Skip to content

Latest commit

 

History

History
 
 

keras

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
baseline
baseline
 
 
data
data
 
 
demo
demo
 
 
note
note
 
 
project
project
 
 
.DS_Store
.DS_Store
 
 
keras_example.ipynb
keras_example.ipynb
 
 
readme.md
readme.md
 
 

readme.md

基于keras的一些东西

Common

  • one-hot
把数值类标签转为10维向量，标签对应位为1其余为0:
y_train = keras.utils.np_utils.to_categorical(y_train, 10)

#这里是先转ckpt再转pb，实测在pb转tnn的时候有点问题，而用下面demo栏的h5转pb就正常

#>First, add after your Keras code model.fit(...) and train your model:

from keras import backend as K
import tensorflow as tf
print(model.output.op.name)
saver = tf.train.Saver()
saver.save(K.get_session(), '/tmp/keras_model.ckpt')

"""
Then cd to your TensorFlow root directory, run:

python freeze_graph.py --input_meta_graph=./tmp/model0.ckpt.meta --input_checkpoint=./tmp/model0.ckpt --output_graph=./tmp/model0.pb --output_node_names="act_6/truediv" --input_binary=true

>路径：~\anaconda\Lib\site-packages\tensorflow\python\tools
"""
  • 指定GPU
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "2"
  • 混合精度训练
os.environ['TF_ENABLE_AUTO_MIXED_PRECISION'] = '1'
opt = Adam()
opt = tf.train.experimental.enable_mixed_precision_graph_rewrite(opt)
# 要把所有模型定义用到的keras全换成tf.keras
# GeForce GTX系列貌似不支持混合精度
  • 剪枝
keras官方貌似是没有剪枝工具的，网上开源第三方实现比较有名的应该是[keras-surgeon](https://github.com/BenWhetton/keras-surgeon)
试了下不支持ReLU和DepthWiseConV
最后的可行方案是把keras换成tf.keras，然后使用tf官方的剪枝工具，可参考我的tensorflow[readme](https://github.com/fire717/Machine-Learning/tree/master/Base/frameworks/tensorflow)中的剪枝部分。
  • 计算class_weights
#ref:https://www.tensorflow.org/tutorials/structured_data/imbalanced_data

neg = 123099
pos = 222467
total = neg+pos
# Scaling by total/2 helps keep the loss to a similar magnitude.
# The sum of the weights of all examples stays the same.
weight_for_0 = (1 / neg)*(total)/2.0 
weight_for_1 = (1 / pos)*(total)/2.0

class_weight = {0: weight_for_0, 1: weight_for_1}

print('Weight for class 0: {:.2f}'.format(weight_for_0))
print('Weight for class 1: {:.2f}'.format(weight_for_1))

Note

Demo

Project

Resource