pycaffe训练的完整组件示例

pycaffe训练的完整组件示例

为什么写这篇博客

1. 需要用到pycaffe

因为用到的开源代码基于Caffe；要维护的项目基于Caffe。基本上是用Caffe的Python接口。

2. 训练中想穿插验证并输出关注的指标

比如每训练完1个epoch就应该在完整的validation集合上执行evaluation，输出测量出的、关注的指标，例如AP、Accuracy、F1-score等。Caffe通过solver.prototxt中配置test_net能执行测试，但基本只能输出Accuracy而且是各个test_batch上的平均Accuracy，而不是想关注的验证集整体上的AP（见Solver.cpp源码）

3. 训练中期望有可视化输出

Caffe训练输出在屏幕终端，也可自行重定向到日志文件。的确可以自行解析日志文件，并结合flask搭建web页面实时显示输出。但是这不够标准和鲁棒。期望有专门的可视化工具，避免自己造难用的轮子。

本文给出很简陋的pyCaffe和VisualDL结合的例子。

解决方案

用pycaffe接管训练接口

通过自行编写python代码来执行训练，而不是用$CAFFE_ROOT/build/tools/caffe train --solver solver.prototxt的方式来启动。

• solver.prototxt中需要配置test_net, test_iter, test_interval，保证solver有test_net对象
• test_interval设置为999999999，以避开Solver.cpp中执行的TestAll()函数，转而在python代码中手动判断和执行validation
• 执行validation之前注意test_net.share_with(train_net)
• 利用solver.step(1)执行训练网络的一次迭代，利用solver.test_net[0].forward()执行测试网络的一次前传
• 利用net.blobs['prob'].data的形式取出网络输出
• 利用sklearn.metrics包，将取出的数据执行evaluation
• 利用VisualDL等可视化工具，将取出的数据执行绘图

依赖项

VisualDL，是PaddlePaddle和ECharts团队联合推出的，应该是对抗谷歌的Tensorboarde的。相信ECharts的实力。

sudo pip install visualdl

看起来VisualDL和Tensorboard类似，不过对于Caffe，用不了Tensorboard，能用VisualDL也是好事。

参考代码

solve.py

#!/usr/bin/env python2
# coding: utf-8

"""
inspired and adapted from:
    - https://github.com/shelhamer/fcn.berkeleyvision.org
    - https://github.com/rbgirshick/py-faster-rcnn
    - https://github.com/PaddlePaddle/VisualDL/blob/develop/docs/quick_start_en.md
"""

from __future__ import print_function
import _init_paths
import caffe
import argparse
import os
import sys
from datetime import datetime
import cv2

from caffe.proto import caffe_pb2
import google.protobuf as pb2
import google.protobuf.text_format
import numpy as np
import perfeval

from visualdl import LogWriter #for visualization during training

def parse_args():
    """Parse input arguments"""
    parser = argparse.ArgumentParser(description='Train a classification network')
    parser.add_argument('--solver', dest='solver',
                        help='solver prototxt',
                        default=None, type=str, required=True)

    parser.add_argument('--weights', dest='pretrained_model',
                        help='initialize with pretrained model weights',
                        default=None, type=str)

    if len(sys.argv) == 1:
        parser.print_help()
        sys.exit(1)

    args = parser.parse_args()
    return args

class SolverWrapper:
    """对于Solver进行封装，便于外部调用"""
    def __init__(self, solver_prototxt, num_epoch, num_example, pretrained_model=None):
        self.solver = caffe.SGDSolver(solver_prototxt)
        if pretrained_model is not None:
            print('Loading pretrained model weights from {:s}'.format(pretrained_model))
            self.solver.net.copy_from(pretrained_model)

        self.solver_param = caffe_pb2.SolverParameter()
        with open(solver_prototxt, 'rt') as f:
            pb2.text_format.Merge(f.read(), self.solver_param)
        self.cur_epoch = 0
        self.test_interval = 100  #用来替代self.solver_param.test_interval
        self.logw = LogWriter("catdog_log", sync_cycle=100)
        with self.logw.mode('train') as logger:
            self.sc_train_loss = logger.scalar("loss")
            self.sc_train_acc = logger.scalar("Accuracy")
        with self.logw.mode('val') as logger:
            self.sc_val_acc = logger.scalar("Accuracy")
            self.sc_val_mAP = logger.scalar("mAP")

