[toc]

『深度学习 7 日打卡营·快速入门特辑』

零基础解锁深度学习神器飞桨框架高层 API，七天时间助你掌握 CV、NLP 领域最火模型及应用。

  1. 课程地址

传送门：https://aistudio.baidu.com/aistudio/course/introduce/6771

  1. 目标
  • 掌握深度学习常用模型基础知识
  • 熟练掌握一种国产开源深度学习框架
  • 具备独立完成相关深度学习任务的能力
  • 能用所学为 AI 加一份年味

DL 万能公式

python
1
2
3
4
5
import paddle
import numpy as np
import matplotlib.pyplot as plt

paddle.__version__
plaintext
1
'2.0.0'

数据加载和预处理

python
1
2
3
4
5
6
7
8
9
10
11
12
13
import paddle.vision.transforms as T

# 数据加载和预处理
# [0-255] -> [0-1]
transform = T.Normalize(mean=[127.5], std=[127.5])

# 训练数据集
train_dataset = paddle.vision.datasets.MNIST(mode='train', transform=transform)

# 评估数据集
eval_dataset = paddle.vision.datasets.MNIST(mode='test', transform=transform)

print(f"训练集样本量:{len(train_dataset)}，验证集样本量:{len(eval_dataset)}")
plaintext
1
训练集样本量:60000，验证集样本量:10000

查看数据

python
1
2
3
4
5
6
7
%matplotlib inline

plt.figure()
plt.imshow(train_dataset[0][0].reshape([28, 28]), cmap=plt.cm.binary)
plt.show()
print("label:", train_dataset[0][1])
print("data shape:", train_dataset[0][0].shape)

在这里插入图片描述

plaintext
1
2
label: [5]
data shape: (1, 28, 28)

搭建 LeNet-5 卷积神经网络

选用 LeNet-5 网络结构。

LeNet-5 模型源于论文“LeCun Y, Bottou L, Bengio Y, et al. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278-2324.”，

论文地址：https://ieeexplore.ieee.org/document/726791

每个阶段用到的 Layer

python
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
# 网络搭建
net = paddle.nn.Sequential(
                            ('C1', paddle.nn.Conv2D(in_channels=1, out_channels=6, kernel_size=3, padding=1, stride=1)),
                            # 6x28x28
                            ('ReLU1', paddle.nn.ReLU()),
                            ('S2', paddle.nn.MaxPool2D(kernel_size=2, stride=2, ceil_mode=True)),
                            # 6x14x14
                            ('C3', paddle.nn.Conv2D(6, 16, kernel_size=5, stride=1, padding=0)),
                            # 16x10x10
                            ('ReLU2', paddle.nn.ReLU()),
                            ('S4', paddle.nn.MaxPool2D(kernel_size=2, stride=2, ceil_mode=True)),
                            # 16x5x5
                            ('C5', paddle.nn.Conv2D(16, 120, kernel_size=5, stride=1, padding=0)),
                            # 120x1x1
                            ('ReLU3', paddle.nn.ReLU()),
                            ('ReLU4', paddle.nn.Flatten()),
                            # 120
                            ('F6', paddle.nn.Linear(120, 84)),
                            # 84
                            ('ReLU5', paddle.nn.ReLU()),
                            ('OUTPUT', paddle.nn.Linear(84, 10))
                          )

网络模型可视化

python
1
2
3
4
5
# 模型封装
model = paddle.Model(net)

# 模型可视化
model.summary((8, 1, 28, 28))  # n c h w
plaintext
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
---------------------------------------------------------------------------
 Layer (type)       Input Shape          Output Shape         Param #
===========================================================================
   Conv2D-1       [[8, 1, 28, 28]]      [8, 6, 28, 28]          60
    ReLU-1        [[8, 6, 28, 28]]      [8, 6, 28, 28]           0
  MaxPool2D-1     [[8, 6, 28, 28]]      [8, 6, 14, 14]           0
   Conv2D-2       [[8, 6, 14, 14]]     [8, 16, 10, 10]         2,416
    ReLU-2       [[8, 16, 10, 10]]     [8, 16, 10, 10]           0
  MaxPool2D-2    [[8, 16, 10, 10]]      [8, 16, 5, 5]            0
   Conv2D-3       [[8, 16, 5, 5]]       [8, 120, 1, 1]        48,120
    ReLU-3        [[8, 120, 1, 1]]      [8, 120, 1, 1]           0
   Flatten-1      [[8, 120, 1, 1]]         [8, 120]              0
   Linear-1          [[8, 120]]            [8, 84]            10,164
    ReLU-4           [[8, 84]]             [8, 84]               0
   Linear-2          [[8, 84]]             [8, 10]              850
===========================================================================
Total params: 61,610
Trainable params: 61,610
Non-trainable params: 0
---------------------------------------------------------------------------
Input size (MB): 0.02
Forward/backward pass size (MB): 0.90
Params size (MB): 0.24
Estimated Total Size (MB): 1.16
---------------------------------------------------------------------------






