parser.add_argument('--load',
type=str,
required=True,
help='where to load a model from')
args = parser.parse_args()
print(args)
init_seeds(args.seed, args.cuda)
print("loading model...")
lm = torch.load(args.load)
if args.cuda:
lm.cuda()
print(lm.model)
print("preparing data...")
def temp_splits_from_fn(fn):
tokens = tokens_from_file(fn, lm.vocab, randomize=False)
return TemporalSplits(tokens, lm.model.in_len, args.target_seq_len)
tss = filelist_to_objects(args.file_list, temp_splits_from_fn)
data = BatchBuilder(tss,
args.batch_size,
discard_h=not args.concat_articles)
if args.cuda:
data = CudaStream(data)
loss = evaluate(lm, data, use_ivecs=False)
print('loss {:5.2f} | ppl {:8.2f}'.format(loss, math.exp(loss)))
lm.cuda()
print(lm.model)
print("preparing data...")
def temp_splits_from_fn(fn):
tokens = tokens_from_file(fn, lm.vocab, randomize=False)
return TemporalSplits(tokens, lm.model.in_len, args.target_seq_len)
print("\ttraining...")
train_tss = filelist_to_objects(args.train_list, temp_splits_from_fn)
train_data = BatchBuilder(train_tss,
args.batch_size,
discard_h=not args.concat_articles)
if args.cuda:
train_data = CudaStream(train_data)
print("\tvalidation...")
valid_tss = filelist_to_objects(args.valid_list, temp_splits_from_fn)
valid_data = BatchBuilder(valid_tss,
args.batch_size,
discard_h=not args.concat_articles)
if args.cuda:
valid_data = CudaStream(valid_data)
print("training...")
lr = args.lr
best_val_loss = None
for epoch in range(1, args.epochs + 1):
if args.keep_shuffling: