车牌识别LPRNet
转载
该项目为v1.0版本的车牌识别项目，对数据集做了更新，先对车牌矫正再识别，降低任务难度，40个epoch训练达验证集98.4%精度。实现模型与batch解耦，保证推理精度不受batch影响，可与车牌检测项目搭配。包含完整训练推理过程、模型转onnx及检查推理、数据集构建等内容。
车牌识别
LPRNet端到端训练车牌识别
本项目包括
完整训练推理过程模型转onnx以及onnx的检查和推理数据集构建
In [1]# 数据解压!unzip -o -q -d /home/aistudio/data /home/aistudio/data/data17968/CCPD2019.zip登录后复制In [4]
"""CCPD数据集的图片名称即是label:0152-4_14-224&551_398&624-388&610_224&624_234&565_398&551-0_0_30_27_31_9_31-97-108.jpg ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ | 框左上角 框右下角 右下角点 左下角点 左上角点 右上角点 车牌号码 亮度 模糊度水平/垂直倾角亮度数值越大，车牌越亮；模糊度数值越小，车牌越模糊。"""import cv2import osimport numpy as npfrom tqdm.notebook import tqdm # 参考 https://blog.csdn.net/qq_36516958/article/details/114274778from PIL import Image# 根据4顶点对图片矫正def four_point_transform(image, pts): rect = pts.astype('float32') br_x, br_y, bl_x, bl_y, tl_x, tl_y, tr_x, tr_y = rect widthA = np.sqrt(((br_x - bl_x) ** 2) + ((br_y - bl_y) ** 2)) widthB = np.sqrt(((tr_x - tl_x) ** 2) + ((tr_y - tl_y) ** 2)) maxWidth = max(int(widthA), int(widthB)) heightA = np.sqrt(((tr_x - br_x) ** 2) + ((tr_y - br_y) ** 2)) heightB = np.sqrt(((tl_x - bl_x) ** 2) + ((tl_y - bl_y) ** 2)) maxHeight = max(int(heightA), int(heightB)) rect = np.array([[tl_x, tl_y], [tr_x, tr_y], [br_x, br_y], [bl_x, bl_y]], dtype='float32') dst = np.array([ [0, 0], [maxWidth - 1, 0], [maxWidth - 1, maxHeight - 1], [0, maxHeight - 1]], dtype = "float32") M = cv2.getPerspectiveTransform(rect, dst) warped = cv2.warpPerspective(image, M, (maxWidth, maxHeight)) return warped# CCPD车牌有重复，应该是不同角度或者模糊程度path = r'data/ccpd_base' # 改成自己的车牌路径provinces = ["皖", "沪", "津", "渝", "冀", "晋", "蒙", "辽", "吉", "黑", "苏", "浙", "京", "闽", "赣", "鲁", "豫", "鄂", "湘", "粤", "桂", "琼", "川", "贵", "云", "藏", "陕", "甘", "青", "宁", "新", "警", "学", "O"]alphabets = ['A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'J', 'K', 'L', 'M', 'N', 'P', 'Q', 'R', 'S', 'T', 'U', 'V', 'W', 'X', 'Y', 'Z', 'O']ads = ['A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'J', 'K', 'L', 'M', 'N', 'P', 'Q', 'R', 'S', 'T', 'U', 'V', 'W', 'X', 'Y', 'Z', '0', '1', '2', '3', '4', '5', '6', '7', '8', '9', 'O']save_path = 'rec_images/data/'if not os.path.exists(save_path): os.makedirs(save_path)num = 0for filename in tqdm(os.listdir(path)): num += 1 result = "" _, _, box, points, plate, brightness, blurriness = filename.split('-') list_plate = plate.split('_') # 读取车牌 result += provinces[int(list_plate[0])] result += alphabets[int(list_plate[1])] result += ads[int(list_plate[2])] + ads[int(list_plate[3])] + ads[int(list_plate[4])] + ads[int(list_plate[5])] + ads[int(list_plate[6])] # 新能源车牌的要求，如果不是新能源车牌可以删掉这个if # if result[2] != 'D' and result[2] != 'F' \ # and result[-1] != 'D' and result[-1] != 'F': # print(filename) # print("Error label, Please check!") # assert 0, "Error label ^~^!!!" # print(result) img_path = os.path.join(path, filename) img = cv2.imread(img_path) assert os.path.exists(img_path), "image file {} dose not exist.".format(img_path) br, bl, tl, tr = points.split('_') br_x, br_y = [float(i) for i in br.split('&')] bl_x, bl_y = [float(i) for i in bl.split('&')] tl_x, tl_y = [float(i) for i in tl.split('&')] tr_x, tr_y = [float(i) for i in tr.split('&')] landmarks = np.array([br_x, br_y, bl_x, bl_y, tl_x, tl_y, tr_x, tr_y], dtype='float32') img = four_point_transform(img, landmarks) img = cv2.resize(img, (94, 24)) cv2.imencode('.jpg', img)[1].tofile(os.path.join(save_path, r"{}.jpg".format(result)))登录后复制0%| | 0/95774 [00:00登录后复制数据集划分
In [5]import osimport randomimage_dir = "rec_images/data" train_file = 'rec_images/train.txt'eval_file = 'rec_images/valid.txt'dataset_list = os.listdir(image_dir)train_num = 0valid_num = 0for img_name in dataset_list: if '.jpg' not in img_name: print(img_name) continue probo = random.randint(1, 100) if(probo <= 80): # train with open(train_file, 'a') as f_train: f_train.write(img_name+'\n') train_num+=1 else: #valid with open(eval_file, 'a') as f_eval: f_eval.write(img_name+'\n') valid_num+=1print(f'train: {train_num}, val:{valid_num}')登录后复制.ipynb_checkpointstrain: 62959, val:15937登录后复制Dataloader
数据读取
In [1]import osfrom paddle.io import Datasetfrom PIL import Imageimport numpy as npCHARS = ['京', '沪', '津', '渝', '冀', '晋', '蒙', '辽', '吉', '黑', '苏', '浙', '皖', '闽', '赣', '鲁', '豫', '鄂', '湘', '粤', '桂', '琼', '川', '贵', '云', '藏', '陕', '甘', '青', '宁', '新', '0', '1', '2', '3', '4', '5', '6', '7', '8', '9', 'A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'J', 'K', 'L', 'M', 'N', 'P', 'Q', 'R', 'S', 'T', 'U', 'V', 'W', 'X', 'Y', 'Z', 'I', 'O', '-' ]CHARS_DICT = {char:i for i, char in enumerate(CHARS)}class LprnetDataloader(Dataset): def __init__(self, target_path, label_text, transforms=None): super().__init__() self.transforms = transforms self.target_path = target_path with open(label_text) as f: self.data = f.readlines() def __getitem__(self, index): img_name = self.data[index].strip() img_path = os.path.join(self.target_path, img_name) data = Image.open(img_path) label = [] img_label = img_name.split('.', 1)[0] for c in img_label: label.append(CHARS_DICT[c]) if len(label) == 8: if self.check(label) == False: print(imgname) assert 0, "Error label ^~^!!!" if self.transforms is not None: data = self.transforms(data) data = np.array(data, dtype=np.float32) np_label = np.array(label, dtype=np.int64) return data, np_label, len(np_label) def __len__(self): return len(self.data) def check(self, label): if label[2] != CHARS_DICT['D'] and label[2] != CHARS_DICT['F'] \ and label[-1] != CHARS_DICT['D'] and label[-1] != CHARS_DICT['F']: print("Error label, Please check!") return False else: return True登录后复制组batch
将不同长度的label，padding为以最大标签长度的同一尺寸，shape为(batch_size,max_label_length)
In [2]def collate_fn(batch): # 图片输入已经规范到相同大小，这里只需要对标签进行padding batch_size = len(batch) # 找出标签最长的 batch_temp = sorted(batch, key=lambda sample: len(sample[1]), reverse=True) max_label_length = len(batch_temp[0][1]) # 以最大的长度创建0张量 labels = np.zeros((batch_size, max_label_length), dtype='int64') label_lens = [] img_list = [] for x in range(batch_size): sample = batch[x] tensor = sample[0] target = sample[1] label_length = sample[2] img_list.append(tensor) # 将数据插入都0张量中，实现了padding labels[x, :label_length] = target[:] label_lens.append(len(target)) label_lens = paddle.to_tensor(label_lens, dtype='int64') # ctcloss需要 imgs = paddle.to_tensor(img_list, dtype='float32') labels = paddle.to_tensor(labels, dtype="int32") # ctcloss仅支持int32的labels return imgs, labels, label_lens登录后复制数据前处理
这里数据集量挺多，各种情况的数据都有（天气，角度，模糊），就不再做数据增强的操作了。
就简单做个归一化操作就好了，训练的时候对数据进行ToTensor + Normalize
import paddle.vision.transforms as Ttrain_transforms = T.Compose([ T.ToTensor(data_format='CHW'), # 这里的CHW是指数据的输出格式 T.Normalize( [0.5, 0.5, 0.5], # 在totensor的时候已经将图片缩放到0-1 [0.5, 0.5, 0.5], data_format='CHW' # 这里是数据输入格式 ), ])登录后复制LPRNet网络
网络结构
