def __init__(self,
rgb_size=RGB_FEATURES_SIZE,
audio_size=AUDIO_FEATURES_SIZE,
output_size=YOUTUBE8M_LABELS_N,
inner_sizes=(2048, 2048)):
super().__init__()
self._impl = FCN(rgb_size + audio_size,
output_size,
inner_sizes,
dropout=0.2,
out_activation=nn.Sigmoid)
def __init__(self,
rgb_size=RGB_FEATURES_SIZE,
audio_size=AUDIO_FEATURES_SIZE,
output_size=YOUTUBE8M_LABELS_N,
inner_size=2048,
layers_number=2):
super().__init__()
self._layers = nn.ModuleList([
nn.GRU(input_size=rgb_size + audio_size if i == 0 else inner_size,
hidden_size=inner_size,
num_layers=1,
batch_first=True,
bidirectional=False) for i in range(layers_number)
])
self._out = FCN(inner_size * layers_number, output_size, (1524, 1524))
self._num_layers = layers_number
self._inner_size = inner_size
def __init__(self,
rgb_size=RGB_FEATURES_SIZE,
audio_size=AUDIO_FEATURES_SIZE,
output_size=YOUTUBE8M_LABELS_N,
layers_number=3,
hidden_size=1024):
super().__init__()
self._linear_layers = nn.ModuleList([
nn.Linear(rgb_size + audio_size + hidden_size * i, hidden_size)
for i in range(layers_number)
])
self._attention_layers = nn.ModuleList(
[nn.Linear(hidden_size, 1) for i in range(layers_number)])
self._bn = nn.BatchNorm1d(rgb_size + audio_size +
hidden_size * layers_number)
self._out = FCN(rgb_size + audio_size + hidden_size * layers_number,
output_size, (4096, 4096))
self._num_layers = layers_number
def __init__(self,
rgb_size=RGB_FEATURES_SIZE,
audio_size=AUDIO_FEATURES_SIZE,
output_size=YOUTUBE8M_LABELS_N,
layers_number=4,
hidden_size=768):
super().__init__()
self._linear_layers = nn.ModuleList([
nn.Linear(rgb_size + audio_size + hidden_size * i, hidden_size)
for i in range(layers_number)
])
self._attention_layers = nn.ModuleList(
[nn.Linear(hidden_size, 1) for i in range(layers_number)])
self._dropout_layers = nn.ModuleList(
[nn.Dropout((i + 1) / 10.) for i in range(layers_number)])
self._out = FCN(rgb_size + audio_size + hidden_size * layers_number,
output_size, (4096, 4096),
dropout=0.3)
self._num_layers = layers_number
def __init__(self,
rgb_size=RGB_FEATURES_SIZE,
audio_size=AUDIO_FEATURES_SIZE,
output_size=YOUTUBE8M_LABELS_N,
rgb_inner_size=1024,
audio_inner_size=128,
layers_number=1):
super().__init__()
self._rgb_layers = nn.ModuleList([
nn.LSTM(input_size=rgb_size + 2 * rgb_inner_size * i,
hidden_size=rgb_inner_size,
num_layers=1,
batch_first=True,
bidirectional=True) for i in range(layers_number)
])
self._audio_layers = nn.ModuleList([
nn.LSTM(input_size=audio_size + 2 * audio_inner_size * i,
hidden_size=audio_inner_size,
num_layers=1,
batch_first=True,
bidirectional=True) for i in range(layers_number)
])
self._first_linear_rgb = nn.Linear(rgb_size, rgb_size)
self._first_linear_audio = nn.Linear(audio_size, audio_size)
self._rgb_attention = nn.Linear(
rgb_size + rgb_inner_size * 2 * layers_number, 1)
self._audio_attention = nn.Linear(
audio_size + audio_inner_size * 2 * layers_number, 1)
self._bn = nn.BatchNorm1d(rgb_size +
rgb_inner_size * 2 * layers_number +
audio_size +
audio_inner_size * 2 * layers_number)
self._out = FCN(
rgb_size + rgb_inner_size * 2 * layers_number + audio_size +
audio_inner_size * 2 * layers_number, output_size, (4096, 4096))
self._num_layers = layers_number