18_wide_and_deep_for_recsys_pt1

The goal of this, and the companion (part 2) notebooks is to illustrate how one could use this library in the context of recommendation systems. In particular, this notebook and the scripts at the wide_deep_for_recsys dir are a response to this issue. Therefore, we will use the Kaggle notebook referred in that issue here.

In order to keep the length of the notebook tractable, we will split this exercise in 2. In this first notebook we will prepare the data in almost the exact same way as it is done in the Kaggle notebook and also show how one could use pytorch-widedeep to build a model almost identical to the one in that notebook.

In a second notebook, we will show how one could use this library to implement other models, still following the same problem formulation.

In [1]:

from pathlib import Path
import warnings

import pandas as pd
from sklearn.model_selection import train_test_split

from pytorch_widedeep.datasets import load_movielens100k

from pathlib import Path import warnings import pandas as pd from sklearn.model_selection import train_test_split from pytorch_widedeep.datasets import load_movielens100k

In [2]:

warnings.filterwarnings("ignore")

warnings.filterwarnings("ignore")

In [3]:

save_path = Path("prepared_data")
if not save_path.exists():
    save_path.mkdir(parents=True, exist_ok=True)

save_path = Path("prepared_data") if not save_path.exists(): save_path.mkdir(parents=True, exist_ok=True)

In [4]:

data, users, items = load_movielens100k(as_frame=True)

data, users, items = load_movielens100k(as_frame=True)

In [5]:

# Alternatively, as specified in the docs: 'The last 19 fields are the genres' so:
# list_of_genres = items.columns.tolist()[-19:]
list_of_genres = [
    "unknown",
    "Action",
    "Adventure",
    "Animation",
    "Children's",
    "Comedy",
    "Crime",
    "Documentary",
    "Drama",
    "Fantasy",
    "Film-Noir",
    "Horror",
    "Musical",
    "Mystery",
    "Romance",
    "Sci-Fi",
    "Thriller",
    "War",
    "Western",
]

# Alternatively, as specified in the docs: 'The last 19 fields are the genres' so: # list_of_genres = items.columns.tolist()[-19:] list_of_genres = [ "unknown", "Action", "Adventure", "Animation", "Children's", "Comedy", "Crime", "Documentary", "Drama", "Fantasy", "Film-Noir", "Horror", "Musical", "Mystery", "Romance", "Sci-Fi", "Thriller", "War", "Western", ]

Let's first start by loading the interactions, user and item data

In [6]:

data.head()

data.head()

Out[6]:

	user_id	movie_id	rating	timestamp
0	196	242	3	881250949
1	186	302	3	891717742
2	22	377	1	878887116
3	244	51	2	880606923
4	166	346	1	886397596

In [7]:

users.head()

users.head()

Out[7]:

	user_id	age	gender	occupation	zip_code
0	1	24	M	technician	85711
1	2	53	F	other	94043
2	3	23	M	writer	32067
3	4	24	M	technician	43537
4	5	33	F	other	15213

In [8]:

items.head()

items.head()

Out[8]:

	movie_id	movie_title	release_date	video_release_date	IMDb_URL	unknown	Action	Adventure	Animation	Children's	...	Fantasy	Film-Noir	Horror	Musical	Mystery	Romance	Sci-Fi	Thriller	War	Western
0	1	Toy Story (1995)	01-Jan-1995	NaN	http://us.imdb.com/M/title-exact?Toy%20Story%2...	0	0	0	1	1	...	0	0	0	0	0	0	0	0	0	0
1	2	GoldenEye (1995)	01-Jan-1995	NaN	http://us.imdb.com/M/title-exact?GoldenEye%20(...	0	1	1	0	0	...	0	0	0	0	0	0	0	1	0	0
2	3	Four Rooms (1995)	01-Jan-1995	NaN	http://us.imdb.com/M/title-exact?Four%20Rooms%...	0	0	0	0	0	...	0	0	0	0	0	0	0	1	0	0
3	4	Get Shorty (1995)	01-Jan-1995	NaN	http://us.imdb.com/M/title-exact?Get%20Shorty%...	0	1	0	0	0	...	0	0	0	0	0	0	0	0	0	0
4	5	Copycat (1995)	01-Jan-1995	NaN	http://us.imdb.com/M/title-exact?Copycat%20(1995)	0	0	0	0	0	...	0	0	0	0	0	0	0	1	0	0

5 rows × 24 columns

In [9]:

# adding a column with the number of movies watched per user
dataset = data.sort_values(["user_id", "timestamp"]).reset_index(drop=True)
dataset["one"] = 1
dataset["num_watched"] = dataset.groupby("user_id")["one"].cumsum()
dataset.drop("one", axis=1, inplace=True)
dataset.head()

# adding a column with the number of movies watched per user dataset = data.sort_values(["user_id", "timestamp"]).reset_index(drop=True) dataset["one"] = 1 dataset["num_watched"] = dataset.groupby("user_id")["one"].cumsum() dataset.drop("one", axis=1, inplace=True) dataset.head()

Out[9]:

	user_id	movie_id	rating	timestamp	num_watched
0	1	168	5	874965478	1
1	1	172	5	874965478	2
2	1	165	5	874965518	3
3	1	156	4	874965556	4
4	1	196	5	874965677	5

In [10]:

# adding a column with the mean rating at a point in time per user
dataset["mean_rate"] = (
    dataset.groupby("user_id")["rating"].cumsum() / dataset["num_watched"]
)
dataset.head()

# adding a column with the mean rating at a point in time per user dataset["mean_rate"] = ( dataset.groupby("user_id")["rating"].cumsum() / dataset["num_watched"] ) dataset.head()

Out[10]:

	user_id	movie_id	rating	timestamp	num_watched	mean_rate
0	1	168	5	874965478	1	5.00
1	1	172	5	874965478	2	5.00
2	1	165	5	874965518	3	5.00
3	1	156	4	874965556	4	4.75
4	1	196	5	874965677	5	4.80

Problem formulation¶

In this particular exercise the problem is formulated as predicting the next movie that will be watched (in consequence the last interactions will be discarded)

In [11]:

dataset["target"] = dataset.groupby("user_id")["movie_id"].shift(-1)

dataset["target"] = dataset.groupby("user_id")["movie_id"].shift(-1)

Following the same processing used by the author in the before-mentioned Kaggle notebook, we build sequences of previous movies watched

In [12]:

# Here the author builds the sequences
dataset["prev_movies"] = dataset["movie_id"].apply(lambda x: str(x))
dataset["prev_movies"] = (
    dataset.groupby("user_id")["prev_movies"]
    .apply(lambda x: (x + " ").cumsum().str.strip())
    .reset_index(drop=True)
)
dataset["prev_movies"] = dataset["prev_movies"].apply(lambda x: x.split())
dataset.head()

# Here the author builds the sequences dataset["prev_movies"] = dataset["movie_id"].apply(lambda x: str(x)) dataset["prev_movies"] = ( dataset.groupby("user_id")["prev_movies"] .apply(lambda x: (x + " ").cumsum().str.strip()) .reset_index(drop=True) ) dataset["prev_movies"] = dataset["prev_movies"].apply(lambda x: x.split()) dataset.head()

Out[12]:

	user_id	movie_id	rating	timestamp	num_watched	mean_rate	target	prev_movies
0	1	168	5	874965478	1	5.00	172.0	[168]
1	1	172	5	874965478	2	5.00	165.0	[168, 172]
2	1	165	5	874965518	3	5.00	156.0	[168, 172, 165]
3	1	156	4	874965556	4	4.75	196.0	[168, 172, 165, 156]
4	1	196	5	874965677	5	4.80	166.0	[168, 172, 165, 156, 196]

And now we add a genre_rate as the mean of all movies rated for a given genre per user

In [13]:

dataset = dataset.merge(items[["movie_id"] + list_of_genres], on="movie_id", how="left")
for genre in list_of_genres:
    dataset[f"{genre}_rate"] = dataset[genre] * dataset["rating"]
    dataset[genre] = dataset.groupby("user_id")[genre].cumsum()
    dataset[f"{genre}_rate"] = (
        dataset.groupby("user_id")[f"{genre}_rate"].cumsum() / dataset[genre]
    )
dataset[list_of_genres] = dataset[list_of_genres].apply(
    lambda x: x / dataset["num_watched"]
)
dataset.head()

dataset = dataset.merge(items[["movie_id"] + list_of_genres], on="movie_id", how="left") for genre in list_of_genres: dataset[f"{genre}_rate"] = dataset[genre] * dataset["rating"] dataset[genre] = dataset.groupby("user_id")[genre].cumsum() dataset[f"{genre}_rate"] = ( dataset.groupby("user_id")[f"{genre}_rate"].cumsum() / dataset[genre] ) dataset[list_of_genres] = dataset[list_of_genres].apply( lambda x: x / dataset["num_watched"] ) dataset.head()

Out[13]:

	user_id	movie_id	rating	timestamp	num_watched	mean_rate	target	prev_movies	unknown	Action	...	Fantasy_rate	Film-Noir_rate	Horror_rate	Musical_rate	Mystery_rate	Romance_rate	Sci-Fi_rate	Thriller_rate	War_rate	Western_rate
0	1	168	5	874965478	1	5.00	172.0	[168]	0.0	0.000000	...	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN
1	1	172	5	874965478	2	5.00	165.0	[168, 172]	0.0	0.500000	...	NaN	NaN	NaN	NaN	NaN	5.0	5.0	NaN	5.0	NaN
2	1	165	5	874965518	3	5.00	156.0	[168, 172, 165]	0.0	0.333333	...	NaN	NaN	NaN	NaN	NaN	5.0	5.0	NaN	5.0	NaN
3	1	156	4	874965556	4	4.75	196.0	[168, 172, 165, 156]	0.0	0.250000	...	NaN	NaN	NaN	NaN	NaN	5.0	5.0	4.0	5.0	NaN
4	1	196	5	874965677	5	4.80	166.0	[168, 172, 165, 156, 196]	0.0	0.200000	...	NaN	NaN	NaN	NaN	NaN	5.0	5.0	4.0	5.0	NaN

5 rows × 46 columns

Adding user features

In [14]:

dataset = dataset.merge(users, on="user_id", how="left")
dataset.head()

dataset = dataset.merge(users, on="user_id", how="left") dataset.head()

Out[14]:

	user_id	movie_id	rating	timestamp	num_watched	mean_rate	target	prev_movies	unknown	Action	...	Mystery_rate	Romance_rate	Sci-Fi_rate	Thriller_rate	War_rate	Western_rate	age	gender	occupation	zip_code
0	1	168	5	874965478	1	5.00	172.0	[168]	0.0	0.000000	...	NaN	NaN	NaN	NaN	NaN	NaN	24	M	technician	85711
1	1	172	5	874965478	2	5.00	165.0	[168, 172]	0.0	0.500000	...	NaN	5.0	5.0	NaN	5.0	NaN	24	M	technician	85711
2	1	165	5	874965518	3	5.00	156.0	[168, 172, 165]	0.0	0.333333	...	NaN	5.0	5.0	NaN	5.0	NaN	24	M	technician	85711
3	1	156	4	874965556	4	4.75	196.0	[168, 172, 165, 156]	0.0	0.250000	...	NaN	5.0	5.0	4.0	5.0	NaN	24	M	technician	85711
4	1	196	5	874965677	5	4.80	166.0	[168, 172, 165, 156, 196]	0.0	0.200000	...	NaN	5.0	5.0	4.0	5.0	NaN	24	M	technician	85711

5 rows × 50 columns

Again, we use the same settings as those in the Kaggle notebook, but COLD_START_TRESH is pretty aggressive

In [15]:

COLD_START_TRESH = 5

filtred_data = dataset[
    (dataset["num_watched"] >= COLD_START_TRESH) & ~(dataset["target"].isna())
].sort_values("timestamp")
train_data, _test_data = train_test_split(filtred_data, test_size=0.2, shuffle=False)
valid_data, test_data = train_test_split(_test_data, test_size=0.5, shuffle=False)

COLD_START_TRESH = 5 filtred_data = dataset[ (dataset["num_watched"] >= COLD_START_TRESH) & ~(dataset["target"].isna()) ].sort_values("timestamp") train_data, _test_data = train_test_split(filtred_data, test_size=0.2, shuffle=False) valid_data, test_data = train_test_split(_test_data, test_size=0.5, shuffle=False)

In [16]:

cols_to_drop = [
    # "rating",
    "timestamp",
    "num_watched",
]

df_train = train_data.drop(cols_to_drop, axis=1)
df_valid = valid_data.drop(cols_to_drop, axis=1)
df_test = test_data.drop(cols_to_drop, axis=1)

df_train.to_pickle(save_path / "df_train.pkl")
df_valid.to_pickle(save_path / "df_valid.pkl")
df_test.to_pickle(save_path / "df_test.pkl")

cols_to_drop = [ # "rating", "timestamp", "num_watched", ] df_train = train_data.drop(cols_to_drop, axis=1) df_valid = valid_data.drop(cols_to_drop, axis=1) df_test = test_data.drop(cols_to_drop, axis=1) df_train.to_pickle(save_path / "df_train.pkl") df_valid.to_pickle(save_path / "df_valid.pkl") df_test.to_pickle(save_path / "df_test.pkl")

Let's now build a model that is nearly identical to the one use in the Kaggle notebook

In [17]:

import numpy as np
import torch
from torch import nn
from scipy.sparse import coo_matrix

from pytorch_widedeep import Trainer
from pytorch_widedeep.models import TabMlp, BasicRNN, WideDeep
from pytorch_widedeep.preprocessing import TabPreprocessor

import numpy as np import torch from torch import nn from scipy.sparse import coo_matrix from pytorch_widedeep import Trainer from pytorch_widedeep.models import TabMlp, BasicRNN, WideDeep from pytorch_widedeep.preprocessing import TabPreprocessor

In [18]:

device = "cuda" if torch.cuda.is_available() else "cpu"

save_path = Path("prepared_data")

PAD_IDX = 0

device = "cuda" if torch.cuda.is_available() else "cpu" save_path = Path("prepared_data") PAD_IDX = 0

Let's use some of the functions the author of the kaggle's notebook uses to prepare the data

In [19]:

def get_coo_indexes(lil):
    rows = []
    cols = []
    for i, el in enumerate(lil):
        if type(el) != list:
            el = [el]
        for j in el:
            rows.append(i)
            cols.append(j)
    return rows, cols


def get_sparse_features(series, shape):
    coo_indexes = get_coo_indexes(series.tolist())
    sparse_df = coo_matrix(
        (np.ones(len(coo_indexes[0])), (coo_indexes[0], coo_indexes[1])), shape=shape
    )
    return sparse_df


def sparse_to_idx(data, pad_idx=-1):
    indexes = data.nonzero()
    indexes_df = pd.DataFrame()
    indexes_df["rows"] = indexes[0]
    indexes_df["cols"] = indexes[1]
    mdf = indexes_df.groupby("rows").apply(lambda x: x["cols"].tolist())
    max_len = mdf.apply(lambda x: len(x)).max()
    return mdf.apply(lambda x: pd.Series(x + [pad_idx] * (max_len - len(x)))).values

def get_coo_indexes(lil): rows = [] cols = [] for i, el in enumerate(lil): if type(el) != list: el = [el] for j in el: rows.append(i) cols.append(j) return rows, cols def get_sparse_features(series, shape): coo_indexes = get_coo_indexes(series.tolist()) sparse_df = coo_matrix( (np.ones(len(coo_indexes[0])), (coo_indexes[0], coo_indexes[1])), shape=shape ) return sparse_df def sparse_to_idx(data, pad_idx=-1): indexes = data.nonzero() indexes_df = pd.DataFrame() indexes_df["rows"] = indexes[0] indexes_df["cols"] = indexes[1] mdf = indexes_df.groupby("rows").apply(lambda x: x["cols"].tolist()) max_len = mdf.apply(lambda x: len(x)).max() return mdf.apply(lambda x: pd.Series(x + [pad_idx] * (max_len - len(x)))).values

For the time being, we will not use a validation set for hyperparameter optimization, and we will simply concatenate the validation and the test set in one test set. I simply splitted the data into train/valid/test in case the reader wants to actually do hyperparameter optimization (and because I know in the future I will).

There is also another caveat worth mentioning, related to the indexing of the movies. To build the matrices of movies watched, we use the entire dataset. A more realistic (and correct) approach would be to use ONLY the movies that appear in the training set and consider unknown or unseen those in the testing set that have not been seen during training. Nonetheless, this will not affect the purposes of this notebook, which is to illustrate how one could use pytorch-widedeep to build a recommendation algorithm. However, if one wanted to explore the performance of different algorithms in a "proper" way, these "details" need to be accounted for.