In [3]import paddle.nn as nnimport paddleCHARS = ['京', '沪', '津', '渝', '冀', '晋', '蒙', '辽', '吉', '黑', '苏', '浙', '皖', '闽', '赣', '鲁', '豫', '鄂', '湘', '粤', '桂', '琼', '川', '贵', '云', '藏', '陕', '甘', '青', '宁', '新', '0', '1', '2', '3', '4', '5', '6', '7', '8', '9', 'A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'J', 'K', 'L', 'M', 'N', 'P', 'Q', 'R', 'S', 'T', 'U', 'V', 'W', 'X', 'Y', 'Z', 'I', 'O', '-' ]class small_basic_block(nn.Layer): def __init__(self, ch_in, ch_out): super(small_basic_block, self).__init__() self.block = nn.Sequential( nn.Conv2D(ch_in, ch_out // 4, kernel_size=1), nn.ReLU(), nn.Conv2D(ch_out // 4, ch_out // 4, kernel_size=(3, 1), padding=(1, 0)), nn.ReLU(), nn.Conv2D(ch_out // 4, ch_out // 4, kernel_size=(1, 3), padding=(0, 1)), nn.ReLU(), nn.Conv2D(ch_out // 4, ch_out, kernel_size=1), ) def forward(self, x): return self.block(x)class maxpool_3d(nn.Layer): def __init__(self, kernel_size, stride): super(maxpool_3d, self).__init__() assert(len(kernel_size)==3 and len(stride)==3) kernel_size2d1 = kernel_size[-2:] stride2d1 = stride[-2:] kernel_size2d2 = (1, kernel_size[0]) stride2d2 = (1, stride[0]) self.maxpool1 = nn.MaxPool2D(kernel_size=kernel_size2d1, stride=stride2d1) self.maxpool2 = nn.MaxPool2D(kernel_size=kernel_size2d2, stride=stride2d2) def forward(self,x): x = self.maxpool1(x) x = x.transpose((0, 3, 2, 1)) x = self.maxpool2(x) x = x.transpose((0, 3, 2, 1)) return xclass LPRNet(nn.Layer): def __init__(self, lpr_max_len, class_num, dropout_rate): super(LPRNet, self).__init__() self.lpr_max_len = lpr_max_len self.class_num = class_num self.backbone = nn.Sequential( nn.Conv2D(in_channels=3, out_channels=64, kernel_size=3, stride=1), # 0 [bs,3,24,94] -> [bs,64,22,92] nn.BatchNorm2D(num_features=64), # 1 -> [bs,64,22,92] nn.ReLU(), # 2 -> [bs,64,22,92] maxpool_3d(kernel_size=(1, 3, 3), stride=(1, 1, 1)), # 3 -> [bs,64,20,90] small_basic_block(ch_in=64, ch_out=128), # 4 -> [bs,128,20,90] nn.BatchNorm2D(num_features=128), # 5 -> [bs,128,20,90] nn.ReLU(), # 6 -> [bs,128,20,90] maxpool_3d(kernel_size=(1, 3, 3), stride=(2, 1, 2)), # 7 -> [bs,64,18,44] small_basic_block(ch_in=64, ch_out=256), # 8 -> [bs,256,18,44] nn.BatchNorm2D(num_features=256), # 9 -> [bs,256,18,44] nn.ReLU(), # 10 -> [bs,256,18,44] small_basic_block(ch_in=256, ch_out=256), # 11 -> [bs,256,18,44] nn.BatchNorm2D(num_features=256), # 12 -> [bs,256,18,44] nn.ReLU(), # 13 -> [bs,256,18,44] maxpool_3d(kernel_size=(1, 3, 3), stride=(4, 1, 2)), # 14 -> [bs,64,16,21] nn.Dropout(dropout_rate), # 15 -> [bs,64,16,21] nn.Conv2D(in_channels=64, out_channels=256, kernel_size=(1, 4), stride=1), # 16 -> [bs,256,16,18] nn.BatchNorm2D(num_features=256), # 17 -> [bs,256,16,18] nn.ReLU(), # 18 -> [bs,256,16,18] nn.Dropout(dropout_rate), # 19 -> [bs,256,16,18] nn.Conv2D(in_channels=256, out_channels=class_num, kernel_size=(13, 1), stride=1), # class_num=68 20 -> [bs,68,4,18] nn.BatchNorm2D(num_features=class_num), # 21 -> [bs,68,4,18] nn.ReLU(), # 22 -> [bs,68,4,18] ) self.container = nn.Sequential( nn.Conv2D(in_channels=448+self.class_num, out_channels=self.class_num, kernel_size=(1, 1), stride=(1, 1)), ) def forward(self, x): keep_features = list() for i, layer in enumerate(self.backbone.children()): x = layer(x) if i in [2, 6, 13, 22]: keep_features.append(x) global_context = list() # keep_features: [bs,64,22,92] [bs,128,20,90] [bs,256,18,44] [bs,68,4,18] for i, f in enumerate(keep_features): if i in [0, 1]: # [bs,64,22,92] -> [bs,64,4,18] # [bs,128,20,90] -> [bs,128,4,18] f = nn.AvgPool2D(kernel_size=5, stride=5)(f) if i in [2]: # [bs,256,18,44] -> [bs,256,4,18] f = nn.AvgPool2D(kernel_size=(4, 10), stride=(4, 2))(f) f_pow = paddle.pow(f, 2) # [bs,64,4,18] 所有元素求平方 # f_mean = paddle.mean(f_pow) # 1 所有元素求平均 f_mean = paddle.mean(f_pow, axis=[1,2,3], keepdim=True) f = paddle.divide(f, f_mean) # [bs,64,4,18] 所有元素除以这个均值 global_context.append(f) x = paddle.concat(global_context, 1) # [bs,516,4,18] x = self.container(x) # -> [bs, 68, 4, 18] head头 logits = paddle.mean(x, axis=2) # -> [bs, 68, 18] # 68 字符类别数 18字符序列长度 return logits登录后复制权重初始化函数
In [4]# 使用model.applay的方法，可以修改到每一个子层def init_weight(model): for name, layer in model.named_sublayers(): if isinstance(layer, nn.Conv2D): weight_attr = nn.initializer.KaimingNormal() bias_attr = nn.initializer.Constant(0.) init_bias = paddle.create_parameter(layer.bias.shape, attr=bias_attr, dtype='float32') init_weight = paddle.create_parameter(layer.weight.shape, attr=weight_attr, dtype='float32') layer.weight = init_weight layer.bias = init_bias elif isinstance(layer, nn.BatchNorm2D): weight_attr = nn.initializer.XavierUniform() bias_attr = nn.initializer.Constant(0.) init_bias = paddle.create_parameter(layer.bias.shape, attr=bias_attr, dtype='float32') init_weight = paddle.create_parameter(layer.weight.shape, attr=weight_attr, dtype='float32') layer.weight = init_weight layer.bias = init_bias登录后复制损失函数
损失函数是CTCLoss，需要传入的参数有：
logits: 概率序列, shape=[max_logit_length, batch_size, num_classes+1]
数据类型仅支持float32
lbels: padding后的标签序列，shape=[batch_size, max_label_length]
数据类型仅支持int32
input_lengths: 输入logits数据中的每个序列的长度，shape=[batch_size]
数据类型仅支持int64
label_lengths: label中每个序列的长度，shape=[batch_size]
数据类型仅支持int64
ctcloss文档：https://www.paddlepaddle.org.cn/documentation/docs/zh/api/paddle/nn/CTCLoss_cn.html
准确率计算函数
In [5]import numpy as npCHARS = ['京', '沪', '津', '渝', '冀', '晋', '蒙', '辽', '吉', '黑', '苏', '浙', '皖', '闽', '赣', '鲁', '豫', '鄂', '湘', '粤', '桂', '琼', '川', '贵', '云', '藏', '陕', '甘', '青', '宁', '新', '0', '1', '2', '3', '4', '5', '6', '7', '8', '9', 'A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'J', 'K', 'L', 'M', 'N', 'P', 'Q', 'R', 'S', 'T', 'U', 'V', 'W', 'X', 'Y', 'Z', 'I', 'O', '-' ]class ACC: def __init__(self): self.Tp = 0 self.Tn_1 = 0 self.Tn_2 = 0 self.acc = 0 def batch_update(self, batch_label, label_lengths, pred): for i, label in enumerate(batch_label): length = label_lengths[i] label = label[:length] pred_i = pred[i, :, :] preb_label = [] for j in range(pred_i.shape[1]): # T preb_label.append(np.argmax(pred_i[:, j], axis=0)) no_repeat_blank_label = [] pre_c = preb_label[0] if pre_c != len(CHARS) - 1: # 非空白 no_repeat_blank_label.append(pre_c) for c in preb_label: # dropout repeate label and blank label if (pre_c == c) or (c == len(CHARS) - 1): if c == len(CHARS) - 1: pre_c = c continue no_repeat_blank_label.append(c) pre_c = c # print('no_repeat_blank_label:', no_repeat_blank_label) # print('gt_label:', label) if len(label) != len(no_repeat_blank_label): self.Tn_1 += 1 elif (np.asarray(label) == np.asarray(no_repeat_blank_label)).all(): self.Tp += 1 else: self.Tn_2 += 1 self.acc = self.Tp * 1.0 / (self.Tp + self.Tn_1 + self.Tn_2) def clear(self): self.Tp = 0 self.Tn_1 = 0 self.Tn_2 = 0 self.acc = 0print(len(CHARS))登录后复制68登录后复制加载预训练参数
一次训练没有到位，在之前的权重参数基础上继续训练，需要加载预训练权重。
