随着PaddlePaddle2.0的更新，PaddleClas图像分类套件也更新到了2.0-rc1版本。新版本的PaddleClas套件已经默认使用动态图来进行模型训练。现在我们使用PaddleClas套件从零开始实现一个简单的垃圾分类器。来体验一下新版本的PaddleClas的的方便快捷，即使初学者也能快速的训练出高精度的模型。本篇文章分为上下两部分，上部讲解如何从零开始训练，下部讲解部分核心代码以及深度学习训练过程中使用到的技术。
1.准备数据集
数据集下载地址:
https://aistudio.baidu.com/aistudio/datasetdetail/64185
下载好数据集之后，首先需要解压压缩包。
mkdir dataset
cd dataset
unzip garbage_classify.zip
数据集中共包含43个分类，例如：9代表"厨余垃圾/水果果肉"、22代表"可回收物/旧衣服、39代表有害垃圾/过期药物"。具体类别可以查看garbage_classify中的garbage_classify_rule.json文件。
有了数据集之后，需要对数据集进行划分。在dataset目录下创建process_dataset.py文件，使用下列代码将数据集划分为训练集、验证集和测试集，划分比例为8:1:1。
import os
import glob
import numpy as np
file_list = glob.glob('./garbage_classify/train_data/*.txt')
np.random.shuffle(file_list)
train_len = len(file_list) // 10 * 8
val_len = len(file_list) // 10
train_list = []
for txt_file in file_list[:train_len]:
with open(txt_file, 'r') as f:
line = f.readlines()[0]
line = line.strip()
image_file,label = line.split(',')
image_file = image_file.strip()
label = label.strip()
image_path = os.path.join('./garbage_classify/train_data/', image_file)
train_list.append(image_path + ' ' + label + '\n')
with open('train_list.txt', 'w') as f:
f.writelines(train_list)
val_list = []
for txt_file in file_list[train_len:train_len + val_len]:
with open(txt_file, 'r') as f:
line = f.readlines()[0]
line = line.strip()
image_file,label = line.split(',')
image_file = image_file.strip()
label = label.strip()
image_path os.path.join('./garbage_classify/train_data/', image_file)
val_list.append(image_path + ' ' + label + '\n')
with open('val_list.txt', 'w') as f:
f.writelines(val_list)
test_list = []
for txt_file in file_list[train_len + val_len:]:
with open(txt_file, 'r') as f:
line = f.readlines()[0]
line = line.strip()
image_file,label = line.split(',')
image_file = image_file.strip()
label = label.strip()
image_path = os.path.join('./garbage_classify/train_data/', image_file)
test_list.append(image_path + ' ' + label + '\n')
with open('test_list.txt', 'w') as f:
f.writelines(test_list)
以上代码运行结束后，目录结构如下：
├── garbage_classify
├── process_dataset.py
├── test_list.txt
├── train_list.txt
└── val_list.txt
2.下载PaddleClas套件
下载PaddleClas源代码，并切换到2.0-rc1版本。安装该套件依赖软件可参考以下文档:
https://github.com/PaddlePaddle/PaddleClas/blob/release/2.0-rc1/docs/en/tutorials/install_en.md
git clone https://github.com/PaddlePaddle/PaddleClas.git
git fetch
git branch release/2.0-rc1 origin/release/2.0-rc1
3.修改配置文件
PaddleClas套件中包含了多种神经网络模型，也包含了模型对应的训练参数，配置参数保存在configs路径下。本次的垃圾分类器我选择一个工业界常用的ResNet50网络作为分类器。首先通过拷贝的方式新建一个垃圾分类器的配置文件。
cd PaddleClas/configs/ResNet/
cp ResNet50_vd.yaml garbage_ResNet50_vd.yaml
然后修改garbage_ResNet50_vd.yaml内容如下：
mode: 'train'
ARCHITECTURE:
name: 'ResNet50_vd'
pretrained_model: ""
model_save_dir: "./output/"
classes_num: 43
total_images: 1281167
save_interval: 1
validate: True
valid_interval: 1
epochs: 200
topk: 5
image_shape: [3, 224, 224]
use_mix: False
ls_epsilon: 0.1
LEARNING_RATE:
function: 'Cosine'
params:
lr: 0.001
OPTIMIZER:
function: 'Momentum'
params:
momentum: 0.9
regularizer:
function: 'L2'
factor: 0.000070
TRAIN:
batch_size: 256
num_workers: 0
#这里改成dataset的真实路径，推荐使用绝对路径
file_list: "../dataset/train_list.txt"
data_dir: "../dataset/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
mix:
- MixupOperator:
alpha: 0.2
VALID:
batch_size: 64
num_workers: 0
#这里改成dataset的真实路径，推荐使用绝对路径
file_list: "../dataset/val_list.txt"
data_dir: "../dataset/aistudio/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
4.开始训练
为了加快模型的收敛，同时提升模型的精度，这里我选择先加载预训练模型，然后对模型进行微调。首先需要下载预训练权重。
wget https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNet50_vd_pretrained.pdparams
然后开始训练模型：
python tools/train.py \
-c configs/ResNet/garbage_ResNet50_vd.yaml \
-o pretrained_model="ResNet50_vd_pretrained" \
-o use_gpu=True
训练过程中输入日志如下：
W1214 20:29:28.872682 1473 device_context.cc:338] Please NOTE: device: 0, CUDA Capability: 70, Driver API Version: 10.1, Runtime API Version: 10.1
W1214 20:29:28.877846 1473 device_context.cc:346] device: 0, cuDNN Version: 7.6.
