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

243 lines
8.8 KiB
Python
Raw Normal View History

from typing import Tuple, List
2022-06-27 15:57:56 +00:00
import torch
2022-07-07 21:18:30 +00:00
from torch import nn, LongTensor, FloatTensor, BoolTensor
from .dalle_bart_encoder import GLU, AttentionBase
2022-06-27 15:57:56 +00:00
2022-07-04 20:06:49 +00:00
IMAGE_TOKEN_COUNT = 256
BLANK_TOKEN = 6965
2022-06-27 15:57:56 +00:00
class DecoderCrossAttention(AttentionBase):
2022-06-27 15:57:56 +00:00
def forward(
self,
decoder_state: FloatTensor,
encoder_state: FloatTensor,
attention_mask: BoolTensor
) -> FloatTensor:
keys = self.k_proj.forward(encoder_state)
values = self.v_proj.forward(encoder_state)
queries = self.q_proj.forward(decoder_state)
return super().forward(keys, values, queries, attention_mask)
class DecoderSelfAttention(AttentionBase):
2022-07-04 20:06:49 +00:00
def __init__(self, head_count: int, embed_count: int):
2022-07-04 12:05:55 +00:00
super().__init__(head_count, embed_count)
2022-07-04 20:06:49 +00:00
token_indices = torch.arange(IMAGE_TOKEN_COUNT)
2022-07-04 12:05:55 +00:00
if torch.cuda.is_available(): token_indices = token_indices.cuda()
self.token_indices = token_indices
def forward(
self,
2022-06-27 15:57:56 +00:00
decoder_state: FloatTensor,
attention_state: FloatTensor,
2022-07-04 12:05:55 +00:00
token_index: LongTensor
) -> Tuple[FloatTensor, FloatTensor]:
2022-06-29 17:48:12 +00:00
keys = self.k_proj.forward(decoder_state)
values = self.v_proj.forward(decoder_state)
queries = self.q_proj.forward(decoder_state)
2022-07-04 12:05:55 +00:00
attn_mask = self.token_indices < token_index + 1
attn_mask = attn_mask[None][[0] * decoder_state.shape[0]]
2022-07-07 12:21:20 +00:00
attn_state_new = torch.cat([keys, values]).to(attention_state.dtype)
attention_state[:, token_index] = attn_state_new
batch_count = decoder_state.shape[0]
2022-06-30 11:41:31 +00:00
keys = attention_state[:batch_count]
values = attention_state[batch_count:]
2022-07-04 12:05:55 +00:00
decoder_state = super().forward(keys, values, queries, attn_mask)
return decoder_state, attention_state
2022-06-27 15:57:56 +00:00
class DecoderLayer(nn.Module):
def __init__(
self,
2022-06-27 15:57:56 +00:00
head_count: int,
embed_count: int,
glu_embed_count: int
):
super().__init__()
self.pre_self_attn_layer_norm = nn.LayerNorm(embed_count)
2022-07-04 20:06:49 +00:00
self.self_attn = DecoderSelfAttention(head_count, embed_count)
2022-06-27 15:57:56 +00:00
self.self_attn_layer_norm = nn.LayerNorm(embed_count)
self.pre_encoder_attn_layer_norm = nn.LayerNorm(embed_count)
self.encoder_attn = DecoderCrossAttention(head_count, embed_count)
2022-06-27 15:57:56 +00:00
self.encoder_attn_layer_norm = nn.LayerNorm(embed_count)
self.glu = GLU(embed_count, glu_embed_count)
2022-06-27 15:57:56 +00:00
2022-06-29 01:28:36 +00:00
def forward(
self,
2022-06-27 15:57:56 +00:00
decoder_state: FloatTensor,
encoder_state: FloatTensor,
attention_state: FloatTensor,
2022-06-27 15:57:56 +00:00
attention_mask: BoolTensor,
token_index: LongTensor
) -> Tuple[FloatTensor, FloatTensor]:
2022-06-27 15:57:56 +00:00
# Self Attention
residual = decoder_state
decoder_state = self.pre_self_attn_layer_norm.forward(decoder_state)
decoder_state, attention_state = self.self_attn.forward(
2022-06-27 15:57:56 +00:00
decoder_state,
attention_state,
2022-07-04 12:05:55 +00:00
token_index
2022-06-27 15:57:56 +00:00
)
decoder_state = self.self_attn_layer_norm.forward(decoder_state)
decoder_state = residual + decoder_state
# Cross Attention
residual = decoder_state
decoder_state = self.pre_encoder_attn_layer_norm.forward(decoder_state)
decoder_state = self.encoder_attn.forward(
decoder_state,
encoder_state,
attention_mask
)
decoder_state = self.encoder_attn_layer_norm.forward(decoder_state)
decoder_state = residual + decoder_state
# Feed forward
residual = decoder_state
decoder_state = self.glu.forward(decoder_state)
decoder_state = residual + decoder_state
return decoder_state, attention_state
2022-06-27 15:57:56 +00:00
class DalleBartDecoder(nn.Module):
def __init__(
self,
image_vocab_count: int,
2022-06-27 15:57:56 +00:00
embed_count: int,
attention_head_count: int,
glu_embed_count: int,
layer_count: int,
start_token: int
2022-06-27 15:57:56 +00:00
):
super().__init__()
self.layer_count = layer_count
self.embed_count = embed_count
2022-07-08 13:19:28 +00:00
self.image_vocab_count = image_vocab_count
self.embed_tokens = nn.Embedding(image_vocab_count + 1, embed_count)
2022-07-04 20:06:49 +00:00
self.embed_positions = nn.Embedding(IMAGE_TOKEN_COUNT, embed_count)
self.layers: List[DecoderLayer] = nn.ModuleList([
DecoderLayer(
2022-06-27 15:57:56 +00:00
attention_head_count,
embed_count,
glu_embed_count
)
for _ in range(layer_count)
])
self.layernorm_embedding = nn.LayerNorm(embed_count)
self.final_ln = nn.LayerNorm(embed_count)
self.lm_head = nn.Linear(embed_count, image_vocab_count + 1, bias=False)
2022-06-29 01:28:36 +00:00
self.zero_prob = torch.zeros([1])
2022-07-04 20:06:49 +00:00
self.token_indices = torch.arange(IMAGE_TOKEN_COUNT)
2022-06-29 01:28:36 +00:00
self.start_token = torch.tensor([start_token]).to(torch.long)
if torch.cuda.is_available():
self.zero_prob = self.zero_prob.cuda()
self.token_indices = self.token_indices.cuda()
self.start_token = self.start_token.cuda()
2022-06-27 15:57:56 +00:00
def decode_step(
self,
2022-07-05 21:23:05 +00:00
log2_k: int,
2022-07-05 01:30:27 +00:00
log2_supercondition_factor: int,
attention_mask: BoolTensor,
2022-06-27 15:57:56 +00:00
encoder_state: FloatTensor,
attention_state: FloatTensor,
prev_tokens: LongTensor,
token_index: LongTensor
) -> Tuple[FloatTensor, FloatTensor]:
image_count = encoder_state.shape[0] // 2
token_index_batched = token_index[[0] * image_count * 2]
prev_tokens = prev_tokens[list(range(image_count)) * 2]
2022-07-08 13:19:28 +00:00
prev_tokens = prev_tokens.clamp(0, self.image_vocab_count)
decoder_state = self.embed_tokens.forward(prev_tokens)
decoder_state += self.embed_positions.forward(token_index_batched)
2022-06-27 15:57:56 +00:00
decoder_state = self.layernorm_embedding.forward(decoder_state)
2022-06-27 17:19:03 +00:00
decoder_state = decoder_state[:, None]
for i in range(self.layer_count):
decoder_state, attention_state[i] = self.layers[i].forward(
2022-06-27 15:57:56 +00:00
decoder_state,
encoder_state,
attention_state[i],
2022-06-27 15:57:56 +00:00
attention_mask,
token_index
2022-06-27 15:57:56 +00:00
)
2022-06-27 17:19:03 +00:00
decoder_state = self.final_ln(decoder_state)
logits = self.lm_head(decoder_state)
2022-07-05 03:29:48 +00:00
a = 2 ** log2_supercondition_factor
logits: FloatTensor = (
logits[:image_count, -1] * (1 - a) +
logits[image_count:, -1] * a
)
2022-06-27 15:57:56 +00:00
2022-07-05 21:23:05 +00:00
top_logits, _ = logits.topk(2 ** log2_k, dim=-1)
2022-06-27 15:57:56 +00:00
probs = torch.where(
logits < top_logits[:, [-1]],
2022-06-29 01:28:36 +00:00
self.zero_prob,
torch.exp(logits - top_logits[:, [0]])
2022-06-27 15:57:56 +00:00
)
probs[:, 2 ** 14:] = 0 # vqgan vocab_count is only 2 ** 14
return probs, attention_state
2022-07-04 20:06:49 +00:00
2022-07-04 13:58:00 +00:00
def decode_row(
self,
row_index: int,
2022-07-05 21:23:05 +00:00
log2_k: int,
2022-07-05 01:30:27 +00:00
log2_supercondition_factor: int,
2022-07-04 13:58:00 +00:00
encoder_state: FloatTensor,
2022-07-04 20:06:49 +00:00
attention_mask: BoolTensor,
2022-07-04 13:58:00 +00:00
attention_state: FloatTensor,
image_tokens_sequence: LongTensor
) -> Tuple[FloatTensor, LongTensor]:
for col_index in range(16):
i = 16 * row_index + col_index
probs, attention_state = self.decode_step(
2022-07-05 21:23:05 +00:00
log2_k = log2_k,
2022-07-05 01:30:27 +00:00
log2_supercondition_factor = log2_supercondition_factor,
2022-07-04 13:58:00 +00:00
attention_mask = attention_mask,
encoder_state = encoder_state,
attention_state = attention_state,
prev_tokens = image_tokens_sequence[:, i],
token_index = self.token_indices[[i]]
)
image_tokens_sequence[:, i + 1] = torch.multinomial(probs, 1)[:, 0]
return attention_state, image_tokens_sequence
2022-06-27 15:57:56 +00:00
2022-07-04 20:06:49 +00:00
def decode_initial(
self,
2022-07-04 20:06:49 +00:00
seed: int,
image_count: int,
2022-06-27 15:57:56 +00:00
text_tokens: LongTensor,
encoder_state: FloatTensor
2022-07-04 20:06:49 +00:00
) -> Tuple[FloatTensor, FloatTensor, FloatTensor, LongTensor]:
expanded_indices = [0] * image_count + [1] * image_count
text_tokens = text_tokens[expanded_indices]
encoder_state = encoder_state[expanded_indices]
attention_mask = text_tokens.not_equal(1)
2022-06-27 15:57:56 +00:00
attention_state_shape = (
self.layer_count,
image_count * 4,
2022-07-04 20:06:49 +00:00
IMAGE_TOKEN_COUNT,
self.embed_count
)
attention_state = torch.zeros(attention_state_shape)
image_tokens_sequence = torch.full(
2022-07-04 20:06:49 +00:00
(image_count, IMAGE_TOKEN_COUNT + 1),
BLANK_TOKEN,
dtype=torch.long
)
if torch.cuda.is_available():
attention_state = attention_state.cuda()
image_tokens_sequence = image_tokens_sequence.cuda()
2022-07-04 13:58:00 +00:00
image_tokens_sequence[:, 0] = self.start_token[0]
2022-06-27 15:57:56 +00:00
2022-07-04 20:06:49 +00:00
if seed > 0: torch.manual_seed(seed)
return encoder_state, attention_mask, attention_state, image_tokens_sequence