tf.train.latest_checkpoint()函数的作用查找最新保存的checkpoint文件的文件名(Finds the filename of latest saved checkpoint file.)。
tf.train.latest_checkpoint(
checkpoint_dir,
latest_filename=None
)
Args:
checkpoint_dir: Directory where the variables were saved.；
latest_filename: Directory where the variables were saved.
Returns:
The full path to the latest checkpoint or None if no checkpoint was found.
经常与模型加载函数checkpoint.restore()共同使用。
例如：
# -*- coding: utf-8 -*-
from __future__ import division, print_function, absolute_import
import tensorflow as tf
import tensorlayer as tl
import os
import numpy
import numpy as np
import librosa
from random import shuffle
import ast
print(os.getcwd())
tf.reset_default_graph()
import argparse
parser = argparse.ArgumentParser()
# quantization level
parser.add_argument('--k', type=int, default=1)
# upper bound
parser.add_argument('--B', type=int, default=2)
# learning rate
parser.add_argument('--learning_rate', type=float, default=0.01)
# resume from previous checkpoint
parser.add_argument('--resume', type=ast.literal_eval, default=False)
# training or inference
parser.add_argument('--mode', type=str, default='training')
args = parser.parse_args()
print(args.k, args.B, args.learning_rate, args.resume, args.mode)
class CNNConfig(object):
"""CNN配置参数"""
#结构体
#learning_rate = 1e-2 # 学习率
learning_rate = args.learning_rate # 学习率
num_epochs = 1000 # 总迭代轮次
batch_size = 200
print_per_batch = 20
save_tb_per_batch = 10
print_freq = 10
k = args.k
B = args.B
class ASRCNN(object):
def __init__(self, input_x, input_y, config, width, height, num_classes, is_train=True, reuse=False): # 20,80
self.config = config
# input_x = tf.reshape(self.input_x, [-1, height, width])
input_x = tf.transpose(input_x, [0, 2, 1])
self.input_x = tf.reshape(input_x, [-1, height, width, 1])
self.input_y = input_y
with tf.variable_scope("binarynet", reuse=reuse):
net_in = tl.layers.InputLayer(self.input_x, name='input')
net0 = tl.layers.Conv2d(net_in, 8, (10, 3), (10, 3), padding='SAME', b_init=None, name='bcnn0')
#net = tl.layers.MaxPool2d(net, (2, 2), (2, 2), padding='SAME', name='pool0')
net0 = tl.layers.BatchNormLayer(net0, act=tf.nn.relu, is_train=is_train, name='bn0')
net0 = tl.layers.Quant_Layer(net0, config.k, config.B)
#需要考虑下是否使用input_frame的tick还是层间的tick_relative
#考虑一下叠加的可行性
net1 = tl.layers.Quant_Conv2d(net0, 16, (3, 3), (1, 1), padding='SAME', b_init=None, name='bcnn1')
#net = tl.layers.MaxPool2d(net, (2, 2), (2, 2), padding='SAME', name='pool1')
net1 = tl.layers.BatchNormLayer(net1, act=tf.nn.relu, is_train=is_train, name='bn1')
net1 = tl.layers.Quant_Layer(net1, config.k, config.B)
#这下子成了标准shorucut，在芯片上的shortcut是elementwisr形式的add，可以灵活处理，充分利用，规则化，简单化，尽管输入输出可能会有些没有充分利用
#可以考虑一下均衡分配
#考虑一下非方形，小型5*5或者其他mapping算子
#通盘考虑，采用特定步长或者填充方式的价值，带来的卷积核的实际意义
#调节系数，匹配运算能力，提升精度，不规则卷积带来的挑战
shortcut0 = tl.layers.Quant_Conv2d(net0, 64, (2, 2), (2, 2), padding='SAME', b_init=None, name='shortcut0')
#net = tl.layers.MaxPool2d(net, (2, 2), (2, 2), padding='SAME', name='pool1')
#shortcut0 = tl.layers.BatchNormLayer(shortcut0, act=tf.nn.relu, is_train=is_train, name='bn_shortcut0')
net2 = tl.layers.Quant_Conv2d(net1, 64, (2, 2), (2, 2), padding='VALID', b_init=None, name='bcnn2')
#net = tl.layers.MaxPool2d(net, (2, 2), (2, 2), padding='SAME', name='pool2')
#net2 = tl.layers.BatchNormLayer(net2, act=tf.nn.relu, is_train=is_train, name='bn2')
#net2 = tl.layers.Quant_Layer(net2, config.k, config.B)
#方便了core上叠加，方便了中值量化
shortcut0 = tl.layers.ElementwiseLayer([shortcut0, net2], combine_fn=tf.add, act=None, name='elementwise0')
shortcut0 = tl.layers.BatchNormLayer(shortcut0, act=tf.nn.relu, is_train=is_train, name='bn_shortcut0')