{'total_params': 61610, 'trainable_params': 61610}

模型配置

  • 优化器：SGD
  • 损失函数：交叉熵（cross entropy）
  • 评估指标：Accuracy
python
1
2
3
4
5
6
7
8
9
10
11
12
# 配置优化器，损失函数，评估指标
model.prepare(optimizer=paddle.optimizer.Adam(learning_rate=0.001, parameters=net.parameters()),
              loss=paddle.nn.CrossEntropyLoss(),
              metrics=paddle.metric.Accuracy())

# 启动模型全流程训练
model.fit(train_data=train_dataset,
          eval_data=eval_dataset,
          batch_size=64,
          epochs=5,
          verbose=1,
          shuffle=True)
plaintext
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
The loss value printed in the log is the current step, and the metric is the average value of previous step.
Epoch 1/5


D:\Anaconda3\envs\paddle2\lib\site-packages\paddle\fluid\layers\utils.py:77: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated since Python 3.3,and in 3.9 it will stop working
  return (isinstance(seq, collections.Sequence) and


step 938/938 [==============================] - loss: 0.0460 - acc: 0.9391 - 14ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 157/157 [==============================] - loss: 0.0032 - acc: 0.9759 - 11ms/step
Eval samples: 10000
Epoch 2/5
step 938/938 [==============================] - loss: 0.0375 - acc: 0.9801 - 14ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 157/157 [==============================] - loss: 0.0014 - acc: 0.9863 - 8ms/step
Eval samples: 10000
Epoch 3/5
step 938/938 [==============================] - loss: 0.0199 - acc: 0.9850 - 13ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 157/157 [==============================] - loss: 0.0128 - acc: 0.9847 - 15ms/step
Eval samples: 10000
Epoch 4/5
step 938/938 [==============================] - loss: 0.0043 - acc: 0.9884 - 21ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 157/157 [==============================] - loss: 0.0019 - acc: 0.9836 - 8ms/step
Eval samples: 10000
Epoch 5/5
step 938/938 [==============================] - loss: 0.0069 - acc: 0.9914 - 14ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 157/157 [==============================] - loss: 2.2102e-04 - acc: 0.9884 - 14ms/step
Eval samples: 10000

模型评估

python
1
2
3
result = model.evaluate(eval_dataset, verbose=1)

print(result)
plaintext
1
2
3
4
5
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10000/10000 [==============================] - loss: 2.0623e-05 - acc: 0.9884 - 4ms/step
Eval samples: 10000
{'loss': [2.0622994e-05], 'acc': 0.9884}

模型预测

批量预测

使用model.predit接口完成对大量数据集的批量预测

python
1
2
3
4
5
6
7
8
9
10
11
12
13
14
result = model.predict(eval_dataset)

# 定义画图方法
def show_img(img, predict):
    plt.figure()
    plt.title('predict:{}'.format(predict))
    plt.imshow(img.reshape([28, 28]), cmap=plt.cm.binary)
    plt.show()

# 抽样展示
indexs = [2, 15, 38, 211]

for idx in indexs:
    show_img(eval_dataset[idx][0], np.argmax(result[0][idx]))
plaintext
1
2
3
Predict begin...
step 10000/10000 [==============================] - 4ms/step
Predict samples: 10000

在这里插入图片描述

在这里插入图片描述

在这里插入图片描述

在这里插入图片描述

单张图片预测

采用model.predict_batch来进行单张或者少量多张图片的预测。

python
1
2
3
4
5
6
7
8
# 读取单张图片
img = eval_dataset[233][0]

result = model.predict_batch([img[np.newaxis, ...]])  # 需要多添加一个batch轴，不然报错

print(result)

show_img(img, np.argmax(result))
plaintext
1
2
3
[array([[-3.4706905, -6.674865 , -1.9018929,  3.8094432, -5.66697  ,
         1.5752668, -6.6928353, -2.2028043,  9.449063 ,  3.296681 ]],
      dtype=float32)]

在这里插入图片描述

部署上线

保存模型

python
1
model.save('finetuning/mnist', training=True)

继续调优训练

python
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
from paddle.static import InputSpec

model_2 = paddle.Model(net, inputs=[InputSpec(shape=[-1, 1, 28, 28], dtype='float32', name='image')])

model_2.load('./finetuning/mnist')

# 配置优化器，损失函数，评估指标
model_2.prepare(optimizer=paddle.optimizer.Adam(learning_rate=0.0001, parameters=net.parameters()),
              loss=paddle.nn.CrossEntropyLoss(),
              metrics=paddle.metric.Accuracy())

# 启动模型全流程训练
model_2.fit(train_data=train_dataset,
            eval_data=eval_dataset,
            batch_size=64,
            epochs=1,
            verbose=1)
plaintext
1
2
3
4
5
6
7
The loss value printed in the log is the current step, and the metric is the average value of previous step.
Epoch 1/1
step 938/938 [==============================] - loss: 3.6613e-04 - acc: 0.9966 - 13ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 157/157 [==============================] - loss: 7.5415e-05 - acc: 0.9908 - 9ms/step
Eval samples: 10000

保存预测模型

python
1
model_2.save('./infer/mnist', training=False)