基于MAX78000FTHR的扑克牌卷积神经网络图像分类识别系统

内容介绍

项目介绍：

参加比赛，同时，让自己在AI时代掌握一种新的技能。用AI来解决身边的一些有趣的任务。

大赛地址：https://www.eetree.cn/page/analog-max78000-2

前面测试了好几种方法，但对项目结果不太满意，测试了好几种模型，最后还是觉得扑克牌大众化，人人都可以玩，分类也足够多（本次分了54个类别，还增加了“unkown”类型来抗干扰）。

很适合大家分享一起玩，看看AI分类的效果，演示该芯片的潜力等场景也很合适。

扑克牌做图像识别可能出现效果较好或较差的原因有很多，以下是一些主要因素：

1. 扑克牌设计的一致性与区分性：
   - 扑克牌在设计上有一定的标准化，如大小、形状、花色和数字样式较为统一。
   - 每张扑克牌有其独特的符号和数字以区分其它牌，如A、K、Q、J以及2到10的数字。
   - 四种不同的花色（黑桃、红心、梅花、方块）有明显的颜色和图案差异。
   这些因素有助于图像识别算法区分和识别各种牌型。

2. 图像质量：
   - 扑克牌图像的分辨率、清晰度和对比度会影响识别算法的性能。高质量的图片更有助于算法准确识别牌面。
   - 照明条件对图像质量也有重要影响，如阴影、反光或不均匀的光照会干扰图像识别。

3. 视角与变形问题：
   - 如果图像捕捉时扑克牌存在倾斜、扭曲或遮挡，则可能对识别效果产生负面影响。
   - 高级的图像识别算法可能会应用透视变换技术以校正视角问题。

4. 环境背景：
   - 复杂或杂乱的背景可能干扰图像识别算法提取扑克牌的特征。
   - 纯色或对比度强烈的背景有助于算法识别和分割前景中的扑克牌。

5. 训练数据质量：
   - 图像识别通常依赖机器学习算法，这些算法的性能取决于训练数据集的多样性和质量。
   - 如果训练数据中包含了多种不同条件下的扑克牌图像，那么识别效果可能会更好。

6. 图像识别算法的复杂性：
   - 使用高级的机器学习和深度学习模型，如卷积神经网络（CNN），通常能提供较高的识别准确率。
   - 算法的泛化能力和对新条件的适应性也决定了在实际应用中的表现。

理想情况下，如果我们能控制上述因素，图像识别应该能够很好地识别出扑克牌。然而在实践中，可能会受到多种因素的影响，图像识别的实际效果可能会有所不同。通过提高输入图像的质量、优化算法和使用高质量的训练数据等措施，可以显著提升图像识别的性能。

我尽量把学习的过程和操作步骤写清楚，方便后来人复现。

模型的选择：在参考了往届的几个图像分类的项目后，也尝试过复现项目，看一看模型的效果。

但好几个项目复现失败，又回过头来看大牛们的模型的定义文件和yaml文件是怎么定义的。

但在复现过程中，遇到很多问题，处理问题的过程中，也加深了对模型的理解。

LCD的采购与杜邦线连接：

外扩元件方面：购买了TFT-LCD.

FiFnlbCoPWuwBWgTQdJ5jJktZCli

把买来的LCD显示器用杜邦线连接开发板，也能用。只是用的不那么舒服而已。

接线引脚定于如下：

我查阅了屏幕驱动库的源代码，找到了接线的方式。

· CLK：P0_7

· MOSI：P0_5

· MISO：P0_6

· CS：P0_11

· DC：P0_8

当然，LCD下面的引脚也要和开发板进行连接才能正常工作：

· 3V3

· GND

· RST

用猫狗分类代码对LCD进行了测试，LCD工作正常, 如下：

外形折腾: 加上外壳是一个简易相机

在“相机”盒子的右上角装了一个大按钮，充当相机按键。

电池输出的3.3V, 可以供开发板用电。

背面插上 MAX78000FTHR 开发板，摄像头正好暴露在外面。

折腾的时候，再这里(https://oshwhub.com/woshinidiegun/Max-Expansion-Board)发现了一个嘉立创开源的PCB。

感谢作者开源，让更多的爱好者有了新玩具。

顺便打样了一个PCB，PCB到了以后，就拆掉了简易相机。最后，用如下PCB测试。

在满足最小可用原则的基础上，对原理图进行了简化，去掉了锂电池的充放电电路，本人使用的电路连接原理图如下：

实物图如下：

开发环境的搭建

按照官方的资料，在windows10下安装了WSL2, 然后按照:

https://github.com/MaximIntegratedAI/ai8x-training

和

https://github.com/MaximIntegratedAI/ai8x-synthesis

的说明进行了搭建。过程还算是挺顺利的。

主要用到的命令如下：

-------------------------------------- venv虚拟环境用法 ----------------------------------
python.exe -m pip install --upgrade pip

