实战Kaggle比赛：狗的品种识别（ImageNet Dogs）
=============================================
本节我们将在Kaggle上实战狗品种识别问题。
本次比赛网址是https://www.kaggle.com/c/dog-breed-identification。
:numref:`fig_kaggle_dog`\ 显示了鉴定比赛网页上的信息。
需要一个Kaggle账户才能提交结果。
在这场比赛中，我们将识别120类不同品种的狗。
这个数据集实际上是著名的ImageNet的数据集子集。与
:numref:`sec_kaggle_cifar10`\ 中CIFAR-10数据集中的图像不同，
ImageNet数据集中的图像更高更宽，且尺寸不一。
.. _fig_kaggle_dog:
.. figure:: ../img/kaggle-dog.jpg
:width: 400px
狗的品种鉴定比赛网站，可以通过单击“数据”选项卡来获得比赛数据集。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
import os
import mindcv
import mindspore
import mindspore.dataset.transforms as transforms
import mindspore.dataset.vision as vision
import mindspore.nn as nn
import pandas as pd
from d2l import mindspore as d2l
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
import os
import torch
import torchvision
from torch import nn
from d2l import torch as d2l
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
transform_train = transforms.Compose([
# 随机裁剪图像，所得图像为原始面积的0.08～1之间，高宽比在3/4和4/3之间。
# 然后，缩放图像以创建224x224的新图像
vision.RandomResizedCrop(224, scale=(0.08, 1.0), ratio=(3.0/4.0, 4.0/3.0)),
vision.RandomHorizontalFlip(),
# 随机更改亮度，对比度和饱和度
vision.RandomColorAdjust(brightness=0.4,
contrast=0.4,
saturation=0.4),
# 标准化图像的每个通道
vision.Normalize(mean=[0.485 * 255, 0.456 * 255, 0.406 * 255],
std=[0.229 * 255, 0.224 * 255, 0.225 * 255]),
vision.HWC2CHW()])
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
transform_train = torchvision.transforms.Compose([
# 随机裁剪图像，所得图像为原始面积的0.08～1之间，高宽比在3/4和4/3之间。
# 然后，缩放图像以创建224x224的新图像
torchvision.transforms.RandomResizedCrop(224, scale=(0.08, 1.0),
ratio=(3.0/4.0, 4.0/3.0)),
torchvision.transforms.RandomHorizontalFlip(),
# 随机更改亮度，对比度和饱和度
torchvision.transforms.ColorJitter(brightness=0.4,
contrast=0.4,
saturation=0.4),
# 添加随机噪声
torchvision.transforms.ToTensor(),
# 标准化图像的每个通道
torchvision.transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])])
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
transform_test = transforms.Compose([
vision.Resize(256),
# 从图像中心裁切224x224大小的图片
vision.CenterCrop(224),
vision.Normalize([0.485 * 255, 0.456 * 255, 0.406 * 255],
[0.229 * 255, 0.224 * 255, 0.225 * 255]),
vision.HWC2CHW()])
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
transform_test = torchvision.transforms.Compose([
torchvision.transforms.Resize(256),
# 从图像中心裁切224x224大小的图片
torchvision.transforms.CenterCrop(224),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])])
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
train_ds, train_valid_ds = [mindspore.dataset.ImageFolderDataset(
os.path.join(data_dir, 'train_valid_test', folder), shuffle=True, decode=True)
for folder in ['train', 'train_valid']]
train_ds = train_ds.map(transform_train, 'image')
train_valid_ds = train_valid_ds.map(transform_train, 'image')
valid_ds, test_ds = [mindspore.dataset.ImageFolderDataset(
os.path.join(data_dir, 'train_valid_test', folder), shuffle=False, decode=True)
for folder in ['valid', 'test']]
valid_ds = valid_ds.map(transform_test, 'image')
test_ds = test_ds.map(transform_test, 'image')
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
train_ds, train_valid_ds = [torchvision.datasets.ImageFolder(
os.path.join(data_dir, 'train_valid_test', folder),
transform=transform_train) for folder in ['train', 'train_valid']]
valid_ds, test_ds = [torchvision.datasets.ImageFolder(
os.path.join(data_dir, 'train_valid_test', folder),
transform=transform_test) for folder in ['valid', 'test']]
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
train_iter, train_valid_iter = [dataset.batch(batch_size=batch_size, drop_remainder=True)
for dataset in (train_ds, train_valid_ds)]
valid_iter = valid_ds.batch(batch_size=batch_size, drop_remainder=True)
test_iter = test_ds.batch(batch_size=batch_size, drop_remainder=False)
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
train_iter, train_valid_iter = [torch.utils.data.DataLoader(
dataset, batch_size, shuffle=True, drop_last=True)
