Why Tensorflow results are different in Python versions 3.5 and 3.7

问题

Why Tensorflow results are different in Python versions 3.5 (SQL server, machine learning services) and 3.7 (local machine, anaconda)?

I found out, it depends on 4 parameters values:

dataset size 
number of epochs 
number of 1st layer (input) neurons 
number of 2nd layer (hidden) neurons

Here is the example: identical results:

dataset size - 50 000
number of epochs - 5/3/2
number of 1st layer (input) neurons  - 300
**number of 2nd layer (hidden) neurons - 80% from 1st layer**

different results:

dataset size - 50 000
number of epochs - 6/5/3/2
number of 1st layers neurons - 300
**number of 2nd layer (hidden) neurons  - 40% from 1st layer**

Seems that the problem is in the 4th parameter, but:

dataset size - 50 000
**number of epochs - 6!**
number of 1st layer (input) neurons  - 300
number of 2nd layer (hidden) neurons - 80% from 1st layer

Gives different results. By decreasing the other 3 parameters values it is possible to achieve identical results, but it is uncomfortable.

Package versions are identical on both the platforms:

2.2.0 tensorflow
0.24.2 pandas 
2.3.1 keras
1.17.5 numpy

Datasets are also absolutely identical.

The code:

#general libraries
import pandas as pd
import numpy as np 
from sklearn.model_selection import train_test_split

#neural network
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense
from keras.optimizers import adam
from sklearn.preprocessing import MinMaxScaler 
from sklearn.metrics import confusion_matrix

np.random.seed(7)
import tensorflow as tf
tf.random.set_seed(7) 
import random
random.seed(7)


layer1neurons = 300
layer2neurons_share = .4
loss="binary_crossentropy"
learning_rate = 0.001
optimizer=tf.keras.optimizers.Adam(lr=learning_rate)
metrics=["accuracy"]
epochs = 6
batch_size = 32
verbose = 1

model = Sequential()
model.add(Dense(layer1neurons, input_dim=len(X_train.columns), activation="relu"))  
model.add(Dense(round(layer1neurons * layer2neurons_share, 0), activation="relu"))
model.add(Dense(1, activation="sigmoid")) 

model.compile(loss=loss,  
              optimizer=optimizer, 
              metrics=metrics) 

model.fit(X_train, y_train, 
          epochs = epochs, 
          batch_size=batch_size, 
          verbose = verbose)

scores = model.evaluate(X_train, y_train)

来源：https://stackoverflow.com/questions/62128556/why-tensorflow-results-are-different-in-python-versions-3-5-and-3-7