When training the cnn_model, I get the following output:
Epoch: 0, Batch: 150, accuracy: 0.4985513 loss: 7.6424022
Epoch: 1, Batch: 163, accuracy: 0.4992854 loss: 7.6783161
Epoch: 2, Batch: 176, accuracy: 0.5000441 loss: 7.6865749
Epoch: 3, Batch: 139, accuracy: 0.4983259 loss: 7.6991839
Epoch: 4, Batch: 152, accuracy: 0.4988766 loss: 7.6995916
%{
"conv_0" => %{
"bias" => #Nx.Tensor<
f32[32]
EXLA.Backend<host:0, 0.1357844422.1979580433.82179>
[NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN]
>,
"kernel" => #Nx.Tensor<
f32[3][3][3][32]
EXLA.Backend<host:0, 0.1357844422.1979580433.82180>
[
[
[
[NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN],
[NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, ...],
...
],
...
],
...
]
>
},
"conv_1" => %{
"bias" => #Nx.Tensor<
f32[64]
EXLA.Backend<host:0, 0.1357844422.1979580433.82181>
[NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, -0.0071477023884654045, NaN, NaN, NaN, NaN, 0.0, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, ...]
>,
"kernel" => #Nx.Tensor<
f32[3][3][32][64]
EXLA.Backend<host:0, 0.1357844422.1979580433.82182>
[
[
[
[NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, ...],
...
],
...
],
...
]
>
},
"conv_2" => %{
"bias" => #Nx.Tensor<
f32[128]
EXLA.Backend<host:0, 0.1357844422.1979580433.82183>
[0.0, NaN, NaN, NaN, NaN, NaN, NaN, NaN, 0.005036031361669302, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, ...]
>,
"kernel" => #Nx.Tensor<
f32[3][3][64][128]
EXLA.Backend<host:0, 0.1357844422.1979580433.82184>
[
[
[
[NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, ...],
...
],
...
],
...
]
>
},
"dense_0" => %{
"bias" => #Nx.Tensor<
f32[128]
EXLA.Backend<host:0, 0.1357844422.1979580433.82185>
[NaN, -0.005992305930703878, -0.006005365401506424, -0.004664595704525709, NaN, NaN, NaN, -5.619042203761637e-4, 0.0, NaN, -0.005999671295285225, -6.131592726887902e-6, NaN, 0.0, NaN, 0.0, 0.0, NaN, NaN, -0.006002828478813171, -0.00600335793569684, 0.0, NaN, NaN, NaN, -0.006002923008054495, -0.006005282513797283, -0.00600528996437788, -0.0060048955492675304, -0.006004981696605682, NaN, -0.006004655733704567, -0.006005233619362116, NaN, -0.006004724185913801, -0.006005335133522749, -0.006005051080137491, -0.006004408933222294, NaN, -0.006005355156958103, 0.0, -0.006005344912409782, 0.0, NaN, -0.005991040728986263, ...]
>,
"kernel" => #Nx.Tensor<
f32[18432][128]
EXLA.Backend<host:0, 0.1357844422.1979580433.82186>
[
[NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN, ...],
...
]
>
},
"dense_1" => %{
"bias" => #Nx.Tensor<
f32[1]
EXLA.Backend<host:0, 0.1357844422.1979580433.82187>
[NaN]
>,
"kernel" => #Nx.Tensor<
f32[128][1]
EXLA.Backend<host:0, 0.1357844422.1979580433.82188>
[
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
[NaN],
...
]
>
}
}
The accuracy of the mlp_model was Batch: 6, accuracy: 0.5078125 and the accuracy of this cnn_model is Batch: 6, accuracy: 0.4944196 which was slightly worse instead of the expected “significantly better.”
I reviewed all the code to make sure I hadn’t missed anything, but I couldn’t find anything that didn’t match.
I’m guessing the NaNs in the trained model state are a problem, but I’m not sure how to fix that.