In [20]:

df_test = pd.concat([df_valid, df_test], ignore_index=True)

df_test = pd.concat([df_valid, df_test], ignore_index=True)

In [21]:

id_cols = ["user_id", "movie_id"]
max_movie_index = max(df_train.movie_id.max(), df_test.movie_id.max())

id_cols = ["user_id", "movie_id"] max_movie_index = max(df_train.movie_id.max(), df_test.movie_id.max())

In [22]:

X_train = df_train.drop(id_cols + ["rating", "prev_movies", "target"], axis=1)
y_train = np.array(df_train.target.values, dtype="int64")
train_movies_watched = get_sparse_features(
    df_train["prev_movies"], (len(df_train), max_movie_index + 1)
)

X_test = df_test.drop(id_cols + ["rating", "prev_movies", "target"], axis=1)
y_test = np.array(df_test.target.values, dtype="int64")
test_movies_watched = get_sparse_features(
    df_test["prev_movies"], (len(df_test), max_movie_index + 1)
)

X_train = df_train.drop(id_cols + ["rating", "prev_movies", "target"], axis=1) y_train = np.array(df_train.target.values, dtype="int64") train_movies_watched = get_sparse_features( df_train["prev_movies"], (len(df_train), max_movie_index + 1) ) X_test = df_test.drop(id_cols + ["rating", "prev_movies", "target"], axis=1) y_test = np.array(df_test.target.values, dtype="int64") test_movies_watched = get_sparse_features( df_test["prev_movies"], (len(df_test), max_movie_index + 1) )

let's have a look to the information in each dataset

In [23]:

X_train.head()

X_train.head()

Out[23]:

	mean_rate	unknown	Action	Adventure	Animation	Children's	Comedy	Crime	Documentary	Drama	...	Mystery_rate	Romance_rate	Sci-Fi_rate	Thriller_rate	War_rate	Western_rate	age	gender	occupation	zip_code
25423	4.000000	0.0	0.400000	0.200000	0.0	0.0	0.400000	0.0	0.0	0.200000	...	NaN	4.0	4.0	4.000000	4.0	NaN	21	M	student	48823
25425	4.000000	0.0	0.285714	0.142857	0.0	0.0	0.428571	0.0	0.0	0.285714	...	NaN	4.0	4.0	4.000000	4.0	NaN	21	M	student	48823
25424	4.000000	0.0	0.333333	0.166667	0.0	0.0	0.333333	0.0	0.0	0.333333	...	NaN	4.0	4.0	4.000000	4.0	NaN	21	M	student	48823
25426	3.875000	0.0	0.250000	0.125000	0.0	0.0	0.375000	0.0	0.0	0.250000	...	NaN	4.0	4.0	3.666667	4.0	NaN	21	M	student	48823
25427	3.888889	0.0	0.222222	0.111111	0.0	0.0	0.333333	0.0	0.0	0.333333	...	NaN	4.0	4.0	3.666667	4.0	NaN	21	M	student	48823

5 rows × 43 columns

In [24]:

y_train

y_train

Out[24]:

array([772, 288, 108, ..., 183, 432, 509])

In [25]:

train_movies_watched

train_movies_watched

Out[25]:

<76228x1683 sparse matrix of type '<class 'numpy.float64'>'
	with 7957390 stored elements in COOrdinate format>

In [26]:

sorted(df_train.prev_movies.tolist()[0])

sorted(df_train.prev_movies.tolist()[0])

Out[26]:

['173', '185', '255', '286', '298']

In [27]:

np.where(train_movies_watched.todense()[0])

np.where(train_movies_watched.todense()[0])

Out[27]:

(array([0, 0, 0, 0, 0]), array([173, 185, 255, 286, 298]))

And from now on is when the specifics related to this library start to appear. The only component that is going to be a bit different is the so-called tabular component, referred as continuous in the notebook.

In the case of pytorch-widedeep we have the TabPreprocessor that allows for a lot of flexibility as to how we would like to process the tabular component of this Wide and Deep model. In other words, here our tabular component is a bit more elaborated than that in the notebook, just a bit...

In [28]:

cat_cols = ["gender", "occupation", "zip_code"]
cont_cols = [c for c in X_train if c not in cat_cols]
tab_preprocessor = TabPreprocessor(
    cat_embed_cols=cat_cols,
    continuous_cols=cont_cols,
)

cat_cols = ["gender", "occupation", "zip_code"] cont_cols = [c for c in X_train if c not in cat_cols] tab_preprocessor = TabPreprocessor( cat_embed_cols=cat_cols, continuous_cols=cont_cols, )

In [29]:

X_train_tab = tab_preprocessor.fit_transform(X_train.fillna(0))
X_test_tab = tab_preprocessor.transform(X_test.fillna(0))

X_train_tab = tab_preprocessor.fit_transform(X_train.fillna(0)) X_test_tab = tab_preprocessor.transform(X_test.fillna(0))

Now, in the notebook, the author moves the sparse matrices to sparse tensors and then turns them into dense tensors. In reality, this is not neccessary, one could feed sparse tensors to nn.Linear layers in pytorch. Nonetheless, this is not the most efficient implementation and is the reason why in our library the wide, linear component is implemented as an embedding layer.

Nonetheless, to reproduce the notebook the best we can and because currently the Wide model in pytorch-widedeep is not designed to receive sparse tensors (we might consider implementing this functionality), we will turn the sparse COO matrices into dense arrays. We will then code a fairly simple, custom Wide component.

In [30]:

X_train_wide = np.array(train_movies_watched.todense())
X_test_wide = np.array(test_movies_watched.todense())

X_train_wide = np.array(train_movies_watched.todense()) X_test_wide = np.array(test_movies_watched.todense())

Finally, the author of the notebook uses a simple Embedding layer to encode the sequences of movies watched, the prev_movies columns. In my opinion, there is an element of information redundancy here. This is because the wide and text components have implicitely the same information, but in different form. Moreover, both of the models used for these two components ignore the sequential element in the data. Nonetheless, we want to reproduce the Kaggle notebook as close as possible, AND as one can explore later (by simply performing simple ablation studies), the wide component seems to carry most of the predictive power.

In [31]:

X_train_text = sparse_to_idx(train_movies_watched, pad_idx=PAD_IDX)
X_test_text = sparse_to_idx(test_movies_watched, pad_idx=PAD_IDX)

X_train_text = sparse_to_idx(train_movies_watched, pad_idx=PAD_IDX) X_test_text = sparse_to_idx(test_movies_watched, pad_idx=PAD_IDX)

Let's now build the models

In [32]:

class Wide(nn.Module):
    def __init__(self, input_dim: int, pred_dim: int):
        super().__init__()

        self.input_dim = input_dim
        self.pred_dim = pred_dim

        # When I coded the library I never though that someone would want to code
        # their own wide component. However, if you do, the wide component must have
        # a 'wide_linear' attribute. In other words, the linear layer must be
        # called 'wide_linear'
        self.wide_linear = nn.Linear(input_dim, pred_dim)

    def forward(self, X):
        out = self.wide_linear(X.type(torch.float32))
        return out


wide = Wide(X_train_wide.shape[1], max_movie_index + 1)

class Wide(nn.Module): def __init__(self, input_dim: int, pred_dim: int): super().__init__() self.input_dim = input_dim self.pred_dim = pred_dim # When I coded the library I never though that someone would want to code # their own wide component. However, if you do, the wide component must have # a 'wide_linear' attribute. In other words, the linear layer must be # called 'wide_linear' self.wide_linear = nn.Linear(input_dim, pred_dim) def forward(self, X): out = self.wide_linear(X.type(torch.float32)) return out wide = Wide(X_train_wide.shape[1], max_movie_index + 1)

In [33]:

wide

wide

Out[33]:

Wide(
  (wide_linear): Linear(in_features=1683, out_features=1683, bias=True)
)

In [34]:

class SimpleEmbed(nn.Module):
    def __init__(self, vocab_size: int, embed_dim: int, pad_idx: int):
        super().__init__()

        self.vocab_size = vocab_size
        self.embed_dim = embed_dim
        self.pad_idx = pad_idx

        # The sequences of movies watched are simply embedded in the Kaggle
        # notebook. No RNN, Transformer or any model is used
        self.embed = nn.Embedding(vocab_size, embed_dim, padding_idx=pad_idx)

    def forward(self, X):
        embed = self.embed(X)
        embed_mean = torch.mean(embed, dim=1)
        return embed_mean

    @property
    def output_dim(self) -> int:
        # All deep components in a custom 'pytorch-widedeep' model must have
        # an output_dim property
        return self.embed_dim


#  In the notebook the author uses simply embeddings
simple_embed = SimpleEmbed(max_movie_index + 1, 16, 0)

class SimpleEmbed(nn.Module): def __init__(self, vocab_size: int, embed_dim: int, pad_idx: int): super().__init__() self.vocab_size = vocab_size self.embed_dim = embed_dim self.pad_idx = pad_idx # The sequences of movies watched are simply embedded in the Kaggle # notebook. No RNN, Transformer or any model is used self.embed = nn.Embedding(vocab_size, embed_dim, padding_idx=pad_idx) def forward(self, X): embed = self.embed(X) embed_mean = torch.mean(embed, dim=1) return embed_mean @property def output_dim(self) -> int: # All deep components in a custom 'pytorch-widedeep' model must have # an output_dim property return self.embed_dim # In the notebook the author uses simply embeddings simple_embed = SimpleEmbed(max_movie_index + 1, 16, 0)

