1 實現場景分析
業務在外呼中經常會遇到接聽者因忙或者空號導致返回的回鈴音被語音識別引擎識別並傳遞給業務流程解析，而這種情況會在外呼后的業務統計中導致接通率的統計較低，為了解決該問題，打算在回鈴音進入語音識別引擎前進行識別，判斷為非接通的則直接丟棄不在接入流程處理。
經過對場景中的錄音音頻及語音識別的文字進行分析，發現大部分的誤識別回鈴音都是客戶忙或者是空號，與正常接通的音頻特征區分很明顯，如下所示采用科大訊飛的語音識別引擎對失敗的回鈴音轉寫的結果
從轉寫結果統計也驗證了我們的分析。（針對回鈴音為視頻彩鈴等的暫時沒有統計到，這里也不作為主要的失敗音頻分析）
2 模型訓練實現基於深度學習的聲音分類
實際實踐參考 keras實現聲音二分類 文章中有對音頻特征mfcc的說明，流程分析也很詳細，可以參考。這里主要貼下在驗證中使用的代碼，模型訓練代碼
import os
import keras
import librosa
import numpy as np
import matplotlib.pyplot as plt
from keras import Sequential
from keras.utils import to_categorical
from keras.layers import Dense
from sklearn.model_selection import train_test_split
import tensorflow as tf
from keras import backend as K
DATA = 'data.npy'
TARGET = 'target.npy'
def load_label(label_path):
"""
遍歷當前給定的目錄，便於后文進行標簽加載
:param label_path:
:return:
"""
label = os.listdir(label_path)
return label
# 提取 mfcc 參數
def wav2mfcc(path, max_pad_size=11):
"""
備注：由於我們拿到的音頻文件，持續時間都不盡相同，所以提取到的 mfcc 大小是不相同的。
但是神經網絡要求待處理的矩陣大小要相同，所以這里我們用到了鋪平操作。我們 mfcc 系數默認提取 20 幀，對於每一幀來說，
如果幀長小於 11，我們就用 0 填滿不滿足要求的幀；如果幀長大於 11，我們就只選取前 11 個參數
:param path: 音頻文件地址
:param max_pad_size: 幀長，最大設置為11
:return:
"""
# 讀取音頻文件，按照音頻本身的采樣率進行讀取
y, sr = librosa.load(path=path, sr=None, mono=True)
y = y[::3] # 不需要太高的采樣率數據，這里進行每三個點選用一個
audio_mac = librosa.feature.mfcc(y=y, sr=16000)
y_shape = audio_mac.shape[1]
if y_shape < max_pad_size:
"""
函數numpy.pad(array, pad_width, mode)，其中 array 是我們需要填充的矩陣，pad_width是各個維度上首尾填充的個數。
舉個例子，假定我們設置的 pad_width 是((0,0), (0,2))，而待處理的 mfcc 系數是 20 * 11 的矩陣。
我們把 mfcc 系數看成 20 行 11 列的矩陣，進行 pad 操作，第一個(0,0)對行進行操作，
表示每一行最前面和最后面增加的數個數為零，也就相當於總共增加了 0 列。第二個(0,2)對列操作，
表示每一列最前面增加的數為 0 個，但最后面要增加兩個數，也就相當於總共增加了 2 行。
mode 設置為 ‘constant’，表明填充的是常數，且默認為 0
"""
pad_size = max_pad_size - y_shape
audio_mac = np.pad(audio_mac, ((0, 0), (0, pad_size)), mode='constant')
else:
audio_mac = audio_mac[:, :max_pad_size]
return audio_mac
def save_data_to_array(label_path, max_pad_size=11):
"""
存儲處理過的數據，方便下一次的使用
:param label_path:
:param max_pad_size:
:return:
"""
mfcc_vectors = []
target = []
labels = load_label(label_path=label_path)
for i, label in enumerate(labels):
path = label_path + '/' + label
wavfiles = [path + '/' + file for file in os.listdir(path)]
for wavfile in wavfiles:
wav = wav2mfcc(wavfile, max_pad_size=max_pad_size)
mfcc_vectors.append(wav)
target.append(i)
np.save(DATA, mfcc_vectors)
np.save(TARGET, target)
# return mfcc_vectors, target
def get_train_test(split_ratio=.6, random_state=42):
"""
使用sklearn 中的train_test_split，把數據集分為訓練集和驗證集。其中訓練集占 6 成，測試集占 4 成
:param split_ratio:
:param random_state:
:return:
"""
X = np.load(DATA)
y = np.load(TARGET)
assert X.shape[0] == y.shape[0]
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=(1 - split_ratio), random_state=random_state,
shuffle=True)
return X_train, X_test, y_train, y_test
def main():
x_train, x_test, y_train, y_test = get_train_test()
# 變成二維矩陣且第二個維度大小為 220
x_train = x_train.reshape(-1, 220)
x_test = x_test.reshape(-1, 220)
# 使用kears中的onehot編碼
y_train_hot = to_categorical(y_train)
y_test_hot = to_categorical(y_test)
model = Sequential()
model.add(Dense(64, activation='relu', input_shape=(220,), name="input_layer"))
model.add(Dense(64, activation='relu', name="dropout_layer1"))
model.add(Dense(64, activation='relu', name="dropout_layer2"))
model.add(Dense(2, activation='softmax', name="output_layer"))
# 模型訓練
sess = tf.Session()
K.set_session(sess)
# 這步找到input_layer和output_layer的完整路徑，在golang中使用時需要用來定義輸入輸出node
for n in sess.graph.as_graph_def().node:
if 'input_layer' in n.name:
print(n.name)
if 'output_layer' in n.name:
print(n.name)
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=keras.optimizers.RMSprop(),
metrics=['accuracy'])
history = model.fit(x_train, y_train_hot, batch_size=100, epochs=100, verbose=1,
validation_data=(x_test, y_test_hot))
# 以下是關鍵代碼
# Use TF to save the graph model instead of Keras save model to load it in Golang
builder = tf.saved_model.builder.SavedModelBuilder("cnnModel")
# Tag the model, required for Go
builder.add_meta_graph_and_variables(sess, ["myTag"])
builder.save()
model.save("classaud.h5")
plot_history(history)
sess.close()
def save():
label_path = 'F:\\doc\\項目\\音頻分類\\audio-heji\\'
save_data_to_array(label_path, max_pad_size=11)
def plot_history(history):
plt.plot(history.history['acc'], label='train')
plt.plot(history.history['val_acc'], label='validation')
plt.legend()
plt.show()
if __name__ == "__main__":
save_data_to_array("F:\\doc\\項目\\音頻分類\\audio-heji\\", max_pad_size=11)
# save_data_to_array("/home/audio/audio-heji/", max_pad_size=11)
main()
音頻文件重命名代碼
import os
def rename(pic_path):
"""
這里對兩個目錄下文件進行排序，第一個fail-audio設置成10001這種，第二個success-audio目錄下設置成了90001這種。
:param pic_path:
:return:
"""
piclist = os.listdir(pic_path)
i = 1
print("ok")
for pic in piclist:
if pic.endswith(".wav"):
old_path = os.path.join(os.path.abspath(pic_path), pic)
new_path = os.path.join(os.path.abspath(pic_path), str(
90000 + (int(i))) + '.wav')
os.renames(old_path, new_path)
print("把原命名格式：" + old_path + u"轉換為新命名格式：" + new_path)
i = i + 1
# 加載標簽
def load_label(label_path):
label = os.listdir(label_path)
return label
if __name__ == '__main__':
rename("F:\\doc\\項目\\音頻分類\\audio-heji\\success-audio")
測試代碼
import librosa
import numpy as np
import os
from keras.models import load_model
# 提取 mfcc 參數
def wav2mfcc(path, max_pad_size=11):
y, sr = librosa.load(path=path, sr=None, mono=1)
y = y[::3] # 每三個點選用一個
audio_mac = librosa.feature.mfcc(y=y, sr=16000)
y_shape = audio_mac.shape[1]
if y_shape < max_pad_size:
pad_size = max_pad_size - y_shape
audio_mac = np.pad(audio_mac, ((0, 0), (0, pad_size)), mode='constant')
else:
audio_mac = audio_mac[:, :max_pad_size]
return audio_mac
def load_label(label_path):
"""
遍歷當前給定的目錄，便於后文進行標簽加載
:param label_path:
:return:
"""
label = os.listdir(label_path)
return label
if __name__ == '__main__':
# 加載模型
model = load_model('classaud.h5') # 加載訓練模型
wavs = [wav2mfcc("F:\\doc\\項目\\音頻分類\\test\\" + file, 11) for file in os.listdir("F:\\doc\\項目\\音頻分類\\test")]
X = np.array(wavs)
