Why a Net inputted with same data could have different output?

1

I have a Net(gluon.Block) network already trained by the training loops

I am using following code to feed the data into the network

net = Net()
# training steps finished here
data = [data1, data2]
output = net(data)

With some same data1,data2, the net could give the different results! what are the possible reasons? I use np.argmax to get the maximum softmax of the output, and each time with the same input it could give the different output

The data batch_size of my data is 15, and data1 has 15 same records

[[0.5579492 0.23680237 0. … 0.20596384 1.0790483 0.11396025]
[0.5579492 0.23680237 0. … 0.20596384 1.0790483 0.11396025]
[0.5579492 0.23680237 0. … 0.20596384 1.0790483 0.11396025]
…
[0.5579492 0.23680237 0. … 0.20596384 1.0790483 0.11396025]
[0.5579492 0.23680237 0. … 0.20596384 1.0790483 0.11396025]
[0.5579492 0.23680237 0. … 0.20596384 1.0790483 0.11396025]]
<NDArray 15x2048 @cpu(0)>

data2 has 5 pieces, each has 3 same records

[[[-0.1518 0.38409 0.8934 … -0.27123 0.22157 0.92112 ]
[-0.54264 0.41476 1.0322 … -1.2969 0.76217 0.46349 ]
[ 0.085703 -0.22201 0.16569 … -0.074273 0.75808 -0.34243 ]
…
[ 0. 0. 0. … 0. 0. 0. ]
[ 0. 0. 0. … 0. 0. 0. ]
[ 0. 0. 0. … 0. 0. 0. ]]

[[-0.1518 0.38409 0.8934 … -0.27123 0.22157 0.92112 ]
[-0.54264 0.41476 1.0322 … -1.2969 0.76217 0.46349 ]
[ 0.085703 -0.22201 0.16569 … -0.074273 0.75808 -0.34243 ]
…
[ 0. 0. 0. … 0. 0. 0. ]
[ 0. 0. 0. … 0. 0. 0. ]
[ 0. 0. 0. … 0. 0. 0. ]]

[[-0.1518 0.38409 0.8934 … -0.27123 0.22157 0.92112 ]
[-0.54264 0.41476 1.0322 … -1.2969 0.76217 0.46349 ]
[ 0.085703 -0.22201 0.16569 … -0.074273 0.75808 -0.34243 ]
…
[ 0. 0. 0. … 0. 0. 0. ]
[ 0. 0. 0. … 0. 0. 0. ]
[ 0. 0. 0. … 0. 0. 0. ]]

…

[[-0.54264 0.41476 1.0322 … -1.2969 0.76217 0.46349 ]
[-0.038194 -0.24487 0.72812 … -0.1459 0.8278 0.27062 ]
[ 0.38709 0.32629 0.64524 … -0.8935 0.26669 -0.61397 ]
…
[ 0.18599 0.37305 0.13079 … -0.48638 1.0193 0.13099 ]
[ 0.31039 0.64859 0.28481 … -0.88554 0.91767 -0.57253 ]
[ 0.40367 0.35096 -0.18594 … -0.44149 0.14828 -0.068031]]

[[-0.54264 0.41476 1.0322 … -1.2969 0.76217 0.46349 ]
[-0.038194 -0.24487 0.72812 … -0.1459 0.8278 0.27062 ]
[ 0.38709 0.32629 0.64524 … -0.8935 0.26669 -0.61397 ]
…
[ 0.18599 0.37305 0.13079 … -0.48638 1.0193 0.13099 ]
[ 0.31039 0.64859 0.28481 … -0.88554 0.91767 -0.57253 ]
[ 0.40367 0.35096 -0.18594 … -0.44149 0.14828 -0.068031]]

[[-0.54264 0.41476 1.0322 … -1.2969 0.76217 0.46349 ]
[-0.038194 -0.24487 0.72812 … -0.1459 0.8278 0.27062 ]
[ 0.38709 0.32629 0.64524 … -0.8935 0.26669 -0.61397 ]
…
[ 0.18599 0.37305 0.13079 … -0.48638 1.0193 0.13099 ]
[ 0.31039 0.64859 0.28481 … -0.88554 0.91767 -0.57253 ]
[ 0.40367 0.35096 -0.18594 … -0.44149 0.14828 -0.068031]]]
<NDArray 15x12x100 @cpu(0)>

2

Does the dropout has an effect on this?

3

Yes dropout has an effect on the result of your network on the same input data. That’s because there is randomness associated with whether a particular weight is zeroed out or not. You can use a random number seed to see effectively make your network deterministic.