t = torch.arange(x.size(1), dtype = x.dtype, device = x.device)[:, None]
j = torch.arange(x.size(2), dtype = x.dtype, device = x.device)[None, :]
k = j%2
t = torch.arange(x.size(1), dtype = x.dtype, device = x.device)[:, None]
j = torch.arange(x.size(2), dtype = x.dtype, device = x.device)[None, :]
k = j%2
- return x + torch.sin(t / (self.len_max ** ((j - k) / x.size(2))) + math.pi/2 * k)[None, :, :]
+ pe = torch.sin(t / (self.len_max ** ((j - k) / x.size(2))) + math.pi/2 * k)
+ return x + pe # Let broadcasting to its job
- def __init__(self, dim_in, dim_qk, dim_v,
- nb_heads = 1, causal = False, attention_dropout = 0.0):
+ def __init__(
+ self,
+ dim_in, dim_qk, dim_v,
+ nb_heads = 1, causal = False, attention_dropout = 0.0
+ ):
self.w_q = randw(nb_heads, dim_qk, dim_in)
self.w_k = randw(nb_heads, dim_qk, dim_in)
self.w_v = randw(nb_heads, dim_v, dim_in)
self.w_q = randw(nb_heads, dim_qk, dim_in)
self.w_k = randw(nb_heads, dim_qk, dim_in)
self.w_v = randw(nb_heads, dim_v, dim_in)
- self.w_o = randw(nb_heads, dim_in, dim_v)
- self.causal = causal
- self.attention_dropout = attention_dropout
+ self.w_o = randw(dim_v * nb_heads, dim_in)
def forward(self, x_q, x_kv = None):
if x_kv is None: x_kv = x_q
def forward(self, x_q, x_kv = None):
if x_kv is None: x_kv = x_q
q = torch.einsum('ntc,hdc->nhtd', x_q, self.w_q)
k = torch.einsum('ntc,hdc->nhtd', x_kv, self.w_k)
v = torch.einsum('ntc,hdc->nhtd', x_kv, self.w_v)
q = torch.einsum('ntc,hdc->nhtd', x_q, self.w_q)
k = torch.einsum('ntc,hdc->nhtd', x_kv, self.w_k)
v = torch.einsum('ntc,hdc->nhtd', x_kv, self.w_v)
a = torch.einsum('nhtd,nhsd->nhts', q, k) / math.sqrt(q.size(3))
a = torch.einsum('nhtd,nhsd->nhts', q, k) / math.sqrt(q.size(3))
if self.causal:
mask = torch.arange(a.size(2), device = q.device)[None, None, :, None] \
< torch.arange(a.size(3), device = q.device)[None, None, None, :]
a = a.masked_fill(mask, float('-inf'))
if self.causal:
mask = torch.arange(a.size(2), device = q.device)[None, None, :, None] \
< torch.arange(a.size(3), device = q.device)[None, None, None, :]
a = a.masked_fill(mask, float('-inf'))
- y = torch.einsum('nhts,nhsd->nhtd', a, v)
- y = torch.einsum('nhtd,hcd->ntc', y, self.w_o)
+ y = torch.einsum('nhts,nhsd->nthd', a, v).flatten(2)
+
+ y = y @ self.w_o
- dim_model = 16, dim_keys = 50, dim_hidden = 100,
+ dim_model = 18, dim_keys = 50, dim_hidden = 100,