IndexError: index out of range in self about pytorch-widedeep

[jrzaurin]

Hey @berkeserol, so, I have two comments, first, this runs well, see if it helps:

import numpy as np
import pandas as pd
from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split

from pytorch_widedeep import Trainer
from pytorch_widedeep.models import TabMlp, Wide, WideDeep  # noqa: F401
from pytorch_widedeep.preprocessing import TabPreprocessor, WidePreprocessor

X, y = make_classification(n_samples=100, n_features=10, n_informative=10, n_redundant=0)
X = pd.DataFrame(X, columns=[f"col_{i}" for i in range(10)])
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, shuffle=False)

tab_preprocessor = TabPreprocessor(continuous_cols=X.columns.tolist())
X_train_processed = tab_preprocessor.fit_transform(X_train)
X_test_processed = tab_preprocessor.transform(X_test)

wide_preprocessor = WidePreprocessor(wide_cols=X.columns.tolist())
X_wide = wide_preprocessor.fit_transform(X_train)

# Define the model
tab_mlp = TabMlp(
    column_idx=tab_preprocessor.column_idx,
    continuous_cols=tab_preprocessor.continuous_cols,
    mlp_hidden_dims=[64, 32]
)

wide = Wide(input_dim=np.unique(X_train).shape[0])
model = WideDeep(wide=wide, deeptabular=tab_mlp)

# Define the trainer and train the model
trainer = Trainer(model, objective="binary")
trainer.fit(
    X_tab=X_train_processed,
    X_wide=X_wide,
    target=y_train,
    n_epochs=10,
    batch_size=32
)

you do not need those .astype(float).

Second, and also important, the wide component is not designed to work only with continuous features. In fact, the way that the WidePreprocessor prepares the data is such that all ends up in a look up table (dictionary) and the linear layer is coded as embeddings (have a look here for an explanation). For example, if you access the encoding_dict attribute in the example above you will see this:

>>> wide_preprocessor.encoding_dict
{'col_0_-3.956915512990843': 1,
 'col_0_1.3006394509583112': 2,
 'col_0_-2.464936360812435': 3,
 'col_0_2.4968884101664472': 4,
 'col_0_-1.0889744597533615': 5,
 'col_0_-1.7765530407990036': 6,
 'col_0_-2.9406750132381694': 7,
 'col_0_-4.13009658776504': 8,
 'col_0_1.3052810320392583': 9,
 'col_0_-0.7368640448231503': 10,
 'col_0_0.5910531307205757': 11,
 'col_0_1.3899029479159295': 12,
 'col_0_4.4319842979409145': 13,
 'col_0_-1.2444478420364677': 14,
 'col_0_1.818636056977371': 15,
 ...

i.e. one encoding per individual value, per column

Just that. Hope this helps and thanks for opening the issue and trying the library!

[berkeserol]

Hi @jrzaurin thank you for the answer. I tried your code exactly and it works fine. However, when I change the input dataframe using the file I attached, I got the error that I mentioned in the first message. Can you try?
widedeep_test_X.csv

[jrzaurin]

ok, so I see what is happening, I will change the example and here is a solution.

ISSUE:

two of the values in your dataset are equal for different columns:

import numpy as np
import pandas as pd

# from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split

X = pd.read_csv("~/Desktop/widedeep_test_X.csv")
y = np.random.randint(2, size=X.shape[0])

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, shuffle=False)

# unique elements column by column
print(len(np.hstack([X_train[col].unique() for col in X_train.columns])))
90

# unique elements consider the whole df at once
print(np.unique(X_train).shape[0])
89

Because of the way that the linear layer (wide component) is implemented, the number that matters is the first one. This is because as I wrote before, is implemented as an Embedding layer where each element will be an entry in the look up table. To that aim we append the column name and the value for each column. In your example, if we refer as the repeated float as repeated_float, and assuming it is in two columns 1 and 2, it will be encoded as:

col_1_repeated_float: encoding_n
...
col_2_repeated_float: encoding_m

Now, if you define your wide model as

wide = Wide(input_dim=np.unique(X_train).shape[0])

we are defining an embedding layer of 90 input dim (89 + index 0 left for 'unseen' categories) when in reality we need an embedding layer of 91 (90 + index 0 left for 'unseen' categories), since col_1_repeated_float and col_2_repeated_float should be encoded differently

SOLUTION

It is safer to define the wide model as: wide = Wide(input_dim=len(wide_preprocessor.encoding_dict))

then this code should run

import numpy as np
import pandas as pd

from sklearn.model_selection import train_test_split

from pytorch_widedeep import Trainer
from pytorch_widedeep.models import TabMlp, Wide, WideDeep  # noqa: F401
from pytorch_widedeep.preprocessing import TabPreprocessor, WidePreprocessor

X = pd.read_csv("~/Desktop/widedeep_test_X.csv")
y = np.random.randint(2, size=X.shape[0])

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, shuffle=False)

tab_preprocessor = TabPreprocessor(continuous_cols=X.columns.tolist())
X_train_processed = tab_preprocessor.fit_transform(X_train)

wide_preprocessor = WidePreprocessor(wide_cols=X.columns.tolist())
X_wide = wide_preprocessor.fit_transform(X_train)

# Define the model
tab_mlp = TabMlp(
    column_idx=tab_preprocessor.column_idx,
    continuous_cols=tab_preprocessor.continuous_cols,
    mlp_hidden_dims=[64, 32],
)


wide = Wide(input_dim=len(wide_preprocessor.encoding_dict))
model = WideDeep(wide=wide, deeptabular=tab_mlp)

# Define the trainer and train the model
trainer = Trainer(model, objective="binary")
trainer.fit(
    X_tab=X_train_processed, X_wide=X_wide, target=y_train, n_epochs=1, batch_size=2
)

[berkeserol]

It is solved. Thanks