mindspore两日集训营202209-金箍棒应用

pip install https://ms-release.obs.cn-north-4.myhuaweicloud.com/1.8.0/GoldenStick/any/mindspore_gs-0.1.0-py3-none-any.whl --trusted-host ms-release.obs.cn-north-4.myhuaweicloud.com -i https://pypi.tuna.tsinghua.edu.cn/simple

1
2

在这里插入图片描述

cd ./golden_stick/scripts/
bash run_standalone_train_gpu.sh ../quantization/simqat/ ../quantization/simqat/lenet_mnist_config.yaml ../../data/
1
2

出现了这样一句报错
run_standalone_train_gpu.sh: line 76: ulimit: max user processes: cannot modify limit: Operation not permitted
正确的做法是，在root下执行命令，
ulimit -u unlimited
然后把sh文件里同样的这句话注释掉
然后就跑起来了，很快的
在这里插入图片描述

获得的ckpt比一般的要小

打印网络结构非常简单，直接print(network)

LeNet5Opt<
  (_handler): LeNet5<
    (conv1): Conv2d<input_channels=1, output_channels=6, kernel_size=(5, 5), stride=(1, 1), pad_mode=valid, padding=0, dilation=(1, 1), group=1, has_bias=False, weight_init=normal, bias_init=zeros, format=NCHW>
    (conv2): Conv2d<input_channels=6, output_channels=16, kernel_size=(5, 5), stride=(1, 1), pad_mode=valid, padding=0, dilation=(1, 1), group=1, has_bias=False, weight_init=normal, bias_init=zeros, format=NCHW>
    (relu): ReLU<>
    (max_pool2d): MaxPool2d<kernel_size=2, stride=2, pad_mode=VALID>
    (flatten): Flatten<>
    (fc1): Dense<input_channels=400, output_channels=120, has_bias=True>
    (fc2): Dense<input_channels=120, output_channels=84, has_bias=True>
    (fc3): Dense<input_channels=84, output_channels=10, has_bias=True>
    >
  (max_pool2d): MaxPool2d<kernel_size=2, stride=2, pad_mode=VALID>
  (flatten): Flatten<>
  (fc1): Dense<input_channels=400, output_channels=120, has_bias=True>
  (fc2): Dense<input_channels=120, output_channels=84, has_bias=True>
  (fc3): Dense<input_channels=84, output_channels=10, has_bias=True>
  (DenseQuant): QuantizeWrapperCell<
    (_handler): DenseQuant<
      in_channels=84, out_channels=10, weight=Parameter (name=DenseQuant._handler.weight, shape=(10, 84), dtype=Float32, requires_grad=True), has_bias=True, bias=Parameter (name=DenseQuant._handler.bias, shape=(10,), dtype=Float32, requires_grad=True)
      (fake_quant_weight): SimulatedFakeQuantizerPerChannel<bit_num=8, symmetric=True, narrow_range=False, ema=False(0.999), per_channel=True(0, 10), quant_delay=900>
      >
    (_input_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    (_output_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    >
  (Conv2dBnFoldQuant): QuantizeWrapperCell<
    (_handler): Conv2dBnFoldQuant<
      in_channels=1, out_channels=6, kernel_size=(5, 5), stride=(1, 1), pad_mode=valid, padding=0, dilation=(1, 1), group=1, fake=True, freeze_bn=10000000, momentum=0.997
      (fake_quant_weight): SimulatedFakeQuantizerPerChannel<bit_num=8, symmetric=True, narrow_range=False, ema=False(0.999), per_channel=True(0, 6), quant_delay=900>
      (batchnorm_fold): BatchNormFoldCell<>
      >
    (_input_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    (_output_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    >
  (Conv2dBnFoldQuant_1): QuantizeWrapperCell<
    (_handler): Conv2dBnFoldQuant<
      in_channels=6, out_channels=16, kernel_size=(5, 5), stride=(1, 1), pad_mode=valid, padding=0, dilation=(1, 1), group=1, fake=True, freeze_bn=10000000, momentum=0.997
      (fake_quant_weight): SimulatedFakeQuantizerPerChannel<bit_num=8, symmetric=True, narrow_range=False, ema=False(0.999), per_channel=True(0, 16), quant_delay=900>
      (batchnorm_fold): BatchNormFoldCell<>
      >
    (_input_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    (_output_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    >
  (DenseQuant_1): QuantizeWrapperCell<
    (_handler): DenseQuant<
      in_channels=400, out_channels=120, weight=Parameter (name=DenseQuant_1._handler.weight, shape=(120, 400), dtype=Float32, requires_grad=True), has_bias=True, bias=Parameter (name=DenseQuant_1._handler.bias, shape=(120,), dtype=Float32, requires_grad=True), activation=ReLU<>
      (activation): ReLU<>
      (fake_quant_weight): SimulatedFakeQuantizerPerChannel<bit_num=8, symmetric=True, narrow_range=False, ema=False(0.999), per_channel=True(0, 120), quant_delay=900>
      >
    (_input_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    (_output_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    >
  (DenseQuant_2): QuantizeWrapperCell<
    (_handler): DenseQuant<
      in_channels=120, out_channels=84, weight=Parameter (name=DenseQuant_2._handler.weight, shape=(84, 120), dtype=Float32, requires_grad=True), has_bias=True, bias=Parameter (name=DenseQuant_2._handler.bias, shape=(84,), dtype=Float32, requires_grad=True), activation=ReLU<>
      (activation): ReLU<>
      (fake_quant_weight): SimulatedFakeQuantizerPerChannel<bit_num=8, symmetric=True, narrow_range=False, ema=False(0.999), per_channel=True(0, 84), quant_delay=900>
      >
    (_input_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    (_output_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    >
  >
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61

