Dataloader for siamese /triplet model

[kaka2nd]

❓ Questions & Help

Hi I am working on building siamese/ triplet model. In this case, I need to input pairs with labels (or triplet without labels). In original pytorch, an easy way is using data to output two imgaes like

    def __getitem__(self, index):

       doing sth

        return (img1, img2), target

Then I can use this to feed into models which accept two inputs like output1,output2 = net(img0,img1)

So I was wondering how I could implement the siamese/ triplet model in your library, which was output 2 or 3 x from the data? Or any indirect idea I can do this ?

Hope for your reply. Thanks.

[rusty1s]

There are two different approaches to do this:

1. Return two data objects in your __getitem__ function and use a custom collate function:

   
   def __getitem__(self, index):
   return data1, data2

from torch_geometric.data import Batch

def custom_collate(data_list): batch_1 = Batch.from_data_list([d[0] for d in data_list]) batch_2 = Batch.from_data_list([d[1] for d in data_list]) return batch_1, batch_2

from torch.utils.data import DataLoader loader = DataLoader(dataset, ..., collate_fn=custom_collate)

for batch_1, batch_2 in loader: ...

2. Put both graph into *a single* data object
```python
def __getitem__(self, index):
    data = Data(edge_index_1=..., edge_index_2=..., x_1=..., x_2=..., num_nodes=None)

and make use of the follow_batch argument in the DataLoader:

from torch_geometric.data import DataLoader
loader = DataLoader(dataset, ..., follow_batch=['x_1', 'x_2'])

for data in loader:
    ...

Note that we explicitly set num_nodes to None due to the data object holding two graphs. This will omit the creation of a data.batch assignment vector. Instead, the follow_batch argument will create two assignment vectors for each mini-batch named data.x_1_batch and data.x_2_batch respectively.

Edit: Another straightforward approach is to initialize two dataloaders:

from torch_geometric.data import DataLoader

loader_1 = DataLoader(dataset_1, ...)
loader_2 = DataLoader(dataset_2, ...)

for data_1, data_2 in zip(loader_1, loader_2):
    ...

[kaka2nd]

Wow! Thanks for the detailed reply. Have a good day (or night).

[gopuvenkat]

Thanks for the straightforward approach mentioned in the Edit. In this case, how is the common label (or target) for the involved graphs specified? One possible method to tackle this would be to keep the common label in both the data_1 and data_2 objects.

[rusty1s]

Yeah, that would be possible. With a more recent PyG release, it's also possible to do the following now:

class PairDataset(torch.utils.data.Dataset):
    def __init__(self, dataset1, dataset2, label):
        ...

    def __getitem__(self, idx):
        return dataset1[idx], dataset2[idx], label[idx]

and use the torch_geometric.data.DataLoader for batching:

loader = DataLoader(PairDataset(dataset1, dataset2, labels), batch_size=32, ...)

for data1, data2, label in loader:
    ...

[gopuvenkat]

Thanks for the quick reply. I tried the new approach and got the following error.

loader = DataLoader(PairDataset(dataset1, dataset2, labels), batch_size=32, ...)

for data1, data2, label in loader:
    ...

TypeError: object of type 'PairDataset' has no len()

Did I miss something?

[rusty1s]

You need to additionally implement the __len__ method (sorry for missing that):

def __len__(self):
    return len(dataset1)

[yyou1996]

I use pytorch_geometric==1.4.0 and method 1 works for me. Really great answer!