Cannot get mnist example working
Opened this issue · 6 comments
I am trying to get the mnist example working.
I have got all the tests passing and am using data downloaded from kaggle.
The output when I run it:
stack runghc examples/main/mnist.hs data/train.csv data/test.csv
Training convolutional neural network...
Failed reading: takeWhile1
Have I got the right dataset? Am I just missing something or just being stupid?
It seems that those CSV files on Kaggle contain an extra (and patently silly) “header” line
label,1x1,1x2,1x3,1x4,1x5,1x6,1x7,1x8,1x9,1x10,1x11,1x12,1x13,1x14,1x15,1x16,1x17,1x18,1x19,1x20,1x21,1x22,1x23,1x24,1x25,1x26,1x27,1x28,2x1,2x2,2x3,2x4,2x5,2x6,2x7,2x8,2x9,2x10,2x11,2x12,2x13,2x14,2x15,2x16,2x17,2x18,2x19,2x20,2x21,2x22,2x23,2x24,2x25,2x26,2x27,2x28,3x1,3x2,3x3,3x4,3x5,3x6,3x7,3x8,3x9,3x10,3x11,3x12,3x13,3x14,3x15,3x16,3x17,3x18,3x19,3x20,3x21,3x22,3x23,3x24,3x25,3x26,3x27,3x28,4x1,4x2,4x3,4x4,4x5,4x6,4x7,4x8,4x9,4x10,4x11,4x12,4x13,4x14,4x15,4x16,4x17,4x18,4x19,4x20,4x21,4x22,4x23,4x24,4x25,4x26,4x27,4x28,5x1,5x2,5x3,5x4,5x5,5x6,5x7,5x8,5x9,5x10,5x11,5x12,5x13,5x14,5x15,5x16,5x17,5x18,5x19,5x20,5x21,5x22,5x23,5x24,5x25,5x26,5x27,5x28,6x1,6x2,6x3,6x4,6x5,6x6,6x7,6x8,6x9,6x10,6x11,6x12,6x13,6x14,6x15,6x16,6x17,6x18,6x19,6x20,6x21,6x22,6x23,6x24,6x25,6x26,6x27,6x28,7x1,7x2,7x3,7x4,7x5,7x6,7x7,7x8,7x9,7x10,7x11,7x12,7x13,7x14,7x15,7x16,7x17,7x18,7x19,7x20,7x21,7x22,7x23,7x24,7x25,7x26,7x27,7x28,8x1,8x2,8x3,8x4,8x5,8x6,8x7,8x8,8x9,8x10,8x11,8x12,8x13,8x14,8x15,8x16,8x17,8x18,8x19,8x20,8x21,8x22,8x23,8x24,8x25,8x26,8x27,8x28,9x1,9x2,9x3,9x4,9x5,9x6,9x7,9x8,9x9,9x10,9x11,9x12,9x13,9x14,9x15,9x16,9x17,9x18,9x19,9x20,9x21,9x22,9x23,9x24,9x25,9x26,9x27,9x28,10x1,10x2,10x3,10x4,10x5,10x6,10x7,10x8,10x9,10x10,10x11,10x12,10x13,10x14,10x15,10x16,10x17,10x18,10x19,10x20,10x21,10x22,10x23,10x24,10x25,10x26,10x27,10x28,11x1,11x2,11x3,11x4,11x5,11x6,11x7,11x8,11x9,11x10,11x11,11x12,11x13,11x14,11x15,11x16,11x17,11x18,11x19,11x20,11x21,11x22,11x23,11x24,11x25,11x26,11x27,11x28,12x1,12x2,12x3,12x4,12x5,12x6,12x7,12x8,12x9,12x10,12x11,12x12,12x13,12x14,12x15,12x16,12x17,12x18,12x19,12x20,12x21,12x22,12x23,12x24,12x25,12x26,12x27,12x28,13x1,13x2,13x3,13x4,13x5,13x6,13x7,13x8,13x9,13x10,13x11,13x12,13x13,13x14,13x15,13x16,13x17,13x18,13x19,13x20,13x21,13x22,13x23,13x24,13x25,13x26,13x27,13x28,14x1,14x2,14x3,14x4,14x5,14x6,14x7,14x8,14x9,14x10,14x11,14x12,14x13,14x14,14x15,14x16,14x17,14x18,14x19,14x20,14x21,14x22,14x23,14x24,14x25,14x26,14x27,14x28,15x1,15x2,15x3,15x4,15x5,15x6,15x7,15x8,15x9,15x10,15x11,15x12,15x13,15x14,15x15,15x16,15x17,15x18,15x19,15x20,15x21,15x22,15x23,15x24,15x25,15x26,15x27,15x28,16x1,16x2,16x3,16x4,16x5,16x6,16x7,16x8,16x9,16x10,16x11,16x12,16x13,16x14,16x15,16x16,16x17,16x18,16x19,16x20,16x21,16x22,16x23,16x24,16x25,16x26,16x27,16x28,17x1,17x2,17x3,17x4,17x5,17x6,17x7,17x8,17x9,17x10,17x11,17x12,17x13,17x14,17x15,17x16,17x17,17x18,17x19,17x20,17x21,17x22,17x23,17x24,17x25,17x26,17x27,17x28,18x1,18x2,18x3,18x4,18x5,18x6,18x7,18x8,18x9,18x10,18x11,18x12,18x13,18x14,18x15,18x16,18x17,18x18,18x19,18x20,18x21,18x22,18x23,18x24,18x25