tensorflow 模型计算中,预测错误;权重参数加载
tensorflow 模型计算主要代码(正确代码)
linear1_kernel_initializer = tf.constant_initializer(numpy.transpose(data["linear1.weight"]))
linear1_bias_initializer = tf.constant_initializer(numpy.transpose(data["linear1.bias"]))
linear1 = layers.Dense(units=400, activation=tf.nn.relu, kernel_initializer=linear1_kernel_initializer, use_bias=True, bias_initializer=linear1_bias_initializer, input_shape=(48,))
linear2_kernel_initializer = tf.constant_initializer(numpy.transpose(data["linear2.weight"]))
linear2_bias_initializer = tf.constant_initializer(numpy.transpose(data["linear2.bias"]))
linear2 = layers.Dense(units=400, activation=tf.nn.relu, kernel_initializer=linear2_kernel_initializer, use_bias=True, bias_initializer=linear2_bias_initializer)
linear3_kernel_initializer = tf.constant_initializer(numpy.transpose(data["linear3.weight"]))
linear3_bias_initializer = tf.constant_initializer(numpy.transpose(data["linear3.bias"]))
linear3 = layers.Dense(units=2, activation=None, kernel_initializer=linear3_kernel_initializer, use_bias=True, bias_initializer=linear3_bias_initializer)
model = tf.keras.Sequential([linear1, linear2, linear3])
input = numpy.ones((2, 48), dtype=float)
predict = model.predict(input)
print(predict[0:100,:])
原本权重参数采用以下代码
linear1_kernel_initializer = tf.constant_initializer(data["linear1.weight"])
linear1_bias_initializer = tf.constant_initializer((data["linear1.bias"])
但模型预测值与Matlab计算值有误。后经过测试定位到 layers.Dense 此处,然后创建 layers.Dense时设置use_bias=False参数,不去考虑偏差参数。改变初始权重参数方式:
input_size = 2
units_p = 3
data = numpy.array([1, 1, 2, 2, 2, 3], dtype=float)
linear1_kernel_initializer = tf.constant_initializer(data)
linear1 = layers.Dense(units=units_p, activation=None, kernel_initializer=linear1_kernel_initializer, use_bias=False, input_shape=(input_size,))
#变化data
data = numpy.array([1, 2, 3, 1, 2, 3], dtype=float)
#或者
data = numpy.array([1, 2, 3, 1, 2, 3], dtype=float).reshape(3, 2)
通过这样的方式,才发现 linear1_kernel_initializer = tf.constant_initializer(data)
中的 data
有问题,通过对预测结果的分析,发现 tf.constant_initializer()
会将传递过来的数据拉成一维,再根据 units
和 不同层
来变更数据矩阵大小,所以传入tf.constant_initializer()
的数据只要总大小是对的就可以传入,而不需要shape一致。
所以,既然之前的数据预测结果有误,那就是数据排列有误,将 data 数据进行矩阵转置 再 传入到tf.constant_initializer() 函数中
问题成功解决。
同时我想说明的是,pytorch
的torch.nn.Linear
是W x + b
而 tensorflow
的 layers.Dense
是 x W + b
。
tensorflow这种情况可以形象的表达为 流动的关系,input -> HL1 -> HL2 -> output
(HL1为隐藏层1)
input 卷上 W1 + b1 => HL1结果
HL1结果 卷上 W2 + b2 => HL2结果
HL2结果 卷上 W3 + b3 => outpu