def __init__(self, n_vocab, n_units):
super(ContinuousBoW, self).__init__()
with self.init_scope():
self.embed = L.EmbedID(n_vocab,
n_units,
initialW=I.Uniform(1. / n_units))
self.out = L.Linear(n_units, n_vocab, initialW=0)
def __init__(self, n_vocab, n_units, loss_func):
super(SkipGram, self).__init__()
with self.init_scope():
self.embed = L.EmbedID(n_vocab,
n_units,
initialW=I.Uniform(1. / n_units))
self.loss_func = loss_func
def __init__(self, emb_dim, n_out, depth=[2, 2, 2, 2]):
super(VDCNN, self).__init__()
with self.init_scope():
self.embed = L.EmbedID(emb_dim,
16,
initializers.Normal(),
ignore_label=-1)
self.conv1 = L.ConvolutionND(1,
16,
64,
3,
1,
1,
initialW=initializers.HeUniform(),
initial_bias=initializers.Uniform(
np.sqrt(6 / (16 * 3))))
self.cb2 = BuildBlock(depth[0], 64, 128)
self.cb3 = BuildBlock(depth[1], 128, 256)
self.cb4 = BuildBlock(depth[2], 256, 512)
self.cb5 = BuildBlock(depth[3], 512, 512)
self.fc6 = L.Linear(
4096,
2048,
initialW=initializers.Uniform(1 / np.sqrt(4096)),
initial_bias=initializers.Uniform(1 / np.sqrt(4096)))
self.fc7 = L.Linear(
2048,
2048,
initialW=initializers.Uniform(1 / np.sqrt(2048)),
initial_bias=initializers.Uniform(1 / np.sqrt(2048)))
self.fc8 = L.Linear(
2048,
n_out,
initialW=initializers.Uniform(1 / np.sqrt(2048)),
initial_bias=initializers.Uniform(1 / np.sqrt(2048)))
def __init__(self, input_space, output_space, embedding_size,\
lstm_size, lstm_dropout, zero_bias=False):
"""
Initialise this network
Args:
input_space -- here space of observations x
output_space -- here space of actions a
embedding_size -- size of the embedding layers, both for
observations -> embeddings and embeddings -> joint embeddings
lstm_size -- size of the lstm hidden layer for both the behaviour
and inference policies
lstm_dropout -- lstm dropout ratio between 0 and 1, i.e. percentage of units
dropped to improve generalisation
zero_bias -- False by default, set True for low uniformly distributed initial bias
"""
super(ClosedLoopEmpowermentNet, self).__init__()
# TODO: Potentially relax assumptions;
# This is presently custom-tailored to the Gridworld environment
assert isinstance(output_space, gym.spaces.Discrete)
assert isinstance(input_space, gym.spaces.Box)
assert input_space.shape is not None
assert embedding_size > 0
assert lstm_size > 0
self._input_space = input_space
self._output_space = output_space
self._lstm_size = lstm_size
in_channels = input_space.shape[0]
initial_bias = None if zero_bias else I.Uniform(1e-4)
# Define the network
with self.init_scope():
# Embeddings are fully connected, one-layer linear networks
# Will be wrapped in rectified linear units later for non-linearity
# Adding initial bias means that most rectified units will have input > 0
# and thus be active in beginning
# Embedding for input observations x_t -> u_t
self.lin_embed_p = L.Linear(in_size=in_channels, out_size=embedding_size,\
initial_bias=initial_bias)
# Embedding for joint embeddings (u_t, u_f) -> v
self.lin_embed_q = L.Linear(in_size=embedding_size*2, out_size=embedding_size,\
initial_bias=initial_bias)
# LSTM for inference policy, taking input of form [h^p_t, v, a_{t-1}]
self.lstm_q = L.NStepLSTM(n_layers=1,\
in_size=(lstm_size + embedding_size + self._output_space.n),\
out_size=lstm_size, dropout=lstm_dropout)
# Linear transformation of LSTM output to actions for inference policy
self.lin_pi_q = L.Linear(in_size=lstm_size,
out_size=self._output_space.n,
initial_bias=initial_bias)
def __init__(self, item_num, hidden_num, rating_num, encoding_size=-1):
if encoding_size <= 0:
super(NadeIn,
self).__init__(a=L.Linear(item_num * rating_num,
hidden_num,
initialW=I.Uniform(0.06),
initial_bias=I.Constant(0)), )
else:
super(NadeIn, self).__init__(
a=L.Linear(item_num * rating_num,
encoding_size,
nobias=True,
initialW=I.Uniform(0.06)),
b=L.Linear(encoding_size,
hidden_num,
initialW=I.Uniform(0.06),
initial_bias=I.Constant(0)),
)
def __init__(self):
super(Actor, self).__init__(
h1 = L.Linear(O_DIM, 400),
h2 = L.Linear(400, 300),
h3 = L.Linear(300, A_DIM, initialW=initializers.Uniform(scale=0.003)),
)
self.optimizer = optimizers.Adam(alpha=ACTOR_LEARNING_RATE)
self.optimizer.setup(self)
self.optimizer.add_hook(optimizer_hooks.GradientClipping(2.0))
def setUp(self):
N = 2
in_channels = 3
out_channels = 2
ndim = len(self.dims)
ksize = (3, ) * ndim
stride = (2, ) * ndim
pad = (1, ) * ndim
if self.used_outsize == 'case1' or self.used_outsize == 'None':
