目录
安装python3
首先安装python3
sudo apt-get install python3
安装pip3，这里可能用 apt get安装
sudo apt-get install python3-pip
执行后可能会报错
The following packages have unmet dependencies:
python3-pip : Depends: python-pip-whl (= 9.0.1-2) but 9.0.1-2.3~ubuntu1.18.04.1 is to be installed
这里是python2的pip的依赖版本高了，导致无法安装
先卸载python2-pip
sudo apt-get remove python-pip
sudo apt-get remove python-pip-whl
再执行安装
sudo apt-get install python3-pip
再安装回python2-pip就行了
sudo apt-get install python-pip
也可以在此链接下载 py脚本来安装
https://bootstrap.pypa.io/get-pip.py
复制内容下来 储存到文件，命名为 get-pip.py
安装命令
sudo python3 get-pip.py
安装完成后可以升下级
python3 -m pip install pip
安装完成后查看版本
pip3 -V
下载tensorflow
这里使用pip3来安装tensorflow-gpu，还有可以去github用源码编译，这里就不展开说。
如果直接和pc上用pip命令直接安装时不行的找不到安装包，我们去nvidia官网下载安装包。
https://developer.nvidia.com/embedded/downloads#?search=tensorflow&tx=$product,jetson_nano
打开页面选择要下载的版本
这里选择了1.13.1 nv19.5的版本
pip安装下载可能有点慢，换国内源
修改 ~/.pip/pip.conf (没有就创建一个文件夹及文件）
cd ~
mkdir .pip
cd .pip
touch pip.conf
pip.conf文件写入，保存
[global]
index-url = https://pypi.tuna.tsinghua.edu.cn/simple
[install]
trusted-host=mirrors.aliyun.com
下载好了后，切换到下载目录，安装
pip3 install tensorflow_gpu-1.13.1+nv19.5-cp36-cp36m-linux_aarch64.whl
等待安装完成
安装依赖
pip3 install matplotlib --user
pip3 install scipy --user
这里scipy安装编译可能会出错，apt安装解决
sudo apt-get install python3-scipy
测试是否安装成功
新建文本 tensorflowTest.py
touch tensorflowTest.py
写入
import tensorflow as tf
from tensorflow import keras
import numpy as np
import matplotlib.pyplot as plt
#初始化一个Tensorflow的常量：Hello Google Tensorflow! 字符串，并命名为greeting作为一个计算模块
greeting = tf.constant('Hello Google Tensorflow!')
#启动一个会话
sess = tf.Session()
#使用会话执行greeting计算模块
result = sess.run(greeting)
print(result)
sess.close()
运行
python3 tensorflowTest.py
看到命令行打印出
Hello Google Tensorflow!
成功
可能遇到错误
运行测时的错误
ImportError: numpy.core._multiarray_umath failed to import
ImportError: numpy.core.umath failed to import
这个是numpy库的问题，我安装1.14.0会出现这个错误，1.13.3也会出现，安装1.17.4就没问题，可以检查看看。
pip安装时的错误
ERROR: Could not install packages due to an EnvironmentError: [Errno 13] Permission denied: '/usr/local/lib/python3.6/dist-packages/cycler.py'
Consider using the `--user` option or check the permissions.