若有预训练权重可以加载预训练权重
In [6]# 保存的权重路径：runs/lprnet_best.pdparamsimport osdef load_pretrained(model, path=None): print('params loading...') if not (os.path.isdir(path) or os.path.exists(path + '.pdparams')): raise ValueError("Model pretrain path {} does not " "exists.".format(path)) param_state_dict = paddle.load(path + ".pdparams") model.set_dict(param_state_dict) print(f'load {path + ".pdparams"} success...') return登录后复制训练
训练得到的模型为：runs/lprnet_best_2.pdparams
矫正后的图片一定程度降低了任务难度，这里进行了40个epoch的训练，最终验证集精度98.4%
In [7]import paddle.vision.transforms as Tfrom paddle.io import DataLoaderimport timefrom statistics import mean# 参数定义EPOCH = 40IMGSIZE = (94, 24)IMGDIR = 'rec_images/data'TRAINFILE = 'rec_images/train.txt'VALIDFILE = 'rec_images/valid.txt'SAVEFOLDER = './runs'DROPOUT = 0.LEARNINGRATE = 0.001LPRMAXLEN = 18TRAINBATCHSIZE = 256EVALBATCHSIZE = 256NUMWORKERS = 2 # 若dataloader报错，调小该参数，或直接改为0WEIGHTDECAY = 0.001# 图片预处理train_transforms = T.Compose([ T.ColorJitter(0.2,0.2,0.2), T.ToTensor(data_format='CHW'), T.Normalize( [0.5, 0.5, 0.5], # 在totensor的时候已经将图片缩放到0-1 [0.5, 0.5, 0.5], data_format='CHW' ), ])eval_transforms = T.Compose([ T.ToTensor(data_format='CHW'), T.Normalize( [0.5, 0.5, 0.5], [0.5, 0.5, 0.5], data_format='CHW' ), ])# 数据加载train_data_set = LprnetDataloader(IMGDIR, TRAINFILE, train_transforms)eval_data_set = LprnetDataloader(IMGDIR, VALIDFILE, eval_transforms)train_loader = DataLoader( train_data_set, batch_size=TRAINBATCHSIZE, shuffle=True, num_workers=NUMWORKERS, drop_last=True, collate_fn=collate_fn)eval_loader = DataLoader( eval_data_set, batch_size=EVALBATCHSIZE, shuffle=False, num_workers=NUMWORKERS, drop_last=False, collate_fn=collate_fn)# 定义lossloss_func = nn.CTCLoss(len(CHARS)-1)# input_length, loss计算需要input_length = np.ones(shape=TRAINBATCHSIZE) * LPRMAXLENinput_length = paddle.to_tensor(input_length, dtype='int64')# LPRNet网络，初始化/加载预训练参数model = LPRNet(LPRMAXLEN, len(CHARS), DROPOUT)model.apply(init_weight) # 首次训练时初始化# 定义优化器def make_optimizer(base_lr, parameters=None): momentum = 0.9 weight_decay = WEIGHTDECAY scheduler = paddle.optimizer.lr.CosineAnnealingDecay( learning_rate=base_lr, eta_min=0.01*base_lr, T_max=EPOCH, verbose=1) scheduler = paddle.optimizer.lr.LinearWarmup( # 第一次训练的时候考虑模型权重不稳定，添加warmup策略 learning_rate=scheduler, warmup_steps=5, start_lr=base_lr/5, end_lr=base_lr, verbose=True) optimizer = paddle.optimizer.Momentum( learning_rate=scheduler, weight_decay=paddle.regularizer.L2Decay(weight_decay), momentum=momentum, parameters=parameters) return optimizer, scheduleroptim, scheduler = make_optimizer(LEARNINGRATE, parameters=model.parameters())# accacc_train = ACC()acc_eval = ACC()BESTACC = 0.5# 训练流程for epoch in range(EPOCH): start_time = time.localtime(time.time()) str_time = time.strftime("%Y-%m-%d %H:%M:%S", start_time) print(f'{str_time} || Epoch {epoch} start:') model.train() for batch_id, bath_data in enumerate(train_loader): img_data, label_data, label_lens = bath_data predict = model(img_data) logits = paddle.transpose(predict, (2,0,1)) # for ctc loss: T x N x C loss = loss_func(logits , label_data, input_length, label_lens) acc_train.batch_update(label_data, label_lens, predict) if batch_id % 50 == 0: print(f'epoch:{epoch}, batch_id:{batch_id}, loss:{loss.item():.4f}, \ acc:{acc_train.acc:.4f} Tp/Tn_1/Tn_2: {acc_train.Tp}/{acc_train.Tn_1}/{acc_train.Tn_2}') loss.backward() optim.step() optim.clear_grad() acc_train.clear() # save if epoch and epoch % 20 == 0: paddle.save(model.state_dict(), os.path.join(SAVEFOLDER,f'lprnet_{epoch}_2.pdparams')) paddle.save(optim.state_dict(), os.path.join(SAVEFOLDER,f'lprnet_{epoch}_2.pdopt')) print(f'Saved log ecpch-{epoch}') # eval with paddle.no_grad(): model.eval() loss_list = [] for batch_id, bath_data in enumerate(eval_loader): img_data, label_data, label_lens = bath_data predict = model(img_data) logits = paddle.transpose(predict, (2,0,1)) loss = loss_func(logits, label_data, input_length, label_lens) acc_eval.batch_update(label_data, label_lens, predict) loss_list.append(loss.item()) print(f'Eval of epoch {epoch} => acc:{acc_eval.acc:.4f}, loss:{mean(loss_list):.4f}') # save best model if acc_eval.acc > BESTACC: paddle.save(model.state_dict(), os.path.join(SAVEFOLDER,f'lprnet_best_2.pdparams')) paddle.save(optim.state_dict(), os.path.join(SAVEFOLDER,f'lprnet_best_2.pdopt')) BESTACC = acc_eval.acc print(f'Saved best model of epoch{epoch}, acc {acc_eval.acc:.4f}, save path "{SAVEFOLDER}"') acc_eval.clear() # 学习率衰减策略 scheduler.step()登录后复制Epoch 0: CosineAnnealingDecay set learning rate to 0.001.Epoch 0: LinearWarmup set learning rate to 0.0002.2024-07-27 13:05:14 || Epoch 0 start:登录后复制W0727 13:05:14.896270 20153 gpu_resources.cc:61] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.2, Runtime API Version: 11.2W0727 13:05:14.900187 20153 gpu_resources.cc:91] device: 0, cuDNN Version: 8.2./opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/nn/layer/norm.py:712: UserWarning: When training, we now always track global mean and variance. "When training, we now always track global mean and variance."/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/math_op_patch.py:277: UserWarning: The dtype of left and right variables are not the same, left dtype is paddle.float32, but right dtype is paddle.int64, the right dtype will convert to paddle.float32 .format(lhs_dtype, rhs_dtype, lhs_dtype))登录后复制epoch:0, batch_id:0, loss:71.7745, acc:0.0000 Tp/Tn_1/Tn_2: 0/242/14epoch:0, batch_id:50, loss:4.1406, acc:0.0000 Tp/Tn_1/Tn_2: 0/11919/1137epoch:0, batch_id:100, loss:2.7202, acc:0.0000 Tp/Tn_1/Tn_2: 0/23156/2700epoch:0, batch_id:150, loss:2.3612, acc:0.0000 Tp/Tn_1/Tn_2: 1/34437/4218epoch:0, batch_id:200, loss:1.9747, acc:0.0003 Tp/Tn_1/Tn_2: 15/45563/5878Eval of epoch 0 => acc:0.0053, loss:1.7451Epoch 1: LinearWarmup set learning rate to 0.00036.2024-07-27 13:07:32 || Epoch 1 start:epoch:1, batch_id:0, loss:1.8066, acc:0.0000 Tp/Tn_1/Tn_2: 0/208/48epoch:1, batch_id:50, loss:1.4053, acc:0.0102 Tp/Tn_1/Tn_2: 133/10502/2421epoch:1, batch_id:100, loss:1.0782, acc:0.0284 Tp/Tn_1/Tn_2: 735/19852/5269epoch:1, batch_id:150, loss:0.8829, acc:0.0548 Tp/Tn_1/Tn_2: 2119/28197/8340epoch:1, batch_id:200, loss:0.6519, acc:0.0886 Tp/Tn_1/Tn_2: 4559/35534/11363Eval of epoch 1 => acc:0.3610, loss:0.5038Epoch 2: LinearWarmup set learning rate to 0.0005200000000000001.2024-07-27 13:11:21 || Epoch 2 start:epoch:2, batch_id:0, loss:0.4789, acc:0.3516 Tp/Tn_1/Tn_2: 90/116/50epoch:2, batch_id:50, loss:0.3491, acc:0.4154 Tp/Tn_1/Tn_2: 5423/5085/2548epoch:2, batch_id:100, loss:0.2775, acc:0.4883 Tp/Tn_1/Tn_2: 12626/8683/4547epoch:2, batch_id:150, loss:0.2271, acc:0.5498 Tp/Tn_1/Tn_2: 21252/11238/6166epoch:2, batch_id:200, loss:0.1476, acc:0.5987 Tp/Tn_1/Tn_2: 30809/13097/7550Eval of epoch 2 => acc:0.8157, loss:0.1444Saved best model of epoch2, acc 0.8157, save path "./runs"Epoch 3: LinearWarmup set learning rate to 0.00068.2024-07-27 13:17:25 || Epoch 3 start:epoch:3, batch_id:0, loss:0.1045, acc:0.8320 Tp/Tn_1/Tn_2: 213/28/15epoch:3, batch_id:50, loss:0.1356, acc:0.8227 Tp/Tn_1/Tn_2: 10741/1188/1127epoch:3, batch_id:100, loss:0.0764, acc:0.8366 Tp/Tn_1/Tn_2: 21631/2146/2079epoch:3, batch_id:150, loss:0.1323, acc:0.8478 Tp/Tn_1/Tn_2: 32772/2902/2982epoch:3, batch_id:200, loss:0.0691, acc:0.8577 Tp/Tn_1/Tn_2: 44132/3486/3838Eval of epoch 3 => acc:0.9084, loss:0.0740Saved best model of epoch3, acc 0.9084, save path "./runs"Epoch 4: LinearWarmup set learning rate to 0.00084.2024-07-27 13:19:41 || Epoch 4 start:epoch:4, batch_id:0, loss:0.1121, acc:0.8867 Tp/Tn_1/Tn_2: 227/15/14epoch:4, batch_id:50, loss:0.0609, acc:0.9091 Tp/Tn_1/Tn_2: 11869/419/768epoch:4, batch_id:100, loss:0.0481, acc:0.9139 Tp/Tn_1/Tn_2: 23629/768/1459epoch:4, batch_id:150, loss:0.0587, acc:0.9171 Tp/Tn_1/Tn_2: 35450/1082/2124epoch:4, batch_id:200, loss:0.0702, acc:0.9203 Tp/Tn_1/Tn_2: 47356/1337/2763Eval of epoch 4 => acc:0.9371, loss:0.0509Saved best model of epoch2, acc 0.9371, save path "./runs"Epoch 0: CosineAnnealingDecay set learning rate to 0.001.Epoch 5: LinearWarmup set learning rate to 0.001.2024-07-27 13:21:59 || Epoch 5 start:epoch:5, batch_id:0, loss:0.0351, acc:0.9336 Tp/Tn_1/Tn_2: 239/6/11epoch:5, batch_id:50, loss:0.0328, acc:0.9372 Tp/Tn_1/Tn_2: 12236/249/571epoch:5, batch_id:100, loss:0.0846, acc:0.9387 Tp/Tn_1/Tn_2: 24270/461/1125epoch:5, batch_id:150, loss:0.0242, acc:0.9410 Tp/Tn_1/Tn_2: 36375/657/1624epoch:5, batch_id:200, loss:0.0518, acc:0.9416 Tp/Tn_1/Tn_2: 48453/859/2144Eval of epoch 5 => acc:0.9513, loss:0.0396Saved best model of epoch5, acc 0.9513, save path "./runs"Epoch 1: CosineAnnealingDecay set learning rate to 0.0009984740801978985.Epoch 6: LinearWarmup set learning rate to 0.0009984740801978985.2024-07-27 13:24:17 || Epoch 6 start:epoch:6, batch_id:0, loss:0.0390, acc:0.9453 Tp/Tn_1/Tn_2: 242/4/10epoch:6, batch_id:50, loss:0.0319, acc:0.9527 Tp/Tn_1/Tn_2: 12438/155/463epoch:6, batch_id:100, loss:0.0267, acc:0.9526 Tp/Tn_1/Tn_2: 24630/296/930epoch:6, batch_id:150, loss:0.0330, acc:0.9526 Tp/Tn_1/Tn_2: 36825/429/1402epoch:6, batch_id:200, loss:0.0202, acc:0.9534 Tp/Tn_1/Tn_2: 49056/560/1840Eval of epoch 6 => acc:0.9589, loss:0.0337Saved best model of epoch6, acc 0.9589, save path "./runs"Epoch 2: CosineAnnealingDecay set learning rate to 0.0009939057285945933.Epoch 7: LinearWarmup set learning rate to 0.0009939057285945933.2024-07-27 13:26:33 || Epoch 7 start:epoch:7, batch_id:0, loss:0.0181, acc:0.9766 Tp/Tn_1/Tn_2: 250/0/6epoch:7, batch_id:50, loss:0.0189, acc:0.9609 Tp/Tn_1/Tn_2: 12546/118/392epoch:7, batch_id:100, loss:0.0424, acc:0.9590 Tp/Tn_1/Tn_2: 24795/234/827epoch:7, batch_id:150, loss:0.0381, acc:0.9611 Tp/Tn_1/Tn_2: 37154/322/1180epoch:7, batch_id:200, loss:0.0206, acc:0.9615 Tp/Tn_1/Tn_2: 49475/431/1550Eval of epoch 7 => acc:0.9644, loss:0.0298Saved best model of epoch7, acc 0.9644, save path "./runs"Epoch 3: CosineAnnealingDecay set learning rate to 0.00098632311059685.Epoch 8: LinearWarmup set learning rate to 0.00098632311059685.2024-07-27 13:28:51 || Epoch 8 start:epoch:8, batch_id:0, loss:0.0230, acc:0.9570 Tp/Tn_1/Tn_2: 245/0/11epoch:8, batch_id:50, loss:0.0259, acc:0.9634 Tp/Tn_1/Tn_2: 12578/91/387epoch:8, batch_id:100, loss:0.0092, acc:0.9656 Tp/Tn_1/Tn_2: 24966/177/713epoch:8, batch_id:150, loss:0.0376, acc:0.9657 Tp/Tn_1/Tn_2: 37332/264/1060epoch:8, batch_id:200, loss:0.0304, acc:0.9662 Tp/Tn_1/Tn_2: 49718/344/1394Eval of epoch 8 => acc:0.9681, loss:0.0268Saved best model of epoch8, acc 0.9681, save path "./runs"Epoch 4: CosineAnnealingDecay set learning rate to 0.0009757729755661011.Epoch 9: LinearWarmup set learning rate to 0.0009757729755661011.2024-07-27 13:31:10 || Epoch 9 start:epoch:9, batch_id:0, loss:0.0178, acc:0.9531 Tp/Tn_1/Tn_2: 244/1/11epoch:9, batch_id:50, loss:0.0270, acc:0.9685 Tp/Tn_1/Tn_2: 12645/74/337epoch:9, batch_id:100, loss:0.0288, acc:0.9699 Tp/Tn_1/Tn_2: 25078/155/623epoch:9, batch_id:150, loss:0.0193, acc:0.9697 Tp/Tn_1/Tn_2: 37483/223/950epoch:9, batch_id:200, loss:0.0158, acc:0.9703 Tp/Tn_1/Tn_2: 49927/287/1242Eval of epoch 9 => acc:0.9695, loss:0.0239Saved best model of epoch9, acc 0.9695, save path "./runs"Epoch 5: CosineAnnealingDecay set learning rate to 0.000962320368593087.Epoch 10: LinearWarmup set learning rate to 0.000962320368593087.2024-07-27 13:33:30 || Epoch 10 start:epoch:10, batch_id:0, loss:0.0241, acc:0.9531 Tp/Tn_1/Tn_2: 244/3/9epoch:10, batch_id:50, loss:0.0086, acc:0.9722 Tp/Tn_1/Tn_2: 12693/66/297epoch:10, batch_id:100, loss:0.0416, acc:0.9717 Tp/Tn_1/Tn_2: 25125/129/602epoch:10, batch_id:150, loss:0.0239, acc:0.9724 Tp/Tn_1/Tn_2: 37588/191/877epoch:10, batch_id:200, loss:0.0162, acc:0.9728 Tp/Tn_1/Tn_2: 50054/263/1139Eval of epoch 10 => acc:0.9715, loss:0.0213Saved best model of epoch10, acc 0.9715, save path "./runs"Epoch 6: CosineAnnealingDecay set learning rate to 0.0009460482294732421.Epoch 11: LinearWarmup set learning rate to 0.0009460482294732421.2024-07-27 13:35:46 || Epoch 11 start:epoch:11, batch_id:0, loss:0.0308, acc:0.9570 Tp/Tn_1/Tn_2: 245/0/11epoch:11, batch_id:50, loss:0.0199, acc:0.9739 Tp/Tn_1/Tn_2: 12715/77/264epoch:11, batch_id:100, loss:0.0190, acc:0.9752 Tp/Tn_1/Tn_2: 25215/125/516epoch:11, batch_id:150, loss:0.0100, acc:0.9751 Tp/Tn_1/Tn_2: 37692/182/782epoch:11, batch_id:200, loss:0.0172, acc:0.9749 Tp/Tn_1/Tn_2: 50166/244/1046Eval of epoch 11 => acc:0.9745, loss:0.0195Saved best model of epoch11, acc 0.9745, save path "./runs"Epoch 7: CosineAnnealingDecay set learning rate to 0.0009270568813552756.Epoch 12: LinearWarmup set learning rate to 0.0009270568813552756.2024-07-27 13:38:05 || Epoch 12 start:epoch:12, batch_id:0, loss:0.0274, acc:0.9609 Tp/Tn_1/Tn_2: 246/3/7epoch:12, batch_id:50, loss:0.0134, acc:0.9765 Tp/Tn_1/Tn_2: 12749/46/261epoch:12, batch_id:100, loss:0.0206, acc:0.9773 Tp/Tn_1/Tn_2: 25269/93/494epoch:12, batch_id:150, loss:0.0182, acc:0.9766 Tp/Tn_1/Tn_2: 37753/151/752epoch:12, batch_id:200, loss:0.0233, acc:0.9762 Tp/Tn_1/Tn_2: 50230/217/1009Eval of epoch 12 => acc:0.9757, loss:0.0177Saved best model of epoch12, acc 0.9757, save path "./runs"Epoch 8: CosineAnnealingDecay set learning rate to 0.000905463412215599.Epoch 13: LinearWarmup set learning rate to 0.000905463412215599.2024-07-27 13:40:19 || Epoch 13 start:epoch:13, batch_id:0, loss:0.0058, acc:0.9922 Tp/Tn_1/Tn_2: 254/1/1epoch:13, batch_id:50, loss:0.0078, acc:0.9788 Tp/Tn_1/Tn_2: 12779/43/234epoch:13, batch_id:100, loss:0.0080, acc:0.9780 Tp/Tn_1/Tn_2: 25287/92/477epoch:13, batch_id:150, loss:0.0355, acc:0.9780 Tp/Tn_1/Tn_2: 37804/141/711epoch:13, batch_id:200, loss:0.0257, acc:0.9784 