GraphRNN/train.py

import networkx as nx
import numpy as np
import torch
import torch.nn as nn
import torch.nn.init as init
from torch.autograd import Variable
import matplotlib.pyplot as plt
import torch.nn.functional as F
from torch import optim
from torch.optim.lr_scheduler import MultiStepLR
from sklearn.decomposition import PCA
import logging
from torch.nn.utils.rnn import pad_packed_sequence, pack_padded_sequence
from time import gmtime, strftime
from sklearn.metrics import roc_curve
from sklearn.metrics import roc_auc_score
from sklearn.metrics import average_precision_score
from random import shuffle
import pickle
from tensorboard_logger import configure, log_value
import scipy.misc
import time as tm

from tensorboard_logger import log_value

from data import *


def train_vae_epoch(epoch, args, rnn, output, data_loader,
                    optimizer_rnn, optimizer_output,
                    scheduler_rnn, scheduler_output):
    rnn.train()
    output.train()
    loss_sum = 0
    for batch_idx, data in enumerate(data_loader):
        rnn.zero_grad()
        output.zero_grad()
        x_unsorted = data['x'].float()
        y_unsorted = data['y'].float()
        y_len_unsorted = data['len']
        y_len_max = max(y_len_unsorted)
        x_unsorted = x_unsorted[:, 0:y_len_max, :]
        y_unsorted = y_unsorted[:, 0:y_len_max, :]
        # initialize lstm hidden state according to batch size
        rnn.hidden = rnn.init_hidden(batch_size=x_unsorted.size(0))

        # sort input
        y_len, sort_index = torch.sort(y_len_unsorted, 0, descending=True)
        y_len = y_len.numpy().tolist()
        x = torch.index_select(x_unsorted, 0, sort_index)
        y = torch.index_select(y_unsorted, 0, sort_index)
        x = Variable(x).cuda()
        y = Variable(y).cuda()

        # if using ground truth to train
        h = rnn(x, pack=True, input_len=y_len)
        y_pred, z_mu, z_lsgms = output(h)
        y_pred = F.sigmoid(y_pred)
        # clean
        y_pred = pack_padded_sequence(y_pred, y_len, batch_first=True)
        y_pred = pad_packed_sequence(y_pred, batch_first=True)[0]
        z_mu = pack_padded_sequence(z_mu, y_len, batch_first=True)
        z_mu = pad_packed_sequence(z_mu, batch_first=True)[0]
        z_lsgms = pack_padded_sequence(z_lsgms, y_len, batch_first=True)
        z_lsgms = pad_packed_sequence(z_lsgms, batch_first=True)[0]
        # use cross entropy loss
        loss_bce = binary_cross_entropy_weight(y_pred, y)
        loss_kl = -0.5 * torch.sum(1 + z_lsgms - z_mu.pow(2) - z_lsgms.exp())
        loss_kl /= y.size(0) * y.size(1) * sum(y_len)  # normalize
        loss = loss_bce + loss_kl
        loss.backward()
        # update deterministic and lstm
        optimizer_output.step()
        optimizer_rnn.step()
        scheduler_output.step()
        scheduler_rnn.step()

        z_mu_mean = torch.mean(z_mu.data)
        z_sgm_mean = torch.mean(z_lsgms.mul(0.5).exp_().data)
        z_mu_min = torch.min(z_mu.data)
        z_sgm_min = torch.min(z_lsgms.mul(0.5).exp_().data)
        z_mu_max = torch.max(z_mu.data)
        z_sgm_max = torch.max(z_lsgms.mul(0.5).exp_().data)

        if epoch % args.epochs_log == 0 and batch_idx == 0:  # only output first batch's statistics
            print(
                'Epoch: {}/{}, train bce loss: {:.6f}, train kl loss: {:.6f}, graph type: {}, num_layer: {}, hidden: {}'.format(
                    epoch, args.epochs, loss_bce.data[0], loss_kl.data[0], args.graph_type, args.num_layers,
                    args.hidden_size_rnn))
            print('z_mu_mean', z_mu_mean, 'z_mu_min', z_mu_min, 'z_mu_max', z_mu_max, 'z_sgm_mean', z_sgm_mean,
                  'z_sgm_min', z_sgm_min, 'z_sgm_max', z_sgm_max)

        # logging
        log_value('bce_loss_' + args.fname, loss_bce.data[0], epoch * args.batch_ratio + batch_idx)
        log_value('kl_loss_' + args.fname, loss_kl.data[0], epoch * args.batch_ratio + batch_idx)
        log_value('z_mu_mean_' + args.fname, z_mu_mean, epoch * args.batch_ratio + batch_idx)
        log_value('z_mu_min_' + args.fname, z_mu_min, epoch * args.batch_ratio + batch_idx)
        log_value('z_mu_max_' + args.fname, z_mu_max, epoch * args.batch_ratio + batch_idx)
        log_value('z_sgm_mean_' + args.fname, z_sgm_mean, epoch * args.batch_ratio + batch_idx)
        log_value('z_sgm_min_' + args.fname, z_sgm_min, epoch * args.batch_ratio + batch_idx)
        log_value('z_sgm_max_' + args.fname, z_sgm_max, epoch * args.batch_ratio + batch_idx)

        loss_sum += loss.data[0]
    return loss_sum / (batch_idx + 1)


def test_vae_epoch(epoch, args, rnn, output, test_batch_size=16, save_histogram=False, sample_time=1):
    rnn.hidden = rnn.init_hidden(test_batch_size)
    rnn.eval()
    output.eval()

    # generate graphs
    max_num_node = int(args.max_num_node)
    y_pred = Variable(
        torch.zeros(test_batch_size, max_num_node, args.max_prev_node)).cuda()  # normalized prediction score
    y_pred_long = Variable(torch.zeros(test_batch_size, max_num_node, args.max_prev_node)).cuda()  # discrete prediction
    x_step = Variable(torch.ones(test_batch_size, 1, args.max_prev_node)).cuda()
    for i in range(max_num_node):
        h = rnn(x_step)
        y_pred_step, _, _ = output(h)
        y_pred[:, i:i + 1, :] = F.sigmoid(y_pred_step)
        x_step = sample_sigmoid(y_pred_step, sample=True, sample_time=sample_time)
        y_pred_long[:, i:i + 1, :] = x_step
        rnn.hidden = Variable(rnn.hidden.data).cuda()
    y_pred_data = y_pred.data
    y_pred_long_data = y_pred_long.data.long()

    # save graphs as pickle
    G_pred_list = []
    for i in range(test_batch_size):
        adj_pred = decode_adj(y_pred_long_data[i].cpu().numpy())
        G_pred = get_graph(adj_pred)  # get a graph from zero-padded adj
        G_pred_list.append(G_pred)

    # save prediction histograms, plot histogram over each time step
    # if save_histogram:
    #     save_prediction_histogram(y_pred_data.cpu().numpy(),
    #                           fname_pred=args.figure_prediction_save_path+args.fname_pred+str(epoch)+'.jpg',
    #                           max_num_node=max_num_node)

    return G_pred_list


def test_vae_partial_epoch(epoch, args, rnn, output, data_loader, save_histogram=False, sample_time=1):
    rnn.eval()
    output.eval()
    G_pred_list = []
    for batch_idx, data in enumerate(data_loader):
        x = data['x'].float()
        y = data['y'].float()
        y_len = data['len']
        test_batch_size = x.size(0)
        rnn.hidden = rnn.init_hidden(test_batch_size)
        # generate graphs
        max_num_node = int(args.max_num_node)
        y_pred = Variable(
            torch.zeros(test_batch_size, max_num_node, args.max_prev_node)).cuda()  # normalized prediction score
        y_pred_long = Variable(
            torch.zeros(test_batch_size, max_num_node, args.max_prev_node)).cuda()  # discrete prediction
        x_step = Variable(torch.ones(test_batch_size, 1, args.max_prev_node)).cuda()
        for i in range(max_num_node):
            print('finish node', i)
            h = rnn(x_step)
            y_pred_step, _, _ = output(h)
            y_pred[:, i:i + 1, :] = F.sigmoid(y_pred_step)
            x_step = sample_sigmoid_supervised(y_pred_step, y[:, i:i + 1, :].cuda(), current=i, y_len=y_len,
                                               sample_time=sample_time)

            y_pred_long[:, i:i + 1, :] = x_step
            rnn.hidden = Variable(rnn.hidden.data).cuda()
        y_pred_data = y_pred.data
        y_pred_long_data = y_pred_long.data.long()

        # save graphs as pickle
        for i in range(test_batch_size):
            adj_pred = decode_adj(y_pred_long_data[i].cpu().numpy())
            G_pred = get_graph(adj_pred)  # get a graph from zero-padded adj
            G_pred_list.append(G_pred)
    return G_pred_list


def train_mlp_epoch(epoch, args, rnn, output, data_loader,
                    optimizer_rnn, optimizer_output,
                    scheduler_rnn, scheduler_output):
    rnn.train()
    output.train()
    loss_sum = 0
    for batch_idx, data in enumerate(data_loader):
        rnn.zero_grad()
        output.zero_grad()
        x_unsorted = data['x'].float()
        y_unsorted = data['y'].float()
        y_len_unsorted = data['len']
        y_len_max = max(y_len_unsorted)
        x_unsorted = x_unsorted[:, 0:y_len_max, :]
        y_unsorted = y_unsorted[:, 0:y_len_max, :]
        # initialize lstm hidden state according to batch size
        rnn.hidden = rnn.init_hidden(batch_size=x_unsorted.size(0))