In [35]:

simple_embed

simple_embed

Out[35]:

SimpleEmbed(
  (embed): Embedding(1683, 16, padding_idx=0)
)

Maybe one would like to use an RNN to account for the sequence nature of the problem. If that was the case it would be as easy as:

In [36]:

basic_rnn = BasicRNN(
    vocab_size=max_movie_index + 1,
    embed_dim=16,
    hidden_dim=32,
    n_layers=2,
    rnn_type="gru",
)

basic_rnn = BasicRNN( vocab_size=max_movie_index + 1, embed_dim=16, hidden_dim=32, n_layers=2, rnn_type="gru", )

And finally, the tabular component, which is the notebook is simply a stak of linear + Rely layers. In our case we have an embedding layer before the linear layers to encode categorial and numerical cols

In [37]:

tab_mlp = TabMlp(
    column_idx=tab_preprocessor.column_idx,
    cat_embed_input=tab_preprocessor.cat_embed_input,
    continuous_cols=tab_preprocessor.continuous_cols,
    cont_norm_layer=None,
    mlp_hidden_dims=[1024, 512, 256],
    mlp_activation="relu",
)

tab_mlp = TabMlp( column_idx=tab_preprocessor.column_idx, cat_embed_input=tab_preprocessor.cat_embed_input, continuous_cols=tab_preprocessor.continuous_cols, cont_norm_layer=None, mlp_hidden_dims=[1024, 512, 256], mlp_activation="relu", )

In [38]:

tab_mlp

tab_mlp

Out[38]:

TabMlp(
  (cat_embed): DiffSizeCatEmbeddings(
    (embed_layers): ModuleDict(
      (emb_layer_gender): Embedding(3, 2, padding_idx=0)
      (emb_layer_occupation): Embedding(22, 9, padding_idx=0)
      (emb_layer_zip_code): Embedding(648, 60, padding_idx=0)
    )
    (embedding_dropout): Dropout(p=0.0, inplace=False)
  )
  (cont_norm): Identity()
  (encoder): MLP(
    (mlp): Sequential(
      (dense_layer_0): Sequential(
        (0): Linear(in_features=111, out_features=1024, bias=True)
        (1): ReLU(inplace=True)
        (2): Dropout(p=0.1, inplace=False)
      )
      (dense_layer_1): Sequential(
        (0): Linear(in_features=1024, out_features=512, bias=True)
        (1): ReLU(inplace=True)
        (2): Dropout(p=0.1, inplace=False)
      )
      (dense_layer_2): Sequential(
        (0): Linear(in_features=512, out_features=256, bias=True)
        (1): ReLU(inplace=True)
        (2): Dropout(p=0.1, inplace=False)
      )
    )
  )
)

Finally, we simply wrap up all models with the WideDeep 'collector' class and we are ready to train.

In [39]:

wide_deep_model = WideDeep(
    wide=wide, deeptabular=tab_mlp, deeptext=simple_embed, pred_dim=max_movie_index + 1
)

wide_deep_model = WideDeep( wide=wide, deeptabular=tab_mlp, deeptext=simple_embed, pred_dim=max_movie_index + 1 )

In [40]:

wide_deep_model

wide_deep_model

Out[40]:

WideDeep(
  (wide): Wide(
    (wide_linear): Linear(in_features=1683, out_features=1683, bias=True)
  )
  (deeptabular): Sequential(
    (0): TabMlp(
      (cat_embed): DiffSizeCatEmbeddings(
        (embed_layers): ModuleDict(
          (emb_layer_gender): Embedding(3, 2, padding_idx=0)
          (emb_layer_occupation): Embedding(22, 9, padding_idx=0)
          (emb_layer_zip_code): Embedding(648, 60, padding_idx=0)
        )
        (embedding_dropout): Dropout(p=0.0, inplace=False)
      )
      (cont_norm): Identity()
      (encoder): MLP(
        (mlp): Sequential(
          (dense_layer_0): Sequential(
            (0): Linear(in_features=111, out_features=1024, bias=True)
            (1): ReLU(inplace=True)
            (2): Dropout(p=0.1, inplace=False)
          )
          (dense_layer_1): Sequential(
            (0): Linear(in_features=1024, out_features=512, bias=True)
            (1): ReLU(inplace=True)
            (2): Dropout(p=0.1, inplace=False)
          )
          (dense_layer_2): Sequential(
            (0): Linear(in_features=512, out_features=256, bias=True)
            (1): ReLU(inplace=True)
            (2): Dropout(p=0.1, inplace=False)
          )
        )
      )
    )
    (1): Linear(in_features=256, out_features=1683, bias=True)
  )
  (deeptext): Sequential(
    (0): SimpleEmbed(
      (embed): Embedding(1683, 16, padding_idx=0)
    )
    (1): Linear(in_features=16, out_features=1683, bias=True)
  )
)

