tutorials on pitchfork data (ULMfit) #2
Conversation
datawrestler
left a comment
datawrestler
left a comment
There was a problem hiding this comment.
Overall, good start - add intro sections to both scripts, take advantage of headers to break things up, change the training process to iteratively unfreeze weights, possibly check out fastprogress, use relative paths, and never put keys/secrets in source code again.
| ], | ||
| "source": [ | ||
| "print(os.getcwd())\n", | ||
| "path='/media/jlealtru/data_files/github/Tutorials/TextAnalytics/pitchfork_data'" |
There was a problem hiding this comment.
Use relative paths - add either a standalone script that secures data from source or run it in an intro section, but show how to download the source data directly so all your steps can be rebuilt.
| "learn_classifier.freeze_to(-2)\n", | ||
| "lr /= 2\n", | ||
| "learn_classifier.fit_one_cycle(1, slice(lr/(2.6**4),lr), moms=(0.8,0.7))\n", | ||
| "#learn_classifier.fit_one_cycle(2, slice(1e-4/2,1e-2/2), moms=(0.8,0.7))" |
There was a problem hiding this comment.
Look at fastprogress - I think folks would find it really interesting to be abel to iteratively build a training graph as you progress.
| "metadata": {}, | ||
| "outputs": [], | ||
| "source": [ | ||
| "learn_classifier.unfreeze()\n", |
There was a problem hiding this comment.
The fastai folks recommend iteratively unfreezing layers sequentially. Start with -1, then -2, then -3, then unfreeze all. That will likely help out.
| "cell_type": "markdown", | ||
| "metadata": {}, | ||
| "source": [ | ||
| "In this tutorial we are going to implement a transfer learning model for text version of the ULMfit. \n", |
There was a problem hiding this comment.
Format this markdown and add TOC with hyperlinks and additional sources to review.
| "# username \n", | ||
| "os.environ['KAGGLE_USERNAME'] = \"jlealtru\" \n", | ||
| "# key\n", | ||
| "os.environ['KAGGLE_KEY'] = \"6c3a4d6b4d8e7804780d6cb02879ac53\"" |
There was a problem hiding this comment.
@jlealtru never post secrets/keys in source code. You have a couple options. The easiest, although not safest, is creating a separate file and import that file in and reference the variable name only in the code. Alternatively, you can leverage something like Azure Key Vault (easy to use, super powerful - think of OnePassword or LastPass except at scale/programatically)
| ], | ||
| "source": [ | ||
| "#learn.fit_one_cycle(10, 2e-3, moms=(0.8,0.7), wd=0.1)\n", | ||
| "learn_pitchfork.fit_one_cycle(12, 2e-3/3, moms=(0.8,0.7), wd= 0.1)" |
There was a problem hiding this comment.
again - iteratively unfreeze layers and train - track progress using something like fastprogress
2 participants
Adding tutorials on pitchfork data and some old code.