- Version 0.3.2 brings i) bug fixes and ii) extension of
.split_data
method by including native DataLoader parameters likenum_workers
,pin_memory
,worker_init_fn
,multiprocessing_context
and others as per PyTorch documentation.
Spltr
is a simple PyTorch-based data loader and splitter.
It may be used to load i) arrays and ii) matrices or iii) Pandas
DataFrames and iv) CSV files containing numerical data with
subsequent split it into Train, Test (Validation)
subsets in
the form of PyTorch DataLoader
objects. The special emphesis was
given to ease of usage and automation of some frequent data-handling procedures.
Originally it was developed in order to speed up a data preparation stage for number of trivial ML tasks. Hope it may be useful for you as well.
Spltr.process_x|y
: converts loaded Input/Target data into PyTorch Tensors
with ability to i) preview, ii) define tensor dtype, iii) set the desired
device of returned tensor (CPU/GPU), iv) use selected rows/columns from
Input/Target data sources or process a single data table (for CSV and
Pandas DataFrame only).
Spltr.reshape_xy
: reshapes subsets.
Spltr.split_data
: splits data subsets into train, test (validation)
PyTorch DataLoader objects.
Spltr.clean_space
: optimizes memory by deleting unnecessary variables.
pip install spltr
OSI Approved :: MIT License
https://github.com/maksymsur/Spltr
- torch >= 1.1.0
- numpy >= 1.16.4
- pandas >= 0.24.2
Hereunder we'll build a simple neural network and describe how Spltr
may be used in the process.
STEP 1: Let's start with loading and reading an Iris dataset to be used as an example. The dataset may be found at: https://github.com/maksymsur/Spltr/blob/master/dataset/iris_num.csv
import pandas as pd
link = 'https://github.com/raw/maksymsur/spltr/master/dataset/iris_num.csv'
db = pd.read_csv(link)
print(db.info())
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 150 entries, 0 to 149
Data columns (total 5 columns):
sepal.length 150 non-null float64
sepal.width 150 non-null float64
petal.length 150 non-null float64
petal.width 150 non-null float64
variety 150 non-null int64
dtypes: float64(4), int64(1)
memory usage: 6.0 KB
None
By building a neural network we'll try to predict a 'variety' column. This column identifies Iris species: Setosa = 0, Versicolor = 1 and Virginica = 2. Let's verify that it is comprised from the mentioned 3 unique values.
print(db.variety.unique())
[0 1 2]
STEP 2: Making X
(Input) dataset and excluding the 5th 'variety' column (our Target).
base_db = db.iloc[:,:-1]
print(base_db.head())
sepal.length sepal.width petal.length petal.width
0 5.1 3.5 1.4 0.2
1 4.9 3.0 1.4 0.2
2 4.7 3.2 1.3 0.2
3 4.6 3.1 1.5 0.2
4 5.0 3.6 1.4 0.2
STEP 3: Building y
(Target) dataset out of 'variety' column.
target_db = db.iloc[:,-1]
print(target_db.head())
0 0
1 0
2 0
3 0
4 0
Name: variety, dtype: int64
STEP 4: Loading necessary packages including Spltr
. Note: for Spltr
to work properly torch
, numpy
and pandas
shall be installed as well.
import torch
import numpy as np
from spltr import Spltr
STEP 5.1: Now let's instantiate a Spltr object by including X,y
datasets into it.
No data normalization procedures will be applied just to avoid overloading of this example. Although, using different types of normalization and creating separate Spltr objects for every of the predefined train-test (validation) datasets may lead to an interesting training mechanics and results.
splt = Spltr(base_db,target_db)
STEP 5.2: Alternatively we may load whole datasets and apply basic preprocessing scenarios (exclude columns, change shape and type of the data, etc.). That allows to quickly iterate and find the best approach to current ML task. To demonstrate that, let's reinstantiate base_db
& target_db
base_db = pd.read_csv(link) # Now 'base_db' is a 5-column dataset that includes the 'target' (5th) column.
target_db = link # And 'target_db' is a simple link to the same 5-column dataset.
splt = Spltr(base_db,target_db)
Here we start with processing X
by selecting only 4 feature columns and, thus, excluding the target (5th column). Pls note that presented vocabulary-type selection {column start : column finish}
works only for DataFrames as for now.
