fin = open(args.dict, 'r')
for line in fin:
tokens = line.split()
dic[int(tokens[1])] = tokens[0]
word_dic = None
if args.word_dict is not None:
fin = open(args.word_dict, 'r')
word_dic = {}
for line in fin:
tokens = line.split()
word_dic[''.join(tokens[1:])] = tokens[0]
device = torch.device('cuda')
mdic = torch.load(args.model_dic)
model = Seq2Seq(**mdic['params']).to(device)
model.load_state_dict(mdic['state'])
model.eval()
reader = ScpStreamReader(args.data_scp,
mean_sub=args.mean_sub,
downsample=args.downsample)
reader.initialize()
space, beam_size, max_len = args.space, args.beam_size, args.max_len
start, win, stable_time = args.start_block, args.incl_block, args.stable_time
head, padding = args.attn_head, args.attn_padding
since = time.time()
fctm = open(args.output, 'w')
total_latency = 0
if args.word_dict is not None:
fin = open(args.word_dict, 'r')
word_dic = {}
for line in fin:
tokens = line.split()
word_dic[''.join(tokens[1:])] = tokens[0]
use_gpu = torch.cuda.is_available()
device = torch.device('cuda' if use_gpu else 'cpu')
models = []
for dic_path in args.model_dic:
sdic = torch.load(dic_path[0])
m_params = sdic['params']
if sdic['type'] == 'lstm':
model = Seq2Seq(**m_params).to(device)
elif sdic['type'] == 'tf':
model = Transformer(**m_params).to(device)
model.load_state_dict(sdic['state'])
model.eval()
models.append(model)
model = Ensemble(models)
reader = ScpStreamReader(args.data_scp,
mean_sub=args.mean_sub,
downsample=args.downsample)
reader.initialize()
since = time.time()
batch_size = args.batch_size
fout = open(args.output, 'w')
if __name__ == '__main__':
args = parser.parse_args()
print(args)
use_gpu = torch.cuda.is_available()
device = torch.device('cuda' if use_gpu else 'cpu')
model = Seq2Seq(input_size=args.d_input,
hidden_size=args.d_model,
output_size=args.n_classes,
n_enc=args.n_enc,
n_dec=args.n_dec,
n_head=args.n_head,
unidirect=args.unidirect,
incl_win=args.incl_win,
time_ds=args.time_ds,
use_cnn=args.use_cnn,
freq_kn=args.freq_kn,
freq_std=args.freq_std,
lm=not args.no_lm,
shared_emb=args.shared_emb,
dropout=args.dropout,
emb_drop=args.emb_drop).to(device)
ScpReader = ScpBatchReader if args.batch else ScpStreamReader
tr_reader = ScpReader(args.train_scp,
args.train_target,
downsample=args.downsample,
sort_src=True,
max_len=args.max_len,
max_utt=args.max_utt,
'input_size': args.d_input,
'hidden_size': args.d_model,
'output_size': args.n_classes,
'n_enc': args.n_enc,
'n_dec': args.n_dec,
'n_head': args.n_head,
'unidirect': args.unidirect,
'incl_win': args.incl_win,
'time_ds': args.time_ds,
'use_cnn': args.use_cnn,
'freq_kn': args.freq_kn,
'freq_std': args.freq_std,
'lm': not args.no_lm,
'shared_emb': args.shared_emb
}
model = Seq2Seq(**m_params)
ext = copy.deepcopy(model)
states = args.states.split(',')
state = "%s/epoch-%s.pt" % (args.model_path, states[0])
model.load_state_dict(torch.load(state, map_location='cpu'))
params = list(model.parameters())
for s in states[1:]:
state = "%s/epoch-%s.pt" % (args.model_path, s)
ext.load_state_dict(torch.load(state, map_location='cpu'))
eparams = list(ext.parameters())
for i in range(len(params)):
params[i].data.add_(eparams[i].data)
scale = 1.0 / len(states)
for p in params: