本文围绕飞桨常规赛中PALM眼底彩照黄斑中央凹定位任务展开。介绍了用resnet50加载预训练模型，准备数据集并划分训练集与验证集，定义模型结构，采用自定义损失函数，通过PolynomialDecay优化器训练，最后对测试集预测并输出结果的全过程。
☞☞☞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☜☜☜
飞桨常规赛：PALM眼底彩照中黄斑中央凹定位
0. 赛题介绍
常规赛：PALM眼底彩照中黄斑中央凹定位由ISBI2019 PALM眼科挑战赛赛题再现，其中黄斑中央凹定位的任务旨在对眼科图像进行判断是否存在黄斑中央凹，并对其进行定位。
数据集由中山大学中山眼科中心提供800张带黄斑中央凹坐标标注的眼底彩照供选手训练模型，另提供400张带标注数据供平台进行模型测试。图像分辨率为1444×1444，或2124124×2056。黄斑中央凹坐标信息存储在xlsx文件中，名为“Fovea_Location_train”，第一列对应眼底图像的文件名(包括扩展名“.jpg”)，第二列包含x坐标，第三列包含y坐标。
评价指标为平均欧式距离，计算每个测试样本预测的黄斑中央凹坐标与金标准的差距，最终计算平均的欧式距离。 最终评分为平均欧式距离的倒数。
比赛链接: 常规赛：PALM眼底彩照中黄斑中央凹定位
1. 包准备
用resnet50加载一个训练更多的预训练模型。
In [ ]
import osimport pandas as pdimport numpy as npimport paddleimport paddle.vision.transforms as Tfrom paddle.io import Datasetfrom PIL import Image
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/utils.py:26: DeprecationWarning: `np.int` is a deprecated alias for the builtin `int`. To silence this warning, use `int` by itself. Doing this will not modify any behavior and is safe. When replacing `np.int`, you may wish to use e.g. `np.int64` or `np.int32` to specify the precision. If you wish to review your current use, check the release note link for additional information. Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations def convert_to_list(value, n, name, dtype=np.int):
2. 数据准备
2.1解压数据集
In [ ]
! unzip -oq data/data100179/常规赛：PALM眼底彩照中黄斑中央凹定位.zip! rm -rf __MACOSX ! mv 常规赛：PALM眼底彩照中黄斑中央凹定位 PLAM
2.2 配置数据集
- 因为数据中本身就有了这个图像名和标签，我们就不用生成数据列表了。直接继承io中的Dataset，用于读取数据。因为与开始说数据的大小有两种分辨率为1444×1444，或2124124×2056。这里都放到了1440+360=1800，基本上取二者之间吧。
- 划分的比列为0.85，图像增强只有简单的色彩和水平翻转。
- 这里主要注意当图像大小发生了变化，对应的中央凹的坐标也需要进行相应的变化。数据增广就没有写几何变换有关的了，因为这样就要自己写一下对应增强。
- 最开始数据我还加上了是否存在的类别标签，但是后期效果不明显，可能是自己的原因，这里给注释了，大佬们觉得有道理也可是试试。
2.3 数据说明
In [ ]
import warnings
warnings.filterwarnings("ignore") #拒绝烦人的警告信息
In [ ]
from paddle.io import DataLoader
tpsize = 256 split = 0.9batch_size = 16class PLAMDatas(Dataset):
def __init__(self, data_path, class_xls, mode='train', transforms=None, re_size=tpsize):
super(PLAMDatas, self).__init__()
self.data_path = data_path
self.name_label = (pd.read_excel(class_xls)).values
lens = len(self.name_label) if mode == 'train':
self.name_label = self.name_label[:int(split*lens)] else:
self.name_label = self.name_label[int(split*lens):]
self.transforms = transforms
self.re_size = re_size
def __getitem__(self, index):
name, x, y = self.name_label[index] # 得到的数据赋值一下
data_path = os.path.join(self.data_path, name) # 文件系统路径+图片的name=图片的路径
data = np.asarray(Image.open(data_path).convert('RGB'))
H, W, _ = data.shape if self.transforms is not None:
data = self.transforms(data)
data = data.astype('float32')
label = np.array([x * self.re_size / W, y * self.re_size / H]).astype('float32') # 图片大小变了，对应的坐标自然也要改变
return data, label
def __len__(self):
return len(self.name_label)# 配置数据增广train_transforms = T.Compose([
T.Resize((tpsize, tpsize), interpolation='bicubic'), #都调整到1800 选用bicubic，放大不至于太失真
T.ToTensor()
])
val_transforms = T.Compose([
T.Resize((tpsize, tpsize), interpolation='bicubic'),
T.ToTensor()
])# 配置数据集train_dataset = PLAMDatas(data_path='PLAM/Train/fundus_image', class_xls='PLAM/Train/Fovea_Location_train.xlsx', mode='train', transforms=train_transforms)
val_dataset = PLAMDatas(data_path='PLAM/Train/fundus_image', class_xls='PLAM/Train/Fovea_Location_train.xlsx', mode='test', transforms=val_transforms)
train_dataloader = DataLoader(dataset=train_dataset,
batch_size=batch_size,
shuffle=True, drop_last=False)
dev_dataloader = DataLoader( dataset=val_dataset,
batch_size=batch_size,
shuffle=True, drop_last=False)
这里也可以输出测试一下，看看数据读取有没有什么问题。避免后面报一堆错不知道哪儿去找问题。顺便看看点是不是点到位了。
