Mirrored_Repos/InvokeAI
1
0
Fork 0
mirror of https://github.com/invoke-ai/InvokeAI synced 2024-08-30 20:32:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

cleanup: merge conflicts

Browse Source
This commit is contained in:
blessedcoolant
2023-09-05 11:37:12 +12:00
parent 6bb378a101
commit 07381e5a26
8 changed files with 86 additions and 69 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
invokeai/backend/ip_adapter/attention_processor.py
View File

@ -12,6 +12,7 @@ class AttnProcessor(nn.Module):
r"""
Default processor for performing attention-related computations.
"""
def __init__(
self,
hidden_size=None,
@ -140,7 +141,10 @@ class IPAttnProcessor(nn.Module):
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
# split hidden states
encoder_hidden_states, ip_hidden_states = encoder_hidden_states[:, :self.text_context_len, :], encoder_hidden_states[:, self.text_context_len:, :]
encoder_hidden_states, ip_hidden_states = (
encoder_hidden_states[:, : self.text_context_len, :],
encoder_hidden_states[:, self.text_context_len :, :],
)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
@ -186,6 +190,7 @@ class AttnProcessor2_0(torch.nn.Module):
r"""
Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
"""
def __init__(
self,
hidden_size=None,
@ -338,7 +343,10 @@ class IPAttnProcessor2_0(torch.nn.Module):
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
# split hidden states
encoder_hidden_states, ip_hidden_states = encoder_hidden_states[:, :self.text_context_len, :], encoder_hidden_states[:, self.text_context_len:, :]
encoder_hidden_states, ip_hidden_states = (
encoder_hidden_states[:, : self.text_context_len, :],
encoder_hidden_states[:, self.text_context_len :, :],
)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)

42
invokeai/backend/ip_adapter/ip_adapter.py
View File

@ -22,6 +22,7 @@ from .resampler import Resampler
class ImageProjModel(torch.nn.Module):
"""Projection Model"""
def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4):
super().__init__()
@ -32,15 +33,15 @@ class ImageProjModel(torch.nn.Module):
def forward(self, image_embeds):
embeds = image_embeds
clip_extra_context_tokens = self.proj(embeds).reshape(-1, self.clip_extra_context_tokens, self.cross_attention_dim)
clip_extra_context_tokens = self.proj(embeds).reshape(
-1, self.clip_extra_context_tokens, self.cross_attention_dim
)
clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
return clip_extra_context_tokens
class IPAdapter:
def __init__(self, sd_pipe, image_encoder_path, ip_ckpt, device, num_tokens=4):
self.device = device
self.image_encoder_path = image_encoder_path
self.ip_ckpt = ip_ckpt
@ -54,7 +55,9 @@ class IPAdapter:
self.set_ip_adapter()
# load image encoder
self.image_encoder = CLIPVisionModelWithProjection.from_pretrained(self.image_encoder_path).to(self.device, dtype=torch.float16)
self.image_encoder = CLIPVisionModelWithProjection.from_pretrained(self.image_encoder_path).to(
self.device, dtype=torch.float16
)
self.clip_image_processor = CLIPImageProcessor()
# image proj model
self.image_proj_model = self.init_proj()
@ -88,8 +91,9 @@ class IPAdapter:
attn_procs[name] = AttnProcessor()
else:
print("swapping in IPAttnProcessor for", name)
attn_procs[name] = IPAttnProcessor(hidden_size=hidden_size, cross_attention_dim=cross_attention_dim,
scale=1.0).to(self.device, dtype=torch.float16)
attn_procs[name] = IPAttnProcessor(
hidden_size=hidden_size, cross_attention_dim=cross_attention_dim, scale=1.0
).to(self.device, dtype=torch.float16)
unet.set_attn_processor(attn_procs)
print("Modified UNet Attn Processors count:", len(unet.attn_processors))
print(unet.attn_processors.keys())
@ -155,7 +159,12 @@ class IPAdapter:
with torch.inference_mode():
prompt_embeds = self.pipe._encode_prompt(
prompt, device=self.device, num_images_per_prompt=num_samples, do_classifier_free_guidance=True, negative_prompt=negative_prompt)
prompt,
device=self.device,
num_images_per_prompt=num_samples,
do_classifier_free_guidance=True,
negative_prompt=negative_prompt,
)
negative_prompt_embeds_, prompt_embeds_ = prompt_embeds.chunk(2)
prompt_embeds = torch.cat([prompt_embeds_, image_prompt_embeds], dim=1)
negative_prompt_embeds = torch.cat([negative_prompt_embeds_, uncond_image_prompt_embeds], dim=1)
@ -212,8 +221,17 @@ class IPAdapterXL(IPAdapter):
uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
with torch.inference_mode():
prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds = self.pipe.encode_prompt(
prompt, num_images_per_prompt=num_samples, do_classifier_free_guidance=True, negative_prompt=negative_prompt)
(
prompt_embeds,
negative_prompt_embeds,
pooled_prompt_embeds,
negative_pooled_prompt_embeds,
) = self.pipe.encode_prompt(
prompt,
num_images_per_prompt=num_samples,
do_classifier_free_guidance=True,
negative_prompt=negative_prompt,
)
prompt_embeds = torch.cat([prompt_embeds, image_prompt_embeds], dim=1)
negative_prompt_embeds = torch.cat([negative_prompt_embeds, uncond_image_prompt_embeds], dim=1)
@ -243,7 +261,7 @@ class IPAdapterPlus(IPAdapter):
num_queries=self.num_tokens,
embedding_dim=self.image_encoder.config.hidden_size,
output_dim=self.pipe.unet.config.cross_attention_dim,
ff_mult=4
ff_mult=4,
).to(self.device, dtype=torch.float16)
return image_proj_model
@ -255,6 +273,8 @@ class IPAdapterPlus(IPAdapter):
clip_image = clip_image.to(self.device, dtype=torch.float16)
clip_image_embeds = self.image_encoder(clip_image, output_hidden_states=True).hidden_states[-2]
image_prompt_embeds = self.image_proj_model(clip_image_embeds)
uncond_clip_image_embeds = self.image_encoder(torch.zeros_like(clip_image), output_hidden_states=True).hidden_states[-2]
uncond_clip_image_embeds = self.image_encoder(
torch.zeros_like(clip_image), output_hidden_states=True
).hidden_states[-2]
uncond_image_prompt_embeds = self.image_proj_model(uncond_clip_image_embeds)
return image_prompt_embeds, uncond_image_prompt_embeds

