TF1 vs TF2，相同的随机种子，不同的结果

如何解决TF1 vs TF2，相同的随机种子，不同的结果

我正在尝试将结果从 / 复制到 <template> <div v-for="content in slider" :key="content.id"> <img :src="content.imageUrl" /> </div> </template> <script> const images = []; export default { props: ['slider'],name: "ImageSlider",data() { return { index: 0,imageUrl: images[0] }; },methods: { next() { this.index = (this.index + 1) % images.length; this.imageUrl = images[this.index]; },autoChangeSlide() { setInterval(() => { this.next(); },3000); },},beforeMount() { this.autoChangeSlide(); },}; </script> 1.15 / 2.4.1。以下是我希望产生类似结果的 2 个示例。第一个示例使用 tf，第二个示例使用 tf.layers.conv2d()。我尝试使用 tf.keras.layers.Conv2D() 并且我也尝试不使用内核初始值设定项，仍然相同。 ================================================== ==========================

TF1

kernel_initializer=some_tf_initializer(seed=seed)

结果

import tensorflow as tf
import numpy as np


if __name__ == '__main__':
    seed = 1
    np.random.seed(seed)
    tf.set_random_seed(seed)
    session = tf.InteractiveSession()
    initializer = tf.initializers.orthogonal(1.0,seed)
    x = np.random.random((2,84,1))
    xp = tf.placeholder(x.dtype,x.shape)
    result = tf.layers.conv2d(xp,32,8,4,kernel_initializer=initializer)
    session.run(tf.global_variables_initializer())
    print(f'tf.layers.conv2d() output:\n{session.run([result],{xp: x})}\n{100 * "="}')
    print(f'x:\n{x}')

