_label, assigned_anchor, grid_offset复制代码
通过yaml文件控制模型大小
这里借鉴了efficientdet的思想，通过两个系数控制模型大小。
def parse_model(self, yaml_dict): anchors, nc, depth_multiple, width_multiple = yaml_dict['anchors'], yaml_dict['nc'], yaml_dict['depth_multiple'], yaml_dict['width_multiple'] num_anchors = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors output_dims = num_anchors * (nc + 5) layers = [] # # from, number, module, args for i, (f, number, module, args) in enumerate(yaml_dict['backbone'] + yaml_dict['head']): module = eval(module) if isinstance(module, str) else module # all component is a Class, initialize here, call in self.forward for j, arg in enumerate(args): try: args[j] = eval(arg) if isinstance(arg, str) else arg # eval strings, like Detect(nc, anchors) except: pass number = max(round(number * depth_multiple), 1) if number > 1 else number # control the model scale, s/m/l/x if module in [Conv2D, Conv, Bottleneck, SPP, DWConv, Focus, BottleneckCSP, BottleneckCSP2, SPPCSP, VoVCSP]: c2 = args[0] c2 = math.ceil(c2 * width_multiple / 8) * 8 if c2 != output_dims else c2 args = [c2, *args[1:]] if module in [BottleneckCSP, BottleneckCSP2, SPPCSP, VoVCSP]: args.insert(1, number) number = 1 modules = tf.keras.Sequential(*[module(*args) for _ in range(number)]) if number > 1 else module(*args) modules.i, modules.f = i, f layers.append(modules) return layers复制代码
损失函数匹配
class YoloLoss(object): def __init__(self, anchors, ignore_iou_threshold, num_classes, img_size, label_smoothing=0): self.anchors = anchors self.strides = [8, 16, 32] self.ignore_iou_threshold = ignore_iou_threshold self.num_classes = num_classes self.img_size = img_size self.bce_conf = tf.keras.losses.BinaryCrossentropy(reduction=tf.keras.losses.Reduction.NONE) self.bce_class = tf.keras.losses.BinaryCrossentropy(reduction=tf.keras.losses.Reduction.NONE, label_smoothing=label_smoothing) def __call__(self, y_true, y_pred): iou_loss_all = obj_loss_all = class_loss_all = 0 balance = [1.0, 1.0, 1.0] if len(y_pred) == 3 else [4.0, 1.0, 0.4, 0.1] # P3-5 or P3-6 for i, (pred, true) in enumerate(zip(y_pred, y_true)): # preprocess, true: batch_size * grid * grid * 3 * 6, pred: batch_size * grid * grid * clss+5 true_box, true_obj, true_class = tf.split(true, (4, 1, -1), axis=-1) pred_box, pred_obj, pred_class = tf.split(pred, (4, 1, -1), axis=-1) if tf.shape(true_class)[-1] == 1 and self.num_classes > 1: true_class = tf.squeeze(tf.one_hot(tf.cast(true_class, tf.dtypes.int32), depth=self.num_classes, axis=-1), -2) # prepare: higher weights to smaller box, true_wh should be normalized to (0,1) box_scale = 2 - 1.0 * true_box[..., 2] * true_box[..., 3] / (self.img_size ** 2) obj_mask = tf.squeeze(true_obj, -1) # # obj or noobj, batch_size * grid * grid * anchors_per_grid background_mask = 1.0 - obj_mask conf_focal = tf.squeeze(tf.math.pow(true_obj - pred_obj, 2), -1) # iou/giou/ciou/diou loss iou = bbox_iou(pred_box, true_box, xyxy=False, giou=True) iou_loss = (1 - iou) * obj_mask * box_scale # batch_size * grid * grid * 3 # confidence loss, Todo: multiply the iou conf_loss = self.bce_conf(true_obj, pred_obj) conf_loss = conf_focal * (obj_mask * conf_loss + background_mask * conf_loss) # batch * grid * grid * 3 # class loss # use binary cross entropy loss for multi class, so every value is independent and sigmoid # please note that the output of tf.keras.losses.bce is origial dim minus the last one class_loss = obj_mask * self.bce_class(true_class, pred_class) iou_loss = tf.reduce_mean(tf.reduce_sum(iou_loss, axis=[1, 2, 3])) conf_loss = tf.reduce_mean(tf.reduce_sum(conf_loss, axis=[1, 2, 3])) class_loss = tf.reduce_mean(tf.reduce_sum(class_loss, axis=[1, 2, 3])) iou_loss_all += iou_loss * balance[i] obj_loss_all += conf_loss * balance[i] class_loss_all += class_loss * self.num_classes * balance[i] # to balance the 3 loss复制代码
损失函数部分，还没有完全和v5的设置一样。v5中做了一些优化，如不同scale的平衡，如目标confidence损失的权重等。
效果
如果想要效果最佳，还是推荐原版pytorch，毕竟一直在更新中，v4和v5的作者也一直还在发力优化。如果对tensorflow有谜之爱好，或者想通过代码了解yolov5，我觉得我的版本写的更清楚一些（与之相应的代价是可能有细节遗漏甚至不对的地方），总之欢迎大家尝试。在MNIST检测数据上的效果：
在voc2012数据集上的效果(效果还有待加强)：
再大的数据集我就跑不动啦，毕竟只有1080Ti可用。