def _test(run):
U.set_random_seeds(11785)
lang = Lang()
trans = U.tokenizeTranscripts('train')
lang.init_lang(trans)
output_size = lang.num_items
batch_size = 1
print("Starting .. ..")
num_layers = 3
hidden_size = 256
input_size = 40
key_size = 128
value_size = 128
bidirectional = True
p = 3
embedding_size = 128
max_len = 496
encoder = EncoderRNN(input_size, hidden_size, key_size, value_size,
num_layers, bidirectional, p)
decoder = DecoderRNN(output_size, embedding_size, hidden_size, key_size,
value_size, num_layers, max_len)
teacher_forcing_ratio = 1.0
las = LAS(encoder, decoder, teacher_forcing_ratio)
model = torch.load('../saved_models/' + run + '-model.pt',
map_location=lambda storage, loc: storage)
las.load_state_dict(model.state_dict())
if U.use_cuda():
las = las.cuda()
# Prediction
test_dataset = SpeechDataset(lang, 'test')
test_dataloader = SpeechDataLoader(test_dataset,
batch_size=batch_size,
shuffle=False)
criterion = CrossEntropyLoss3D(reduce=False, ignore_index=C.PAD_TOKEN_IDX)
predictor = Predictor(las, lang, criterion)
outFile = open('../predictions-' + run + '.txt', 'w')
predictor.predict(test_dataloader, outFile)
def __init__(self, lang, mode):
self.mode = mode
features_path = (C.DATA_PATH + "%s.npy") % (self.mode)
self.feats = np.load(features_path)
self.trans = []
if self.mode is not 'test':
trans_tokenized = U.tokenizeTranscripts(self.mode)
for trans_items in trans_tokenized:
trans_indices = lang.items2indices(trans_items)
self.trans.append(trans_indices)
self.trans = np.asarray(self.trans)
assert (self.feats.shape[0] == self.trans.shape[0])
self.len = self.feats.shape[0]
print("%s Dataset Loading Completed" % (self.mode))
from data.language import Lang
from data.dataloader import SpeechDataLoader
from data.dataset import SpeechDataset
from models.encoder import EncoderRNN
from models.decoder import DecoderRNN
from models.las import LAS
from evaluator.evaluator import Evaluator
import torch.nn as nn
if __name__ == "__main__":
batch_size = 32
lang = Lang()
trans = U.tokenizeTranscripts('train')
lang.init_lang(trans)
output_size = lang.num_items
dev_dataset = SpeechDataset(lang, 'dev')
dev_dataloader = SpeechDataLoader(dev_dataset, batch_size=batch_size)
num_layers = 3
hidden_size = 256
input_size = 40
key_size = 128
value_size = 128
bidirectional = True
p = 3
assert (self.feats.shape[0] == self.trans.shape[0])
self.len = self.feats.shape[0]
print("%s Dataset Loading Completed" % (self.mode))
def __getitem__(self, index):
if self.mode is not 'test':
trans = np.append([C.SOS_TOKEN_IDX], self.trans[index])
trans = np.append(trans, [C.EOS_TOKEN_IDX])
else:
trans = np.asarray([C.SOS_TOKEN_IDX])
return (self.feats[index], trans)
def __len__(self):
return self.len
if __name__ == "__main__":
lang = Lang()
trans = U.tokenizeTranscripts('dev')
lang.init_lang(trans)
dataset = SpeechDataset(lang, 'dev')
assert (len(dataset) == 1139)
print(dataset[0])
print(''.join(lang.indices2items(dataset[0][1])))