【复现】在细粒度CUB_200_2011数据集上复现分类程序IELT
我是在程序中，IELT-main/utils/dataset.py中设置download=True让它自己下载的。或者直接下zip也挺好用，有的时候git clone会出问题，下zip目前没什么问题。唯独还没看到测试的代码，没找得到测试方法，没找到可视化方法，未能应用于自己的数据集。缺啥装啥，由于代码比较新（复现于2023/12/8），我没遇到版本问题。由于代码中使用了较多的 os.path.jo
·
几个网站
细粒度领域的资源整合：Awesome Fine-Grained Image Analysis – Papers, Codes and Datasets (weixiushen.com)
深度学习各个领域的排行榜网站：Browse the State-of-the-Art in Machine Learning | Papers With Code
在CUB-200-2011上的排行：
CUB-200-2011 Benchmark (Fine-Grained Image Classification) | Papers With Code
IELT效果差不多，年份也近，所以选择了该网络。
项目情况
Article: Fine-Grained Visual Classification via Internal Ensemble Learning Transformer
Published in: IEEE Transactions on Multimedia ( Early Access ）
注：什么是Early Access的论文？ 为了学术成果更快更广泛的传播、交流，有些论文在正式见刊（印刷）之前便已网络见刊，这些论文已经被数据库收录，除了没有分配卷、期、页码外，论文的其他信息（内容、DOI等）均已确定，可以检索、下载、阅读和引用，被称为“Early Access论文”。
-------------------------------------------------------------------------------------------------------------------------------
我的复现
由于代码中使用了较多的 os.path.join ，会有正反斜杠问题，在windows10上训练时出了很多问题，我直接换平台，问题一个没有，代码很好跑。特别是没有版本问题，很好，这可能就是代码年份较新的好处吧。
唯独还没看到测试的代码，没找得到测试方法，没找到可视化方法，未能应用于自己的数据集。
TBC
硬件环境：
数据集： CUB_200_2011
我是在程序中，IELT-main/utils/dataset.py中设置download=True让它自己下载的。（建议）
也能自己下载好放到合适的目录。
在dataset.py中，作者给出的参考网址：
http://www.vision.caltech.edu/visipedia-data/CUB-200-2011/CUB_200_2011.tgz
目录组织形式参考：
clone到本地
git clone https://github.com/mobulan/ielt或者直接下zip也挺好用，有的时候git clone会出问题，下zip目前没什么问题。
准备虚拟环境
conda create -n IELT python=3.9
conda activate IELT
直接pytorch官网安装，这样最保险、最全面、最好用：Start Locally | PyTorch
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118缺啥装啥，由于代码比较新（复现于2023/12/8），我没遇到版本问题。
pip install timm tensorboard ml_collections scipy yacs pandas 下载预训练模型
需要一个谷歌账号，没有就直接买一个吧hh。
参考组织形式
修改配置文件
cub.yaml
参考
根目录setup.py文件
主要是选一下cfg_file,显卡
from settings.defaults import _C
from settings.setup_functions import *
config = _C.clone()
cfg_file = os.path.join('configs', 'cub.yaml')
# 可以修改log名称
# cfg_file = os.path.join('..','configs', 'cub.yaml')
config = SetupConfig(config, cfg_file)
config.defrost()
## Log Name and Perferences
config.write = True # comment it to disable all the log writing
config.train.checkpoint = True # comment it to disable saving the checkpoint
config.misc.exp_name = f'{config.data.dataset}'
config.misc.log_name = f'IELT'
config.cuda_visible = '0'
# 可以选择使用的显卡
# Environment Settings
config.data.log_path = os.path.join(config.misc.output, config.misc.exp_name, config.misc.log_name
+ time.strftime(' %m-%d_%H-%M', time.localtime()))
config.model.pretrained = os.path.join(config.model.pretrained,
config.model.name + config.model.pre_version + config.model.pre_suffix)
os.environ['CUDA_VISIBLE_DEVICES'] = config.cuda_visible
os.environ['OMP_NUM_THREADS'] = '1'
# Setup Functions
config.nprocess, config.local_rank = SetupDevice()
config.data.data_root, config.data.batch_size = LocateDatasets(config)
config.train.lr = ScaleLr(config)
log = SetupLogs(config, config.local_rank)
if config.write and config.local_rank in [-1, 0]:
with open(config.data.log_path + '/config.json', "w") as f:
