DL-Art-School/codes/models/lucidrains/x_transformers.py

1262 lines
42 KiB
Python
Raw Normal View History

import functools
2022-04-06 06:21:58 +00:00
import math
import torch
from torch import nn, einsum
import torch.nn.functional as F
from functools import partial
from inspect import isfunction
from collections import namedtuple
from einops import rearrange, repeat, reduce
from einops.layers.torch import Rearrange
from entmax import entmax15
from x_transformers.autoregressive_wrapper import AutoregressiveWrapper
# constants
from utils.util import checkpoint
2022-04-06 06:21:58 +00:00
DEFAULT_DIM_HEAD = 64
Intermediates = namedtuple('Intermediates', [
'pre_softmax_attn',
'post_softmax_attn'
])
LayerIntermediates = namedtuple('Intermediates', [
'hiddens',
'attn_intermediates'
])
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
# helpers
def exists(val):
return val is not None
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
def default(val, d):
if exists(val):
return val
return d() if isfunction(d) else d
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
def cast_tuple(val, depth):
return val if isinstance(val, tuple) else (val,) * depth
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
class always():
def __init__(self, val):
self.val = val
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
def __call__(self, *args, **kwargs):
return self.val
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
class not_equals():
def __init__(self, val):
self.val = val
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
def __call__(self, x, *args, **kwargs):
return x != self.val
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
class equals():
def __init__(self, val):
self.val = val
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
def __call__(self, x, *args, **kwargs):
return x == self.val
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
def max_neg_value(tensor):
return -torch.finfo(tensor.dtype).max
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
def l2norm(t):
2022-04-08 06:08:03 +00:00
return F.normalize(t, p=2, dim=-1)
2022-04-06 06:21:58 +00:00
# init helpers
def init_zero_(layer):
nn.init.constant_(layer.weight, 0.)
if exists(layer.bias):
nn.init.constant_(layer.bias, 0.)
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
# keyword argument helpers
def pick_and_pop(keys, d):
values = list(map(lambda key: d.pop(key), keys))
return dict(zip(keys, values))
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
def group_dict_by_key(cond, d):
2022-04-08 06:08:03 +00:00
return_val = [dict(), dict()]
2022-04-06 06:21:58 +00:00
for key in d.keys():
match = bool(cond(key))
ind = int(not match)
return_val[ind][key] = d[key]
return (*return_val,)
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
def string_begins_with(prefix, str):
return str.startswith(prefix)
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
def group_by_key_prefix(prefix, d):
return group_dict_by_key(partial(string_begins_with, prefix), d)
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
def groupby_prefix_and_trim(prefix, d):
kwargs_with_prefix, kwargs = group_dict_by_key(partial(string_begins_with, prefix), d)
kwargs_without_prefix = dict(map(lambda x: (x[0][len(prefix):], x[1]), tuple(kwargs_with_prefix.items())))
return kwargs_without_prefix, kwargs
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
# activations
class ReluSquared(nn.Module):
def forward(self, x):
return F.relu(x) ** 2
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
# positional embeddings
class AbsolutePositionalEmbedding(nn.Module):
def __init__(self, dim, max_seq_len):
super().__init__()
self.scale = dim ** -0.5
self.emb = nn.Embedding(max_seq_len, dim)
def forward(self, x):
2022-04-08 06:08:03 +00:00
n = torch.arange(x.shape[1], device=x.device)
2022-04-06 06:21:58 +00:00
pos_emb = self.emb(n)
pos_emb = rearrange(pos_emb, 'n d -> () n d')
return pos_emb * self.scale
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
class FixedPositionalEmbedding(nn.Module):
def __init__(self, dim):
super().__init__()
inv_freq = 1. / (10000 ** (torch.arange(0, dim, 2).float() / dim))
self.register_buffer('inv_freq', inv_freq)
2022-04-08 06:08:03 +00:00
def forward(self, x, seq_dim=1, offset=0):
t = torch.arange(x.shape[seq_dim], device=x.device).type_as(self.inv_freq) + offset
2022-04-06 06:21:58 +00:00
sinusoid_inp = torch.einsum('i , j -> i j', t, self.inv_freq)
emb = torch.cat((sinusoid_inp.sin(), sinusoid_inp.cos()), dim=-1)
return rearrange(emb, 'n d -> () n d')
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
class RelativePositionBias(nn.Module):
2022-04-08 06:08:03 +00:00
def __init__(self, scale, causal=False, num_buckets=32, max_distance=128, heads=8):
2022-04-06 06:21:58 +00:00
super().__init__()
self.scale = scale
self.causal = causal
self.num_buckets = num_buckets
self.max_distance = max_distance
self.relative_attention_bias = nn.Embedding(num_buckets, heads)
@staticmethod
2022-04-08 06:08:03 +00:00
def _relative_position_bucket(relative_position, causal=True, num_buckets=32, max_distance=128):
2022-04-06 06:21:58 +00:00
ret = 0
n = -relative_position
if not causal:
num_buckets //= 2
ret += (n < 0).long() * num_buckets
n = torch.abs(n)
else:
n = torch.max(n, torch.zeros_like(n))
max_exact = num_buckets // 2
is_small = n < max_exact
val_if_large = max_exact + (
2022-04-08 06:08:03 +00:00
torch.log(n.float() / max_exact) / math.log(max_distance / max_exact) * (num_buckets - max_exact)
2022-04-06 06:21:58 +00:00
).long()
val_if_large = torch.min(val_if_large, torch.full_like(val_if_large, num_buckets - 1))
ret += torch.where(is_small, n, val_if_large)
return ret
def forward(self, qk_dots):
i, j, device = *qk_dots.shape[-2:], qk_dots.device
2022-04-08 06:08:03 +00:00
q_pos = torch.arange(i, dtype=torch.long, device=device)
k_pos = torch.arange(j, dtype=torch.long, device=device)
2022-04-06 06:21:58 +00:00
rel_pos = k_pos[None, :] - q_pos[:, None]
2022-04-08 06:08:03 +00:00
rp_bucket = self._relative_position_bucket(rel_pos, causal=self.causal, num_buckets=self.num_buckets,
max_distance=self.max_distance)
2022-04-06 06:21:58 +00:00
values = self.relative_attention_bias(rp_bucket)
bias = rearrange(values, 'i j h -> () h i j')
return qk_dots + (bias * self.scale)
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
class AlibiPositionalBias(nn.Module):
def __init__(self, heads, **kwargs):
super().__init__()
self.heads = heads
slopes = torch.Tensor(self._get_slopes(heads))
slopes = rearrange(slopes, 'h -> () h () ()')
2022-04-08 06:08:03 +00:00
self.register_buffer('slopes', slopes, persistent=False)
self.register_buffer('bias', None, persistent=False)
2022-04-06 06:21:58 +00:00
@staticmethod
def _get_slopes(heads):
def get_slopes_power_of_2(n):
2022-04-08 06:08:03 +00:00
start = (2 ** (-2 ** -(math.log2(n) - 3)))
2022-04-06 06:21:58 +00:00
ratio = start
2022-04-08 06:08:03 +00:00
return [start * ratio ** i for i in range(n)]
2022-04-06 06:21:58 +00:00
if math.log2(heads).is_integer():
return get_slopes_power_of_2(heads)
closest_power_of_2 = 2 ** math.floor(math.log2(heads))
2022-04-08 06:08:03 +00:00
return get_slopes_power_of_2(closest_power_of_2) + get_slopes_power_of_2(2 * closest_power_of_2)[0::2][
:heads - closest_power_of_2]
2022-04-06 06:21:58 +00:00
def forward(self, qk_dots):
h, i, j, device = *qk_dots.shape[-3:], qk_dots.device
if exists(self.bias) and self.bias.shape[-1] >= j:
return qk_dots + self.bias[..., :j]
2022-04-08 06:08:03 +00:00
bias = torch.arange(j, device=device)
2022-04-06 06:21:58 +00:00
bias = rearrange(bias, 'j -> () () () j')
bias = bias * self.slopes
num_heads_unalibied = h - bias.shape[1]
bias = F.pad(bias, (0, 0, 0, 0, 0, num_heads_unalibied))
2022-04-08 06:08:03 +00:00
self.register_buffer('bias', bias, persistent=False)
2022-04-06 06:21:58 +00:00
return qk_dots + self.bias
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
class LearnedAlibiPositionalBias(AlibiPositionalBias):
2022-04-08 06:08:03 +00:00
def __init__(self, heads, bidirectional=False):
2022-04-06 06:21:58 +00:00
super().__init__(heads)
los_slopes = torch.log(self.slopes)
self.learned_logslopes = nn.Parameter(los_slopes)
self.bidirectional = bidirectional
if self.bidirectional:
self.learned_logslopes_future = nn.Parameter(los_slopes)
def forward(self, qk_dots):
h, i, j, device = *qk_dots.shape[-3:], qk_dots.device
def get_slopes(param):
return F.pad(param.exp(), (0, 0, 0, 0, 0, h - param.shape[1]))
if exists(self.bias) and self.bias.shape[-1] >= j:
bias = self.bias[..., :i, :j]
else:
2022-04-08 06:08:03 +00:00
i_arange = torch.arange(i, device=device)
j_arange = torch.arange(j, device=device)
2022-04-06 06:21:58 +00:00
bias = rearrange(j_arange, 'j -> 1 1 1 j') - rearrange(i_arange, 'i -> 1 1 i 1')
2022-04-08 06:08:03 +00:00
self.register_buffer('bias', bias, persistent=False)
2022-04-06 06:21:58 +00:00
if self.bidirectional:
past_slopes = get_slopes(self.learned_logslopes)
future_slopes = get_slopes(self.learned_logslopes_future)
bias = torch.tril(bias * past_slopes) + torch.triu(bias * future_slopes)
else:
slopes = get_slopes(self.learned_logslopes)
bias = bias * slopes
return qk_dots + bias
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
class RotaryEmbedding(nn.Module):
def __init__(self, dim):
super().__init__()
inv_freq = 1. / (10000 ** (torch.arange(0, dim, 2).float() / dim))
self.register_buffer('inv_freq', inv_freq)
def forward(self, max_seq_len, device):
2022-04-08 06:08:03 +00:00
t = torch.arange(max_seq_len, device=device).type_as(self.inv_freq)
2022-04-06 06:21:58 +00:00
freqs = torch.einsum('i , j -> i j', t, self.inv_freq)
emb = torch.cat((freqs, freqs), dim=-1)
return rearrange(emb, 'n d -> () () n d')
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
def rotate_half(x):
2022-04-08 06:08:03 +00:00
x = rearrange(x, '... (j d) -> ... j d', j=2)
x1, x2 = x.unbind(dim=-2)
return torch.cat((-x2, x1), dim=-1)
2022-04-06 06:21:58 +00:00
def apply_rotary_pos_emb(t, freqs):
seq_len = t.shape[-2]
freqs = freqs[:, :, -seq_len:]
return (t * freqs.cos()) + (rotate_half(t) * freqs.sin())
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
# norms
class Scale(nn.Module):
def __init__(self, value, fn):
super().__init__()
self.value = value
self.fn = fn
def forward(self, x, **kwargs):
out = self.fn(x, **kwargs)
scale_fn = lambda t: t * self.value
