mirror of
https://github.com/invoke-ai/InvokeAI
synced 2024-08-30 20:32:17 +00:00
Refactoring simplet2i (#387)
* start refactoring -not yet functional * first phase of refactor done - not sure weighted prompts working * Second phase of refactoring. Everything mostly working. * The refactoring has moved all the hard-core inference work into ldm.dream.generator.*, where there are submodules for txt2img and img2img. inpaint will go in there as well. * Some additional refactoring will be done soon, but relatively minor work. * fix -save_orig flag to actually work * add @neonsecret attention.py memory optimization * remove unneeded imports * move token logging into conditioning.py * add placeholder version of inpaint; porting in progress * fix crash in img2img * inpainting working; not tested on variations * fix crashes in img2img * ported attention.py memory optimization #117 from basujindal branch * added @torch_no_grad() decorators to img2img, txt2img, inpaint closures * Final commit prior to PR against development * fixup crash when generating intermediate images in web UI * rename ldm.simplet2i to ldm.generate * add backward-compatibility simplet2i shell with deprecation warning * add back in mps exception, addresses @vargol comment in #354 * replaced Conditioning class with exported functions * fix wrong type of with_variations attribute during intialization * changed "image_iterator()" to "get_make_image()" * raise NotImplementedError for calling get_make_image() in parent class * Update ldm/generate.py better error message Co-authored-by: Kevin Gibbons <bakkot@gmail.com> * minor stylistic fixes and assertion checks from code review * moved get_noise() method into img2img class * break get_noise() into two methods, one for txt2img and the other for img2img * inpainting works on non-square images now * make get_noise() an abstract method in base class * much improved inpainting Co-authored-by: Kevin Gibbons <bakkot@gmail.com>
This commit is contained in:
Lincoln Stein
GitHub
@ -13,7 +13,7 @@ def exists(val):
|
||||
|
||||
|
||||
def uniq(arr):
|
||||
return {el: True for el in arr}.keys()
|
||||
return{el: True for el in arr}.keys()
|
||||
|
||||
|
||||
def default(val, d):
|
||||
@ -45,18 +45,19 @@ class GEGLU(nn.Module):
|
||||
|
||||
|
||||
class FeedForward(nn.Module):
|
||||
def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.0):
|
||||
def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
|
||||
super().__init__()
|
||||
inner_dim = int(dim * mult)
|
||||
dim_out = default(dim_out, dim)
|
||||
project_in = (
|
||||
nn.Sequential(nn.Linear(dim, inner_dim), nn.GELU())
|
||||
if not glu
|
||||
else GEGLU(dim, inner_dim)
|
||||
)
|
||||
project_in = nn.Sequential(
|
||||
nn.Linear(dim, inner_dim),
|
||||
nn.GELU()
|
||||
) if not glu else GEGLU(dim, inner_dim)
|
||||
|
||||
self.net = nn.Sequential(
|
||||
project_in, nn.Dropout(dropout), nn.Linear(inner_dim, dim_out)
|
||||
project_in,
|
||||
nn.Dropout(dropout),
|
||||
nn.Linear(inner_dim, dim_out)
|
||||
)
|
||||
|
||||
def forward(self, x):
|
||||
@ -73,9 +74,7 @@ def zero_module(module):
|
||||
|
||||
|
||||
def Normalize(in_channels):
|
||||
return torch.nn.GroupNorm(
|
||||
num_groups=32, num_channels=in_channels, eps=1e-6, affine=True
|
||||
)
|
||||
return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
|
||||
|
||||
|
||||
class LinearAttention(nn.Module):
|
||||
@ -83,28 +82,17 @@ class LinearAttention(nn.Module):
|
||||
super().__init__()
|
||||
self.heads = heads
|
||||
hidden_dim = dim_head * heads
|
||||
self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias=False)
|
||||
self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
|
||||
self.to_out = nn.Conv2d(hidden_dim, dim, 1)
|
||||
|
||||
def forward(self, x):
|
||||
b, c, h, w = x.shape
|
||||
qkv = self.to_qkv(x)
|
||||
q, k, v = rearrange(
|
||||
qkv,
|
||||
'b (qkv heads c) h w -> qkv b heads c (h w)',
|
||||
heads=self.heads,
|
||||
qkv=3,
|
||||
)
|
||||
k = k.softmax(dim=-1)
|
||||
q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
|
||||
k = k.softmax(dim=-1)
|
||||
context = torch.einsum('bhdn,bhen->bhde', k, v)
|
||||
out = torch.einsum('bhde,bhdn->bhen', context, q)
|
||||
out = rearrange(
|
||||
out,
|
||||
'b heads c (h w) -> b (heads c) h w',
|
||||
heads=self.heads,
|
||||
h=h,
|
||||
w=w,
|
||||
)
|
||||
out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
|
||||
return self.to_out(out)
|
||||
|
||||
|
||||
@ -114,18 +102,26 @@ class SpatialSelfAttention(nn.Module):
|
||||
self.in_channels = in_channels
|
||||
|
||||
self.norm = Normalize(in_channels)
|
||||
self.q = torch.nn.Conv2d(
|
||||
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
||||
)
|
||||
self.k = torch.nn.Conv2d(
|
||||
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
||||
)
|
||||
self.v = torch.nn.Conv2d(
|
||||
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
||||
)
|
||||
self.proj_out = torch.nn.Conv2d(
|
||||
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
||||
)
|
||||
self.q = torch.nn.Conv2d(in_channels,
|
||||
in_channels,
|
||||
kernel_size=1,
|
||||
stride=1,
|
||||
padding=0)
|
||||
self.k = torch.nn.Conv2d(in_channels,
|
||||
in_channels,
|
||||
kernel_size=1,
|
||||
stride=1,
|
||||
padding=0)
|
||||
self.v = torch.nn.Conv2d(in_channels,
|
||||
in_channels,
|
||||
kernel_size=1,
|
||||
stride=1,
|
||||
padding=0)
|
||||
self.proj_out = torch.nn.Conv2d(in_channels,
|
||||
in_channels,
|
||||
kernel_size=1,
|
||||
stride=1,
|
||||
padding=0)
|
||||
|
||||
def forward(self, x):
|
||||
h_ = x
|
||||
@ -135,12 +131,12 @@ class SpatialSelfAttention(nn.Module):
|
||||
v = self.v(h_)
|
||||
|
||||
# compute attention
|
||||
b, c, h, w = q.shape
|
||||
b,c,h,w = q.shape
|
||||
q = rearrange(q, 'b c h w -> b (h w) c')
|
||||
k = rearrange(k, 'b c h w -> b c (h w)')
|
||||
w_ = torch.einsum('bij,bjk->bik', q, k)
|
||||
|
||||
w_ = w_ * (int(c) ** (-0.5))
|
||||
w_ = w_ * (int(c)**(-0.5))
|
||||
w_ = torch.nn.functional.softmax(w_, dim=2)
|
||||
|
||||
# attend to values
|
||||
@ -150,18 +146,16 @@ class SpatialSelfAttention(nn.Module):
|
||||
h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
|
||||
h_ = self.proj_out(h_)
|
||||
|
||||
return x + h_
|
||||
return x+h_
|
||||
|
||||
|
||||
class CrossAttention(nn.Module):
|
||||
def __init__(
|
||||
self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0
|
||||
):
|
||||
def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.):
|
||||
super().__init__()
|
||||
inner_dim = dim_head * heads
|
||||
context_dim = default(context_dim, query_dim)
|
||||
|
||||
self.scale = dim_head**-0.5
|
||||
self.scale = dim_head ** -0.5
|
||||
self.heads = heads
|
||||
|
||||
self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
|
||||
@ -169,7 +163,8 @@ class CrossAttention(nn.Module):
|
||||
self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
|
||||
|
||||
self.to_out = nn.Sequential(
|
||||
nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)
|
||||
nn.Linear(inner_dim, query_dim),
|
||||
nn.Dropout(dropout)
|
||||
)
|
||||
|
||||
def forward(self, x, context=None, mask=None):
|
||||
@ -179,59 +174,43 @@ class CrossAttention(nn.Module):
|
||||
context = default(context, x)
|
||||
k = self.to_k(context)
|
||||
v = self.to_v(context)
|
||||
del context, x
|
||||
|
||||
q, k, v = map(
|
||||
lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v)
|
||||
)
|
||||
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
|
||||
|
||||
sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
|
||||
sim = einsum('b i d, b j d -> b i j', q, k) * self.scale # (8, 4096, 40)
|
||||
del q, k
|
||||
|
||||
if exists(mask):
|
||||
mask = rearrange(mask, 'b ... -> b (...)')