f.write(config.dump())
config.freeze()
SetSeed(config)
开始训练
上述官方命令有点小问题，修改如下：
python main.py
或者
python -m main 输出
log
================================================================================
Fine-Grained Visual Classification via Internal Ensemble Learning Transformer
Pytorch Implementation
================================================================================
Author: Xu Qin, Wang Jiahui, Jiang Bo, Luo Bin
Institute: Anhui University Date: 2023-02-13
--------------------------------------------------------------------------------
Python Version: 3.9.18 Pytorch Version: 2.1.1+cu118 Cuda Version: 11.8
--------------------------------------------------------------------------------
============================ Data Settings ============================
dataset cub batch_size 4
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
data_root ./data/ img_size 448
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
resize 600 padding 0
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
no_crop 0 autoaug 0
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
blur 0.0 color 0.4
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
saturation 0.0 hue 0.0
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
mixup 0.0 cutmix 0.0
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
rotate 0.0 log_path ./output/cub/IELT 12-07_23-43
============================ Hyper Parameters ============================
vote_perhead 24 update_warm 10 loss_alpha 0.4
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
total_num 126 fix 1 dsm 1
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
cam 1 assess 0
============================ Training Settings ============================
start_epoch 0 epochs 50 warmup_epochs 0
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
weight_decay 0 clip_grad None checkpoint 1
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
lr 0.0025 scheduler cosine lr_epoch_update0
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
optimizer SGD freeze_backbone0 eps 1e-08
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
betas 0.9,0.999 momentum 0.9
============================ Other Settings ============================
amp 1 output ./output exp_name cub
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
log_name IELT eval_every 1 seed 42
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
eval_mode 0 throughput 0 fused_window 1
============================ Model Structure ============================
type ViT name ViT-B_16 baseline_model0
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
pretrained pretrained/ViT-B_16.npz pre_version pre_suffix .npz
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
resume num_classes 200 drop_path 0.0
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
dropout 0.0 label_smooth 0.0 parameters 93.493M
============================ Training Information ============================
Train samples 5994 Test samples 5794
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Total Batch Size 4 Load Time 3s
