from torch.utils.data import Dataset
from torch.utils.data import DataLoader, WeightedRandomSampler
import torch
import numpy as np
import scipy
import json
from sklearn.utils import check_array
import pandas as pd
import warnings
[docs]class TorchDataset(Dataset):
"""
Format for numpy array
Parameters
----------
X : 2D array
The input matrix
y : 2D array
The one-hot encoded target
"""
def __init__(self, x, y):
self.x = x
self.y = y
def __len__(self):
return len(self.x)
def __getitem__(self, index):
x, y = self.x[index], self.y[index]
return x, y
[docs]class SparseTorchDataset(Dataset):
"""
Format for csr_matrix
Parameters
----------
X : CSR matrix
The input matrix
y : 2D array
The one-hot encoded target
"""
def __init__(self, x, y):
self.x = x
self.y = y
def __len__(self):
return self.x.shape[0]
def __getitem__(self, index):
x = torch.from_numpy(self.x[index].toarray()[0]).float()
y = self.y[index]
return x, y
[docs]class PredictDataset(Dataset):
"""
Format for numpy array
Parameters
----------
X : 2D array
The input matrix
"""
def __init__(self, x):
self.x = x
def __len__(self):
return len(self.x)
def __getitem__(self, index):
x = self.x[index]
return x
[docs]class SparsePredictDataset(Dataset):
"""
Format for csr_matrix
Parameters
----------
X : CSR matrix
The input matrix
"""
def __init__(self, x):
self.x = x
def __len__(self):
return self.x.shape[0]
def __getitem__(self, index):
x = torch.from_numpy(self.x[index].toarray()[0]).float()
return x
[docs]def create_sampler(weights, y_train):
"""
This creates a sampler from the given weights
Parameters
----------
weights : either 0, 1, dict or iterable
if 0 (default) : no weights will be applied
if 1 : classification only, will balanced class with inverse frequency
if dict : keys are corresponding class values are sample weights
if iterable : list or np array must be of length equal to nb elements
in the training set
y_train : np.array
Training targets
"""
if isinstance(weights, int):
if weights == 0:
need_shuffle = True
sampler = None
elif weights == 1:
need_shuffle = False
class_sample_count = np.array(
[len(np.where(y_train == t)[0]) for t in np.unique(y_train)]
)
weights = 1.0 / class_sample_count
samples_weight = np.array([weights[t] for t in y_train])
samples_weight = torch.from_numpy(samples_weight)
samples_weight = samples_weight.double()
sampler = WeightedRandomSampler(samples_weight, len(samples_weight))
else:
raise ValueError("Weights should be either 0, 1, dictionnary or list.")
elif isinstance(weights, dict):
# custom weights per class
need_shuffle = False
samples_weight = np.array([weights[t] for t in y_train])
sampler = WeightedRandomSampler(samples_weight, len(samples_weight))
else:
# custom weights
if len(weights) != len(y_train):
raise ValueError("Custom weights should match number of train samples.")
need_shuffle = False
samples_weight = np.array(weights)
sampler = WeightedRandomSampler(samples_weight, len(samples_weight))
return need_shuffle, sampler
[docs]def create_dataloaders(
X_train, y_train, eval_set, weights, batch_size, num_workers, drop_last, pin_memory
):
"""
Create dataloaders with or without subsampling depending on weights and balanced.
Parameters
----------
X_train : np.ndarray
Training data
y_train : np.array
Mapped Training targets
eval_set : list of tuple
List of eval tuple set (X, y)
weights : either 0, 1, dict or iterable
if 0 (default) : no weights will be applied
if 1 : classification only, will balanced class with inverse frequency
if dict : keys are corresponding class values are sample weights
if iterable : list or np array must be of length equal to nb elements
in the training set
batch_size : int
how many samples per batch to load
num_workers : int
how many subprocesses to use for data loading. 0 means that the data
will be loaded in the main process
drop_last : bool
set to True to drop the last incomplete batch, if the dataset size is not
divisible by the batch size. If False and the size of dataset is not
divisible by the batch size, then the last batch will be smaller
pin_memory : bool
Whether to pin GPU memory during training
Returns
-------
train_dataloader, valid_dataloader : torch.DataLoader, torch.DataLoader
Training and validation dataloaders
"""
need_shuffle, sampler = create_sampler(weights, y_train)
if scipy.sparse.issparse(X_train):
train_dataloader = DataLoader(
SparseTorchDataset(X_train.astype(np.float32), y_train),
batch_size=batch_size,
sampler=sampler,
shuffle=need_shuffle,
num_workers=num_workers,
drop_last=drop_last,
pin_memory=pin_memory,
)
else:
train_dataloader = DataLoader(
TorchDataset(X_train.astype(np.float32), y_train),
batch_size=batch_size,
sampler=sampler,
shuffle=need_shuffle,
num_workers=num_workers,
drop_last=drop_last,
pin_memory=pin_memory,
)
valid_dataloaders = []
for X, y in eval_set:
if scipy.sparse.issparse(X):
valid_dataloaders.append(
DataLoader(
SparseTorchDataset(X.astype(np.float32), y),
batch_size=batch_size,
shuffle=False,
num_workers=num_workers,
pin_memory=pin_memory,
)
)
else:
valid_dataloaders.append(
DataLoader(
TorchDataset(X.astype(np.float32), y),
batch_size=batch_size,
shuffle=False,
num_workers=num_workers,
pin_memory=pin_memory,
)
)
return train_dataloader, valid_dataloaders
[docs]def create_explain_matrix(input_dim, cat_emb_dim, cat_idxs, post_embed_dim):
"""
This is a computational trick.
In order to rapidly sum importances from same embeddings
to the initial index.
Parameters
----------
input_dim : int
Initial input dim
cat_emb_dim : int or list of int
if int : size of embedding for all categorical feature
if list of int : size of embedding for each categorical feature
cat_idxs : list of int
Initial position of categorical features
post_embed_dim : int
Post embedding inputs dimension
Returns
-------
reducing_matrix : np.array
Matrix of dim (post_embed_dim, input_dim) to performe reduce
"""
if isinstance(cat_emb_dim, int):
all_emb_impact = [cat_emb_dim - 1] * len(cat_idxs)
else:
all_emb_impact = [emb_dim - 1 for emb_dim in cat_emb_dim]
acc_emb = 0
nb_emb = 0
indices_trick = []
for i in range(input_dim):
if i not in cat_idxs:
indices_trick.append([i + acc_emb])
else:
indices_trick.append(
range(i + acc_emb, i + acc_emb + all_emb_impact[nb_emb] + 1)
)
acc_emb += all_emb_impact[nb_emb]
nb_emb += 1
reducing_matrix = np.zeros((post_embed_dim, input_dim))
for i, cols in enumerate(indices_trick):
reducing_matrix[cols, i] = 1
return scipy.sparse.csc_matrix(reducing_matrix)
[docs]def create_group_matrix(list_groups, input_dim):
"""
Create the group matrix corresponding to the given list_groups
Parameters
----------
- list_groups : list of list of int
Each element is a list representing features in the same group.
One feature should appear in maximum one group.
Feature that don't get assigned a group will be in their own group of one feature.
- input_dim : number of feature in the initial dataset
Returns
-------
- group_matrix : torch matrix
A matrix of size (n_groups, input_dim)
where m_ij represents the importance of feature j in group i
The rows must some to 1 as each group is equally important a priori.
"""
check_list_groups(list_groups, input_dim)
if len(list_groups) == 0:
group_matrix = torch.eye(input_dim)
return group_matrix
else:
n_groups = input_dim - int(np.sum([len(gp) - 1 for gp in list_groups]))
group_matrix = torch.zeros((n_groups, input_dim))
remaining_features = [feat_idx for feat_idx in range(input_dim)]
current_group_idx = 0
for group in list_groups:
group_size = len(group)
for elem_idx in group:
# add importrance of element in group matrix and corresponding group
group_matrix[current_group_idx, elem_idx] = 1 / group_size
# remove features from list of features
remaining_features.remove(elem_idx)
# move to next group
current_group_idx += 1
# features not mentionned in list_groups get assigned their own group of singleton
for remaining_feat_idx in remaining_features:
group_matrix[current_group_idx, remaining_feat_idx] = 1
current_group_idx += 1
return group_matrix
[docs]def check_list_groups(list_groups, input_dim):
"""
Check that list groups:
- is a list of list
- does not contain twice the same feature in different groups
- does not contain unknown features (>= input_dim)
- does not contain empty groups
Parameters
----------
- list_groups : list of list of int
Each element is a list representing features in the same group.
One feature should appear in maximum one group.
Feature that don't get assign a group will be in their own group of one feature.
- input_dim : number of feature in the initial dataset
"""
assert isinstance(list_groups, list), "list_groups must be a list of list."
if len(list_groups) == 0:
return
else:
for group_pos, group in enumerate(list_groups):
msg = f"Groups must be given as a list of list, but found {group} in position {group_pos}." # noqa
assert isinstance(group, list), msg
assert len(group) > 0, "Empty groups are forbidding please remove empty groups []"
n_elements_in_groups = np.sum([len(group) for group in list_groups])
flat_list = []
for group in list_groups:
flat_list.extend(group)
unique_elements = np.unique(flat_list)
n_unique_elements_in_groups = len(unique_elements)
msg = f"One feature can only appear in one group, please check your grouped_features."
assert n_unique_elements_in_groups == n_elements_in_groups, msg
highest_feat = np.max(unique_elements)
assert highest_feat < input_dim, f"Number of features is {input_dim} but one group contains {highest_feat}." # noqa
return
[docs]def filter_weights(weights):
"""
This function makes sure that weights are in correct format for
regression and multitask TabNet
Parameters
----------
weights : int, dict or list
Initial weights parameters given by user
Returns
-------
None : This function will only throw an error if format is wrong
"""
err_msg = """Please provide a list or np.array of weights for """
err_msg += """regression, multitask or pretraining: """
if isinstance(weights, int):
if weights == 1:
raise ValueError(err_msg + "1 given.")
if isinstance(weights, dict):
raise ValueError(err_msg + "Dict given.")
return
[docs]def validate_eval_set(eval_set, eval_name, X_train, y_train):
"""Check if the shapes of eval_set are compatible with (X_train, y_train).
