Why is it slow for the first inference

General Discussion
,
#1

I have two simple network implementations:

import numpy as np
import tensorflow as tf
import time


class SimpleNet(tf.keras.Model):
    def __init__(self):
        super(SimpleNet, self).__init__()
        self.mod1 = tf.keras.Sequential([
            tf.keras.layers.Conv2D(64, (3, 3), padding="same"),
            tf.keras.layers.ReLU(),
        ])
        self.final = tf.keras.layers.Conv2D(3, (3, 3), padding="same")

    def call(self, x):
        x = self.mod1(x)
        x = self.final(x)
        return x


def simple_net():
    x = tf.keras.Input(shape=(256, 256, 3))

    y = tf.keras.Sequential([
        tf.keras.layers.Conv2D(64, (3, 3), padding="same"),
        tf.keras.layers.ReLU(),
    ])(x)

    y = tf.keras.layers.Conv2D(3, (3, 3), padding="same")(y)

    return tf.keras.Model(x, y)


if __name__ == "__main__":
    tf.compat.v1.disable_eager_execution()

    x_test = np.random.random((32, 256, 256, 3))
    x_test = tf.convert_to_tensor(x_test)
    print(x_test.shape)

    model1 = SimpleNet()
    model1.build(tf.TensorShape((None, 256, 256, 3)))

    model2 = simple_net()

    print(model1.summary())
    print(model2.summary())

    for i in range(4):
        s = time.time()
        pred = model1(x_test)
        e = time.time()
        print(f"time: {e-s} s")

    print("") 

    for i in range(4):
        s = time.time()
        pred = model2(x_test)
        e = time.time()
        print(f"time: {e-s} s")

The result is:

2022-07-14 11:32:45.087770: I tensorflow/core/util/util.cc:175] Experimental oneDNN custom operations are on. If you experience issues, please turn them off by setting the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
(32, 256, 256, 3)
Model: "simple_net"
_________________________________________________________________
 Layer (type)                Output Shape              Param #
=================================================================
 sequential (Sequential)     multiple                  1792

 conv2d_1 (Conv2D)           multiple                  1731

=================================================================
Total params: 3,523
Trainable params: 3,523
Non-trainable params: 0
_________________________________________________________________
None
Model: "model"
_________________________________________________________________
 Layer (type)                Output Shape              Param #
=================================================================
 input_1 (InputLayer)        [(None, 256, 256, 3)]     0

 sequential_1 (Sequential)   multiple                  1792

 conv2d_3 (Conv2D)           (None, 256, 256, 3)       1731

=================================================================
Total params: 3,523
Trainable params: 3,523
Non-trainable params: 0
_________________________________________________________________
None
time: 0.029999732971191406 s
time: 0.006798267364501953 s
time: 0.0066530704498291016 s
time: 0.00666046142578125 s

time: 0.006479501724243164 s
time: 0.006349802017211914 s
time: 0.006381988525390625 s
time: 0.006410121917724609 s

I feel that the first inference time of SimpleNet is a bit strange. The first time is significantly longer than the laters. Also, I notice that the first implementation is always slower than the second
So I want to ask for help.

  1. Do I write the code wrong? How to make the first inference as fast as others.
  2. Why the two implementations take different times (the first implementation slower the second)
#3

Seems like this is exactly the issue on mine.