if not isinstance(out, tuple):
return scale_fn(out)
return (scale_fn(out[0]), *out[1:])
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
class Rezero(nn.Module):
def __init__(self, fn):
super().__init__()
self.fn = fn
self.g = nn.Parameter(torch.zeros(1))
def forward(self, x, **kwargs):
out = self.fn(x, **kwargs)
rezero_fn = lambda t: t * self.g
if not isinstance(out, tuple):
return rezero_fn(out)
return (rezero_fn(out[0]), *out[1:])
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
class ScaleNorm(nn.Module):
2022-04-08 06:08:03 +00:00
def __init__(self, dim, eps=1e-5):
2022-04-06 06:21:58 +00:00
super().__init__()
self.scale = dim ** -0.5
self.eps = eps
self.g = nn.Parameter(torch.ones(1))
def forward(self, x):
2022-04-08 06:08:03 +00:00
norm = torch.norm(x, dim=-1, keepdim=True) * self.scale
return x / norm.clamp(min=self.eps) * self.g
2022-04-06 06:21:58 +00:00
class RMSNorm(nn.Module):
2022-04-08 06:08:03 +00:00
def __init__(self, dim, eps=1e-8):
2022-04-06 06:21:58 +00:00
super().__init__()
self.scale = dim ** -0.5
self.eps = eps
self.g = nn.Parameter(torch.ones(dim))
def forward(self, x):
2022-04-08 06:08:03 +00:00
norm = torch.norm(x, dim=-1, keepdim=True) * self.scale
return x / norm.clamp(min=self.eps) * self.g
2022-04-06 06:21:58 +00:00
2022-04-08 02:10:57 +00:00
class RMSScaleShiftNorm(nn.Module):
2022-04-08 06:08:03 +00:00
def __init__(self, dim, eps=1e-8):
2022-04-08 02:10:57 +00:00
super().__init__()
self.scale = dim ** -0.5
self.eps = eps
self.g = nn.Parameter(torch.ones(dim))
2022-04-08 06:08:03 +00:00
self.scale_shift_process = nn.Linear(dim * 2, dim * 2)
2022-04-08 02:10:57 +00:00
def forward(self, x, norm_scale_shift_inp):
2022-04-08 06:08:03 +00:00
norm = torch.norm(x, dim=-1, keepdim=True) * self.scale
norm = x / norm.clamp(min=self.eps) * self.g
2022-04-08 02:10:57 +00:00
ss_emb = self.scale_shift_process(norm_scale_shift_inp)
scale, shift = torch.chunk(ss_emb, 2, dim=1)
h = norm * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
return h
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
# residual and residual gates
class Residual(nn.Module):
2022-04-08 06:08:03 +00:00
def __init__(self, dim, scale_residual=False):
2022-04-06 06:21:58 +00:00
super().__init__()
self.residual_scale = nn.Parameter(torch.ones(dim)) if scale_residual else None
def forward(self, x, residual):
if exists(self.residual_scale):
residual = residual * self.residual_scale
return x + residual
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
class GRUGating(nn.Module):
2022-04-08 06:08:03 +00:00
def __init__(self, dim, scale_residual=False):
2022-04-06 06:21:58 +00:00
super().__init__()
self.gru = nn.GRUCell(dim, dim)
self.residual_scale = nn.Parameter(torch.ones(dim)) if scale_residual else None
def forward(self, x, residual):
if exists(self.residual_scale):
residual = residual * self.residual_scale
gated_output = self.gru(
rearrange(x, 'b n d -> (b n) d'),
rearrange(residual, 'b n d -> (b n) d')
)
return gated_output.reshape_as(x)
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
# token shifting
2022-04-08 06:08:03 +00:00
def shift(t, amount, mask=None):
2022-04-06 06:21:58 +00:00
if amount == 0:
return t
if exists(mask):
t = t.masked_fill(~mask[..., None], 0.)
2022-04-08 06:08:03 +00:00
return F.pad(t, (0, 0, amount, -amount), value=0.)
2022-04-06 06:21:58 +00:00
class ShiftTokens(nn.Module):
def __init__(self, shifts, fn):
super().__init__()
self.fn = fn
self.shifts = tuple(shifts)
def forward(self, x, **kwargs):
mask = kwargs.get('mask', None)
shifts = self.shifts
segments = len(shifts)
feats_per_shift = x.shape[-1] // segments
2022-04-08 06:08:03 +00:00
splitted = x.split(feats_per_shift, dim=-1)
2022-04-06 06:21:58 +00:00
segments_to_shift, rest = splitted[:segments], splitted[segments:]
2022-04-08 06:08:03 +00:00
segments_to_shift = list(map(lambda args: shift(*args, mask=mask), zip(segments_to_shift, shifts)))
x = torch.cat((*segments_to_shift, *rest), dim=-1)
2022-04-06 06:21:58 +00:00
return self.fn(x, **kwargs)
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
# feedforward
class GLU(nn.Module):
def __init__(self, dim_in, dim_out, activation):
super().__init__()
self.act = activation
self.proj = nn.Linear(dim_in, dim_out * 2)
def forward(self, x):
2022-04-08 06:08:03 +00:00
x, gate = self.proj(x).chunk(2, dim=-1)
2022-04-06 06:21:58 +00:00
return x * self.act(gate)
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
class FeedForward(nn.Module):
def __init__(
2022-04-08 06:08:03 +00:00
self,
dim,
dim_out=None,
mult=4,
glu=False,
relu_squared=False,
post_act_ln=False,
dropout=0.,
zero_init_output=False
2022-04-06 06:21:58 +00:00
):
super().__init__()
inner_dim = int(dim * mult)
dim_out = default(dim_out, dim)
activation = ReluSquared() if relu_squared else nn.GELU()
project_in = nn.Sequential(
nn.Linear(dim, inner_dim),
activation
) if not glu else GLU(dim, inner_dim, activation)
self.net = nn.Sequential(
project_in,
nn.LayerNorm(inner_dim) if post_act_ln else nn.Identity(),
nn.Dropout(dropout),
nn.Linear(inner_dim, dim_out)
)
# init last linear layer to 0
if zero_init_output:
init_zero_(self.net[-1])
def forward(self, x):
return self.net(x)