|
||||
max_neg_value = -torch.finfo(sim.dtype).max
|
||||
mask = repeat(mask, 'b j -> (b h) () j', h=h)
|
||||
sim.masked_fill_(~mask, max_neg_value)
|
||||
del mask
|
||||
|
||||
# attention, what we cannot get enough of
|
||||
attn = sim.softmax(dim=-1)
|
||||
# attention, what we cannot get enough of, by halves
|
||||
sim[4:] = sim[4:].softmax(dim=-1)
|
||||
sim[:4] = sim[:4].softmax(dim=-1)
|
||||
|
||||
out = einsum('b i j, b j d -> b i d', attn, v)
|
||||
out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
|
||||
return self.to_out(out)
|
||||
sim = einsum('b i j, b j d -> b i d', sim, v)
|
||||
sim = rearrange(sim, '(b h) n d -> b n (h d)', h=h)
|
||||
return self.to_out(sim)
|
||||
|
||||
|
||||
class BasicTransformerBlock(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
dim,
|
||||
n_heads,
|
||||
d_head,
|
||||
dropout=0.0,
|
||||
context_dim=None,
|
||||
gated_ff=True,
|
||||
checkpoint=True,
|
||||
):
|
||||
def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True):
|
||||
super().__init__()
|
||||
self.attn1 = CrossAttention(
|
||||
query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout
|
||||
) # is a self-attention
|
||||
self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout) # is a self-attention
|
||||
self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
|
||||
self.attn2 = CrossAttention(
|
||||
query_dim=dim,
|
||||
context_dim=context_dim,
|
||||
heads=n_heads,
|
||||
dim_head=d_head,
|
||||
dropout=dropout,
|
||||
) # is self-attn if context is none
|
||||
self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim,
|
||||
heads=n_heads, dim_head=d_head, dropout=dropout) # is self-attn if context is none
|
||||
self.norm1 = nn.LayerNorm(dim)
|
||||
self.norm2 = nn.LayerNorm(dim)
|
||||
self.norm3 = nn.LayerNorm(dim)
|
||||
self.checkpoint = checkpoint
|
||||
|
||||
def forward(self, x, context=None):
|
||||
return checkpoint(
|
||||
self._forward, (x, context), self.parameters(), self.checkpoint
|
||||
)
|
||||
return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint)
|
||||
|
||||
def _forward(self, x, context=None):
|
||||
x = x.contiguous() if x.device.type == 'mps' else x
|
||||
@ -249,43 +228,29 @@ class SpatialTransformer(nn.Module):
|
||||
Then apply standard transformer action.
|
||||
Finally, reshape to image
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
in_channels,
|
||||
n_heads,
|
||||
d_head,
|
||||
depth=1,
|
||||
dropout=0.0,
|
||||
context_dim=None,
|
||||
):
|
||||
def __init__(self, in_channels, n_heads, d_head,
|
||||
depth=1, dropout=0., context_dim=None):
|
||||
super().__init__()
|
||||
self.in_channels = in_channels
|
||||
inner_dim = n_heads * d_head
|
||||
self.norm = Normalize(in_channels)
|
||||
|
||||
self.proj_in = nn.Conv2d(
|
||||
in_channels, inner_dim, kernel_size=1, stride=1, padding=0
|
||||
)
|
||||
self.proj_in = nn.Conv2d(in_channels,
|
||||
inner_dim,
|
||||
kernel_size=1,
|
||||
stride=1,
|
||||
padding=0)
|
||||
|
||||
self.transformer_blocks = nn.ModuleList(
|
||||
[
|
||||
BasicTransformerBlock(
|
||||
inner_dim,
|
||||
n_heads,
|
||||
d_head,
|
||||
dropout=dropout,
|
||||
context_dim=context_dim,
|
||||
)
|
||||
for d in range(depth)
|
||||
]
|
||||
[BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim)
|
||||
for d in range(depth)]
|
||||
)
|
||||
|
||||
self.proj_out = zero_module(
|
||||
nn.Conv2d(
|
||||
inner_dim, in_channels, kernel_size=1, stride=1, padding=0
|
||||
)
|
||||
)
|
||||
self.proj_out = zero_module(nn.Conv2d(inner_dim,
|
||||
in_channels,
|
||||
kernel_size=1,
|
||||
stride=1,
|
||||
padding=0))
|
||||
|
||||
def forward(self, x, context=None):
|
||||
# note: if no context is given, cross-attention defaults to self-attention
|
||||
|
||||
Reference in New Issue
Block a user