def train(config):
train_set, dev_set, test_set = mod.SpeechDataset.splits(config)
if config["input_file"]:
model = mod.SpeechModel(config)
model.load(config["input_file"])
else:
model = mod.SpeechModel(config)
if not config["no_cuda"]:
torch.cuda.set_device(config["gpu_no"])
model.cuda()
optimizer = torch.optim.SGD(model.parameters(), lr=config["lr"])
criterion = nn.CrossEntropyLoss()
min_loss = sys.float_info.max
train_loader = data.DataLoader(train_set, batch_size=config["batch_size"], shuffle=True, drop_last=True)
dev_loader = data.DataLoader(dev_set, batch_size=min(len(dev_set), 100), shuffle=True)
test_loader = data.DataLoader(test_set, batch_size=min(len(test_set), 100), shuffle=True)
step_no = 0
for epoch_idx in range(config["n_epochs"]):
for batch_idx, (model_in, labels) in enumerate(train_loader):
model.train()
optimizer.zero_grad()
if not config["no_cuda"]:
model_in = model_in.cuda()
labels = labels.cuda()
model_in = Variable(model_in, requires_grad=False)
scores = model(model_in)
labels = Variable(labels, requires_grad=False)
loss = criterion(scores, labels)
loss.backward()
optimizer.step()
step_no += 1
print_eval("train step #{}".format(step_no), scores, labels, loss)
if epoch_idx % config["dev_every"] == config["dev_every"] - 1:
model.eval()
for model_in, labels in dev_loader:
model_in = Variable(model_in, requires_grad=False)
if not config["no_cuda"]:
model_in = model_in.cuda()
labels = labels.cuda()
scores = model(model_in)
labels = Variable(labels, requires_grad=False)
loss = criterion(scores, labels)
loss_numeric = loss.cpu().data.numpy()[0]
if loss_numeric < min_loss:
min_loss = loss_numeric
model.save(config["output_file"])
print_eval("dev", scores, labels, loss)
evaluate(config, model, test_loader)
def evaluate(config, model=None, test_loader=None):
if not test_loader:
_, _, test_set = mod.SpeechDataset.splits(config)
test_loader = data.DataLoader(test_set, batch_size=len(test_set))
if not config["no_cuda"]:
torch.cuda.set_device(config["gpu_no"])
if not model:
model = mod.SpeechModel(config)
model.load(config["input_file"])
if not config["no_cuda"]:
torch.cuda.set_device(config["gpu_no"])
model.cuda()
model.eval()
criterion = nn.CrossEntropyLoss()
results = []
total = 0
for model_in, labels in test_loader:
model_in = Variable(model_in, requires_grad=False)
if not config["no_cuda"]:
model_in = model_in.cuda()
labels = labels.cuda()
scores = model(model_in)
labels = Variable(labels, requires_grad=False)
loss = criterion(scores, labels)
results.append(print_eval("test", scores, labels, loss) * model_in.size(0))
total += model_in.size(0)
print("final test accuracy: {}".format(sum(results) / total))
def __init__(self):
self.CHUNK = 1024 * 3
self.THRESHOLD = 0.1
self.FORMAT = pyaudio.paFloat32
self.CHANNELS = 1
self.SAMPLE_RATE = 16000
self.model = md.SpeechModel()
self.p = pyaudio.PyAudio()
import model
import librosa
import os
path_name = "/home/minhhiu/MyProjects/Compressed Speech Data/full_command_data/test/wav/"
cmd_list = []
m = model.SpeechModel()
correct = 0
total = 0
with open("./model/cmd_list.txt", "r", encoding="utf8") as cmd_r:
for i in range(71):
cmd = cmd_r.readline()
cmd = cmd.lower().strip("\n")
cmd_list.append(cmd)
print(cmd_list)
for d in os.listdir(path_name):
path = os.path.join(path_name, d)
for file in os.listdir(path):
wav_path = os.path.join(path, file)
txt_path = wav_path.replace("wav", "txt")
audio, _ = librosa.load(wav_path, sr=16000, mono=True)
audio = audio.reshape(-1, 1)
pred_index = m.predict(audio)
with open(txt_path, "r", encoding="utf8") as fr:
n_coms = np.arange(3, 6)
n_mixs = np.arange(3, 6)
window_sizes = np.arange(0.015, 0.045, 0.01)
strides = np.arange(0.005, 0.015, 0.005)
n_mfccs = np.array([6, 9, 13])
params = itertools.product(n_coms, n_mixs, window_sizes, strides, n_mfccs)
max_setting = {}
max_acc = -np.inf
for n_com, n_mix, window_size, stride, n_mfcc in params:
print('______')
print('n_com = {n_com}, n_mix = {n_mix}, window_size = {window_size}, stride = {stride}, \
n_mfcc = {n_mfcc}'.format(n_com=n_com, n_mix=n_mix, window_size=window_size, stride=stride, n_mfcc=n_mfcc))
data_loader = model.DataLoader(path='tiengviet', n_mfcc=n_mfcc, window_size=window_size,
overlap=stride, test_size=0.1 , shuffle = True)
data_loader.load_data()
m = model.SpeechModel(data_loader, n_com=n_com, n_mix=n_mix)
m.fit()
acc = m.cal_accuracy()
print(acc)
if acc > max_acc:
max_acc = acc
max_setting['n_com'] = n_com
max_setting['n_mix'] = n_mix
max_setting['window_size'] = window_size
max_setting['stride'] = stride
max_setting['n_mfcc'] = n_mfcc