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
# attention.
class Attention(nn.Module):
def __init__(
2022-04-08 06:08:03 +00:00
self,
dim,
dim_head=DEFAULT_DIM_HEAD,
heads=8,
causal=False,
talking_heads=False,
head_scale=False,
collab_heads=False,
collab_compression=.3,
sparse_topk=None,
use_entmax15=False,
num_mem_kv=0,
dropout=0.,
on_attn=False,
gate_values=False,
zero_init_output=False,
max_attend_past=None,
qk_norm=False,
scale_init_value=None,
rel_pos_bias=False,
rel_pos_num_buckets=32,
rel_pos_max_distance=128,
2022-04-06 06:21:58 +00:00
):
super().__init__()
self.scale = dim_head ** -0.5
self.heads = heads
self.causal = causal
self.max_attend_past = max_attend_past
qk_dim = v_dim = dim_head * heads
# collaborative heads
self.collab_heads = collab_heads
if self.collab_heads:
qk_dim = int(collab_compression * qk_dim)
self.collab_mixing = nn.Parameter(torch.randn(heads, qk_dim))
2022-04-08 06:08:03 +00:00
self.to_q = nn.Linear(dim, qk_dim, bias=False)
self.to_k = nn.Linear(dim, qk_dim, bias=False)
self.to_v = nn.Linear(dim, v_dim, bias=False)
2022-04-06 06:21:58 +00:00
self.dropout = nn.Dropout(dropout)
# add GLU gating for aggregated values, from alphafold2
self.to_v_gate = None
if gate_values:
self.to_v_gate = nn.Linear(dim, v_dim)
nn.init.constant_(self.to_v_gate.weight, 0)
nn.init.constant_(self.to_v_gate.bias, 1)
# cosine sim attention
self.qk_norm = qk_norm
if qk_norm:
2022-04-08 06:08:03 +00:00
scale_init_value = default(scale_init_value,
-3) # if not provided, initialize as though it were sequence length of 1024
2022-04-06 06:21:58 +00:00
self.scale = nn.Parameter(torch.ones(1, heads, 1, 1) * scale_init_value)
# talking heads
self.talking_heads = talking_heads
if talking_heads:
self.pre_softmax_proj = nn.Parameter(torch.randn(heads, heads))
self.post_softmax_proj = nn.Parameter(torch.randn(heads, heads))
# head scaling
self.head_scale = head_scale
if head_scale:
self.head_scale_params = nn.Parameter(torch.ones(1, heads, 1, 1))
# explicit topk sparse attention
self.sparse_topk = sparse_topk
# entmax
self.attn_fn = entmax15 if use_entmax15 else F.softmax
# add memory key / values
self.num_mem_kv = num_mem_kv
if num_mem_kv > 0:
self.mem_k = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head))
self.mem_v = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head))
# attention on attention
self.attn_on_attn = on_attn
self.to_out = nn.Sequential(nn.Linear(v_dim, dim * 2), nn.GLU()) if on_attn else nn.Linear(v_dim, dim)
self.rel_pos_bias = rel_pos_bias
if rel_pos_bias:
assert rel_pos_num_buckets <= rel_pos_max_distance, 'number of relative position buckets must be less than the relative position max distance'
2022-04-08 06:08:03 +00:00
self.rel_pos = RelativePositionBias(scale=dim_head ** 0.5, causal=causal, heads=heads,
num_buckets=rel_pos_num_buckets, max_distance=rel_pos_max_distance)
2022-04-06 06:21:58 +00:00
# init output projection 0
if zero_init_output:
init_zero_(self.to_out)
def forward(
2022-04-08 06:08:03 +00:00
self,
x,
context=None,
mask=None,
context_mask=None,
attn_mask=None,
sinusoidal_emb=None,
rotary_pos_emb=None,
prev_attn=None,
mem=None
2022-04-06 06:21:58 +00:00
):
2022-04-08 06:08:03 +00:00
b, n, _, h, talking_heads, collab_heads, head_scale, scale, device, has_context = *x.shape, self.heads, self.talking_heads, self.collab_heads, self.head_scale, self.scale, x.device, exists(
context)
2022-04-06 06:21:58 +00:00
kv_input = default(context, x)
q_input = x
k_input = kv_input
v_input = kv_input
if exists(mem):
2022-04-08 06:08:03 +00:00
k_input = torch.cat((mem, k_input), dim=-2)
v_input = torch.cat((mem, v_input), dim=-2)
2022-04-06 06:21:58 +00:00
if exists(sinusoidal_emb):
# in shortformer, the query would start at a position offset depending on the past cached memory
offset = k_input.shape[-2] - q_input.shape[-2]
2022-04-08 06:08:03 +00:00
q_input = q_input + sinusoidal_emb(q_input, offset=offset)
2022-04-06 06:21:58 +00:00
k_input = k_input + sinusoidal_emb(k_input)
q = self.to_q(q_input)
k = self.to_k(k_input)
v = self.to_v(v_input)
if not collab_heads:
2022-04-08 06:08:03 +00:00
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h=h), (q, k, v))
2022-04-06 06:21:58 +00:00
else:
q = einsum('b i d, h d -> b h i d', q, self.collab_mixing)
k = rearrange(k, 'b n d -> b () n d')
2022-04-08 06:08:03 +00:00
v = rearrange(v, 'b n (h d) -> b h n d', h=h)
2022-04-06 06:21:58 +00:00
if exists(rotary_pos_emb) and not has_context:
l = rotary_pos_emb.shape[-1]
(ql, qr), (kl, kr), (vl, vr) = map(lambda t: (t[..., :l], t[..., l:]), (q, k, v))
ql, kl, vl = map(lambda t: apply_rotary_pos_emb(t, rotary_pos_emb), (ql, kl, vl))
2022-04-08 06:08:03 +00:00
q, k, v = map(lambda t: torch.cat(t, dim=-1), ((ql, qr), (kl, kr), (vl, vr)))
2022-04-06 06:21:58 +00:00
input_mask = None
if any(map(exists, (mask, context_mask))):
2022-04-08 06:08:03 +00:00
q_mask = default(mask, lambda: torch.ones((b, n), device=device).bool())
2022-04-06 06:21:58 +00:00
k_mask = q_mask if not exists(context) else context_mask
2022-04-08 06:08:03 +00:00
k_mask = default(k_mask, lambda: torch.ones((b, k.shape[-2]), device=device).bool())
2022-04-06 06:21:58 +00:00
q_mask = rearrange(q_mask, 'b i -> b () i ()')
k_mask = rearrange(k_mask, 'b j -> b () () j')
