You cannot build your model by calling `build` if your layers do not support float type inputs

General Discussion
, ,
#1

I am trying to code a coordinate attention mechanism.But when I instantiate the model, an error occurred while building，my code below

import numpy as np
from tensorflow.keras.layers import AvgPool2D, Conv2D, BatchNormalization, Activation
from tensorflow.keras import Model
from tensorflow.keras.activations import sigmoid
import tensorflow as tf

def ac_swish(x):
    temp = min(max(0, x + 3), 6) / 6
    x = x * temp
    return x

def coord_act(x):
    tmpx = tf.nn.relu6(x+3) / 6
    x = x * tmpx
    return x

class Coordinate_Attention(Model):
    def __init__(self, w, h, inp, oup, groups=32):
        """
        :param w:   width
        :param h:   height
        :param inp: input channels
        :param oup: output channels
        :param groups:
        """
        super(Coordinate_Attention, self).__init__()
        self.w = w
        self.h = h
        self.inp = inp
        self.oup = oup
        self.groups = groups

        self.pool_h = AvgPool2D(pool_size=(1, self.w), strides=1, padding='same')
        self.pool_w = AvgPool2D(pool_size=(self.h, 1), strides=1, padding='same')

        self.mip = max(8, self.inp // self.groups)

        self.conv1 = Conv2D(filters=self.mip, kernel_size=(1,1), strides=1, padding='same')
        self.bn1 = BatchNormalization()
        self.conv2 = Conv2D(filters=self.oup, kernel_size=(1,1), strides=1, padding='same')
        self.conv3 = Conv2D(filters=self.oup, kernel_size=(1,1), strides=1, padding='same')

        self.ac = Activation(ac_swish)

    def call(self, inputs):
        residual = inputs
        x = residual
        n, c, h, w = x.shape

        x_h = self.pool_h(x)
        x_w = self.pool_w(x)

        x_w = tf.transpose(x_w, [0, 2, 1, 3])

        y = tf.concat([x_h, x_w], axis=1)
        y = self.conv1(y)
        y = self.bn1(y)
        y = self.ac(y)

        x_h, x_w = tf.split(y, 2, axis=1)
        x_w = tf.transpose(x_w, [0, 2, 1, 3])

        x_h = sigmoid(self.conv2(x_h))
        x_w = sigmoid(self.conv3(x_w))

        y = residual * x_w * x_h
        print(np.shape(y))
        return y


if __name__ == '__main__':
    model = Coordinate_Attention(w=14, h=14, inp=512, oup=512)

    model.build((1, 14, 14, 512))
    model.summary()

This is the error message：

Traceback (most recent call last):
  File "E:/Python_Subject/Study_All/DeepLearning_Study/Other/Attention/Coordinate Attention.py", line 81, in <module>
    model.build((1, 14, 14, 512))
  File "E:\anaconda\envs\py3.6_tensorflow2.2\lib\site-packages\tensorflow\python\keras\engine\network.py", line 687, in build
    raise ValueError('You cannot build your model by calling `build` '
ValueError: You cannot build your model by calling `build` if your layers do not support float type inputs. Instead, in order to instantiate and build your model, `call` your model on real tensor data (of the correct dtype).

But when I change the activation function to coord_act it works fine,

self.ac = Activation(ac_swish)

I don’t know why

#2

You could use TF ops:

def ac_swish(x):
    temp = tf.math.minimum(tf.math.maximum(0.0, x + 3), 6) / 6
    x = x * temp
    return x
1 Like
#4

Yes，I know it works.But I don’t understand why it has to use tensorflow’s built-in functions

#5

Is that internally it is something like:

@tf.function
def test(a):
  return max(a,1)

a=2
test(tf.convert_to_tensor(a))
#6

You could check python max impl:

And:

1 Like
#8

If you are interested we are tracking this at:

1 Like
#10

Thanks for your reply :grinning: