def extract_features():
audio_txt = 'audio_files.txt'
model = SoundNet()
model.load_weights()
# Extract Feature
sound_samples, audio_paths = load_from_txt(audio_txt, config=local_config)
print(LEN_WAVEFORM / 6)
print(model)
features = {}
features['feats'] = []
features['paths'] = []
model.eval()
for idx, sound_sample in enumerate(sound_samples):
print(audio_paths[idx])
new_sample = torch.from_numpy(sound_sample)
output = model.forward(new_sample)
features['feats'].append(output)
features['paths'].append(audio_paths[idx])
return features
# Setup visible device
os.environ["CUDA_VISIBLE_DEVICES"] = args.cuda_device
# Load pre-trained model
G_name = './models/sound8.npy'
param_G = np.load(G_name, encoding='latin1').item()
if args.phase == 'demo':
# Demo
sound_samples = [
np.reshape(np.load('data/demo.npy', encoding='latin1'),
[1, -1, 1, 1])
]
else:
# Extract Feature
sound_samples = load_from_txt(args.audio_txt, config=local_config)
# Make path
if not os.path.exists(args.outpath):
os.mkdir(args.outpath)
# Init. Session
sess_config = tf.ConfigProto()
sess_config.allow_soft_placement = True
sess_config.gpu_options.allow_growth = True
with tf.Session(config=sess_config) as session:
# Build model
model = Model(session, config=local_config, param_G=param_G)
init = tf.global_variables_initializer()
session.run(init)
'sample_rate': 22050,
'load_size': 22050 * 20,
'name_scope': 'SoundNet_TF',
'phase': 'extract',
}
#load model and weights
from soundnet2 import SoundNet8_pytorch
model = SoundNet8_pytorch()
model.load_state_dict(torch.load("sound8.pth"))
#summary model
from torchsummaryX import summary
# summary(model, torch.zeros(1,1,22050 * 20,1))
#load data and extract feature
audio_txt = os.path.abspath(os.path.join(
BASE_DIR, 'mydata', "audio_files.txt")) #path of audio_files.txt
sound_samples, audio_paths = load_from_txt(audio_txt, config=local_config)
for idx, sound_sample in enumerate(sound_samples):
print(audio_paths[idx])
new_sample = torch.from_numpy(sound_sample)
output = model.forward(new_sample)
#classification
softmax = nn.Softmax(dim=1)
id_obj = torch.max(softmax(output[0]), 1)
id_scn = torch.max(softmax(output[1]), 1)
print('#####objects class: %s' % torch.squeeze(id_obj[1]))
print('#####places class: %s' % torch.squeeze(id_scn[1]))
#average poolling
avgpool_layer = nn.AvgPool2d((4, 1))
avgpool_obj = avgpool_layer(softmax(output[0]))
#tensor-->ndarry
a_feature = avgpool_obj.detach()
# Build model
model = Model(session, config=local_config, param_G=param_G)
init = tf.global_variables_initializer()
session.run(init)
model.load()
while position < n_files:
if args.phase == 'demo':
# Demo
sound_samples = [
np.reshape(np.load('data/demo.npy', encoding='latin1'),
[1, -1, 1, 1])
]
else:
# Extract Feature
sound_samples = load_from_txt(args.audio_txt,
position,
config=local_config,
batch_size=batch_size)
# Make path
if not os.path.exists(args.outpath):
os.mkdir(args.outpath)
for idx in range(len(sound_samples)):
sound_sample = sound_samples[idx]
output = extract_feat(model,
sound_sample,
args,
filename=all_files[position + idx])
position += batch_size