input_mask = q_mask * k_mask
if self.num_mem_kv > 0:
2022-04-08 06:08:03 +00:00
mem_k, mem_v = map(lambda t: repeat(t, 'h n d -> b h n d', b=b), (self.mem_k, self.mem_v))
k = torch.cat((mem_k, k), dim=-2)
v = torch.cat((mem_v, v), dim=-2)
2022-04-06 06:21:58 +00:00
if exists(input_mask):
2022-04-08 06:08:03 +00:00
input_mask = F.pad(input_mask, (self.num_mem_kv, 0), value=True)
2022-04-06 06:21:58 +00:00
if collab_heads:
k = k.expand(-1, h, -1, -1)
if self.qk_norm:
q, k = map(l2norm, (q, k))
2022-04-08 06:08:03 +00:00
scale = 1 / (self.scale.exp().clamp(min=1e-2))
2022-04-06 06:21:58 +00:00
dots = einsum('b h i d, b h j d -> b h i j', q, k) * scale
mask_value = max_neg_value(dots)
if exists(prev_attn):
dots = dots + prev_attn
pre_softmax_attn = dots.clone()
if talking_heads:
dots = einsum('b h i j, h k -> b k i j', dots, self.pre_softmax_proj).contiguous()
if self.rel_pos_bias:
dots = self.rel_pos(dots)
2022-04-06 06:21:58 +00:00
if exists(input_mask):
dots.masked_fill_(~input_mask, mask_value)
del input_mask
if exists(attn_mask):
assert 2 <= attn_mask.ndim <= 4, 'attention mask must have greater than 2 dimensions but less than or equal to 4'
if attn_mask.ndim == 2:
attn_mask = rearrange(attn_mask, 'i j -> () () i j')
elif attn_mask.ndim == 3:
attn_mask = rearrange(attn_mask, 'h i j -> () h i j')
dots.masked_fill_(~attn_mask, mask_value)
if exists(self.max_attend_past):
i, j = dots.shape[-2:]
2022-04-08 06:08:03 +00:00
range_q = torch.arange(j - i, j, device=device)
range_k = torch.arange(j, device=device)
2022-04-06 06:21:58 +00:00
dist = rearrange(range_q, 'i -> () () i ()') - rearrange(range_k, 'j -> () () () j')
mask = dist > self.max_attend_past
dots.masked_fill_(mask, mask_value)
del mask
if self.causal:
i, j = dots.shape[-2:]
2022-04-08 06:08:03 +00:00
r = torch.arange(i, device=device)
2022-04-06 06:21:58 +00:00
mask = rearrange(r, 'i -> () () i ()') < rearrange(r, 'j -> () () () j')
2022-04-08 06:08:03 +00:00
mask = F.pad(mask, (j - i, 0), value=False)
2022-04-06 06:21:58 +00:00
dots.masked_fill_(mask, mask_value)
del mask
if exists(self.sparse_topk) and self.sparse_topk < dots.shape[-1]:
2022-04-08 06:08:03 +00:00
top, _ = dots.topk(self.sparse_topk, dim=-1)
2022-04-06 06:21:58 +00:00
vk = top[..., -1].unsqueeze(-1).expand_as(dots)
mask = dots < vk
dots.masked_fill_(mask, mask_value)
del mask
2022-04-08 06:08:03 +00:00
attn = self.attn_fn(dots, dim=-1)
2022-04-06 06:21:58 +00:00
post_softmax_attn = attn.clone()
attn = self.dropout(attn)
if talking_heads:
attn = einsum('b h i j, h k -> b k i j', attn, self.post_softmax_proj).contiguous()
out = einsum('b h i j, b h j d -> b h i d', attn, v)
if head_scale:
out = out * self.head_scale_params
out = rearrange(out, 'b h n d -> b n (h d)')
if exists(self.to_v_gate):
gates = self.to_v_gate(x)
out = out * gates.sigmoid()
intermediates = Intermediates(
2022-04-08 06:08:03 +00:00
pre_softmax_attn=pre_softmax_attn,
post_softmax_attn=post_softmax_attn
2022-04-06 06:21:58 +00:00
)
return self.to_out(out), intermediates
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
class AttentionLayers(nn.Module):
def __init__(
2022-04-08 06:08:03 +00:00
self,
dim,
depth,
heads=8,
causal=False,
cross_attend=False,
only_cross=False,
use_scalenorm=False,
use_rms_scaleshift_norm=False,
use_rmsnorm=False,
use_rezero=False,
alibi_pos_bias=False,
alibi_num_heads=None,
alibi_learned=False,
position_infused_attn=False,
rotary_pos_emb=False,
rotary_emb_dim=None,
custom_layers=None,
sandwich_coef=None,
par_ratio=None,
residual_attn=False,
cross_residual_attn=False,
macaron=False,
pre_norm=True,
gate_residual=False,
scale_residual=False,
shift_tokens=0,
sandwich_norm=False,
use_qk_norm_attn=False,
qk_norm_attn_seq_len=None,
zero_init_branch_output=False,
**kwargs
2022-04-06 06:21:58 +00:00
):
super().__init__()
ff_kwargs, kwargs = groupby_prefix_and_trim('ff_', kwargs)
attn_kwargs, _ = groupby_prefix_and_trim('attn_', kwargs)
dim_head = attn_kwargs.get('dim_head', DEFAULT_DIM_HEAD)
self.dim = dim
self.depth = depth
self.layers = nn.ModuleList([])
rel_pos_bias = 'rel_pos_bias' in attn_kwargs
2022-04-06 06:21:58 +00:00
self.has_pos_emb = position_infused_attn or rel_pos_bias or rotary_pos_emb
self.pia_pos_emb = FixedPositionalEmbedding(dim) if position_infused_attn else None
rotary_emb_dim = max(default(rotary_emb_dim, dim_head // 2), 32)
self.rotary_pos_emb = RotaryEmbedding(rotary_emb_dim) if rotary_pos_emb else None
2022-04-08 06:08:03 +00:00
assert not (
alibi_pos_bias and rel_pos_bias), 'you can only choose Alibi positional bias or T5 relative positional bias, not both'
2022-04-06 06:21:58 +00:00
if alibi_pos_bias:
2022-04-06 06:21:58 +00:00
alibi_num_heads = default(alibi_num_heads, heads)
assert alibi_num_heads <= heads, 'number of ALiBi heads must be less than the total number of heads'
alibi_pos_klass = LearnedAlibiPositionalBias if alibi_learned or not causal else AlibiPositionalBias
2022-04-08 06:08:03 +00:00
self.rel_pos = alibi_pos_klass(heads=alibi_num_heads, bidirectional=not causal)
2022-04-06 06:21:58 +00:00
else:
self.rel_pos = None
assert not (not pre_norm and sandwich_norm), 'sandwich norm cannot be used when not using prenorm'
self.pre_norm = pre_norm
self.sandwich_norm = sandwich_norm
self.residual_attn = residual_attn
self.cross_residual_attn = cross_residual_attn
self.cross_attend = cross_attend
norm_class = ScaleNorm if use_scalenorm else nn.LayerNorm
norm_class = RMSNorm if use_rmsnorm else norm_class
2022-04-08 02:10:57 +00:00
norm_class = RMSScaleShiftNorm if use_rms_scaleshift_norm else norm_class
2022-04-06 06:21:58 +00:00
norm_fn = partial(norm_class, dim)
norm_fn = nn.Identity if use_rezero else norm_fn
branch_fn = Rezero if use_rezero else None
if cross_attend and not only_cross:
default_block = ('a', 'c', 'f')
elif cross_attend and only_cross:
default_block = ('c', 'f')
else:
default_block = ('a', 'f')
if macaron:
default_block = ('f',) + default_block
# qk normalization
if use_qk_norm_attn:
2022-04-08 06:08:03 +00:00
attn_scale_init_value = -math.log(math.log2(qk_norm_attn_seq_len ** 2 - qk_norm_attn_seq_len)) if exists(
qk_norm_attn_seq_len) else None
2022-04-06 06:21:58 +00:00
attn_kwargs = {**attn_kwargs, 'qk_norm': True, 'scale_init_value': attn_scale_init_value}
# zero init
if zero_init_branch_output:
2022-04-08 06:08:03 +00:00
attn_kwargs = {**attn_kwargs, 'zero_init_output': True}
ff_kwargs = {**ff_kwargs, 'zero_init_output': True}
2022-04-06 06:21:58 +00:00
# calculate layer block order
if exists(custom_layers):
layer_types = custom_layers
elif exists(par_ratio):
par_depth = depth * len(default_block)
assert 1 < par_ratio <= par_depth, 'par ratio out of range'
default_block = tuple(filter(not_equals('f'), default_block))
2022-04-08 06:08:03 +00:00
par_attn = par_depth // par_ratio
2022-04-06 06:21:58 +00:00
depth_cut = par_depth * 2 // 3 # 2 / 3 attention layer cutoff suggested by PAR paper
par_width = (depth_cut + depth_cut // par_attn) // par_attn
assert len(default_block) <= par_width, 'default block is too large for par_ratio'
par_block = default_block + ('f',) * (par_width - len(default_block))
par_head = par_block * par_attn
layer_types = par_head + ('f',) * (par_depth - len(par_head))
elif exists(sandwich_coef):
assert sandwich_coef > 0 and sandwich_coef <= depth, 'sandwich coefficient should be less than the depth'
layer_types = ('a',) * sandwich_coef + default_block * (depth - sandwich_coef) + ('f',) * sandwich_coef
else:
layer_types = default_block * depth
self.layer_types = layer_types
self.num_attn_layers = len(list(filter(equals('a'), layer_types)))
# calculate token shifting
shift_tokens = cast_tuple(shift_tokens, len(layer_types))
# iterate and construct layers
for ind, (layer_type, layer_shift_tokens) in enumerate(zip(self.layer_types, shift_tokens)):
is_last_layer = ind == (len(self.layer_types) - 1)
if layer_type == 'a':
2022-04-08 06:08:03 +00:00
layer = Attention(dim, heads=heads, causal=causal, **attn_kwargs)
2022-04-06 06:21:58 +00:00
elif layer_type == 'c':
2022-04-08 06:08:03 +00:00
layer = Attention(dim, heads=heads, **attn_kwargs)
2022-04-06 06:21:58 +00:00
elif layer_type == 'f':
layer = FeedForward(dim, **ff_kwargs)
layer = layer if not macaron else Scale(0.5, layer)
else:
raise Exception(f'invalid layer type {layer_type}')
if layer_shift_tokens > 0:
shift_range_upper = layer_shift_tokens + 1
shift_range_lower = -layer_shift_tokens if not causal else 0
layer = ShiftTokens(range(shift_range_lower, shift_range_upper), layer)
if exists(branch_fn):
layer = branch_fn(layer)
residual_fn = GRUGating if gate_residual else Residual
2022-04-08 06:08:03 +00:00
residual = residual_fn(dim, scale_residual=scale_residual)
2022-04-06 06:21:58 +00:00
layer_uses_qk_norm = use_qk_norm_attn and layer_type in ('a', 'c')
pre_branch_norm = norm_fn() if pre_norm and not layer_uses_qk_norm else None
post_branch_norm = norm_fn() if sandwich_norm or layer_uses_qk_norm else None
post_main_norm = norm_fn() if not pre_norm and not is_last_layer else None
norms = nn.ModuleList([
pre_branch_norm,
post_branch_norm,
post_main_norm
])
self.layers.append(nn.ModuleList([
norms,
layer,
residual
]))
def forward(
2022-04-08 06:08:03 +00:00
self,
x,
context=None,
full_context=None, # for passing a list of hidden states from an encoder
mask=None,
context_mask=None,
attn_mask=None,
mems=None,
return_hiddens=False,
norm_scale_shift_inp=None,
2022-04-06 06:21:58 +00:00
):
2022-04-08 06:08:03 +00:00
assert not (self.cross_attend ^ (exists(context) or exists(
full_context))), 'context must be passed in if cross_attend is set to True'
2022-04-07 03:04:23 +00:00
assert context is None or full_context is None, 'only one of full_context or context can be provided'
2022-04-06 06:21:58 +00:00
hiddens = []
intermediates = []
prev_attn = None
prev_cross_attn = None
mems = mems.copy() if exists(mems) else [None] * self.num_attn_layers
2022-04-08 02:10:57 +00:00
norm_args = {}
if exists(norm_scale_shift_inp):
norm_args['norm_scale_shift_inp'] = norm_scale_shift_inp
2022-04-06 06:21:58 +00:00
rotary_pos_emb = None
if exists(self.rotary_pos_emb):
max_rotary_emb_length = max(list(map(lambda m: (m.shape[1] if exists(m) else 0) + x.shape[1], mems)))
rotary_pos_emb = self.rotary_pos_emb(max_rotary_emb_length, x.device)
2022-04-07 03:04:23 +00:00
cross_attn_count = 0
2022-04-06 06:21:58 +00:00
for ind, (layer_type, (norm, block, residual_fn)) in enumerate(zip(self.layer_types, self.layers)):
if layer_type == 'a':
layer_mem = mems.pop(0) if mems else None
