Problems arising after S4TF V0.3 -> v0.4

[F20170604]

Tutorial 3 Colab Link : https://colab.research.google.com/drive/1XfT75glapWdCp0Zj_4mkiMrjT6odD-Gc
Tutorial 5 Colab Link : https://colab.research.google.com/drive/13lBsht3Wa4GjKKkA47JCrd54XikhNX2E

I am facing Out Of Memory(OOM) while allocating tensor during training of classifier in Tutorial 5. I've tried 12 times and all of the time, it's happening after Epoch 4 out of 100.
Further, I couldn't get the packages to install. The error I am getting is 'toolchain is invalid'
It all started happening after S4TF update to v0.4

Following are the screenshots of the issue regarding OOM


Screenshot regarding Package Install

[F20170604]

OOM gets resolved after modifying the callAsFunction
I changed it to :

public func callAsFunction(_ input: Input) -> Output {
        var convolved1 = pool1(conv1a(input))
        var convolved2 = pool2(conv1b(convolved1))
        var convolved3 = pool3(conv1c(convolved2))
        var convolved4 = pool4(conv1d(convolved3))
        return layer1b(layer1a(flatten(convolved4)))
        
    }

[F20170604]

Package installation is still giving errors. I looked through fast.ai new notebooks and exactly copied the import commands. But I am still getting the same error.

[RahulBhalley]

Any updates on memory leak @rxwei? I am [also] still experiencing this issue in my own notebook where I was previously able to train my GAN successfully in v0.3 but now in v0.4 it just shows the error of "memory limits reached" as shown in this issue above.

Or @Ayush517 did you find some way to fix this issue in your own code?

[F20170604]

@RahulBhalley Earlier my network was this :

func call(_ input: Input) -> Output {
        var convolved1 = input.sequenced(through: conv1a, pool1)
        var convolved2 = convolved1.sequenced(through: conv1b, pool2)
        var convolved3 = convolved2.sequenced(through: conv1c, pool3)
        var convolved4 = convolved3.sequenced(through: conv1d, pool4)
        return convolved4.sequenced(through: flatten, layer1a, layer1b)
        
    }

After I changed to the following code, It started working fine

public func callAsFunction(_ input: Input) -> Output {
        var convolved1 = pool1(conv1a(input))
        var convolved2 = pool2(conv1b(convolved1))
        var convolved3 = pool3(conv1c(convolved2))
        var convolved4 = pool4(conv1d(convolved3))
        return layer1b(layer1a(flatten(convolved4)))
        
    }

[RahulBhalley]

Let's check. Fingers crossed🤞. I hope this be the only problem with sequenced method.

Edit: Nope, crashed again @rxwei. Please help me resolve this issue. ☹️

[rxwei]

Hi @RahulBhalley , sorry for the delay. Could you try to build a toolchain against the latest HEAD of the tensorflow branch? swiftlang/swift#25967 fixed the major issue and it could resolve your OOM problem. We'll release another 0.4 release candidate later today.

[RahulBhalley]

Sorry, I can't build because I refrain from heating up my MacBook Pro (recently got its battery replaced). I'll wait for newer 0.4 release to go live today on Google Colab (if I get you correctly 😃). Thanks for replying btw @rxwei.