图像分类（image classification）是计算机视觉领域中最简单最基础的任务，学习研究图像分类是每个计算机视觉研究者的必经之路，图像分类网络也是很多更复杂任务（如目标检测、语义分割等）算法的基础。本练习赛旨在让选手们用图像分类任务来以赛代练、熟悉深度学习框架和比赛流程。
☞☞☞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☜☜☜
赛题介绍
赛题背景
图像分类（image classification）是计算机视觉领域中最简单最基础的任务，学习研究图像分类是每个计算机视觉研究者的必经之路，图像分类网络也是很多更复杂任务（如目标检测、语义分割等）算法的基础。本练习赛旨在让选手们用图像分类任务来以赛代练、熟悉深度学习框架和比赛流程。在图像分类学习中，MNIST数据集常被用来作为入门教学数据集。但是，MNIST数据集存在一些问题：首先，MNIST数据集对于现在的卷积神经网络来说过于简单，SOTA模型的分类精度达到了99.84%，甚至传统机器学习方法也能达到97%的精度，因此模型的精度在此达到了饱和，几乎没有提升的空间；再者，有些专家对MNIST数据集提出了质疑，比如谷歌的深度学习专家、Keras的作者François Chollet曾表示：“MNIST存在很多问题，但最重要的是，它真的不具有计算机视觉任务的代表性。”并补充道：“很多好点子（比如batch norm）在MNIST上效果差，但相反的，一些差的方法可能在MNIST产生好效果，却不能迁移到真实计算机视觉任务中。”
数据说明
本练习赛采用和MNIST同等规模但更有难度的数据集Fashion-MNIST（github链接：https://github.com/zalandoresearch/fashion-mnist），Fashion-MNIST由60000张训练集图像、10000张测试集图像及对应的标签构成，每张图像是分辨率为28x28的灰度图像，包含10种分类：T恤、裤子、套头衫、连衣裙、大衣、凉鞋、衬衫、运动鞋、包、短靴。
本练习赛的参赛者可以使用Tensorflow、Keras、Pytorch、Paddlepaddle等开源深度学习框架来进行模型的搭建、训练和预测。
In [1]
import pandas as pdimport numpy as np# 数据读取train_df = pd.read_csv('fashion-mnist_train.csv')
test_df = pd.read_csv('fashion-mnist_test_data.csv')
In [2]
train_df.head()
label pixel1 pixel2 pixel3 pixel4 pixel5 pixel6 pixel7 pixel8 \ 0 2 0 0 0 0 0 0 0 0 1 9 0 0 0 0 0 0 0 0 2 6 0 0 0 0 0 0 0 5 3 0 0 0 0 1 2 0 0 0 4 3 0 0 0 0 0 0 0 0 pixel9 ... pixel775 pixel776 pixel777 pixel778 pixel779 pixel780 \ 0 0 ... 0 0 0 0 0 0 1 0 ... 0 0 0 0 0 0 2 0 ... 0 0 0 30 43 0 3 0 ... 3 0 0 0 0 1 4 0 ... 0 0 0 0 0 0 pixel781 pixel782 pixel783 pixel784 0 0 0 0 0 1 0 0 0 0 2 0 0 0 0 3 0 0 0 0 4 0 0 0 0 [5 rows x 785 columns]
数据可视化
In [3]
%pylab inline# 绘制10*10的图片plt.figure(figsize=(10, 10))for idx in range(100): # 图片尺寸从784 -》 28 * 28
xy = train_df.iloc[idx].values[1:].reshape(28,28)
plt.subplot(10, 10, idx+1) # 展示图片
plt.imshow(xy, cmap='gray')
plt.xticks([]); plt.yticks([])
Populating the interactive namespace from numpy and matplotlib
<Figure size 720x720 with 100 Axes>
In [4]
import paddle paddle.__version__
'2.2.0'
数据读取
In [5]
from paddle.io import DataLoader, Datasetfrom PIL import Image# 自定义数据集完成数据读取class MyDataset(Dataset):
def __init__(self, img, label):
super(MyDataset, self).__init__()
self.img = img
self.label = label
def __getitem__(self, index):
img = self.img[index] # 对数据集进行归一化
return img/255, int(self.label[index]) def __len__(self):
return len(self.label)# 使用5.9w进行训练train_dataset = MyDataset(
train_df.iloc[:-1000, 1:].values.reshape(59000, 28, 28).astype(np.float32),
paddle.to_tensor(train_df.label.iloc[:-1000].values.astype(np.float32))
)
train_loader = DataLoader(train_dataset, batch_size=300, shuffle=True)# 最后1k作为验证集val_dataset = MyDataset(
train_df.iloc[-1000:, 1:].values.reshape(1000, 28, 28).astype(np.float32),
paddle.to_tensor(train_df.label.iloc[-1000:].values.astype(np.float32))
)
val_loader = DataLoader(val_dataset, batch_size=300, shuffle=False)# 最终测试集test_dataset = MyDataset(