TF2

tf.layers.conv2d() output:
[array([[[[-4.92592295e-01,5.84946930e-01,6.45957303e-01,...,3.65328020e-01,4.03869755e-01,-1.06886940e+00],[ 5.37549724e-02,5.47611947e-01,3.91425858e-01,1.20702194e-01,9.21035825e-02,-1.10599980e+00],[-4.68301348e-01,8.54547125e-01,2.99738041e-01,-7.47656723e-02,-3.29838560e-01,-6.43979360e-01],[-3.61929553e-01,5.70854964e-01,1.49096153e-01,4.05001572e-01,-5.70764458e-01,-8.50054931e-01],[-5.23792632e-01,5.79328891e-01,4.36402398e-01,4.66264733e-01,-8.00687780e-02,-1.05017933e+00],[-9.59200260e-01,9.77023018e-01,7.22006476e-01,3.29199395e-01,-6.48041695e-02,-9.65927090e-01]],[[-3.05192775e-01,5.62304495e-01,2.66985564e-01,4.01866526e-01,-5.27853938e-01,-1.36294176e+00],[-3.33627733e-01,8.70909716e-01,-1.90685411e-01,5.82781510e-01,-9.54447222e-04,-5.93152081e-01],[ 1.05312097e-01,1.34995604e+00,4.18403345e-01,7.29398371e-01,-1.52729852e-01,-1.44087877e+00],[-2.65877698e-01,1.30702457e+00,2.12075863e-01,4.02235128e-01,-4.83213829e-02,-9.84811507e-01],[-2.33065665e-01,8.06240360e-01,2.04250988e-01,5.11907848e-01,-2.97176027e-01,-5.89994050e-01],[-6.06616196e-01,7.52128441e-01,6.05931022e-01,8.25940123e-01,-8.11965401e-01,-1.38039418e+00]],[[-1.14268620e-02,9.44562211e-01,7.30843200e-01,7.71483717e-01,-3.62299259e-01,-5.19022458e-01],[-1.25248650e-01,8.31288483e-01,3.01203507e-01,6.10087249e-02,-1.37678504e-01,-1.12305468e+00],[-2.37495350e-01,5.80913482e-01,2.48314393e-01,6.06333851e-01,-3.30380816e-01,-1.00515721e+00],[ 1.95822525e-01,9.85462620e-01,9.30753642e-02,5.07316387e-01,-2.72563623e-01,-3.44717273e-01],[-7.58518148e-01,7.18211754e-01,3.56392246e-01,5.41345861e-01,-2.84327606e-01,-5.78967512e-01],[-9.36528091e-01,4.51066108e-01,3.60469945e-01,6.06252149e-01,-9.82606865e-02,-1.12673980e+00]],[[-3.23906425e-01,4.89133765e-01,5.51718357e-01,-1.33718077e-01,1.23015200e-03,-5.54514943e-01],[-5.04556877e-01,7.36618066e-01,2.02579467e-01,-3.74142562e-01,6.00494771e-03,-1.31111289e+00],[-4.16236220e-01,5.98196128e-01,6.97848461e-01,6.23122747e-01,-1.07252840e-01,-4.83301913e-01],[ 1.06692401e-01,1.73388947e+00,3.22292122e-01,-1.75720933e-01,-2.96764992e-01,-1.06044579e+00],[-5.62972482e-01,1.36246077e+00,9.13179694e-01,6.32824517e-01,-4.73498606e-01,-9.47552266e-01],[-5.11697389e-01,6.66774542e-01,3.82947609e-01,5.67110785e-01,-7.11579248e-03,-7.77461939e-01]],[[-1.11410557e-01,1.10325702e+00,5.50558807e-01,5.51697720e-01,-3.75442814e-01,-4.68653786e-01],[-3.26447172e-01,1.19972335e+00,3.82313314e-01,2.20346417e-01,-3.39678412e-01,-8.66913575e-01],[-3.98474415e-01,7.14095329e-01,8.04559140e-02,1.19488448e-01,-2.52660129e-01,-9.91879947e-01],[-5.70385227e-01,8.52797253e-01,4.66251675e-01,-3.24400644e-01,3.20423484e-03,-1.39903236e+00],[-2.07675682e-01,6.29120258e-01,6.79053796e-01,1.07336154e-01,1.87886060e-01,-3.86794041e-01],[-2.67700849e-02,4.49128504e-01,3.40760390e-01,6.36000637e-01,-3.77890278e-01,-3.30207391e-01]],[[-3.10700139e-01,8.18823907e-01,3.67007768e-03,4.90558846e-01,-8.36860336e-01,-1.03869365e+00],[-4.88181365e-01,8.00080028e-01,3.56097390e-01,4.76738039e-01,-3.42056012e-01,-5.30593649e-01],[-4.60314405e-01,7.07439805e-01,4.22496599e-01,5.05574451e-01,-1.24705048e-01,-2.87646769e-01],[-2.57950491e-01,1.20126622e+00,5.56477074e-02,6.11127028e-01,-1.93377726e-01,-1.07765356e+00],[-6.50455755e-01,1.29336304e+00,8.44637456e-01,1.81637465e-01,-2.88107846e-01,-4.49444781e-01],[-4.21331048e-01,3.20901642e-01,8.83963968e-01,8.65426315e-01,-4.92428131e-01,-8.49800993e-01]]],[[[-5.16387221e-01,6.99757393e-01,7.85561831e-01,1.03644383e-01,-3.70012470e-01,-5.32289379e-01],[-7.05045607e-02,6.10474017e-01,3.49680420e-01,7.92201453e-01,-5.39308419e-01,-3.42154387e-01],[-4.22583634e-01,8.01291482e-01,1.48925846e-01,7.68131504e-01,-7.38695500e-01,-1.09516989e+00],[-7.26389962e-01,6.26188865e-01,5.42793169e-01,3.52803865e-01,-3.46388144e-01,-7.60162696e-01],[-9.04680334e-02,1.75359623e-01,8.00145598e-01,2.51379785e-01,-2.79905131e-01,-8.72223633e-01],[-4.20990087e-01,3.09422635e-01,4.11481559e-01,3.69259084e-02,-1.60866470e-01,-5.05859647e-01]],[[-4.38932989e-01,9.72496331e-01,2.50341307e-01,2.83248632e-01,-3.26580435e-01,-1.03773839e+00],[-8.19208455e-01,8.96083502e-01,4.24397325e-01,-5.64658536e-02,-6.09974865e-01,-1.23144756e+00],[-9.45132686e-01,9.13248242e-01,8.86299832e-01,2.15502049e-01,-5.58705913e-01,-2.25369791e-01],[-8.56722975e-01,7.94114365e-01,5.81065297e-01,4.96281428e-01,-1.03129758e+00,-1.05092158e+00],[-2.82233182e-02,9.48533077e-01,6.68768609e-01,5.78499983e-01,-6.95651561e-01,-9.76414384e-01],[-8.06302266e-01,1.05147010e+00,4.22578075e-01,2.94475123e-01,1.74168406e-01,-1.14952206e+00]],[[-3.06847178e-01,9.29719098e-01,2.68849689e-01,3.96790007e-01,-3.59629268e-01,-9.25076133e-01],[-2.38372207e-01,5.78896860e-01,2.17575482e-01,-6.97642069e-02,-2.63976905e-01,-9.10845525e-01],[-5.40468423e-01,1.20147694e+00,4.04620974e-01,1.76649501e-01,1.46967870e-01,-6.74134783e-01],[-6.27551970e-01,5.05884731e-01,2.10479188e-01,3.09780840e-01,1.46682047e-01,-1.07307698e+00],[-6.44653887e-01,8.90779216e-01,3.82747674e-01,1.72155001e-01,-4.48190676e-01,-8.62522695e-01],[-1.50221285e-01,9.35658435e-01,3.59703911e-01,-8.34232530e-02,-1.89649715e-01,-7.73705172e-01]],[[-3.47146995e-01,6.34597952e-02,4.73544353e-01,-8.20012972e-02,-1.78752555e-01,-1.36296041e+00],[-8.95298221e-01,8.66479595e-01,5.61457267e-01,2.22826074e-01,-4.86448503e-01,-8.56547191e-01],[-8.62118822e-01,4.83077034e-01,3.60908309e-02,3.29259093e-01,-4.08073639e-02,-6.45683881e-01],[-5.01558985e-01,4.62753228e-01,4.01966678e-01,6.35652593e-01,-7.45519465e-02,-5.39741380e-01],[ 1.29982837e-01,7.31941977e-01,-2.05750614e-01,3.16325608e-01,-3.28495177e-01,-1.09927128e+00],[-4.90335504e-01,4.94757973e-01,9.40801327e-02,5.66634378e-02,-7.30613447e-01,-7.25730750e-01]],[[-2.75066335e-01,7.35209409e-01,5.99839688e-01,6.71254510e-02,7.97677772e-02,-7.72461196e-01],[-1.58777502e-01,7.51857910e-01,3.80584693e-01,-1.01868390e-01,-7.12568409e-02,-6.42932271e-01],[-1.45951405e-01,1.03692197e+00,4.91333873e-01,2.98796942e-01,-5.46416283e-01,-1.04881221e+00],[-9.75652891e-03,8.28500896e-01,3.48450207e-01,4.07241092e-01,-2.34265134e-01,-5.27081486e-01],[-5.65917326e-01,7.37496827e-01,1.65005917e-01,6.61606291e-01,-2.20420580e-01,-1.11307865e+00],[-2.21167732e-01,4.83734785e-01,4.59793140e-01,4.19304500e-01,-2.96987262e-01,-2.01353157e-01]],[[-6.33535626e-01,1.11787318e+00,6.41380845e-01,1.12652086e-01,-1.99377096e-02,-6.69645437e-01],[-1.05662073e-01,3.87718948e-01,4.30142659e-01,3.27637078e-01,-3.59568870e-01,-9.63431155e-01],[-1.26793132e-01,1.29664648e+00,3.28922837e-01,3.21313848e-01,-7.53446525e-01,-7.93674733e-01],[-5.32937597e-01,1.09915270e+00,6.23443352e-01,9.96585500e-01,-6.21343220e-01,-1.01232184e+00],[ 5.05322966e-02,1.18874480e+00,4.57358272e-01,4.80935716e-01,-2.04122013e-01,-1.13864994e+00],[-1.03084733e-01,1.14916096e+00,2.73508528e-01,6.76093153e-01,-3.34324702e-01,-1.28436283e+00]]]])]
====================================================================================================
x:
[[[[4.17022005e-01]
   [7.20324493e-01]
   [1.14374817e-04]
   ...
   [6.23672207e-01]
   [7.50942434e-01]
   [3.48898342e-01]]

