dgnn = dgnn.to(device)
num_instance = len(train_data.sources)
num_batch = math.ceil(num_instance / BATCH_SIZE)
logger.info('num of training instances: {}'.format(num_instance))
logger.info('num of batches per epoch: {}'.format(num_batch))
idx_list = np.arange(num_instance)
new_nodes_val_aps = []
val_aps = []
epoch_times = []
total_epoch_times = []
train_losses = []
early_stopper = EarlyStopMonitor(max_round=args.patience)
for epoch in range(NUM_EPOCH):
start_epoch = time.time()
### Training
# Reinitialize memory of the model at the start of each epoch
dgnn.memory_s.__init_memory__()
dgnn.memory_g.__init_memory__()
# Train using only training graph
dgnn.set_neighbor_finder(train_ngh_finder)
m_loss = []
logger.info('start {} epoch'.format(epoch))
for k in range(0, num_batch, args.backprop_every):
loss = 0