def __init__(self,
input_dim,
in_channel,
channels,
kernel_sizes,
dropout,
bottleneck_dim=0,
param_init=0.1):
super(GatedConvEncoder, self).__init__()
(channels, kernel_sizes, _,
_), _ = parse_cnn_config(channels, kernel_sizes, '', '')
self.in_channel = in_channel
assert input_dim % in_channel == 0
self.input_freq = input_dim // in_channel
self.bridge = None
assert len(channels) > 0
assert len(channels) == len(kernel_sizes)
layers = OrderedDict()
for lth in range(len(channels)):
layers['conv%d' % lth] = ConvGLUBlock(kernel_sizes[lth][0],
input_dim,
channels[lth],
weight_norm=True,
dropout=0.2)
input_dim = channels[lth]
# weight normalization + GLU for the last fully-connected layer
self.fc_glu = nn.utils.weight_norm(nn.Linear(input_dim, input_dim * 2),
name='weight',
dim=0)
self._odim = int(input_dim)
if bottleneck_dim > 0:
self.bridge = nn.Linear(self._odim, bottleneck_dim)
self._odim = bottleneck_dim
self.layers = nn.Sequential(layers)
self._factor = 1
self.reset_parameters(param_init)
def __init__(self, args, save_path=None):
super(LMBase, self).__init__()
logger.info(self.__class__.__name__)
self.lm_type = args.lm_type
self.save_path = save_path
self.emb_dim = args.emb_dim
self.n_units = args.n_units
self.n_layers = args.n_layers
self.lsm_prob = args.lsm_prob
self.vocab = args.vocab
self.eos = 2
self.pad = 3
# NOTE: reserved in advance
# for cache
self.cache_theta = 0.2 # smoothing parameter
self.cache_lambda = 0.2 # cache weight
self.cache_ids = []
self.cache_keys = []
self.cache_attn = []
self.embed_cache = None
self.embed = nn.Embedding(self.vocab,
args.emb_dim,
padding_idx=self.pad)
self.dropout_embed = nn.Dropout(p=args.dropout_in)
model_size = args.lm_type.replace('gated_conv_', '')
blocks = OrderedDict()
dropout = args.dropout_hidden
if model_size == 'custom':
blocks['conv1'] = ConvGLUBlock(args.kernel_size,
args.emb_dim,
args.n_units,
bottlececk_dim=args.n_projs,
dropout=dropout)
for lth in range(args.n_layers - 1):
blocks['conv%d' % (lth + 2)] = ConvGLUBlock(
args.kernel_size,
args.n_units,
args.n_units,
bottlececk_dim=args.n_projs,
dropout=dropout)
last_dim = args.n_units
elif model_size == '8':
blocks['conv1'] = ConvGLUBlock(4,
args.emb_dim,
900,
dropout=dropout)
for i in range(1, 8, 1):
blocks['conv2-%d' % i] = ConvGLUBlock(4,
900,
900,
dropout=dropout)
last_dim = 900
elif model_size == '8B':
blocks['conv1'] = ConvGLUBlock(1,
args.emb_dim,
512,
dropout=dropout)
for i in range(1, 4, 1):
blocks['conv2-%d' % i] = ConvGLUBlock(5,
512,
512,
bottlececk_dim=128,
dropout=dropout)
for i in range(1, 4, 1):
blocks['conv3-%d' % i] = ConvGLUBlock(5,
512,
512,
bottlececk_dim=256,
dropout=dropout)
blocks['conv4'] = ConvGLUBlock(1,
512,
2048,
bottlececk_dim=1024,
dropout=dropout)
last_dim = 2048
elif model_size == '9':
blocks['conv1'] = ConvGLUBlock(4,
args.emb_dim,
807,
dropout=dropout)
for i in range(1, 4, 1):
blocks['conv2-%d-1' % i] = ConvGLUBlock(4,
807,
807,
dropout=dropout)
blocks['conv2-%d-2' % i] = ConvGLUBlock(4,
807,
807,
dropout=dropout)
last_dim = 807
elif model_size == '13':
blocks['conv1'] = ConvGLUBlock(4,
args.emb_dim,
1268,
dropout=dropout)
for i in range(1, 13, 1):
blocks['conv2-%d' % i] = ConvGLUBlock(4,
1268,
1268,
dropout=dropout)
last_dim = 1268
elif model_size == '14':
for i in range(1, 4, 1):
blocks['conv1-%d' % i] = ConvGLUBlock(
6, args.emb_dim if i == 1 else 850, 850, dropout=dropout)
blocks['conv2'] = ConvGLUBlock(1, 850, 850, dropout=dropout)
for i in range(1, 5, 1):
blocks['conv3-%d' % i] = ConvGLUBlock(5,
850,
850,
dropout=dropout)
blocks['conv4'] = ConvGLUBlock(1, 850, 850, dropout=dropout)
for i in range(1, 4, 1):
blocks['conv5-%d' % i] = ConvGLUBlock(4,
850,
850,
dropout=dropout)
blocks['conv6'] = ConvGLUBlock(4, 850, 1024, dropout=dropout)
blocks['conv7'] = ConvGLUBlock(4, 1024, 2048, dropout=dropout)
last_dim = 2048
elif model_size == '14B':
blocks['conv1'] = ConvGLUBlock(5,
args.emb_dim,
512,
dropout=dropout)
for i in range(1, 4, 1):
blocks['conv2-%d' % i] = ConvGLUBlock(5,
512,
512,
bottlececk_dim=128,
dropout=dropout)
for i in range(1, 4, 1):
blocks['conv3-%d' % i] = ConvGLUBlock(5,
512 if i == 1 else 1024,
1024,
bottlececk_dim=512,
dropout=dropout)
for i in range(1, 7, 1):
blocks['conv4-%d' % i] = ConvGLUBlock(5,
1024 if i == 1 else 2048,
2048,
bottlececk_dim=1024,
dropout=dropout)
blocks['conv5'] = ConvGLUBlock(5,
2048,
4096,
bottlececk_dim=1024,
dropout=dropout)
last_dim = 4096
else:
raise NotImplementedError(model_size)
self.blocks = nn.Sequential(blocks)
if args.adaptive_softmax:
self.adaptive_softmax = nn.AdaptiveLogSoftmaxWithLoss(
last_dim,
self.vocab,
# cutoffs=[self.vocab // 10, 3 * self.vocab // 10],
cutoffs=[self.vocab // 25, self.vocab // 5],
div_value=4.0)
self.output = None
else:
self.adaptive_softmax = None
self.output = nn.Linear(last_dim, self.vocab)
if args.tie_embedding:
if args.n_units != args.emb_dim:
raise ValueError(
'When using the tied flag, n_units must be equal to emb_dim.'
)
self.output.weight = self.embed.weight
self.reset_parameters(args.param_init)
