由于本人是刚接触python
然后这是我在gitee上面找的一篇有关python 布匹瑕疵检测 的源码想拿到自己电脑上复现。虚拟环境已经配置好了（python3.6）
这是main.py
# coding=utf-8
#################################################
# finetune
#################################################
import pandas as pd
import numpy as np
import os
import random
import cv2
import datetime
from tqdm import tqdm
from keras.callbacks import ModelCheckpoint, ReduceLROnPlateau, EarlyStopping
from keras.preprocessing.image import ImageDataGenerator, img_to_array, load_img
from keras.models import Model
from keras.layers import Dense
from keras.layers import Dropout
from keras.layers.pooling import GlobalAveragePooling2D
from keras.optimizers import Adam, SGD
from keras.applications.densenet import DenseNet169, DenseNet121
from keras.applications.densenet import preprocess_input
from focal_loss import focal_loss
from attetion import add_attention
#######################################
# 在训练的时候置为1
from keras import backend as K
K.set_learning_phase(1)
#######################################
EPOCHS = 60
RANDOM_STATE = 2018
learning_rate = 0.001 # 0.001
TRAIN_DIR = '../data/data_for_train'
VALID_DIR = '../data/data_for_valid'
# 获得每一类的图片张数，索引顺序与keras中的读取顺序对应
def get_cls_num(directory):
'''
'''
classes = list()
for subdir in sorted(os.listdir(directory)):
if os.path.isdir(os.path.join(directory, subdir)):
classes.append(subdir)
classes_num = list()
for cls_ in classes:
path = os.path.join(directory, cls_)
pics = os.listdir(path)
classes_num.append(len(pics))
return classes_num
def get_classes_indice(directory):
'''
得到索引与label对应的dict
'''
classes = []
for subdir in sorted(os.listdir(directory)):
if os.path.isdir(os.path.join(directory, subdir)):
classes.append(subdir)
num_classes = len(classes)
class_indices = dict(zip(classes, range(len(classes))))
return class_indices
def get_callbacks(filepath, patience=2):
lr_reduce = ReduceLROnPlateau(monitor='val_acc', factor=0.1, epsilon=5e-6, patience=patience, verbose=1,
min_lr=0.00001)
msave = ModelCheckpoint(filepath, monitor='val_acc', save_best_only=True) # 该回调函数将在每个epoch后保存模型到filepath
earlystop = EarlyStopping(monitor='val_acc', min_delta=0, patience=patience * 3 + 4, verbose=1, mode='auto')
return [lr_reduce, msave, earlystop]
def add_new_last_layer(base_model, nb_classes, drop_rate=0.):
"""Add last layer to the convnet
Args:
base_model: keras model excluding top
nb_classes: # of classes
Returns:
new keras model with last layer
"""
bn_axis = 3 if K.image_data_format() == 'channels_last' else 1
x = base_model.output
# x = BatchNormalization(axis=bn_axis, epsilon=1.001e-5,name='bn_final')(x)
x = Dropout(0.5)(x)
x = GlobalAveragePooling2D()(x)
predictions = Dense(nb_classes, activation='softmax')(x) # new softmax layer
model = Model(input=base_model.input, output=predictions)
return model
# LSR(label smoothing regularization)
def mycrossentropy(e=0.25, nb_classes=11):
'''
https://spaces.ac.cn/archives/4493
'''
def mycrossentropy_fixed(y_true, y_pred):
return (1 - e) * K.categorical_crossentropy(y_true, y_pred) + e * K.categorical_crossentropy(
K.ones_like(y_pred) / nb_classes, y_pred)
return mycrossentropy_fixed
# LSR(label smoothing regularization) + focal_loss
def myloss(classes_num, e=0.25, nb_classes=11):
def myloss_fixed(y_true, y_pred):
import tensorflow as tf
from tensorflow.python.ops import array_ops
from keras import backend as K
zeros = array_ops.zeros_like(y_pred, dtype=y_pred.dtype)
crossentropy_loss = -1 * y_true * tf.log(tf.clip_by_value(y_pred, 1e-8, 1.0))
classes_weight = array_ops.zeros_like(y_true, dtype=y_true.dtype)
total_num = float(sum(classes_num))
classes_w_t1 = [total_num / ff for ff in classes_num]
sum_ = sum(classes_w_t1)
classes_w_t2 = [ff / sum_ for ff in classes_w_t1] # scale
classes_w_tensor = tf.convert_to_tensor(classes_w_t2, dtype=y_true.dtype)
classes_weight += classes_w_tensor
alpha = array_ops.where(y_true > zeros, classes_weight, zeros)
balanced_fl = alpha * crossentropy_loss
balanced_fl = tf.reduce_sum(balanced_fl)
