def test_graphsage():
n_train = 1024
n_val = 256
train_uvvis = UVvis(masked=True)
train_uvvis.df = train_uvvis.df.iloc[:n_train, :]
mu, sigma = np.mean(train_uvvis.df['computational']), np.std(
train_uvvis.df['computational'])
train_uvvis.df['computational'] = (train_uvvis.df['computational'] -
mu) / sigma
val_uvvis = UVvis(masked=True)
val_uvvis.df = val_uvvis.df.iloc[n_train:n_train + n_val, :]
val_uvvis.df['computational'] = (val_uvvis.df['computational'] -
mu) / sigma
train_loader = DataLoader(train_uvvis, batch_size=64)
val_loader = DataLoader(val_uvvis, batch_size=64)
model = ToyGraphSAGEModel()
optimizer = optim.Adam(model.parameters())
MSE = nn.MSELoss()
print(model, '\n')
for epoch in range(256):
train_losses = []
train_metrics = []
model.train()
for idx, (A, x, masks, y_true) in enumerate(train_loader):
optimizer.zero_grad()
y_pred = model(A, x, masks)
loss = MSE(y_true, y_pred)
train_losses.append(loss.item())
y_pred = (y_pred * sigma) + mu
y_true = (y_true * sigma) + mu
train_metrics.append(nn.L1Loss()(y_true, y_pred).item())
loss.backward()
optimizer.step()
# break
# break
val_losses = []
val_metrics = []
model.eval()
for idx, (A, x, masks, y_true) in enumerate(val_loader):
y_pred = model(A, x, masks)
y_pred = (y_pred * sigma) + mu
y_true = (y_true * sigma) + mu
val_losses.append(MSE(y_true, y_pred).item())
val_metrics.append(nn.L1Loss()(y_true, y_pred).item())
print(
'Epoch: {} Loss: {:.4f} Train MAE: {:.4f} Val Loss: {:.4f} Val MAE: {:.4f}'
.format(epoch + 1, np.mean(train_losses), np.mean(train_metrics),
np.mean(val_losses), np.mean(val_metrics)))
def get_uvvis_dataloaders(batch_size=4):
train_uvvis = UVvis(masked=True)
train_uvvis.df = train_uvvis.df.iloc[:4096, :]
mu, sigma = np.mean(train_uvvis.df['computational']), np.std(
train_uvvis.df['computational'])
train_uvvis.df['computational'] = (train_uvvis.df['computational'] -
mu) / sigma
val_uvvis = UVvis(masked=True)
val_uvvis.df = val_uvvis.df.iloc[4096:4096 + 256, :]
val_uvvis.df['computational'] = (val_uvvis.df['computational'] -
mu) / sigma
train_loader = DataLoader(train_uvvis, batch_size=batch_size)
val_loader = DataLoader(val_uvvis, batch_size=batch_size)
return train_uvvis, val_uvvis, train_loader, val_loader
def test_kipf_conv():
train_uvvis = UVvis(masked=True)
train_uvvis.df = train_uvvis.df.iloc[:1024, :]
val_uvvis = UVvis(masked=True)
val_uvvis.df = val_uvvis.df.iloc[1024:1280, :]
train_loader = DataLoader(train_uvvis, batch_size=64)
val_loader = DataLoader(val_uvvis, batch_size=64)
model = ToyModel()
optimizer = optim.Adam(model.parameters())
MSE = nn.MSELoss()
print(model, '\n')
for epoch in range(256):
train_losses = []
train_metrics = []
model.train()
for idx, (A, x, masks, y_true) in enumerate(train_loader):
optimizer.zero_grad()
y_pred = model(A, x, masks)
loss = MSE(y_true, y_pred)
train_losses.append(loss.item())
train_metrics.append(nn.L1Loss()(y_true, y_pred).item())
loss.backward()
optimizer.step()
val_losses = []
val_metrics = []
model.eval()
for idx, (A, x, masks, y_true) in enumerate(val_loader):
y_pred = model(A, x, masks)
val_losses.append(MSE(y_true, y_pred).item())
val_metrics.append(nn.L1Loss()(y_true, y_pred).item())
print(
'Epoch: {} Loss: {:.4f} Train MAE: {:.4f} Val Loss: {:.4f} Val MAE: {:.4f}'
.format(epoch + 1, np.mean(train_losses), np.mean(train_metrics),
np.mean(val_losses), np.mean(val_metrics)))
def test_uvvis_dataset():
uvvis = UVvis(masked=True)
x, node_masks, A, y = uvvis[0]
print(x.shape, A.shape)
gc = SpectralGraphConv(in_features=121, out_features=64)
out = gc(A, x, node_masks)