高层 API 实现 RepVGG 模型微调

本文介绍如何用Paddle 2.0高层API微调RepVGG模型。先导入必要包，构建RepVGG模型及模块，封装模型预设配置，通过paddle.Model配置模型，加载Cifar10数据集，经训练后用model.predict_batch对图片预测，还可借助VisualDL可视化训练数据。

☞☞☞AI 智能聊天, 问答助手, AI 智能搜索, 免费无限量使用 DeepSeek R1 模型☜☜☜

引入

上一个项目介绍了【如何构建 RepVGG 模型】自从去年底尝试过高层 API 的使用之后，就没怎么使用过这个功能了Paddle 2.0 正式版更新之后，高层 API 功能也更加完善了所以，本次就介绍如何使用高层 API 完成 RepVGG 模型微调

相关项目

Paddle 2.0 图像分类微调的相关项目PaddleClas：复现 Vision Transformer 实现鲜花图像分类任务的微调PaddleHub：十行代码完成图像分类任务的微调PaddleHapi：高层 API 实现 RepVGG 模型微调

开始

废话少说，直接开始

准备

导入必要的包可选择开启静态图模式In [1]

# 导入 Paddleimport paddleimport paddle.nn as nnfrom paddle.static import InputSpecfrom paddle.vision.datasets import Cifar10from paddle.vision.transforms import Resize, CenterCrop, Transpose, Normalize, Compose# 导入其他包import cv2import numpy as npfrom IPython.display import Image# 开启静态图# 也可以直接用动态图进行模型训练paddle.enable_static()

模型组网

常规的继承 paddle.nn.Layer 来进行模型组网更多细节请参考【如何构建 RepVGG 模型】In [2]

