Hello, I’m facing “out of memory” error of my network on GPU and I checked that main problem is in EncoderBlock layer. Every iteration of the first cycle in the layer allocates new memory, so it seems that there is a history leak. As I am a novice in PyTorch, I don’t really understand what is actually wrong. Please, give an advice what part of the code is wrong.
class EncoderBlock(nn.Module):
def __init__(self, batch_size, d_model, conv_num, n_heads, kern_sz, p=0.1):
super(EncoderBlock, self).__init__()
self.convs = nn.ModuleList([DepthwiseSeparableConv(d_model, d_model, kern_sz)
for _ in range(conv_num)])
self.self_att = SelfAttention(n_heads, d_model)
self.W = nn.Linear(d_model, d_model).to(device)
self.relu = nn.ReLU()
self.dropout = nn.Dropout(p)
def forward(self, x, num_rep=1):
out = self.pos_encoding(x)
for _ in range(num_rep):
out = res = out.transpose(1, 2)
for i, conv in enumerate(self.convs):
out = norm(out)
out = conv(out)
out = self.relu(out)
out = res + out
if (i + 1) % 2 == 0:
out = self.dropout(out)
res = out
res = out = norm(out).transpose(1, 2)
out = self.self_att(out, out)
out = res + out
out = self.dropout(out)
res = out
out = norm(out)
out = self.W(out)
out = self.relu(out)
out = res + out
out = self.dropout(out)
return out
@staticmethod
def pos_encoding(x):
_, max_len, model_dim = x.shape
encoding = np.array([
[pos / np.power(10000, 2 * i / model_dim) for i in range(model_dim)]
if pos != 0 else np.zeros(model_dim) for pos in range(max_len)])
encoding[1:, 0::2] = np.sin(encoding[1:, 0::2])
encoding[1:, 1::2] = np.cos(encoding[1:, 1::2])
return x + torch.from_numpy(encoding).float().to(device)
class ScaledDotProductAttention(nn.Module):
def __init__(self, size, p=0.1):
super(ScaledDotProductAttention, self).__init__()
self.scaling = 1 / (np.sqrt(size))
self.dropout = nn.Dropout(p)
def forward(self, q, k, v, mask=None):
attention = torch.bmm(q, k.transpose(1, 2)) * self.scaling
if mask is not None:
attention.data.masked_fill_(mask, -float('inf'))
attention = F.softmax(attention, dim=2)
return torch.bmm(self.dropout(attention), v)
class MultiHeadAttention(nn.Module):
def __init__(self, n_heads, h_size, k_size, v_size, p=0.1):
super(MultiHeadAttention, self).__init__()
self.n_heads = n_heads
self.h_size = h_size
self.q_proj = nn.Parameter(torch.empty((n_heads, h_size, k_size), device=device))
self.k_v_proj = nn.Parameter(torch.empty((2 * n_heads, h_size, v_size)))
self.attention = ScaledDotProductAttention(k_size, p)
self.out = nn.Linear(n_heads * v_size, h_size).to(device)
self.layer_norm = nn.LayerNorm(h_size).to(device)
self.dropout = nn.Dropout(p)
def repeat_n_heads(self, input):
return input.repeat(self.n_heads, 1, 1).view(self.n_heads, -1, self.h_size)
def forward(self, q, k, v, mask=None):
batch_size = q.size(0)
q_len = q.size(1)
seq_len = k.size(1)
residual = q
q = self.repeat_n_heads(q)
k = self.repeat_n_heads(k)
v = self.repeat_n_heads(v)
k_v = torch.cat([k, v], dim=0)
q = torch.bmm(q, self.q_proj).view(-1, q_len, self.q_proj.size(2))
k, v = torch.split(torch.bmm(k_v, self.k_v_proj), self.n_heads, 0)
k = k.view(-1, seq_len, self.k_v_proj.size(2))
v = k.view(-1, seq_len, self.k_v_proj.size(2))
if mask is not None:
mask = mask.repeat(self.n_heads, 1, 1)
result = self.attention(q, k, v, mask)
result = torch.split(result, batch_size, dim=0)
result = torch.cat(result, dim=-1)
result = self.out(result)
result = self.dropout(result)
return self.layer_norm(result + residual)
class SelfAttention(nn.Module):
def __init__(self, n_heads, input_dim):
super(SelfAttention, self).__init__()
self.d_k = input_dim // n_heads
self.multihead = MultiHeadAttention(n_heads, input_dim, self.d_k, self.d_k)
def forward(self, context, questions, mask=None):
return self.multihead(context, questions, questions, mask)