def __init__(self, **kwargs):
super().__init__(**kwargs)
self.nfeat_factor = kwargs.get("nfeat_factor")
self.cfiltsz = kwargs.get("cfiltsz", kwargs.get("filtsz", [3]))
self.max_feat = kwargs.get("max_feat", 30)
self.gating = kwargs.get("gating", "skip")
self.num_gates = kwargs.get("num_gates", 1)
self.activation = kwargs.get("activation", "tanh")
self.wsz = kwargs.get("wsz", 30)
self.projsz = kwargs.get("projsz", 0)
self.pdrop = kwargs.get("pdrop", 0.5)
self.filtsz, self.nfeats = calc_nfeats(self.cfiltsz, self.nfeat_factor,
self.max_feat, self.wsz)
self.conv_outsz = int(np.sum(self.nfeats))
self.outsz = self.conv_outsz
if self.projsz > 0:
self.outsz = self.projsz
self.proj = pytorch_linear(self.conv_outsz, self.outsz)
self.embeddings = LookupTableEmbeddings(**kwargs)
self.char_comp = WithDropout(
ParallelConv(self.embeddings.output_dim, self.nfeats, self.filtsz,
self.activation), self.pdrop)
GatingConnection = SkipConnection if self.gating == "skip" else Highway
self.gating_seq = nn.Sequential(
OrderedDict([("gate-{}".format(i),
GatingConnection(self.char_comp.output_dim))
for i in range(self.num_gates)]))
def __init__(self, trainable=True, name=None, dtype=tf.float32, **kwargs):
trainable = kwargs.get("finetune", trainable)
super().__init__(trainable=trainable, name=name, dtype=dtype)
self._name = name
self.scope = kwargs.get("scope", "CharConv")
self.finetune = kwargs.get("finetune", trainable)
self.nfeat_factor = kwargs.get("nfeat_factor", None)
self.cfiltsz = kwargs.get("cfiltsz", kwargs.get("filtsz", [3]))
self.max_feat = kwargs.get("max_feat", 30)
self.gating = kwargs.get("gating", "skip")
self.num_gates = kwargs.get("num_gates", 1)
self.activation = kwargs.get("activation", "tanh")
self.wsz = kwargs.get("wsz", 30)
self.projsz = kwargs.get("projsz")
self.x = None
# These are the actual final filter sizes and num features
self.filtsz, self.nfeats = calc_nfeats(self.cfiltsz, self.nfeat_factor,
self.max_feat, self.wsz)
self.conv_outsz = np.sum(self.nfeats)
self.outsz = self.conv_outsz
if self.projsz is not None:
self.outsz = self.projsz
self.proj = tf.keras.layers.Dense(
self.outsz, bias_initializer=tf.constant_initializer(0.0))
self.embed = LookupTableEmbeddings(name=f"{self.name}/CharLUT",
finetune=self.finetune,
**kwargs)
def __init__(self, trainable=True, name=None, dtype=tf.float32, **kwargs):
trainable = kwargs.get("finetune", trainable)
super().__init__(trainable=trainable, name=name, dtype=dtype, **kwargs)
self.cpu_placement = bool(kwargs.get('cpu_placement', False))
self.scope = kwargs.get("scope", "CharConv")
self.finetune = kwargs.get("finetune", trainable)
self.nfeat_factor = kwargs.get("nfeat_factor", None)
self.cfiltsz = kwargs.get("cfiltsz", kwargs.get("filtsz", [3]))
self.max_feat = kwargs.get("max_feat", 30)
gating = kwargs.get("gating", "skip")
num_gates = kwargs.get("num_gates", 1)
activation = kwargs.get("activation", "tanh")
self.wsz = kwargs.get("wsz", 30)
self.projsz = kwargs.get("projsz")
self.x = None