Tp/Tn_1/Tn_2: 50345/190/921Eval of epoch 13 => acc:0.9772, loss:0.0168Saved best model of epoch13, acc 0.9772, save path "./runs"Epoch 9: CosineAnnealingDecay set learning rate to 0.0008814009529720154.Epoch 14: LinearWarmup set learning rate to 0.0008814009529720154.2024-07-27 13:42:34 || Epoch 14 start:epoch:14, batch_id:0, loss:0.0075, acc:0.9883 Tp/Tn_1/Tn_2: 253/0/3epoch:14, batch_id:50, loss:0.0097, acc:0.9809 Tp/Tn_1/Tn_2: 12807/48/201epoch:14, batch_id:100, loss:0.0088, acc:0.9803 Tp/Tn_1/Tn_2: 25346/96/414epoch:14, batch_id:150, loss:0.0239, acc:0.9795 Tp/Tn_1/Tn_2: 37862/148/646epoch:14, batch_id:200, loss:0.0091, acc:0.9794 Tp/Tn_1/Tn_2: 50397/194/865Eval of epoch 14 => acc:0.9778, loss:0.0162Saved best model of epoch14, acc 0.9778, save path "./runs"Epoch 10: CosineAnnealingDecay set learning rate to 0.000855017856687341.Epoch 15: LinearWarmup set learning rate to 0.000855017856687341.2024-07-27 13:44:53 || Epoch 15 start:epoch:15, batch_id:0, loss:0.0075, acc:0.9844 Tp/Tn_1/Tn_2: 252/0/4epoch:15, batch_id:50, loss:0.0075, acc:0.9814 Tp/Tn_1/Tn_2: 12813/28/215epoch:15, batch_id:100, loss:0.0178, acc:0.9817 Tp/Tn_1/Tn_2: 25383/66/407epoch:15, batch_id:150, loss:0.0193, acc:0.9820 Tp/Tn_1/Tn_2: 37959/100/597epoch:15, batch_id:200, loss:0.0099, acc:0.9818 Tp/Tn_1/Tn_2: 50517/147/792Eval of epoch 15 => acc:0.9785, loss:0.0155Saved best model of epoch15, acc 0.9785, save path "./runs"Epoch 11: CosineAnnealingDecay set learning rate to 0.0008264767839234411.Epoch 16: LinearWarmup set learning rate to 0.0008264767839234411.2024-07-27 13:47:12 || Epoch 16 start:epoch:16, batch_id:0, loss:0.0216, acc:0.9688 Tp/Tn_1/Tn_2: 248/0/8epoch:16, batch_id:50, loss:0.0061, acc:0.9833 Tp/Tn_1/Tn_2: 12838/35/183epoch:16, batch_id:100, loss:0.0131, acc:0.9831 Tp/Tn_1/Tn_2: 25420/70/366epoch:16, batch_id:150, loss:0.0040, acc:0.9834 Tp/Tn_1/Tn_2: 38013/117/526epoch:16, batch_id:200, loss:0.0064, acc:0.9831 Tp/Tn_1/Tn_2: 50587/160/709Eval of epoch 16 => acc:0.9786, loss:0.0148Saved best model of epoch16, acc 0.9786, save path "./runs"Epoch 12: CosineAnnealingDecay set learning rate to 0.0007959536998847742.Epoch 17: LinearWarmup set learning rate to 0.0007959536998847742.2024-07-27 13:49:29 || Epoch 17 start:epoch:17, batch_id:0, loss:0.0169, acc:0.9844 Tp/Tn_1/Tn_2: 252/0/4epoch:17, batch_id:50, loss:0.0070, acc:0.9845 Tp/Tn_1/Tn_2: 12853/42/161epoch:17, batch_id:100, loss:0.0067, acc:0.9825 Tp/Tn_1/Tn_2: 25403/80/373epoch:17, batch_id:150, loss:0.0193, acc:0.9827 Tp/Tn_1/Tn_2: 37989/129/538epoch:17, batch_id:200, loss:0.0085, acc:0.9833 Tp/Tn_1/Tn_2: 50599/155/702Eval of epoch 17 => acc:0.9797, loss:0.0144Saved best model of epoch17, acc 0.9797, save path "./runs"Epoch 13: CosineAnnealingDecay set learning rate to 0.0007636367895343947.Epoch 18: LinearWarmup set learning rate to 0.0007636367895343947.2024-07-27 13:51:45 || Epoch 18 start:epoch:18, batch_id:0, loss:0.0047, acc:0.9961 Tp/Tn_1/Tn_2: 255/0/1epoch:18, batch_id:50, loss:0.0091, acc:0.9848 Tp/Tn_1/Tn_2: 12857/32/167epoch:18, batch_id:100, loss:0.0121, acc:0.9844 Tp/Tn_1/Tn_2: 25452/59/345epoch:18, batch_id:150, loss:0.0130, acc:0.9846 Tp/Tn_1/Tn_2: 38059/90/507epoch:18, batch_id:200, loss:0.0141, acc:0.9846 Tp/Tn_1/Tn_2: 50663/130/663Eval of epoch 18 => acc:0.9804, loss:0.0138Saved best model of epoch18, acc 0.9804, save path "./runs"Epoch 14: CosineAnnealingDecay set learning rate to 0.0007297252973710757.Epoch 19: LinearWarmup set learning rate to 0.0007297252973710757.2024-07-27 13:54:00 || Epoch 19 start:epoch:19, batch_id:0, loss:0.0076, acc:0.9844 Tp/Tn_1/Tn_2: 252/0/4epoch:19, batch_id:50, loss:0.0050, acc:0.9849 Tp/Tn_1/Tn_2: 12859/32/165epoch:19, batch_id:100, loss:0.0146, acc:0.9853 Tp/Tn_1/Tn_2: 25476/63/317epoch:19, batch_id:150, loss:0.0035, acc:0.9858 Tp/Tn_1/Tn_2: 38107/87/462epoch:19, batch_id:200, loss:0.0061, acc:0.9856 Tp/Tn_1/Tn_2: 50713/120/623Eval of epoch 19 => acc:0.9812, loss:0.0138Saved best model of epoch19, acc 0.9812, save path "./runs"Epoch 15: CosineAnnealingDecay set learning rate to 0.0006944282990207195.Epoch 20: LinearWarmup set learning rate to 0.0006944282990207195.2024-07-27 13:56:15 || Epoch 20 start:epoch:20, batch_id:0, loss:0.0080, acc:0.9883 Tp/Tn_1/Tn_2: 253/0/3epoch:20, batch_id:50, loss:0.0089, acc:0.9860 Tp/Tn_1/Tn_2: 12873/25/158epoch:20, batch_id:100, loss:0.0161, acc:0.9858 Tp/Tn_1/Tn_2: 25488/55/313epoch:20, batch_id:150, loss:0.0112, acc:0.9864 Tp/Tn_1/Tn_2: 38132/77/447epoch:20, batch_id:200, loss:0.0057, acc:0.9862 Tp/Tn_1/Tn_2: 50744/107/605Saved log ecpch-20Eval of epoch 20 => acc:0.9813, loss:0.0134Saved best model of epoch20, acc 0.9813, save path "./runs"Epoch 16: CosineAnnealingDecay set learning rate to 0.000657963412215599.Epoch 21: LinearWarmup set learning rate to 0.000657963412215599.2024-07-27 13:58:31 || Epoch 21 start:epoch:21, batch_id:0, loss:0.0319, acc:0.9844 Tp/Tn_1/Tn_2: 252/3/1epoch:21, batch_id:50, loss:0.0204, acc:0.9850 Tp/Tn_1/Tn_2: 12860/29/167epoch:21, batch_id:100, loss:0.0085, acc:0.9858 Tp/Tn_1/Tn_2: 25488/53/315epoch:21, batch_id:150, loss:0.0323, acc:0.9862 Tp/Tn_1/Tn_2: 38124/79/453epoch:21, batch_id:200, loss:0.0075, acc:0.9866 Tp/Tn_1/Tn_2: 50764/104/588Eval of epoch 21 => acc:0.9824, loss:0.0128Saved best model of epoch21, acc 0.9824, save path "./runs"Epoch 17: CosineAnnealingDecay set learning rate to 0.0006205554551086733.Epoch 22: LinearWarmup set learning rate to 0.0006205554551086733.2024-07-27 14:00:46 || Epoch 22 start:epoch:22, batch_id:0, loss:0.0116, acc:0.9844 Tp/Tn_1/Tn_2: 252/0/4epoch:22, batch_id:50, loss:0.0096, acc:0.9880 Tp/Tn_1/Tn_2: 12899/23/134epoch:22, batch_id:100, loss:0.0068, acc:0.9870 Tp/Tn_1/Tn_2: 25521/63/272epoch:22, batch_id:150, loss:0.0063, acc:0.9876 Tp/Tn_1/Tn_2: 38175/84/397epoch:22, batch_id:200, loss:0.0104, acc:0.9878 Tp/Tn_1/Tn_2: 50826/102/528Eval of epoch 22 => acc:0.9822, loss:0.0129Epoch 18: CosineAnnealingDecay set learning rate to 0.0005824350601949143.Epoch 23: LinearWarmup set learning rate to 0.0005824350601949143.2024-07-27 14:03:02 || Epoch 23 start:epoch:23, batch_id:0, loss:0.0049, acc:0.9922 Tp/Tn_1/Tn_2: 254/1/1epoch:23, batch_id:50, loss:0.0056, acc:0.9878 Tp/Tn_1/Tn_2: 12897/21/138epoch:23, batch_id:100, loss:0.0058, acc:0.9876 Tp/Tn_1/Tn_2: 25535/50/271epoch:23, batch_id:150, loss:0.0023, acc:0.9876 Tp/Tn_1/Tn_2: 38175/77/404epoch:23, batch_id:200, loss:0.0041, acc:0.9878 Tp/Tn_1/Tn_2: 50827/99/530Eval of epoch 23 => acc:0.9822, loss:0.0125Epoch 19: CosineAnnealingDecay set learning rate to 0.0005438372523852833.Epoch 24: LinearWarmup set learning rate to 0.0005438372523852833.2024-07-27 14:05:22 || Epoch 24 start:epoch:24, batch_id:0, loss:0.0029, acc:0.9961 Tp/Tn_1/Tn_2: 255/0/1epoch:24, batch_id:50, loss:0.0065, acc:0.9869 Tp/Tn_1/Tn_2: 