The same feature dimension limitation when using PGExplainer

[transparency066]

The same feature dimension limitation when using PGExplainer

[BarclayII]

Could you add a Linear layer at the beginning of your model:

nn.Linear(in_feats, out_feats)

So that it transforms the input feature dimensions to the same size of later layers?

[transparency066]

So this should only be considered when designing the model? Can PGExplainer be used If a fixed-structure model is finished training?

Could you add a Linear layer at the beginning of your model:

nn.Linear(in_feats, out_feats)

So that it transforms the input feature dimensions to the same size of later layers?

[transparency066]

I am sorry. I misunderstand the meaning of the parameter 'num_feature', I change it to the feature dimension of the final representation of the node and it works.

[transparency066]

However, whether it is my own model or the official example, the edge_weight obtained is close to 0. Do I need to normalize it?

[Rhett-Ying]

which official example you're using? please share more details? which dataset? what are the arguments including hyperparameters?

[transparency066]

#%%
import torch as th
import torch.nn as nn
import dgl
from dgl.nn import GraphConv
from dgl.data import GINDataset
from dgl.dataloading import GraphDataLoader
from dgl.nn.pytorch.explain import PGExplainer
import numpy as np
from tqdm import tqdm

# Set the device to GPU 2
device = th.device("cuda:2" if th.cuda.is_available() else "cpu")

# Define the model
class Model(nn.Module):
    def __init__(self, in_feats, out_feats):
        super().__init__()
        self.conv = GraphConv(in_feats, out_feats)
        self.fc = nn.Linear(out_feats, out_feats)
        nn.init.xavier_uniform_(self.fc.weight)

    def forward(self, g, h, embed=False, edge_weight=None):
        h = self.conv(g, h, edge_weight=edge_weight)
        if embed:
            return h
        with g.local_scope():
            g.ndata["h"] = h
            hg = dgl.mean_nodes(g, "h")
            return self.fc(hg)

# Load dataset
data = GINDataset("MUTAG", self_loop=True)
dataloader = GraphDataLoader(data, batch_size=16, shuffle=True)

# Initialize the model and move it to the device
feat_size = data[0][0].ndata["attr"].shape[1]
model = Model(feat_size, data.gclasses).to(device)
criterion = nn.CrossEntropyLoss()
optimizer = th.optim.Adam(model.parameters(), lr=1e-2)

# Training loop
for epoch in tqdm(range(200)):
    for bg, labels in dataloader:
        bg, labels = bg.to(device), labels.to(device)
        preds = model(bg, bg.ndata["attr"].to(device))
        loss = criterion(preds, labels)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

# Initialize the explainer
explainer = PGExplainer(model, data.gclasses).to(device)

#%%
# Train the explainer
init_tmp, final_tmp = 5.0, 1.0
optimizer_exp = th.optim.Adam(explainer.parameters(), lr=0.01)
for epoch in tqdm(range(100)):
    tmp = float(init_tmp * np.power(final_tmp / init_tmp, epoch / 20))
    for bg, labels in dataloader:
        bg, labels = bg.to(device), labels.to(device)
        loss = explainer.train_step(bg, bg.ndata["attr"].to(device), tmp)
        optimizer_exp.zero_grad()
        loss.backward()
        optimizer_exp.step()

# Explain the prediction for graph 0
graph, l = data[0]
graph = graph.to(device)
graph_feat = graph.ndata.pop("attr").to(device)
probs, edge_weight = explainer.explain_graph(graph, graph_feat)

Here is the offical example. The edge_weight is like

tensor([9.2120e-13, 9.2120e-13, 6.8172e-13, 9.2120e-13, 5.4758e-13, 1.2444e-12,
        5.4758e-13, 5.4758e-13, 4.0203e-13, 2.3260e-13, 5.4758e-13, 1.2444e-12,
        1.2444e-12, 1.2444e-12, 7.1015e-13, 1.2444e-12, 7.1015e-13, 1.1197e-10,
        5.2539e-13, 4.0185e-13, 1.1197e-10, 1.5726e-10, 4.4363e-10, 5.7454e-10,
        1.5726e-10, 1.2444e-12, 1.2444e-12, 1.2444e-12, 7.1015e-13, 1.2444e-12,
        7.1015e-13, 5.2539e-13, 4.0185e-13, 4.0185e-13, 5.2539e-13, 5.2539e-13,
        6.8172e-13, 6.8172e-13, 4.0203e-13, 6.8172e-13, 5.2539e-13, 6.8172e-13,
        5.2539e-13, 7.1015e-13, 4.0185e-13, 4.0185e-13, 9.2120e-13, 7.1015e-13,
        1.2444e-12, 4.0185e-13, 4.0185e-13, 7.1015e-13, 4.0185e-13, 4.0185e-13,
        4.0185e-13, 7.1015e-13, 4.0185e-13, 7.1015e-13, 9.2120e-13, 1.2444e-12,
        9.2120e-13, 6.8172e-13, 6.8172e-13, 6.8172e-13, 9.2120e-13, 6.8172e-13,
        7.1015e-13, 9.2120e-13, 1.2444e-12, 4.4363e-10, 2.0517e-10, 2.0517e-10,
        1.1654e-10, 2.0517e-10, 2.0139e-11, 2.0517e-10, 2.0139e-11],
       device='cuda:2', grad_fn=<DivBackward0>)

I also try to modify feature dimension of the final node representation and change the second argument 'num_feature‘ correspondingly like this. Am i right?

class Model(nn.Module):
    def __init__(self, in_feats, out_feats):
        super().__init__()
        self.conv = GraphConv(in_feats, out_feats*2)
        self.fc = nn.Linear(out_feats*2, out_feats)
        nn.init.xavier_uniform_(self.fc.weight)

    def forward(self, g, h, embed=False, edge_weight=None):
        h = self.conv(g, h, edge_weight=edge_weight)
        if embed:
            return h
        with g.local_scope():
            g.ndata["h"] = h
            hg = dgl.mean_nodes(g, "h")
            return self.fc(hg)

explainer = PGExplainer(model, data.gclasses*2).to(device)

[Rhett-Ying]

@mufeili Hi could you help give some guidance on how to dive deep into this issue?

[github-actions]

This issue has been automatically marked as stale due to lack of activity. It will be closed if no further activity occurs. Thank you