for dataset in (train_ds, train_valid_ds)]
valid_iter = torch.utils.data.DataLoader(valid_ds, batch_size, shuffle=False,
drop_last=True)
test_iter = torch.utils.data.DataLoader(test_ds, batch_size, shuffle=False,
drop_last=False)
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
import mindspore.common.initializer as initializer
def get_net(devices):
finetune_net = nn.SequentialCell()
finetune_net.feature = mindcv.create_model('resnet34', pretrained=True)
#finetune_net.append(feature)
# 定义一个新的输出网络，共有120个输出类别
output_new = nn.SequentialCell([nn.Dense(1000, 256),
nn.ReLU(),
nn.Dense(256, 120)])
for name, cell in output_new.cells_and_names():
if isinstance(cell, nn.Dense):
k = 1 / cell.in_channels
k = k ** 0.5
cell.weight.set_data(
initializer.initializer(initializer.Uniform(k), cell.weight.shape, cell.weight.dtype))
if cell.bias is not None:
cell.bias.set_data(
initializer.initializer(initializer.Uniform(k), cell.bias.shape, cell.bias.dtype))
finetune_net.append(output_new)
#finetune_net.append(output_new)
# 冻结参数
for param in finetune_net.feature.get_parameters():
param.requires_grad = False
return finetune_net
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
def get_net(devices):
finetune_net = nn.Sequential()
finetune_net.features = torchvision.models.resnet34(pretrained=True)
# 定义一个新的输出网络，共有120个输出类别
finetune_net.output_new = nn.Sequential(nn.Linear(1000, 256),
nn.ReLU(),
nn.Linear(256, 120))
# 将模型参数分配给用于计算的CPU或GPU
finetune_net = finetune_net.to(devices[0])
# 冻结参数
for param in finetune_net.features.parameters():
param.requires_grad = False
return finetune_net
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
loss = nn.CrossEntropyLoss(reduction='none')
def evaluate_loss(data_iter, net):
l_sum, n = 0.0, 0
for features, labels in data_iter:
outputs = net(features)
l = loss(outputs, labels)
l_sum += l.sum()
n += labels.numel()
return (l_sum / n)
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
loss = nn.CrossEntropyLoss(reduction='none')
def evaluate_loss(data_iter, net, devices):
l_sum, n = 0.0, 0
for features, labels in data_iter:
features, labels = features.to(devices[0]), labels.to(devices[0])
outputs = net(features)
l = loss(outputs, labels)
l_sum += l.sum()
n += labels.numel()
return (l_sum / n).to('cpu')
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
def train(net, train_iter, valid_iter, num_epochs, lr, wd, lr_period, lr_decay):
devices = None
lr_list = mindspore.Tensor([lr*(lr_decay**(i//lr_period))
for i in range(num_epochs)
for j in range(train_iter.get_dataset_size())])
trainer = nn.SGD((param for param in net.get_parameters() if param.requires_grad),
learning_rate=lr_list, momentum=0.9, weight_decay=wd)
def forward_fn(inputs, targets):
logits = net(inputs)
# print(logits.shape, targets.shape)
l = loss(logits, targets)
return l, logits
grad_fn = mindspore.value_and_grad(forward_fn, None, trainer.parameters, has_aux=True)
def train_step(inputs, targets):
(l, logits), grads = grad_fn(inputs, targets)
trainer(grads)
return l.sum(), logits
num_batches, timer = train_iter.get_dataset_size(), d2l.Timer()
legend = ['train loss']
if valid_iter is not None:
legend.append('valid loss')
animator = d2l.Animator(xlabel='epoch', xlim=[1, num_epochs],
legend=legend)
for epoch in range(num_epochs):
net.set_train()
metric = d2l.Accumulator(2)
for i, (features, labels) in enumerate(train_iter):
timer.start()
l, logits = train_step(features, labels)
metric.add(l, labels.shape[0])
timer.stop()
if (i + 1) % (num_batches // 5) == 0 or i == num_batches - 1:
animator.add(epoch + (i + 1) / num_batches,
(metric[0] / metric[1], None))
measures = f'train loss {metric[0] / metric[1]:.3f}'
if valid_iter is not None:
valid_loss = evaluate_loss(valid_iter, net)
animator.add(epoch + 1, (None, valid_loss))
if valid_iter is not None:
measures += f', valid loss {float(valid_loss):.3f}'
print(measures + f'\n{metric[1] * num_epochs / timer.sum():.1f}'
f' examples/sec on {str(devices)}')
