Python中deep tensorflow keras模型的输出与tensorflowjs转换之间的显著差异
我一直在尝试使用tensorflowjs在浏览器中直接在预先训练好的Keras模型中运行推理，我正在努力获得相同的性能 在调查中，我注意到，在所有层面上，产量的累积差异都很小，到最后变得相当糟糕 下面的简化示例可以用来说明。这是用tensorflow 2.1.0、tensorflowjs 1.7.2和Chrome 80.0.3987.163测试的 首先，我们在tf.keras中构建一个简单的deep模型，将权重初始化为随机值，并将其转换为tensorflowjsPython中deep tensorflow keras模型的输出与tensorflowjs转换之间的显著差异,keras,tf.keras,tensorflow.js,Keras,Tf.keras,Tensorflow.js,我一直在尝试使用tensorflowjs在浏览器中直接在预先训练好的Keras模型中运行推理，我正在努力获得相同的性能 在调查中，我注意到，在所有层面上，产量的累积差异都很小，到最后变得相当糟糕 下面的简化示例可以用来说明。这是用tensorflow 2.1.0、tensorflowjs 1.7.2和Chrome 80.0.3987.163测试的 首先，我们在tf.keras中构建一个简单的deep模型，将权重初始化为随机值，并将其转换为tensorflowjs 将tensorflowjs导入为
将tensorflowjs导入为tfjs
从tensorflow.keras.models导入模型，保存_模型
从tensorflow.keras将图层导入为L
从tensorflow.keras导入后端为K
从tensorflow.keras导入初始值设定项
K.清除会话（）
def build_模型（深度、大小）：
x=L.Input（（大小），name='Input'）
输入=[x]
输出=[]
uniform_init=I.RandomUniform（minval=-1，maxval=1，seed=123）
对于范围内的i（深度）：
x=L.密度（尺寸，
数据类型='float32'，
内核初始化器=统一初始化，
偏差\初始值设定项=一致\初始值，
名称='densite_uz'+str（i））（x）
输出+=[x]
模型=模型（输入=输入，输出=输出）
返回（模型、输入、输出）
模型、输入、输出=构建模型（10,50）
compile（优化器='adam'，loss='mean_squared_error'）
model.summary（）
保存_模型（模型“./keras_模型.h5”）
tfjs.converters.save_keras_model（model'。/keras_converted'）
Model: "model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input (InputLayer) [(None, 50)] 0
_________________________________________________________________
dense_0 (Dense) (None, 50) 2550
_________________________________________________________________
dense_1 (Dense) (None, 50) 2550
_________________________________________________________________
dense_2 (Dense) (None, 50) 2550
_________________________________________________________________
dense_3 (Dense) (None, 50) 2550
_________________________________________________________________
dense_4 (Dense) (None, 50) 2550
_________________________________________________________________
dense_5 (Dense) (None, 50) 2550
_________________________________________________________________
dense_6 (Dense) (None, 50) 2550
_________________________________________________________________
dense_7 (Dense) (None, 50) 2550
_________________________________________________________________
dense_8 (Dense) (None, 50) 2550
_________________________________________________________________
dense_9 (Dense) (None, 50) 2550
=================================================================
Total params: 25,500
Trainable params: 25,500
Non-trainable params: 0
dense_0_weights[0]: different values = 0, average difference = 0.0
dense_0_weights[1]: different values = 0, average difference = 0.0
dense_0_output[0]: different values = 0, average difference = 0.0
dense_1_weights[0]: different values = 0, average difference = 0.0
dense_1_weights[1]: different values = 0, average difference = 0.0
dense_1_output[0]: different values = 40, average difference = 1.7833709716796876e-06
dense_2_weights[0]: different values = 0, average difference = 0.0
dense_2_weights[1]: different values = 0, average difference = 0.0
dense_2_output[0]: different values = 43, average difference = 1.2047290802001953e-05
dense_3_weights[0]: different values = 0, average difference = 0.0
dense_3_weights[1]: different values = 0, average difference = 0.0
dense_3_output[0]: different values = 38, average difference = 5.3539276123046876e-05
dense_4_weights[0]: different values = 0, average difference = 0.0
dense_4_weights[1]: different values = 0, average difference = 0.0
dense_4_output[0]: different values = 43, average difference = 0.00024005889892578125
dense_5_weights[0]: different values = 0, average difference = 0.0
dense_5_weights[1]: different values = 0, average difference = 0.0
dense_5_output[0]: different values = 45, average difference = 0.0010586166381835937
dense_6_weights[0]: different values = 0, average difference = 0.0
dense_6_weights[1]: different values = 0, average difference = 0.0
dense_6_output[0]: different values = 46, average difference = 0.00515625
dense_7_weights[0]: different values = 0, average difference = 0.0
dense_7_weights[1]: different values = 0, average difference = 0.0
dense_7_output[0]: different values = 47, average difference = 0.0254345703125
dense_8_weights[0]: different values = 0, average difference = 0.0
dense_8_weights[1]: different values = 0, average difference = 0.0
dense_8_output[0]: different values = 46, average difference = 0.11546875
dense_9_weights[0]: different values = 0, average difference = 0.0
dense_9_weights[1]: different values = 0, average difference = 0.0
dense_9_output[0]: different values = 44, average difference = 0.42203125
异步函数运行（）{
const model=await tf.loadLayersModel（'/keras\u converted/model.json'）
model.summary（）
让input_size=model.inputLayers[0]。batchInputShape[1]
let one=[…数组（输入大小）].map（x=>1）
设x=tf.张量（个）。重塑（[1，输入大小]）
让yhat=model.predict（x.map）（x=>x.arraySync（））
设i=0
设vals={}
model.outputLayers.forEach（输出=>{
VAL[output.name+'_output']=yhat[i]
i+=1
})
model.layers.forEach（层=>{
VAL[layer.name+'_weights']=layer.getWeights（）.map（w=>w.arraySync（））
})
设a=document.createElement（'a'）
let file=new Blob（[JSON.stringify（vals）]，{type:'text/plain'}）
a、 href=URL.createObjectURL（文件）
a、 下载='keras\u model\u tfjs\u output.json'
a、 单击（）
}
运行（）
重新导入
导入json
将numpy作为np导入
从tensorflow.keras.models导入负载_模型
模型=负荷模型（'./keras_模型.h5'）
#负载tensorflow js输出和重量
将open（'./keras_model_tfjs_output.json'，'r'）作为json_文件：
tfjs_vals=json.load（json_文件）
#通过在Python中运行推理创建类似的字典
keras_vals=dict（）
#在1的单个向量上运行模型
yhat=model.predict（np.one（（1，输入[0]。形状[1]））
对于i，枚举中的层（model.layers[1:]）：
keras_vals[layer.name+'_weights']=layer.get_weights（）
keras_vals[layer.name+'_output']=yhat[i]
#比较keras\u vals和tfjs\u vals中的值
def比较值（键a、b）：
对于范围内的i（len（a））：
print（{0}[{1}]：不同的值={2}，平均差={3}\
格式（键，i，
np.sum（a[i]！=b[i]），
np.sum（np.abs（a[i]-b[i]））/np.size（a[i]）
))
对于dict.keys中的键（keras_vals）：
比较变量（键、keras变量[key]、tfjs变量[key]）
Model: "model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input (InputLayer) [(None, 50)] 0
_________________________________________________________________
dense_0 (Dense) (None, 50) 2550
_________________________________________________________________
dense_1 (Dense) (None, 50) 2550
_________________________________________________________________
dense_2 (Dense) (None, 50) 2550
_________________________________________________________________
dense_3 (Dense) (None, 50) 2550
_________________________________________________________________
dense_4 (Dense) (None, 50) 2550
_________________________________________________________________
dense_5 (Dense) (None, 50) 2550
_________________________________________________________________
dense_6 (Dense) (None, 50) 2550
_________________________________________________________________
dense_7 (Dense) (None, 50) 2550
_________________________________________________________________
dense_8 (Dense) (None, 50) 2550
_________________________________________________________________
dense_9 (Dense) (None, 50) 2550
=================================================================
Total params: 25,500
Trainable params: 25,500
Non-trainable params: 0
dense_0_weights[0]: different values = 0, average difference = 0.0
dense_0_weights[1]: different values = 0, average difference = 0.0
dense_0_output[0]: different values = 0, average difference = 0.0
dense_1_weights[0]: different values = 0, average difference = 0.0
dense_1_weights[1]: different values = 0, average difference = 0.0
dense_1_output[0]: different values = 40, average difference = 1.7833709716796876e-06
dense_2_weights[0]: different values = 0, average difference = 0.0
dense_2_weights[1]: different values = 0, average difference = 0.0
dense_2_output[0]: different values = 43, average difference = 1.2047290802001953e-05
dense_3_weights[0]: different values = 0, average difference = 0.0
dense_3_weights[1]: different values = 0, average difference = 0.0
dense_3_output[0]: different values = 38, average difference = 5.3539276123046876e-05
dense_4_weights[0]: different values = 0, average difference = 0.0
dense_4_weights[1]: different values = 0, average difference = 0.0
dense_4_output[0]: different values = 43, average difference = 0.00024005889892578125
dense_5_weights[0]: different values = 0, average difference = 0.0
dense_5_weights[1]: different values = 0, average difference = 0.0
dense_5_output[0]: different values = 45, average difference = 0.0010586166381835937
dense_6_weights[0]: different values = 0, average difference = 0.0
dense_6_weights[1]: different values = 0, average difference = 0.0
dense_6_output[0]: different values = 46, average difference = 0.00515625
dense_7_weights[0]: different values = 0, average difference = 0.0
dense_7_weights[1]: different values = 0, average difference = 0.0
dense_7_output[0]: different values = 47, average difference = 0.0254345703125
dense_8_weights[0]: different values = 0, average difference = 0.0
dense_8_weights[1]: different values = 0, average difference = 0.0
dense_8_output[0]: different values = 46, average difference = 0.11546875
dense_9_weights[0]: different values = 0, average difference = 0.0
dense_9_weights[1]: different values = 0, average difference = 0.0
dense_9_output[0]: different values = 44, average difference = 0.42203125