X = X.reshape(-1, 220)
print(X.shape)
for j in range(X.shape[0]):
print(j)
print(X[j:j+1])
result = model.predict(X[j:j+1])[0] #
print("識別結果", result)
# 因為在訓練的時候，標簽集的名字 為： 0：fail-audio 1：success-audio
name = ["fail-audio", "success-audio"] # 創建一個跟訓練時一樣的標簽集
ind = 0 # 結果中最大的一個數
for i in range(len(result)):
if result[i] > result[ind]:
ind = 1
print("識別的語音結果是：", name[ind])
在模型訓練中，為了后續golang中加載訓練好的模型，增加了部分代碼，主要是如下
# 模型訓練
sess = tf.Session()
K.set_session(sess)
# 這步找到input_layer和output_layer的完整路徑，在golang中使用時需要用來定義輸入輸出node
for n in sess.graph.as_graph_def().node:
if 'input_layer' in n.name:
print(n.name)
if 'output_layer' in n.name:
print(n.name)
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=keras.optimizers.RMSprop(),
metrics=['accuracy'])
history = model.fit(x_train, y_train_hot, batch_size=100, epochs=100, verbose=1,
validation_data=(x_test, y_test_hot))
# 以下是關鍵代碼
# Use TF to save the graph model instead of Keras save model to load it in Golang
builder = tf.saved_model.builder.SavedModelBuilder("cnnModel")
# Tag the model, required for Go
builder.add_meta_graph_and_variables(sess, ["myTag"])
builder.save()
模型訓練完成后會生成響應的模型，其中cnnModel文件夾包含pd模型及variables文件夾，為后續golang調度使用的模型，h5模型為這里測試使用的模型，結果如下圖
另外在模型訓練時，我們打印出了每層神經網絡的名字，這塊需要關注，因為在后續的golang環境中因為加載節點名錯誤導致的問題，打印如下
WARNING:tensorflow:From E:\pycharm\nlu-algorithm-package\venv\lib\site-packages\tensorflow\python\framework\op_def_library.py:263: colocate_with (from tensorflow.python.framework.ops) is deprecated and will be removed in a future version.
Instructions for updating:
Colocations handled automatically by placer.
2020-08-27 19:06:19.437965: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2
input_layer_input
input_layer/random_uniform/shape
input_layer/random_uniform/min
input_layer/random_uniform/max
input_layer/random_uniform/RandomUniform
input_layer/random_uniform/sub
input_layer/random_uniform/mul
input_layer/random_uniform
input_layer/kernel
input_layer/kernel/Assign
input_layer/kernel/read
input_layer/Const
input_layer/bias
input_layer/bias/Assign
input_layer/bias/read
input_layer/MatMul
input_layer/BiasAdd
input_layer/Relu
output_layer/random_uniform/shape
output_layer/random_uniform/min
output_layer/random_uniform/max
output_layer/random_uniform/RandomUniform
output_layer/random_uniform/sub
output_layer/random_uniform/mul
output_layer/random_uniform
output_layer/kernel
output_layer/kernel/Assign
output_layer/kernel/read
output_layer/Const
output_layer/bias
output_layer/bias/Assign
output_layer/bias/read
output_layer/MatMul
output_layer/BiasAdd
output_layer/Softmax
測試中的一條結果記錄為
25
[[-1.66098328e+02 -2.12202209e+02 -4.31954193e+02 -4.47424835e+02
-3.33869904e+02 -3.58775604e+02 -5.43935608e+02 -6.60088867e+02
-2.30052383e+02 -6.87607422e+01 -1.21050777e+01 4.23143768e+01
4.26393929e+01 4.03902130e+01 3.61016235e+01 3.74344673e+01
3.76606216e+01 2.93862953e+01 1.91747296e+00 -9.67822647e+00
-1.10491867e+01 -1.35342636e+01 -6.71506195e+01 -6.73550262e+01