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
def train(net, train_iter, valid_iter, num_epochs, lr, wd, devices, lr_period,
lr_decay):
# 只训练小型自定义输出网络
net = nn.DataParallel(net, device_ids=devices).to(devices[0])
trainer = torch.optim.SGD((param for param in net.parameters()
if param.requires_grad), lr=lr,
momentum=0.9, weight_decay=wd)
scheduler = torch.optim.lr_scheduler.StepLR(trainer, lr_period, lr_decay)
num_batches, timer = len(train_iter), d2l.Timer()
legend = ['train loss']
if valid_iter is not None:
legend.append('valid loss')
animator = d2l.Animator(xlabel='epoch', xlim=[1, num_epochs],
legend=legend)
for epoch in range(num_epochs):
metric = d2l.Accumulator(2)
for i, (features, labels) in enumerate(train_iter):
timer.start()
features, labels = features.to(devices[0]), labels.to(devices[0])
trainer.zero_grad()
output = net(features)
l = loss(output, labels).sum()
l.backward()
trainer.step()
metric.add(l, labels.shape[0])
timer.stop()
if (i + 1) % (num_batches // 5) == 0 or i == num_batches - 1:
animator.add(epoch + (i + 1) / num_batches,
(metric[0] / metric[1], None))
measures = f'train loss {metric[0] / metric[1]:.3f}'
if valid_iter is not None:
valid_loss = evaluate_loss(valid_iter, net, devices)
animator.add(epoch + 1, (None, valid_loss.detach().cpu()))
scheduler.step()
if valid_iter is not None:
measures += f', valid loss {valid_loss:.3f}'
print(measures + f'\n{metric[1] * num_epochs / timer.sum():.1f}'
f' examples/sec on {str(devices)}')
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
num_epochs, lr, wd = 10, 1e-4, 1e-4
lr_period, lr_decay, net = 2, 0.9, get_net(None)
train(net, train_iter, valid_iter, num_epochs, lr, wd, lr_period,
lr_decay)
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
train loss 1.744, valid loss 1.946
2013.1 examples/sec on None
.. figure:: output_kaggle-dog_571091_79_1.svg
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
devices, num_epochs, lr, wd = d2l.try_all_gpus(), 10, 1e-4, 1e-4
lr_period, lr_decay, net = 2, 0.9, get_net(devices)
train(net, train_iter, valid_iter, num_epochs, lr, wd, devices, lr_period,
lr_decay)
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
train loss 1.055, valid loss 1.340
3706.6 examples/sec on [device(type='cuda', index=0)]
.. figure:: output_kaggle-dog_571091_82_1.svg
.. raw:: html
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
net, preds = get_net(None), []
train(net, train_valid_iter, None, num_epochs, lr, wd, lr_period,
lr_decay)
for X, _ in test_iter:
y_hat = net(X)
preds.extend(y_hat.argmax(axis=1).numpy().astype('int32'))
sorted_ids = list(range(1, test_ds.get_dataset_size() + 1))
sorted_ids.sort(key=lambda x: str(x))
df = pd.DataFrame({'id': sorted_ids, 'label': preds})
class_indexing = train_ds.get_class_indexing()
classes = sorted(class_indexing.items(), key=lambda x: x[0])
df['label'] = df['label'].apply(lambda x: classes[x][0])
df.to_csv('submission.csv', index=False)
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
train loss 1.586
2707.0 examples/sec on None
.. figure:: output_kaggle-dog_571091_88_1.svg
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
net = get_net(devices)
train(net, train_valid_iter, None, num_epochs, lr, wd, devices, lr_period,
lr_decay)
preds = []
for data, label in test_iter:
output = torch.nn.functional.softmax(net(data.to(devices[0])), dim=1)
preds.extend(output.cpu().detach().numpy())
ids = sorted(os.listdir(
os.path.join(data_dir, 'train_valid_test', 'test', 'unknown')))
with open('submission.csv', 'w') as f:
f.write('id,' + ','.join(train_valid_ds.classes) + '\n')
for i, output in zip(ids, preds):
f.write(i.split('.')[0] + ',' + ','.join(
[str(num) for num in output]) + '\n')
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
train loss 1.121
3817.9 examples/sec on [device(type='cuda', index=0)]
.. figure:: output_kaggle-dog_571091_91_1.svg
.. raw:: html
.. raw:: html
.. raw:: html
`讨论 `__
.. raw:: html
.. raw:: html
`讨论 `__
.. raw:: html
.. raw:: html