Fastai v2 tabular

sgugger · January 22, 2020, 8:50pm

Ok, changed test_dl to DataBunch.test_dl and in the case of tabular data it has a different behavior.

muellerzr · January 22, 2020, 8:55pm

Thanks!

muellerzr · January 22, 2020, 9:05pm

@sgugger I seem to be getting an error, I think this line in test_dl should be:

test_ds = test_set(self.valid_ds, test_items, rm_tfms=rm_type_tfms) if isinstance(self.valid_ds, DataSource) else test_items

As we no longer have dbunch (originally it was ifinstance(dbunch.valid_ds...)

sgugger · January 22, 2020, 9:08pm

Missed one indeed, should be fixed.

muellerzr · January 22, 2020, 9:21pm

Tabular is still showing the DataFrame object has no attribute with_cont. Should it be now:

dl = dbunch.test_dl(df)

If I look at the type of dbunch it’s just showing the generic DataBunch, (fastai2.data.core.DataBunch) which may be a hint of what is going on.
(also doing dbunch.test_dl?? shows the non-tabular version

(part of this may be because I’m doing to = TabularPandas(df, procs, cat_names, cont_names, y_names="salary", splits=splits) and then dbunch = to.dbunch instead of TabularDataBunch)

When I try going through and ensuring it is a TabDataLoader (start with TabularDataBunch.from_df) I still get the DataFrame has no attribute with_cont

sgugger · January 22, 2020, 10:50pm

Ah right, .databunch doesn’t use the subclass (yet). Will fix tonight or tomorrow.

muellerzr · January 22, 2020, 10:54pm

Thanks! Appreciate all the help

sgugger · January 22, 2020, 11:00pm

Should be fixed now.

muellerzr · January 23, 2020, 2:43pm

Can verify @sgugger! Also nice that we can have them be labeled too (assuming they can be labelled the same as the original dbunch) Thank you for this!

navneetkrch · February 3, 2020, 1:56pm

Hey Guys,
I am working on a dataset similar to sale prediction task.
I am getting this error when I want to use metrics like rmse, mse.

If I use mae it works but I need my metrics proportional to the cost.
GIST for error using rmse
Execution using MAE

Javi · February 3, 2020, 2:17pm

I got a similar error (fastai2 0.0.7). I think that the dataloader does not encode the data.

On:

I Get

muellerzr · February 3, 2020, 2:18pm

How are you guys making your DataLoaders? I just used tabular the other day and it was working fine

Javi · February 3, 2020, 2:19pm

I use this code as reference:

github.com

muellerzr/fastai2-Tabular-Baselines/blob/master/Tabular_fastai_Baseline.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "Tabular fastai Baseline.ipynb",
      "provenance": [],
      "machine_shape": "hm"
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "accelerator": "GPU"
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "EurvOUfOPWtz",

This file has been truncated. show original

muellerzr · February 3, 2020, 2:21pm

At least I know it’s a trusted source try running this notebook:

github.com

fastai/fastai2/blob/master/nbs/course/lesson4-tabular.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Tabular models"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from fastai2.tabular.all import *"
   ]
  },
  {
   "cell_type": "markdown",

This file has been truncated. show original

muellerzr · February 3, 2020, 2:22pm

What do your procs look like?

Javi · February 3, 2020, 2:24pm

procs = [Categorify, FillMissing, Normalize]

I’m going to try that notebook with fastai2 0.0.7 from pip

muellerzr · February 3, 2020, 2:36pm

No, I use pip. The tabular baseline notebook shows my install script

s.s.o · February 3, 2020, 2:39pm

The new one has regression block its not fully documented yet. So, you might need to wait for documents updated or browse the code…

muellerzr · February 3, 2020, 2:42pm

@s.s.o it looks like it’s just adults he’s running, which shouldn’t be a RegressionBlock. Though I got regression to work too. @Javi are you installing fastcore too?

Javi · February 3, 2020, 2:44pm

Now it works! I had pytorch 1.2.0 sorry