noise=0)
G_target_normalized = out_normalizer.normalize(G_target)
loss = build_graph_loss2(G_out, G_target_normalized)
loss.backward()
if step % 10 == 0:
writer.add_scalar('loss', loss.data.item(), step)
step += 1
for param in gn.parameters():
if param.grad is not None:
param.grad.clamp_(-3, 3)
optimizer.step()
schedular.step()
if step % 10000 == 0:
torch.save(gn.state_dict(),
savedir + '/model_{}.pth'.format(step))
iter = 0
sum_loss = 0
# evaluation loop, done every epoch
for i, data in tqdm(enumerate(dl_eval)):
action, delta_state, last_state = data
action, delta_state, last_state = action.float(), \
delta_state.float(), last_state.float()
if use_cuda:
action, delta_state, last_state = action.cuda(),\
delta_state.cuda(), last_state.cuda()
init_graph_features(G1,
G_target_normalized = out_normalizer.normalize(G_target)
loss = build_graph_loss2(G_out, G_target_normalized)
loss.backward()
if step % 10 == 0:
writer.add_scalar('loss', loss.data.item(), step)
step += 1
for param in gn.parameters():
if not param.grad is None:
param.grad.clamp_(-3,3)
optimizer.step()
schedular.step()
if step % 10000 == 0:
torch.save(
gn.state_dict(),
savedir +
'/model_{}.pth'.format(step))
iter = 0
sum_loss = 0
#evaluation loop, done every epoch
for i,data in tqdm(enumerate(dl_eval)):
action, delta_state, last_state = data
action, delta_state, last_state = action.float(), delta_state.float(), last_state.float()
if use_cuda:
action, delta_state, last_state = action.cuda(), delta_state.cuda(), last_state.cuda()
init_graph_features(G1, graph_feat_size, node_feat_size, edge_feat_size, cuda=True, bs = 200)
load_graph_features(G1, action, last_state, delta_state, bs=200, noise = 0)