,18x26,18x27,18x28,19x1,19x2,19x3,19x4,19x5,19x6,19x7,19x8,19x9,19x10,19x11,19x12,19x13,19x14,19x15,19x16,19x17,19x18,19x19,19x20,19x21,19x22,19x23,19x24,19x25,19x26,19x27,19x28,20x1,20x2,20x3,20x4,20x5,20x6,20x7,20x8,20x9,20x10,20x11,20x12,20x13,20x14,20x15,20x16,20x17,20x18,20x19,20x20,20x21,20x22,20x23,20x24,20x25,20x26,20x27,20x28,21x1,21x2,21x3,21x4,21x5,21x6,21x7,21x8,21x9,21x10,21x11,21x12,21x13,21x14,21x15,21x16,21x17,21x18,21x19,21x20,21x21,21x22,21x23,21x24,21x25,21x26,21x27,21x28,22x1,22x2,22x3,22x4,22x5,22x6,22x7,22x8,22x9,22x10,22x11,22x12,22x13,22x14,22x15,22x16,22x17,22x18,22x19,22x20,22x21,22x22,22x23,22x24,22x25,22x26,22x27,22x28,23x1,23x2,23x3,23x4,23x5,23x6,23x7,23x8,23x9,23x10,23x11,23x12,23x13,23x14,23x15,23x16,23x17,23x18,23x19,23x20,23x21,23x22,23x23,23x24,23x25,23x26,23x27,23x28,24x1,24x2,24x3,24x4,24x5,24x6,24x7,24x8,24x9,24x10,24x11,24x12,24x13,24x14,24x15,24x16,24x17,24x18,24x19,24x20,24x21,24x22,24x23,24x24,24x25,24x26,24x27,24x28,25x1,25x2,25x3,25x4,25x5,25x6,25x7,25x8,25x9,25x10,25x11,25x12,25x13,25x14,25x15,25x16,25x17,25x18,25x19,25x20,25x21,25x22,25x23,25x24,25x25,25x26,25x27,25x28,26x1,26x2,26x3,26x4,26x5,26x6,26x7,26x8,26x9,26x10,26x11,26x12,26x13,26x14,26x15,26x16,26x17,26x18,26x19,26x20,26x21,26x22,26x23,26x24,26x25,26x26,26x27,26x28,27x1,27x2,27x3,27x4,27x5,27x6,27x7,27x8,27x9,27x10,27x11,27x12,27x13,27x14,27x15,27x16,27x17,27x18,27x19,27x20,27x21,27x22,27x23,27x24,27x25,27x26,27x27,27x28,28x1,28x2,28x3,28x4,28x5,28x6,28x7,28x8,28x9,28x10,28x11,28x12,28x13,28x14,28x15,28x16,28x17,28x18,28x19,28x20,28x21,28x22,28x23,28x24,28x25,28x26,28x27,28x28
This needs to be stripped before the parser can do its work; can be done with
diff --git a/examples/main/mnist.hs b/examples/main/mnist.hs
index 28be8cd..8805c8c 100644
--- a/examples/main/mnist.hs
+++ b/examples/main/mnist.hs
@@ -110,7 +110,7 @@ main = do
readMNIST :: FilePath -> ExceptT String IO [(S ('D2 28 28), S ('D1 10))]
readMNIST mnist = ExceptT $ do
mnistdata <- T.readFile mnist
- return $ traverse (A.parseOnly parseMNIST) (T.lines mnistdata)
+ return $ traverse (A.parseOnly parseMNIST) (tail $ T.lines mnistdata)
parseMNIST :: A.Parser (S ('D2 28 28), S ('D1 10))
parseMNIST = doNot sure if that's really sufficient to get it working though... for me, the process runs out of memory...
Yeah, it does seem to work now – when I reduce the size of the datasets, it starts and improves accuracy during the training. Unfortunately, it doesn't run in constant memory, which is a big problem (but another issue).
Fortunately, the memory leak actually seems to be specific to the Inception-style layers. With the simpler model from the README, it runs seemingly in constant space. Nice!
It seems that those CSV files on Kaggle contain an extra (and patently silly) “header” line
Header lines are a common feature for CSV fines and they are not silly.
@erikd I don't mean the fact that there is a header, but the XxY information it contains.
Anyway... this issue has apparently been amply addressed with #97. I'll do some cherry-picking there and maybe try to submit that in form of usable PRs...