華為雲AI-深度學習糖尿病預測
阿新 • • 發佈：2018-12-04
num eval 鍵值 ges validate amp () slim step
#!/usr/bin/env python2
# -*- coding: utf-8 -*-
"""
Created on Sat Sep 15 10:54:53 2018
@author: myhaspl
@email:[email protected]
糖尿病預測（多層）
csv格式：懷孕次數、葡萄糖、血壓、皮膚厚度，胰島素,bmi,糖尿病血統函數,年齡,結果
"""
import tensorflow as tf
import os
trainCount=10000
inputNodeCount=8
validateCount=50
sampleCount=200
testCount=10
outputNodeCount=1
g=tf.Graph()
with g.as_default():
def getWeights(shape,wname):
weights=tf.Variable(tf.truncated_normal(shape,stddev=0.1),name=wname)
return weights
def getBias(shape,bname):
biases=tf.Variable(tf.constant(0.1,shape=shape),name=bname)
return biases
def inferenceInput(x):
layer1=tf.nn.relu(tf.add(tf.matmul(x,w1),b1))
result=tf.add(tf.matmul(layer1,w2),b2)
return result
def inference(x):
yp=inferenceInput(x)
return tf.sigmoid(yp)
def loss():
yp=inferenceInput(x)
return tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels=y,logits=yp))
def train(learningRate,trainLoss,trainStep):
trainOp=tf.train.AdamOptimizer(learningRate).minimize(trainLoss,global_step=trainStep)
return trainOp
def evaluate(x):
return tf.cast(inference(x)>0.5,tf.float32)
def accuracy(x,y,count):
yp=evaluate(x)
return tf.reduce_mean(tf.cast(tf.equal(yp,y),tf.float32))
def inputFromFile(fileName,skipLines=1):
#生成文件名隊列
fileNameQueue=tf.train.string_input_producer([fileName])
#生成記錄鍵值對
reader=tf.TextLineReader(skip_header_lines=skipLines)
key,value=reader.read(fileNameQueue)
return value
def getTestData(fileName,skipLines=1,n=10):
#生成文件名隊列
testFileNameQueue=tf.train.string_input_producer([fileName])
#生成記錄鍵值對
testReader=tf.TextLineReader(skip_header_lines=skipLines)
testKey,testValue=testReader.read(testFileNameQueue)
testRecordDefaults=[[1.],[1.],[1.],[1.],[1.],[1.],[1.],[1.],[1.]]
testDecoded=tf.decode_csv(testValue,record_defaults=testRecordDefaults)
pregnancies,glucose,bloodPressure,skinThickness,insulin,bmi,diabetespedigreefunction,age,outcome=tf.train.shuffle_batch(testDecoded,batch_size=n,capacity=1000,min_after_dequeue=1)
testFeatures=tf.transpose(tf.stack([pregnancies,glucose,bloodPressure,skinThickness,insulin,bmi,diabetespedigreefunction,age]))
testY=tf.transpose([outcome])
return (testFeatures,testY)
def getNextBatch(n,values):
recordDefaults=[[1.],[1.],[1.],[1.],[1.],[1.],[1.],[1.],[1.]]
decoded=tf.decode_csv(values,record_defaults=recordDefaults)
pregnancies,glucose,bloodPressure,skinThickness,insulin,bmi,diabetespedigreefunction,age,outcome=tf.train.shuffle_batch(decoded,batch_size=n,capacity=1000,min_after_dequeue=1)
features=tf.transpose(tf.stack([pregnancies,glucose,bloodPressure,skinThickness,insulin,bmi,diabetespedigreefunction,age]))
y=tf.transpose([outcome])
return (features,y)
with tf.name_scope("inputSample"):
samples=inputFromFile("s3://myhaspl/tf_learn/diabetes.csv",1)
inputDs=getNextBatch(sampleCount,samples)
with tf.name_scope("validateSamples"):
validateInputs=getNextBatch(validateCount,samples)
with tf.name_scope("testSamples"):
testInputs=getTestData("s3://myhaspl/tf_learn/diabetes_test.csv")
with tf.name_scope("inputDatas"):
x=tf.placeholder(dtype=tf.float32,shape=[None,inputNodeCount],name="input_x")
y=tf.placeholder(dtype=tf.float32,shape=[None,outputNodeCount],name="input_y")
with tf.name_scope("Variable"):
w1=getWeights([inputNodeCount,12],"w1")
b1=getBias((),"b1")
w2=getWeights([12,outputNodeCount],"w2")
b2=getBias((),"b2")
trainStep=tf.Variable(0,dtype=tf.int32,name="tcount",trainable=False)
with tf.name_scope("train"):
trainLoss=loss()
trainOp=train(0.005,trainLoss,trainStep)
init=tf.global_variables_initializer()
with tf.Session(graph=g) as sess:
sess.run(init)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
while trainStep.eval()<trainCount:
sampleX,sampleY=sess.run(inputDs)
sess.run(trainOp,feed_dict={x:sampleX,y:sampleY})
nowStep=sess.run(trainStep)
if nowStep%500==0:
validate_acc=sess.run(accuracy(sampleX,sampleY,sampleCount))
print "%d次後=>正確率%g"%(nowStep,validate_acc)
if nowStep>trainCount:
break
testInputX,testInputY=sess.run(testInputs)
print "測試樣本正確率%g"%sess.run(accuracy(testInputX,testInputY,testCount))
print testInputX,testInputY
print sess.run(evaluate(testInputX))
coord.request_stop()
coord.join(threads)
500次後=>正確率0.67 1000次後=>正確率0.75 1500次後=>正確率0.81 2000次後=>正確率0.75 2500次後=>正確率0.775 3000次後=>正確率0.765 3500次後=>正確率0.84 4000次後=>正確率0.85 4500次後=>正確率0.77 5000次後=>正確率0.78 5500次後=>正確率0.775 6000次後=>正確率0.835 6500次後=>正確率0.84 7000次後=>正確率0.785 7500次後=>正確率0.805 8000次後=>正確率0.765 8500次後=>正確率0.83 9000次後=>正確率0.835 9500次後=>正確率0.78 10000次後=>正確率0.775 測試樣本正確率0.7 [[1.00e+01 1.01e+02 7.60e+01 4.80e+01 1.80e+02 3.29e+01 1.71e-01 6.30e+01] [3.00e+00 7.80e+01 5.00e+01 3.20e+01 8.80e+01 3.10e+01 2.48e-01 2.60e+01] [2.00e+00 1.22e+02 7.00e+01 2.70e+01 0.00e+00 3.68e+01 3.40e-01 2.70e+01] [2.00e+00 8.80e+01 5.80e+01 2.60e+01 1.60e+01 2.84e+01 7.66e-01 2.20e+01] [1.00e+01 1.01e+02 7.60e+01 4.80e+01 1.80e+02 3.29e+01 1.71e-01 6.30e+01] [2.00e+00 1.22e+02 7.00e+01 2.70e+01 0.00e+00 3.68e+01 3.40e-01 2.70e+01] [1.00e+00 8.90e+01 6.60e+01 2.30e+01 9.40e+01 2.81e+01 1.67e-01 2.10e+01] [6.00e+00 1.48e+02 7.20e+01 3.50e+01 0.00e+00 3.36e+01 6.27e-01 5.00e+01] [1.00e+00 9.30e+01 7.00e+01 3.10e+01 0.00e+00 3.04e+01 3.15e-01 2.30e+01] [2.00e+00 1.22e+02 7.00e+01 2.70e+01 0.00e+00 3.68e+01 3.40e-01 2.70e+01]] [[0.] [1.] [0.] [0.] [0.] [0.] [0.] [1.] [0.] [0.]] [[1.] [0.] [0.] [0.] [1.] [0.] [0.] [1.] [0.] [0.]]
感覺華為雲中提供的深度學習服務，就是給你提供一個強大的服務器，然後，你自己編寫代碼。可能還提供了一些更多的功能
另外，提供了一個訓練用戶自定義數據的代碼
補充一個概念：
MoXing是華為雲深度學習服務提供的網絡模型開發API。相對於TensorFlow和MXNet等原生API而言，MoXing API讓模型的代碼編寫更加簡單，而且能夠自動獲取高性能的分布式執行能力。
MoXing允許用戶只需要關心數據輸入(input_fn)和模型構建(model_fn)的代碼，就可以實現任意模型在多GPU和分布式下的高性能運行。MoXing-TensorFlow支持原生TensorFlow、Keras、slim等API，幫助構建圖像分類、物體檢測、生成對抗、自然語言處理和OCR等多種模型。
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import tensorflow as tf
import moxing.tensorflow as mox
slim = tf.contrib.slim
# 用TensorFlow原生的方式定義超參
tf.flags.DEFINE_string(‘data_url‘, None, ‘‘)
tf.flags.DEFINE_string(‘train_dir‘, None, ‘‘)
flags = tf.flags.FLAGS
def train_my_model():
def input_fn(run_mode, **kwargs):
# 從TFRecord中獲取輸入數據集
keys_to_features = {
‘image/encoded‘: tf.FixedLenFeature((), tf.string, default_value=‘‘),
‘image/format‘: tf.FixedLenFeature((), tf.string, default_value=‘raw‘),
‘image/class/label‘: tf.FixedLenFeature(
[1], tf.int64, default_value=tf.zeros([1], dtype=tf.int64)),
}
items_to_handlers = {
‘image‘: slim.tfexample_decoder.Image(shape=[28, 28, 1], channels=1),
‘label‘: slim.tfexample_decoder.Tensor(‘image/class/label‘, shape=[]),
}
# 數據集中包含60000張訓練集圖像（數據文件名為mnist_train.tfrecord）
# 以及10000張驗證集圖像（數據文件名為mnist_test.tfrecord）
dataset = mox.get_tfrecord(dataset_dir=flags.data_url,
file_pattern=‘mnist_train.tfrecord‘ if run_mode == mox.ModeKeys.TRAIN else ‘mnist_test.tfrecord‘,
num_samples=60000 if run_mode == mox.ModeKeys.TRAIN else 10000,
keys_to_features=keys_to_features,
items_to_handlers=items_to_handlers,
capacity=1000)
image, label = dataset.get([‘image‘, ‘label‘])
# 將圖像像素值轉換為float並統一大小
image = tf.to_float(image)
image = tf.image.resize_image_with_crop_or_pad(image, 28, 28)
return image, label
def model_fn(inputs, run_mode, **kwargs):
# 獲取一批輸入數據
images, labels = inputs
# 將輸入圖像進行歸一化
images = tf.subtract(images, 128.0)
images = tf.div(images, 128.0)
# 定義函數參數作用域：
# 1. 所有的卷積和全鏈接L2正則項系數為0
# 2. 所有的卷積和全鏈接使用截斷正態分布初始化待訓練變量
# 3. 所有的卷積和全鏈接的激活層采用ReLU
with slim.arg_scope(
[slim.conv2d, slim.fully_connected],
weights_regularizer=slim.l2_regularizer(scale=0.0),
weights_initializer=tf.truncated_normal_initializer(stddev=0.1),
activation_fn=tf.nn.relu):
# 定義網絡
net = slim.conv2d(images, 32, [5, 5])
net = slim.max_pool2d(net, [2, 2], 2)
net = slim.conv2d(net, 64, [5, 5])
net = slim.max_pool2d(net, [2, 2], 2)
net = slim.flatten(net)
net = slim.fully_connected(net, 1024)
net = slim.dropout(net, 0.5, is_training=True)
logits = slim.fully_connected(net, 10, activation_fn=None)
labels_one_hot = slim.one_hot_encoding(labels, 10)
# 定義交叉熵損失值
loss = tf.losses.softmax_cross_entropy(
logits=logits, onehot_labels=labels_one_hot,
label_smoothing=0.0, weights=1.0)
# 由於函數參數作用域定義了所有L2正則項系數為0，所以這裏將不會獲取到任何L2正則項
regularization_losses = mox.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
if len(regularization_losses) > 0:
regularization_loss = tf.add_n(regularization_losses)
loss += regularization_loss
# 定義評價指標
accuracy_top_1 = tf.reduce_mean(tf.cast(tf.nn.in_top_k(logits, labels, 1), tf.float32))
accuracy_top_5 = tf.reduce_mean(tf.cast(tf.nn.in_top_k(logits, labels, 5), tf.float32))
# 必須返回mox.ModelSpec
return mox.ModelSpec(loss=loss,
log_info={‘loss‘: loss, ‘top1‘: accuracy_top_1, ‘top5‘: accuracy_top_5})
# 獲取一個內置的Optimizer
optimizer_fn = mox.get_optimizer_fn(‘sgd‘, learning_rate=0.01)
# 啟動訓練
mox.run(input_fn=input_fn,
model_fn=model_fn,
optimizer_fn=optimizer_fn,
run_mode=mox.ModeKeys.TRAIN,
batch_size=50,
log_dir=flags.train_dir,
max_number_of_steps=2000,
log_every_n_steps=10,
save_summary_steps=50,
save_model_secs=60)
if __name__ == ‘__main__‘:
train_my_model()華為雲AI-深度學習糖尿病預測