stable-diffusion-webui/modules/interrogate.py
Patryk Wychowaniec 7157e5d064 interrogate: Fix CLIP-interrogation on CPU
Currently, trying to perform CLIP interrogation on a CPU fails, saying:

```
RuntimeError: "slow_conv2d_cpu" not implemented for 'Half'
```

This merge request fixes this issue by detecting whether the target
device is CPU and, if so, force-enabling `--no-half` and passing
`device="cpu"` to `clip.load()` (which then does some extra tricks to
ensure it works correctly on CPU).
2022-10-21 09:52:12 +03:00

177 lines
6.6 KiB
Python

import contextlib
import os
import sys
import traceback
from collections import namedtuple
import re
import torch
from torchvision import transforms
from torchvision.transforms.functional import InterpolationMode
import modules.shared as shared
from modules import devices, paths, lowvram
blip_image_eval_size = 384
blip_model_url = 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model_base_caption_capfilt_large.pth'
clip_model_name = 'ViT-L/14'
Category = namedtuple("Category", ["name", "topn", "items"])
re_topn = re.compile(r"\.top(\d+)\.")
class InterrogateModels:
blip_model = None
clip_model = None
clip_preprocess = None
categories = None
dtype = None
running_on_cpu = None
def __init__(self, content_dir):
self.categories = []
self.running_on_cpu = devices.device_interrogate == torch.device("cpu")
if os.path.exists(content_dir):
for filename in os.listdir(content_dir):
m = re_topn.search(filename)
topn = 1 if m is None else int(m.group(1))
with open(os.path.join(content_dir, filename), "r", encoding="utf8") as file:
lines = [x.strip() for x in file.readlines()]
self.categories.append(Category(name=filename, topn=topn, items=lines))
def load_blip_model(self):
import models.blip
blip_model = models.blip.blip_decoder(pretrained=blip_model_url, image_size=blip_image_eval_size, vit='base', med_config=os.path.join(paths.paths["BLIP"], "configs", "med_config.json"))
blip_model.eval()
return blip_model
def load_clip_model(self):
import clip
if self.running_on_cpu:
model, preprocess = clip.load(clip_model_name, device="cpu")
else:
model, preprocess = clip.load(clip_model_name)
model.eval()
model = model.to(devices.device_interrogate)
return model, preprocess
def load(self):
if self.blip_model is None:
self.blip_model = self.load_blip_model()
if not shared.cmd_opts.no_half and not self.running_on_cpu:
self.blip_model = self.blip_model.half()
self.blip_model = self.blip_model.to(devices.device_interrogate)
if self.clip_model is None:
self.clip_model, self.clip_preprocess = self.load_clip_model()
if not shared.cmd_opts.no_half and not self.running_on_cpu:
self.clip_model = self.clip_model.half()
self.clip_model = self.clip_model.to(devices.device_interrogate)
self.dtype = next(self.clip_model.parameters()).dtype
def send_clip_to_ram(self):
if not shared.opts.interrogate_keep_models_in_memory:
if self.clip_model is not None:
self.clip_model = self.clip_model.to(devices.cpu)
def send_blip_to_ram(self):
if not shared.opts.interrogate_keep_models_in_memory:
if self.blip_model is not None:
self.blip_model = self.blip_model.to(devices.cpu)
def unload(self):
self.send_clip_to_ram()
self.send_blip_to_ram()
devices.torch_gc()
def rank(self, image_features, text_array, top_count=1):
import clip
if shared.opts.interrogate_clip_dict_limit != 0:
text_array = text_array[0:int(shared.opts.interrogate_clip_dict_limit)]
top_count = min(top_count, len(text_array))
text_tokens = clip.tokenize([text for text in text_array], truncate=True).to(devices.device_interrogate)
text_features = self.clip_model.encode_text(text_tokens).type(self.dtype)
text_features /= text_features.norm(dim=-1, keepdim=True)
similarity = torch.zeros((1, len(text_array))).to(devices.device_interrogate)
for i in range(image_features.shape[0]):
similarity += (100.0 * image_features[i].unsqueeze(0) @ text_features.T).softmax(dim=-1)
similarity /= image_features.shape[0]
top_probs, top_labels = similarity.cpu().topk(top_count, dim=-1)
return [(text_array[top_labels[0][i].numpy()], (top_probs[0][i].numpy()*100)) for i in range(top_count)]
def generate_caption(self, pil_image):
gpu_image = transforms.Compose([
transforms.Resize((blip_image_eval_size, blip_image_eval_size), interpolation=InterpolationMode.BICUBIC),
transforms.ToTensor(),
transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711))
])(pil_image).unsqueeze(0).type(self.dtype).to(devices.device_interrogate)
with torch.no_grad():
caption = self.blip_model.generate(gpu_image, sample=False, num_beams=shared.opts.interrogate_clip_num_beams, min_length=shared.opts.interrogate_clip_min_length, max_length=shared.opts.interrogate_clip_max_length)
return caption[0]
def interrogate(self, pil_image):
res = None
try:
if shared.cmd_opts.lowvram or shared.cmd_opts.medvram:
lowvram.send_everything_to_cpu()
devices.torch_gc()
self.load()
caption = self.generate_caption(pil_image)
self.send_blip_to_ram()
devices.torch_gc()
res = caption
clip_image = self.clip_preprocess(pil_image).unsqueeze(0).type(self.dtype).to(devices.device_interrogate)
precision_scope = torch.autocast if shared.cmd_opts.precision == "autocast" else contextlib.nullcontext
with torch.no_grad(), precision_scope("cuda"):
image_features = self.clip_model.encode_image(clip_image).type(self.dtype)
image_features /= image_features.norm(dim=-1, keepdim=True)
if shared.opts.interrogate_use_builtin_artists:
artist = self.rank(image_features, ["by " + artist.name for artist in shared.artist_db.artists])[0]
res += ", " + artist[0]
for name, topn, items in self.categories:
matches = self.rank(image_features, items, top_count=topn)
for match, score in matches:
if shared.opts.interrogate_return_ranks:
res += f", ({match}:{score/100:.3f})"
else:
res += ", " + match
except Exception:
print(f"Error interrogating", file=sys.stderr)
print(traceback.format_exc(), file=sys.stderr)
res += "<error>"
self.unload()
return res