residual = x
pre_branch_norm, post_branch_norm, post_main_norm = norm
if exists(pre_branch_norm):
2022-04-08 02:10:57 +00:00
x = pre_branch_norm(x, **norm_args)
2022-04-06 06:21:58 +00:00
if layer_type == 'a':
2022-04-08 06:08:03 +00:00
out, inter = checkpoint(block, x, None, mask, None, attn_mask, self.pia_pos_emb, rotary_pos_emb,
prev_attn, layer_mem)
2022-04-06 06:21:58 +00:00
elif layer_type == 'c':
2022-04-07 03:04:23 +00:00
if exists(full_context):
2022-04-08 06:08:03 +00:00
out, inter = checkpoint(block, x, full_context[cross_attn_count], mask, context_mask, None, None,
None, prev_attn)
2022-04-07 03:04:23 +00:00
else:
2022-04-08 06:08:03 +00:00
out, inter = checkpoint(block, x, context, mask, context_mask, None, None, None, prev_attn)
2022-04-06 06:21:58 +00:00
elif layer_type == 'f':
out = checkpoint(block, x)
2022-04-06 06:21:58 +00:00
if exists(post_branch_norm):
2022-04-08 02:10:57 +00:00
out = post_branch_norm(out, **norm_args)
2022-04-06 06:21:58 +00:00
x = residual_fn(out, residual)
if layer_type in ('a', 'c'):
intermediates.append(inter)
if layer_type == 'a' and self.residual_attn:
prev_attn = inter.pre_softmax_attn
elif layer_type == 'c' and self.cross_residual_attn:
prev_cross_attn = inter.pre_softmax_attn
if exists(post_main_norm):
2022-04-08 02:10:57 +00:00
x = post_main_norm(x, **norm_args)
2022-04-06 06:21:58 +00:00
2022-04-07 03:04:23 +00:00
if layer_type == 'c':
cross_attn_count += 1
if layer_type == 'f':
hiddens.append(x)
2022-04-06 06:21:58 +00:00
if return_hiddens:
intermediates = LayerIntermediates(
2022-04-08 06:08:03 +00:00
hiddens=hiddens,
attn_intermediates=intermediates
2022-04-06 06:21:58 +00:00
)
return x, intermediates
return x
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
class Encoder(AttentionLayers):
def __init__(self, **kwargs):
assert 'causal' not in kwargs, 'cannot set causality on encoder'
2022-04-08 06:08:03 +00:00
super().__init__(causal=False, **kwargs)
2022-04-06 06:21:58 +00:00
class Decoder(AttentionLayers):
def __init__(self, **kwargs):
assert 'causal' not in kwargs, 'cannot set causality on decoder'
2022-04-08 06:08:03 +00:00
super().__init__(causal=True, **kwargs)
2022-04-06 06:21:58 +00:00
class CrossAttender(AttentionLayers):
def __init__(self, **kwargs):
2022-04-08 06:08:03 +00:00
super().__init__(cross_attend=True, only_cross=True, **kwargs)
2022-04-06 06:21:58 +00:00
class ViTransformerWrapper(nn.Module):
def __init__(
2022-04-08 06:08:03 +00:00
self,
*,
image_size,
patch_size,
attn_layers,
num_classes=None,
dropout=0.,
emb_dropout=0.
2022-04-06 06:21:58 +00:00
):
super().__init__()
assert isinstance(attn_layers, Encoder), 'attention layers must be an Encoder'
assert image_size % patch_size == 0, 'image dimensions must be divisible by the patch size'
dim = attn_layers.dim
num_patches = (image_size // patch_size) ** 2
patch_dim = 3 * patch_size ** 2
self.patch_size = patch_size
self.pos_embedding = nn.Parameter(torch.randn(1, num_patches + 1, dim))
self.patch_to_embedding = nn.Linear(patch_dim, dim)
self.cls_token = nn.Parameter(torch.randn(1, 1, dim))
self.dropout = nn.Dropout(emb_dropout)
self.attn_layers = attn_layers
self.norm = nn.LayerNorm(dim)
2022-04-08 06:08:03 +00:00
self.mlp_head = FeedForward(dim, dim_out=num_classes, dropout=dropout) if exists(num_classes) else None
2022-04-06 06:21:58 +00:00
def forward(
2022-04-08 06:08:03 +00:00
self,
img,
return_embeddings=False
2022-04-06 06:21:58 +00:00
):
p = self.patch_size
2022-04-08 06:08:03 +00:00
x = rearrange(img, 'b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1=p, p2=p)
2022-04-06 06:21:58 +00:00
x = self.patch_to_embedding(x)
b, n, _ = x.shape
2022-04-08 06:08:03 +00:00
cls_tokens = repeat(self.cls_token, '() n d -> b n d', b=b)
2022-04-06 06:21:58 +00:00
x = torch.cat((cls_tokens, x), dim=1)
x = x + self.pos_embedding[:, :(n + 1)]
x = self.dropout(x)
x = self.attn_layers(x)
x = self.norm(x)
if not exists(self.mlp_head) or return_embeddings:
return x
return self.mlp_head(x[:, 0])
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
class TransformerWrapper(nn.Module):
def __init__(
2022-04-08 06:08:03 +00:00
self,
*,
num_tokens,
max_seq_len,
attn_layers,
emb_dim=None,
max_mem_len=0.,
shift_mem_down=0,
emb_dropout=0.,
num_memory_tokens=None,
tie_embedding=False,
use_pos_emb=True
2022-04-06 06:21:58 +00:00
):
super().__init__()
assert isinstance(attn_layers, AttentionLayers), 'attention layers must be one of Encoder or Decoder'
dim = attn_layers.dim
emb_dim = default(emb_dim, dim)
self.max_seq_len = max_seq_len
self.max_mem_len = max_mem_len
self.shift_mem_down = shift_mem_down
self.token_emb = nn.Embedding(num_tokens, emb_dim)
2022-04-08 06:08:03 +00:00
self.pos_emb = AbsolutePositionalEmbedding(emb_dim, max_seq_len) if (
use_pos_emb and not attn_layers.has_pos_emb) else always(0)
2022-04-06 06:21:58 +00:00
self.emb_dropout = nn.Dropout(emb_dropout)
self.project_emb = nn.Linear(emb_dim, dim) if emb_dim != dim else nn.Identity()
self.attn_layers = attn_layers
self.norm = nn.LayerNorm(dim)
self.init_()
self.to_logits = nn.Linear(dim, num_tokens) if not tie_embedding else lambda t: t @ self.token_emb.weight.t()
# memory tokens (like [cls]) from Memory Transformers paper
num_memory_tokens = default(num_memory_tokens, 0)
