商品检测数据集训练应用API完成商品数据集的训练
Posted ZSYL
tags:
篇首语：本文由小常识网(cha138.com)小编为大家整理，主要介绍了商品检测数据集训练应用API完成商品数据集的训练相关的知识，希望对你有一定的参考价值。
应用API完成商品数据集的训练
1. 案例训练结果
- 文件
2. 案例思路
- image_generator：获取图片数据标注数据生成器
- 标注数据分割
from utils.detection_generate import Generator
- 初始化模型参数以及冻结部分结构
- compile与fit_generator
2.1 获取Generator
- class Generator(object):
def init(self, gt, bbox_util,
batch_size, path_prefix,
train_keys, val_keys, image_size,
saturation_var=0.5,
brightness_var=0.5,
contrast_var=0.5,
lighting_std=0.5,
hflip_prob=0.5,
vflip_prob=0.5,
do_crop=True,
crop_area_range=[0.75, 1.0],
aspect_ratio_range=[3. / 4., 4. / 3.]):- gt:groundtruth结果
- bbox_util：解码工具，priors: Priors and variances, numpy tensor of shape (num_priors, 8),
priors[i] = [xmin, ymin, xmax, ymax, varxc, varyc, varw, varh]. - batch_size:批处理大小
- path_frefix: 图片的路径
- train_keys：训练图片名字列表
- val_keys：测试图片名字列表
导入工具
from utils.detection_generate import Generator
from utils.ssd_utils import BBoxUtility
from nets.ssd_net import SSD300
import numpy as np
import pickle
定义类，进行初始化网络基础参数
class SSDTrain(object):
def __init__(self, num_classes=9, input_shape=(300, 300, 3), epoch=50):
self.num_classes = num_classes
self.input_shape = input_shape
self.epoch = epoch
# prior box读取工具，初始化网络中prior boxes 配置
priors = pickle.load(open('./datasets/prior_boxes_ssd300.pkl', 'rb'))
self.bbox_util = BBoxUtility(self.num_classes, priors)
self.path_prefix = './datasets/commodity/JPEGImages/'
self.model = SSD300(self.input_shape, num_classes=self.num_classes)
def image_generator(self):
# 获取标记数据，分成训练集与测试集
gt = pickle.load(open('./datasets/commodity_gt.pkl', 'rb'))
keys = sorted(gt.keys())
num_train = int(round(0.8 * len(keys)))
train_keys = keys[:num_train]
val_keys = keys[num_train:]
# Generator获取数据
gen = Generator(gt, self.bbox_util, 16, self.path_prefix,
train_keys, val_keys,
(self.input_shape[0], self.input_shape[1]), do_crop=False)
2.2 初始化网络参数，微调网络
进行模型参数加载以及模型的结构freeze
def init_model_param(self):
"""
初始化模型参数
:return:
"""
self.model.load_weights('./ckpt/pre_trained/weights_SSD300.hdf5', by_name=True)
# 选择freeze部分结构
freeze = ['input_1', 'conv1_1', 'conv1_2', 'pool1',
'conv2_1', 'conv2_2', 'pool2',
'conv3_1', 'conv3_2', 'conv3_3', 'pool3']
for L in self.model.layers:
if L.name in freeze:
L.trainable = False
2.3 设置训练参数以及fit
- 使用adam默认算法
需要导入相关库，计算损失
from utils.ssd_losses import MultiboxLoss
MultiboxLoss的计算：多个priorbox的损失值计算
def compile(self):
"""
配置训练参数
:return:
"""
# MultiboxLoss:N个类别+1背景类别
# TensorFlow.python.keras.optimizers.Adam() 出现问题
# keras 1.2.2 optimizers.Adam() 这个版本的函数可以
optimizer = keras.optimizers.Adam()
self.model.compile(optimizer=optimizer,
loss=MultiboxLoss(self.num_classes, neg_pos_ratio=2.0).compute_loss)
def fit_generator(self, gen):
"""
训练
:param gen: 图片数据生成器
:return:
"""
# 配置回调
callbacks = [
keras.callbacks.ModelCheckpoint('./ckpt/fine_tuning/weights.{epoch:02d}-{val_acc:.2f}.hdf5',monitor='val_acc',
save_weights_only=True,
save_best_only=True,
mode='auto',
period=1),keras.callbacks.TensorBoard(log_dir='./graph', histogram_freq=1,