        # sort input
        y_len, sort_index = torch.sort(y_len_unsorted, 0, descending=True)
        y_len = y_len.numpy().tolist()
        x = torch.index_select(x_unsorted, 0, sort_index)
        y = torch.index_select(y_unsorted, 0, sort_index)
        x = Variable(x).cuda()
        y = Variable(y).cuda()

        h = rnn(x, pack=True, input_len=y_len)
        y_pred = output(h)
        y_pred = F.sigmoid(y_pred)
        # clean
        y_pred = pack_padded_sequence(y_pred, y_len, batch_first=True)
        y_pred = pad_packed_sequence(y_pred, batch_first=True)[0]
        # use cross entropy loss
        loss = binary_cross_entropy_weight(y_pred, y)
        loss.backward()
        # update deterministic and lstm
        optimizer_output.step()
        optimizer_rnn.step()
        scheduler_output.step()
        scheduler_rnn.step()

        if epoch % args.epochs_log == 0 and batch_idx == 0:  # only output first batch's statistics
            print('Epoch: {}/{}, train loss: {:.6f}, graph type: {}, num_layer: {}, hidden: {}'.format(
                epoch, args.epochs, loss.data[0], args.graph_type, args.num_layers, args.hidden_size_rnn))

        # logging
        log_value('loss_' + args.fname, loss.data[0], epoch * args.batch_ratio + batch_idx)

        loss_sum += loss.data[0]
    return loss_sum / (batch_idx + 1)


def test_mlp_epoch(epoch, args, rnn, output, test_batch_size=16, save_histogram=False, sample_time=1):
    rnn.hidden = rnn.init_hidden(test_batch_size)
    rnn.eval()
    output.eval()

    # generate graphs
    max_num_node = int(args.max_num_node)
    y_pred = Variable(
        torch.zeros(test_batch_size, max_num_node, args.max_prev_node)).cuda()  # normalized prediction score
    y_pred_long = Variable(torch.zeros(test_batch_size, max_num_node, args.max_prev_node)).cuda()  # discrete prediction
    x_step = Variable(torch.ones(test_batch_size, 1, args.max_prev_node)).cuda()
    for i in range(max_num_node):
        h = rnn(x_step)
        y_pred_step = output(h)
        y_pred[:, i:i + 1, :] = F.sigmoid(y_pred_step)
        x_step = sample_sigmoid(y_pred_step, sample=True, sample_time=sample_time)
        y_pred_long[:, i:i + 1, :] = x_step
        rnn.hidden = Variable(rnn.hidden.data).cuda()
    y_pred_data = y_pred.data
    y_pred_long_data = y_pred_long.data.long()

    # save graphs as pickle
    G_pred_list = []
    for i in range(test_batch_size):
        adj_pred = decode_adj(y_pred_long_data[i].cpu().numpy())
        G_pred = get_graph(adj_pred)  # get a graph from zero-padded adj
        G_pred_list.append(G_pred)

    # # save prediction histograms, plot histogram over each time step
    # if save_histogram:
    #     save_prediction_histogram(y_pred_data.cpu().numpy(),
    #                           fname_pred=args.figure_prediction_save_path+args.fname_pred+str(epoch)+'.jpg',
    #                           max_num_node=max_num_node)
    return G_pred_list


def test_mlp_partial_epoch(epoch, args, rnn, output, data_loader, save_histogram=False, sample_time=1):
    rnn.eval()
    output.eval()
    G_pred_list = []
    for batch_idx, data in enumerate(data_loader):
        x = data['x'].float()
        y = data['y'].float()
        y_len = data['len']
        test_batch_size = x.size(0)
        rnn.hidden = rnn.init_hidden(test_batch_size)
        # generate graphs
        max_num_node = int(args.max_num_node)
        y_pred = Variable(
            torch.zeros(test_batch_size, max_num_node, args.max_prev_node)).cuda()  # normalized prediction score
        y_pred_long = Variable(
            torch.zeros(test_batch_size, max_num_node, args.max_prev_node)).cuda()  # discrete prediction
        x_step = Variable(torch.ones(test_batch_size, 1, args.max_prev_node)).cuda()
        for i in range(max_num_node):
            print('finish node', i)
            h = rnn(x_step)
            y_pred_step = output(h)
            y_pred[:, i:i + 1, :] = F.sigmoid(y_pred_step)
            x_step = sample_sigmoid_supervised(y_pred_step, y[:, i:i + 1, :].cuda(), current=i, y_len=y_len,
                                               sample_time=sample_time)

            y_pred_long[:, i:i + 1, :] = x_step
            rnn.hidden = Variable(rnn.hidden.data).cuda()
        y_pred_data = y_pred.data
        y_pred_long_data = y_pred_long.data.long()

        # save graphs as pickle
        for i in range(test_batch_size):
            adj_pred = decode_adj(y_pred_long_data[i].cpu().numpy())
            G_pred = get_graph(adj_pred)  # get a graph from zero-padded adj
            G_pred_list.append(G_pred)
    return G_pred_list


def test_mlp_partial_simple_epoch(epoch, args, rnn, output, data_loader, save_histogram=False, sample_time=1):
    rnn.eval()
    output.eval()
    G_pred_list = []
    for batch_idx, data in enumerate(data_loader):
        x = data['x'].float()
        y = data['y'].float()
        y_len = data['len']
        test_batch_size = x.size(0)
        rnn.hidden = rnn.init_hidden(test_batch_size)
        # generate graphs
        max_num_node = int(args.max_num_node)
        y_pred = Variable(
            torch.zeros(test_batch_size, max_num_node, args.max_prev_node)).cuda()  # normalized prediction score
        y_pred_long = Variable(
            torch.zeros(test_batch_size, max_num_node, args.max_prev_node)).cuda()  # discrete prediction
        x_step = Variable(torch.ones(test_batch_size, 1, args.max_prev_node)).cuda()
        for i in range(max_num_node):
            print('finish node', i)
            h = rnn(x_step)
            y_pred_step = output(h)
            y_pred[:, i:i + 1, :] = F.sigmoid(y_pred_step)
            x_step = sample_sigmoid_supervised_simple(y_pred_step, y[:, i:i + 1, :].cuda(), current=i, y_len=y_len,
                                                      sample_time=sample_time)

            y_pred_long[:, i:i + 1, :] = x_step
            rnn.hidden = Variable(rnn.hidden.data).cuda()
        y_pred_data = y_pred.data
        y_pred_long_data = y_pred_long.data.long()

        # save graphs as pickle
        for i in range(test_batch_size):
            adj_pred = decode_adj(y_pred_long_data[i].cpu().numpy())
            G_pred = get_graph(adj_pred)  # get a graph from zero-padded adj
            G_pred_list.append(G_pred)
    return G_pred_list


def train_mlp_forward_epoch(epoch, args, rnn, output, data_loader):
    rnn.train()
    output.train()
    loss_sum = 0
    for batch_idx, data in enumerate(data_loader):
        rnn.zero_grad()
        output.zero_grad()
        x_unsorted = data['x'].float()
        y_unsorted = data['y'].float()
        y_len_unsorted = data['len']
        y_len_max = max(y_len_unsorted)
        x_unsorted = x_unsorted[:, 0:y_len_max, :]
        y_unsorted = y_unsorted[:, 0:y_len_max, :]
        # initialize lstm hidden state according to batch size
        rnn.hidden = rnn.init_hidden(batch_size=x_unsorted.size(0))

        # sort input
        y_len, sort_index = torch.sort(y_len_unsorted, 0, descending=True)
        y_len = y_len.numpy().tolist()
        x = torch.index_select(x_unsorted, 0, sort_index)
        y = torch.index_select(y_unsorted, 0, sort_index)
        x = Variable(x).cuda()
        y = Variable(y).cuda()

        h = rnn(x, pack=True, input_len=y_len)
        y_pred = output(h)
        y_pred = F.sigmoid(y_pred)
        # clean
        y_pred = pack_padded_sequence(y_pred, y_len, batch_first=True)
        y_pred = pad_packed_sequence(y_pred, batch_first=True)[0]
        # use cross entropy loss

        loss = 0
        for j in range(y.size(1)):
            # print('y_pred',y_pred[0,j,:],'y',y[0,j,:])
            end_idx = min(j + 1, y.size(2))
            loss += binary_cross_entropy_weight(y_pred[:, j, 0:end_idx], y[:, j, 0:end_idx]) * end_idx

        if epoch % args.epochs_log == 0 and batch_idx == 0:  # only output first batch's statistics
            print('Epoch: {}/{}, train loss: {:.6f}, graph type: {}, num_layer: {}, hidden: {}'.format(
                epoch, args.epochs, loss.data[0], args.graph_type, args.num_layers, args.hidden_size_rnn))

        # logging
        log_value('loss_' + args.fname, loss.data[0], epoch * args.batch_ratio + batch_idx)

        loss_sum += loss.data[0]
    return loss_sum / (batch_idx + 1)


## too complicated, deprecated
# def test_mlp_partial_bfs_epoch(epoch, args, rnn, output, data_loader, save_histogram=False,sample_time=1):
#     rnn.eval()
#     output.eval()
#     G_pred_list = []
#     for batch_idx, data in enumerate(data_loader):
#         x = data['x'].float()
#         y = data['y'].float()
#         y_len = data['len']
#         test_batch_size = x.size(0)
#         rnn.hidden = rnn.init_hidden(test_batch_size)
#         # generate graphs
#         max_num_node = int(args.max_num_node)
#         y_pred = Variable(torch.zeros(test_batch_size, max_num_node, args.max_prev_node)).cuda() # normalized prediction score
#         y_pred_long = Variable(torch.zeros(test_batch_size, max_num_node, args.max_prev_node)).cuda() # discrete prediction
#         x_step = Variable(torch.ones(test_batch_size,1,args.max_prev_node)).cuda()
#         for i in range(max_num_node):
#             # 1 back up hidden state
#             hidden_prev = Variable(rnn.hidden.data).cuda()
#             h = rnn(x_step)
#             y_pred_step = output(h)
#             y_pred[:, i:i + 1, :] = F.sigmoid(y_pred_step)
#             x_step = sample_sigmoid_supervised(y_pred_step, y[:,i:i+1,:].cuda(), current=i, y_len=y_len, sample_time=sample_time)
#             y_pred_long[:, i:i + 1, :] = x_step
#
#             rnn.hidden = Variable(rnn.hidden.data).cuda()
#
#             print('finish node', i)
#         y_pred_data = y_pred.data
#         y_pred_long_data = y_pred_long.data.long()
#
#         # save graphs as pickle
#         for i in range(test_batch_size):
#             adj_pred = decode_adj(y_pred_long_data[i].cpu().numpy())
#             G_pred = get_graph(adj_pred) # get a graph from zero-padded adj
#             G_pred_list.append(G_pred)
#     return G_pred_list


def train_rnn_epoch(epoch, args, rnn, output, data_loader,
                    optimizer_rnn, optimizer_output,
                    scheduler_rnn, scheduler_output):
    rnn.train()
    output.train()
    loss_sum = 0
    for batch_idx, data in enumerate(data_loader):
        rnn.zero_grad()
        output.zero_grad()
        x_unsorted = data['x'].float()
        y_unsorted = data['y'].float()
        y_len_unsorted = data['len']
        y_len_max = max(y_len_unsorted)
        x_unsorted = x_unsorted[:, 0:y_len_max, :]
        y_unsorted = y_unsorted[:, 0:y_len_max, :]
        # initialize lstm hidden state according to batch size
        rnn.hidden = rnn.init_hidden(batch_size=x_unsorted.size(0))
        # output.hidden = output.init_hidden(batch_size=x_unsorted.size(0)*x_unsorted.size(1))

        # sort input
        y_len, sort_index = torch.sort(y_len_unsorted, 0, descending=True)
        y_len = y_len.numpy().tolist()
        x = torch.index_select(x_unsorted, 0, sort_index)
        y = torch.index_select(y_unsorted, 0, sort_index)

        # input, output for output rnn module
        # a smart use of pytorch builtin function: pack variable--b1_l1,b2_l1,...,b1_l2,b2_l2,...
        y_reshape = pack_padded_sequence(y, y_len, batch_first=True).data
        # reverse y_reshape, so that their lengths are sorted, add dimension
        idx = [i for i in range(y_reshape.size(0) - 1, -1, -1)]
        idx = torch.LongTensor(idx)
        y_reshape = y_reshape.index_select(0, idx)
        y_reshape = y_reshape.view(y_reshape.size(0), y_reshape.size(1), 1)

        output_x = torch.cat((torch.ones(y_reshape.size(0), 1, 1), y_reshape[:, 0:-1, 0:1]), dim=1)
        output_y = y_reshape
        # batch size for output module: sum(y_len)
        output_y_len = []
        output_y_len_bin = np.bincount(np.array(y_len))
        for i in range(len(output_y_len_bin) - 1, 0, -1):
            count_temp = np.sum(output_y_len_bin[i:])  # count how many y_len is above i
            output_y_len.extend(
                [min(i, y.size(2))] * count_temp)  # put them in output_y_len; max value should not exceed y.size(2)
        # pack into variable
        x = Variable(x).cuda()
        y = Variable(y).cuda()
        output_x = Variable(output_x).cuda()
        output_y = Variable(output_y).cuda()
        # print(output_y_len)
        # print('len',len(output_y_len))
        # print('y',y.size())
        # print('output_y',output_y.size())

        # if using ground truth to train
        h = rnn(x, pack=True, input_len=y_len)
        h = pack_padded_sequence(h, y_len, batch_first=True).data  # get packed hidden vector
        # reverse h
        idx = [i for i in range(h.size(0) - 1, -1, -1)]
        idx = Variable(torch.LongTensor(idx)).cuda()
        h = h.index_select(0, idx)
        hidden_null = Variable(torch.zeros(args.num_layers - 1, h.size(0), h.size(1))).cuda()
        output.hidden = torch.cat((h.view(1, h.size(0), h.size(1)), hidden_null),
                                  dim=0)  # num_layers, batch_size, hidden_size
        y_pred = output(output_x, pack=True, input_len=output_y_len)
        y_pred = F.sigmoid(y_pred)
        # clean
        y_pred = pack_padded_sequence(y_pred, output_y_len, batch_first=True)
        y_pred = pad_packed_sequence(y_pred, batch_first=True)[0]
        output_y = pack_padded_sequence(output_y, output_y_len, batch_first=True)
        output_y = pad_packed_sequence(output_y, batch_first=True)[0]
        # use cross entropy loss
        loss = binary_cross_entropy_weight(y_pred, output_y)
        loss.backward()
        # update deterministic and lstm
        optimizer_output.step()
        optimizer_rnn.step()
        scheduler_output.step()
        scheduler_rnn.step()

        if epoch % args.epochs_log == 0 and batch_idx == 0:  # only output first batch's statistics
            print('Epoch: {}/{}, train loss: {:.6f}, graph type: {}, num_layer: {}, hidden: {}'.format(
                epoch, args.epochs, loss.data, args.graph_type, args.num_layers, args.hidden_size_rnn))

        # logging
        log_value('loss_' + args.fname, loss.data, epoch * args.batch_ratio + batch_idx)
        feature_dim = y.size(1) * y.size(2)
        loss_sum += loss.data * feature_dim
    return loss_sum / (batch_idx + 1)


def test_rnn_epoch(epoch, args, rnn, output, test_batch_size=16):
    rnn.hidden = rnn.init_hidden(test_batch_size)
    rnn.eval()
    output.eval()

    # generate graphs
    max_num_node = int(args.max_num_node)
    y_pred_long = Variable(torch.zeros(test_batch_size, max_num_node, args.max_prev_node)).cuda()  # discrete prediction
    x_step = Variable(torch.ones(test_batch_size, 1, args.max_prev_node)).cuda()
    for i in range(max_num_node):
        h = rnn(x_step)
        # output.hidden = h.permute(1,0,2)
        hidden_null = Variable(torch.zeros(args.num_layers - 1, h.size(0), h.size(2))).cuda()
        output.hidden = torch.cat((h.permute(1, 0, 2), hidden_null),
                                  dim=0)  # num_layers, batch_size, hidden_size
        x_step = Variable(torch.zeros(test_batch_size, 1, args.max_prev_node)).cuda()
        output_x_step = Variable(torch.ones(test_batch_size, 1, 1)).cuda()
        for j in range(min(args.max_prev_node, i + 1)):
            output_y_pred_step = output(output_x_step)
            output_x_step = sample_sigmoid(output_y_pred_step, sample=True, sample_time=1)
            x_step[:, :, j:j + 1] = output_x_step
            output.hidden = Variable(output.hidden.data).cuda()
        y_pred_long[:, i:i + 1, :] = x_step
        rnn.hidden = Variable(rnn.hidden.data).cuda()
    y_pred_long_data = y_pred_long.data.long()

    # save graphs as pickle
    G_pred_list = []
    for i in range(test_batch_size):
        adj_pred = decode_adj(y_pred_long_data[i].cpu().numpy())
        G_pred = get_graph(adj_pred)  # get a graph from zero-padded adj
        G_pred_list.append(G_pred)

    return G_pred_list


def train_rnn_forward_epoch(epoch, args, rnn, output, data_loader):
    rnn.train()
    output.train()
    loss_sum = 0
    for batch_idx, data in enumerate(data_loader):
        rnn.zero_grad()
        output.zero_grad()
        x_unsorted = data['x'].float()
        y_unsorted = data['y'].float()
        y_len_unsorted = data['len']
        y_len_max = max(y_len_unsorted)
        x_unsorted = x_unsorted[:, 0:y_len_max, :]
        y_unsorted = y_unsorted[:, 0:y_len_max, :]
        # initialize lstm hidden state according to batch size
        rnn.hidden = rnn.init_hidden(batch_size=x_unsorted.size(0))
        # output.hidden = output.init_hidden(batch_size=x_unsorted.size(0)*x_unsorted.size(1))

        # sort input
        y_len, sort_index = torch.sort(y_len_unsorted, 0, descending=True)
        y_len = y_len.numpy().tolist()
        x = torch.index_select(x_unsorted, 0, sort_index)
        y = torch.index_select(y_unsorted, 0, sort_index)

        # input, output for output rnn module
        # a smart use of pytorch builtin function: pack variable--b1_l1,b2_l1,...,b1_l2,b2_l2,...
        y_reshape = pack_padded_sequence(y, y_len, batch_first=True).data
        # reverse y_reshape, so that their lengths are sorted, add dimension
        idx = [i for i in range(y_reshape.size(0) - 1, -1, -1)]
        idx = torch.LongTensor(idx)
        y_reshape = y_reshape.index_select(0, idx)
        y_reshape = y_reshape.view(y_reshape.size(0), y_reshape.size(1), 1)

        output_x = torch.cat((torch.ones(y_reshape.size(0), 1, 1), y_reshape[:, 0:-1, 0:1]), dim=1)
        output_y = y_reshape
        # batch size for output module: sum(y_len)
        output_y_len = []
        output_y_len_bin = np.bincount(np.array(y_len))
        for i in range(len(output_y_len_bin) - 1, 0, -1):
            count_temp = np.sum(output_y_len_bin[i:])  # count how many y_len is above i
            output_y_len.extend(
                [min(i, y.size(2))] * count_temp)  # put them in output_y_len; max value should not exceed y.size(2)
        # pack into variable
        x = Variable(x).cuda()
        y = Variable(y).cuda()
        output_x = Variable(output_x).cuda()
        output_y = Variable(output_y).cuda()
        # print(output_y_len)
        # print('len',len(output_y_len))
        # print('y',y.size())
        # print('output_y',output_y.size())

        # if using ground truth to train
        h = rnn(x, pack=True, input_len=y_len)
        h = pack_padded_sequence(h, y_len, batch_first=True).data  # get packed hidden vector
        # reverse h
        idx = [i for i in range(h.size(0) - 1, -1, -1)]
        idx = Variable(torch.LongTensor(idx)).cuda()
        h = h.index_select(0, idx)
        hidden_null = Variable(torch.zeros(args.num_layers - 1, h.size(0), h.size(1))).cuda()
        output.hidden = torch.cat((h.view(1, h.size(0), h.size(1)), hidden_null),
                                  dim=0)  # num_layers, batch_size, hidden_size
        y_pred = output(output_x, pack=True, input_len=output_y_len)
        y_pred = F.sigmoid(y_pred)
        # clean
        y_pred = pack_padded_sequence(y_pred, output_y_len, batch_first=True)
        y_pred = pad_packed_sequence(y_pred, batch_first=True)[0]
        output_y = pack_padded_sequence(output_y, output_y_len, batch_first=True)
        output_y = pad_packed_sequence(output_y, batch_first=True)[0]
        # use cross entropy loss
        loss = binary_cross_entropy_weight(y_pred, output_y)

        if epoch % args.epochs_log == 0 and batch_idx == 0:  # only output first batch's statistics
            print('Epoch: {}/{}, train loss: {:.6f}, graph type: {}, num_layer: {}, hidden: {}'.format(
                epoch, args.epochs, loss.data, args.graph_type, args.num_layers, args.hidden_size_rnn))

        # logging
        log_value('loss_' + args.fname, loss.data, epoch * args.batch_ratio + batch_idx)
        # print(y_pred.size())
        feature_dim = y_pred.size(0) * y_pred.size(1)
        loss_sum += loss.data * feature_dim / y.size(0)
    return loss_sum / (batch_idx + 1)


########### train function for LSTM + VAE
def train(args, dataset_train, rnn, output):
    # check if load existing model
    if args.load:
        fname = args.model_save_path + args.fname + 'lstm_' + str(args.load_epoch) + '.dat'
        rnn.load_state_dict(torch.load(fname))
        fname = args.model_save_path + args.fname + 'output_' + str(args.load_epoch) + '.dat'
        output.load_state_dict(torch.load(fname))

        args.lr = 0.00001
        epoch = args.load_epoch
        print('model loaded!, lr: {}'.format(args.lr))
    else:
        epoch = 1

    # initialize optimizer
    optimizer_rnn = optim.Adam(list(rnn.parameters()), lr=args.lr)
    optimizer_output = optim.Adam(list(output.parameters()), lr=args.lr)

    scheduler_rnn = MultiStepLR(optimizer_rnn, milestones=args.milestones, gamma=args.lr_rate)
    scheduler_output = MultiStepLR(optimizer_output, milestones=args.milestones, gamma=args.lr_rate)

    # start main loop
    time_all = np.zeros(args.epochs)
    while epoch <= args.epochs:
        time_start = tm.time()
        # train
        if 'GraphRNN_VAE' in args.note:
            train_vae_epoch(epoch, args, rnn, output, dataset_train,
                            optimizer_rnn, optimizer_output,
                            scheduler_rnn, scheduler_output)
        elif 'GraphRNN_MLP' in args.note:
            train_mlp_epoch(epoch, args, rnn, output, dataset_train,
                            optimizer_rnn, optimizer_output,
                            scheduler_rnn, scheduler_output)
        elif 'GraphRNN_RNN' in args.note:
            train_rnn_epoch(epoch, args, rnn, output, dataset_train,
                            optimizer_rnn, optimizer_output,
                            scheduler_rnn, scheduler_output)
        time_end = tm.time()
        time_all[epoch - 1] = time_end - time_start
        # test
        if epoch % args.epochs_test == 0 and epoch >= args.epochs_test_start:
            for sample_time in range(1, 4):
                G_pred = []
                while len(G_pred) < args.test_total_size:
                    if 'GraphRNN_VAE' in args.note:
                        G_pred_step = test_vae_epoch(epoch, args, rnn, output, test_batch_size=args.test_batch_size,
                                                     sample_time=sample_time)
                    elif 'GraphRNN_MLP' in args.note:
                        G_pred_step = test_mlp_epoch(epoch, args, rnn, output, test_batch_size=args.test_batch_size,
                                                     sample_time=sample_time)
                    elif 'GraphRNN_RNN' in args.note:
                        G_pred_step = test_rnn_epoch(epoch, args, rnn, output, test_batch_size=args.test_batch_size)
                    G_pred.extend(G_pred_step)
                # save graphs
                fname = args.graph_save_path + args.fname_pred + str(epoch) + '_' + str(sample_time) + '.dat'
                save_graph_list(G_pred, fname)
                if 'GraphRNN_RNN' in args.note:
                    break
            print('test done, graphs saved')

        # save model checkpoint
        if args.save:
            if epoch % args.epochs_save == 0:
                fname = args.model_save_path + args.fname + 'lstm_' + str(epoch) + '.dat'
                torch.save(rnn.state_dict(), fname)
                fname = args.model_save_path + args.fname + 'output_' + str(epoch) + '.dat'
                torch.save(output.state_dict(), fname)
        epoch += 1
    np.save(args.timing_save_path + args.fname, time_all)


########### for graph completion task
def train_graph_completion(args, dataset_test, rnn, output):
    fname = args.model_save_path + args.fname + 'lstm_' + str(args.load_epoch) + '.dat'
    rnn.load_state_dict(torch.load(fname))
    fname = args.model_save_path + args.fname + 'output_' + str(args.load_epoch) + '.dat'
    output.load_state_dict(torch.load(fname))

    epoch = args.load_epoch
    print('model loaded!, epoch: {}'.format(args.load_epoch))

    for sample_time in range(1, 4):
        if 'GraphRNN_MLP' in args.note:
            G_pred = test_mlp_partial_simple_epoch(epoch, args, rnn, output, dataset_test, sample_time=sample_time)
        if 'GraphRNN_VAE' in args.note:
            G_pred = test_vae_partial_epoch(epoch, args, rnn, output, dataset_test, sample_time=sample_time)
        # save graphs
        fname = args.graph_save_path + args.fname_pred + str(epoch) + '_' + str(sample_time) + 'graph_completion.dat'
        save_graph_list(G_pred, fname)
    print('graph completion done, graphs saved')


########### for NLL evaluation
def train_nll(args, dataset_train, dataset_test, rnn, output, graph_validate_len, graph_test_len, max_iter=1000):
    fname = args.model_save_path + args.fname + 'lstm_' + str(args.load_epoch) + '.dat'
    rnn.load_state_dict(torch.load(fname))
    fname = args.model_save_path + args.fname + 'output_' + str(args.load_epoch) + '.dat'
    output.load_state_dict(torch.load(fname))

    epoch = args.load_epoch
    print('model loaded!, epoch: {}'.format(args.load_epoch))
    fname_output = args.nll_save_path + args.note + '_' + args.graph_type + '.csv'
    with open(fname_output, 'w+') as f:
        f.write(str(graph_validate_len) + ',' + str(graph_test_len) + '\n')
        f.write('train,test\n')
        for iter in range(max_iter):
            if 'GraphRNN_MLP' in args.note:
                nll_train = train_mlp_forward_epoch(epoch, args, rnn, output, dataset_train)
                nll_test = train_mlp_forward_epoch(epoch, args, rnn, output, dataset_test)
            if 'GraphRNN_RNN' in args.note:
                nll_train = train_rnn_forward_epoch(epoch, args, rnn, output, dataset_train)
                nll_test = train_rnn_forward_epoch(epoch, args, rnn, output, dataset_test)
            print('train', nll_train, 'test', nll_test)
            f.write(str(nll_train) + ',' + str(nll_test) + '\n')

    print('NLL evaluation done')