splt.process_x(preview=True, usecols={0:4})
[X]: The following DataFrame was loaded:
sepal.length sepal.width petal.length petal.width
0 5.1 3.5 1.4 0.2
1 4.9 3.0 1.4 0.2
[X]: Tensor of shape torch.Size([150, 4]) was created
And continue with processing y
by using just the 5th column named 'variety'. Note that CSV columns may be selected by [Int] or [String] as per official pd.read_csv documentation.
splt.process_y(preview=True, usecols=['variety'])
[y]: The following CSV was loaded:
variety
0 0
1 0
[y]: Tensor of shape torch.Size([150, 1]) was created
splt.reshape_xy(y_shape=-1) # Reshaping 'y' to be easily used in a classification task
print(splt.shape_y_)
torch.Size([150])
STEP 6: Splitting the dataset into train - 30%, test - 20%, which makes validation set equal to 50% (calculated automatically: val = 100% - train - test), and initializing permutation of the data.
Spltr.split_data
method may use native DataLoader
methods like num_workers
, pin_memory
, worker_init_fn
, multiprocessing_context
and others as per PyTorch documentation.
splt.split_data(splits=(0.3,0.2), perm_seed=3, batch_size=1, num_workers = 1, shuffle=True)
[X,y]: The Data is splitted into 3 datasets of length: Train 45, Test 30, Validation 75.
STEP 7: Now let's clean unnecessary variables saved in the memory. This step may be especially useful if you are dealing with huge datasets and don't want for X,y tensors
to share memory with X,y DataLoader objects
.
splt.clean_space()
All variables are deleted. Only Train-Test (Validation) sets are left.
STEP 8: Setting up a very simple neural network. Pls mind that the presented network architecture is comprised only to demonstrate how Spltr
may be adopted. That's not an optimal way to solve an Iris classification problem.
Note that depending on your system configuration a bug may be encountered if using torch v.1.6
. It's a known issue caused by inaproproate behaviour of torch.max
function.
from torch import nn, optim
import torch.nn.functional as F
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.input = nn.Linear(4,8)
self.output = nn.Linear(8,3)
def forward(self,x):
x = F.softsign(self.input(x))
x = self.output(x)
return x
model = Net()
criterion = nn.CrossEntropyLoss(reduction='none')
optimizer = optim.RMSprop(model.parameters(), lr=0.01, alpha=0.9, eps=1e-08)
for n in range(7):
train_loss = 0.0
# Fitting a dataset for Training
for train_data, train_target in splt.xy_train:
model.zero_grad()
train_result = model(train_data.float())
loss = criterion(train_result, train_target)
train_loss += loss.item()
loss.backward()
optimizer.step()
test_correct = 0
with torch.no_grad():
# Fitting a dataset for Testing
for test_data, test_target in splt.xy_test:
test_result = model(test_data.float())
_, predicted = torch.max(test_result, 1)
test_correct += (predicted == test_target).sum().item()
print(f'Epoch {n+1}. Train loss: ', round(train_loss/len(splt.xy_train), 5))
print(f'Testing: Correctly identified samples {test_correct} out of {len(splt.xy_test)}', end='\n\n')
val_correct = 0
with torch.no_grad():
# Fitting a dataset for Validation
for val_data, val_target in splt.xy_val:
val_result = model(val_data.float())
_, val_predicted = torch.max(val_result, 1)
val_correct += (val_predicted == val_target).sum().item()
print(f'VALIDATION: Correctly identified samples {val_correct} out of {len(splt.xy_val)}', end='\n\n')
Epoch 1. Train loss: 1.09528
Testing: Correctly identified samples 21 out of 30
Epoch 2. Train loss: 0.80778
Testing: Correctly identified samples 21 out of 30
Epoch 3. Train loss: 0.58353
Testing: Correctly identified samples 24 out of 30
Epoch 4. Train loss: 0.48647
Testing: Correctly identified samples 26 out of 30
Epoch 5. Train loss: 0.43741
Testing: Correctly identified samples 21 out of 30
Epoch 6. Train loss: 0.41516
Testing: Correctly identified samples 27 out of 30
Epoch 7. Train loss: 0.38361
Testing: Correctly identified samples 30 out of 30
VALIDATION: Correctly identified samples 73 out of 75