Skip to content

Latest commit

 

History

History
executable file
·
158 lines (134 loc) · 7.26 KB

README.md

File metadata and controls

executable file
·
158 lines (134 loc) · 7.26 KB
Raw

Torchy Docs

Following is the official documentation of torchy, a pytorch wrapper, and all the additional functions of the wrapper. This documentation contains detail descriptions of all the arguments that can be passed on the new utilities of nn.Module as well as guides and tutorials to make the best use of torchy wrapper.

Using model.fit()

Argument Default Description
train_dataloader N/A Pass either the training dataset in form of TensorDataset or the DataLoader for training. Note: It's upon you to pass the validation dataloader if you choose to pass a training dataloader. You don't have to worry about validation dataloader if you passed the TensorDataset.
loss_fn N/A The loss function that should be used to calculate the loss function. Ideal loss function should be from nn.functional but a custom loss function that can handle tensors will work fine.
opt N/A The optimizer that the model should use to fine tune the parameters. Initialize optimizer form torch.optim, set appropirate hyperparameters and pass to .fit().
epochs N/A The number os epochs. Pass integer.
valid_dataloader None Pass a validation DataLoader if you have passed a training DataLoader instead of a Dataset.
valid_pct 30 The percentage of the training TensorDataset that should be the validation dataset. Usual values between 10 and 30.
batch_size 32 The batch size of the Training DataLoader that will be created when you upload a TensorDataset. Ignore if you passed a DataLoader by default.
accuracy False Either to calculate the accuracy of model or not. Pass Boolean True or False.
device CPU The device that the given model, and its dataset should be converted into. Only provide device if you passed a TensorDataset. Value should be any device available ('cpu' or'cuda').

Using other utility tools on torchy.nn.Module

model.get_loss()

Argument Description
batch The TensorDataset or batch or tuple containing the input and output tensors in the form of (x,y) whose loss is to be calculated
loss_fn The loss function from nn.functional should be passed.

model._accuracy()

Argument Description
labels The actual outputs.
preds The predicted outputs.

model.validate()

Argument Description
valid_dl The validation dataloader which should be used to do validation
loss_fn The loss function from nn.functional should be passed.

PS: When using model.validate() you don't need to do model.eval() beforehand as the method is already decorated with @torch.no_grad() and has self.eval() implemented whithin the method.

What additional utilities are in torchy.utils.data?

DeviceDL

DeviceDL is a helper tool to put your DataLoader into given device in the most efficient way. It's not recommended to put the entire DataLoader into given device in pytorch, so DeviceDL would only put the current batch in the specified device and pytorch will automatically remove the batch from device after its processed.

DeviceDL can be used in the following way:

from torchy.utils.data import DeviceDL, DataLoader, TensorDataset
# create a TensorDataset
dataset = TensorDataset(x,y)
# create your desired dataloader with the hyperparameters
dataloader = DataLoader(dataset, ...)
# now put your dataloader into the appropirate device using DeviceDL
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
dataloader_device = DeviceDL(dataloader, device)
Argument Description
dl The dataloader that you have created using for your model.
dev The device that the given dataloader should be kept in.

SplitPCT

Do you find it a meh experience when you have to count the number of examples you want to use in your training and validation set when using random_split? you don't want to install sklearn in your virtual environment just to split your dataset into train and validation sets?

If so, SplitPCT is a native pytorch's implementation to split the given dataset into training and validaiton set based on the percentage of data you want in training set and not the number of examples that the dataset should be divided into.

SplitPCT can be used in the following way:

from torchy.utils.data import SplitPCT, TensorDataset
# create a dataset
dataset = TensorDataset(x,y)
# determine what percentage of data should be on your training dataset
training_pct = 75
dataset = SplitPCT(dataset,training_pct)
# get training dataset and validation dataset as attributes
train_ds, validaiton_ds = dataset.train_ds, dataset.validation_ds
# get the original TensorDataset
orig_dataset = dataset.tensor_dataset

PS: the dataset passed to SplitPCT can be any type of pytorch's dataset and should not be limited to TensorDataset.

Argument Description
tensor_dataset The dataset that you have created using for your model.
train_pct The percentage of data that sould be on the training dataset; rest will be validation dataset.

Effect on torch.nn

By definition, torchy is a pytorch wrapper, so there will be no changes on torch.nn or any other torch functionality. torchy.nn can replacetorch.nn there will not be any unsolvable errors.

Torchy user guides and tutorials

Examples and quick start guide to use torchy can also be found on the project readme at github.

Since torchy is just a Wrapper and doesn't implement everything from scratch its recommended to implement just the nn.Module using the wrapper.

Recommended

import torch
import torchy.nn as nn
import torch.nn.functional as F

Make your models as you would using torch.nn

class Model(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(1, 1)

    def forward(self,x):
        return self.linear(x)


model = Model()

Choose your loss function and optimizers

loss_fn = F.mse_loss
opt = torch.optim.SGD(model.parameters(), lr=0.001)

Then, you can use torchy's DeviceDL to put your DataLoader in the given device

from torchy.utils.data import DeviceDL

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
train_dataloader = DeviceDL(old_dataloader, device)
valid_dataloader = DeviceDL(old_valid_dataloader, device)

Now you can fit the model

hist = model.fit(
    train_dataloader,
    loss_fn, 
    opt, 
    epochs, 
    valid_dataloader, 
    batch_size=64, 
    accuracy=True, 
    device=device
)

If you don't want to go through the hassle of making a dataloader then don't worry, torchy will do it for you.

PS: Torchy requires a TensorDataset to be passed for the following implementation to work.

hist = model.fit(
    dataset,
    loss_fn,
    opt,
    epochs,
    valic_pct=30,
    batch_size=64,
    accuracy=True,
    device=device
)
print(hist)

Looks pretty same and simple with the only change being you have to provide the percentage of the dataset that should be in the validation dataset and eventually the validation dataloader.

Why torchy is better than other pytorch .fit() implementations

Because the wheel doesn't need to be reinvented when using torchy. The end user can just use torchy as torch and just learn some new methods in the nn.Module that are handy.