def main(argv):
data = np.loadtxt(FLAGS.input, dtype=str, delimiter=",")
# calculate the number of loops to run the test
num = len(data[0])
batch_size = FLAGS.batch_size
num_loops = num / batch_size
assert num % batch_size == 0
with tf.device("/gpu:0"):
tf.set_random_seed(1234)
tfconf = tf.ConfigProto(allow_soft_placement=True)
with tf.Session(config=tfconf) as sess:
params = model_registry.GetParams(
'asr.librispeech.Librispeech960Wpm', 'Test')
params.cluster.worker.gpus_per_replica = 1
cluster = cluster_factory.Cluster(params.cluster)
with cluster, tf.device(cluster.GetPlacer()):
params.vn.global_vn = False
params.random_seed = 1234
params.is_eval = True
model = params.cls(params)
task = model.GetTask()
saver = tf.train.Saver()
saver.restore(sess, FLAGS.checkpoint)
# define the placeholders
input_tf = tf.placeholder(tf.float32, shape=[batch_size, None])
tgt_tf = tf.placeholder(tf.string)
sample_rate_tf = tf.placeholder(tf.int32)
mask_tf = tf.placeholder(tf.float32,
shape=[batch_size, None, 80])
# generate the features and inputs
features = create_features(input_tf, sample_rate_tf, mask_tf)
shape = tf.shape(features)
inputs = create_inputs(model, features, tgt_tf, batch_size,
mask_tf)
# loss
metrics = task.FPropDefaultTheta(inputs)
loss = tf.get_collection("per_loss")[0]
# prediction
decoded_outputs = task.Decode(inputs)
dec_metrics_dict = task.CreateDecoderMetrics()
correct = 0
for l in range(num_loops):
data_sub = data[:, l * batch_size:(l + 1) * batch_size]
audios_np, sample_rate, tgt_np, mask_freq = Read_input(
data_sub, batch_size)
feed_dict = {
input_tf: audios_np,
sample_rate_tf: sample_rate,
tgt_tf: tgt_np,
mask_tf: mask_freq
}
losses = sess.run(loss, feed_dict)
predictions = sess.run(decoded_outputs, feed_dict)
task.PostProcessDecodeOut(predictions, dec_metrics_dict)
wer_value = dec_metrics_dict['wer'].value * 100.
for i in range(batch_size):
print("pred:{}".format(predictions['topk_decoded'][i,
0]))
print("targ:{}".format(tgt_np[i].lower()))
print("true: {}".format(data_sub[1, i].lower()))
if predictions['topk_decoded'][i,
0] == tgt_np[i].lower():
correct += 1
print("------------------------------")
print("example {} succeeds".format(i))
print("Now, the WER is: {0:.2f}%".format(wer_value))
print("num of examples succeed: {}".format(correct))
print("success rate: {}%".format(correct / float(num) * 100))
def __init__(
self,
sess,
batch_size=1,
lr_stage1=100,
lr_stage2=0.1,
num_iter_stage1=1000,
num_iter_stage2=4000,
th=None,
psd_max_ori=None,
):
self.sess = sess
self.num_iter_stage1 = num_iter_stage1
self.num_iter_stage2 = num_iter_stage2
self.batch_size = batch_size
self.lr_stage1 = lr_stage1
self.lr_stage2 = lr_stage2
tf.set_random_seed(1234)
params = model_registry.GetParams("asr.librispeech.Librispeech960Wpm",
"Test")
params.random_seed = 1234
params.is_eval = True
params.cluster.worker.gpus_per_replica = 1
cluster = cluster_factory.Cluster(params.cluster)
with cluster, tf.device(cluster.GetPlacer()):
model = params.cls(params)
self.delta_large = tf.Variable(
np.zeros((batch_size, FLAGS.max_length_dataset),
dtype=np.float32),
name="qq_delta",
)
# placeholders
self.input_tf = tf.placeholder(tf.float32,
shape=[batch_size, None],
name="qq_input")
self.tgt_tf = tf.placeholder(tf.string)
self.rir = tf.placeholder(tf.float32)
self.sample_rate_tf = tf.placeholder(tf.int32,
name="qq_sample_rate")
self.mask = tf.placeholder(dtype=np.float32,
shape=[batch_size, None],