1. 创建虚拟环境：在终端中输入以下命令，创建一个名为 venv 的虚拟环境：
   python -m venv venv

2.激活虚拟环境：
在 Windows 上，可以使用以下命令激活虚拟环境：
venv\Scripts\activate.bat

而在 Linux 或 macOS 上，可以使用以下命令激活虚拟环境：
source venv/bin/activate
激活虚拟环境后，终端的提示符会显示虚拟环境名称，表示当前环境已切换到虚拟环境中。

3. 在虚拟环境中安装第三方库：在激活虚拟环境后，可以使用 pip 命令安装需要的第三方库，例如：
   pip install pytest
这样可以在虚拟环境中安装 其它需要的库。

4.退出虚拟环境：在虚拟环境中完成工作后，可以使用以下命令退出虚拟环境：
   deactivate
------------------------------------------------------------------------------------------
pip install pyenv
pyenv install 3.8.11
pyenv local 3.8.11    # 设置使用那个版本的python
python --version      # 检查版本

$ python -m venv venv --prompt ai8x-training
$ source venv/bin/activate
(ai8x-training) $ pip3 install -U pip wheel setuptools
(ai8x-training) $ pip3 install -r requirements-cu11.txt -i https://mirrors.aliyun.com/pypi/simple/

(ai8x-training) $ git pull
(ai8x-training) $ git submodule update --init
(ai8x-training) $ pip3 install -U pip setuptools
(ai8x-training) $ pip3 install -U -r requirements.txt  # or requirements-xxx.txt, as shown above

(ai8x-training) $ deactivate   # 退出虚拟环境

nvidia-smi         查看 CUDA 版本

ai8x-training$ ./check_cuda.py
System:            linux
Python version:    3.8.11 (default, Nov 26 2023, 00:09:03) [GCC 9.4.0]
PyTorch version:   1.8.1+cu111
CUDA acceleration: available in PyTorch
---------------------------------------- WSL2 ----------------------------------------
https://github.com/MaximIntegratedAI/ai8x-training              按照这里的说明安装
curl -L https://github.com/pyenv/pyenv-installer/raw/master/bin/pyenv-installer | bash

在 Ubuntu 20.04 的 WSL2 上，将以下内容添加到 ~/.bashrc ：
# WSL2
export PYENV_ROOT="$HOME/.pyenv"
export PATH="$PYENV_ROOT/bin:$PATH"
eval "$(pyenv init --path)"
eval "$(pyenv virtualenv-init -)"

windows与虚拟环境的文件共享:

wsl2安装后可以通过网络路径访问wsl2 Linux的文件，路径是：\\wsl$

模型的打印调试思路可用快速定位问题

我在模型定义的文件中，增加了模型的各个层的打印语句，通过打印。

就能知道模型一层和一层之间大概是怎么传递参数的。

例如：我的模型文件如下：

"""
playcards classification network for AI85
"""
import torch
import torch.nn as nn

import ai8x

class AI85PlayingCardsNet(nn.Module):
    """
    Define CNN model for image classification.
    """
    def __init__(self, num_classes=54, num_channels=3, dimensions=(128, 128),
                 fc_inputs=16, bias=False, **kwargs):
        super().__init__()

        # AI85 Limits
        assert dimensions[0] == dimensions[1]  # Only square supported

        # Keep track of image dimensions so one constructor works for all image sizes
        dim = dimensions[0]

        self.conv1 = ai8x.FusedConv2dReLU(num_channels, 16, 3,
                                          padding=1, bias=bias, **kwargs)
        # padding 1 -> no change in dimensions -> 16x128x128

        pad = 2 if dim == 28 else 1
        self.conv2 = ai8x.FusedMaxPoolConv2dReLU(16, 32, 3, pool_size=2, pool_stride=2,
                                                 padding=pad, bias=bias, **kwargs)
        dim //= 2     # pooling, padding 0 -> 32x64x64
        if pad == 2:
            dim += 2  # padding 2 -> 32x32x32

        self.conv3 = ai8x.FusedMaxPoolConv2dReLU(32, 64, 3, pool_size=2, pool_stride=2, padding=1,
                                                 bias=bias, **kwargs)
        dim //= 2  # pooling, padding 0 -> 64x32x32

        self.conv4 = ai8x.FusedMaxPoolConv2dReLU(64, 32, 3, pool_size=2, pool_stride=2, padding=1,
                                                 bias=bias, **kwargs)
        dim //= 2  # pooling, padding 0 -> 32x16x16

        self.conv5 = ai8x.FusedMaxPoolConv2dReLU(32, 32, 3, pool_size=2, pool_stride=2, padding=1,
                                                 bias=bias, **kwargs)
        dim //= 2  # pooling, padding 0 -> 32x8x8

        self.conv6 = ai8x.FusedConv2dReLU(32, fc_inputs, 3, padding=1, bias=bias, **kwargs)

        self.fc = ai8x.Linear(fc_inputs*dim*dim, num_classes, bias=True, wide=True, **kwargs)

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')

    def forward(self, x):  # pylint: disable=arguments-differ
        """Forward prop"""
        x = self.conv1(x)
        print('After conv1 shape:', x.shape)
        x = self.conv2(x)
        print('After conv2 shape:', x.shape)
        x = self.conv3(x)
        print('After conv3 shape:', x.shape)
        x = self.conv4(x)
        print('After conv4 shape:', x.shape)
        x = self.conv5(x)
        print('After conv5 shape:', x.shape)
        x = self.conv6(x)
        print('After conv6 shape:', x.shape)
        x = x.view(x.size(0), -1)
        print('After flatten shape:', x.shape)
        x = self.fc(x)
        print('After fc shape:', x.shape)

        return x
        
def ai85net_playingcards(pretrained=False, **kwargs):
    assert not pretrained
    return AI85PlayingCardsNet(**kwargs)

models = [
    {
        'name': 'ai85net_playingcards',
        'min_input': 1,
        'dim': 2,
    },
]

if __name__ == '__main__':
    ai8x.set_device(device=85, simulate=True, round_avg=False)

    model = AI85PlayingCardsNet()
    input_tensor = torch.randn(54, 3, 128, 128)
    output_tensor = model(input_tensor)
    print('Output shape:', output_tensor.shape)

我通过运行模型文件，让模型文件打印各层大小的参数。就能做到对模型心中有数。

python models/ai85net-playingcards.py

Configuring device: MAX78000, simulate=True.
After conv1 shape: torch.Size([54, 16, 128, 128])
After conv2 shape: torch.Size([54, 32, 64, 64])
After conv3 shape: torch.Size([54, 64, 32, 32])
After conv4 shape: torch.Size([54, 32, 16, 16])
After conv5 shape: torch.Size([54, 32, 8, 8])
After conv6 shape: torch.Size([54, 16, 8, 8])
After flatten shape: torch.Size([54, 1024])
After fc shape: torch.Size([54, 54])
Output shape: torch.Size([54, 54])

好几次，发现了别人的模型，用上面的方法一打印，发现那个模型他用是可用的，我用是有问题的。

例如，水果分类的那个模型，原作者在64x64分辨率是可用的，如下：

(ai8x-training) ml@DESKTOP-UHGOPTO:~/ai8x-training$ python models/ai85net-fruit.py
Configuring device: MAX78000, simulate=True.
Input shape: torch.Size([54, 3, 64, 64])
After conv1 shape: torch.Size([54, 16, 32, 32])
After conv2 shape: torch.Size([54, 16, 32, 32])
After conv3 shape: torch.Size([54, 20, 32, 32])
After resid1 shape: torch.Size([54, 20, 32, 32])
After conv4 shape: torch.Size([54, 20, 32, 32])
After conv5 shape: torch.Size([54, 20, 32, 32])
After conv6 shape: torch.Size([54, 20, 16, 16])
After resid2 shape: torch.Size([54, 20, 16, 16])
After conv7 shape: torch.Size([54, 44, 16, 16])
After conv8 shape: torch.Size([54, 44, 16, 16])
After conv9 shape: torch.Size([54, 48, 8, 8])
After resid3 shape: torch.Size([54, 48, 8, 8])
After conv10 shape: torch.Size([54, 32, 2, 2])
After flatten shape: torch.Size([54, 128])
After fc shape: torch.Size([54, 54])
Output shape: torch.Size([54, 54])