# These are the actual final filter sizes and num features
filtsz, nfeats = calc_nfeats(self.cfiltsz, self.nfeat_factor, self.max_feat, self.wsz)
self.embed = LookupTableEmbeddings(name=f"{self.name}/CharLUT", finetune=self.finetune, **kwargs)
dsz = self.embed.output_dim
self.parallel_conv = ParallelConv(dsz, nfeats, filtsz, activation)
self.gating_fns = tf.keras.Sequential()
for _ in range(num_gates):
if gating == 'skip':
self.gating_fns.add(SkipConnection(self.parallel_conv.output_dim, activation))
else:
self.gating_fns.add(Highway(self.parallel_conv.output_dim))
self.outsz = self.parallel_conv.output_dim
if self.projsz is not None:
self.outsz = self.projsz
self.proj = tf.keras.layers.Dense(self.outsz, bias_initializer=tf.constant_initializer(0.0))
def test_feat_factor_capped():
gold_filtsz = [1, 2, 3, 4, 5]
feat_factor = 10
gold_nfeats = [10, 20, 30, 30, 30]
filtsz, nfeats = calc_nfeats(gold_filtsz, feat_factor, 30)
assert filtsz == gold_filtsz
assert nfeats == gold_nfeats
def test_feat_factor_manual():
gold_filtsz = [1, 2, 3, 4, 5]
feat_factor = 10
gold_nfeats = [10, 20, 30, 40, 50]
filtsz, nfeats = calc_nfeats(gold_filtsz, feat_factor, float("Infinity"))
assert filtsz == gold_filtsz
assert nfeats == gold_nfeats
def test_feat_factor():
gold_filtsz = [random.randint(1, 10) for _ in range(random.randint(2, 6))]
feat_factor = random.randint(10, 25)
max_feat = random.randint(30, 40)
gold_nfeats = np.minimum(np.array(gold_filtsz) * feat_factor, max_feat)
filtsz, nfeats = calc_nfeats(gold_filtsz, feat_factor, max_feat)
np.testing.assert_equal(filtsz, gold_filtsz)
np.testing.assert_equal(nfeats, gold_nfeats)
def test_extract_tuple():
filt_feat = [(random.randint(1, 10), random.randint(10, 20))
for _ in range(random.randint(2, 6))]
gold_filtsz = tuple(filter_and_size[0] for filter_and_size in filt_feat)
gold_nfeats = tuple(filter_and_size[1] for filter_and_size in filt_feat)
filtsz, nfeats = calc_nfeats(filt_feat)
assert filtsz == gold_filtsz
assert nfeats == gold_nfeats
def test_feat_factor_max_none():
filtsz = [random.randint(1, 10) for _ in range(random.randint(2, 6))]
feat_factor = 10
with pytest.raises(AssertionError):
calc_nfeats(filtsz, nfeat_factor=feat_factor, max_feat=None)
def test_use_nfeats():
filtsz = [random.randint(1, 10) for _ in range(random.randint(2, 6))]
input_nfeat = random.randint(1, 100)
gold_nfeats = [input_nfeat] * len(filtsz)
_, nfeat = calc_nfeats(filtsz, None, None, nfeats=input_nfeat)
assert nfeat == gold_nfeats
def test_use_nfeats_list():
filtsz = [random.randint(1, 10) for _ in range(random.randint(2, 6))]
nfeats = [random.randint(1, 10) for _ in range(len(filtsz))]
with pytest.raises(AssertionError):
_, nfeat = calc_nfeats(filtsz, None, None, nfeats=nfeats)
def test_use_nfeats_none():
filtsz = [random.randint(1, 10) for _ in range(random.randint(2, 6))]
with pytest.raises(AssertionError):
calc_nfeats(filtsz)
def test_use_nfeats_filtsz_unchanged():
gold_filtsz = [random.randint(1, 10) for _ in range(random.randint(2, 6))]
nfeat = random.randint(1, 100)
filtsz, _ = calc_nfeats(gold_filtsz, None, None, nfeats=nfeat)
assert filtsz == gold_filtsz