test_df.iloc[:, 1:].values.reshape(10000, 28, 28).astype(np.float32),
paddle.to_tensor(np.zeros((test_df.shape[0])))
)
test_loader = DataLoader(test_dataset, batch_size=64, shuffle=False)
全连接模型
In [6]
# 构建全连接模型，先需要从多维转变为二维model = paddle.nn.Sequential(
paddle.nn.Flatten(),
paddle.nn.Linear(28*28,128),
paddle.nn.LeakyReLU(),
paddle.nn.Linear(128, 10)
)
paddle.summary(model, (64, 28, 28))
W1207 20:53:04.353837 104 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1 W1207 20:53:04.359086 104 device_context.cc:465] device: 0, cuDNN Version: 7.6.
--------------------------------------------------------------------------- Layer (type) Input Shape Output Shape Param # =========================================================================== Flatten-1 [[64, 28, 28]] [64, 784] 0 Linear-1 [[64, 784]] [64, 128] 100,480 LeakyReLU-1 [[64, 128]] [64, 128] 0 Linear-2 [[64, 128]] [64, 10] 1,290 =========================================================================== Total params: 101,770 Trainable params: 101,770 Non-trainable params: 0 --------------------------------------------------------------------------- Input size (MB): 0.19 Forward/backward pass size (MB): 0.51 Params size (MB): 0.39 Estimated Total Size (MB): 1.09 ---------------------------------------------------------------------------
{'total_params': 101770, 'trainable_params': 101770}
In [7]
# 定义优化器，损失函数optimizer = paddle.optimizer.Adam(parameters=model.parameters(), learning_rate=0.0001)
criterion = paddle.nn.CrossEntropyLoss()# 每个epoch迭代训练# 每个batch -> 正向传播 -> 计算损失 -> 更新参数for epoch in range(0, 5):
Train_Loss, Val_Loss = [], []
Train_ACC, Val_ACC = [], [] # 训练部分
model.train() for i, (x, y) in enumerate(train_loader):
pred = model(x)
loss = criterion(pred, y)
Train_Loss.append(loss.item())
loss.backward()
optimizer.step()
optimizer.clear_grad()
Train_ACC.append((pred.numpy().argmax(1) == y.numpy()).mean())
if i % 100 == 0: print(f'{i}/{len(train_loader)}\t Loss {np.mean(Train_Loss):3.5f} {np.mean(Train_ACC):3.5f}') # 验证部分
model.eval() for i, (x, y) in enumerate(val_loader):
pred = model(x)
loss = criterion(pred, y)
Val_Loss.append(loss.item())
Val_ACC.append((pred.numpy().argmax(1) == y.numpy()).mean())
if epoch % 1 == 0: print(f'\nEpoch: {epoch}') print(f'Loss {np.mean(Train_Loss):3.5f}/{np.mean(Val_Loss):3.5f}') print(f'ACC {np.mean(Train_ACC):3.5f}/{np.mean(Val_ACC):3.5f}')
0/197 Loss 2.33518 0.17000 100/197 Loss 1.50995 0.55713 Epoch: 0 Loss 1.20884/0.83340 ACC 0.63190/0.71083 0/197 Loss 0.77190 0.74667 100/197 Loss 0.73694 0.76122 Epoch: 1 Loss 0.69631/0.65332 ACC 0.77584/0.77250 0/197 Loss 0.61251 0.80333 100/197 Loss 0.60704 0.80515 Epoch: 2 Loss 0.59082/0.57583 ACC 0.80878/0.80167 0/197 Loss 0.52229 0.84333 100/197 Loss 0.55234 0.81921 Epoch: 3 Loss 0.53963/0.53572 ACC 0.82233/0.80417 0/197 Loss 0.54005 0.81667 100/197 Loss 0.51667 0.82782 Epoch: 4 Loss 0.50867/0.50455 ACC 0.83046/0.82083