12885/28/143epoch:24, batch_id:100, loss:0.0044, acc:0.9877 Tp/Tn_1/Tn_2: 25537/52/267epoch:24, batch_id:150, loss:0.0028, acc:0.9891 Tp/Tn_1/Tn_2: 38234/65/357epoch:24, batch_id:200, loss:0.0007, acc:0.9887 Tp/Tn_1/Tn_2: 50875/83/498Eval of epoch 24 => acc:0.9834, loss:0.0121Saved best model of epoch24, acc 0.9834, save path "./runs"Epoch 20: CosineAnnealingDecay set learning rate to 0.000505.Epoch 25: LinearWarmup set learning rate to 0.000505.2024-07-27 14:07:41 || Epoch 25 start:epoch:25, batch_id:0, loss:0.0111, acc:0.9844 Tp/Tn_1/Tn_2: 252/0/4epoch:25, batch_id:50, loss:0.0031, acc:0.9896 Tp/Tn_1/Tn_2: 12920/21/115epoch:25, batch_id:100, loss:0.0046, acc:0.9896 Tp/Tn_1/Tn_2: 25586/45/225epoch:25, batch_id:150, loss:0.0075, acc:0.9891 Tp/Tn_1/Tn_2: 38236/73/347epoch:25, batch_id:200, loss:0.0018, acc:0.9892 Tp/Tn_1/Tn_2: 50901/97/458Eval of epoch 25 => acc:0.9830, loss:0.0120Epoch 21: CosineAnnealingDecay set learning rate to 0.0004661627476147168.Epoch 26: LinearWarmup set learning rate to 0.0004661627476147168.2024-07-27 14:09:58 || Epoch 26 start:epoch:26, batch_id:0, loss:0.0084, acc:0.9922 Tp/Tn_1/Tn_2: 254/0/2epoch:26, batch_id:50, loss:0.0024, acc:0.9897 Tp/Tn_1/Tn_2: 12922/29/105epoch:26, batch_id:100, loss:0.0021, acc:0.9894 Tp/Tn_1/Tn_2: 25583/52/221epoch:26, batch_id:150, loss:0.0017, acc:0.9892 Tp/Tn_1/Tn_2: 38239/66/351epoch:26, batch_id:200, loss:0.0042, acc:0.9893 Tp/Tn_1/Tn_2: 50906/81/469Eval of epoch 26 => acc:0.9829, loss:0.0120Epoch 22: CosineAnnealingDecay set learning rate to 0.0004275649398050859.Epoch 27: LinearWarmup set learning rate to 0.0004275649398050859.2024-07-27 14:12:15 || Epoch 27 start:epoch:27, batch_id:0, loss:0.0072, acc:0.9844 Tp/Tn_1/Tn_2: 252/0/4epoch:27, batch_id:50, loss:0.0022, acc:0.9881 Tp/Tn_1/Tn_2: 12900/32/124epoch:27, batch_id:100, loss:0.0169, acc:0.9889 Tp/Tn_1/Tn_2: 25568/57/231epoch:27, batch_id:150, loss:0.0024, acc:0.9893 Tp/Tn_1/Tn_2: 38243/71/342epoch:27, batch_id:200, loss:0.0031, acc:0.9894 Tp/Tn_1/Tn_2: 50909/89/458Eval of epoch 27 => acc:0.9833, loss:0.0119Epoch 23: CosineAnnealingDecay set learning rate to 0.0003894445448913269.Epoch 28: LinearWarmup set learning rate to 0.0003894445448913269.2024-07-27 14:14:31 || Epoch 28 start:epoch:28, batch_id:0, loss:0.0051, acc:0.9844 Tp/Tn_1/Tn_2: 252/0/4epoch:28, batch_id:50, loss:0.0014, acc:0.9895 Tp/Tn_1/Tn_2: 12919/20/117epoch:28, batch_id:100, loss:0.0064, acc:0.9897 Tp/Tn_1/Tn_2: 25589/44/223epoch:28, batch_id:150, loss:0.0022, acc:0.9896 Tp/Tn_1/Tn_2: 38253/61/342epoch:28, batch_id:200, loss:0.0049, acc:0.9895 Tp/Tn_1/Tn_2: 50918/85/453Eval of epoch 28 => acc:0.9833, loss:0.0117Epoch 24: CosineAnnealingDecay set learning rate to 0.0003520365877844011.Epoch 29: LinearWarmup set learning rate to 0.0003520365877844011.2024-07-27 14:16:48 || Epoch 29 start:epoch:29, batch_id:0, loss:0.0044, acc:0.9883 Tp/Tn_1/Tn_2: 253/1/2epoch:29, batch_id:50, loss:0.0087, acc:0.9891 Tp/Tn_1/Tn_2: 12914/22/120epoch:29, batch_id:100, loss:0.0026, acc:0.9901 Tp/Tn_1/Tn_2: 25600/35/221epoch:29, batch_id:150, loss:0.0053, acc:0.9900 Tp/Tn_1/Tn_2: 38271/59/326epoch:29, batch_id:200, loss:0.0129, acc:0.9900 Tp/Tn_1/Tn_2: 50942/79/435Eval of epoch 29 => acc:0.9838, loss:0.0116Saved best model of epoch29, acc 0.9838, save path "./runs"Epoch 25: CosineAnnealingDecay set learning rate to 0.0003155717009792806.Epoch 30: LinearWarmup set learning rate to 0.0003155717009792806.2024-07-27 14:19:02 || Epoch 30 start:epoch:30, batch_id:0, loss:0.0021, acc:0.9961 Tp/Tn_1/Tn_2: 255/0/1epoch:30, batch_id:50, loss:0.0052, acc:0.9897 Tp/Tn_1/Tn_2: 12921/20/115epoch:30, batch_id:100, loss:0.0035, acc:0.9902 Tp/Tn_1/Tn_2: 25602/46/208epoch:30, batch_id:150, loss:0.0025, acc:0.9903 Tp/Tn_1/Tn_2: 38281/73/302epoch:30, batch_id:200, loss:0.0030, acc:0.9902 Tp/Tn_1/Tn_2: 50954/88/414Eval of epoch 30 => acc:0.9839, loss:0.0116Saved best model of epoch30, acc 0.9839, save path "./runs"Epoch 26: CosineAnnealingDecay set learning rate to 0.0002802747026289244.Epoch 31: LinearWarmup set learning rate to 0.0002802747026289244.2024-07-27 14:21:17 || Epoch 31 start:epoch:31, batch_id:0, loss:0.0085, acc:0.9844 Tp/Tn_1/Tn_2: 252/0/4epoch:31, batch_id:50, loss:0.0040, acc:0.9926 Tp/Tn_1/Tn_2: 12960/7/89epoch:31, batch_id:100, loss:0.0104, acc:0.9915 Tp/Tn_1/Tn_2: 25636/32/188epoch:31, batch_id:150, loss:0.0017, acc:0.9908 Tp/Tn_1/Tn_2: 38300/55/301epoch:31, batch_id:200, loss:0.0060, acc:0.9902 Tp/Tn_1/Tn_2: 50952/81/423Eval of epoch 31 => acc:0.9842, loss:0.0115Saved best model of epoch31, acc 0.9842, save path "./runs"Epoch 27: CosineAnnealingDecay set learning rate to 0.0002463632104656054.Epoch 32: LinearWarmup set learning rate to 0.0002463632104656054.2024-07-27 14:23:36 || Epoch 32 start:epoch:32, batch_id:0, loss:0.0040, acc:0.9961 Tp/Tn_1/Tn_2: 255/0/1epoch:32, batch_id:50, loss:0.0020, acc:0.9917 Tp/Tn_1/Tn_2: 12947/19/90epoch:32, batch_id:100, loss:0.0056, acc:0.9910 Tp/Tn_1/Tn_2: 25623/36/197epoch:32, batch_id:150, loss:0.0016, acc:0.9906 Tp/Tn_1/Tn_2: 38291/57/308epoch:32, batch_id:200, loss:0.0043, acc:0.9906 Tp/Tn_1/Tn_2: 50974/80/402Eval of epoch 32 => acc:0.9838, loss:0.0115Epoch 28: CosineAnnealingDecay set learning rate to 0.00021404630011522585.Epoch 33: LinearWarmup set learning rate to 0.00021404630011522585.2024-07-27 14:25:51 || Epoch 33 start:epoch:33, batch_id:0, loss:0.0026, acc:0.9961 Tp/Tn_1/Tn_2: 255/0/1epoch:33, batch_id:50, loss:0.0023, acc:0.9904 Tp/Tn_1/Tn_2: 12931/23/102epoch:33, batch_id:100, loss:0.0130, acc:0.9909 Tp/Tn_1/Tn_2: 25620/41/195epoch:33, batch_id:150, loss:0.0057, acc:0.9908 Tp/Tn_1/Tn_2: 38302/55/299epoch:33, batch_id:200, loss:0.0029, acc:0.9911 Tp/Tn_1/Tn_2: 50999/74/383Eval of epoch 33 => acc:0.9839, loss:0.0115Epoch 29: CosineAnnealingDecay set learning rate to 0.00018352321607655915.Epoch 34: LinearWarmup set learning rate to 0.00018352321607655915.2024-07-27 14:28:06 || Epoch 34 start:epoch:34, batch_id:0, loss:0.0104, acc:0.9883 Tp/Tn_1/Tn_2: 253/1/2epoch:34, batch_id:50, loss:0.0019, acc:0.9918 Tp/Tn_1/Tn_2: 12949/21/86epoch:34, batch_id:100, loss:0.0054, acc:0.9913 Tp/Tn_1/Tn_2: 25632/40/184epoch:34, batch_id:150, loss:0.0043, acc:0.9911 Tp/Tn_1/Tn_2: 38312/52/292epoch:34, batch_id:200, loss:0.0021, acc:0.9910 Tp/Tn_1/Tn_2: 50994/75/387Eval of epoch 34 => acc:0.9836, loss:0.0114Epoch 30: CosineAnnealingDecay set learning rate to 0.000154982143312659.Epoch 35: LinearWarmup set learning rate to 0.000154982143312659.2024-07-27 14:30:23 || Epoch 35 start:epoch:35, batch_id:0, loss:0.0033, acc:0.9961 Tp/Tn_1/Tn_2: 255/1/0epoch:35, batch_id:50, loss:0.0076, acc:0.9917 Tp/Tn_1/Tn_2: 12948/18/90epoch:35, batch_id:100, loss:0.0009, acc:0.9917 Tp/Tn_1/Tn_2: 25642/38/176epoch:35, batch_id:150, loss:0.0035, acc:0.9915 Tp/Tn_1/Tn_2: 38328/57/271epoch:35, batch_id:200, loss:0.0037, acc:0.9914 Tp/Tn_1/Tn_2: 51013/72/371Eval of epoch 35 => acc:0.9843, loss:0.0113Saved best model of epoch35, acc 0.9843, save path "./runs"Epoch 31: CosineAnnealingDecay set learning rate to 0.0001285990470279847.Epoch 36: LinearWarmup set learning rate to 0.0001285990470279847.2024-07-27 14:32:39 || Epoch 36 start:epoch:36, batch_id:0, loss:0.0057, acc:0.9883 Tp/Tn_1/Tn_2: 253/0/3epoch:36, batch_id:50, loss:0.0112, acc:0.9913 Tp/Tn_1/Tn_2: 12942/22/92epoch:36, batch_id:100, loss:0.0022, acc:0.9909 Tp/Tn_1/Tn_2: 25621/42/193epoch:36, batch_id:150, loss:0.0024, acc:0.9913 Tp/Tn_1/Tn_2: 38319/55/282epoch:36, batch_id:200, loss:0.0041, acc:0.9913 Tp/Tn_1/Tn_2: 51007/76/373Eval of epoch 36 => acc:0.9841, loss:0.0113Epoch 32: CosineAnnealingDecay set learning rate to 0.00010453658778440107.Epoch 37: LinearWarmup set learning rate to 0.00010453658778440107.2024-07-27 14:34:57 || Epoch 37 start:epoch:37, batch_id:0, loss:0.0035, acc:0.9961 Tp/Tn_1/Tn_2: 255/0/1epoch:37, batch_id:50, loss:0.0025, acc:0.9913 Tp/Tn_1/Tn_2: 12943/18/95epoch:37, batch_id:100, loss:0.0018, acc:0.9918 Tp/Tn_1/Tn_2: 25644/36/176epoch:37, batch_id:150, loss:0.0049, acc:0.9919 Tp/Tn_1/Tn_2: 38341/52/263epoch:37, batch_id:200, loss:0.0031, acc:0.9915 Tp/Tn_1/Tn_2: 51020/71/365Eval of epoch 37 => acc:0.9843, loss:0.0113Saved best model of epoch37, acc 0.9843, save path "./runs"Epoch 33: CosineAnnealingDecay set learning rate to 8.294311864472437e-05.Epoch 38: LinearWarmup set learning rate to 8.294311864472437e-05.2024-07-27 14:37:16 || Epoch 38 start:epoch:38, batch_id:0, loss:0.0019, acc:0.9961 Tp/Tn_1/Tn_2: 255/0/1epoch:38, batch_id:50, loss:0.0070, acc:0.9921 Tp/Tn_1/Tn_2: 12953/16/87epoch:38, batch_id:100, loss:0.0009, acc:0.9921 Tp/Tn_1/Tn_2: 25653/41/162epoch:38, batch_id:150, loss:0.0053, acc:0.9922 Tp/Tn_1/Tn_2: 38356/56/244epoch:38, batch_id:200, loss:0.0121, acc:0.9921 Tp/Tn_1/Tn_2: 51048/72/336Eval of epoch 38 => acc:0.9839, loss:0.0113Epoch 34: CosineAnnealingDecay set learning rate to 6.395177052675794e-05.Epoch 39: LinearWarmup set learning rate to 6.395177052675794e-05.2024-07-27 14:39:32 || Epoch 39 start:epoch:39, batch_id:0, loss:0.0059, acc:0.9883 Tp/Tn_1/Tn_2: 253/0/3epoch:39, batch_id:50, loss:0.0037, acc:0.9926 Tp/Tn_1/Tn_2: 12960/17/79epoch:39, batch_id:100, loss:0.0056, acc:0.9923 Tp/Tn_1/Tn_2: 25657/30/169epoch:39, batch_id:150, loss:0.0050, acc:0.9924 Tp/Tn_1/Tn_2: 38361/48/247epoch:39, batch_id:200, loss:0.0080, acc:0.9920 Tp/Tn_1/Tn_2: 51044/61/351Eval of epoch 39 => acc:0.9843, loss:0.0113Epoch 35: CosineAnnealingDecay set learning rate to 4.7679631406913064e-05.Epoch 40: LinearWarmup set learning rate to 4.7679631406913064e-05.登录后复制验证，测试
当batchsize从256设置为1时，验证集的准确率从95.23%降低到93.96%，可能是网络中如下代码的问题
# line 103f_pow = paddle.pow(f, 2)f_mean = paddle.mean(f_pow)f = paddle.divide(f, f_mean)登录后复制这里的mean方法与batch耦合，可以考虑限制维度来解耦（只在每个batch内做平均）比如：
f_mean = paddle.mean(f_pow, axis=[1,2,3], keepdim=True)登录后复制模型对batch解耦后，改变batchsize大小，不会影响最终精度
验证
In [9]import paddle.vision.transforms as Tfrom paddle.io import DataLoaderimport timefrom statistics import mean# 参数定义IMGSIZE = (94, 24)IMGDIR = 'rec_images/data'VALIDFILE = 'rec_images/valid.txt'LPRMAXLEN = 18EVALBATCHSIZE = 1 # batch sizeNUMWORKERS = 2# 图片预处理eval_transforms = T.Compose([ T.ToTensor(data_format='CHW'), T.Normalize( [0.5, 0.5, 0.5], # 在totensor的时候已经将图片缩放到0-1 [0.5, 0.5, 0.5], data_format='CHW' ), ])# 数据加载eval_data_set = LprnetDataloader(IMGDIR, VALIDFILE, eval_transforms)eval_loader = DataLoader( eval_data_set, batch_size=EVALBATCHSIZE, shuffle=False, num_workers=NUMWORKERS, drop_last=False, collate_fn=collate_fn)# 定义lossloss_func = nn.CTCLoss(len(CHARS)-1)# input_length, loss计算需要input_length = np.ones(shape=TRAINBATCHSIZE) * LPRMAXLENinput_length = paddle.to_tensor(input_length, dtype='int64')# LPRNet网络，添加模型权重model = LPRNet(LPRMAXLEN, len(CHARS), DROPOUT)load_pretrained(model, 'runs/lprnet_best_2') # accacc_eval = ACC()# 验证# evalwith paddle.no_grad(): model.eval() loss_list = [] for batch_id, bath_data in enumerate(eval_loader): img_data, label_data, label_lens = bath_data predict = model(img_data) logits = paddle.transpose(predict, (2,0,1)) loss = loss_func(logits, label_data, input_length, label_lens) acc_eval.batch_update(label_data, label_lens, predict) loss_list.append(loss.item()) print(f'Eval from {VALIDFILE} => acc:{acc_eval.acc:.4f}, loss:{mean(loss_list):.4f}') acc_eval.clear()登录后复制params loading...load runs/lprnet_best_2.pdparams success...Eval from rec_images/valid.txt => acc:0.9843, loss:0.0114登录后复制预测结果可视化
这里仍然是以动态图的方式进行预测，想要部署的话建议转静态图
https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/jit/index_cn.html
In [11]"""此部分为测试的可视化代码, 后处理可参考"""import cv2import matplotlib.pyplot as pltimport numpy as npimport paddle%matplotlib inlineimg_path = 'rec_images/data/皖AF358Z.jpg'img_data = cv2.imread(img_path)img_data = img_data[:,:,::-1] # BGR to RGBplt.imshow(img_data)plt.axis('off')plt.show()# 数据前处理img_data = cv2.resize(img_data,(94, 24))img_data = (img_data - 127.5) / 127.5 # 归一化img_data = np.transpose(img_data, (2,0,1)) # HWC to CHWimg_data = np.expand_dims(img_data, 0) # to BCHWimg_tensor = paddle.to_tensor(img_data, dtype='float32') # shape == [1, 3, 24, 94]print(img_tensor.shape)# 加载模型, 预测CHARS = ['京', '沪', '津', '渝', '冀', '晋', '蒙', '辽', '吉', '黑', '苏', '浙', '皖', '闽', '赣', '鲁', '豫', '鄂', '湘', '粤', '桂', '琼', '川', '贵', '云', '藏', '陕', '甘', '青', '宁', '新', '0', '1', '2', '3', '4', '5', '6', '7', '8', '9', 'A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'J', 'K', 'L', 'M', 'N', 'P', 'Q', 'R', 'S', 'T', 'U', 'V', 'W', 'X', 'Y', 'Z', 'I', 'O', '-' ]LPRMAXLEN = 18model = LPRNet(LPRMAXLEN, len(CHARS), dropout_rate=0)load_pretrained(model, 'runs/lprnet_best_2')out_data = model(img_tensor) # out_data.shape == [1, 68, 18]# 后处理，单张图片数据def reprocess(pred): pred_data = pred[0] pred_label = np.argmax(pred_data, axis=0) no_repeat_blank_label = [] pre_c = pred_label[0] if pre_c != len(CHARS) - 1: # 非空白 no_repeat_blank_label.append(pre_c) for c in pred_label: # dropout repeate label and blank label if (pre_c == c) or (c == len(CHARS) - 1): if c == len(CHARS) - 1: pre_c = c continue no_repeat_blank_label.append(c) pre_c = c char_list = [CHARS[i] for i in no_repeat_blank_label] return ''.join(char_list)rep_result = reprocess(out_data)print(rep_result) # 皖AF358Z登录后复制登录后复制登录后复制[1, 3, 24, 94]params loading...load runs/lprnet_best_2.pdparams success...皖AF358Z登录后复制TODO
数据集分布不均，这里只用了ccpd2019的蓝牌，可以使用ccpd2020包含绿牌数据
大部分车牌都是“皖”，可以适当添加其他省份的车牌数据
3. 车牌裁切没有做矫正，想提高精度，可考虑加上车牌的矫正算法
4. 本车牌识别网络模型与batch数据耦合，可以尝试解耦后再训练
网络模型已经固定了输出序列的长度18，考虑修改为能自定义长度，让模型能适用于更多场景模型导出
导出onnx
这里将模型从动态图导出onnx文件, 直接使用api:paddle.onnx.export
https://www.paddlepaddle.org.cn/documentation/docs/zh/api/paddle/onnx/export_cn.html