这样就可以了。
然后评估

bash run_eval_gpu.sh ../quantization/simqat/ ../quantization/simqat/lenet_mnist_config.yaml ../../data/ ./train/ckpt/checkpoint_lenet-10_2187.ckpt
1

在这里插入图片描述
精度非常高的。
第一个作业完成了。
然后是第二个作业。
alexnet
先把全精度跑通

bash run_standalone_train_gpu.sh cifar10 ~/haahsummer/pro57msgs/models/official/cv/alexnet/data/cifar-10-batches-bin ckpt

1
2

30轮有点太久，跑了4~5个小时。一个模型427.9M
在这里插入图片描述

然后是测试

bash run_standalone_eval_gpu.sh cifar10 ~/haahsummer/pro57msgs/models/official/cv/alexnet/data/cifar-10-batches-bin ckpt/checkpoint_alexnet-30_1875.ckpt 0

1
2

在这里插入图片描述
网络结构就打印出来了。
最后acc只有0.1,不知道原因是什么，在学校里有个课题我们也尝试用了女alexnet，效果也不是很好，那我们恐怕需要微调参数了，而且仔细看训练的loss，波动后也没有明显的下降。

要使用金箍棒
在这里插入图片描述
写一个algorithm.py
train和eval分别导入
eval中添加

现在看来像alexnet这种比较大的网络，还是不要在本地比较好。

在这里插入图片描述

上面是全精度训练的结果。fp32训练后精度为0.88
在这里插入图片描述

我们添加量化训练的eval和train逻辑，把对应部分量化。
在这里插入图片描述
fp32训练结果如图，15s一个epoch
loss有序下降，并且训练速度比本地快得多。
量化训练用时明显比全精度训练慢了很多，一百多s一个epoch

量化测试采用刚刚生成的ckpt，并在测试脚本里添加打印量化后的网络的逻辑。
在这里插入图片描述

AlexNetOpt<
  (_handler): AlexNet<
    (conv1): Conv2d<input_channels=3, output_channels=64, kernel_size=(11, 11), stride=(4, 4), pad_mode=same, padding=0, dilation=(1, 1), group=1, has_bias=True, weight_init=normal, bias_init=zeros, format=NCHW>
    (conv2): Conv2d<input_channels=64, output_channels=128, kernel_size=(5, 5), stride=(1, 1), pad_mode=same, padding=0, dilation=(1, 1), group=1, has_bias=True, weight_init=normal, bias_init=zeros, format=NCHW>
    (conv3): Conv2d<input_channels=128, output_channels=192, kernel_size=(3, 3), stride=(1, 1), pad_mode=same, padding=0, dilation=(1, 1), group=1, has_bias=True, weight_init=normal, bias_init=zeros, format=NCHW>
    (conv4): Conv2d<input_channels=192, output_channels=256, kernel_size=(3, 3), stride=(1, 1), pad_mode=same, padding=0, dilation=(1, 1), group=1, has_bias=True, weight_init=normal, bias_init=zeros, format=NCHW>
    (conv5): Conv2d<input_channels=256, output_channels=256, kernel_size=(3, 3), stride=(1, 1), pad_mode=same, padding=0, dilation=(1, 1), group=1, has_bias=True, weight_init=normal, bias_init=zeros, format=NCHW>
    (max_pool2d): MaxPool2d<kernel_size=3, stride=2, pad_mode=VALID>
    (flatten): Flatten<>
    (fc1): Dense<input_channels=9216, output_channels=4096, has_bias=True>
    (fc2): Dense<input_channels=4096, output_channels=4096, has_bias=True>
    (fc3): Dense<input_channels=4096, output_channels=10, has_bias=True>
    (dropout): Dropout<keep_prob=1.0>
    >
  (max_pool2d): MaxPool2d<kernel_size=3, stride=2, pad_mode=VALID>
  (flatten): Flatten<>
  (fc1): Dense<input_channels=9216, output_channels=4096, has_bias=True>
  (fc2): Dense<input_channels=4096, output_channels=4096, has_bias=True>
  (fc3): Dense<input_channels=4096, output_channels=10, has_bias=True>
  (dropout): Dropout<keep_prob=1.0>
  (Conv2dQuant): QuantizeWrapperCell<
    (_handler): Conv2dQuant<
      in_channels=3, out_channels=64, kernel_size=(11, 11), stride=(4, 4), pad_mode=same, padding=0, dilation=(1, 1), group=1, has_bias=True
      (fake_quant_weight): SimulatedFakeQuantizerPerChannel<bit_num=8, symmetric=True, narrow_range=False, ema=False(0.999), per_channel=True(0, 64), quant_delay=900>
      >
    (_input_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    (_output_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    >
  (Conv2dQuant_1): QuantizeWrapperCell<
    (_handler): Conv2dQuant<
      in_channels=64, out_channels=128, kernel_size=(5, 5), stride=(1, 1), pad_mode=same, padding=0, dilation=(1, 1), group=1, has_bias=True
      (fake_quant_weight): SimulatedFakeQuantizerPerChannel<bit_num=8, symmetric=True, narrow_range=False, ema=False(0.999), per_channel=True(0, 128), quant_delay=900>
      >
    (_input_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    (_output_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    >
  (Conv2dQuant_2): QuantizeWrapperCell<
    (_handler): Conv2dQuant<
      in_channels=128, out_channels=192, kernel_size=(3, 3), stride=(1, 1), pad_mode=same, padding=0, dilation=(1, 1), group=1, has_bias=True
      (fake_quant_weight): SimulatedFakeQuantizerPerChannel<bit_num=8, symmetric=True, narrow_range=False, ema=False(0.999), per_channel=True(0, 192), quant_delay=900>
      >
    (_input_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    (_output_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    >
  (Conv2dQuant_3): QuantizeWrapperCell<
    (_handler): Conv2dQuant<
      in_channels=192, out_channels=256, kernel_size=(3, 3), stride=(1, 1), pad_mode=same, padding=0, dilation=(1, 1), group=1, has_bias=True
      (fake_quant_weight): SimulatedFakeQuantizerPerChannel<bit_num=8, symmetric=True, narrow_range=False, ema=False(0.999), per_channel=True(0, 256), quant_delay=900>
      >
    (_input_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    (_output_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    >
  (Conv2dQuant_4): QuantizeWrapperCell<
    (_handler): Conv2dQuant<
      in_channels=256, out_channels=256, kernel_size=(3, 3), stride=(1, 1), pad_mode=same, padding=0, dilation=(1, 1), group=1, has_bias=True
      (fake_quant_weight): SimulatedFakeQuantizerPerChannel<bit_num=8, symmetric=True, narrow_range=False, ema=False(0.999), per_channel=True(0, 256), quant_delay=900>
      >
    (_input_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    (_output_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    >
  (DenseQuant): QuantizeWrapperCell<
    (_handler): DenseQuant<
      in_channels=9216, out_channels=4096, weight=Parameter (name=DenseQuant._handler.weight, shape=(4096, 9216), dtype=Float32, requires_grad=True), has_bias=True, bias=Parameter (name=DenseQuant._handler.bias, shape=(4096,), dtype=Float32, requires_grad=True)
      (fake_quant_weight): SimulatedFakeQuantizerPerChannel<bit_num=8, symmetric=True, narrow_range=False, ema=False(0.999), per_channel=True(0, 4096), quant_delay=900>
      >
    (_input_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    (_output_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    >
  (DenseQuant_1): QuantizeWrapperCell<
    (_handler): DenseQuant<
      in_channels=4096, out_channels=4096, weight=Parameter (name=DenseQuant_1._handler.weight, shape=(4096, 4096), dtype=Float32, requires_grad=True), has_bias=True, bias=Parameter (name=DenseQuant_1._handler.bias, shape=(4096,), dtype=Float32, requires_grad=True)
      (fake_quant_weight): SimulatedFakeQuantizerPerChannel<bit_num=8, symmetric=True, narrow_range=False, ema=False(0.999), per_channel=True(0, 4096), quant_delay=900>
      >
    (_input_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    (_output_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    >
  (DenseQuant_2): QuantizeWrapperCell<
    (_handler): DenseQuant<
      in_channels=4096, out_channels=10, weight=Parameter (name=DenseQuant_2._handler.weight, shape=(10, 4096), dtype=Float32, requires_grad=True), has_bias=True, bias=Parameter (name=DenseQuant_2._handler.bias, shape=(10,), dtype=Float32, requires_grad=True)
      (fake_quant_weight): SimulatedFakeQuantizerPerChannel<bit_num=8, symmetric=True, narrow_range=False, ema=False(0.999), per_channel=True(0, 10), quant_delay=900>
      >
    (_input_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    (_output_quantizer): SimulatedFakeQuantizerPerLayer<bit_num=8, symmetric=False, narrow_range=False, ema=False(0.999), per_channel=False, quant_delay=900>
    >
  >
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85

打印网络结构
训练后的精度为0.87几乎没有变化
在这里插入图片描述
ok顺利完成。

相关阅读:
帧结构的串行数据接收器——Verilog实现
 Tomcat介绍和配置使用
 【MySQL系列】-B+树索引和HASH索引有什么区别
 网络协议与攻击模拟_17HTTPS 协议
 单点登录SSO
【Python零基础入门篇 · 34】：进程间的通信-Queue、进程池的构建
 如何从 OpenVINO 的主分支构建 Python Wheel 包
 centos 7.9离线安装wget
FPGA之手把手教你做多路信号发生器（STM32与FPGA数据互传控制波形生成）
三十七、【进阶】验证索引的效率
原文地址：https://blog.csdn.net/weixin_54227557/article/details/126800367