Parameters
----------
eval_set : list of tuple
List of eval tuple set (X, y).
The last one is used for early stopping
eval_name : list of str
List of eval set names.
X_train : np.ndarray
Train owned products
y_train : np.array
Train targeted products
Returns
-------
eval_names : list of str
Validated list of eval_names.
eval_set : list of tuple
Validated list of eval_set.
"""
eval_name = eval_name or [f"val_{i}" for i in range(len(eval_set))]
assert len(eval_set) == len(
eval_name
), "eval_set and eval_name have not the same length"
if len(eval_set) > 0:
assert all(
len(elem) == 2 for elem in eval_set
), "Each tuple of eval_set need to have two elements"
for name, (X, y) in zip(eval_name, eval_set):
check_input(X)
msg = (
f"Dimension mismatch between X_{name} "
+ f"{X.shape} and X_train {X_train.shape}"
)
assert len(X.shape) == len(X_train.shape), msg
msg = (
f"Dimension mismatch between y_{name} "
+ f"{y.shape} and y_train {y_train.shape}"
)
assert len(y.shape) == len(y_train.shape), msg
msg = (
f"Number of columns is different between X_{name} "
+ f"({X.shape[1]}) and X_train ({X_train.shape[1]})"
)
assert X.shape[1] == X_train.shape[1], msg
if len(y_train.shape) == 2:
msg = (
f"Number of columns is different between y_{name} "
+ f"({y.shape[1]}) and y_train ({y_train.shape[1]})"
)
assert y.shape[1] == y_train.shape[1], msg
msg = (
f"You need the same number of rows between X_{name} "
+ f"({X.shape[0]}) and y_{name} ({y.shape[0]})"
)
assert X.shape[0] == y.shape[0], msg
return eval_name, eval_set
[docs]def define_device(device_name):
"""
Define the device to use during training and inference.
If auto it will detect automatically whether to use cuda or cpu
Parameters
----------
device_name : str
Either "auto", "cpu" or "cuda"
Returns
-------
str
Either "cpu" or "cuda"
"""
if device_name == "auto":
if torch.cuda.is_available():
return "cuda"
else:
return "cpu"
elif device_name == "cuda" and not torch.cuda.is_available():
return "cpu"
else:
return device_name
[docs]class ComplexEncoder(json.JSONEncoder):
[docs] def default(self, obj):
if isinstance(obj, (np.generic, np.ndarray)):
return obj.tolist()
# Let the base class default method raise the TypeError
return json.JSONEncoder.default(self, obj)
[docs]def check_warm_start(warm_start, from_unsupervised):
"""
Gives a warning about ambiguous usage of the two parameters.
"""
if warm_start and from_unsupervised is not None:
warn_msg = "warm_start=True and from_unsupervised != None: "
warn_msg = "warm_start will be ignore, training will start from unsupervised weights"
warnings.warn(warn_msg)
return
[docs]def check_embedding_parameters(cat_dims, cat_idxs, cat_emb_dim):
"""
Check parameters related to embeddings and rearrange them in a unique manner.
"""
if (cat_dims == []) ^ (cat_idxs == []):
if cat_dims == []:
msg = "If cat_idxs is non-empty, cat_dims must be defined as a list of same length."
else:
msg = "If cat_dims is non-empty, cat_idxs must be defined as a list of same length."
raise ValueError(msg)
elif len(cat_dims) != len(cat_idxs):
msg = "The lists cat_dims and cat_idxs must have the same length."
raise ValueError(msg)
if isinstance(cat_emb_dim, int):
cat_emb_dims = [cat_emb_dim] * len(cat_idxs)
else:
cat_emb_dims = cat_emb_dim
# check that all embeddings are provided
if len(cat_emb_dims) != len(cat_dims):
msg = f"""cat_emb_dim and cat_dims must be lists of same length, got {len(cat_emb_dims)}
and {len(cat_dims)}"""
raise ValueError(msg)
# Rearrange to get reproducible seeds with different ordering
if len(cat_idxs) > 0:
sorted_idxs = np.argsort(cat_idxs)
cat_dims = [cat_dims[i] for i in sorted_idxs]
cat_emb_dims = [cat_emb_dims[i] for i in sorted_idxs]
return cat_dims, cat_idxs, cat_emb_dims