卷积模型
In [11]
# 卷积模型model = paddle.nn.Sequential(
paddle.nn.Conv2D(1, 10, (5, 5)),
paddle.nn.ReLU(),
paddle.nn.Conv2D(10, 20, (5, 5)),
paddle.nn.ReLU(),
paddle.nn.MaxPool2D((2, 2)),
paddle.nn.Flatten(),
paddle.nn.Linear(2000, 10),
)
paddle.summary(model, (64, 1, 28, 28))
---------------------------------------------------------------------------
Layer (type) Input Shape Output Shape Param #
===========================================================================
Conv2D-5 [[64, 1, 28, 28]] [64, 10, 24, 24] 260
ReLU-5 [[64, 10, 24, 24]] [64, 10, 24, 24] 0
Conv2D-6 [[64, 10, 24, 24]] [64, 20, 20, 20] 5,020
ReLU-6 [[64, 20, 20, 20]] [64, 20, 20, 20] 0
MaxPool2D-4 [[64, 20, 20, 20]] [64, 20, 10, 10] 0
Flatten-4 [[64, 20, 10, 10]] [64, 2000] 0
Linear-5 [[64, 2000]] [64, 10] 20,010
===========================================================================
Total params: 25,290
Trainable params: 25,290
Non-trainable params: 0
---------------------------------------------------------------------------
Input size (MB): 0.19
Forward/backward pass size (MB): 15.40
Params size (MB): 0.10
Estimated Total Size (MB): 15.68
---------------------------------------------------------------------------
{'total_params': 25290, 'trainable_params': 25290}
In [12]
# 优化器与损失函数optimizer = paddle.optimizer.Adam(parameters=model.parameters(), learning_rate=0.0001)
criterion = paddle.nn.CrossEntropyLoss()# 训练部分for epoch in range(0, 5):
Train_Loss, Val_Loss = [], []
Train_ACC, Val_ACC = [], []
model.train() for i, (x, y) in enumerate(train_loader):
pred = model(x.reshape((-1, 1, 28, 28)))
loss = criterion(pred, y)
Train_Loss.append(loss.item())
loss.backward()
optimizer.step()
optimizer.clear_grad()
Train_ACC.append((pred.numpy().argmax(1) == y.numpy()).mean())
if i % 100 == 0: print(f'{i}/{len(train_loader)}\t Loss {np.mean(Train_Loss):3.5f} {np.mean(Train_ACC):3.5f}')
model.eval() for i, (x, y) in enumerate(val_loader):
pred = model(x.reshape((-1, 1, 28, 28)))
loss = criterion(pred, y)
Val_Loss.append(loss.item())
Val_ACC.append((pred.numpy().argmax(1) == y.numpy()).mean())
if epoch % 1 == 0: print(f'\nEpoch: {epoch}') print(f'Loss {np.mean(Train_Loss):3.5f}/{np.mean(Val_Loss):3.5f}') print(f'ACC {np.mean(Train_ACC):3.5f}/{np.mean(Val_ACC):3.5f}')
0/197 Loss 2.29863 0.15333 100/197 Loss 1.61717 0.53125 Epoch: 0 Loss 1.23357/0.74845 ACC 0.62601/0.74167 0/197 Loss 0.73336 0.76000 100/197 Loss 0.66099 0.76261 Epoch: 1 Loss 0.63249/0.61879 ACC 0.77187/0.77333 0/197 Loss 0.56483 0.80000 100/197 Loss 0.55956 0.79855 Epoch: 2 Loss 0.54548/0.56240 ACC 0.80436/0.79167 0/197 Loss 0.58850 0.77333 100/197 Loss 0.51179 0.81703 Epoch: 3 Loss 0.50036/0.52460 ACC 0.82221/0.80583 0/197 Loss 0.41658 0.86667 100/197 Loss 0.47553 0.83274 Epoch: 4 Loss 0.47199/0.50074 ACC 0.83300/0.82333