# 卷积 + 批归一化class ConvBN(nn.Layer):    def __init__(self, in_channels, out_channels, kernel_size, stride, padding, groups=1):        super(ConvBN, self).__init__()        self.conv = nn.Conv2D(in_channels=in_channels, out_channels=out_channels,                              kernel_size=kernel_size, stride=stride, padding=padding, groups=groups, bias_attr=False)        self.bn = nn.BatchNorm2D(num_features=out_channels)    def forward(self, x):        y = self.conv(x)        y = self.bn(y)        return y# RepVGG 模块class RepVGGBlock(nn.Layer):    def __init__(self, in_channels, out_channels, kernel_size,                 stride=1, padding=0, dilation=1, groups=1, padding_mode='zeros'):        super(RepVGGBlock, self).__init__()        self.in_channels = in_channels        self.out_channels = out_channels        self.kernel_size = kernel_size        self.stride = stride        self.padding = padding        self.dilation = dilation        self.groups = groups        self.padding_mode = padding_mode        assert kernel_size == 3        assert padding == 1        padding_11 = padding - kernel_size // 2        self.nonlinearity = nn.ReLU()        self.rbr_identity = nn.BatchNorm2D(            num_features=in_channels) if out_channels == in_channels and stride == 1 else None        self.rbr_dense = ConvBN(in_channels=in_channels, out_channels=out_channels,                                kernel_size=kernel_size, stride=stride, padding=padding, groups=groups)        self.rbr_1x1 = ConvBN(in_channels=in_channels, out_channels=out_channels,                              kernel_size=1, stride=stride, padding=padding_11, groups=groups)    def forward(self, inputs):        if not self.training:            return self.nonlinearity(self.rbr_reparam(inputs))        if self.rbr_identity is None:            id_out = 0        else:            id_out = self.rbr_identity(inputs)        return self.nonlinearity(self.rbr_dense(inputs) + self.rbr_1x1(inputs) + id_out)    def eval(self):        if not hasattr(self, 'rbr_reparam'):            self.rbr_reparam = nn.Conv2D(in_channels=self.in_channels, out_channels=self.out_channels, kernel_size=self.kernel_size, stride=self.stride,                                         padding=self.padding, dilation=self.dilation, groups=self.groups, padding_mode=self.padding_mode)        self.training = False        kernel, bias = self.get_equivalent_kernel_bias()        self.rbr_reparam.weight.set_value(kernel)        self.rbr_reparam.bias.set_value(bias)        for layer in self.sublayers():            layer.eval()    def get_equivalent_kernel_bias(self):        kernel3x3, bias3x3 = self._fuse_bn_tensor(self.rbr_dense)        kernel1x1, bias1x1 = self._fuse_bn_tensor(self.rbr_1x1)        kernelid, biasid = self._fuse_bn_tensor(self.rbr_identity)        return kernel3x3 + self._pad_1x1_to_3x3_tensor(kernel1x1) + kernelid, bias3x3 + bias1x1 + biasid    def _pad_1x1_to_3x3_tensor(self, kernel1x1):        if kernel1x1 is None:            return 0        else:            return nn.functional.pad(kernel1x1, [1, 1, 1, 1])    def _fuse_bn_tensor(self, branch):        if branch is None:            return 0, 0        if isinstance(branch, ConvBN):            kernel = branch.conv.weight            running_mean = branch.bn._mean            running_var = branch.bn._variance            gamma = branch.bn.weight            beta = branch.bn.bias            eps = branch.bn._epsilon        else:            assert isinstance(branch, nn.BatchNorm2D)            if not hasattr(self, 'id_tensor'):                input_dim = self.in_channels // self.groups                kernel_value = np.zeros(                    (self.in_channels, input_dim, 3, 3), dtype=np.float32)                for i in range(self.in_channels):                    kernel_value[i, i % input_dim, 1, 1] = 1                self.id_tensor = paddle.to_tensor(kernel_value)            kernel = self.id_tensor            running_mean = branch._mean            running_var = branch._variance            gamma = branch.weight            beta = branch.bias            eps = branch._epsilon        std = (running_var + eps).sqrt()        t = (gamma / std).reshape((-1, 1, 1, 1))        return kernel * t, beta - running_mean * gamma / std# RepVGG 模型class RepVGG(nn.Layer):    def __init__(self, num_blocks, width_multiplier=None, override_groups_map=None, in_channels=3, class_dim=1000):        super(RepVGG, self).__init__()        assert len(width_multiplier) == 4        self.override_groups_map = override_groups_map or dict()        assert 0 not in self.override_groups_map        self.in_planes = min(64, int(64 * width_multiplier[0]))        self.stage0 = RepVGGBlock(            in_channels=in_channels, out_channels=self.in_planes, kernel_size=3, stride=2, padding=1)        self.cur_layer_idx = 1        self.stage1 = self._make_stage(            int(64 * width_multiplier[0]), num_blocks[0], stride=2)        self.stage2 = self._make_stage(            int(128 * width_multiplier[1]), num_blocks[1], stride=2)        self.stage3 = self._make_stage(            int(256 * width_multiplier[2]), num_blocks[2], stride=2)        self.stage4 = self._make_stage(            int(512 * width_multiplier[3]), num_blocks[3], stride=2)        self.gap = nn.AdaptiveAvgPool2D(output_size=1)        self.linear = nn.Linear(int(512 * width_multiplier[3]), class_dim)    def _make_stage(self, planes, num_blocks, stride):        strides = [stride] + [1]*(num_blocks-1)        blocks = []        for stride in strides:            cur_groups = self.override_groups_map.get(self.cur_layer_idx, 1)            blocks.append(RepVGGBlock(in_channels=self.in_planes, out_channels=planes, kernel_size=3,                                      stride=stride, padding=1, groups=cur_groups))            self.in_planes = planes            self.cur_layer_idx += 1        return nn.Sequential(*blocks)    def forward(self, x):        out = self.stage0(x)        out = self.stage1(out)        out = self.stage2(out)        out = self.stage3(out)        out = self.stage4(out)        out = self.gap(out)        out = paddle.flatten(out, start_axis=1)        out = self.linear(out)        return out

模型封装

使用函数形式将模型进行封装，将预设配置固定In [3]