6
invokeai/backend/ip_adapter/resampler.py
View File

@ -20,7 +20,7 @@ def FeedForward(dim, mult=4):
def reshape_tensor(x, heads):
bs, length, width = x.shape
#(bs, length, width) --> (bs, length, n_heads, dim_per_head)
# (bs, length, width) --> (bs, length, n_heads, dim_per_head)
x = x.view(bs, length, heads, -1)
# (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
x = x.transpose(1, 2)
@ -44,7 +44,6 @@ class PerceiverAttention(nn.Module):
self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
self.to_out = nn.Linear(inner_dim, dim, bias=False)
def forward(self, x, latents):
"""
Args:
@ -68,7 +67,7 @@ class PerceiverAttention(nn.Module):
# attention
scale = 1 / math.sqrt(math.sqrt(self.dim_head))
weight = (q * scale) @ (k * scale).transpose(-2, -1) # More stable with f16 than dividing afterwards
weight = (q * scale) @ (k * scale).transpose(-2, -1) # More stable with f16 than dividing afterwards
weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
out = weight @ v
@ -110,7 +109,6 @@ class Resampler(nn.Module):
)
def forward(self, x):
latents = self.latents.repeat(x.size(0), 1, 1)
x = self.proj_in(x)

9
invokeai/backend/ip_adapter/utils.py
View File

@ -15,7 +15,6 @@ from diffusers.models import ControlNetModel
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
def is_torch2_available():
return hasattr(F, "scaled_dot_product_attention")
@ -150,9 +149,7 @@ def generate(
control_guidance_end = len(control_guidance_start) * [control_guidance_end]
elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
control_guidance_start, control_guidance_end = mult * [control_guidance_start], mult * [
control_guidance_end
]
control_guidance_start, control_guidance_end = mult * [control_guidance_start], mult * [control_guidance_end]
# 1. Check inputs. Raise error if not correct
self.check_inputs(
@ -192,9 +189,7 @@ def generate(
guess_mode = guess_mode or global_pool_conditions
# 3. Encode input prompt
text_encoder_lora_scale = (
cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
)
text_encoder_lora_scale = cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
prompt_embeds = self._encode_prompt(
prompt,
device,