When using the torch.nn.modules.transformer.Transformer module/object, the first layer is the encoder.layers.0.self_attn layer that is a MultiheadAttention layer, i.e.
from torch.nn.modules.transformer import Transformer
bumblebee = Transformer()
bumblee.parameters
[out]:
<bound method Module.parameters of Transformer(
(encoder): TransformerEncoder(
(layers): ModuleList(
(0): TransformerEncoderLayer(
(self_attn): MultiheadAttention(
(out_proj): Linear(in_features=512, out_features=512, bias=True)
)
(linear1): Linear(in_features=512, out_features=2048, bias=True)
(dropout): Dropout(p=0.1, inplace=False)
(linear2): Linear(in_features=2048, out_features=512, bias=True)
(norm1): LayerNorm((512,), eps=1e-05, elementwise_affine=True)
(norm2): LayerNorm((512,), eps=1e-05, elementwise_affine=True)
(dropout1): Dropout(p=0.1, inplace=False)
(dropout2): Dropout(p=0.1, inplace=False)
)
And if we print out the size of the layer, we see:
for name in bumblebee.encoder.state_dict():
print(name, '\t', bumblebee.encoder.state_dict()[name].shape)
[out]:
layers.0.self_attn.in_proj_weight torch.Size([1536, 512])
layers.0.self_attn.in_proj_bias torch.Size([1536])
layers.0.self_attn.out_proj.weight torch.Size([512, 512])
layers.0.self_attn.out_proj.bias torch.Size([512])
layers.0.linear1.weight torch.Size([2048, 512])
layers.0.linear1.bias torch.Size([2048])
layers.0.linear2.weight torch.Size([512, 2048])
layers.0.linear2.bias torch.Size([512])
layers.0.norm1.weight torch.Size([512])
layers.0.norm1.bias torch.Size([512])
layers.0.norm2.weight torch.Size([512])
layers.0.norm2.bias torch.Size([512])
It seems like 1536 is 512 * 3 and somehow the layers.0.self_attn.in_proj_weight parameter might be storing all three QKV tensors in the transformer architecture in one matrix.
From https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/activation.py#L649
class MultiheadAttention(Module):
def __init__(self, embed_dim, num_heads, dropout=0., bias=True, add_bias_kv=False, add_zero_attn=False, kdim=None, vdim=None):
super(MultiheadAttention, self).__init__()
self.embed_dim = embed_dim
self.kdim = kdim if kdim is not None else embed_dim
self.vdim = vdim if vdim is not None else embed_dim
self._qkv_same_embed_dim = self.kdim == embed_dim and self.vdim == embed_dim
self.num_heads = num_heads
self.dropout = dropout
self.head_dim = embed_dim // num_heads
assert self.head_dim * num_heads == self.embed_dim, "embed_dim must be divisible by num_heads"
if self._qkv_same_embed_dim is False:
self.q_proj_weight = Parameter(torch.Tensor(embed_dim, embed_dim))
self.k_proj_weight = Parameter(torch.Tensor(embed_dim, self.kdim))
self.v_proj_weight = Parameter(torch.Tensor(embed_dim, self.vdim))
else:
self.in_proj_weight = Parameter(torch.empty(3 * embed_dim, embed_dim))
And the note in the docstring of the MultiheadAttention says:
Note: if kdim and vdim are None, they will be set to embed_dim such that query, key, and value have the same number of features.
Is that correct?
From the nn.Transformer definition with the default values, EncoderLayer is instantiated with d_model=512, nhead=8.
The MultiheadAttention is instantiated with d_model, nhead equal to those values and k_dim, v_dim are left to the default value of None.
If they are None, self._qkv_same_embed_dim at this line evaluates to True. When that happens, as you correctly pointed out self.in_proj_weight is defined as a Tensor of shape (3 x embed_dim, embed_dim).
In short: yes, that's correct.