# 模型超参数配置optional_groupwise_layers = [2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26]g2_map = {l: 2 for l in optional_groupwise_layers}g4_map = {l: 4 for l in optional_groupwise_layers}# 各种模型预设配置def RepVGG_A0(**kwargs):    return RepVGG(num_blocks=[2, 4, 14, 1], width_multiplier=[0.75, 0.75, 0.75, 2.5], override_groups_map=None, **kwargs)def RepVGG_A1(**kwargs):    return RepVGG(num_blocks=[2, 4, 14, 1], width_multiplier=[1, 1, 1, 2.5], override_groups_map=None, **kwargs)def RepVGG_A2(**kwargs):    return RepVGG(num_blocks=[2, 4, 14, 1], width_multiplier=[1.5, 1.5, 1.5, 2.75], override_groups_map=None, **kwargs)def RepVGG_B0(**kwargs):    return RepVGG(num_blocks=[4, 6, 16, 1], width_multiplier=[1, 1, 1, 2.5], override_groups_map=None, **kwargs)def RepVGG_B1(**kwargs):    return RepVGG(num_blocks=[4, 6, 16, 1], width_multiplier=[2, 2, 2, 4], override_groups_map=None, **kwargs)def RepVGG_B1g2(**kwargs):    return RepVGG(num_blocks=[4, 6, 16, 1], width_multiplier=[2, 2, 2, 4], override_groups_map=g2_map, **kwargs)def RepVGG_B1g4(**kwargs):    return RepVGG(num_blocks=[4, 6, 16, 1], width_multiplier=[2, 2, 2, 4], override_groups_map=g4_map, **kwargs)def RepVGG_B2(**kwargs):    return RepVGG(num_blocks=[4, 6, 16, 1], width_multiplier=[2.5, 2.5, 2.5, 5], override_groups_map=None, **kwargs)def RepVGG_B2g2(**kwargs):    return RepVGG(num_blocks=[4, 6, 16, 1], width_multiplier=[2.5, 2.5, 2.5, 5], override_groups_map=g2_map, **kwargs)def RepVGG_B2g4(**kwargs):    return RepVGG(num_blocks=[4, 6, 16, 1], width_multiplier=[2.5, 2.5, 2.5, 5], override_groups_map=g4_map, **kwargs)def RepVGG_B3(**kwargs):    return RepVGG(num_blocks=[4, 6, 16, 1], width_multiplier=[3, 3, 3, 5], override_groups_map=None, **kwargs)def RepVGG_B3g2(**kwargs):    return RepVGG(num_blocks=[4, 6, 16, 1], width_multiplier=[3, 3, 3, 5], override_groups_map=g2_map, **kwargs)def RepVGG_B3g4(**kwargs):    return RepVGG(num_blocks=[4, 6, 16, 1], width_multiplier=[3, 3, 3, 5], override_groups_map=g4_map, **kwargs)

模型配置

高层 API 通过 paddle.Model 类来配置模型In [4]

# 设置模型的输入和标签images = InputSpec(shape=[-1, 3, 32, 32], dtype='float32', name='images')labels = InputSpec(shape=[-1], dtype='int64', name='labels')# 初始化模型model = paddle.Model(RepVGG_A0(in_channels=3, class_dim=10), inputs=images, labels=labels)# 加载预训练模型参数model.load(path='data/data69662/RepVGG_A0', skip_mismatch=True, reset_optimizer=True)# 打印模型结构model.summary()# 配置优化器opt = paddle.optimizer.Adam(learning_rate=0.001, parameters=model.parameters())# 配置模型model.prepare(optimizer=opt, loss=nn.CrossEntropyLoss(), metrics=paddle.metric.Accuracy(topk=(1, 5)))

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/hapi/model.py:1206: UserWarning: Skip loading for linear.weight. linear.weight receives a shape [1280, 1000], but the expected shape is [1280, 10].  ("Skip loading for {}. ".format(key) + str(err)))/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/hapi/model.py:1206: UserWarning: Skip loading for linear.bias. linear.bias receives a shape [1000], but the expected shape is [10].  ("Skip loading for {}. ".format(key) + str(err)))/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/hapi/model_summary.py:107: UserWarning: Your model was created in static mode, this may not get correct summary information!  "Your model was created in static mode, this may not get correct summary information!"/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/nn/layer/norm.py:636: UserWarning: When training, we now always track global mean and variance.  "When training, we now always track global mean and variance.")

