7 ms·

> in reality, much expert help is needed to make it work. I totally agree. Having been taking classes and read ML/CNN articles online, this Google Cloud blog p

by devy 10y ago

> in reality, much expert help is needed to make it work. I totally agree. Having been taking classes and read ML/CNN articles online, this Google Cloud blog post made this project seem to be significantly easier than the actual effort that went into it. In particular, the blog simply covered the secret sauce to make it work in one line: > Makoto used the sample TensorFlow code Deep MNIST for Experts with > minor modifications to the convolution, pooling and last layers, > changing the network design to adapt to the pixel format of cucumber > images and the number of cucumber classes. I doubt this is a one man operation in just a few months of time, considering the computing power that's required to accomplish such feat (which it was said to take 2-3 days for each complete training set). Let alone the fact that they admit 7000 cucumber images are NOT enough to train the model to achieve the typical 95% accuracy that the original MNIST model it was based off. Although promising, this blog post reads more like a PR campaign for GCP's ML offering (with the custom hardware devices to do the compute, perhaps TPUs?)

Eridrus 10y ago
I just wish they'd move their CloudML thing out of alpha, I've given AWS too much of my money for their shitty GPUs.
- jonbarker 10y ago
  What will the pricing for CloudML be? It will be difficult to bootstrap anything without a free tier of some sort.
  - Eridrus 10y ago
    I would be super surprised if it wasn't at least cost competitive with AWS/Azure GPU instances. They didn't approve my request for alpha access though, so no real insight on my end.
    - jonbarker 10y ago
      What can you guess are their criteria for approval? I applied as well.
      - Eridrus 10y ago
        Probably how likely you are to spend a significant chunk of change on their cloud offering
dgacmu 10y ago
I disagree. One of the first projects I tried in deep learning was to tweak the LeNet-5 style MNIST example to handle 64x64 color images. It was pretty straightforward. Pretty straightforward doesn't mean it took ten minutes; but it didn't take more than a couple of hours, for someone clueless about DNN models. If you look at: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/models/image/mnist/convolutional.py https://github.com/tensorflow/tensorflow/blob/master/tensorf... you could actually manage it just by changing some of the image size and classes constants. In earlier versions of that code, you'd have had to manually change a few constants that weren't extracted well. This is approximately easy, once you know your way around the python side of TensorFlow. (Which itself takes some work, but it's not rocket science.) A better approach would probably add a third convolutional layer to continue to shrink the 80x80 input down to something easier to handle in the fully connected layer. What's surprised me is actually how slow it is, knowing nothing about their model. My guess is they could speed it up by adding that extra convolutional layer to reduce the amount of work done in the final FC layers. The fact that they're that slow suggests the opposite of what you're noting -- I doubt they had serious assistance on the machine learning side. I'm more impressed by the amount of time it must have taken to get the physical parts working and to collect 7,000 training images! Training MNIST itself is trivial. You can get > 99% accuracy in only a few minutes of training using a single GPU. This model should be at most 10-20x slower than MNIST (but, again, that FC layer would scale badly). If you look at the stackoverflow tags for tensorflow, you'll see a lot of very similar problems being solved. ("I want to reuse model X on problem Y, where such and such is different"). TPUs are awesome, but you don't need them for this problem. You could run the inference at 1 cucumber per second on a laptop. And you could train it on a Titan X in a matter of hours at most. (Source: I was and am still slightly on the Google Brain team, as a non-machine-learning person.)
blazespin 10y ago
Hardly.. the guy is only getting 70% accuracy. The hard part is the assembly line and he probably has a background in that.

by devy 10y ago

Eridrus 10y ago

I just wish they'd move their CloudML thing out of alpha, I've given AWS too much of my money for their shitty GPUs.

jonbarker 10y ago
What will the pricing for CloudML be? It will be difficult to bootstrap anything without a free tier of some sort.
- Eridrus 10y ago
  I would be super surprised if it wasn't at least cost competitive with AWS/Azure GPU instances. They didn't approve my request for alpha access though, so no real insight on my end.
  - jonbarker 10y ago
    What can you guess are their criteria for approval? I applied as well.
    - Eridrus 10y ago
      Probably how likely you are to spend a significant chunk of change on their cloud offering

dgacmu 10y ago

I disagree. One of the first projects I tried in deep learning was to tweak the LeNet-5 style MNIST example to handle 64x64 color images. It was pretty straightforward. Pretty straightforward doesn't mean it took ten minutes; but it didn't take more than a couple of hours, for someone clueless about DNN models. If you look at: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/models/image/mnist/convolutional.py https://github.com/tensorflow/tensorflow/blob/master/tensorf... you could actually manage it just by changing some of the image size and classes constants. In earlier versions of that code, you'd have had to manually change a few constants that weren't extracted well. This is approximately easy, once you know your way around the python side of TensorFlow. (Which itself takes some work, but it's not rocket science.) A better approach would probably add a third convolutional layer to continue to shrink the 80x80 input down to something easier to handle in the fully connected layer. What's surprised me is actually how slow it is, knowing nothing about their model. My guess is they could speed it up by adding that extra convolutional layer to reduce the amount of work done in the final FC layers. The fact that they're that slow suggests the opposite of what you're noting -- I doubt they had serious assistance on the machine learning side. I'm more impressed by the amount of time it must have taken to get the physical parts working and to collect 7,000 training images! Training MNIST itself is trivial. You can get > 99% accuracy in only a few minutes of training using a single GPU. This model should be at most 10-20x slower than MNIST (but, again, that FC layer would scale badly). If you look at the stackoverflow tags for tensorflow, you'll see a lot of very similar problems being solved. ("I want to reuse model X on problem Y, where such and such is different"). TPUs are awesome, but you don't need them for this problem. You could run the inference at 1 cucumber per second on a laptop. And you could train it on a Titan X in a matter of hours at most. (Source: I was and am still slightly on the Google Brain team, as a non-machine-learning person.)

blazespin 10y ago

Hardly.. the guy is only getting 70% accuracy. The hard part is the assembly line and he probably has a background in that.