快速开始

1. 安装环境

  • python版本: 2.7 OR >=3.4
  • TensorFlow版本: 开源TF >= 1.8(安装) OR PAI-TF

2. 安装步骤

  • pip install cython
  • pip install http://pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com/release/easy_vision-1.0.7-py2.py3-none-any.whl
  • 验证:
import easy_vision

如下安装第三方库下载缓慢，可以采取如下方式更改pip源加速

新建 ~/.config/pip/pip.conf 文件

[global]
index-url = https://pypi.tuna.tsinghua.edu.cn/simple

3. 使用步骤

3.1 数据准备

下载pascal数据集(在当前目录):

osscmd downloadallobject oss://pai-vision-data-hz/data/voc0712_tfrecord/ data/voc0712_tfrecord --host=oss-cn-zhangjiakou.aliyuncs.com
# or use oss util, 需要在配置文件中设置host为oss-cn-zhangjiakou.aliyuncs.com
ossutil  cp -r  oss://pai-vision-data-hz/data/voc0712_tfrecord/ data/voc0712_tfrecord

下载resnet50预训练模型(在当前目录):

mkdir -p pretrained_models/
ossutil cp -r oss://pai-vision-data-hz/pretrained_models/resnet_v1d_50/ pretrained_models/resnet_v1d_50

3.2 启动训练任务

普通配置模式

import easy_vision
easy_vision.train_and_evaluate(easy_vision.RFCN_SAMPLE_CONFIG)

参数配置模式

import easy_vision
param_config = """
  --model_type RFCN
  --backbone  resnet_v1d_50
  --num_classes 20
  --model_dir experiments/pascal_voc/resnet50_rfcn_model
  --train_data data/voc0712_tfrecord/voc0712_part_*.tfrecord
  --test_data data/voc0712_tfrecord/VOC2007_test.tfrecord
  --num_test_example 2
  --train_batch_size 32
  --test_batch_size 1
  --image_min_sizes 600
  --image_max_sizes 1024
  --lr_type exponential_decay
  --initial_learning_rate 0.001
  --decay_epochs 20
  --staircase true"""
easy_vision.train_and_evaluate_with_param_config(param_config)

PS模式

需要在具有至少两张GPU卡的机器上才能运行，一共需要启动3个子进程，分别是ps(parameter server), master(训练过程，负责写summary，保存checkpoint, 定期evaluation)，worker(训练过程)，脚本如下:

#-*- encoding:utf-8 -*-
import multiprocessing
import sys
import os
import easy_vision
import json
import logging
import subprocess
import time

# train config under distributed settings
config=easy_vision.RFCN_DISTRIBUTE_SAMPLE_CONFIG
print('config path: %s' % config)

# cluster spec 
TF_CONFIG={'cluster':{
             'ps': ['localhost:12921'],
             'master': ['localhost:12922'],
             'worker': ['localhost:12923']
            }
          }

def job(task, gpu):
  task_name = task['type']
  # redirect python log and tf log to log_file_name
  # [logs/master.log, logs/worker.log, logs/ps.log]
  log_file_name = "logs/%s.log" % task_name

  TF_CONFIG['task'] = task
  os.environ['TF_CONFIG'] = json.dumps(TF_CONFIG)
  os.environ['CUDA_VISIBLE_DEVICES'] = gpu
  train_cmd = 'python -m easy_vision.python.train_eval --pipeline_config_path %s' % config 
  logging.info('%s > %s 2>&1 ' % (train_cmd, log_file_name))
  with open(log_file_name, 'w') as lfile:
    return subprocess.Popen(train_cmd.split(' '), stdout= lfile, stderr=subprocess.STDOUT)

if __name__ == '__main__':
  procs = {}
  # start ps job on cpu
  task = {'type':'ps', 'index':0}
  procs['ps'] = job(task, '')
  # start master job on gpu 0
  task = {'type':'master', 'index':0}
  procs['master'] = job(task, '0')
  # start worker job on gpu 1
  task = {'type':'worker', 'index':0}
  procs['worker'] = job(task, '1')

  num_worker = 2
  for k, proc in procs.items():
    logging.info('%s pid: %d' %(k, proc.pid))
  task_failed = None 
  task_finish_cnt = 0
  task_has_finished = {k:False for k in procs.keys()}
  while True:
    for k, proc in procs.items():
      if proc.poll() is None:
        if task_failed is not None:
          logging.error('task %s failed, %s quit' % (task_failed, k))
          proc.terminate()
          if k != 'ps':
            task_has_finished[k] = True
            task_finish_cnt += 1
          logging.info('task_finish_cnt %d' % task_finish_cnt) 
      else:
        if not task_has_finished[k]:
          #process quit by itself
          if k != 'ps':
            task_finish_cnt += 1
            task_has_finished[k] = True
          logging.info('task_finish_cnt %d' % task_finish_cnt) 
          if proc.returncode != 0:
            logging.error('%s failed' %k)
            task_failed = k 
          else:
            logging.info('%s run successfuly' % k)

    if task_finish_cnt >= num_worker:
      break
    time.sleep(1)

3.3 使用tensorboard观察训练过程

在模型目录下面可以看到保存的模型的checkpoint和event file，通过tensorboard可以查看loss, mAP等相关信息

tensorboard --port 6006 --logdir ${model_dir}  [ --host 0.0.0.0 ]
  • 训练Loss

如下图所示，loss是总的loss, loss/loss/rcnn_cls是分类loss， loss/loss/rcnn_reg是回归的loss，loss/loss/rpn_cls是RPN(RegionProposalNetwork)的分类loss，loss/loss/rpn_reg是RPN的回归loss。_images/tensorboard_detection_loss.png

  • 测试mAP

如下图所示，这里使用了两个常用物体检测的metric：PascalBoxes07和PascalBoxes_images/tensorboard_detection_metric.png

3.4 评估和预测

训练完成之后可以在其它数据集上测试，得到每张图的检测结果

import easy_vision
detect_results = easy_vision.predict(easy_vision.RFCN_SAMPLE_CONFIG)

也可以在训练完以后再次评估，评估的结果

import easy_vision
eval_metrics = easy_vision.evaluate(easy_vision.RFCN_SAMPLE_CONFIG)

3.5 导出模型

EasyVision可以方便地将模型导出成savedmodel格式，方便后续使用python、c++等语言进行预测使用

import easy_vision
easy_vision.export(export_dir, easy_vision.RFCN_SAMPLE_CONFIG, checkpoint_path)

导出程序会在export_dir下以当前unix时间戳创建模型目录，把checkpoint导出成savedmodel存放在目录下。

3.6 模型预测

我们提供了savedmodel的python版本的预测器，下面以检测模型为例，其余的如分类、分割、OCR等预测器详见模型预测文档

import easy_vision
detector = ev.Detector(saved_model_path)
image = np.zeros([640, 480, 3],  dtype=np.float32)
output_dict = detector.predict([image])

预测结果以List of Json形式输出，List的Length与输入图像的张数相等，一下为各模型Json结果的示例与说明，Json结果样例如下：

{
  "detection_boxes": [[243.5308074951172, 197.69570922851562, 385.59625244140625, 247.7247772216797], [292.1929931640625, 114.28043365478516, 571.2748413085938, 165.09771728515625]], 
  "detection_scores": [0.9942291975021362, 0.9940272569656372],
  "detection_classes": [1, 1],
  "detection_classe_names": ["text", "text"]
}

3.7 SavedModel评估

评估已导出的savedmodel，metric会打印在控制台中，predictor_eval_config的样例配置参见sample_config

from easy_vision.python.main import predictor_evaluate
predictor_evaluate(predictor_eval_config)