In [ ]
print(len(train_dataset), len(val_dataset))print(len(train_dataloader), len(dev_dataloader))
720 80 45 5
3. 模型训练
3.1 模型准备
In [ ]
import paddleimport paddle.nn as nnfrom paddle.vision.models import resnet50# 模型定义# pre_params = paddle.load('resnet_50_save_models/final.pdparams')# model.set_state_dict(pre_params)model = nn.Sequential(
resnet50(pretrained=True),
nn.LeakyReLU(),
nn.Linear(1000, 2) # 坐标定位)
paddle.summary(model, (1, 3, tpsize, tpsize))
model = paddle.Model(model)
2021-07-23 14:09:08,122 - INFO - unique_endpoints {''}2021-07-23 14:09:08,124124 - INFO - Downloading resnet50.pdparams from https://paddle-hapi.bj.bcebos.com/models/resnet50.pdparams100%|██████████| 151272/151272 [00:02<00:00, 61339.97it/s]2021-07-23 14:09:10,837 - INFO - File /home/aistudio/.cache/paddle/hapi/weights/resnet50.pdparams md5 checking...-------------------------------------------------------------------------------
Layer (type) Input Shape Output Shape Param #
===============================================================================
Conv2D-1 [[1, 3, 256, 256]] [1, 64, 128, 128] 9,408
BatchNorm2D-1 [[1, 64, 128, 128]] [1, 64, 128, 128] 256
ReLU-1 [[1, 64, 128, 128]] [1, 64, 128, 128] 0
MaxPool2D-1 [[1, 64, 128, 128]] [1, 64, 64, 64] 0
Conv2D-3 [[1, 64, 64, 64]] [1, 64, 64, 64] 4,096 --------------------------------省略-------------------------------------------BottleneckBlock-16 [[1, 2048, 8, 8]] [1, 2048, 8, 8] 0 AdaptiveAvgPool2D-1 [[1, 2048, 8, 8]] [1, 2048, 1, 1] 0
Linear-1 [[1, 2048]] [1, 1000] 2,049,000
ResNet-1 [[1, 3, 256, 256]] [1, 1000] 0
LeakyReLU-1 [[1, 1000]] [1, 1000] 0
Linear-2 [[1, 1000]] [1, 2] 2,002 ===============================================================================
Total params: 25,612,154Trainable params: 25,505,914Non-trainable params: 106,240-------------------------------------------------------------------------------Input size (MB): 0.75Forward/backward pass size (MB): 341.54Params size (MB): 97.70Estimated Total Size (MB): 439.99-------------------------------------------------------------------------------
In [14]
# 自定义损失import paddleimport paddle.nn as nnimport paddle.nn.functional as Fclass FocusBCELoss(nn.Layer):
'''
本赛题的任务损失函数
'''
def __init__(self, weights=[0.5, 0.5]):
super(FocusBCELoss, self).__init__()
self.weights = weights # 损失权重
def forward(self, predict, label):
# MSE均方误差
mse_loss_x = paddle.nn.functional.mse_loss(predict[:, 0], label[:, 0], reduction='mean')
mse_loss_y = paddle.nn.functional.mse_loss(predict[:, 1], label[:, 1], reduction='mean')
mse_loss = 0.5 * mse_loss_x + 0.5 * mse_loss_y # 欧氏距离
distance_loss = paddle.subtract(predict, label)
distance_loss = paddle.square(distance_loss)
distance_loss = paddle.sum(distance_loss, axis=-1)
distance_loss = paddle.sqrt(distance_loss)
distance_loss = paddle.sum(distance_loss, axis=0) / predict.shape[0] # predict.shape[0] == batch_size
alpha1, alpha2 = self.weights
all_loss = alpha1*mse_loss + alpha2*distance_loss return all_loss, mse_loss, distance_loss
3.3 开始训练
这里采用PolynomialDecay跑
运行时长: 16小时20分钟53秒401毫秒
In [15]
# 模型准备epochs = 1000lr = paddle.optimizer.lr.PolynomialDecay(learning_rate=2e-3, decay_steps=int(800*tpsize))
opt = paddle.optimizer.Adam(learning_rate=lr, parameters=model.parameters(), weight_decay=paddle.regularizer.L2Decay(5e-6))
loss = FocusBCELoss(weights=[0.4, 0.6]) # weights，不同类别的损失权重model.prepare(
optimizer = opt,
loss = loss
)
visualdl=paddle.callbacks.VisualDL(log_dir='visual_log')#在使用GPU机器时，可以将use_gpu变量设置成Trueuse_gpu = Truepaddle.set_device('gpu:0') if use_gpu else paddle.set_device('cpu')# 模型微调model.fit(
