You now can train a really good image classification model on 25k images in less than 2 hours, entirely for free, on a Colab GPU. And it's dead easy. What a time to be alive I remember when you had to write your own CUDA and C++ and you needed a hardware budget
-
Show this thread
-
Here's a Colab that does just that. Download a zip file with 25k images. Get a model training in seconds, with all the best practices baked in. One hour and a half to convergence on the free Colab GPU.https://colab.research.google.com/drive/1umJnCp8tZ7UDTYSQsuWdKRhqbHts38AC …
14 replies 129 retweets 764 likesShow this thread -
Replying to @fchollet
One thing about “from dir”. Is there a way to enforce an ordering of the files before train/Val split. E.g. you don’t want to totally randomize esp if u know some images are “correlated” and u don’t want them to end up in train and val separately to inflate accuracy.
2 replies 0 retweets 0 likes
shuffle=False will give you alphanumerical order, but then of course samples from each class will be contiguous segments in the data.
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.