if args.anneal_lr and (len(all_val_ppls) > args.nonmono and
val_ppl > min(all_val_ppls[:-args.nonmono])
and lr > 1e-5):
optimizer.param_groups[0]['lr'] /= args.lr_factor
if epoch % args.visualize_every == 0:
visualize(model)
all_val_ppls.append(val_ppl)
with open(ckpt, 'rb') as f:
model = torch.load(f)
model = model.to(device)
val_ppl = evaluate(model, 'val')
else:
with open(ckpt, 'rb') as f:
model = torch.load(f)
model = model.to(device)
model.eval()
with torch.no_grad():
## get document completion perplexities
test_ppl = evaluate(model, 'test', tc=args.tc, td=args.td)
## get most used topics
indices = torch.tensor(range(args.num_docs_train))
indices = torch.split(indices, args.batch_size)
thetaAvg = torch.zeros(1, args.num_topics).to(device)
thetaWeightedAvg = torch.zeros(1, args.num_topics).to(device)
cnt = 0
for idx, ind in enumerate(indices):
data_batch = data.get_batch(train_tokens, train_counts, ind,
args.vocab_size, device)
sums = data_batch.sum(1).unsqueeze(1)
visualize(etm_model)
# save perplexities
all_val_ppls.append(val_ppl)
print("Training finished.")
print("## -------------------------------------")
print("##\t TESTING THE MODEL ")
print("## -------------------------------------")
# load trained model and evaluate it
with open(ckpt, 'rb') as f:
etm_model = torch.load(f)
etm_model = etm_model.to(device)
etm_model.eval()
with torch.no_grad():
## ---------------
## Idea : get document completion perplexities
test_ppl = evaluate(etm_model, test_corpus, num_docs_test)
## ----------------
## Idea : get most used topics
indices = torch.tensor(range(num_docs_test)) # training documents indices
indices = torch.split(indices, _training_batch_size)
# just initialising data structures
thetaAvg = torch.zeros(1, num_topics).to(device)
thetaWeightedAvg = torch.zeros(1, num_topics).to(device)
cnt = 0
for idx, ind in enumerate(indices):