self.num_memory_tokens = num_memory_tokens
if num_memory_tokens > 0:
self.memory_tokens = nn.Parameter(torch.randn(num_memory_tokens, dim))
def init_(self):
nn.init.kaiming_normal_(self.token_emb.weight)
def forward(
2022-04-08 06:08:03 +00:00
self,
x,
return_embeddings=False,
mask=None,
return_hiddens=False,
return_attn=False,
mems=None,
**kwargs
2022-04-06 06:21:58 +00:00
):
b, n, device, num_mem = *x.shape, x.device, self.num_memory_tokens
x = self.token_emb(x)
x = x + self.pos_emb(x)
x = self.emb_dropout(x)
x = self.project_emb(x)
if num_mem > 0:
2022-04-08 06:08:03 +00:00
mem = repeat(self.memory_tokens, 'n d -> b n d', b=b)
x = torch.cat((mem, x), dim=1)
2022-04-06 06:21:58 +00:00
# auto-handle masking after appending memory tokens
if exists(mask):
2022-04-08 06:08:03 +00:00
mask = F.pad(mask, (num_mem, 0), value=True)
2022-04-06 06:21:58 +00:00
if self.shift_mem_down and exists(mems):
mems_l, mems_r = mems[:self.shift_mem_down], mems[self.shift_mem_down:]
mems = [*mems_r, *mems_l]
2022-04-08 06:08:03 +00:00
x, intermediates = self.attn_layers(x, mask=mask, mems=mems, return_hiddens=True, **kwargs)
2022-04-06 06:21:58 +00:00
x = self.norm(x)
mem, x = x[:, :num_mem], x[:, num_mem:]
out = self.to_logits(x) if not return_embeddings else x
2022-04-07 03:04:23 +00:00
if return_hiddens:
2022-04-06 06:21:58 +00:00
hiddens = intermediates.hiddens
2022-04-07 03:04:23 +00:00
return out, hiddens
2022-04-06 06:21:58 +00:00
if return_attn:
attn_maps = list(map(lambda t: t.post_softmax_attn, intermediates.attn_intermediates))
return out, attn_maps
return out
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
class ContinuousTransformerWrapper(nn.Module):
def __init__(
2022-04-08 06:08:03 +00:00
self,
*,
max_seq_len,
attn_layers,
dim_in=None,
dim_out=None,
emb_dim=None,
emb_dropout=0.,
use_pos_emb=True
2022-04-06 06:21:58 +00:00
):
super().__init__()
assert isinstance(attn_layers, AttentionLayers), 'attention layers must be one of Encoder or Decoder'
dim = attn_layers.dim
self.max_seq_len = max_seq_len
2022-04-08 06:08:03 +00:00
self.pos_emb = AbsolutePositionalEmbedding(dim, max_seq_len) if (
use_pos_emb and not attn_layers.has_pos_emb) else always(0)
2022-04-06 06:21:58 +00:00
self.emb_dropout = nn.Dropout(emb_dropout)
self.project_in = nn.Linear(dim_in, dim) if exists(dim_in) else nn.Identity()
self.attn_layers = attn_layers
self.norm = nn.LayerNorm(dim)
self.project_out = nn.Linear(dim, dim_out) if exists(dim_out) else nn.Identity()
def forward(
2022-04-08 06:08:03 +00:00
self,
x,
return_embeddings=False,
mask=None,
return_attn=False,
mems=None,
**kwargs
2022-04-06 06:21:58 +00:00
):
b, n, _, device = *x.shape, x.device
x = self.project_in(x)
x = x + self.pos_emb(x)
x = self.emb_dropout(x)
2022-04-08 06:08:03 +00:00
x, intermediates = self.attn_layers(x, mask=mask, mems=mems, return_hiddens=True, **kwargs)
2022-04-06 06:21:58 +00:00
x = self.norm(x)
out = self.project_out(x) if not return_embeddings else x
if return_attn:
attn_maps = list(map(lambda t: t.post_softmax_attn, intermediates.attn_intermediates))
return out, attn_maps
return out
2022-04-08 06:08:03 +00:00
2022-04-06 06:21:58 +00:00
class XTransformer(nn.Module):
def __init__(
2022-04-08 06:08:03 +00:00
self,
*,
dim,
tie_token_emb=False,
**kwargs
2022-04-06 06:21:58 +00:00
):
super().__init__()
enc_kwargs, kwargs = groupby_prefix_and_trim('enc_', kwargs)
dec_kwargs, kwargs = groupby_prefix_and_trim('dec_', kwargs)
assert 'dim' not in enc_kwargs and 'dim' not in dec_kwargs, 'dimension of either encoder or decoder must be set with `dim` keyword'
enc_transformer_kwargs = pick_and_pop(['num_tokens', 'max_seq_len'], enc_kwargs)
enc_transformer_kwargs['emb_dropout'] = enc_kwargs.pop('emb_dropout', 0)
enc_transformer_kwargs['num_memory_tokens'] = enc_kwargs.pop('num_memory_tokens', None)
enc_transformer_kwargs['use_pos_emb'] = enc_kwargs.pop('use_pos_emb', True)
dec_transformer_kwargs = pick_and_pop(['num_tokens', 'max_seq_len'], dec_kwargs)
dec_transformer_kwargs['emb_dropout'] = dec_kwargs.pop('emb_dropout', 0)
dec_transformer_kwargs['use_pos_emb'] = dec_kwargs.pop('use_pos_emb', True)
self.encoder = TransformerWrapper(
**enc_transformer_kwargs,
2022-04-08 06:08:03 +00:00
attn_layers=Encoder(dim=dim, **enc_kwargs)
2022-04-06 06:21:58 +00:00
)
self.decoder = TransformerWrapper(
**dec_transformer_kwargs,
2022-04-08 06:08:03 +00:00
attn_layers=Decoder(dim=dim, cross_attend=True, **dec_kwargs)
2022-04-06 06:21:58 +00:00
)
if tie_token_emb:
self.decoder.token_emb = self.encoder.token_emb
self.decoder = AutoregressiveWrapper(self.decoder)
@torch.no_grad()
2022-04-08 06:08:03 +00:00
def generate(self, seq_in, seq_out_start, seq_len, src_mask=None, src_attn_mask=None, **kwargs):
encodings = self.encoder(seq_in, mask=src_mask, attn_mask=src_attn_mask, return_embeddings=True)
return self.decoder.generate(seq_out_start, seq_len, context=encodings, context_mask=src_mask, **kwargs)
2022-04-06 06:21:58 +00:00
2022-04-08 06:08:03 +00:00
def forward(self, src, tgt, src_mask=None, tgt_mask=None, src_attn_mask=None):
enc = self.encoder(src, mask=src_mask, attn_mask=src_attn_mask, return_embeddings=True)
out = self.decoder(tgt, context=enc, mask=tgt_mask, context_mask=src_mask)
2022-04-06 06:21:58 +00:00
return out