shortcut0 = tl.layers.Quant_Layer(shortcut0, config.k, config.B)
net3= tl.layers.Quant_Conv2d(shortcut0, 128, (3, 3), (1, 1), padding='VALID', b_init=None, name='bcnn3')
#net = tl.layers.MaxPool2d(net, (2, 2), (2, 2), padding='SAME', name='pool3')
net3 = tl.layers.BatchNormLayer(net3, act=tf.nn.relu, is_train=is_train, name='bn3')
net3 = tl.layers.Quant_Layer(net3, config.k, config.B)
net3 = tl.layers.FlattenLayer(net3)
# net = tl.layers.DropoutLayer(net, 0.8, True, is_train, name='drop0')
net4 = tl.layers.DenseLayer(net3, n_units=num_classes, b_init=None, name='dense')
net4 = tl.layers.BatchNormLayer(net4, act=tf.identity, is_train=is_train, name='bn4')
#这个激活函数很重要
# 分类器
self.logits = net4.outputs
self.y_pred_cls = tf.argmax(tf.nn.softmax(self.logits), 1) # 预测类别
# 损失函数，交叉熵
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=self.logits, labels=self.input_y)
self.loss = tf.reduce_mean(cross_entropy)
# 优化器
self.optim = tf.train.AdamOptimizer(learning_rate=self.config.learning_rate).minimize(self.loss)
# 准确率
correct_pred = tf.equal(tf.argmax(self.input_y, 1), self.y_pred_cls)
self.acc = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
self.net = net4
def dense_to_one_hot(labels_dense, num_classes=10):
"""Convert class labels from scalars to one-hot vectors."""
return numpy.eye(num_classes)[labels_dense]
def read_files(files):
labels = []
features = []
for ans, files in files.items():
for file in files:
wave, sr = librosa.load(file, mono=True)
label = dense_to_one_hot(ans, 10)
# label = [float(value) for value in label]
labels.append(label)
mfcc = librosa.feature.mfcc(wave, sr, n_mfcc=24)
l = len(mfcc)
# print(np.array(mfcc).shape)
mfcc = np.pad(mfcc, ((0, 0), (0, 80 - len(mfcc[0]))), mode='constant', constant_values=0)
features.append(np.array(mfcc))
# print('reading '+file)
return np.array(features), np.array(labels)
def load_files(path='data/spoken_numbers_pcm/'):
files = os.listdir(path)
cls_files = {}
for wav in files:
if not wav.endswith(".wav"): continue
ans = int(wav[0])
cls_files.setdefault(ans, [])
cls_files[ans].append(path + wav)
train_files = {}
valid_files = {}
test_files = {}
for ans, file_list in cls_files.items():
shuffle(file_list)
all_len = len(file_list)
train_len = int(all_len * 0.7)
valid_len = int(all_len * 0.2)
test_len = all_len - train_len - valid_len
train_files[ans] = file_list[0:train_len]
valid_files[ans] = file_list[train_len:train_len + valid_len]
test_files[ans] = file_list[all_len - test_len:all_len]
return train_files, valid_files, test_files
def batch_iter(X, Y, batch_size=128):
data_len = len(X)
num_batch = int((data_len - 1) / batch_size) + 1
indices = np.random.permutation(np.arange(data_len))
x_shuffle = X[indices]
y_shuffle = Y[indices]
for i in range(num_batch):
start_id = i * batch_size
end_id = min((i + 1) * batch_size, data_len)
yield x_shuffle[start_id:end_id], y_shuffle[start_id:end_id]
def feed_data(input_x, input_y, x_batch, y_batch):
feed_dict = {
input_x: x_batch,
input_y: y_batch,
}
return feed_dict
def mean_normalize(features):
std_value = features.std()
mean_value = features.mean()
return (features - mean_value) / std_value
def train(args=None):
'''batch = mfcc_batch_generator()
X, Y = next(batch)
trainX, trainY = X, Y
testX, testY = X, Y # overfit for now'''
train_files, valid_files, test_files = load_files()
train_features, train_labels = read_files(train_files)
train_features = mean_normalize(train_features)
print('read train files down')
valid_features, valid_labels = read_files(valid_files)
valid_features = mean_normalize(valid_features)
print('read valid files down')
test_features, test_labels = read_files(test_files)
test_features = mean_normalize(test_features)