name="qq_mask")
self.mask_freq = tf.placeholder(dtype=np.float32,
shape=[batch_size, None, 80])
self.noise = tf.placeholder(np.float32,
shape=[batch_size, None],
name="qq_noise")
self.maxlen = tf.placeholder(np.int32)
self.lr = tf.placeholder(np.float32)
self.lengths = tf.placeholder(
np.int32,
shape=[
batch_size,
],
)
# variable
self.rescale = tf.Variable(
np.ones(
(batch_size, 1), dtype=np.float32) * FLAGS.initial_bound,
name="qq_rescale",
)
# extract the delta
self.delta = tf.slice(tf.identity(self.delta_large), [0, 0],
[batch_size, self.maxlen])
self.apply_delta = tf.clip_by_value(self.delta, -self.rescale,
self.rescale)
self.before_rir = tf.clip_by_value(
self.apply_delta * self.mask + self.input_tf, -(2**15),
2**15 - 1)
self.new_input = (create_speech_rir(
self.before_rir,
self.rir,
self.lengths,
self.maxlen,
self.batch_size,
) * self.mask)
self.pass_in = tf.clip_by_value(self.new_input + self.noise,
-(2**15), 2**15 - 1)
# generate the inputs that are needed for the lingvo model
self.features = create_features(self.pass_in, self.sample_rate_tf,
self.mask_freq)
self.inputs = create_inputs(model, self.features, self.tgt_tf,
self.batch_size, self.mask_freq)
task = model.GetTask()
metrics = task.FPropDefaultTheta(self.inputs)
# self.celoss with the shape (batch_size)
self.celoss = tf.get_collection("per_loss")[0]
self.decoded = task.Decode(self.inputs)
self.optimizer1 = tf.train.AdamOptimizer(self.lr)
grad1, var1 = self.optimizer1.compute_gradients(
self.celoss, [self.delta_large])[0]
self.train1 = self.optimizer1.apply_gradients([(tf.sign(grad1), var1)])
def main(argv):
data = np.loadtxt(FLAGS.input, dtype=str, delimiter=",")
# calculate the number of loops to run the test
num = len(data[0])
batch_size = FLAGS.batch_size
num_loops = num / batch_size
assert num % batch_size == 0
with tf.device("/gpu:0"):
tf.set_random_seed(1234)
tfconf = tf.ConfigProto(allow_soft_placement=True)
with tf.Session(config=tfconf) as sess:
params = model_registry.GetParams(
"asr.librispeech.Librispeech960Wpm", "Test")
params.cluster.worker.gpus_per_replica = 1
cluster = cluster_factory.Cluster(params.cluster)
with cluster, tf.device(cluster.GetPlacer()):
params.vn.global_vn = False
params.random_seed = 1234
params.is_eval = True
model = params.cls(params)
task = model.GetTask()
saver = tf.train.Saver()
saver.restore(sess, FLAGS.checkpoint)
# define the placeholders
input_tf = tf.placeholder(tf.float32, shape=[batch_size, None])
tgt_tf = tf.placeholder(tf.string)
sample_rate_tf = tf.placeholder(tf.int32)
mask_tf = tf.placeholder(tf.float32,
shape=[batch_size, None, 80])
rir_tf = tf.placeholder(tf.float32)
lengths = tf.placeholder(
np.int32,
shape=[
batch_size,
],
)
maxlen = tf.placeholder(np.int32)
mask = tf.placeholder(dtype=np.float32,
shape=[batch_size, None])
# generate the features and inputs
new_input = (create_speech_rir(input_tf, rir_tf, lengths,
maxlen, batch_size) * mask)
features = create_features(new_input, sample_rate_tf, mask_tf)
shape = tf.shape(features)
inputs = create_inputs(model, features, tgt_tf, batch_size,
mask_tf)
# loss
metrics = task.FPropDefaultTheta(inputs)
loss = tf.get_collection("per_loss")[0]
# prediction
decoded_outputs = task.Decode(inputs)
dec_metrics_dict = task.CreateDecoderMetrics()
success_rates = []
for num_room in range(FLAGS.num_test_rooms):
correct = 0
rir = Readrir(num_room)
for l in range(num_loops):
data_sub = data[:, l * batch_size:(l + 1) * batch_size]
(
audios_np,
sample_rate,
tgt_np,
mask_freq,
lengths_np,
max_len,
masks,
) = Read_input(data_sub, batch_size)
feed_dict = {
input_tf: audios_np,
sample_rate_tf: sample_rate,
tgt_tf: tgt_np,
mask_tf: mask_freq,
rir_tf: rir,
lengths: lengths_np,
maxlen: max_len,
mask: masks,
}
losses = sess.run(loss, feed_dict)
predictions = sess.run(decoded_outputs, feed_dict)
task.PostProcessDecodeOut(predictions,
dec_metrics_dict)
wer_value = dec_metrics_dict["wer"].value * 100.0
for i in range(batch_size):
print("example: {}, loss_ce: {}".format(
l * batch_size + i, losses[i]))
print("pred:{}".format(
predictions["topk_decoded"][i, 0]))
print("targ:{}".format(tgt_np[i].lower()))
print("true: {}".format(data_sub[1, i].lower()))
if predictions["topk_decoded"][
i, 0] == tgt_np[i].lower():
correct += 1
print("--------------------------------")
print("Now, the WER is: {0:.2f}%".format(wer_value))
print("num of examples succeed for room {}: {}".format(
num_room, correct))
success_rate = correct / float(num) * 100
print("success rate for room {}: {}%".format(
num_room, success_rate))
success_rates.append(success_rate)
success_ave = float(sum(success_rates)) / len(success_rates)
print("success rate overall: {}%".format(success_ave))
def __init__(self, sess, batch_size=1,
lr_stage1=100, lr_stage2=0.1, num_iter_stage1=1000, num_iter_stage2=4000, th=None,
psd_max_ori=None):
self.sess = sess
self.num_iter_stage1 = num_iter_stage1
self.num_iter_stage2 = num_iter_stage2
self.batch_size = batch_size
self.lr_stage1 = lr_stage1
tf.set_random_seed(1234)
params = model_registry.GetParams('asr.librispeech.Librispeech960Wpm', 'Test')
params.random_seed = 1234
params.is_eval = True
params.cluster.worker.gpus_per_replica = 1
cluster = cluster_factory.Cluster(params.cluster)
with cluster, tf.device(cluster.GetPlacer()):
model = params.cls(params)
self.delta_large = tf.Variable(np.zeros((batch_size, FLAGS.max_length_dataset), dtype=np.float32), name='qq_delta')
# placeholders
self.input_tf = tf.placeholder(tf.float32, shape=[batch_size, None], name='qq_input')
self.tgt_tf = tf.placeholder(tf.string)
self.sample_rate_tf = tf.placeholder(tf.int32, name='qq_sample_rate')
self.th = tf.placeholder(tf.float32, shape=[batch_size, None, None], name='qq_th')
self.psd_max_ori = tf.placeholder(tf.float32, shape=[batch_size], name='qq_psd')
self.mask = tf.placeholder(dtype=np.float32, shape=[batch_size, None], name='qq_mask')
self.mask_freq = tf.placeholder(dtype=np.float32, shape=[batch_size, None, 80])
self.noise = tf.placeholder(np.float32, shape=[batch_size, None], name="qq_noise")
self.maxlen = tf.placeholder(np.int32)
self.lr_stage2 = tf.placeholder(np.float32)
# variable
self.rescale = tf.Variable(np.ones((batch_size,1), dtype=np.float32), name='qq_rescale')
self.alpha = tf.Variable(np.ones((batch_size), dtype=np.float32) * 0.05, name='qq_alpha')
# extract the delta
self.delta = tf.slice(tf.identity(self.delta_large), [0, 0], [batch_size, self.maxlen])
self.apply_delta = tf.clip_by_value(self.delta, -FLAGS.initial_bound, FLAGS.initial_bound) * self.rescale
self.new_input = self.apply_delta * self.mask + self.input_tf
self.pass_in = tf.clip_by_value(self.new_input + self.noise, -2**15, 2**15-1)
# generate the inputs that are needed for the lingvo model
self.features = create_features(self.pass_in, self.sample_rate_tf, self.mask_freq)
self.inputs = create_inputs(model, self.features, self.tgt_tf, self.batch_size, self.mask_freq)