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Train Steps 74900 Warm Epochs 0
============================ Start Training ============================
Epoch 1 /50 : Accuracy 66.051 BA 66.051 BE 1 Loss 1.1079 TA 45.91
Epoch 2 /50 : Accuracy 69.814 BA 69.814 BE 2 Loss 0.9985 TA 62.67
Epoch 3 /50 : Accuracy 71.246 BA 71.246 BE 3 Loss 1.0025 TA 71.04
Epoch 4 /50 : Accuracy 72.006 BA 72.006 BE 4 Loss 1.0287 TA 78.81
Epoch 5 /50 : Accuracy 72.523 BA 72.523 BE 5 Loss 1.0503 TA 82.93
Epoch 6 /50 : Accuracy 75.768 BA 75.768 BE 6 Loss 0.9947 TA 87.40
Epoch 7 /50 : Accuracy 74.836 BA 75.768 BE 6 Loss 1.0967 TA 88.85
Epoch 8 /50 : Accuracy 78.961 BA 78.961 BE 8 Loss 0.9542 TA 91.51
Epoch 9 /50 : Accuracy 81.101 BA 81.101 BE 9 Loss 0.8824 TA 92.99
Epoch 10 /50 : Accuracy 80.290 BA 81.101 BE 9 Loss 0.9127 TA 95.11
Epoch 11 /50 : Accuracy 79.703 BA 81.101 BE 9 Loss 0.9738 TA 94.04
Epoch 12 /50 : Accuracy 78.840 BA 81.101 BE 9 Loss 1.0418 TA 95.51
Epoch 13 /50 : Accuracy 82.171 BA 82.171 BE 13 Loss 0.8814 TA 96.76
Epoch 14 /50 : Accuracy 82.137 BA 82.171 BE 13 Loss 0.9301 TA 97.55
Epoch 15 /50 : Accuracy 82.206 BA 82.206 BE 15 Loss 0.9095 TA 97.91
Epoch 16 /50 : Accuracy 82.171 BA 82.206 BE 15 Loss 0.9413 TA 98.87
Epoch 17 /50 : Accuracy 83.345 BA 83.345 BE 17 Loss 0.8929 TA 98.93
Epoch 18 /50 : Accuracy 83.828 BA 83.828 BE 18 Loss 0.8763 TA 99.33
Epoch 19 /50 : Accuracy 84.311 BA 84.311 BE 19 Loss 0.8594 TA 99.73
Epoch 20 /50 : Accuracy 84.001 BA 84.311 BE 19 Loss 0.8794 TA 99.65
Epoch 21 /50 : Accuracy 84.122 BA 84.311 BE 19 Loss 0.8902 TA 99.47
Epoch 22 /50 : Accuracy 85.243 BA 85.243 BE 22 Loss 0.8365 TA 99.93
Epoch 23 /50 : Accuracy 84.984 BA 85.243 BE 22 Loss 0.8427 TA 99.77
Epoch 24 /50 : Accuracy 84.829 BA 85.243 BE 22 Loss 0.8708 TA 99.77
Epoch 25 /50 : Accuracy 85.243 BA 85.243 BE 22 Loss 0.8370 TA 99.77
Epoch 26 /50 : Accuracy 86.055 BA 86.055 BE 26 Loss 0.8122 TA 99.97
Epoch 27 /50 : Accuracy 86.210 BA 86.210 BE 27 Loss 0.8063 TA 99.98
Epoch 28 /50 : Accuracy 85.882 BA 86.210 BE 27 Loss 0.8244 TA 99.95
Epoch 29 /50 : Accuracy 85.658 BA 86.210 BE 27 Loss 0.8194 TA 99.97
Epoch 30 /50 : Accuracy 85.934 BA 86.210 BE 27 Loss 0.8201 TA 99.97
Epoch 31 /50 : Accuracy 85.968 BA 86.210 BE 27 Loss 0.8277 TA 99.93
Epoch 32 /50 : Accuracy 85.986 BA 86.210 BE 27 Loss 0.8212 TA 99.97
Epoch 33 /50 : Accuracy 86.210 BA 86.210 BE 27 Loss 0.8157 TA 99.95
Epoch 34 /50 : Accuracy 86.124 BA 86.210 BE 27 Loss 0.8106 TA 100.00
Epoch 35 /50 : Accuracy 85.986 BA 86.210 BE 27 Loss 0.8110 TA 99.98
Epoch 36 /50 : Accuracy 86.158 BA 86.210 BE 27 Loss 0.8093 TA 100.00
Epoch 37 /50 : Accuracy 86.020 BA 86.210 BE 27 Loss 0.8084 TA 99.98
Epoch 38 /50 : Accuracy 86.141 BA 86.210 BE 27 Loss 0.8073 TA 100.00
Epoch 39 /50 : Accuracy 86.158 BA 86.210 BE 27 Loss 0.8032 TA 99.93
Epoch 40 /50 : Accuracy 86.175 BA 86.210 BE 27 Loss 0.8025 TA 100.00
Epoch 41 /50 : Accuracy 86.210 BA 86.210 BE 27 Loss 0.8016 TA 100.00
Epoch 42 /50 : Accuracy 86.193 BA 86.210 BE 27 Loss 0.8015 TA 100.00
Epoch 43 /50 : Accuracy 86.193 BA 86.210 BE 27 Loss 0.8014 TA 100.00
Epoch 44 /50 : Accuracy 86.175 BA 86.210 BE 27 Loss 0.8024 TA 99.97
Epoch 45 /50 : Accuracy 86.175 BA 86.210 BE 27 Loss 0.8018 TA 99.95
Epoch 46 /50 : Accuracy 86.193 BA 86.210 BE 27 Loss 0.8020 TA 100.00
Epoch 47 /50 : Accuracy 86.227 BA 86.227 BE 47 Loss 0.8017 TA 99.98
Epoch 48 /50 : Accuracy 86.175 BA 86.227 BE 47 Loss 0.8017 TA 100.00
Epoch 49 /50 : Accuracy 86.210 BA 86.227 BE 47 Loss 0.8018 TA 99.98
Epoch 50 /50 : Accuracy 86.175 BA 86.227 BE 47 Loss 0.8017 TA 99.97
============================ Finish Training ============================
Best Accuracy 86.227 Best Epoch 47
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Training Time 126.95 min Testing Time 44.73 min
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Total Time 171.68 min 开始测试
源代码没有提供测试程序，所以自己写了一个test.py放在根目录中，需要在./data/ForTest放几张测试图片。
import os
import torch
from torchvision import transforms
from PIL import Image
import matplotlib.pyplot as plt
from models.build import build_models # 导入模型构建函数
from setup import config # 导入配置
from utils.data_loader import build_transforms
def load_model(model_path, num_classes):
model = build_models(config, num_classes)
checkpoint = torch.load(model_path)
model.load_state_dict(checkpoint['model']) # 只加载模型部分
model.eval()
model.cuda()
return model
def classify_images(model, image_folder):
transform = transforms.Compose([
transforms.Resize((config.data.img_size, config.data.img_size)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
image_paths = [os.path.join(image_folder, f) for f in os.listdir(image_folder) if f.endswith(('.png', '.jpg', '.jpeg'))]
results = []
for img_path in image_paths:
img = Image.open(img_path).convert('RGB')
img_tensor = transform(img).unsqueeze(0).cuda() # 添加 batch 维度并转到 GPU
with torch.no_grad():
logits = model(img_tensor)
PRED = torch.argmax(logits, dim=1).item()
print("PRED:", PRED)
class_map = load_class_mapping(classes_file_path)
pred = PRED + 1 # 预测编号和classes.txt相对应
preds = class_map[pred]
print("preds:", preds)
results.append((img_path, preds))
print("results:", results)
return results
def visualize_results(results):
for img_path, pred in results:
img = Image.open(img_path)
plt.imshow(img)
plt.title(f'Predicted Class: {pred}')
plt.axis('off')
plt.savefig(f'{img_path}.png')
plt.show() # 可视化分类情况，并存储
def load_class_mapping(file_path):
class_mapping = {}
with open(file_path, 'r') as file:
for line in file:
index, name = line.strip().split(' ', 1)
class_mapping[int(index)] = name
return class_mapping
if __name__ == '__main__':
classes_file_path = './data/CUB_200_2011/CUB_200_2011/classes.txt'
model_path = './output/cub/IELT 09-14_15-30/checkpoint.bin' # 替换为你的模型路径
image_folder = './data/ForTest' # 替换为待测试的文件夹路径
num_classes = 200 # 根据你的模型设置类别数
model = load_model(model_path, num_classes)
results = classify_images(model, image_folder)
visualize_results(results)
可视化情况：
找了几张网络图片试一试：
下一步制作自己的细粒度数据集，然后放里面跑跑看情况。
TBC
开放原子开发者工作坊旨在鼓励更多人参与开源活动，与志同道合的开发者们相互交流开发经验、分享开发心得、获取前沿技术趋势。工作坊有多种形式的开发者活动，如meetup、训练营等，主打技术交流，干货满满，真诚地邀请各位开发者共同参与！
更多推荐
23
0- 0
已为社区贡献1条内容
所有评论(0)