这里是权限问题可以用 sudo或者在末尾加 --user
例如
pip3 install scipy --user
sudo python3 -m pip install scipy
SSD.Tensorflow训练自己的数据集
参考自
https://blog.csdn.net/czksnk/article/details/83010533
https://tensorflow.google.cn/install/source
https://blog.csdn.net/weixin_39881922/article/details/80569803
https://blog.csdn.net/w5688414/article/details/78395177
https://blog.csdn.net/liuyan20062010/article/details/78905517
https://blog.csdn.net/weixin_39881922/article/details/80569803
根据以上这些网站的信息，配置后只能跑他的模型来预测，训练自己的数据就不收敛，预测出来满屏的框框，不知道是否是我操作不对还是怎样，这个github工程的星星几千个，我看也有人反馈收敛问题，后来我换了个工程，就可以了，配置起来比上面那些还简单点。
github页面：https://github.com/HiKapok/SSD.TensorFlow/tree/AbsoluteCoord
他有其他版本，不过我自己标出来的数据集是绝对座标，voc也是绝对座标的，所以下载绝对座标版本。
这里可以下载我改好过的版本，里面有标好的数据集 和 测试图片
链接：https://download.csdn.net/download/ourkix/12038648
制作voc数据集
介绍
目录结构是这样的
|--VOCseahorse
|--VOC2007
|--Annotations
|--ImageSets
|--Layout
|--Main
|--Segmentation
|--JPEGImages
·这里解释下Annotations是放xml的，就是放置标注（物体位置 类别）好了的文件，用的labelImg软件标注的。
·ImageSets下面的Main放的是文本，里面有4个文件 test.txt val,txt train.txt trainval.txt,分别为测试图片，验证图片，训练图片，训练验证图片的文件名。其他Layout和Segmentation占时用不到，可以放Main中一样的文件进去。
·JPEGImages是放置图片的
下载我上传的文件的话，这个数据已经放在dataset里面了。
生成训练数据文件
首先切换到SSD.TensorFlow-AbsoluteCoord目录，进入dataset目录。
新建一个文件 splitImage.py,这个是用来把图片分几份一部分训练一部分测试。
import os
import random
xmlfilepath=r'./VOCseahorse/VOC2007/Annotations'
saveBasePath=r"./VOCseahorse/VOC2007/"
trainval_percent=0.8
train_percent=0.7
total_xml = os.listdir(xmlfilepath)
num=len(total_xml)
list=range(num)
tv=int(num*trainval_percent)
tr=int(tv*train_percent)
trainval= random.sample(list,tv)
train=random.sample(trainval,tr)
print("train and val size",tv)
print("traub suze",tr)
ftrainval = open(os.path.join(saveBasePath,'VOC2007/ImageSets/Main/trainval.txt'), 'w')
ftest = open(os.path.join(saveBasePath,'VOC2007/ImageSets/Main/test.txt'), 'w')
ftrain = open(os.path.join(saveBasePath,'VOC2007/ImageSets/Main/train.txt'), 'w')
fval = open(os.path.join(saveBasePath,'VOC2007/ImageSets/Main/val.txt'), 'w')
for i in list:
name=total_xml[i][:-4]+'\n'
if i in trainval:
ftrainval.write(name)
if i in train:
ftrain.write(name)
else:
fval.write(name)
else:
ftest.write(name)
ftrainval.close()
ftrain.close()
fval.close()
ftest .close()
切换到ImageSets/Main下，新建四个文件
touch train.txt
touch test.txt
touch val.txt
touch trainval.txt
回到dataset目录，运行（前提是有数据 有标注好的xml文件）
python3 splitImage.py
这样就分好了。
修改dataset_common.py文件，修改类别个数，训练图个数 验证图个数。如果用的我的数据集，注释掉原来的，改成这样就行了。至于train数和val数，去看你文件里面有 Main文件夹里的train.txt val.txt这两个文件，数一数几张图就知道了。
#VOC_LABELS = {
# 'none': (0, 'Background'),
# 'aeroplane': (1, 'Vehicle'),
# 'bicycle': (2, 'Vehicle'),
# 'bird': (3, 'Animal'),
# 'boat': (4, 'Vehicle'),
# 'bottle': (5, 'Indoor'),
# 'bus': (6, 'Vehicle'),
# 'car': (7, 'Vehicle'),
# 'cat': (8, 'Animal'),
# 'chair': (9, 'Indoor'),
# 'cow': (10, 'Animal'),
# 'diningtable': (11, 'Indoor'),
# 'dog': (12, 'Animal'),
# 'horse': (13, 'Animal'),
# 'motorbike': (14, 'Vehicle'),
# 'person': (15, 'Person'),
# 'pottedplant': (16, 'Indoor'),
# 'sheep': (17, 'Animal'),
# 'sofa': (18, 'Indoor'),
# 'train': (19, 'Vehicle'),
# 'tvmonitor': (20, 'Indoor'),
#}
#COCO_LABELS = {
# "bench": (14, 'outdoor') ,
# "skateboard": (37, 'sports') ,
# "toothbrush": (80, 'indoor') ,
# "person": (1, 'person') ,
# "donut": (55, 'food') ,
# "none": (0, 'background') ,
# "refrigerator": (73, 'appliance') ,
# "horse": (18, 'animal') ,
# "elephant": (21, 'animal') ,
# "book": (74, 'indoor') ,
# "car": (3, 'vehicle') ,
# "keyboard": (67, 'electronic') ,
# "cow": (20, 'animal') ,
# "microwave": (69, 'appliance') ,
# "traffic light": (10, 'outdoor') ,
# "tie": (28, 'accessory') ,
# "dining table": (61, 'furniture') ,
# "toaster": (71, 'appliance') ,
# "baseball glove": (36, 'sports') ,
# "giraffe": (24, 'animal') ,
# "cake": (56, 'food') ,
# "handbag": (27, 'accessory') ,
# "scissors": (77, 'indoor') ,
# "bowl": (46, 'kitchen') ,
# "couch": (58, 'furniture') ,
# "chair": (57, 'furniture') ,
# "boat": (9, 'vehicle') ,
# "hair drier": (79, 'indoor') ,
# "airplane": (5, 'vehicle') ,
# "pizza": (54, 'food') ,
# "backpack": (25, 'accessory') ,
# "kite": (34, 'sports') ,
# "sheep": (19, 'animal') ,
# "umbrella": (26, 'accessory') ,
# "stop sign": (12, 'outdoor') ,
# "truck": (8, 'vehicle') ,
# "skis": (31, 'sports') ,
# "sandwich": (49, 'food') ,
# "broccoli": (51, 'food') ,
# "wine glass": (41, 'kitchen') ,
# "surfboard": (38, 'sports') ,
# "sports ball": (33, 'sports') ,
# "cell phone": (68, 'electronic') ,
# "dog": (17, 'animal') ,
# "bed": (60, 'furniture') ,
# "toilet": (62, 'furniture') ,
# "fire hydrant": (11, 'outdoor') ,
# "oven": (70, 'appliance') ,
# "zebra": (23, 'animal') ,
# "tv": (63, 'electronic') ,
# "potted plant": (59, 'furniture') ,
# "parking meter": (13, 'outdoor') ,
# "spoon": (45, 'kitchen') ,
# "bus": (6, 'vehicle') ,
# "laptop": (64, 'electronic') ,
# "cup": (42, 'kitchen') ,
# "bird": (15, 'animal') ,
# "sink": (72, 'appliance') ,
# "remote": (66, 'electronic') ,
# "bicycle": (2, 'vehicle') ,
# "tennis racket": (39, 'sports') ,
# "baseball bat": (35, 'sports') ,
# "cat": (16, 'animal') ,
# "fork": (43, 'kitchen') ,
# "suitcase": (29, 'accessory') ,
# "snowboard": (32, 'sports') ,
# "clock": (75, 'indoor') ,
# "apple": (48, 'food') ,
# "mouse": (65, 'electronic') ,
# "bottle": (40, 'kitchen') ,
# "frisbee": (30, 'sports') ,
# "carrot": (52, 'food') ,
# "bear": (22, 'animal') ,
# "hot dog": (53, 'food') ,
# "teddy bear": (78, 'indoor') ,
# "knife": (44, 'kitchen') ,
# "train": (7, 'vehicle') ,
# "vase": (76, 'indoor') ,
# "banana": (47, 'food') ,
# "motorcycle": (4, 'vehicle') ,
# "orange": (50, 'food')
# }
VOC_LABELS = {
'none': (0, 'Background'),
'seahorse': (1, 'animal')
}
# use dataset_inspect.py to get these summary
data_splits_num = {
'train': 5,
'val': 3,
}
修改convert_tfrecords.py文件，这里是修改默认值，也可以不修改，到时运行时改输入参数也行。
tf.app.flags.DEFINE_string('dataset_directory', './dataset/VOC',
'All datas directory')
tf.app.flags.DEFINE_string('train_splits', 'VOC2007',
'Comma-separated list of the training data sub-directory')
tf.app.flags.DEFINE_string('validation_splits', 'VOC2007TEST',
'Comma-separated list of the validation data sub-directory')
tf.app.flags.DEFINE_string('output_directory', './dataset/tfrecords',
'Output data directory')
这里我们做个sh脚本，方便数据路径更改。换到SSD.TensorFlow-AbsoluteCoord目录。
新建tf_converdata.sh
touch tf_converdata.sh
chmod +x tf_converdata.sh
内容写，数据目录dataset_directory 输出训练用数据目录output_directory，没有的目录自己建。
python3 dataset/convert_tfrecords.py --dataset_directory=dataset/VOCseahorse/ --output_directory=./dataset/tfrecords/
运行
./tf_converdata.sh
这样数据集就准备好了。
训练
下载预训练模型。如果用的我上传的项目，已经在model文件夹里了。
下载链接：百度网盘 提取码：tg64
如果是TX2内存是没问题的，可以训练，nano就有点吃力。
nano关闭图形界面，训练时再关闭
# ubuntu关闭图形用户界面
sudo systemctl set-default multi-user.target
sudo reboot
# ubuntu启用图形用户界面
sudo systemctl set-default graphical.target
sudo reboot
切换到SSD.TensorFlow-AbsoluteCoord目录，新建文件train.sh脚本
touch train.sh
chmod +x train.sh
写入,保存
python3 train_ssd.py \
--num_classes=2 \
--batch_size=4 \
--max_number_of_steps=5000 \
--data_dir=./dataset/tfrecords \
--model_dir=./logs \
--save_summary_steps=2500 \
--save_checkpoints_steps=5000 \
--learning_rate=1e-3 \
--checkpoint_path=./model \
开始训练
./train.sh
这里也许会报错，一般是内存不足，重新运行train.sh就行了，多试几次。
测试
这里我训练了5000次。
修改simple_ssd_demo.py文件，一样下载的是我上传的话不用改。这里有个目录images，里面放的是测试用的出图片。
# Copyright 2018 Changan Wang
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
# http://www.apache.org/licenses/LICENSE-2.0
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# =============================================================================
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import sys
import cv2
import tensorflow as tf
from scipy.misc import imread, imsave, imshow, imresize
import numpy as np
from net import ssd_net
from dataset import dataset_common
from preprocessing import ssd_preprocessing
from utility import anchor_manipulator
from utility import draw_toolbox
from utility import bbox_util
# scaffold related configuration
tf.app.flags.DEFINE_integer(
'num_classes', 21, 'Number of classes to use in the dataset.')
# model related configuration
tf.app.flags.DEFINE_integer(
'train_image_size', 300,
'The size of the input image for the model to use.')
tf.app.flags.DEFINE_string(
'data_format', 'channels_last', # 'channels_first' or 'channels_last'
'A flag to override the data format used in the model. channels_first '
'provides a performance boost on GPU but is not always compatible '
'with CPU. If left unspecified, the data format will be chosen '
'automatically based on whether TensorFlow was built for CPU or GPU.')
tf.app.flags.DEFINE_float(
'select_threshold', 0.2, 'Class-specific confidence score threshold for selecting a box.')
tf.app.flags.DEFINE_float(
'min_size', 4., 'The min size of bboxes to keep.')
tf.app.flags.DEFINE_float(
'nms_threshold', 0.45, 'Matching threshold in NMS algorithm.')
tf.app.flags.DEFINE_integer(
'nms_topk', 20, 'Number of total object to keep after NMS.')
tf.app.flags.DEFINE_integer(
'keep_topk', 200, 'Number of total object to keep for each image before nms.')