# Use output size determined with get_deconv_outsize.
outs = tuple(
conv.get_deconv_outsize(d, k, s, p)
for (d, k, s, p) in zip(self.dims, ksize, stride, pad))
elif self.used_outsize == 'case2':
# Use possible output size other than the one determined with
# get_deconv_outsize.
outs = tuple(
conv.get_deconv_outsize(d, k, s, p) + 1
for (d, k, s, p) in zip(self.dims, ksize, stride, pad))
if self.used_outsize != 'None':
outsize = outs
else:
outsize = None
if not self.nobias:
initial_bias = initializers.Uniform(scale=1, dtype=self.dtype)
else:
initial_bias = None
self.link = deconvolution_nd.DeconvolutionND(ndim,
in_channels,
out_channels,
ksize,
stride=stride,
pad=pad,
outsize=outsize,
initial_bias=initial_bias,
nobias=self.nobias)
self.link.cleargrads()
x_shape = (N, in_channels) + self.dims
self.x = numpy.random.uniform(-1, 1, x_shape).astype(self.dtype)
gy_shape = (N, out_channels) + outs
self.gy = numpy.random.uniform(-1, 1, gy_shape).astype(self.dtype)
self.check_forward_options = {}
self.check_backward_options = {'eps': 1e-2, 'atol': 1e-4, 'rtol': 1e-3}
if self.dtype == numpy.float16:
self.check_forward_options = {'atol': 5e-3, 'rtol': 5e-2}
self.check_backward_options = {
'eps': 2**-3,
'atol': 1e-2,
'rtol': 1e-1
}
def __init__(self, grid_weight_share=False):
"""
"""
super(CapsNet, self).__init__()
self.grid_weight_share = grid_weight_share
self.recon_loss_weight = 0.0005
init_scale = 0.1
# init_W = initializers.HeNormal()
init_W = initializers.Uniform(scale=init_scale)
conv1_param = dict(in_channels=1,
out_channels=256,
ksize=9,
stride=1,
nobias=True,
initialW=init_W)
pcaps_param = dict(in_caps=1,
in_dims=256,
out_caps=32,
out_dims=8,
n_iters=1,
ksize=9,
stride=2,
nobias=True,
initialW=init_W)
if self.grid_weight_share:
dcapsconv_param = dict(in_caps=32,
in_dims=8,
out_caps=10,
out_dims=16,
n_iters=3,
flat_output=True,
ksize=1,
stride=1,
nobias=True,
initialW=init_W)
else:
dcapslin_param = dict(in_caps=32 * 6 * 6,
in_dims=8,
out_caps=10,
out_dims=16,
n_iters=3,
initialW=init_W)
with self.init_scope():
self.conv1 = L.Convolution2D(**conv1_param)
self.primarycaps = CapsConv(**pcaps_param)
if self.grid_weight_share:
self.digitcaps = CapsConv(**dcapsconv_param)
else:
self.digitcaps = CapsLinear(**dcapslin_param)
# for reconstruction
self.fc1 = L.Linear(16 * 10, 512, initialW=init_W)
self.fc2 = L.Linear(512, 1024, initialW=init_W)
self.fc3 = L.Linear(1024, 784, initialW=init_W)
def __init__(self, n_vocab, n_embed, n_units):
super(CandW, self).__init__()
with self.init_scope():
self.embed = L.EmbedID(n_vocab,
n_embed,
initialW=initializers.Uniform(1. / n_embed))
self.l1 = L.Linear(n_embed * (args.window * 2 + 1),
n_units) # n_embed * (window*2 + 1)-> n_units
self.l2 = L.Linear(n_units, 1) # n_units -> 1
def setUp(self):
ndim = len(self.dims)
self.ksize = (3, ) * ndim
self.stride = (2, ) * ndim
self.pad = (1, ) * ndim
if self.in_channels == 'omit':
self.link = convolution_nd.ConvolutionND(
ndim,
2,
self.ksize,
stride=self.stride,
pad=self.pad,
groups=self.groups,
initial_bias=initializers.Uniform(scale=1., dtype=self.dtype))
else:
self.link = convolution_nd.ConvolutionND(
ndim,
self.in_channels,
2,
self.ksize,
stride=self.stride,
pad=self.pad,
groups=self.groups,
initial_bias=initializers.Uniform(scale=1., dtype=self.dtype))
self.link.cleargrads()
x_shape = (2, 4) + self.dims
self.x = numpy.random.uniform(-1, 1, x_shape).astype(self.dtype)
gy_shape = (2, 2) + tuple(
conv.get_conv_outsize(d, k, s, p)
for (d, k, s,
p) in zip(self.dims, self.ksize, self.stride, self.pad))
self.gy = numpy.random.uniform(-1, 1, gy_shape).astype(self.dtype)
self.check_backward_options = {'eps': 1e-2, 'atol': 1e-3, 'rtol': 1e-3}
if self.dtype == numpy.float16:
self.check_backward_options = {
'eps': 2**-4,
'atol': 2**-4,
'rtol': 2**-4
}
def __init__(self, item_num, hidden_num, rating_num, encoding_size=-1):
if encoding_size <= 0:
super(NadeOut,
self).__init__(p=L.Linear(hidden_num,
rating_num * item_num,
initialW=I.Uniform(0.06),
initial_bias=I.Constant(0)), )
else:
super(NadeOut, self).__init__(
p=L.Linear(hidden_num,
encoding_size,
nobias=True,
initialW=I.Uniform(0.06)),
q=L.Linear(encoding_size,
rating_num * item_num,
initialW=I.Uniform(0.06),
initial_bias=I.Constant(0)),
)
self.item_num = item_num
self.rating_num = rating_num
def __init__(self, n_vocab, n_units, loss_func, ori_con_data):
super(ContinuousBoW, self).__init__()
with self.init_scope():
self.embed = L.EmbedID(n_vocab, n_units, initialW=I.Uniform(1. / n_units))
# 派生単語と元の単語の初期ベクトルを統一する
for i in range(len(ori_con_data)):
self.embed.W.data[ori_con_data[i][0]] = self.embed.W.data[ori_con_data[i][1]]
self.loss_func = loss_func
def __init__(self):
super(Critic, self).__init__(
h1=L.Linear(O_DIM, 400),
h2_s=L.Linear(400, 300, nobias=True),
h2_a=L.Linear(A_DIM, 300),
h3=L.Linear(300, 1, initialW=initializers.Uniform(scale=0.003)),
)
self.optimizer = optimizers.Adam(alpha=CRITIC_LEARNING_RATE)
self.optimizer.setup(self)
#self.optimizer.add_hook(optimizer_hooks.WeightDecay(CRITIC_WEIGHT_DECAY))
self.optimizer.add_hook(optimizer_hooks.GradientClipping(2.0))
def __init__(self, n_vocab_char, n_units, n_units_char):
super(RNN, self).__init__()
with self.init_scope():
self.embed = L.EmbedID(n_vocab_char,
n_units_char,
initialW=I.Uniform(
1. / n_units_char)) # word embedding
self.mid = L.LSTM(n_units_char,
n_units_char) # the first LSTM layer
self.out = L.Linear(n_units_char,
n_units) # the feed-forward output layer
def __init__(self, nc_size, base=32):
self.nc_size = nc_size
wscale = math.sqrt(1 / (base * 32) / 3 / 3)
wout = initializers.Uniform(scale=wscale)
wscale = math.sqrt(1 / (base * 32) / 2 / 2)
wcls = initializers.Uniform(scale=wscale)
super(Discriminator, self).__init__()
with self.init_scope():
self.c0 = CBR(3, base, 4, 2, 1, down=True, act=F.leaky_relu)
self.c1 = CBR(base, base * 2, 4, 2, 1, down=True, act=F.leaky_relu)
self.c2 = CBR(base * 2,
base * 4,
4,
2,
1,
down=True,
act=F.leaky_relu)
self.c3 = CBR(base * 4,
base * 8,
4,
2,
1,
down=True,
act=F.leaky_relu)
self.c4 = CBR(base * 8,
base * 16,
4,
2,
1,
down=True,
act=F.leaky_relu)
self.c5 = CBR(base * 16,
base * 32,
4,
2,
1,
down=True,
act=F.leaky_relu)
self.lembed = L.Linear(None, base * 32, initialW=wout)
self.l1 = L.Linear(None, 1, initialW=wout)
def __init__(self, n_vocab, n_embed, n_units, n_senti):
super(SentimentEmbed, self).__init__()
with self.init_scope():
self.embed = L.EmbedID(n_vocab,
n_embed,
initialW=initializers.Uniform(1. / n_embed))
self.l1 = L.Linear(n_embed * (args.window * 2 + 1),
n_units) # n_embed * (window*2 + 1)-> n_units
self.lc = L.Linear(n_units, 1) # n_units -> 1
self.ls = L.Linear(n_units, 2) # n_units -> 1
self.lws = L.Linear(n_embed, n_senti) # n_embed -> n_senti
def make_network(self, scale, bias=0):
dim = self.dim
self.lmNet = chainer.Chain()
self.lmNet.add_link(
'Embed',
L.EmbedID(len(self.vocab), self.dim, initialW=I.Uniform(scale)))
for i in range(self.layerNum):
self.lmNet.add_link(
'x2h%s_0' % (i),
L.Linear(dim, dim, initialW=I.Uniform(scale), nobias=True))
self.lmNet.add_link(
'x2t%s_0' % (i),
L.Linear(dim, dim, initialW=I.Uniform(scale), nobias=True))
for j in range(self.depth):
self.lmNet.add_link(
'h2h%s_%s' % (i, j),
L.Linear(dim, dim, initialW=I.Uniform(scale)))
self.lmNet.add_link(
'h2t%s_%s' % (i, j),
L.Linear(dim,
dim,
initialW=I.Uniform(scale),
initial_bias=bias))
self.lmNet.add_link(
'Output', L.Linear(dim, len(self.vocab),
initialW=I.Uniform(scale)))
def __init__(self, nc, dtype=np.float32):
super(InstanceNormalization, self).__init__()
self.nc = nc
self.dtype = dtype
self.bn = None
self.prev_batch = None
# self._device_id = cuda.get_device()
self.add_param('gamma', nc, dtype=dtype)
initializers.Uniform(self.gamma.data)
self.add_param('beta', nc, dtype=dtype)
initializers.Zero(self.beta.data)
def get_initializers(self):
if self.initialW == 'zero':
weight_initializer = initializers.constant.Zero()
elif self.initialW == 'random':
weight_initializer = initializers.GlorotUniform(
rng=numpy.random.RandomState(seed=0))
if self.initial_bias == 'zero':
bias_initializer = initializers.constant.Zero()
elif self.initial_bias == 'random':
bias_initializer = initializers.Uniform(
rng=numpy.random.RandomState(seed=0))
return weight_initializer, bias_initializer
def __init__(self, obs_size, num_actions, nhidden):
"""Initialize weights"""
# use LeCunUniform weight initialization for weights
self.initializer = initializers.LeCunUniform()
self.bias_initializer = initializers.Uniform(1e-4)
super(QNetwork, self).__init__(
feature_layer = L.Linear(obs_size, nhidden,
initialW = self.initializer,
initial_bias = self.bias_initializer),
action_values = L.Linear(nhidden, num_actions,
initialW=self.initializer,
initial_bias = self.bias_initializer)
)
def __init__(self,
input_space,
output_space,
zero_bias=False,
in_channels=None):
super(ChainerDQN, self).__init__()
assert isinstance(input_space, gym.Space)
assert input_space.shape is not None
if isinstance(output_space, gym.spaces.Discrete):
out_units = output_space.n
elif isinstance(output_space, gym.spaces.Box):
assert len(output_space.shape) == 1
out_units = output_space.shape[0]
else:
raise NotImplementedError
self._input_space = input_space
self._output_space = output_space
initial_bias = None if zero_bias else I.Uniform(1e-4)
in_channels = get_in_channels(input_space, in_channels)
with self.init_scope():
self.conv0 = L.Convolution2D(in_channels,
32,
ksize=8,
stride=4,
initialW=I.HeUniform(),
initial_bias=initial_bias)
self.conv1 = L.Convolution2D(None,
64,
ksize=4,
stride=2,
initialW=I.HeUniform(),
initial_bias=initial_bias)
self.conv2 = L.Convolution2D(None,
64,
ksize=3,
stride=1,
initialW=I.HeUniform(),
initial_bias=initial_bias)
self.fc0 = L.Linear(None,
512,
initialW=I.HeNormal(scale=0.01),
initial_bias=initial_bias)
self.fc1 = L.Linear(None,
out_units,
initialW=I.HeNormal(scale=0.01),
initial_bias=initial_bias)
self(np.zeros((1, ) + input_space.shape, input_space.dtype))
def __init__(self,
wordCounts,
embed_size,
lossfun=negative_sampling.NegativeSampling):
# 这句话的定义是实现其父类chainer.link.Chain的init
super(SkipModel, self).__init__()
with self.init_scope():
# self.embedings = np.random.uniform(0,1,(dataMes.vocabSize,embed_size)).astype(np.float32)
self.embedings = L.EmbedID(len(wordCounts),
embed_size,
initialW=I.Uniform(1. / embed_size))
# 负采样函数需要获取词频数据 self.wordCnt, 和词向量维度信息embed_size,这样才能构建一个第二层映射矩阵
self.lossfun = negative_sampling.NegativeSampling(embed_size,
wordCounts,
sample_size=5)
def __init__(self,
n_vocab,
n_units,
loss_func,
n_vocab_char=None,
n_units_char=None):
super(SkipGram, self).__init__()
with self.init_scope():
if args.subword == 'none':
self.embed = L.EmbedID(n_vocab,
n_units,
initialW=I.Uniform(1. / n_units))
if args.subword == 'rnn':
self.rnn = RNN(n_vocab_char, n_units, n_units_char)
self.loss_func = loss_func
def __init__(self, in_channels, out_channels, use_tanh=False):
if True:
initialW = initializers.Uniform(scale=0.05)
else:
wstd = math.sqrt(in_channels)
initialW = initializers.Normal(wstd)
super().__init__(W=Convolution1D(in_channels,
3 * out_channels,
1,
stride=1,
pad=0,
initialW=initialW),
bf=Bias(shape=(out_channels, )),
br=Bias(shape=(out_channels, )))
self.use_highway_connections = in_channels == out_channels
self.use_tanh = use_tanh
self.reset_state()
def __init__(self, vocab, vocab_ngram_tokens, n_units, n_units_char, dropout,
subword): # dropout ratio, zero indicates no dropout
super(SUMAVG, self).__init__()
with self.init_scope():
if subword.startswith('sum'):
self.f_sumavg = F.sum
if subword.startswith('avg'):
self.f_sumavg = F.average
self.embed = L.EmbedID(
len(vocab_ngram_tokens.lst_words) + 2, n_units_char,
initialW=I.Uniform(1. / n_units_char)) # ngram tokens embedding plus 2 for OOV and end symbol.
self.n_ngram = vocab_ngram_tokens.metadata["max_gram"] - vocab_ngram_tokens.metadata["min_gram"] + 1
self.dropout = dropout
self.vocab = vocab
self.vocab_ngram_tokens = vocab_ngram_tokens
def __init__(self,
input_space,
output_space,
hidden_layer_sizes=(512, 512, 512),
activation=F.relu,
zero_bias=False):
"""
Args:
input_space -- the space that inputs will be drawn from
output_space -- the space that outputs will be drawn from.
