Week 1: Just go with the (Tensor)Flow

Feb 20, 2019

I was thrown a bit of a curveball before this project began, as my external advisor had visa issues, and could not return until Friday. As this was meant to be the learning week, I was given a couple of resources to look at – namely a Keras tutorial and a set of MIT lectures to watch. From Monday to Thursday, I spent time going through some of the lectures, getting a hold of some of the machine learning theory, and reading a bit into using Keras. I also crash coursed myself on python and read a bit about NVIDIA digits, an alternative to Keras that I may or may not end up using. I’ll get into that later.

Friday was when the gears started turning. It turned out that I could run some low level AI off my own laptop, at least for learning purposes. As such I was tasked with getting Keras and TensorFlow (a Keras dependency) working by next week (which I did over the weekend; shoutout to Colin for helping me out). To test it, my advisor gave me a basic classification follow-along exercise for TensorFlow [https://www.tensorflow.org/tutorials/keras/basic_classification] and a simple model follow-along exercise for Keras [https://www.tensorflow.org/guide/keras#eager_execution]. On top of these was a similar exercise for learning about overfitting and underfitting, and (more importantly) seeing them in practice [https://www.tensorflow.org/tutorials/keras/overfit_and_underfit].

We also had a discussion on digits vs keras. For now, we were using Keras because a) I could run it easily on my computer and b) it was lightweight and easier for learning. Digits on the other hand, is neither of those – but has the advantage of being able to easily visualize the AI learning process. It is usable on the several GPU workstations in the building, but it has its own learning curve, so we still need to discuss which we will go through with.

In regards to next week, the plan is to keep familiarizing myself with this environment and (if it rains) collect some data on which we will train and test the AI.

(For those commenting – how is this style? Should I get more or less technical with my explanations? Was it too long-winded or too brief? Let me know. Thanks)

4 Replies to “Week 1: Just go with the (Tensor)Flow”

  1. Aarushi N. says:

    I like this style because although it’s on the more technical side, you do a really good job of explaining it to the reader.

    Do you think you would be using digits later on in your project as you start processing more data? Or, would Keras be enough?

    I imagine you might just use stock photos and whatnot if it doesn’t rain while you’re collecting data.

    1. Travis S. says:

      There isn’t any real difference between the two functionally, it’s just digits has a built in visualizer for the learning process. In all likelihood I will end up using it.

      Data shouldn’t be an issue as it’s supposed to rain between Feb 25th and Feb 27th.

  2. Alex Y. says:

    A very detailed blog. I like the way you presented your problems to the readers. If you can post the specs of your workstation, I feel like it would be better for readers to help you decide which program you should be running or whether the processing power would be enough.

    Looking forward to more of these problems/conflicts we can solve together.

  3. alexisjefferson says:

    I really enjoy the title of this post. I think 3 detailed paragraphs make for a good blog post. I thought you gave a good amount of information and focused on struggles as well, which shows me how you were working.

Leave a Reply

Your email address will not be published.