TFLiteConverter saved model error with

TensorFlow General Discussion
,
1

Hello,
I’ve converted a tensorflow model to a tensorflow lite model and I’ve used tflite built-in ops(Exp, Greater, Mul, Square, Sub, Sum, Transpose), but I get error:

TF Select ops: AddV2, Exp, Greater, MatMul, Mul, Square, Sub, Sum, Transpose
Details:
        tf.AddV2(tensor<1x1xf64>, tensor<f64>) -> (tensor<1x1xf64>) : {device = ""}
        tf.Exp(tensor<133x1xf64>) -> (tensor<133x1xf64>) : {device = ""}
        tf.Greater(tensor<1x1xf64>, tensor<f64>) -> (tensor<1x1xi1>) : {device = ""}
        tf.MatMul(tensor<1x133xf64>, tensor<1x133xf64>) -> (tensor<1x1xf64>) : {transpose_a = false, transpose_b = true}
        tf.Mul(tensor<133x1xf64>, tensor<f64>) -> (tensor<133x1xf64>) : {device = ""}
        tf.Square(tensor<133x72xf64>) -> (tensor<133x72xf64>) : {device = ""}
        tf.Sub(tensor<72xf64>, tensor<133x72xf64>) -> (tensor<133x72xf64>) : {device = ""}
        tf.Sum(tensor<133x72xf64>, tensor<i32>) -> (tensor<133x1xf64>) : {device = "", keep_dims = true}
        tf.Transpose(tensor<133x1xf64>, tensor<2xi32>) -> (tensor<1x133xf64>)

Why these become the custom ops?

code:

class oneClassSVM(tf.Module):
    def __init__(self, dual_coef, intercept, support_vector, gamma):
        self.dual_coef = dual_coef
        self.intercept = intercept
        self.support_vector = support_vector
        self.gamma = gamma    
	@tf.function(input_signature=[tf.TensorSpec(shape=(72), dtype=tf.float64)])
    def svm_predict(self, x):
        distance = tf.reduce_sum(tf.square(tf.subtract(x, self.support_vector)), axis=1, keepdims=True)
        summation = tf.matmul(self.dual_coef, tf.math.exp(tf.multiply(tf.negative(self.gamma),  distance)))
        predict = 1 if tf.add(summation, self.intercept) > 0 else -1
        return predict
1 Like
3

Hi @wei-ying_Wang

Please provide the error and the code for conversion.

Regards

5

Hi @George_Soloupis ,
I just call tf.lite.TFLiteConverter.from_save_model() and convert(), then show error above.
BUT! After I modify the data type form float64 to float32, it can converter success now.
I think TFLite doesn’t support float64, so need select OPs when converter

1 Like