return (1 - e) * balanced_fl + e * K.categorical_crossentropy(K.ones_like(y_pred) / nb_classes, y_pred)
return myloss_fixed
def get_model(INT_HEIGHT, IN_WIDTH):
'''
获得模型
'''
base_model = DenseNet169(include_top=False, weights='imagenet', input_shape=(INT_HEIGHT, IN_WIDTH, 3))
# base_model = DenseNet169(include_top=False, weights='../save_model/densenet169_weights_tf_dim_ordering_tf_kernels_notop.h5', input_shape=(INT_HEIGHT,IN_WIDTH, 3))
# add attention
# base_model = add_attention(base_model)
model = add_new_last_layer(base_model, 11)
model.compile(optimizer=SGD(lr=learning_rate, momentum=0.9), loss=[mycrossentropy()], metrics=['accuracy'])
# classes_num = get_cls_num('../data/data_for_train')
# model.compile(optimizer=SGD(lr=learning_rate, momentum=0.9), loss=[myloss(classes_num)], metrics=['accuracy'])
model.summary()
return model
def train_model(save_model_path, BATCH_SIZE, IN_SIZE):
INT_HEIGHT = IN_SIZE[0]
IN_WIDTH = IN_SIZE[1]
callbacks = get_callbacks(filepath=save_model_path, patience=3)
model = get_model(INT_HEIGHT, IN_WIDTH)
train_datagen = ImageDataGenerator(
preprocessing_function=preprocess_input,
horizontal_flip=True,
vertical_flip=True
)
valid_datagen = ImageDataGenerator(
preprocessing_function=preprocess_input
)
train_generator = train_datagen.flow_from_directory(
directory=TRAIN_DIR,
target_size=(INT_HEIGHT, IN_WIDTH),
batch_size=BATCH_SIZE,
class_mode='categorical',
seed=RANDOM_STATE,
interpolation='antialias', # PIL默认插值下采样的时候会模糊
)
valid_generator = valid_datagen.flow_from_directory(
directory=VALID_DIR,
target_size=(INT_HEIGHT, IN_WIDTH),
batch_size=BATCH_SIZE,
class_mode='categorical',
seed=RANDOM_STATE,
interpolation='antialias', # PIL默认插值下采样的时候会模糊
)
model.fit_generator(
train_generator,
steps_per_epoch=2 * (train_generator.samples // BATCH_SIZE + 1),
epochs=EPOCHS,
max_queue_size=100,
workers=1,
verbose=1,
validation_data=valid_generator, # valid_generator,
validation_steps=valid_generator.samples // BATCH_SIZE,
# valid_generator.samples // BATCH_SIZE + 1, #len(valid_datagen)+1,
callbacks=callbacks
)
def predict(weights_path, IN_SIZE):
'''
对测试数据进行预测
'''
INT_HEIGHT = IN_SIZE[0]
IN_WIDTH = IN_SIZE[1]
K.set_learning_phase(0)
test_pic_root_path = '../data'
filename = []
probability = []
# 1#得到模型
model = get_model(INT_HEIGHT, IN_WIDTH)
model.load_weights(weights_path)
# 2#预测
for parent, _, files in os.walk(test_pic_root_path):
for line in tqdm(files):
pic_path = os.path.join(test_pic_root_path, line.strip())
img = load_img(pic_path, target_size=(INT_HEIGHT, IN_WIDTH, 3), interpolation='antialias')
img = img_to_array(img)
img = preprocess_input(img)
img = np.expand_dims(img, axis=0)
prediction = model.predict(img)[0]
prediction = list(prediction)
classes_dict = get_classes_indice(TRAIN_DIR)
classes_dict = {classes_dict[key]: key for key in classes_dict.keys()}
for index, pro in enumerate(prediction):
# 获得对应label名称
label = classes_dict[index]
pro = round(pro, 5)
if pro == 0:
pro = 0.00001
if pro == 1:
pro = 0.99999
# 存入list
filename.append(line.strip() + '|' + label)
probability.append(' ' + str(pro))
# 3#写入csv
res_path = "../submit/submit_" + datetime.datetime.now().strftime('%Y%m%d_%H%M%S') + ".csv"
dataframe = pd.DataFrame({'filename|defect': filename, 'probability': probability})
dataframe.to_csv(res_path, index=False, header=True)
def main():
# ～～～～～～～～～～～～～～～～～ 复现线上最优结果 模型预测 ～～～～～～～～～～～～～～～～～
batch_size = 4
in_size = [550, 600] # [H,W]
weights_path1 = '../models/final/model_weight_0.7296.hdf5' # dense169, [550,600] submit_1.csv
weights_path2 = '../models/final/model_weight_0.7239.hdf5' # dense169, [550,600] submit_2.csv
## 预测得到线上746的结果
predict(weights_path=weights_path1, IN_SIZE=in_size)
predict(weights_path=weights_path2, IN_SIZE=in_size)
# ～～～～～～～～～～～～～～～～～ 训练及预测全过程 ～～～～～～～～～～～～～～～～～
# weights_path = '../models/model_weight.hdf5'
# train_model(save_model_path=weights_path, BATCH_SIZE=batch_size, IN_SIZE=in_size)
# predict(weights_path=weights_path, IN_SIZE=in_size)
if __name__ == '__main__':
main()
然后会出现以下结果：
由于本人是刚接触python不知道问题的原因，求助大佬！！
