def __init__(self,
word_vectors,
char_vectors,
hidden_size,
num_heads,
char_embed_drop_prob,
drop_prob=0.1):
super(SketchyReader, self).__init__()
'''class QANet(nn.Module):
def __init__(self, word_vectors, char_vectors, hidden_size, device, drop_prob=0.):
super(QANet, self).__init__()
self.device = device'''
self.emb = layers.Embedding(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=hidden_size,
char_embed_drop_prob=char_embed_drop_prob,
word_embed_drop_prob=drop_prob)
hidden_size *= 2 # update hidden size for other layers due to char embeddings
self.c_resizer = layers.Initialized_Conv1d(hidden_size, 128)
self.q_resizer = layers.Initialized_Conv1d(hidden_size, 128)
self.model_resizer = layers.Initialized_Conv1d(512, 128)
self.enc = layers.StackedEncoder(
num_conv_blocks=4,
kernel_size=7,
num_heads=num_heads,
dropout=drop_prob) # embedding encoder layer
self.att = layers.BiDAFAttention(
hidden_size=128,
drop_prob=drop_prob) # context-query attention layer
# self.mod1 = layers.StackedEncoder(num_conv_blocks=2,
# kernel_size=7,
# dropout=drop_prob) # model layer
# self.mod2 = layers.StackedEncoder(num_conv_blocks=2,
# kernel_size=7,
# dropout=drop_prob) # model layer
# self.mod3 = layers.StackedEncoder(num_conv_blocks=2,
# kernel_size=7,
# dropout=drop_prob) # model layer
self.model_encoder_layers = nn.ModuleList([
layers.StackedEncoder(num_conv_blocks=2,
kernel_size=7,
dropout=drop_prob) for _ in range(7)
])
self.out = layers.SketchyOutput(hidden_size=128) # output layer
def __init__(self,
word_vectors,
char_vectors,
context_max_len,
query_max_len,
d_model,
train_cemb=False,
pad=0,
dropout=0.1,
num_head=8):
"""
"""
super(QANet, self).__init__()
if train_cemb:
self.char_emb = nn.Embedding.from_pretrained(char_vectors,
freeze=False)
print("Training char_embeddings")
else:
self.char_emb = nn.Embedding.from_pretrained(char_vectors)
self.word_emb = nn.Embedding.from_pretrained(word_vectors)
self.LC = context_max_len
self.LQ = query_max_len
self.num_head = num_head
self.pad = pad
self.dropout = dropout
wemb_dim = word_vectors.size()[1]
cemb_dim = char_vectors.size()[1]
#print("Word vector dim-%d, Char vector dim-%d" % (wemb_dim, cemb_dim))
#Layer Declarations
self.emb = layers.Embedding(wemb_dim, cemb_dim, d_model)
self.emb_enc = layers.Encoder(num_conv=4,
d_model=d_model,
num_head=num_head,
k=7,
dropout=0.1)
self.cq_att = layers.CQAttention(d_model=d_model)
self.cq_resizer = layers.Initialized_Conv1d(
d_model * 4, d_model
) #Foward layer to reduce dimension of cq_att output back to d_dim
self.model_enc_blks = nn.ModuleList([
layers.Encoder(num_conv=2,
d_model=d_model,
num_head=num_head,
k=5,
dropout=0.1) for _ in range(7)
])
self.out = layers.QAOutput(d_model)
def __init__(self,
word_vectors,
char_vectors,
context_max_len,
query_max_len,
d_model,
d_head,
mem_len=0,
same_length=False,
clamp_len=-1,
train_cemb=False,
pad=0,
dropout=0.1,
num_head=8):
"""
"""
super(QANet, self).__init__()
if train_cemb:
self.char_emb = nn.Embedding.from_pretrained(char_vectors,
freeze=False)
else:
self.char_emb = nn.Embedding.from_pretrained(char_vectors)
self.word_emb = nn.Embedding.from_pretrained(word_vectors)
self.LC = context_max_len
self.LQ = query_max_len
self.num_head = num_head
self.pad = pad
self.dropout = dropout
self.mem_len = mem_len
self.d_head = d_head
self.d_model = d_model
self.num_head = num_head
self.same_length = same_length
self.clamp_len = clamp_len
self.ext_len = 0
wemb_dim = word_vectors.size()[1]
cemb_dim = char_vectors.size()[1]
#Layer Declarations
self.emb = layers.Embedding(wemb_dim, cemb_dim, d_model)
self.emb_enc = layers.Encoder(4,
num_head,
d_model,
d_head,
d_inner=d_model * 4,
k=7,
dropout=0.1) #Hard coded
self.cq_att = layers.CQAttention(d_model=d_model)
self.cq_resizer = layers.Initialized_Conv1d(
d_model * 4, d_model
) #Foward layer to reduce dimension of cq_att output back to d_dim
self.model_enc_blks = nn.ModuleList([
layers.Encoder(2,
num_head,
d_model,
d_head,
d_inner=d_model * 4,
k=5,
dropout=0.1) for _ in range(7)
])
self.out = layers.QAOutput(d_model)
self.drop = nn.Dropout(dropout)
self._create_parameters()