-6.63572083e+01 -4.63606071e+01 -4.91609383e+01 -5.08716202e+01
-4.48681793e+01 -1.87636101e+00 -4.61852417e+01 -4.17765961e+01
-2.01916142e+01 -9.79832268e+00 -9.91659927e+00 -1.03536911e+01
-3.72100983e+01 -3.87253189e+01 -3.71894073e+01 -2.30876770e+01
7.12325764e+00 -1.86448917e+01 -1.23296547e+01 -1.36949463e+01
2.36775169e+01 2.40340843e+01 2.65298653e+01 -1.50987358e+01
-1.75523911e+01 -1.85126324e+01 -1.25076523e+01 7.13483989e-01
1.25870705e+01 9.64732361e+00 1.51968069e+01 1.57215271e+01
1.60245552e+01 1.79179173e+01 8.83283997e+00 6.01854324e+00
5.03554726e+00 7.35849476e+00 7.32028627e+00 2.61785851e+01
2.36798325e+01 2.17165947e+01 -8.92044258e+00 -8.98554516e+00
-9.36619282e+00 8.72706318e+00 7.63292933e+00 8.11320019e+00
1.23650103e+01 4.05273724e+00 -2.74963975e+00 6.03793526e+00
3.32019657e-01 -1.07288876e+01 -1.08059797e+01 -1.12256699e+01
3.14207101e+00 2.12598038e+00 3.41194320e+00 1.52846613e+01
3.58989859e+00 -7.22187281e+00 -2.92357159e+00 -1.40336580e+01
1.25538235e+01 1.30331860e+01 1.53311596e+01 4.33768129e+00
3.44467473e+00 3.40012264e+00 8.20392323e+00 4.26902437e+00
2.08825417e+01 1.70654850e+01 6.18888092e+00 1.60812531e+01
1.66913948e+01 1.89202042e+01 -3.88430738e+00 -3.84901094e+00
-2.63745117e+00 -1.11108208e+00 6.36666417e-01 -2.25304246e+00
-8.85197830e+00 -1.91202374e+01 -1.46577859e+00 -1.23925185e+00
7.83551395e-01 2.47491241e+00 1.99022865e+00 1.92004573e+00
-4.81319517e-01 5.03325987e+00 -3.67527580e+00 1.00166121e+01
1.32022190e+01 1.41424477e+00 1.61442292e+00 1.70821917e+00
1.69235497e+01 1.71120987e+01 1.50358353e+01 1.89076972e+00
6.39788628e+00 -1.38607597e+00 1.02029294e-01 -1.02280178e+01
4.86864090e+00 5.27380610e+00 7.52170134e+00 -5.28688669e+00
-4.54772520e+00 -1.99729645e+00 9.84556198e+00 6.01770210e+00
6.42514515e+00 5.05019569e+00 -3.58427215e+00 2.53060675e+00
2.77301073e+00 3.36337566e+00 -3.74559736e+00 -3.88710737e+00
-2.85192370e+00 8.00442696e+00 5.90060949e-01 4.97644138e+00
8.34950066e+00 6.98132086e+00 -2.52411366e+00 -2.62051678e+00
-2.87881041e+00 7.85154676e+00 7.58860874e+00 7.21512508e+00
6.53992605e+00 -2.57507980e-01 -4.69269657e+00 -3.40787172e+00
1.32537198e+00 -4.12547922e+00 -4.38115311e+00 -5.22326088e+00
2.03997135e+00 2.89152718e+00 4.46722126e+00 6.32854557e+00
5.27089882e+00 8.66948891e+00 5.42871141e+00 1.18013754e+01
-1.06842923e+00 -1.21782100e+00 -1.62583649e+00 4.68027020e+00
4.82862568e+00 5.17801666e+00 5.02924442e+00 3.57280898e+00
-1.60658951e+01 -3.89933228e+00 5.28476810e+00 7.85575271e-01
7.23506689e-01 9.12119806e-01 1.29786149e-01 -1.08789623e+00
-2.01344442e+00 -3.12455368e+00 4.22501802e+00 -2.05132604e-01
4.64199352e+00 1.28417645e+01 -2.14332151e+00 -2.28186941e+00
-2.92554092e+00 -7.33241290e-02 -4.54517424e-01 -9.12135363e-01
-1.92673039e+00 1.32999837e+00 -5.95955181e+00 -1.38899193e+01
-1.53170991e+00 -3.57915735e+00 -3.69184279e+00 -3.76903296e+00
-3.22209269e-01 -6.59340382e-01 5.86225927e-01 1.03645115e+01
2.81656504e+00 -1.55326450e+00 -1.87907255e+00 -2.12706447e+00]]
識別結果 [1.4359492e-26 1.0000000e+00]
識別的語音結果是： success-audio
后續的golang調用tensorflow模型中我們以此結果作為測試。
備注：此處使用的tensorflow為1.13.1版本、kears為2.2.4、python3.6.0，音頻特征抽取使用的librosa庫，版本0.8.0。
其他方式實現的音頻分類，在實踐中也參考了 Python Project – Music Genre Classification 該文章時使用的K近鄰方式實現的，在實踐中也做了此方式，准確率在70%左右，也可以參考。
3 golang調用tensorflow/keras訓練的模型
在 安裝 go 版 TensorFlow的問題記錄 我們已經將相關環境設置好，此處只需要完成相關的go代碼即可上線測試。代碼驗證參考了 golang調用tensorflow/keras訓練的模型 此處先給出基於此文章實現的驗證代碼
package main
import (
"fmt"
tf "github.com/tensorflow/tensorflow/tensorflow/go"
)
func main() {
// 特征長度
const MAXLEN int = 220
// 將文本轉換為id序列，為了實驗方便直接使用轉換好的ID序列即可，此處是使用上文中測試中打印出來的音頻特征
input_data := [1][MAXLEN]float32{{-1.66098328e+02,-2.12202209e+02,-4.31954193e+02,-4.47424835e+02,-3.33869904e+02,-3.58775604e+02,-5.43935608e+02,-6.60088867e+02,-2.30052383e+02,-6.87607422e+01,-1.21050777e+01,4.23143768e+01,4.26393929e+01,4.03902130e+01,3.61016235e+01,3.74344673e+01,3.76606216e+01,2.93862953e+01,1.91747296e+00,-9.67822647e+00,-1.10491867e+01,-1.35342636e+01,-6.71506195e+01,-6.73550262e+01,-6.63572083e+01,-4.63606071e+01,-4.91609383e+01,-5.08716202e+01,-4.48681793e+01,-1.87636101e+00,-4.61852417e+01,-4.17765961e+01,-2.01916142e+01,-9.79832268e+00,-9.91659927e+00,-1.03536911e+01,-3.72100983e+01,-3.87253189e+01,-3.71894073e+01,-2.30876770e+01,7.12325764e+00,-1.86448917e+01,-1.23296547e+01,-1.36949463e+01,2.36775169e+01,2.40340843e+01,2.65298653e+01,-1.50987358e+01,-1.75523911e+01,-1.85126324e+01,-1.25076523e+01,7.13483989e-01,1.25870705e+01,9.64732361e+00,1.51968069e+01,1.57215271e+01,1.60245552e+01,1.79179173e+01,8.83283997e+00,6.01854324e+00,5.03554726e+00,7.35849476e+00,7.32028627e+00,2.61785851e+01,2.36798325e+01,2.17165947e+01,-8.92044258e+00,-8.98554516e+00,-9.36619282e+00,8.72706318e+00,7.63292933e+00,8.11320019e+00,1.23650103e+01,4.05273724e+00,-2.74963975e+00,6.03793526e+00,3.32019657e-01,-1.07288876e+01,-1.08059797e+01,-1.12256699e+01,3.14207101e+00,2.12598038e+00,3.41194320e+00,1.52846613e+01,3.58989859e+00,-7.22187281e+00,-2.92357159e+00,-1.40336580e+01,1.25538235e+01,1.30331860e+01,1.53311596e+01,4.33768129e+00,3.44467473e+00,3.40012264e+00,8.20392323e+00,4.26902437e+00,2.08825417e+01,1.70654850e+01,6.18888092e+00,1.60812531e+01,1.66913948e+01,1.89202042e+01,-3.88430738e+00,-3.84901094e+00,-2.63745117e+00,-1.11108208e+00,6.36666417e-01,-2.25304246e+00,-8.85197830e+00,-1.91202374e+01,-1.46577859e+00,-1.23925185e+00,7.83551395e-01,2.47491241e+00,1.99022865e+00,1.92004573e+00,-4.81319517e-01,5.03325987e+00,-3.67527580e+00,1.00166121e+01,1.32022190e+01,1.41424477e+00,1.61442292e+00,1.70821917e+00,1.69235497e+01,1.71120987e+01,1.50358353e+01,1.89076972e+00,6.39788628e+00,-1.38607597e+00,1.02029294e-01,-1.02280178e+01,4.86864090e+00,5.27380610e+00,7.52170134e+00,-5.28688669e+00,-4.54772520e+00,-1.99729645e+00,9.84556198e+00,6.01770210e+00,6.42514515e+00,5.05019569e+00,-3.58427215e+00,2.53060675e+00,2.77301073e+00,3.36337566e+00,-3.74559736e+00,-3.88710737e+00,-2.85192370e+00,8.00442696e+00,5.90060949e-01,4.97644138e+00,8.34950066e+00,6.98132086e+00,-2.52411366e+00,-2.62051678e+00,-2.87881041e+00,7.85154676e+00,7.58860874e+00,7.21512508e+00,6.53992605e+00,-2.57507980e-01,-4.69269657e+00,-3.40787172e+00,1.32537198e+00,-4.12547922e+00,-4.38115311e+00,-5.22326088e+00,2.03997135e+00,2.89152718e+00,4.46722126e+00,6.32854557e+00,5.27089882e+00,8.66948891e+00,5.42871141e+00,1.18013754e+01,-1.06842923e+00,-1.21782100e+00,-1.62583649e+00,4.68027020e+00,4.82862568e+00,5.17801666e+00,5.02924442e+00,3.57280898e+00,-1.60658951e+01,-3.89933228e+00,5.28476810e+00,7.85575271e-01,7.23506689e-01,9.12119806e-01,1.29786149e-01,-1.08789623e+00,-2.01344442e+00,-3.12455368e+00,4.22501802e+00,-2.05132604e-01,4.64199352e+00,1.28417645e+01,-2.14332151e+00,-2.28186941e+00,-2.92554092e+00,-7.33241290e-02,-4.54517424e-01,-9.12135363e-01,-1.92673039e+00,1.32999837e+00,-5.95955181e+00,-1.38899193e+01,-1.53170991e+00,-3.57915735e+00,-3.69184279e+00,-3.76903296e+00,-3.22209269e-01,-6.59340382e-01,5.86225927e-01,1.03645115e+01,2.81656504e+00,-1.55326450e+00,-1.87907255e+00,-2.12706447e+00}}
tensor, err := tf.NewTensor(input_data)
if err != nil {
fmt.Printf("Error NewTensor: err: %s", err.Error())
return
}
//讀取模型
model, err := tf.LoadSavedModel("cnnModel", []string{"myTag"}, nil)
if err != nil {
fmt.Printf("Error loading Saved Model: %s\n", err.Error())
return
}
// 識別
result, err := model.Session.Run(
map[tf.Output]*tf.Tensor{
// python版tensorflow/keras中定義的輸入層input_layer
model.Graph.Operation("input_layer").Output(0): tensor,
},
[]tf.Output{
// python版tensorflow/keras中定義的輸出層output_layer
model.Graph.Operation("output_layer/Softmax").Output(0),
},
nil,
)
if err != nil {
fmt.Printf("Error running the session with input, err: %s ", err.Error())
return
}
// 輸出結果，interface{}格式
fmt.Printf("Result value: %v", result[0].Value())
}
將訓練好的模型及驗證程序上傳到虛擬機的gopath目錄
[root@localhost gopath]# cd /home/gopath/
[root@localhost gopath]#
[root@localhost gopath]# ll
總用量 8
-rw-r--r--. 1 root root 5072 8月 27 21:21 cnn.go
drwxr-xr-x. 3 root root 45 8月 27 19:27 cnnModel
drwxr-xr-x. 3 root root 25 8月 27 15:40 pkg
drwxr-xr-x. 5 root root 65 8月 27 15:40 src
執行程序后報錯，如下所示
[root@localhost gopath]# go run cnn.go
2020-08-27 21:22:55.306065: I tensorflow/cc/saved_model/reader.cc:31] Reading SavedModel from: cnnModel
2020-08-27 21:22:55.312739: I tensorflow/cc/saved_model/reader.cc:54] Reading meta graph with tags { myTag }
2020-08-27 21:22:55.320111: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2
2020-08-27 21:22:55.335002: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 1799995000 Hz
2020-08-27 21:22:55.335970: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x1ba6ea0 initialized for platform Host (this does not guarantee that XLA will be used). Devices:
2020-08-27 21:22:55.336091: I tensorflow/compiler/xla/service/service.cc:176] StreamExecutor device (0): Host, Default Version
2020-08-27 21:22:55.364460: I tensorflow/cc/saved_model/loader.cc:202] Restoring SavedModel bundle.
2020-08-27 21:22:55.449190: I tensorflow/cc/saved_model/loader.cc:311] SavedModel load for tags { myTag }; Status: success. Took 143138 microseconds.
模型read成功panic: nil-Operation. If the Output was created with a Scope object, see Scope.Err() for details.
goroutine 1 [running]:
github.com/tensorflow/tensorflow/tensorflow/go.Output.c(...)
/home/gopath/src/github.com/tensorflow/tensorflow/tensorflow/go/operation.go:130
github.com/tensorflow/tensorflow/tensorflow/go.newCRunArgs(0xc000097e50, 0xc000097e20, 0x1, 0x1, 0x0, 0x0, 0x0, 0xc000097668)
/home/gopath/src/github.com/tensorflow/tensorflow/tensorflow/go/session.go:369 +0x594
github.com/tensorflow/tensorflow/tensorflow/go.(*Session).Run(0xc00000e0c0, 0xc000097e50, 0xc000097e20, 0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, ...)
/home/gopath/src/github.com/tensorflow/tensorflow/tensorflow/go/session.go:143 +0x1e2
main.main()
/home/gopath/cnn.go:27 +0x1199
exit status 2
[root@localhost gopath]#
錯誤顯示找不到輸出節點，可是我們在代碼中已經設置了相關的輸出節點信息，如下所示
map[tf.Output]*tf.Tensor{
// python版tensorflow/keras中定義的輸入層input_layer
model.Graph.Operation("input_layer").Output(0): tensor,
},
[]tf.Output{
// python版tensorflow/keras中定義的輸出層output_layer
model.Graph.Operation("output_layer/Softmax").Output(0),
},
那找不到相關的節點是否是節點名字綁定問題，我們回看在模型訓練時打印的各層節點名稱，發現是以“input_layer_input”開始“output_layer/Softmax”結束，查看我們代碼中是按照訓練中設置的name進行標記的，故將輸入的開始節點從“input_layer”修改為“input_layer_input”，如下
map[tf.Output]*tf.Tensor{
// python版tensorflow/keras中定義的輸入層input_layer
model.Graph.Operation("input_layer_input").Output(0): tensor,
},
[]tf.Output{
// python版tensorflow/keras中定義的輸出層output_layer
model.Graph.Operation("output_layer/Softmax").Output(0),
},
重新運行程序，結果如下
[root@localhost gopath]# go run cnn.go
2020-08-27 21:28:33.161277: I tensorflow/cc/saved_model/reader.cc:31] Reading SavedModel from: cnnModel
2020-08-27 21:28:33.166881: I tensorflow/cc/saved_model/reader.cc:54] Reading meta graph with tags { myTag }
2020-08-27 21:28:33.173316: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2
2020-08-27 21:28:33.186007: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 1799995000 Hz
2020-08-27 21:28:33.187068: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x13fbea0 initialized for platform Host (this does not guarantee that XLA will be used). Devices:
2020-08-27 21:28:33.187126: I tensorflow/compiler/xla/service/service.cc:176] StreamExecutor device (0): Host, Default Version
2020-08-27 21:28:33.210363: I tensorflow/cc/saved_model/loader.cc:202] Restoring SavedModel bundle.
2020-08-27 21:28:33.294067: I tensorflow/cc/saved_model/loader.cc:311] SavedModel load for tags { myTag }; Status: success. Took 132796 microseconds.
模型read成功模型識別成功Result value: [[1.4359492e-26 1]]
對比上文給出的測試中的識別結果是保持一致的，故這里golang調度keras訓練的模型成功，后續就是實際場景如何進行使用的問題規划。