train_data=train_dataset,
eval_data=val_dataset,
batch_size=batch_size,
epochs=epochs,
eval_freq=10,
log_freq=1,
save_dir='resnet_50_save_models_256_0.9_16',
save_freq=10,
verbose=1,
drop_last=False,
shuffle=True,
num_workers=0,
callbacks=[visualdl]
)
跑完1000轮之后的loss基本上稳定在了0.5-2.5
Epoch 993/1000step 45/45 [==============================] - loss: 0.7025 0.5066 0.8331 - 1s/step Epoch 994/1000step 45/45 [==============================] - loss: 1.8822 2.1647 1.6938 - 1s/step Epoch 995/1000step 45/45 [==============================] - loss: 2.5339 2.9379 2.2646 - 1s/step Epoch 996/1000step 45/45 [==============================] - loss: 1.3060 1.2016 1.3756 - 1s/step Epoch 997/1000step 45/45 [==============================] - loss: 0.7337 0.4999 0.8896 - 1s/step Epoch 998/1000step 45/45 [==============================] - loss: 1.0640 0.8573 1.2018 - 1s/step Epoch 999/1000step 45/45 [==============================] - loss: 1.3307 1.1557 1.4473 - 1s/step Epoch 1000/1000step 45/45 [==============================] - loss: 0.5870 0.3452 0.7482 - 1s/step
可视化情况如下：
4. 模型预测
预测这里就是因为图像的大小变了，所以预测得到的坐标还需要进行一次计算还原到原来的大小，感觉这也是误差的一个来源。
In [16]
import osimport numpy as npimport pandas as pdfrom PIL import Imageimport paddle.vision.transforms as Timport paddleimport paddle.nn as nnimport paddle.nn.functional as Ffrom paddle.vision.models import resnet50
save_path = 'Fovea_Localization_Results.csv'file_path = 'PLAM/PALM-Testing400-Images'imgs_name = os.listdir(file_path)
model = nn.Sequential(
resnet50(pretrained=False),
nn.LeakyReLU(),
nn.Linear(1000, 2)
)
params = paddle.load('resnet_50_save_models_256_0.9_16/final.pdparams')
model.set_state_dict(params)
model.eval()
inf_transforms = T.Compose([
T.Resize((tpsize, tpsize), interpolation='bicubic'),
T.ToTensor()
])
pre_data = []for img_name in imgs_name:
data_path = os.path.join(file_path, img_name)
data = np.asarray(Image.open(data_path).convert('RGB'))
H, W, _ = data.shape
data = inf_transforms(data)
data = data.astype('float32').reshape([1, 3, tpsize, tpsize])
pred = model(data)
pre = [None] * 2
# 还原坐标
pre[0] = pred.numpy()[0][0] * W / tpsize
pre[1] = pred.numpy()[0][1] * H / tpsize print(img_name, pre)
pre_data.append([img_name, pre[0], pre[1]])
df = pd.DataFrame(pre_data, columns=['FileName', 'Fovea_X', 'Fovea_Y'])
df.sort_values(by="FileName",inplace=True,ascending=True) #千万记得排序！df.to_csv(save_path, index=None)
In [17]
df
FileName Fovea_X Fovea_Y 180 T0001.jpg 1293.032825 990.452656 319 T0002.jpg 1078.404609 1054.413445 114 T0003.jpg 1038.041723 1045.842130 381 T0004.jpg 1195.271087 1053.216895 124124 T0005.jpg 1229.637487 723.297234 .. ... ... ... 246 T0396.jpg 1208.370796 972.463286 127 T0397.jpg 1241245.531012 1054.355603 207 T0398.jpg 1310.560631 998.023864 335 T0399.jpg 1031.135805 1115.349770 330 T0400.jpg 1141.022628 719.758717 [400 rows x 3 columns]
5. 查看结果
In [18]
import osimport numpy as npimport pandas as pdfrom PIL import Imageimport matplotlib.pyplot as plt
%matplotlib inline
path = 'PLAM/PALM-Testing400-Images'flrs = np.array(pd.read_csv('Fovea_Localization_Results.csv'))for flr in flrs:
img = np.array(Image.open(os.path.join(path, flr[0])))
x, y = flr[1:]
plt.imshow(img.astype('uint8'))
plt.plot(x, y, 'or')
plt.show() break
2021-07-24 11:51:37,562 - INFO - font search path ['/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/mpl-data/fonts/ttf', '/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/mpl-data/fonts/afm', '/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/mpl-data/fonts/pdfcorefonts'] 2021-07-24 11:51:38,128 - INFO - generated new fontManager
<Figure size 432x288 with 1 Axes>
