what is the meaning of iterations in dataloader?

[ppsmk388]

I have found a word "iterations" in
https://pytorch-geometric.readthedocs.io/en/latest/modules/loader.html?highlight=HGTLoader#torch_geometric.loader.HGTLoader

It say “ Sample 512 nodes per type and per iteration for 4 iterations”，can you tell me what is the meaning of "iterations " in dataloader?

loader = HGTLoader(
hetero_data,
# Sample 512 nodes per type and per iteration for 4 iterations
num_samples={key: [512] * 4 for key in hetero_data.node_types},
# Use a batch size of 128 for sampling training nodes of type paper
batch_size=128,
input_nodes=('paper', hetero_data['paper'].train_mask),
)

[rusty1s]

The word iterations here mean how many steps are being used to sample neighbors. HGTLoader is a layer-wise sampling strategy, and so we will sample 512 nodes in 4 iterations/layers. Everytime we have sampled nodes in one layer, they will be added as seed nodes so that we can also sample their neighbors in the follow-up iteration.

[ppsmk388]

Thanks a lot！

[ppsmk388]

and for the code in https://github.com/pyg-team/pytorch_geometric/blob/master/examples/hetero/to_hetero_mag.py
I found this code:

train_loader = HGTLoader(data, num_samples=[1024] * 4, shuffle=True,
input_nodes=train_input_nodes, **kwargs)
val_loader = HGTLoader(data, num_samples=[1024] * 4,
input_nodes=val_input_nodes, **kwargs)

if the data from the training set and the data from the validation set all storage in the same graph :data

Does the data obtained from train_loader contain the data in the validation set?

[EdisonLeeeee]

There shall be an overlap on the neighborhoods between the data obtained from the two loaders.

[ppsmk388]

thank you！

[ppsmk388]

but what is the meaning of the train_mask in train_input_nodes = ('paper', data['paper'].train_mask)

[EdisonLeeeee]

The train_mask only specifies the root nodes (or seed nodes) in a sampled subgraph. For an inductive scenario, we only care about whether the root nodes overlap between training and test graphs. It's okay if there is an overlap between neighborhoods.

[ppsmk388]

thank you very much！！！

[ppsmk388]

Or the data obtained from train_loader only contain the data from training set(beacuse of the train_mask and val_mask)?
I found the code below in https://github.com/pyg-team/pytorch_geometric/blob/master/examples/hetero/to_hetero_mag.py

train_input_nodes = ('paper', data['paper'].train_mask)
val_input_nodes = ('paper', data['paper'].val_mask)

train_loader = HGTLoader(data, num_samples=[1024] * 4, shuffle=True,
input_nodes=train_input_nodes, **kwargs)
val_loader = HGTLoader(data, num_samples=[1024] * 4,
input_nodes=val_input_nodes, **kwargs)