print('read test files down')
width = 24 # mfcc features
height = 80 # (max) length of utterance
classes = 10 # digits
config = CNNConfig
input_x = tf.placeholder(tf.float32, [None, width, height], name='input_x')
input_y = tf.placeholder(tf.float32, [None, classes], name='input_y')
cnn_train = ASRCNN(input_x, input_y, config, width, height, classes, is_train=True, reuse=False)
cnn_test = ASRCNN(input_x, input_y, config, width, height, classes, is_train=False, reuse=True)
#session = tf.Session()
session = tf.InteractiveSession()
session.run(tf.global_variables_initializer())
cnn_train.net.print_params()
cnn_train.net.print_layers()
saver = tf.train.Saver(tf.global_variables())
checkpoint_path = os.path.join('cnn_model', 'model.ckpt')
tensorboard_train_dir = 'tensorboard/train'
tensorboard_valid_dir = 'tensorboard/valid'
if not os.path.exists(tensorboard_train_dir):
os.makedirs(tensorboard_train_dir)
if not os.path.exists(tensorboard_valid_dir):
os.makedirs(tensorboard_valid_dir)
tf.summary.scalar("loss", cnn_train.loss)
tf.summary.scalar("accuracy", cnn_train.acc)
merged_summary = tf.summary.merge_all()
train_writer = tf.summary.FileWriter(tensorboard_train_dir)
valid_writer = tf.summary.FileWriter(tensorboard_valid_dir)
model_file_name = tf.train.latest_checkpoint('./cnn_model/')
#globe, global_step可能是个问题
if args.resume:
print("Load existing model " + "!" * 10)
saver.restore(session, model_file_name)
if args.mode == 'training':
total_batch = 0
for epoch in range(config.num_epochs):
# print('Epoch:', epoch + 1)
batch_train = batch_iter(train_features, train_labels)
for x_batch, y_batch in batch_train:
total_batch += 1
#神奇的feed_dict
feed_dict = feed_data(input_x, input_y, x_batch, y_batch)
session.run(cnn_train.optim, feed_dict=feed_dict)
if total_batch % config.print_per_batch == 0:
train_loss, train_accuracy = session.run([cnn_train.loss, cnn_train.acc], feed_dict=feed_dict)
valid_loss, valid_accuracy = session.run([cnn_test.loss, cnn_test.acc], feed_dict={input_x: valid_features,
input_y: valid_labels})
print('Steps:' + str(total_batch))
print(
'train_loss:' + str(train_loss) + ' train accuracy:' + str(train_accuracy) + '\tvalid_loss:' + str(
valid_loss) + ' valid accuracy:' + str(valid_accuracy))
if total_batch % config.save_tb_per_batch == 0:
train_s = session.run(merged_summary, feed_dict=feed_dict)
train_writer.add_summary(train_s, total_batch)
valid_s = session.run(merged_summary, feed_dict={input_x: valid_features, input_y: valid_labels})
valid_writer.add_summary(valid_s, total_batch)
saver.save(session, checkpoint_path, global_step=epoch)
if (epoch + 1) % (config.print_freq) == 0:
print("Save npz model " + "!" * 10)
#saver = tf.train.Saver()
#save_path = saver.save(sess, model_file_name)
# you can also save model into npz
tl.files.save_npz(cnn_train.net.all_params, name='model_kws.npz', sess=session)
# and restore it as follow:
# tl.files.load_and_assign_npz(sess=sess, name='model.npz', network=network)
test_loss, test_accuracy = session.run([cnn_test.loss, cnn_test.acc],
feed_dict={input_x: test_features, input_y: test_labels})
print('test_loss:' + str(test_loss) + ' test accuracy:' + str(test_accuracy))
if __name__ == '__main__':
train(args)
# test('data/spoken_numbers_pcm/9_Alex_260.wav')
更多参考
https://www.jianshu.com/p/5006be1c5f59
https://blog.csdn.net/weixin_38145317/article/details/92815106
https://blog.csdn.net/admin_maxin/article/details/89393399
https://www.w3cschool.cn/doc_tensorflow_python/tensorflow_python-tf-train-latest_checkpoint.html
https://www.bookstack.cn/read/TensorFlow2.0/spilt.1.3b87bc87b85cbe5d.md
3340
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