我做扑克牌识别，考虑到芯片的内存大小，芯片能支持的最大分辨率128x128，修改后运行就报错了。报错如下：

有了报错，修改不可用模型为可用，也有了对应的修改思路；

测试模型的可用性也多了一种快速、有效的手段：

(ai8x-training) ml@DESKTOP-UHGOPTO:~/ai8x-training$ python models/ai85net-fruit.py
Configuring device: MAX78000, simulate=True.
Input shape: torch.Size([54, 3, 128, 128])
After conv1 shape: torch.Size([54, 16, 64, 64])
After conv2 shape: torch.Size([54, 16, 64, 64])
After conv3 shape: torch.Size([54, 20, 64, 64])
After resid1 shape: torch.Size([54, 20, 64, 64])
After conv4 shape: torch.Size([54, 20, 64, 64])
After conv5 shape: torch.Size([54, 20, 64, 64])
After conv6 shape: torch.Size([54, 20, 32, 32])
After resid2 shape: torch.Size([54, 20, 32, 32])
After conv7 shape: torch.Size([54, 44, 32, 32])
After conv8 shape: torch.Size([54, 44, 32, 32])
After conv9 shape: torch.Size([54, 48, 16, 16])
After resid3 shape: torch.Size([54, 48, 16, 16])
After conv10 shape: torch.Size([54, 32, 6, 6])
After flatten shape: torch.Size([54, 1152])
Traceback (most recent call last):
  File "models/ai85net-fruit.py", line 90, in <module>
    output_tensor = model(input_tensor)
  File "/home/ml/ai8x-training/venv/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "models/ai85net-fruit.py", line 64, in forward
    x = self.fc(x)
  File "/home/ml/ai8x-training/venv/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/ml/ai8x-training/models/ai8x.py", line 450, in forward
    x = self.op(x)
  File "/home/ml/ai8x-training/venv/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/ml/ai8x-training/venv/lib/python3.8/site-packages/torch/nn/modules/linear.py", line 94, in forward
    return F.linear(input, self.weight, self.bias)
  File "/home/ml/ai8x-training/venv/lib/python3.8/site-packages/torch/nn/functional.py", line 1753, in linear
    return torch._C._nn.linear(input, weight, bias)
RuntimeError: mat1 and mat2 shapes cannot be multiplied (54x1152 and 128x54)

报错了以后，再回头看芯片的参数，发现还是具有参考意义的。

上面的报错原因，个人理解为：受芯片资源限制通道数，不能超过1024。

那我们就要考虑设计更加精简的网络结构，既能满足芯片要求，又能满足我们的训练要求。

后来又尝试参考了“AI宝可梦图鉴”，他的分辨率也是128x128，符合做扑克牌识别的要求。但后来发现有坑的。

这个过程中，积累了很多有意思的经验。

最后，参考了厂家的经典的猫狗分类模型的案例。终于把扑克牌识别训练到了满意的效果。

卷积神经网络中各层的设计如下：

训练和移植步骤

经过这段时间的学习，我觉得把一种基于卷积神经网络的图像分类技术用到嵌入式产品里，需要经过下面5个步骤：

（0）收集可用数据集。

（1）训练。

（2）量化。

（3）生成样本测试文件(sample_playingcards.npy)。

（4）量化完成后可以对量化后结果进行评估，评估量化后的模型后续是否可用。

（5）把量化后的模型转化为嵌入式C语言的CNN代码。

训练和量化用到的几个程序之间的关系如下：

（0）收集合适的扑克牌的数据集。

我找到的主要参考数据集如下:

https://github.com/lordloh/playing-cards

https://www.kaggle.com/datasets/gpiosenka/cards-image-datasetclassification

我对两个扑克牌数据集合并的过程中，我自己用手机拍照的方法新增了3000多张扑克牌图片。

手机拍照有一个选项就是正方形，拍照时尽可能的拍扑克牌的全貌。

每一种花色和点数都在原来的基础上，合并了精心裁剪的带有不同背景的扑克牌照片。增加抗干扰能力。

（1）准备好数据集、模型，进行训练

命令如下:

python train.py --epochs 500  --optimizer Adam --lr 0.001 --wd 0.001 --batch-size 200 --gpus 0 --deterministic --compress policies/schedule-playingcards.yaml --model ai85net_playingcards --dataset puke --param-hist --pr-curves --embedding --device MAX78000

中间又几次wsl的linux闪退了，好在可以断点继续。

这样就可以在上次炼丹的基础上，进行炼丹了。可以大大节约时间。区别在于增加一个参数“--resume-from ../ai8x-synthesis/trained/playingcard_qat_best.pth.tar”。

继续训练的命令如下：

python train.py --epochs 800  --optimizer Adam --lr 0.001 --wd 0.001 --batch-size 210 --gpus 0 --deterministic --compress policies/schedule-playingcards.yaml --model ai85net_playingcards --dataset puke  --resume-from  ../ai8x-synthesis/trained/playingcard_qat_best.pth.tar --param-hist --pr-curves --embedding --device MAX78000

又崩了。

python train.py --epochs 900  --optimizer Adam --lr 0.001 --wd 0.001 --batch-size 200 --gpus 0 --deterministic --compress policies/schedule-playingcards.yaml --model ai85net_playingcards --dataset puke  --resume-from  logs/2023.12.31-141013/qat_best.pth.tar --param-hist --pr-curves --embedding --device MAX78000