    def train_model(self):
        """执行训练的整个流程，穿插了validation"""
        cur_iter = 0
        test_batch_size, num_classes = self.solver.test_nets[0].blobs['prob'].shape
        num_test_images_tot = test_batch_size * self.solver_param.test_iter[0]
        while cur_iter < self.solver_param.max_iter:
            #self.solver.step(self.test_interval)
            for i in range(self.test_interval):
                self.solver.step(1)
                loss = self.solver.net.blobs['loss'].data
                acc = self.solver.net.blobs['accuracy'].data
                step = self.solver.iter
                self.sc_train_loss.add_record(step, loss)
                self.sc_train_acc.add_record(step, acc)

            self.eval_on_val(num_classes, num_test_images_tot, test_batch_size)
            cur_iter += self.test_interval

    def eval_on_val(self, num_classes, num_test_images_tot, test_batch_size):
        """在整个验证集上执行inference和evaluation"""
        self.solver.test_nets[0].share_with(self.solver.net)
        self.cur_epoch += 1
        scores = np.zeros((num_classes, num_test_images_tot), dtype=float)
        gt_labels = np.zeros((1, num_test_images_tot), dtype=float).squeeze()
        for t in range(self.solver_param.test_iter[0]):
            output = self.solver.test_nets[0].forward()
            probs = output['prob']
            labels = self.solver.test_nets[0].blobs['label'].data

            gt_labels[t*test_batch_size:(t+1)*test_batch_size] = labels.T.astype(float)
            scores[:,t*test_batch_size:(t+1)*test_batch_size] = probs.T

        ap, acc = perfeval.cls_eval(scores, gt_labels)
        print('====================================================================\n')
        print('\tDo validation after the {:d}-th training epoch\n'.format(self.cur_epoch))
        print('>>>>', end='\t')  #设定标记，方便于解析日志获取出数据
        for i in range(num_classes):
            print('AP[{:d}]={:.2f}'.format(i, ap[i]), end=', ')
        mAP = np.average(ap)
        print('mAP={:.2f}, Accuracy={:.2f}'.format(mAP, acc))
        print('\n====================================================================\n')
        step = self.solver.iter
        self.sc_val_mAP.add_record(step, mAP)
        self.sc_val_acc.add_record(step, acc)

if __name__ == '__main__':
    args = parse_args()
    solver_prototxt = args.solver
    num_epoch = args.num_epoch
    num_batch = args.num_batch
    pretrained_model = args.pretrained_model

    # init
    caffe.set_mode_gpu()
    caffe.set_device(0)

    sw = SolverWrapper(solver_prototxt, num_epoch, num_batch, pretrained_model)
    sw.train_model()

perfeval.py

#!/usr/bin/env python2
# coding: utf-8

from __future__ import print_function
import numpy as np

import sklearn.metrics as metrics

def cls_eval(scores, gt_labels):
    """
    分类任务的evaluation
    @param scores: cxm　np-array, m为样本数量(例如一个epoch)
    @param gt_labels: 1xm np-array, 元素属于{0,1,2,...,K-1}，表示K个类别的索引
    """
    num_classes, num_test_imgs = scores.shape

    pred_labels = scores.argmax(axis=0)

    ap = np.zeros((num_classes, 1), dtype=float).squeeze()
    for i in range(num_classes):
        cls_labels = np.zeros((1, num_test_imgs), dtype=float).squeeze()
        for j in range(num_test_imgs):
            if gt_labels[j]==i:
                cls_labels[j]=1
        ap[i] = metrics.average_precision_score(cls_labels, scores[i])

    acc = metrics.accuracy_score(gt_labels, pred_labels)

    return ap, acc

样例输出

首先需要开启训练，比如:

python solve.py

然后启动VisualDL：

visualDL --logdir=catdog_log --port=8080

打开浏览器获取训练的实时更新的绘图输出：http://localhost:8080。这里仅截图展示：