# checkpoint related configuration
tf.app.flags.DEFINE_string(
'checkpoint_path', './logs',
'The path to a checkpoint from which to fine-tune.')
tf.app.flags.DEFINE_string(
'model_scope', 'ssd300',
'Model scope name used to replace the name_scope in checkpoint.')
FLAGS = tf.app.flags.FLAGS
#CUDA_VISIBLE_DEVICES
def get_checkpoint():
if tf.gfile.IsDirectory(FLAGS.checkpoint_path):
checkpoint_path = tf.train.latest_checkpoint(FLAGS.checkpoint_path)
else:
checkpoint_path = FLAGS.checkpoint_path
return checkpoint_path
def main(_):
with tf.Graph().as_default():
out_shape = [FLAGS.train_image_size] * 2
image_input = tf.placeholder(tf.uint8, shape=(None, None, 3))
shape_input = tf.placeholder(tf.int32, shape=(2,))
features, output_shape = ssd_preprocessing.preprocess_for_eval(image_input, out_shape, data_format=FLAGS.data_format, output_rgb=False)
features = tf.expand_dims(features, axis=0)
output_shape = tf.expand_dims(output_shape, axis=0)
all_anchor_scales = [(30.,), (60.,), (112.5,), (165.,), (217.5,), (270.,)]
all_extra_scales = [(42.43,), (82.17,), (136.23,), (189.45,), (242.34,), (295.08,)]
all_anchor_ratios = [(1., 2., .5), (1., 2., 3., .5, 0.3333), (1., 2., 3., .5, 0.3333), (1., 2., 3., .5, 0.3333), (1., 2., .5), (1., 2., .5)]
# all_anchor_ratios = [(2., .5), (2., 3., .5, 0.3333), (2., 3., .5, 0.3333), (2., 3., .5, 0.3333), (2., .5), (2., .5)]
with tf.variable_scope(FLAGS.model_scope, default_name=None, values=[features], reuse=tf.AUTO_REUSE):
backbone = ssd_net.VGG16Backbone(FLAGS.data_format)
feature_layers = backbone.forward(features, training=False)
with tf.device('/cpu:0'):
anchor_encoder_decoder = anchor_manipulator.AnchorEncoder(positive_threshold=None, ignore_threshold=None, prior_scaling=[0.1, 0.1, 0.2, 0.2])
if FLAGS.data_format == 'channels_first':
all_layer_shapes = [tf.shape(feat)[2:] for feat in feature_layers]
else:
all_layer_shapes = [tf.shape(feat)[1:3] for feat in feature_layers]
all_layer_strides = [8, 16, 32, 64, 100, 300]
total_layers = len(all_layer_shapes)
anchors_height = list()
anchors_width = list()
anchors_depth = list()
for ind in range(total_layers):
_anchors_height, _anchors_width, _anchor_depth = anchor_encoder_decoder.get_anchors_width_height(all_anchor_scales[ind], all_extra_scales[ind], all_anchor_ratios[ind], name='get_anchors_width_height{}'.format(ind))
anchors_height.append(_anchors_height)
anchors_width.append(_anchors_width)
anchors_depth.append(_anchor_depth)
anchors_ymin, anchors_xmin, anchors_ymax, anchors_xmax, _ = anchor_encoder_decoder.get_all_anchors(tf.squeeze(output_shape, axis=0),
anchors_height, anchors_width, anchors_depth,
[0.5] * total_layers, all_layer_shapes, all_layer_strides,
[0.] * total_layers, [False] * total_layers)
location_pred, cls_pred = ssd_net.multibox_head(feature_layers, FLAGS.num_classes, anchors_depth, data_format=FLAGS.data_format)
if FLAGS.data_format == 'channels_first':
cls_pred = [tf.transpose(pred, [0, 2, 3, 1]) for pred in cls_pred]
