Using multiple dataloaders in the training_step?

[christofer-f]

Hi!

In the pseudo-code below I have two models that I want to fit with two different datasets.
I have tried to figure out if this is possible by reading test_dataloaders.py with no success...

In the documentation, it states that:
Multiple training dataloaders
For training, the best way to use multiple-dataloaders is to create a Dataloader class which wraps both your dataloaders. (This of course also works for testing and validation dataloaders).

But that doesn't really help me...

I guess that this already has been discussed in: #1089
And that I should study: https://gist.github.com/Dref360/2524e524244569ed47428f19c487f264

But it would be nice with a dataloader_idx like just like the optimizer_idx parameter...

Or perhaps a batch could have a dictionary-like structure where you sample data into different "baskets"
so that I could write something like:

        if optimizer_idx == 0:
            # REQUIRED
            x, y = batch[0]

class FashionMNIST_and_MNISTModel(pl.LightningModule):

    def __init__(self):
        super(FashionMNIST_and_MNISTModel, self).__init__()

       # l1 should be fit to MNIST dataset
        self.l1 = torch.nn.Linear(28 * 28, 10)

       # l2 should be fit to FashionMNIST dataset
        self.l2 = torch.nn.Linear(28 * 28, 10)  

    def training_step(self, batch, batch_nb, optimizer_idx):
        if optimizer_idx == 0:
            # REQUIRED
            x, y = batch
            y_hat = torch.relu(self.l1(x.view(x.size(0), -1)))
            loss_l1 = F.cross_entropy(y_hat, y)
            tensorboard_logs = {'train_loss': loss_l1}
            return {'loss': loss_l1, 'log': tensorboard_logs}
        if optimizer_idx == 1:            
            # REQUIRED
            x, y = batch
            y_hat = torch.relu(self.l2(x.view(x.size(0), -1)))
            loss_l2 = F.cross_entropy(y_hat, y)
            tensorboard_logs = {'train_loss': loss_l2}
            return {'loss': loss_l2, 'log': tensorboard_logs}

    def train_dataloader(self):
        # REQUIRED
        return [
            DataLoader(MNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32),
            DataLoader(FashionMNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)
        ]

//Christofer

[github-actions]

Hi! thanks for your contribution!, great first issue!

[justusschock]

Hi @christofer-f, I've actually prototyped this feature already (in #1959). If your dataset has a length, this is already working (the failing tests are due to the case, where your dataset does not have a defined length).
Any feedback is highly appreciated :)

[christofer-f]

Hi! This looks very promising.

In my case, I have several processes that create datasets on the fly.
So the requirement with a minimum length is not a problem. At least not for me.

I will try to apply your code to the toy example above.

Br,
Christofer

[christofer-f]

This is exactly what I wanted!!! Great job.

pip install -e git://github.com/PyTorchLightning/pytorch-lightning.git@train_loaders#egg=pytorch-lightning_train_loaders

import os

import torch
from torch.nn import functional as F
from torch.utils.data import DataLoader
from torchvision.datasets import MNIST, FashionMNIST
from torchvision import transforms
import pytorch_lightning as pl

class FashionMNIST_and_MNISTModel(pl.LightningModule):

    def __init__(self):
        super(FashionMNIST_and_MNISTModel, self).__init__()

        self.l_mnist = torch.nn.Linear(28 * 28, 10)
        self.l_fashion_mnist = torch.nn.Linear(28 * 28, 10)        

    def forward(self, x):
        # called with self(x)
        return torch.relu(self.l_mnist(x.view(x.size(0), -1)))

    def training_step(self, batch, batch_idx, optimizer_idx):
        if optimizer_idx == 0:
            x, y = batch['mnist']
            y_hat = torch.relu(self.l_mnist(x.view(x.size(0), -1)))
            loss_mnist = F.cross_entropy(y_hat, y)
            tensorboard_logs = {'train_loss': loss_mnist}
            return {'loss': loss_mnist, 'log': tensorboard_logs}
        if optimizer_idx == 1:
            x, y = batch['fashion_mnist']
            y_hat = torch.relu(self.l_fashion_mnist(x.view(x.size(0), -1)))
            loss_fashion_mnist = F.cross_entropy(y_hat, y)
            tensorboard_logs = {'train_loss': loss_fashion_mnist}
            return {'loss': loss_fashion_mnist, 'log': tensorboard_logs}

    def configure_optimizers(self):
        opt_mnist = torch.optim.Adam(self.l_mnist.parameters(), lr=0.02)
        opt_fashion_mnist = torch.optim.Adam(self.l_fashion_mnist.parameters(), lr=0.02)
        return [opt_mnist, opt_fashion_mnist], []

    def train_dataloader(self):
        loader_mnist = DataLoader(MNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)
        loader_fashion_mnist = DataLoader(FashionMNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)       
        loaders = {"mnist": loader_mnist, "fashion_mnist": loader_fashion_mnist}
        return loaders


def main():
    mnist_model = FashionMNIST_and_MNISTModel()

    trainer = pl.Trainer(gpus=1, fast_dev_run=True)    
    trainer.fit(mnist_model)   

if __name__ == "__main__":
    main()

[justusschock]

great to hear that! So the feature should almost be ready to merge to master. There is also a trainer flag, that controls, how to deal with datasets of different lengths

[christofer-f]

I close this. It is a very good and useful feature. I need to tinker a bit with this before I understand how it really works...

[omiita]

Thank your for this wonderful work @justusschock !
I play around this new feature a bit, and I encounter some problem.
I used following code (from @christofer-f ) to check how many steps in one epoch.

pip install -e git://github.com/PyTorchLightning/pytorch-lightning.git@train_loaders#egg=pytorch-lightning_train_loaders

import os

import torch
from torch.nn import functional as F
from torch.utils.data import DataLoader
from torchvision.datasets import MNIST, FashionMNIST
from torchvision import transforms
import pytorch_lightning as pl

class FashionMNIST_and_MNISTModel(pl.LightningModule):
    def __init__(self):
        super(FashionMNIST_and_MNISTModel, self).__init__()

        self.l_mnist = torch.nn.Linear(28 * 28, 10)
        self.l_fashion_mnist = torch.nn.Linear(28 * 28, 10)        

    def forward(self, x):
        return torch.relu(self.l_mnist(x.view(x.size(0), -1)))

    def training_step(self, batch, batch_idx, optimizer_idx):
        if optimizer_idx == 0:
            print('mnist')
            x, y = batch['mnist']
            y_hat = self(x)
            loss_mnist = F.cross_entropy(y_hat, y)
            tensorboard_logs = {'train_loss': loss_mnist}
            return {'loss': loss_mnist, 'log': tensorboard_logs}
        if optimizer_idx == 1:
            print('fashion')
            x, y = batch['fashion_mnist']
            y_hat = torch.relu(self.l_fashion_mnist(x.view(x.size(0), -1)))
            loss_fashion_mnist = F.cross_entropy(y_hat, y)
            tensorboard_logs = {'train_loss': loss_fashion_mnist}
            return {'loss': loss_fashion_mnist, 'log': tensorboard_logs}

    def configure_optimizers(self):
        opt_mnist = torch.optim.Adam(self.l_mnist.parameters(), lr=0.02)
        opt_fashion_mnist = torch.optim.Adam(self.l_fashion_mnist.parameters(), lr=0.02)
        return [opt_mnist, opt_fashion_mnist], []

    def train_dataloader(self):
        loader_mnist = DataLoader(MNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)
        loader_fashion_mnist = DataLoader(FashionMNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)       
        loaders = {"mnist": loader_mnist, "fashion_mnist": loader_fashion_mnist}
        return loaders

mnist_model = FashionMNIST_and_MNISTModel()

trainer = pl.Trainer(gpus=0, fast_dev_run=False, max_epochs=1)    
trainer.fit(mnist_model)

And, I got the following output.

GPU available: False, used: False
TPU available: False, using: 0 TPU cores

  | Name            | Type   | Params
-------------------------------------------
0 | l_mnist         | Linear | 7 K   
1 | l_fashion_mnist | Linear | 7 K   

Epoch 1: 100%
2/2 [00:00<00:00, 30.03it/s, loss=2.272, v_num=16]

mnist
fashion
mnist
fashion

1

It seems there are only two steps in one epoch according to the print messages.
There should be
(#samples) / (batch_size) = (50,000) / (32) => 1563 steps
Am I missing something? If so, please correct me!

Thanks in advance!

[christofer-f]

Hi,

I think @omiita is right.
Comparing the training loops when only running the mnist dataset, the difference is obvious.
For details:
#1959 (comment)

[christofer-f]

I close this again... the problem has been identified...