##############################
-class Residual(nn.Module):
+class WithResidual(nn.Module):
def __init__(self, *f):
super().__init__()
self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
##############################
-class PositionalEncoding(nn.Module):
+class AddPositionalEncoding(nn.Module):
def __init__(self, len_max):
super().__init__()
self.len_max = len_max
- # From Vaswani et al 2018
- # PE_{t,2i} = sin(t/(L^{2i/D}))
- # PE_{t,2i+1} = cos(t/(L^{2i/D}))
+ # [Vaswani et al 2018] PE_{t,2i} = sin(t/(L^{2i/D})), PE_{t,2i+1} = cos(t/(L^{2i/D}))
def forward(self, x):
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
##############################
class QKVAttention(nn.Module):
- def __init__(self, dim_in, dim_qk, dim_v,
+ def __init__(self,
+ dim_in, dim_qk, dim_v,
nb_heads = 1, causal = False, attention_dropout = 0.0):
super().__init__()
def randw(*d):
- return nn.Parameter(torch.empty(*d).normal_(0, 1 / math.sqrt(d[-1])))
+ return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
+
+ self.causal = causal
+ self.attention_dropout = attention_dropout
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(dim_in, dim_v * nb_heads)
- 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
+
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))
+
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'))
+
a = a.softmax(dim = 3)
a = F.dropout(a, self.attention_dropout, self.training)
- y = torch.einsum('nhts,nhsd->nthd', a, v)
- y = y.flatten(2) @ self.w_o
+ y = torch.einsum('nhts,nhsd->nthd', a, v).flatten(2)
+
+ y = y @ self.w_o
return y
def __init__(self,
vocabulary_size,
dim_model, dim_keys, dim_hidden,
- nb_heads, nb_blocks, dropout = 0.):
+ nb_heads, nb_blocks,
+ dropout = 0.0, len_max = 1e5):
super().__init__()
self.embedding = nn.Sequential(
nn.Embedding(vocabulary_size, dim_model),
nn.Dropout(dropout),
- PositionalEncoding(len_max = 1e5),
+ AddPositionalEncoding(len_max),
)
+ # Small embedding initialization
+ with torch.no_grad():
+ self.embedding[0].weight.normal_(0, 2e-2)
+
trunk_blocks = [ ]
for _ in range(nb_blocks):
trunk_blocks += [
- Residual(
- nn.LayerNorm(dim_model),
+ WithResidual(
+ nn.LayerNorm((dim_model,)),
QKVAttention(
dim_in = dim_model,
- dim_qk = dim_keys, dim_v = dim_model // nb_heads,
+ dim_qk = dim_keys,
+ dim_v = dim_model // nb_heads,
nb_heads = nb_heads,
causal = True, attention_dropout = dropout
),
- nn.Linear(in_features = dim_model, out_features = dim_model),
),
- Residual(
- nn.LayerNorm(dim_model),
+ WithResidual(
+ nn.LayerNorm((dim_model,)),
nn.Linear(in_features = dim_model, out_features = dim_hidden),
nn.ReLU(),
nn.Linear(in_features = dim_hidden, out_features = dim_model),
self.readout = nn.Linear(in_features = dim_model, out_features = vocabulary_size)
def forward(self, x):
+ x = F.pad(x, (1, -1))
x = self.embedding(x)
x = self.trunk(x)
x = self.readout(x)
model = MyGPT(
vocabulary_size = vocabulary_size,
- dim_model = 16, dim_keys = 50, dim_hidden = 100,
+ dim_model = 18, dim_keys = 50, dim_hidden = 100,
nb_heads = 2, nb_blocks = 3,
dropout = 0.1
)