  [[2.69927892e-01]
   [8.95886218e-01]
   [4.28091190e-01]
   ...
   [1.85762022e-02]
   [7.00221437e-02]
   [4.86345111e-01]]

  [[6.06329462e-01]
   [5.68851437e-01]
   [3.17362409e-01]
   ...
   [9.18601778e-01]
   [4.02024891e-04]
   [9.76759149e-01]]

  ...

  [[5.89549934e-01]
   [3.89137609e-01]
   [5.05975232e-01]
   ...
   [4.35888475e-01]
   [7.89075202e-01]
   [4.66467704e-01]]

  [[6.73554921e-01]
   [8.84836452e-01]
   [9.38138449e-01]
   ...
   [7.93970466e-01]
   [2.13784215e-01]
   [6.41105035e-01]]

  [[7.31134736e-01]
   [9.50619892e-02]
   [7.00729238e-02]
   ...
   [9.95522026e-01]
   [4.81429517e-01]
   [8.37812754e-01]]]


 [[[6.03655452e-01]
   [6.64374944e-01]
   [2.72461392e-01]
   ...
   [1.14069927e-01]
   [2.93705095e-01]
   [8.78904978e-03]]

  [[2.53263696e-01]
   [8.37712781e-01]
   [8.07756027e-01]
   ...
   [7.37152445e-01]
   [6.00521471e-01]
   [7.37999367e-01]]

  [[3.75760828e-01]
   [9.11106703e-01]
   [8.72308594e-01]
   ...
   [9.80268932e-01]
   [1.13198035e-01]
   [4.65678949e-01]]

  ...

  [[4.94241516e-02]
   [6.34450548e-01]
   [8.93053413e-01]
   ...
   [7.97760317e-01]
   [3.18871974e-01]
   [6.47314782e-01]]

  [[4.65136696e-01]
   [5.07669096e-01]
   [4.23295851e-01]
   ...
   [2.98177206e-01]
   [5.32380132e-01]
   [6.12348886e-01]]

  [[2.58528146e-01]
   [5.20561003e-02]
   [7.82628170e-01]
   ...
   [8.26242775e-03]
   [7.43071396e-01]
   [3.29652868e-01]]]]

结果

import tensorflow as tf
import numpy as np
from tensorflow.keras.layers import Conv2D


if __name__ == '__main__':
    seed = 1
    np.random.seed(seed)
    tf.random.set_seed(seed)
    x = np.random.random((2,1))
    initializer = tf.initializers.Orthogonal(1.0,seed)
    print(f'tf.keras.layers.Conv2D() output:\n{Conv2D(32,kernel_initializer=initializer)(x)}\n{100 * "="}')
    print(f'x:\n{x}')