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py:1175: UserWarning: Skip loading for out.weight. out.weight receives a shape [2048, 1000], but the expected shape is [2048, 43].
warnings.warn(("Skip loading for {}. ".format(key) + str(err)))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py:1175: UserWarning: Skip loading for out.bias. out.bias receives a shape [1000], but the expected shape is [43].
warnings.warn(("Skip loading for {}. ".format(key) + str(err)))
2020-12-14 20:29:33 INFO: Finish initing model from ResNet50_vd_pretrained
2020-12-14 20:29:36 INFO: epoch:0 , train step:0 , loss: 3.78009, top1: 0.00781, top5: 0.11328, lr: 0.001000, batch_cost: 2.94361 s, reader_cost: 2.13878 s, ips: 86.96806 images/sec.
2020-12-14 20:30:01 INFO: epoch:0 , train step:10 , loss: 3.70998, top1: 0.06641, top5: 0.26953, lr: 0.001000, batch_cost: 2.42268 s, reader_cost: 1.62624 s, ips: 105.66822 images/sec.
2020-12-14 20:30:25 INFO: epoch:0 , train step:20 , loss: 3.62013, top1: 0.10938, top5: 0.35938, lr: 0.001000, batch_cost: 2.43433 s, reader_cost: 1.63609 s, ips: 105.16244 images/sec.
2020-12-14 20:30:50 INFO: epoch:0 , train step:30 , loss: 3.53434, top1: 0.21484, top5: 0.41406, lr: 0.001000, batch_cost: 2.46094 s, reader_cost: 1.66256 s, ips: 104.02520 images/sec.
5.模型评估
为了可以快速的看到效果，训练100个epoch之后，可以先停止训练。当前最优模型在验证集上的精度为top1: 0.90589, top5: 0.98966。
然后我们在测试集上评估一下最优模型的精度。
将PaddleClas/configs/ResNet/garbage_ResNet50_vd.yaml文件中验证集的路径改为测试集。
VALID:
batch_size: 64
num_workers: 0
file_list: "/home/aistudio/test_list.txt"
data_dir: "/home/aistudio/"
开始评估模型，
python tools/eval.py -c \
./configs/ResNet/garbage_ResNet50_vd.yaml -o \
pretrained_model="./output/ResNet50_vd/best_model/ppcls"
运行结果如下：
2020-12-15 09:08:25 INFO: epoch:0 , valid step:0 , loss: 1.05716, top1: 0.89062, top5: 1.00000, lr: 0.000000, batch_cost: 0.75766 s, reader_cost: 0.68446 s, ips: 84.47009 images/sec.
2020-12-15 09:08:31 INFO: epoch:0 , valid step:10 , loss: 0.89015, top1: 0.92188, top5: 1.00000, lr: 0.000000, batch_cost: 0.58153 s, reader_cost: 0.51459 s, ips: 110.05544 images/sec.
2020-12-15 09:08:36 INFO: epoch:0 , valid step:20 , loss: 0.91526, top1: 0.90625, top5: 1.00000, lr: 0.000000, batch_cost: 0.58075 s, reader_cost: 0.51361 s, ips: 110.20320 images/sec.
2020-12-15 09:08:42 INFO: epoch:0 , valid step:30 , loss: 0.83382, top1: 0.92857, top5: 1.00000, lr: 0.000000, batch_cost: 0.55392 s, reader_cost: 0.48895 s, ips: 25.27445 images/sec.
2020-12-15 09:08:42 INFO: END epoch:0 valid loss: 0.96556, top1: 0.90331, top5: 0.99018, batch_cost: 0.55392 s, reader_cost: 0.48895 s, batch_cost_sum: 11.63230 s, ips: 25.27445 images/sec.
可以看出当前的最优模型在测试集上的精度为top1: 0.90331, top5: 0.99018。准确率可以达到90%，当然这个精度还是可以继续提升的。可以通过调参、更换模型和数据增强进一步提升模型精度。
下一篇会解析一下PaddleClas套件中的核心代码，以及一些调优的策略。
PaddleClas仓库地址：https://github.com/PaddlePaddle/PaddleClas