-------------------------------------------------------------------------------   Layer (type)         Input Shape          Output Shape         Param #    ===============================================================================     Conv2D-1         [[1, 3, 32, 32]]     [1, 48, 16, 16]         1,296        BatchNorm2D-1     [[1, 48, 16, 16]]     [1, 48, 16, 16]          192           ConvBN-1         [[1, 3, 32, 32]]     [1, 48, 16, 16]         1,488          Conv2D-2         [[1, 3, 32, 32]]     [1, 48, 16, 16]          144         BatchNorm2D-2     [[1, 48, 16, 16]]     [1, 48, 16, 16]          192           ConvBN-2         [[1, 3, 32, 32]]     [1, 48, 16, 16]          336            ReLU-1         [[1, 48, 16, 16]]     [1, 48, 16, 16]           0          RepVGGBlock-1      [[1, 3, 32, 32]]     [1, 48, 16, 16]         1,824          Conv2D-3        [[1, 48, 16, 16]]      [1, 48, 8, 8]         20,736        BatchNorm2D-3      [[1, 48, 8, 8]]       [1, 48, 8, 8]           192           ConvBN-3        [[1, 48, 16, 16]]      [1, 48, 8, 8]         20,928          Conv2D-4        [[1, 48, 16, 16]]      [1, 48, 8, 8]          2,304        BatchNorm2D-4      [[1, 48, 8, 8]]       [1, 48, 8, 8]           192           ConvBN-4        [[1, 48, 16, 16]]      [1, 48, 8, 8]          2,496           ReLU-2          [[1, 48, 8, 8]]       [1, 48, 8, 8]            0          RepVGGBlock-2     [[1, 48, 16, 16]]      [1, 48, 8, 8]         23,424        BatchNorm2D-5      [[1, 48, 8, 8]]       [1, 48, 8, 8]           192           Conv2D-5         [[1, 48, 8, 8]]       [1, 48, 8, 8]         20,736        BatchNorm2D-6      [[1, 48, 8, 8]]       [1, 48, 8, 8]           192           ConvBN-5         [[1, 48, 8, 8]]       [1, 48, 8, 8]         20,928          Conv2D-6         [[1, 48, 8, 8]]       [1, 48, 8, 8]          2,304        BatchNorm2D-7      [[1, 48, 8, 8]]       [1, 48, 8, 8]           192           ConvBN-6         [[1, 48, 8, 8]]       [1, 48, 8, 8]          2,496           ReLU-3          [[1, 48, 8, 8]]       [1, 48, 8, 8]            0          RepVGGBlock-3      [[1, 48, 8, 8]]       [1, 48, 8, 8]         23,616          Conv2D-7         [[1, 48, 8, 8]]       [1, 96, 4, 4]         41,472        BatchNorm2D-8      [[1, 96, 4, 4]]       [1, 96, 4, 4]           384           ConvBN-7         [[1, 48, 8, 8]]       [1, 96, 4, 4]         41,856          Conv2D-8         [[1, 48, 8, 8]]       [1, 96, 4, 4]          4,608        BatchNorm2D-9      [[1, 96, 4, 4]]       [1, 96, 4, 4]           384           ConvBN-8         [[1, 48, 8, 8]]       [1, 96, 4, 4]          4,992           ReLU-4          [[1, 96, 4, 4]]       [1, 96, 4, 4]            0          RepVGGBlock-4      [[1, 48, 8, 8]]       [1, 96, 4, 4]         46,848       BatchNorm2D-10      [[1, 96, 4, 4]]       [1, 96, 4, 4]           384           Conv2D-9         [[1, 96, 4, 4]]       [1, 96, 4, 4]         82,944       BatchNorm2D-11      [[1, 96, 4, 4]]       [1, 96, 4, 4]           384           ConvBN-9         [[1, 96, 4, 4]]       [1, 96, 4, 4]         83,328          Conv2D-10        [[1, 96, 4, 4]]       [1, 96, 4, 4]          9,216       BatchNorm2D-12      [[1, 96, 4, 4]]       [1, 96, 4, 4]           384           ConvBN-10        [[1, 96, 4, 4]]       [1, 96, 4, 4]          9,600           ReLU-5          [[1, 96, 4, 4]]       [1, 96, 4, 4]            0          RepVGGBlock-5      [[1, 96, 4, 4]]       [1, 96, 4, 4]         93,312       BatchNorm2D-13      [[1, 96, 4, 4]]       [1, 96, 4, 4]           384           Conv2D-11        [[1, 96, 4, 4]]       [1, 96, 4, 4]         82,944       BatchNorm2D-14      [[1, 96, 4, 4]]       [1, 96, 4, 4]           384           ConvBN-11        [[1, 96, 4, 4]]       [1, 96, 4, 4]         83,328          Conv2D-12        [[1, 96, 4, 4]]       [1, 96, 4, 4]          