write_graph=True, write_images=True)]
self.model.fit_generator(gen.generate(True), gen.train_batches,
self.epoch, verbose=1,
callbacks=callbacks,
validation_data=gen.generate(False),
nb_val_samples=gen.val_batches)
3. 多GPU训练代码修改
- 在tf.keras中直接使用DistributionStrategy
def compile(self):
"""
配置训练参数
:return:
"""
distribution = tf.contrib.distribute.MirroredStrategy()
optimizer = keras.optimizers.Adam()
self.model.compile(optimizer=optimizer,
loss=MultiboxLoss(self.num_classes, neg_pos_ratio=2.0).compute_loss,
distribution=distribution)
4. 预测代码
修改源代码self.classes_name的目标个数：按照建立one_hot编码的顺序。
self.classes_name = ['clothes', 'pants', 'shoes', 'watch', 'phone',
'audio', 'computer', 'books']
修改读取训练过后的模型
from tensorflow import keras
from keras.applications.imagenet_utils import preprocess_input
from keras.preprocessing.image import load_img, img_to_array
import matplotlib.pyplot as plt
import numpy as np
from nets.ssd_net import SSD300
from utils.ssd_utils import BBoxUtility
from scipy.misc import imread
import os
class SSDTrain(object):
def __init__(self):
self.classes_name = ['Aeroplane', 'Bicycle', 'Bird', 'Boat', 'Bottle',
'Bus', 'Car', 'Cat', 'Chair', 'Cow', 'Diningtable',
'Dog', 'Horse', 'Motorbike', 'Person', 'Pottedplant',
'Sheep', 'Sofa', 'Train', 'Tvmonitor']
self.classes_nums = len(self.classes_name) + 1
self.input_shape = (300, 300, 3)
def test(self):
model = SSD300(self.input_shape, num_classes=self.classes_nums)
model.load_weights('./ckpt/weights_SSD300.hdf5', by_name=True)
# 循环读取图片进行多个图片输出检测
feature = []
images = []
for pic_name in os.listdir("./image/"):
img_path = os.path.join("./image/", pic_name)
print(img_path)
# 读取图片
# 转换成数组
# 模型输入
img = load_img(img_path, target_size=(self.input_shape[0], self.input_shape[1]))
img = img_to_array(img)
feature.append(img)
images.append(imread(img_path))
# 处理图片数据,ndarray数组输入
inputs = preprocess_input(np.array(feature))
# 预测
preds = model.predict(inputs, batch_size=1, verbose=1)
print(preds)
# 定义BBox工具
bbox_util = BBoxUtility(self.classes_nums)
# 使用非最大抑制算法过滤
results = bbox_util.detection_out(preds)
print(results[0].shape, results[1].shape)
return images, results
def tag_picture(self, images, results):
"""
对图片预测结果画图显示
:param images:
:param results:
:return:
"""
for i, img in enumerate(images):
# 解析输出结果,每张图片的标签，置信度和位置
pre_label = results[i][:, 0]
pre_conf = results[i][:, 1]
pre_xmin = results[i][:, 2]
pre_ymin = results[i][:, 3]
pre_xmax = results[i][:, 4]
pre_ymax = results[i][:, 5]
print("label:{}, probability:{}, xmin:{}, ymin:{}, xmax:{}, ymax:{}".
format(pre_label, pre_conf, pre_xmin, pre_ymin, pre_xmax, pre_ymax))
# 过滤置信度低的结果
top_indices = [i for i, conf in enumerate(pre_conf) if conf >= 0.6]
top_conf = pre_conf[top_indices]
top_label_indices = pre_label[top_indices].tolist()
top_xmin = pre_xmin[top_indices]
top_ymin = pre_ymin[top_indices]
top_xmax = pre_xmax[top_indices]
top_ymax = pre_ymax[top_indices]
# 定义21中颜色，显示图片
# currentAxis增加图中文本显示和标记显示
colors = plt.cm.hsv(np.linspace(0, 1, 21)).tolist()
plt.imshow(img / 255.)
currentAxis = plt.gca()
for i in range(top_conf.shape[0]):
xmin = int(round(top_xmin[i] * img.shape[1]))
ymin = int(round(top_ymin[i] * img.shape[0]))
xmax = int(round(top_xmax[i] * img.shape[1]))
ymax = int(round(top_ymax[i] * img.shape[0]))
# 获取该图片预测概率，名称，定义显示颜色
score = top_conf[i]
label = int(top_label_indices[i])
label_name = self.classes_name[label - 1]
display_txt = '{:0.2f}, {}'.format(score, label_name)
coords = (xmin, ymin), xmax - xmin + 1, ymax - ymin + 1
color = colors[label]
# 显示方框
currentAxis.add_patch(plt.Rectangle(*coords, fill=False, edgecolor=color, linewidth=2))
# 左上角显示概率以及名称
currentAxis.text(xmin, ymin, display_txt, bbox={'facecolor': color, 'alpha': 0.5})
plt.show()
if __name__ == '__main__':
ssd = SSDTrain()
images, results = ssd.test()
ssd.tag_picture(images, results)
5. test_ssd.py
"""SSD物体检测案例"""
from computerVision.nets.ssd_net import SSD300
from computerVision.utils.ssd_utils import BBoxUtility
from tensorflow.python.keras.preprocessing.image import load_img, img_to_array
from tensorflow.python.keras.applications.imagenet_utils import preprocess_input
import matplotlib.pyplot as plt
from imageio import imread
import numpy as np
import os
"""
- 定义好类别数量以及输出
- 模型预测流程
- SSD300模型输入以及加载参数
- 读取多个本地路径测试图片，preprocess_input以及保存图像像素值（显示需要）
- 模型预测结果，得到7308个priorbox
- 进行非最大抑制算法处理
- 图片的预测结果显示
"""
class SSDTest(object):
def __init__(self):
# 定义识别类别
self.classes_name = ['Aeroplane', 'Bicycle', 'Bird', 'Boat', 'Bottle',
'Bus', 'Car', 'Cat', 'Chair', 'Cow', 'Diningtable',
'Dog', 'Horse', 'Motorbike', 'Person', 'Pottedplant',
'Sheep', 'Sofa', 'Train', 'Tvmonitor']
# 定义模型的输入参数 + 1 背景
self.classes_nums = len(self.classes_name) + 1
self.input_shape = (300, 300, 3)
def test(self):
"""
对于输入图片进行预测物体位置
:return:
"""
# - SSD300模型输入以及加载参数
model = SSD300(self.input_shape, num_classes=self.classes_nums)
model.load_weights('./ckpt/weights_SSD300.hdf5', by_name=True) # 按照名字加载参数
# 模型特征
feature = []
images_data = []
# - 读取多个本地路径测试图片，preprocess_input以及保存图像像素值（显示需要）
for path in os.listdir('./images'):
img_path = os.path.join('./images/', path)
# 1. 输入到SSD网络中，数组
image = load_img(img_path, target_size=(self.input_shape[0], self.input_shape[1]))
image = img_to_array(image)
feature.append(image)
# 2. 读取图片二进制数据，matplotlib显示使用
images_data.append(imread(img_path))
# - 模型预测结果，得到7308个priorbox
# 图片特征处理
inputs = preprocess_input(np.asarray(feature))
# print(inputs)
predict = model.predict(inputs)
# (2, 7308, 33) 2代表图片数量，7308代表每个图片预测的default boxes数量，33： 4（位置）+ 21（预测概率） + 8（其他default boxes参数）
print(predict.shape)
# - 进行非最大抑制算法处理NMS 21类别
bb = BBoxUtility(self.classes_nums)
res = bb.detection_out(predict) # 将图片解析格式
# (200, 6) (200, 6)
print(res[0].shape, res[1].shape)
# 200个候选框，每个候选框位置，类别
return res, images_data
def tag_picture(self, images_data, outputs):
"""
显示预测结果到图片中
:return:
"""
# 1. 获取每张图片的预测结果中的值
for i, img in enumerate(images_data):
# 获取res当中对应的记过label，location，xmin，ymin，xmax，ymax
pre_label = outputs[i][:, 0]
pre_conf = outputs[i][:, 1]
pre_xmin = outputs[i][:, 2]
pre_ymin = outputs[i][:, 3]
pre_xmax = outputs[i][:, 4]
pre_ymax = outputs[i][:, 5]
# print('pre_label:{}, pre_conf:{}, pre_xmin:{}, pre_ymin:{}, pre_xmax:{}, pre_ymax:{}'.
# format(pre_label, pre_conf, pre_xmin, pre_ymin, pre_xmax, pre_ymax))
# 由于检测出的物体还是很多，所以进行显示过滤（%90）
top_indices = [i for i, conf in enumerate(pre_conf) if conf > 0.6]
top_conf = pre_conf[top_indices]
top_label_indices = pre_label[top_indices].tolist()
top_xmin = pre_xmin[top_indices]
top_ymin = pre_ymin[top_indices]
top_xmax = pre_xmax[top_indices]
top_ymax = pre_ymax[top_indices]
print('after filter top_label_indices:{}, top_conf:{}, top_xmin:{}, top_ymin:{}, top_xmax:{}, top_ymax:{}'.
format(top_label_indices, top_conf, top_xmin, top_ymin, top_xmax, top_ymax))
# matplotlib画图显示结果
# 定义21中颜色，显示图片
# currentAxis增加图中文本显示和标记显示
colors = plt.cm.hsv(np.linspace(0, 1, 21)).tolist()
plt.imshow(img / 255.)
currentAxis = plt.gca()
for i in range(top_conf.shape[0]):
xmin = int(round(top_xmin[i] * img.shape[1]))
ymin = int(round(top_ymin[i] * img.shape[0]))
xmax = int(商品检测数据集训练项目训练结构介绍
我应该使用训练数据集的函数来处理训练数据集和测试数据集的缺失值吗