贵在坚持，最后训练了1100个epochs结束，我用的GTX3050，总的训练时间大概是两天两夜。

python train.py --epochs 1100  --optimizer Adam --lr 0.001 --wd 0.001 --batch-size 200 --gpus 0 --deterministic --compress policies/schedule-playingcards.yaml --model ai85net_playingcards --dataset puke  --resume-from  logs/2024.01.04-210257/qat_best.pth.tar --param-hist --pr-curves --embedding --device MAX78000

新的发现和感想：--batch-size 200 这个参数很有用，我的的GTX3050是4G内存，默认参数一运行就崩了。

我发现参数是210或者200的时候，显存刚刚好。这一发现，也给我省了不少时间。不然，去云上再搭一套环境肯定不如本地训练方便快捷。

因为有54个分类，实际分担到每个扑克花色的训练图片，其实并不多，也就120-180个左右的图片。数据集一共有7063个图片。

最终完成了1100个epoch的训练后的结果如下：

Training epoch: 10928 samples (200 per mini-batch)
Epoch: [1099][   10/   55]    Overall Loss 0.459331    Objective Loss 0.459331                                        LR 0.000250    Time 2.016463
Epoch: [1099][   20/   55]    Overall Loss 0.445955    Objective Loss 0.445955                                        LR 0.000250    Time 1.775931
Epoch: [1099][   30/   55]    Overall Loss 0.439101    Objective Loss 0.439101                                        LR 0.000250    Time 1.886492
Epoch: [1099][   40/   55]    Overall Loss 0.437805    Objective Loss 0.437805                                        LR 0.000250    Time 1.784566
Epoch: [1099][   50/   55]    Overall Loss 0.440521    Objective Loss 0.440521                                        LR 0.000250    Time 1.832200
Epoch: [1099][   55/   55]    Overall Loss 0.445095    Objective Loss 0.445095    Top1 86.280488    Top5 97.865854    LR 0.000250    Time 1.756030
--- validate (epoch=1099)-----------
1214 samples (200 per mini-batch)
Epoch: [1099][    7/    7]    Loss 0.868613    Top1 74.299835    Top5 91.598023
==> Top1: 74.300    Top5: 91.598    Loss: 0.869

==> Best [Top1: 77.842   Top5: 92.339   Sparsity:0.00   Params: 111024 on epoch: 1084]
Saving checkpoint to: logs/2024.01.05-232502/qat_checkpoint.pth.tar
--- test ---------------------
549 samples (200 per mini-batch)
Test: [    3/    3]    Loss 0.518195    Top1 84.335155    Top5 95.264117
==> Top1: 84.335    Top5: 95.264    Loss: 0.518

拷贝 ai8x-training/logs/2024.01.05-232502/qat_best.pth.tar 为 ai8x-synthesis/trained/playingcards5_qat_best_q8.pth.tar

可用看出，一次性识别成功率为: 84.335%, Top5: 95.264% 这个结果已经很不错了，够我们一般的玩耍和某些场合的演示了。

（2）模型的量化.

(ai8x-synthesis) ml@DESKTOP-UHGOPTO:~/ai8x-synthesis$ python quantize.py trained/playingcards5_qat_best.pth.tar trained/playingcards5_qat_best_q8.pth.tar --device MAX78000 -v -c
 networks/playingcards.yaml "$@"
Configuring device: MAX78000
Reading networks/playingcards.yaml to configure network...
Converting checkpoint file trained/playingcards5_qat_best.pth.tar to trained/playingcards5_qat_best_q8.pth.tar