Note that the main difference between this wide and deep model and the Wide and Deep model in the Kaggle notebook is that in that notebook, the author concatenates the embedings and the tabular features, then passes this concatenation through a stack of linear + Relu layers with a final output dim of 256. Then concatenates this output with the binary features and connects this concatenation with the final linear layer (so the final weights are of dim (batch_size, 256 + 1683)). Our implementation follows the notation of the original paper and instead of concatenating the tabular, text and wide components and then connect them to the output neurons, we first compute their output, and then add it (see here: https://arxiv.org/pdf/1606.07792.pdf, their Eq 3). Note that this is effectively the same, with the caveat that while in one case one initialises a big weight matrix "at once", in our implementation we initialise different matrices for different components. Anyway, let's give it a go.

In [41]:

trainer = Trainer(
    model=wide_deep_model,
    objective="multiclass",
    custom_loss_function=nn.CrossEntropyLoss(ignore_index=PAD_IDX),
    optimizers=torch.optim.Adam(wide_deep_model.parameters(), lr=1e-3),
)

trainer = Trainer( model=wide_deep_model, objective="multiclass", custom_loss_function=nn.CrossEntropyLoss(ignore_index=PAD_IDX), optimizers=torch.optim.Adam(wide_deep_model.parameters(), lr=1e-3), )

In [42]:

trainer.fit(
    X_train={
        "X_wide": X_train_wide,
        "X_tab": X_train_tab,
        "X_text": X_train_text,
        "target": y_train,
    },
    X_val={
        "X_wide": X_test_wide,
        "X_tab": X_test_tab,
        "X_text": X_test_text,
        "target": y_test,
    },
    n_epochs=5,
    batch_size=512,
    shuffle=False,
)

trainer.fit( X_train={ "X_wide": X_train_wide, "X_tab": X_train_tab, "X_text": X_train_text, "target": y_train, }, X_val={ "X_wide": X_test_wide, "X_tab": X_test_tab, "X_text": X_test_text, "target": y_test, }, n_epochs=5, batch_size=512, shuffle=False, )

epoch 1: 100%|████████████████████████████████████████████████████████████████████████████████████| 149/149 [00:19<00:00,  7.66it/s, loss=6.66]
valid: 100%|█████████████████████████████████████████████████████████████████████████████████████████| 38/38 [00:02<00:00, 18.75it/s, loss=6.6]
epoch 2: 100%|████████████████████████████████████████████████████████████████████████████████████| 149/149 [00:21<00:00,  6.95it/s, loss=5.97]
valid: 100%|████████████████████████████████████████████████████████████████████████████████████████| 38/38 [00:01<00:00, 21.03it/s, loss=6.52]
epoch 3: 100%|████████████████████████████████████████████████████████████████████████████████████| 149/149 [00:19<00:00,  7.51it/s, loss=5.65]
valid: 100%|████████████████████████████████████████████████████████████████████████████████████████| 38/38 [00:01<00:00, 20.16it/s, loss=6.53]
epoch 4: 100%|████████████████████████████████████████████████████████████████████████████████████| 149/149 [00:23<00:00,  6.29it/s, loss=5.41]
valid: 100%|████████████████████████████████████████████████████████████████████████████████████████| 38/38 [00:02<00:00, 13.97it/s, loss=6.57]
epoch 5: 100%|█████████████████████████████████████████████████████████████████████████████████████| 149/149 [00:19<00:00,  7.58it/s, loss=5.2]
valid: 100%|████████████████████████████████████████████████████████████████████████████████████████| 38/38 [00:02<00:00, 18.82it/s, loss=6.63]

Now one could continue to the 'compare' metrics section of the Kaggle notebook. However, for the purposes of illustrating how one could use pytorch-widedeep to build recommendation algorithms we consider this notebook completed and move onto part 2