> so I should run `OUTPUT fit__pipe INPUT` 2 or 3 more times.
Yes, I think so. After 2 or three more times, you should get all correct. 100%
accuracy.
>What does the other output mean? For example what is alternating 1 and 2,
> what is 1...20, what is 10?
There are 15 images. When we constructed the 'NNPipeline' class, we chose a
batch size of 10,
and an epoch count of 10 (epoch is one complete iteration through the whole
dataset).
So 15 / 10 = 1.5, which we round up to 2. So one epoch needs 2 iterations to
complete (with some images getting an extra pass).
As we selected 10 epochs, the total number of iterations is 2 * 10 = 20. 1 and
2 are the iterations for the current epoch.
For these small cases, this information is hardly useful. But, for example,
cifar-10 has 50,000 images, and batch sizes of 100 (in the model I made),
so one epoch is 500 iterations. It helps to know how many iterations have been
completed so far.
It might help to play around with the parameters of 'NNPipeline' constructor.
the learning rate 0.001 could be increased, for instance. I don't know whether
that would improve anything, but could be interesting.
Also, assuming you can use plot.
plot _3 (+/ % #) \ lossKeep__pipe
will give you a rolling average (of 3) of the loss during training. This is not
accuracy, but cross entropy loss. It should be a downward curve.
On Friday, April 19, 2019, 2:18:52 PM GMT+9, Brian Schott
<[email protected]> wrote:
Ok. This is what I got.
load jpath '~temp/simple_conv_test.ijs'
batchnorm
┌─┬─┬──────┬──┬────┐
│0│1│0.0001│10│0.01│
└─┴─┴──────┴──┴────┘
batchnorm
┌─┬─┬──────┬─┬────┐
│0│1│0.0001│5│0.01│
└─┴─┴──────┴─┴────┘
batchnorm
┌─┬─┬──────┬─┬────┐
│0│1│0.0001│2│0.01│
└─┴─┴──────┴─┴────┘
Added Conv2D. Network depth is 1.
Added PoolLayer. Network depth is 2.
Added Activation. Network depth is 3.
Added Conv2D. Network depth is 4.
Added Activation. Network depth is 5.
Added FlattenLayer. Network depth is 6.
Added SimpleLayer. Network depth is 7.
Added Activation. Network depth is 8.
OUTPUT fit__pipe INPUT
Iteration complete: 1, total: 1
Iteration complete: 2, total: 2
Iteration complete: 1, total: 3
Iteration complete: 2, total: 4
Iteration complete: 1, total: 5
Iteration complete: 2, total: 6
Iteration complete: 1, total: 7
Iteration complete: 2, total: 8
Iteration complete: 1, total: 9
Iteration complete: 2, total: 10
Iteration complete: 1, total: 11
Iteration complete: 2, total: 12
Iteration complete: 1, total: 13
Iteration complete: 2, total: 14
Iteration complete: 1, total: 15
Iteration complete: 2, total: 16
Iteration complete: 1, total: 17
Iteration complete: 2, total: 18
Iteration complete: 1, total: 19
Iteration complete: 2, total: 20
10
OUTPUT -:"1 1 (=>./)"1 >{: predict__pipe INPUT
1 0 1 1 1 1 0 0 0 1 0 0 0 0 1
NB. so this tells me that 8 times the prediction was wrong and 7 times
right.
NB. so I should run `OUTPUT fit__pipe INPUT` 2 or 3 more times.
What does the other output mean? For example what is alternating 1 and 2,
what is 1...20, what is 10?
Thanks,
--
(B=)
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