In [12]model = LPRNet(18, 68, dropout_rate=0)load_pretrained(model, 'runs/lprnet_best_2')save_path = 'save_onnx/lprnet' # 需要保存的路径x_spec = paddle.static.InputSpec([1, 3, 24, 94], 'float32', 'image') paddle.onnx.export(model, save_path, input_spec=[x_spec], opset_version=11)登录后复制params loading...load runs/lprnet_best_2.pdparams success...登录后复制/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/nn/layer/norm.py:712: UserWarning: When training, we now always track global mean and variance. "When training, we now always track global mean and variance."/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/nn/layer/norm.py:712: UserWarning: When training, we now always track global mean and variance. "When training, we now always track global mean and variance."/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/nn/layer/norm.py:712: UserWarning: When training, we now always track global mean and variance. "When training, we now always track global mean and variance."/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/nn/layer/norm.py:712: UserWarning: When training, we now always track global mean and variance. "When training, we now always track global mean and variance."/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/nn/layer/norm.py:712: UserWarning: When training, we now always track global mean and variance. "When training, we now always track global mean and variance."登录后复制2024-07-27 15:03:51 [INFO]Static PaddlePaddle model saved in save_onnx/paddle_model_static_onnx_temp_dir.[Paddle2ONNX] Start to parse PaddlePaddle model...[Paddle2ONNX] Model file path: save_onnx/paddle_model_static_onnx_temp_dir/model.pdmodel[Paddle2ONNX] Paramters file path: save_onnx/paddle_model_static_onnx_temp_dir/model.pdiparams[Paddle2ONNX] Start to parsing Paddle model...[Paddle2ONNX] Use opset_version = 11 for ONNX export.[Paddle2ONNX] PaddlePaddle model is exported as ONNX format now.2024-07-27 15:03:52 [INFO]ONNX model saved in save_onnx/lprnet.onnx.登录后复制onnx测试
可以在https://netron.app/ 中查看可视化结构
更多参考：https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/advanced/model_to_onnx_cn.html
In [ ]!pip install onnx==1.10.2!pip install onnxruntime==1.9.0登录后复制In [14]"""检查onnx是否合理，模型的版本、图的结构、节点及其输入和输出"""import onnxonnx_model = onnx.load("save_onnx/lprnet.onnx")check = onnx.checker.check_model(onnx_model)print('check: ', check)登录后复制check: None登录后复制In [15]# 导入所需的库import numpy as npimport onnxruntimeimport paddle# 随机生成输入，用于验证 Paddle 和 ONNX 的推理结果是否一致x = np.random.random((1, 3, 24, 94)).astype('float32')# predict by ONNXRuntimeonnx_path = "save_onnx/lprnet.onnx"ort_sess = onnxruntime.InferenceSession(onnx_path)ort_inputs = {ort_sess.get_inputs()[0].name: x}ort_outs = ort_sess.run(None, ort_inputs)print("Exported model has been predicted by ONNXRuntime!")# predict by Paddlemodel = paddle.jit.load("save_onnx/paddle_model_static_onnx_temp_dir/model") # 上一步中导出onnx的时候会保存静态图文件到输出目录model.eval()paddle_input = paddle.to_tensor(x)paddle_outs = model(paddle_input)print("Original model has been predicted by Paddle!")# compare ONNXRuntime and Paddle resultsnp.testing.assert_allclose(ort_outs[0], paddle_outs.numpy(), rtol=1.0, atol=1e-05)print("The difference of results between ONNXRuntime and Paddle looks good!")登录后复制Exported model has been predicted by ONNXRuntime!Original model has been predicted by Paddle!The difference of results between ONNXRuntime and Paddle looks good!登录后复制onnx推理
推理与上面相同，只是添加了实际数据的前处理和模型输出的后处理部分
In [16]import onnxruntimeimport cv2import matplotlib.pyplot as pltimport numpy as np%matplotlib inline# 数据预处理img_path = 'rec_images/data/皖AF358Z.jpg'img_data = cv2.imread(img_path)img_data = img_data[:,:,::-1] # BGR to RGBplt.imshow(img_data)plt.axis('off')plt.show()img_data = cv2.resize(img_data,(94, 24))img_data = (img_data - 127.5) / 127.5 # 归一化img_data = np.transpose(img_data, (2,0,1)) # HWC to CHWimg_data = np.expand_dims(img_data, 0) # to BCHWnp_data = np.array(img_data, dtype=np.float32)# 加载 ONNX 模型生成推理用 sessonnx_path = "save_onnx/lprnet.onnx"sess = onnxruntime.InferenceSession(onnx_path)# 使用 ONNXRuntime 推理ort_inputs = {sess.get_inputs()[0].name: np_data}result, = sess.run(None, ort_inputs)# 推理结果后处理CHARS = ['京', '沪', '津', '渝', '冀', '晋', '蒙', '辽', '吉', '黑', '苏', '浙', '皖', '闽', '赣', '鲁', '豫', '鄂', '湘', '粤', '桂', '琼', '川', '贵', '云', '藏', '陕', '甘', '青', '宁', '新', '0', '1', '2', '3', '4', '5', '6', '7', '8', '9', 'A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'J', 'K', 'L', 'M', 'N', 'P', 'Q', 'R', 'S', 'T', 'U', 'V', 'W', 'X', 'Y', 'Z', 'I', 'O', '-' ]def reprocess(pred): pred_data = pred[0] pred_label = np.argmax(pred_data, axis=0) no_repeat_blank_label = [] pre_c = pred_label[0] if pre_c != len(CHARS) - 1: # 非空白 no_repeat_blank_label.append(pre_c) for c in pred_label: # dropout repeate label and blank label if (pre_c == c) or (c == len(CHARS) - 1): if c == len(CHARS) - 1: pre_c = c continue no_repeat_blank_label.append(c) pre_c = c char_list = [CHARS[i] for i in no_repeat_blank_label] return ''.join(char_list)plate_str = reprocess(result)print(plate_str) # 皖AF358Z登录后复制登录后复制登录后复制皖AF358Z登录后复制
免责声明
游乐网为非赢利性网站，所展示的游戏/软件/文章内容均来自于互联网或第三方用户上传分享，版权归原作者所有，本站不承担相应法律责任。如您发现有涉嫌抄袭侵权的内容，请联系youleyoucom@outlook.com。
同类文章
余承东：华为招聘顶级AI人才，共攀通用人工智能高峰
华为近日通过最新招聘渠道发布“全球顶尖AI人才招募计划”，旨在组建一支具备国际竞争力的AI研发团队，加速推进大模型技术创新，向通用人工智能（AGI）领域发起冲刺。公司常务董事、终端BG董事长余承东在
2025-10-21.
JEDEC发布SOCAMM2标准：为AI数据中心提供9600MT/s高速内存
JEDEC固态技术协会近日发布消息，面向数据中心AI应用场景的SOCAMM2小尺寸内存模块外形规范即将完成标准化进程，其对应的JESD328技术标准已进入最终审定阶段。这项基于LPDDR5X DRA
2025-10-21.
余承东诚邀AI青年才俊，华为广发人工智能招募令
华为近日正式启动“全球顶尖AI人才招募计划”，面向海内外高校广发英雄帖，旨在吸引一批具有创新潜力的年轻人才投身人工智能领域。此次招聘的覆盖范围十分广泛。国内高校方面，2026年1月1日至2026年1
2025-10-21.
2025上半年中国AI用户破5亿，高学历中青年成主力军
2025年10月18日，第六届中国互联网基础资源大会在北京召开。会上，中国互联网络信息中心正式发布《生成式人工智能应用发展报告（2025）》，数据显示，截至2025年6月，我国生成式人工智能用户规模
2025-10-21.
马斯克预测Grok下月有10%可能实现AGI，即将到来的里程碑
埃隆·马斯克近日在社交媒体平台X（原Twitter）上公开表示，其旗下人工智能公司xAI即将推出的新一代大型语言模型Grok 5，存在10%的可能性实现通用人工智能（AGI）。他同时强调，这一概率仍
2025-10-21.
相关攻略
热门教程
更多- 游戏攻略
- 安卓教程
- 苹果教程
- 电脑教程
植物大战僵尸在线版入口：网页版秒玩直接畅游
发布于 2025-10-21
在线下中国象棋，网页版对弈入口与技巧
发布于 2025-10-21
三国归心国家队阵容搭配全解析
发布于 2025-10-21
鸣潮成就荆刺之鸟怎么完成：具体位置与攻略
发布于 2025-10-21
三国归心阵营战玩法指南：5步掌握实战技巧
发布于 2025-10-21
全球游戏玩家平均年龄新趋势：41岁玩家占比攀升
发布于 2025-10-21
最终幻想7蒂法3D穿搭指南：新服装塑造惊艳身型
发布于 2025-10-21
前PlayStation高管质疑XGP涨价：背后逻辑深度解析
发布于 2025-10-21
《释放阿凡达》Steam公测：印度首款3A，1060显卡可流畅运行
发布于 2025-10-21
初音未来联动支付宝：蓝花社区化身虚拟偶像应援圣地
发布于 2025-10-21
深圳联通上门办理eSIM卡：预约指南与流程详解
发布于 2025-10-21
苹果多项服务中断：TestFlight及音乐播客受波及
发布于 2025-10-21
Counterpoint：iPhone 17中美上市首10天销量较上代增长14%
发布于 2025-10-21
苹果eSIM转换功能即将上线，国行iPhone换机更便捷
发布于 2025-10-21
苹果M4 MacBook Air国补4773元：学生无需验证直接购机
发布于 2025-10-21
Windows 11紧急更新修复USB失灵！键鼠无法使用解决办法
发布于 2025-10-21
Win11年度版本对比：24H2与25H2核心差异解析
发布于 2025-10-21
Redmi投影仪4 Pro发布：1499元起，600CVIA流明百吋巨幕
发布于 2025-10-21
韩国廉价硅脂慎用：硬件刺客导致CPU损坏的真相
发布于 2025-10-21
AMD锐龙9000G APU即将发布：核显性能仍存短板
发布于 2025-10-21