location_pred = [tf.transpose(pred, [0, 2, 3, 1]) for pred in location_pred]
cls_pred = [tf.reshape(pred, [-1, FLAGS.num_classes]) for pred in cls_pred]
location_pred = [tf.reshape(pred, [-1, 4]) for pred in location_pred]
cls_pred = tf.concat(cls_pred, axis=0)
location_pred = tf.concat(location_pred, axis=0)
with tf.device('/cpu:0'):
bboxes_pred = anchor_encoder_decoder.decode_anchors(location_pred, anchors_ymin, anchors_xmin, anchors_ymax, anchors_xmax)
selected_bboxes, selected_scores = bbox_util.parse_by_class(tf.squeeze(output_shape, axis=0), cls_pred, bboxes_pred,
FLAGS.num_classes, FLAGS.select_threshold, FLAGS.min_size,
FLAGS.keep_topk, FLAGS.nms_topk, FLAGS.nms_threshold)
labels_list = []
scores_list = []
bboxes_list = []
for k, v in selected_scores.items():
labels_list.append(tf.ones_like(v, tf.int32) * k)
scores_list.append(v)
bboxes_list.append(selected_bboxes[k])
all_labels = tf.concat(labels_list, axis=0)
all_scores = tf.concat(scores_list, axis=0)
all_bboxes = tf.concat(bboxes_list, axis=0)
saver = tf.train.Saver()
with tf.Session() as sess:
init = tf.global_variables_initializer()
sess.run(init)
saver.restore(sess, get_checkpoint())
path = './images/'
image_names = sorted(os.listdir(path))
#def detecte(image_names):
for f in os.listdir(path):
np_image = imread(path + "/" + f)
labels_, scores_, bboxes_, output_shape_ = sess.run([all_labels, all_scores, all_bboxes, output_shape], feed_dict = {image_input : np_image, shape_input : np_image.shape[:-1]})
bboxes_[:, 0] = bboxes_[:, 0] * np_image.shape[0] / output_shape_[0, 0]
bboxes_[:, 1] = bboxes_[:, 1] * np_image.shape[1] / output_shape_[0, 1]
bboxes_[:, 2] = bboxes_[:, 2] * np_image.shape[0] / output_shape_[0, 0]
bboxes_[:, 3] = bboxes_[:, 3] * np_image.shape[1] / output_shape_[0, 1]
img_to_draw = draw_toolbox.bboxes_draw_on_img(np_image, labels_, scores_, bboxes_, thickness=2)
#print(bboxes_)
#imshow(img_to_draw)
cv2.imshow("SSD", img_to_draw)
k = cv2.waitKey(0) & 0xff
#Exit if ESC pressed
if k == 27 : break
# np_image = imread('./demo/test.jpg')
# labels_, scores_, bboxes_, output_shape_ = sess.run([all_labels, all_scores, all_bboxes, output_shape], feed_dict = {image_input : np_image, shape_input : np_image.shape[:-1]})
# bboxes_[:, 0] = bboxes_[:, 0] * np_image.shape[0] / output_shape_[0, 0]
# bboxes_[:, 1] = bboxes_[:, 1] * np_image.shape[1] / output_shape_[0, 1]
# bboxes_[:, 2] = bboxes_[:, 2] * np_image.shape[0] / output_shape_[0, 0]
# bboxes_[:, 3] = bboxes_[:, 3] * np_image.shape[1] / output_shape_[0, 1]
# img_to_draw = draw_toolbox.bboxes_draw_on_img(np_image, labels_, scores_, bboxes_, thickness=2)
#print(bboxes_)
# imshow(img_to_draw)
#imsave('./demo/test_out.jpg', img_to_draw)
if __name__ == '__main__':
tf.logging.set_verbosity(tf.logging.INFO)
tf.app.run()
新建脚本ssd.sh
touch ssd.sh
chmod +x ssd.sh
写入
python3 simple_ssd_demo.py \
--num_classes=2 \
运行测试
./ssd.sh
任意键下一张图片，Esc退出。
