I have a set of independent graphs that have features “degree” and “strat” (either 1 or 0), and I’m trying to predict feature * strat. I trained an MLP to do this based on the features of individual nodes. No problem.
I try to train a GCN to do the same thing:
But the loss doesn’t improve and on inspecting the logits, I notice they’re all very similar and nowhere close to the labels:
The GCN example works like this:
def forward(self, features):
And I edited it to be able to pass it a graph like this:
def forward(self, g, features):
And it breaks. I have no idea why. But I do need it to take a diff graph each time, so how do I accomplish that?
create a list of graphs in iterate over it
graphs = [g1, g2, ...]
for graph in graphs:
sgd.zero_grad()
n_feats = graph.ndata['feat']
labels = graph.ndata['label']
out = gcn(graph, n_feats)
loss = lloss_fnc(out, labels)
loss.backward()
sgd.step()
that’s what I do in training_loop
def train_loop(model, train_data, loss_fcn, optimizer, num_epochs=1):
dur = []
for epoch in range(num_epochs):
for step, g in enumerate(train_data):
model.train() # sets training mode rather than evaluation
if step >= 3:
t0 = time.time()
# forward
logits = model(g)
loss = loss_fcn(logits, labels)
optimizer.zero_grad()
loss.backward()
optimizer.step()
it fails to learn, as described
a batch full of graphs works perfectly, yet passing them sequentially doesn’t!
I need to pass them sequential to use GCN in RL. Any thoughts?
I narrowed the problem down so I made a new, clearer thread