batch_size=seqGANs.batch_size,
shuffle=False,
num_workers=2)
nll = oracle.NLL(dataloader=dataloader_gen,
start_input=seqGANs.start_input,
start_h=seqGANs.start_h)
print('nll: ' + str(nll))
seqGANs = seqGANs.cuda()
return nll
if __name__ == '__main__':
vis = visdom.Visdom(port=2424, env='seqGANs-syn')
seqGANs = SEQGANs_syn().cuda()
oracle = Oracle().cuda()
oracle.load_state_dict(torch.load('../datas/Synthetic_Data/oracle.pkl'))
start_epoch = 0
seqGANs.load_state_dict(
torch.load('../syn_save_pretrained/pretrained_290.pkl'))
'''
for j in range(300):
total_loss = seqGANs.pretraining()
vis.line(X=torch.tensor([j]), Y=torch.unsqueeze(torch.tensor(total_loss), 0), win='G_pre_loss',
opts=dict(legend=['G_pre_loss']), update='append' if j > 0 else None)
if j%10==0:
torch.save(seqGANs.state_dict(), '../syn_save_pretrained/pretrained_'+str(j)+'.pkl')
if j%5==0:
nll = output_nll(seqGANs=seqGANs, oracle=oracle)
vis.line(X=torch.tensor([j]), Y=torch.unsqueeze(torch.tensor(nll), 0), win='nll',
opts=dict(legend=['nll']), update='append' if j > 0 else None)
'''
