min-dalle-test/min_dalle/models/dalle_bart_encoder.py

150 lines
5.5 KiB
Python
Raw Normal View History

2022-06-27 15:57:56 +00:00
from typing import List
import torch
from torch import nn, BoolTensor, FloatTensor, LongTensor
2022-06-29 01:28:36 +00:00
torch.set_grad_enabled(False)
2022-06-27 15:57:56 +00:00
2022-06-27 18:34:10 +00:00
class GLU(nn.Module):
2022-06-27 15:57:56 +00:00
def __init__(self, count_in_out, count_middle):
super().__init__()
self.gelu = nn.GELU()
self.ln0 = nn.LayerNorm(count_in_out)
self.ln1 = nn.LayerNorm(count_middle)
self.fc0 = nn.Linear(count_in_out, count_middle, bias=False)
self.fc1 = nn.Linear(count_in_out, count_middle, bias=False)
self.fc2 = nn.Linear(count_middle, count_in_out, bias=False)
def forward(self, z: FloatTensor) -> FloatTensor:
z = self.ln0.forward(z)
w = self.fc0.forward(z)
w = self.gelu.forward(w)
v = self.fc1.forward(z)
z = self.ln1.forward(w * v)
z = self.fc2.forward(z)
return z
2022-06-27 18:34:10 +00:00
class AttentionBase(nn.Module):
2022-06-27 15:57:56 +00:00
def __init__(self, head_count: int, embed_count: int):
super().__init__()
self.head_count = head_count
self.embed_count = embed_count
2022-06-27 17:19:03 +00:00
self.k_proj = nn.Linear(embed_count, embed_count, bias=False)
self.v_proj = nn.Linear(embed_count, embed_count, bias=False)
self.q_proj = nn.Linear(embed_count, embed_count, bias=False)
self.out_proj = nn.Linear(embed_count, embed_count, bias=False)
2022-06-29 01:28:36 +00:00
self.one = torch.ones((1, 1))
if torch.cuda.is_available(): self.one = self.one.cuda()
2022-06-27 15:57:56 +00:00
def forward(
self,
2022-06-27 15:57:56 +00:00
keys: FloatTensor,
values: FloatTensor,
queries: FloatTensor,
attention_mask: BoolTensor
) -> FloatTensor:
2022-06-29 17:48:12 +00:00
keys = keys.reshape(keys.shape[:2] + (self.head_count, -1))
values = values.reshape(values.shape[:2] + (self.head_count, -1))
queries = queries.reshape(queries.shape[:2] + (self.head_count, -1))
queries /= queries.shape[-1] ** 0.5
2022-06-27 15:57:56 +00:00
attention_bias = torch.where(
attention_mask,
2022-06-29 01:28:36 +00:00
self.one * 0,
self.one * (-torch.inf),
2022-06-27 15:57:56 +00:00
)
attention_weights: FloatTensor = torch.einsum(
2022-06-27 17:19:03 +00:00
'bqhc,bkhc->bhqk',
queries,
2022-06-27 15:57:56 +00:00
keys
)
2022-06-27 17:19:03 +00:00
attention_weights += attention_bias[:, None, None, :]
attention_weights = torch.softmax(attention_weights, -1)
attention_output: FloatTensor = torch.einsum(
"bhqk,bkhc->bqhc",
2022-06-27 15:57:56 +00:00
attention_weights,
values
)
2022-06-27 17:19:03 +00:00
shape = attention_output.shape[:2] + (self.embed_count,)
attention_output = attention_output.reshape(shape)
attention_output = self.out_proj.forward(attention_output)
return attention_output
2022-06-27 15:57:56 +00:00
class EncoderSelfAttention(AttentionBase):
2022-06-27 15:57:56 +00:00
def forward(
self,
encoder_state: FloatTensor,
attention_mask: BoolTensor
) -> FloatTensor:
2022-06-29 17:48:12 +00:00
keys = self.k_proj.forward(encoder_state)
values = self.v_proj.forward(encoder_state)
queries = self.q_proj.forward(encoder_state)
2022-06-27 15:57:56 +00:00
return super().forward(keys, values, queries, attention_mask)
class EncoderLayer(nn.Module):
2022-06-27 15:57:56 +00:00
def __init__(self, embed_count: int, head_count: int, glu_embed_count: int):
super().__init__()
self.pre_self_attn_layer_norm = nn.LayerNorm(embed_count)
self.self_attn = EncoderSelfAttention(head_count, embed_count)
2022-06-27 15:57:56 +00:00
self.self_attn_layer_norm = nn.LayerNorm(embed_count)
self.glu = GLU(embed_count, glu_embed_count)
2022-06-27 15:57:56 +00:00
def forward(
self,
encoder_state: FloatTensor,
attention_mask: BoolTensor
) -> FloatTensor:
residual = encoder_state
encoder_state = self.pre_self_attn_layer_norm.forward(encoder_state)
encoder_state = self.self_attn.forward(encoder_state, attention_mask)
encoder_state = self.self_attn_layer_norm.forward(encoder_state)
encoder_state = residual + encoder_state
residual = encoder_state
encoder_state = self.glu.forward(encoder_state)
encoder_state = residual + encoder_state
return encoder_state
class DalleBartEncoder(nn.Module):
def __init__(
self,
2022-06-27 15:57:56 +00:00
layer_count: int,
embed_count: int,
attention_head_count: int,
text_vocab_count: int,
text_token_count: int,
glu_embed_count: int
):
super().__init__()
self.embed_tokens = nn.Embedding(text_vocab_count, embed_count)
self.embed_positions = nn.Embedding(text_token_count, embed_count)
self.layers: List[EncoderLayer] = nn.ModuleList([
EncoderLayer(
2022-06-27 15:57:56 +00:00
embed_count = embed_count,
head_count = attention_head_count,
glu_embed_count = glu_embed_count
)
for _ in range(layer_count)
])
self.layernorm_embedding = nn.LayerNorm(embed_count)
self.final_ln = nn.LayerNorm(embed_count)
2022-06-29 01:28:36 +00:00
self.token_indices = torch.arange(text_token_count).to(torch.long)
if torch.cuda.is_available():
self.token_indices = self.token_indices.cuda()
2022-06-27 15:57:56 +00:00
def forward(self, text_tokens: LongTensor) -> FloatTensor:
attention_mask = text_tokens.not_equal(1)
2022-06-29 01:28:36 +00:00
batch_count = text_tokens.shape[0]
pose_tokens = torch.stack([self.token_indices] * batch_count)
2022-06-27 15:57:56 +00:00
encoder_state = (
self.embed_tokens.forward(text_tokens) +
self.embed_positions.forward(pose_tokens)
)
encoder_state = self.layernorm_embedding.forward(encoder_state)
for layer in self.layers:
encoder_state = layer.forward(encoder_state, attention_mask)
encoder_state = self.final_ln.forward(encoder_state)
return encoder_state