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
##############################
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.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
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),
self.readout = nn.Linear(in_features = dim_model, out_features = vocabulary_size)
def forward(self, x):
+ x = F.pad(x, (1, 0))
x = self.embedding(x)
x = self.trunk(x)
x = self.readout(x)
- return x
+ return x[:, :-1]
######################################################################
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
)