config = args.parse_args()
N_CLASSES = 11
# 1. Loading a saved model
x = tf.keras.layers.Input(shape=(257, None, 4))
model = getattr(model, config.model)(
weights=None,
input_tensor=x,
classes=N_CLASSES,
backend=tf.keras.backend,
layers=tf.keras.layers,
models=tf.keras.models,
utils=tf.keras.utils,
)
model.load_weights(config.name)
if config.verbose:
model.summary()
PATH = '/media/data1/datasets/ai_challenge/icassp/'
# 2. loading evaluation dataset
if config.dataset == 'challenge':
eval_x = np.load(os.path.join(PATH, 'final_x.npy'))
eval_y = np.load(os.path.join(PATH, 'final_y.npy'))
else:
eval_x = np.concatenate([
np.load(os.path.join(PATH, 'test_x.npy')),
np.load(os.path.join(PATH, 'noise_test_x.npy'))
],
out = tf.keras.layers.Flatten()(out)
out = tf.keras.layers.ReLU()(out)
else:
for i in range(config.n_layers):
# out = tf.keras.layers.Dropout(0.1)(out)
out = tf.keras.layers.Dense(config.n_dim)(out)
out = tf.keras.layers.Activation('sigmoid')(out) * out
out = tf.keras.layers.Dense(config.n_classes, activation='relu')(out)
model = tf.keras.models.Model(inputs=model.input, outputs=out)
specs = None
for name in config.name.split(','):
NAME = name if name.endswith('.h5') else name + '.h5'
model.load_weights(NAME)
if specs is None:
specs = model.predict(wavs, verbose=True)
else:
specs += model.predict(wavs, verbose=True)
specs /= len(config.name.split(','))
specs = specs / config.multiplier
specs = tf.reshape(specs, [*specs.shape[:2], 3, 10])
angles = tf.cast(tf.round(tf.reduce_sum(specs, axis=(1, 2))), tf.int32)
classes = tf.cast(tf.round(tf.reduce_sum(specs, axis=(1, 3))), tf.int32)
d_dir = D_direction(tf.cast(gt_angle, tf.float32),
tf.cast(angles, tf.float32))
def main(config):
wavs = load_wav(config)
import tensorflow as tf
import efficientnet.model as model
x = tf.keras.layers.Input(shape=(config.n_mels, None, 2))
model = getattr(model, config.model)(
include_top=False,
weights=None,
input_tensor=x,
backend=tf.keras.backend,
layers=tf.keras.layers,
models=tf.keras.models,
utils=tf.keras.utils,
)
out = tf.transpose(model.output, perm=[0, 2, 1, 3])
out = tf.keras.layers.Reshape([-1, out.shape[-1] * out.shape[-2]])(out)
if config.n_layers > 0:
if config.mode == 'GRU':
out = tf.keras.layers.Dense(config.n_dim)(out)
for i in range(config.n_layers):
# out = transformer_layer(config.n_dim, config.n_heads)(out)
out = tf.keras.layers.Bidirectional(
tf.keras.layers.GRU(config.n_dim, return_sequences=True),
backward_layer=tf.keras.layers.GRU(config.n_dim,
return_sequences=True,
go_backwards=True))(out)
elif config.mode == 'transformer':
out = tf.keras.layers.Dense(config.n_dim)(out)
out = encoder(config.n_layers, config.n_dim, config.n_heads)(out)
out = tf.keras.layers.Flatten()(out)
out = tf.keras.layers.ReLU()(out)
out = tf.keras.layers.Dense(config.n_classes, activation='relu')(out)
model = tf.keras.models.Model(inputs=model.input, outputs=out)
model.load_weights('model.h5')
wavs = model.predict(wavs)
wavs = wavs / config.multiplier
wavs = tf.reshape(wavs, [*wavs.shape[:2], 3, 10])
angles = tf.cast(tf.round(tf.reduce_sum(wavs, axis=(1, 2))), tf.int8)
classes = tf.cast(tf.round(tf.reduce_sum(wavs, axis=(1, 3))), tf.int8)
# import numpy as np
# from glob import glob
# num = len(sorted(glob('dataset/3rd_track3/*.wav')))
# angles = tf.convert_to_tensor(np.zeros((num, 10), dtype=np.int))
# classes = tf.convert_to_tensor(np.ones((num, 3), dtype=np.int) * 2)
data = {'track3_results': list()}
for idx, (ag, cl) in enumerate(zip(angles, classes)):
_data = {
'id': idx,
'angle': ag.numpy().tolist(),
'class': cl.numpy().tolist()
}
data['track3_results'].append(_data)
tojson(data)