task = model.GetTask()
metrics = task.FPropDefaultTheta(self.inputs)
# self.celoss with the shape (batch_size)
self.celoss = tf.get_collection("per_loss")[0]
self.decoded = task.Decode(self.inputs)
# compute the loss for masking threshold
self.loss_th_list = []
self.transform = Transform(FLAGS.window_size)
for i in range(self.batch_size):
logits_delta = self.transform((self.apply_delta[i, :]), (self.psd_max_ori)[i])
loss_th = tf.reduce_mean(tf.nn.relu(logits_delta - (self.th)[i]))
loss_th = tf.expand_dims(loss_th, dim=0)
self.loss_th_list.append(loss_th)
self.loss_th = tf.concat(self.loss_th_list, axis=0)
self.optimizer1 = tf.train.AdamOptimizer(self.lr_stage1)
self.optimizer2 = tf.train.AdamOptimizer(self.lr_stage2)
grad1, var1 = self.optimizer1.compute_gradients(self.celoss, [self.delta_large])[0]
grad21, var21 = self.optimizer2.compute_gradients(self.celoss, [self.delta_large])[0]
grad22, var22 = self.optimizer2.compute_gradients(self.alpha * self.loss_th, [self.delta_large])[0]
self.train1 = self.optimizer1.apply_gradients([(tf.sign(grad1), var1)])
self.train21 = self.optimizer2.apply_gradients([(grad21, var21)])
self.train22 = self.optimizer2.apply_gradients([(grad22, var22)])
self.train2 = tf.group(self.train21, self.train22)
def main():
checkpoint = "./model/ckpt-00908156"
parser = argparse.ArgumentParser(description=None)
parser.add_argument('--dirs',
type=str,
nargs='+',
required=True,
help='Filepath of original input audio')
args = parser.parse_args()
while len(sys.argv) > 1:
sys.argv.pop()
with tf.device("/gpu:0"):
tf.set_random_seed(1234)
tfconf = tf.ConfigProto(allow_soft_placement=True)
with tf.Session(config=tfconf) as sess:
params = model_registry.GetParams(
'asr.librispeech.Librispeech960Wpm', 'Test')
params.cluster.worker.gpus_per_replica = 1
cluster = cluster_factory.Cluster(params.cluster)
with cluster, tf.device(cluster.GetPlacer()):
params.vn.global_vn = False
params.random_seed = 1234
params.is_eval = True
model = params.cls(params)
task = model.GetTask()
saver = tf.train.Saver()
saver.restore(sess, checkpoint)
input_tf = tf.placeholder(tf.float32, shape=[1, None])
tgt_tf = tf.placeholder(tf.string)
sample_rate_tf = tf.placeholder(tf.int32)
mask_tf = tf.placeholder(tf.float32, shape=[1, None, 80])
features = create_features(input_tf, sample_rate_tf, mask_tf)
shape = tf.shape(features)
inputs = create_inputs(model, features, tgt_tf, 1, mask_tf)
metrics = task.FPropDefaultTheta(inputs)
loss = tf.get_collection("per_loss")[0]
# prediction
decoded_outputs = task.Decode(inputs)
dec_metrics_dict = task.CreateDecoderMetrics()
for audio_dir in args.dirs:
file_names = _get_file_names(audio_dir)
transcriptions = {}
for fidx, file_name in enumerate(file_names):
audios_np, sample_rate, tgt_np, mask_freq = _decode_audio(
audio_dir, file_name)
feed_dict = {
input_tf: audios_np,
sample_rate_tf: sample_rate,
tgt_tf: tgt_np,
mask_tf: mask_freq
}
try:
losses = sess.run(loss, feed_dict)
predictions = sess.run(decoded_outputs, feed_dict)
except:
print("Error in transcribing: ", file_name)
continue
task.PostProcessDecodeOut(predictions,
dec_metrics_dict)
wer_value = dec_metrics_dict['wer'].value * 100.
transcriptions[file_name] = predictions[
'topk_decoded'][0, 0].lower()
print(
fidx, "pred-{},{} : {}".format(
audio_dir, file_name,
predictions['topk_decoded'][0, 0]))
with open(join(audio_dir, "transcriptions.json"),
'w') as f:
f.write(json.dumps(transcriptions))