9,216       BatchNorm2D-15      [[1, 96, 4, 4]]       [1, 96, 4, 4]           384           ConvBN-12        [[1, 96, 4, 4]]       [1, 96, 4, 4]          9,600           ReLU-6          [[1, 96, 4, 4]]       [1, 96, 4, 4]            0          RepVGGBlock-6      [[1, 96, 4, 4]]       [1, 96, 4, 4]         93,312       BatchNorm2D-16      [[1, 96, 4, 4]]       [1, 96, 4, 4]           384           Conv2D-13        [[1, 96, 4, 4]]       [1, 96, 4, 4]         82,944       BatchNorm2D-17      [[1, 96, 4, 4]]       [1, 96, 4, 4]           384           ConvBN-13        [[1, 96, 4, 4]]       [1, 96, 4, 4]         83,328          Conv2D-14        [[1, 96, 4, 4]]       [1, 96, 4, 4]          9,216       BatchNorm2D-18      [[1, 96, 4, 4]]       [1, 96, 4, 4]           384           ConvBN-14        [[1, 96, 4, 4]]       [1, 96, 4, 4]          9,600           ReLU-7          [[1, 96, 4, 4]]       [1, 96, 4, 4]            0          RepVGGBlock-7      [[1, 96, 4, 4]]       [1, 96, 4, 4]         93,312          Conv2D-15        [[1, 96, 4, 4]]       [1, 192, 2, 2]        165,888      BatchNorm2D-19      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-15        [[1, 96, 4, 4]]       [1, 192, 2, 2]        166,656         Conv2D-16        [[1, 96, 4, 4]]       [1, 192, 2, 2]        18,432       BatchNorm2D-20      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-16        [[1, 96, 4, 4]]       [1, 192, 2, 2]        19,200           ReLU-8          [[1, 192, 2, 2]]      [1, 192, 2, 2]           0          RepVGGBlock-8      [[1, 96, 4, 4]]       [1, 192, 2, 2]        185,856      BatchNorm2D-21      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           Conv2D-17        [[1, 192, 2, 2]]      [1, 192, 2, 2]        331,776      BatchNorm2D-22      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-17        [[1, 192, 2, 2]]      [1, 192, 2, 2]        332,544         Conv2D-18        [[1, 192, 2, 2]]      [1, 192, 2, 2]        36,864       BatchNorm2D-23      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-18        [[1, 192, 2, 2]]      [1, 192, 2, 2]        37,632           ReLU-9          [[1, 192, 2, 2]]      [1, 192, 2, 2]           0          RepVGGBlock-9      [[1, 192, 2, 2]]      [1, 192, 2, 2]        370,944      BatchNorm2D-24      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           Conv2D-19        [[1, 192, 2, 2]]      [1, 192, 2, 2]        331,776      BatchNorm2D-25      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-19        [[1, 192, 2, 2]]      [1, 192, 2, 2]        332,544         Conv2D-20        [[1, 192, 2, 2]]      [1, 192, 2, 2]        36,864       BatchNorm2D-26      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-20        [[1, 192, 2, 2]]      [1, 192, 2, 2]        37,632           ReLU-10         [[1, 192, 2, 2]]      [1, 192, 2, 2]           0         RepVGGBlock-10      [[1, 192, 2, 2]]      [1, 192, 2, 2]        370,944      BatchNorm2D-27      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           Conv2D-21        [[1, 192, 2, 2]]      [1, 192, 2, 2]        331,776      BatchNorm2D-28      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-21        [[1, 192, 2, 2]]      [1, 192, 2, 2]        332,544         Conv2D-22        [[1, 192, 2, 2]]      [1, 192, 2, 2]        36,864       BatchNorm2D-29      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-22        [[1, 192, 2, 2]]      [1, 192, 2, 2]        37,632           ReLU-11         [[1, 192, 2, 2]]      [1, 192, 2, 2]           0         RepVGGBlock-11      [[1, 192, 2, 2]]      [1, 192, 2, 2]        370,944      BatchNorm2D-30      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           Conv2D-23        [[1, 192, 2, 2]]      [1, 192, 2, 2]        331,776      BatchNorm2D-31      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-23        [[1, 192, 2, 2]]      [1, 192, 2, 2]        332,544         Conv2D-24        [[1, 192, 2, 2]]      [1, 192, 2, 2]        36,864       BatchNorm2D-32      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-24        [[1, 192, 2, 2]]      [1, 192, 2, 2]        37,632           ReLU-12         [[1, 192, 2, 2]]      [1, 192, 2, 2]           0         RepVGGBlock-12      [[1, 192, 2, 2]]      [1, 192, 2, 2]        370,944      BatchNorm2D-33      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           Conv2D-25        [[1, 192, 2, 2]]      [1, 192, 2, 2]        331,776      BatchNorm2D-34      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-25        [[1, 192, 2, 2]]      [1, 192, 2, 2]        332,544         Conv2D-26        [[1, 192, 2, 2]]      [1, 192, 2, 2]        36,864       BatchNorm2D-35      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-26        [[1, 192, 2, 2]]      [1, 192, 2, 2]        37,632           ReLU-13         [[1, 192, 2, 2]]      [1, 192, 2, 2]           0         RepVGGBlock-13      [[1, 192, 2, 2]]      [1, 192, 2, 2]        370,944      BatchNorm2D-36      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           Conv2D-27        [[1, 192, 2, 2]]      [1, 192, 2, 2]        331,776      BatchNorm2D-37      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-27        [[1, 192, 2, 2]]      [1, 192, 2, 2]        332,544         Conv2D-28        [[1, 192, 2, 2]]      [1, 192, 2, 2]        36,864       BatchNorm2D-38      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-28        [[1, 192, 2, 2]]      [1, 192, 2, 2]        37,632           ReLU-14         [[1, 192, 2, 2]]      [1, 192, 2, 2]           0         RepVGGBlock-14      [[1, 192, 2, 2]]      [1, 192, 2, 2]        370,944      BatchNorm2D-39      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           Conv2D-29        [[1, 192, 2, 2]]      [1, 192, 2, 2]        331,776      BatchNorm2D-40      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-29        [[1, 192, 2, 2]]      [1, 192, 2, 2]        332,544         Conv2D-30        [[1, 192, 2, 2]]      [1, 192, 2, 2]        36,864       BatchNorm2D-41      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-30        [[1, 192, 2, 2]]      [1, 192, 2, 2]        37,632           ReLU-15         [[1, 192, 2, 2]]      [1, 192, 2, 2]           0         RepVGGBlock-15      [[1, 192, 2, 2]]      [1, 192, 2, 2]        370,944      BatchNorm2D-42      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           Conv2D-31        [[1, 192, 2, 2]]      [1, 192, 2, 2]        331,776      BatchNorm2D-43      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-31        [[1, 192, 2, 2]]      [1, 192, 2, 2]        332,544         Conv2D-32        [[1, 192, 2, 2]]      [1, 192, 2, 2]        36,864       BatchNorm2D-44      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-32        [[1, 192, 2, 2]]      [1, 192, 2, 2]        37,632           ReLU-16         [[1, 192, 2, 2]]      [1, 192, 2, 2]           0         RepVGGBlock-16      [[1, 192, 2, 2]]      [1, 192, 2, 2]        370,944      BatchNorm2D-45      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           Conv2D-33        [[1, 192, 2, 2]]      [1, 192, 2, 2]        331,776      BatchNorm2D-46      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-33        [[1, 192, 2, 2]]      [1, 192, 2, 2]        332,544         Conv2D-34        [[1, 192, 2, 2]]      [1, 192, 2, 2]        36,864       BatchNorm2D-47      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-34        [[1, 192, 2, 2]]      [1, 192, 2, 2]        37,632           ReLU-17         [[1, 192, 2, 2]]      [1, 192, 2, 2]           0         RepVGGBlock-17      [[1, 192, 2, 2]]      [1, 192, 2, 2]        370,944      BatchNorm2D-48      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           Conv2D-35        [[1, 192, 2, 2]]      [1, 192, 2, 2]        331,776      BatchNorm2D-49      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-35        [[1, 192, 2, 2]]      [1, 192, 2, 2]        332,544         Conv2D-36        [[1, 192, 2, 2]]      [1, 192, 2, 2]        36,864       BatchNorm2D-50      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-36        [[1, 192, 2, 2]]      [1, 192, 2, 2]        37,632           ReLU-18         [[1, 192, 2, 2]]      [1, 192, 2, 2]           0         RepVGGBlock-18      [[1, 192, 2, 2]]      [1, 192, 2, 2]        370,944      BatchNorm2D-51      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           Conv2D-37        [[1, 192, 2, 2]]      [1, 192, 2, 2]        331,776      BatchNorm2D-52      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-37        [[1, 192, 2, 2]]      [1, 192, 2, 2]        332,544         Conv2D-38        [[1, 192, 2, 2]]      [1, 192, 2, 2]        36,864       BatchNorm2D-53      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-38        [[1, 192, 2, 2]]      [1, 192, 2, 2]        37,632           ReLU-19         [[1, 192, 2, 2]]      [1, 192, 2, 2]           0         RepVGGBlock-19      [[1, 192, 2, 2]]      [1, 192, 2, 2]        370,944      BatchNorm2D-54      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           Conv2D-39        [[1, 192, 2, 2]]      [1, 192, 2, 2]        331,776      BatchNorm2D-55      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-39        [[1, 192, 2, 2]]      [1, 192, 2, 2]        332,544         Conv2D-40        [[1, 192, 2, 2]]      [1, 192, 2, 2]        36,864       BatchNorm2D-56      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-40        [[1, 192, 2, 2]]      [1, 192, 2, 2]        37,632           ReLU-20         [[1, 192, 2, 2]]      [1, 192, 2, 2]           0         RepVGGBlock-20      [[1, 192, 2, 2]]      [1, 192, 2, 2]        370,944      BatchNorm2D-57      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           Conv2D-41        [[1, 192, 2, 2]]      [1, 192, 2, 2]        331,776      BatchNorm2D-58      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-41        [[1, 192, 2, 2]]      [1, 192, 2, 2]        332,544         Conv2D-42        [[1, 192, 2, 2]]      [1, 192, 2, 2]        36,864       BatchNorm2D-59      [[1, 192, 2, 2]]      [1, 192, 2, 2]          768           ConvBN-42        [[1, 192, 2, 2]]      [1, 192, 2, 2]        37,632           ReLU-21         [[1, 192, 2, 2]]      [1, 192, 2, 2]           0         RepVGGBlock-21      [[1, 192, 2, 2]]      [1, 192, 2, 2]        370,944         Conv2D-43        [[1, 192, 2, 2]]     [1, 1280, 1, 1]       2,211,840     BatchNorm2D-60     [[1, 1280, 1, 1]]     [1, 1280, 1, 1]         5,120          ConvBN-43        [[1, 192, 2, 2]]     [1, 1280, 1, 1]       2,216,960        Conv2D-44        [[1, 192, 2, 2]]     [1, 1280, 1, 1]        245,760      BatchNorm2D-61     [[1, 1280, 1, 1]]     [1, 1280, 1, 1]         5,120          ConvBN-44        [[1, 192, 2, 2]]     [1, 1280, 1, 1]        250,880          ReLU-22        [[1, 1280, 1, 1]]     [1, 1280, 1, 1]           0         RepVGGBlock-22      [[1, 192, 2, 2]]     [1, 1280, 1, 1]       2,467,840   AdaptiveAvgPool2D-1  [[1, 1280, 1, 1]]     [1, 1280, 1, 1]           0            Linear-1           [[1, 1280]]            [1, 10]            12,810     ===============================================================================Total params: 23,556,330Trainable params: 23,509,034Non-trainable params: 47,296-------------------------------------------------------------------------------Input size (MB): 0.01Forward/backward pass size (MB): 2.38Params size (MB): 89.86Estimated Total Size (MB): 92.25-------------------------------------------------------------------------------

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:298: UserWarning: :53The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.  op_type, op_type, EXPRESSION_MAP[method_name]))/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:298: UserWarning: /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/metric/metrics.py:270The behavior of expression A == B has been unified with equal(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use equal(X, Y, axis=0) instead of A == B. This transitional warning will be dropped in the future.  op_type, op_type, EXPRESSION_MAP[method_name]))

数据集配置

本次使用经典的 Cifar10 分类数据集进行演示本数据集已集成在 Paddle 中，可以直接通过 API 进行调用In [5]

# 配置数据预处理# 通道转置 + 归一化transform = Compose([Transpose(), Normalize(mean=127.5, std=127.5)])# 加载数据集train_dataset = Cifar10(mode='train', transform=transform)val_dataset = Cifar10(mode='test',  transform=transform)

Cache file /home/aistudio/.cache/paddle/dataset/cifar/cifar-10-python.tar.gz not found, downloading https://dataset.bj.bcebos.com/cifar/cifar-10-python.tar.gz Begin to downloadDownload finished

模型训练

通过 model.fit 接口进行模型训练还可以通过添加回调函数进行训练数据可视化等操作可视化训练数据样例如下图：

In [6]

# 配置 VisualDL 可视化回调函数# 训练启动后可通过右侧可视化查看训练数据vdl_callback = paddle.callbacks.VisualDL(log_dir='log')# 模型训练# train_data 训练数据# eval_data 测试数据# batch_size 数据批大小# epochs 训练轮次# eval_freq 评测间隔# log_freq log 间隔# save_dir 保存目录# verbose log 方式# drop_last 是否丢弃末尾数据# num_workers 读取线程# callbacks 回调函数model.fit(    train_data=train_dataset,     eval_data=val_dataset,     batch_size=256,     epochs=2,     eval_freq=1,     log_freq=20,     save_dir='save_models',     save_freq=1,     verbose=1,     drop_last=False,     shuffle=True,    num_workers=8,     callbacks=vdl_callback)

The loss value printed in the log is the current step, and the metric is the average value of previous step.Epoch 1/2

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/utils.py:77: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working  return (isinstance(seq, collections.Sequence) and

step 196/196 [==============================] - loss: 1.0396 - acc_top1: 0.6414 - acc_top5: 0.9535 - 51ms/step         save checkpoint at /home/aistudio/save_models/0Eval begin...The loss value printed in the log is the current batch, and the metric is the average value of previous step.step 40/40 [==============================] - loss: 2.1800 - acc_top1: 0.4489 - acc_top5: 0.9243 - 34ms/step        Eval samples: 10000Epoch 2/2step 196/196 [==============================] - loss: 0.5899 - acc_top1: 0.7437 - acc_top5: 0.9820 - 48ms/step        save checkpoint at /home/aistudio/save_models/1Eval begin...The loss value printed in the log is the current batch, and the metric is the average value of previous step.step 40/40 [==============================] - loss: 2.4723 - acc_top1: 0.4512 - acc_top5: 0.9275 - 37ms/step        Eval samples: 10000save checkpoint at /home/aistudio/save_models/final

模型预测

通过 model.predict_batch 可单独对一张图片进行预测In [8]

# 标签列表classes = ['飞机', '汽车', '鸟', '猫', '鹿', '狗', '青蛙', '马', '船', '卡车']# 预测图像路径test_img_path = 'cat.jpg'# 显示预测图像display(Image(test_img_path))# 读取测试图像test_img = cv2.imread(test_img_path)# 数据预处理# 缩放 + 通道转置 + 归一化 + 新增维度test_img = cv2.resize(test_img, (32, 32))test_img = transform(test_img)test_img = test_img[np.newaxis, ...]# 模型预测result = model.predict_batch(test_img)# 结果后处理# 取置信度最大的标签下标 + 标签转换index = np.argmax(result)predict_label = classes[index]# 打印结果print('该图片的预测结果为：%s' % predict_label)

该图片的预测结果为：猫

以上就是高层 API 实现 RepVGG 模型微调的详细内容，更多请关注创想鸟其它相关文章！