Collaborative filtering - different size of Embedding matrix

Part 1 (2019)
1

Hi,

I’ve tried collaborative filtering, by below example using files from MovieLens-Latest.

from fastai.collab import *
from fastai.tabular import *


#'http://files.grouplens.org/datasets/movielens/ml-latest-small.zip'
#path = untar_data(URLs.ML_SAMPLE)
#print(path.ls())

ratings = pd.read_csv('ml-latest-small/ratings.csv')
all_movies = ratings['movieId'].unique().astype(str)
all_users = ratings['userId'].unique().astype(str)

print('Number of movies:', len(all_movies),'Number of users:', len(all_users))
ratings.head()

image

I hope to get the embedding size for items: 9724, and for user 610, but when I divide data and split what I’ve got is only: 8974:

user,title,rating = 'userId','movieId','rating'
data = CollabList.from_df(ratings, cat_names=[user,title],procs=Categorify)
data_split = data.split_by_rand_pct(valid_pct=0.2, seed=200).label_from_df(cols=rating)
print('classes: ', len(data_split.x.classes['movieId']))

data_bunch = data_split.databunch()
y_range = [0,5.5]
learn = collab_learner(data_bunch, n_factors=40, y_range=y_range, wd=1e-1)


print(len(learn.data.x.classes['movieId']))
print(learn.data.get_emb_szs())
learn.model

image

and a model size is not that I expected. I notice that if I use valid_pct=0 than I get the correct size. I thought that size for weights and biases should be the same as the size of the classes. What could cause the different size of the weights ?

Link to the Colab file:

https://colab.research.google.com/drive/1qVwvgsMa1UMG5cqxZQu9r8zWznAsg2ay

2

Copying my answer from GitHub to help anyone that has the same question:

Hi there! Classes that are only present in the validation set as considered as unknown. This is the expected behavior and not a bug: if we put some embedding for those, they will stay in their random initial state since the model doesn’t get any sample with those classes in the training set. It can’t properly learn.

This is why you see the right number when setting a validation pct to 0. Another way to solve this would be to carefully compute validation indices so that all the classes are present in the training set and not use a random split. You would then use split_by_idx instead of rand_split_by_pct.

2 Likes
3

Thanks @sgugger! I was trying to train my own collaborative fitlering following lesson 4. I didn’t realise I could you CollabList.from_df() and split my dataset the same way Jeremy did with images (split_from_df())