Keyword Args:
hidden_layer_sizes -- the number of units in each hidden layer in order
[(512, 512, 512)]
activation -- the activation function used on all layers [relu]
zero_bias -- whether the bias should be initialized to zero [False]
"""
assert isinstance(input_space, gym.Space)
assert input_space.shape is not None
if isinstance(output_space, gym.spaces.Discrete):
out_units = output_space.n
elif isinstance(output_space, gym.spaces.Box):
assert len(output_space.shape) == 1
out_units = output_space.shape[0]
else:
raise NotImplementedError
self._input_space = input_space
self._output_space = output_space
self._num_hidden = len(hidden_layer_sizes)
assert self._num_hidden > 0
initial_weights = I.HeUniform()
initial_bias = None if zero_bias else I.Uniform(1e-4)
links = []
for units in hidden_layer_sizes:
links.append(
L.Linear(None, units, False, initial_weights, initial_bias))
links.append(
L.Linear(None, out_units, False, initial_weights, initial_bias))
self._activation = activation
super(ChainerMLP, self).__init__(*links)
self(np.zeros((1, ) + input_space.shape, input_space.dtype))
def __init__(self,
n_vocab_char,
n_units,
n_units_char,
index2charIds,
dropout=.2): #dropout ratio, zero indicates no dropout
super(RNN, self).__init__()
with self.init_scope():
self.embed = L.EmbedID(n_vocab_char,
n_units_char,
initialW=I.Uniform(
1. / n_units_char)) # word embedding
self.mid = L.LSTM(n_units_char,
n_units_char) # the first LSTM layer
self.out = L.Linear(n_units_char,
n_units) # the feed-forward output layer
self.dropout = dropout
self.index2charIds = index2charIds
def __init__(self,
n_vocab,
n_units,
filter_sizes,
n_filter,
drop_rate,
n_class,
init_E=None):
"""
"""
self.n_vocab = n_vocab
self.n_units = n_units
self.filter_sizes = filter_sizes
self.n_filter = n_filter
self.drop_rate = drop_rate
self.n_class = n_class
super(CNNSentenceClassifier, self).__init__()
with self.init_scope():
# Embedding
if init_E is None:
init_E = initializers.Uniform(1. / n_units)
self.add_link("embed", L.EmbedID(n_vocab, n_units,
initialW=init_E))
# Convolutions
for sz in filter_sizes:
self.add_link(
"conv_{}".format(sz),
L.Convolution2D(1,
n_filter,
ksize=(sz, n_units),
stride=1,
pad=0,
initialW=initializers.HeNormal()))
# FC
self.add_link(
"fc",
L.Linear(len(filter_sizes) * n_filter,
n_class,
initialW=initializers.HeNormal()))
def __init__(self, vocab, vocab_ngram_tokens, n_units, n_units_char, dropout,
subword): # dropout ratio, zero indicates no dropout
super(RNN, self).__init__()
with self.init_scope():
self.embed = L.EmbedID(
len(vocab_ngram_tokens.lst_words) + 2, n_units_char,
initialW=I.Uniform(1. / n_units_char)) # ngram tokens embedding plus 2 for OOV and end symbol.
if 'lstm' in subword:
self.mid = L.LSTM(n_units_char, n_units_char * 2)
self.out = L.Linear(n_units_char * 2, n_units_char) # the feed-forward output layer
if 'bilstm' in subword:
self.mid_b = L.LSTM(n_units_char, n_units_char * 2)
self.out_b = L.Linear(n_units_char * 2, n_units_char)
self.n_ngram = vocab_ngram_tokens.metadata["max_gram"] - vocab_ngram_tokens.metadata["min_gram"] + 1
self.final_out = L.Linear(n_units * (self.n_ngram), n_units)
self.dropout = dropout
self.vocab = vocab
self.vocab_ngram_tokens = vocab_ngram_tokens
self.subword = subword
def __init__(self,
subword,
vocab,
vocab_ngram_tokens,
dimensions,
loss_func,
dropout=0): # dropout ratio, zero indicates no dropout
super(SkipGram, self).__init__()
with self.init_scope():
self.subword = subword
self.vocab = vocab
self.vocab_ngram_tokens = vocab_ngram_tokens
self.n_ngram = vocab_ngram_tokens.metadata[
"max_gram"] - vocab_ngram_tokens.metadata["min_gram"] + 1
if 'none' in subword:
self.word_embed = L.EmbedID(
len(vocab.lst_words) + 2,
dimensions,
initialW=I.Uniform(
1. / dimensions)) # plus 2 for OOV and end symbol.
else:
self.word_embed = None
if subword.startswith('_none'):
self.f = None
# if subword.startswith('cnn_'):
# self.f = CNN(vocab, vocab_ngram_tokens, dimensions, dimensions, dropout)
if subword.startswith('cnn1d'):
self.f = CNN1D(vocab, vocab_ngram_tokens, dimensions,
dimensions, dropout, args.subword)
if subword.startswith('bilstm') or subword.startswith('lstm'):
self.f = RNN(vocab, vocab_ngram_tokens, dimensions, dimensions,
dropout, args.subword)
if subword.startswith('avg') or subword.startswith('sum'):
self.f = SUMAVG(vocab, vocab_ngram_tokens, dimensions,
dimensions, dropout, args.subword)
self.loss_func = loss_func