Model keys (state_dict):
conv1.output_shift, conv1.weight_bits, conv1.bias_bits, conv1.quantize_activation, conv1.adjust_output_shift, conv1.shift_quantile, conv1.op.weight, conv2.output_shift, conv2.weight_bits, conv2.bias_bits, conv2.quantize_activation, conv2.adjust_output_shift, conv2.shift_quantile, conv2.op.weight, conv3.output_shift, conv3.weight_bits, conv3.bias_bits, conv3.quantize_activation, conv3.adjust_output_shift, conv3.shift_quantile, conv3.op.weight, conv4.output_shift, conv4.weight_bits, conv4.bias_bits, conv4.quantize_activation, conv4.adjust_output_shift, conv4.shift_quantile, conv4.op.weight, conv5.output_shift, conv5.weight_bits, conv5.bias_bits, conv5.quantize_activation, conv5.adjust_output_shift, conv5.shift_quantile, conv5.op.weight, conv6.output_shift, conv6.weight_bits, conv6.bias_bits, conv6.quantize_activation, conv6.adjust_output_shift, conv6.shift_quantile, conv6.op.weight, fc.output_shift, fc.weight_bits, fc.bias_bits, fc.quantize_activation, fc.adjust_output_shift, fc.shift_quantile, fc.op.weight, fc.op.bias
conv1.op.weight avg_max: 0.3080426 max: 0.50178313 mean: -0.0016146562 factor: [128.] bits: 8
conv2.op.weight avg_max: 0.21638438 max: 0.37782183 mean: -0.004252241 factor: [256.] bits: 8
conv3.op.weight avg_max: 0.10264358 max: 0.18610927 mean: -0.00275823 factor: [512.] bits: 8
conv4.op.weight avg_max: 0.18799132 max: 0.52683735 mean: -0.0013671168 factor: [128.] bits: 8
conv5.op.weight avg_max: 0.2335556 max: 0.3548111 mean: 0.0019469503 factor: [256.] bits: 8
conv6.op.weight avg_max: 0.31261 max: 0.40786323 mean: 0.004781828 factor: [256.] bits: 8
fc.op.weight avg_max: 0.11929286 max: 0.19588341 mean: -0.00024378698 factor: [512.] bits: 8
fc.op.bias avg_max: 0.00069900794 max: 0.10763672 mean: 0.00069900794 factor: [512.] bits: 8

（3）生成样本测试文件(sample_playingcards.npy)

使用的命令如下：

(ai8x-training) ml@DESKTOP-UHGOPTO:~/ai8x-training$ python train.py --batch-size 200 --optimizer Adam --lr 0.0001 --model ai85net_playingcards --save-sample 100 --dataset puke --evaluate --exp-load-weights-from ../ai8x-synthesis/trained/playingcards2_qat_best_q8.pth.tar -8 --device MAX78000 --use-bias "$@"
警告:打开文件限制为1024.请提高限额(见文档).
Configuring device: MAX78000, simulate=True.
Log file for this run: /home/ml/ai8x-training/logs/2023.12.31-212714/2023.12.31-212714.log
{'start_epoch': 10, 'weight_bits': 8}
Model `../ai8x-synthesis/trained/playingcards2_qat_best_q8.pth.tar` is old. Missing parameters added with default values!
=> loading checkpoint ../ai8x-synthesis/trained/playingcards2_qat_best_q8.pth.tar
=> Checkpoint contents:
+----------------------+-------------+----------------------+
| Key                  | Type        | Value                |
|----------------------+-------------+----------------------|
| arch                 | str         | ai85net_playingcards |
| compression_sched    | dict        |                      |
| epoch                | int         | 401                  |
| extras               | dict        |                      |
| optimizer_state_dict | dict        |                      |
| optimizer_type       | type        | Adam                 |
| state_dict           | OrderedDict |                      |
+----------------------+-------------+----------------------+

=> Checkpoint['extras'] contents:
+-----------------+--------+-------------------+
| Key             | Type   | Value             |
|-----------------+--------+-------------------|
| best_epoch      | int    | 401               |
| best_mAP        | int    | 0                 |
| best_top1       | float  | 82.90816326530613 |
| clipping_method | str    | MAX_BIT_SHIFT     |
| current_mAP     | int    | 0                 |
| current_top1    | float  | 82.90816326530613 |
+-----------------+--------+-------------------+

Loaded compression schedule from checkpoint (epoch 401)
=> loaded 'state_dict' from checkpoint '../ai8x-synthesis/trained/playingcards2_qat_best_q8.pth.tar'
Optimizer Type: <class 'torch.optim.adam.Adam'>
Optimizer Args: {'lr': 0.0001, 'betas': (0.9, 0.999), 'eps': 1e-08, 'weight_decay': 0.0001, 'amsgrad': False}
Dataset sizes:
        test=318
--- test ---------------------
318 samples (200 per mini-batch)
==> Saving sample at index 100 to sample_puke.npy
Test: [    2/    2]    Loss 0.532525    Top1 89.937107    Top5 97.169811
==> Top1: 89.937    Top5: 97.170    Loss: 0.533

运行后，把本目录下的文件 sample_playingcards.npy 复制到ai8x-synthesis\tests下，在生成嵌入式CNN代码的时候会用到。

（4）量化完成后可以对量化后结果进行评估

评估量化后的模型后续是否可用。

脚本路径: ai8x-training\scripts\evaluate_playingcards.sh

(ai8x-training) ml@DESKTOP-UHGOPTO:~/ai8x-training$ python train.py --model ai85net_playingcards --dataset puke --confusion --evaluate --exp-load-weights-from ../ai8x-synthesis/trained/playingcards5_qat_best_q8.pth.tar -8 --device MAX78000 --use-bias "$@"
警告:打开文件限制为1024.请提高限额(见文档).
Configuring device: MAX78000, simulate=True.
Log file for this run: /home/ml/ai8x-training/logs/2024.01.06-073135/2024.01.06-073135.log
{'start_epoch': 10, 'weight_bits': 8}
Model `../ai8x-synthesis/trained/playingcards5_qat_best_q8.pth.tar` is old. Missing parameters added with default values!
=> loading checkpoint ../ai8x-synthesis/trained/playingcards5_qat_best_q8.pth.tar
=> Checkpoint contents:
+----------------------+-------------+----------------------+
| Key                  | Type        | Value                |
|----------------------+-------------+----------------------|
| arch                 | str         | ai85net_playingcards |
| compression_sched    | dict        |                      |
| epoch                | int         | 1084                 |
| extras               | dict        |                      |
| optimizer_state_dict | dict        |                      |
| optimizer_type       | type        | Adam                 |
| state_dict           | OrderedDict |                      |
+----------------------+-------------+----------------------+

=> Checkpoint['extras'] contents:
+-----------------+--------+-------------------+
| Key             | Type   | Value             |
|-----------------+--------+-------------------|
| best_epoch      | int    | 1084              |
| best_mAP        | int    | 0                 |
| best_top1       | float  | 77.84184514003294 |
| clipping_method | str    | MAX_BIT_SHIFT     |
| current_mAP     | int    | 0                 |
| current_top1    | float  | 77.84184514003294 |
+-----------------+--------+-------------------+

Loaded compression schedule from checkpoint (epoch 1084)
=> loaded 'state_dict' from checkpoint '../ai8x-synthesis/trained/playingcards5_qat_best_q8.pth.tar'
Optimizer Type: <class 'torch.optim.sgd.SGD'>
Optimizer Args: {'lr': 0.1, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0.0001, 'nesterov': False}
Dataset sizes:
        test=549
--- test ---------------------
549 samples (256 per mini-batch)
Test: [    3/    3]    Loss 0.772329    Top1 85.245902    Top5 95.992714
==> Top1: 85.246    Top5: 95.993    Loss: 0.772

==> Confusion:
[[ 8  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0]
 [ 0 10  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0 10  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0 10  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  8  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  2  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  9  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  9  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  7  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  1  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  6  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  1  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0]
 [ 0  1  0  0  0  0  0  0  0  8  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  9  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  1  0  0  0  0  0  0  0  8  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0 10  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  1  0  0  0  0  0  0  0  0  0  0  0  9  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  1  0  0  0  8  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  8  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  8  1  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  9  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0 10  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  6  0  1  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0 12  0  0  0  0  1  0  0  0  3  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  9  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  8  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  9  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  9  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  8  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  2  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  8  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  1  7  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0 10  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  8  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  1]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0 10  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0 10  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  1  0  0  1  0  0  0  0  0  0  0  7  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  9  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0 10  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  9  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0 10  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  1  0  0  0  0  0  0  0  8  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  9  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  8  0  0  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  9  0  0  0  0  0  0  0  0  0  0  0  1  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0 10  0  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  1  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  7  0  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0 10  0  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  9  0  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  7  0  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  2  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  7  0  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  8  0  0  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  9  1  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0 10  0  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0 10  0  0  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  8  0  0]
 [ 0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  1  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  2  0  0  6  0]
 [ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0 10]]

发现个别在验证集上好像有问题，进去看了一下，验证集(test)的图片设置的不合理，部分图片看不清，进行了合理的替换。

可以看到：经过1100个epoch的训练，在第1084个epoch的训练中，发现一次性识别成功率最高。为: 85.246%, Top5: 95.993% 这个准确率已经很高，够我们愉快的玩耍和一般场合的演示了。

（5）把量化后的模型转化为嵌入式C语言的CNN代码

把 ai8x-training/puke.npy 文件拷贝为 ai8x-synthesis/tests/sample_playingcard5.npy

然后，执行下面的命令进行转化。

ml@DESKTOP-UHGOPTO:~$ cd ai8x-synthesis/
ml@DESKTOP-UHGOPTO:~/ai8x-synthesis$ source venv/bin/activate
(ai8x-synthesis) ml@DESKTOP-UHGOPTO:~/ai8x-synthesis$ python ai8xize.py --test-dir "sdk/Examples/$DEVICE/CNN" --prefix playingcard --checkpoint-file trained/playingcards2_qat_best_q8.pth.tar --config-file networks/playingcards.yaml --fifo --softmax --device MAX78000 --timer 0 --display-checkpoint --verbose --compact-data --mexpress --sample-input 'tests/sample_playingcards2.npy'  --boost 2.5 "$@"

Configuring device: MAX78000
Reading networks/playingcards.yaml to configure network...
WARNING: Cannot run "yamllint" linter to check networks/playingcards.yaml
Reading trained/playingcards2_qat_best_q8.pth.tar to configure network weights...
Checkpoint for epoch 401, model ai85net_playingcards - weight and bias data:
 InCh OutCh  Weights         Quant Shift  Min  Max    Size Key                                       Bias       Quant  Min  Max Size Key
    3    16  (48, 3, 3)          8    -1  -89  112     432 conv1.op.weight                           (16,)          8   -1    0   16 conv1.op.bias
   16    32  (512, 3, 3)         8    -1  -78   94    4608 conv2.op.weight                           (32,)          8   -1    0   32 conv2.op.bias
   32    64  (2048, 3, 3)        8    -2  -81   96   18432 conv3.op.weight                           (64,)          8   -1    0   64 conv3.op.bias
   64    32  (2048, 3, 3)        8    -1  -59  102   18432 conv4.op.weight                           (32,)          8   -1    0   32 conv4.op.bias
   32    32  (1024, 3, 3)        8    -1  -69   85    9216 conv5.op.weight                           (32,)          8   -1    0   32 conv5.op.bias
   32    16  (512, 3, 3)         8    -1 -103   91    4608 conv6.op.weight                           (16,)          8   -1    0   16 conv6.op.bias
 1024    54  (1, 54, 1024)       8    -2  -90  109   55296 fc.op.weight                              (54,)          8  -29   37   54 fc.op.bias
TOTAL: 7 parameter layers, 111,270 parameters, 111,270 bytes
Configuring data set: puke.
playingcard...
Arranging weights... ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100%
Storing weights...   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100%
WARNING: Layer 0: Combining streaming and a bias. THIS COMBINATION MIGHT NOT FUNCTION CORRECTLY!!!
WARNING: Layer 1: Combining streaming and a bias. THIS COMBINATION MIGHT NOT FUNCTION CORRECTLY!!!
Creating network...  ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100%

(6) 移植CNN代码到嵌入式Demo项目中

增加中文打印的支持：

这是给家里的小朋友演示的时候，小朋友说他有点看不懂英文。于是，就增加了中文的支持。调用过程请参考附件的项目代码。

TFT LCD如下文件"MaximSDK\Libraries\MiscDrivers\Display\tft_ili9341.c" 增加了中文的打印部分，主要源码如下：

// 下面函数:新加的一个字的中文显示函数
void tft_Print_CN_char(unsigned int x,unsigned int y, char *str)
{
	unsigned int column;
	unsigned char tm=0,temp=0,xxx=0;

	write_command(0x2a);   //Column address set
	write_data(x>>8);      //start column
	write_data(x);
	x=x+15;
	write_data(x>>8);      //end column
	write_data(x);

	write_command(0x2b);   //Row address set
	write_data(y>>8);      //start row
	write_data(y); 
	y=y+15;
	write_data(y>>8);      //end row
	write_data(y);
	write_command(0x2C);   //Memory write
    
    for(column=0;column<32;column++)  //column loop
    {
        temp=str[xxx];
        for(tm=0;tm<8;tm++)
        {
			if(temp&0x01)
			{
			  write_data(g_foreground_color>>8);
			  write_data(g_foreground_color);
			}
			else 
			{
			  write_data(g_background_color>>8);
			  write_data(g_background_color);
			}
            temp>>=1;
        }
        xxx++;  
    }
}

经过1100轮训练，识别率如下: Top1: 85.246 Top5: 95.993 Loss: 0.714

实际使用也证明识别率尚可接受。目前对常见的扑克牌识别率比较高。

但因为没有专门对残缺的扑克进行训练。所以，对半个扑克拍照暂时无法识别，因为没有对半个扑克进行更多的训练。

整体扑克的识别率较高，侧面拍照，扑克有浅一点的形变，也可以正确识别扑克牌的花色和点数。

项目的简要工作思路的流程如下：

测试结果和显示效果

识别方块8的结果和显示界面如下：

我们测试识别一张梅花Q，如下:

不足:

本次训练图片中，歪斜的图片比较少。歪一定的角度后，识别就没有原来那么精确了。

实际商品化的产品训练，要增加很多抗干扰措施。

对着空的东西，也能分类到54种扑克牌类别中，但是置信度比较低（小于50%）。

抗干扰措施的一种就是：发现置信度低于50%的情况，设置为“未识别”。

本项目基于演示和学习的场景，没有加上这一规则。

参加本活动过程中，学习到了很多有用的经验。