update

src/pipeline.py

@@ -27,7 +27,7 @@ import numpy as np
 import torch.nn as nn
 import torch.nn.functional as F
 from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only
-from model import _load
+from utils import _load
 import torchvision 
 import os
 
@@ -92,4 +92,4 @@ def infer(request: TextToImageRequest, pipeline: Pipeline, generator: Generator)
         sample=1
         empty_cache()
     image=pipeline(request.prompt,generator=generator, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256, height=request.height, width=request.width, output_type="pt").images[0]
-    return torchvision.transforms.functional.to_pil_image(image.to(torch.float32).mul_(2).sub_(1))# torchvision.transforms.functional.to_pil_image(image)
+    return torchvision.transforms.functional.to_pil_image(image.to(torch.float32).mul_(2).sub_(1))# torchvision.transforms.functional.to_pil_image(image)

src/utils.py

@@ -0,0 +1,64 @@
+A=None
+e_sd_pt="ko.pth"
+d_sd_pt="ok.pth"
+import torch as t, torch.nn as nn, torch.nn.functional as F
+def C(n_in, n_out, **kwargs):
+    return nn.Conv2d(n_in, n_out, 3, padding=1, **kwargs)
+class Clamp(nn.Module):
+    def forward(self, x):
+        return t.tanh(x / 3) * 3
+class B(nn.Module):
+    def __init__(self, n_in, n_out):
+        super().__init__()
+        self.conv = nn.Sequential(C(n_in, n_out), nn.ReLU(), C(n_out, n_out), nn.ReLU(), C(n_out, n_out))
+        self.skip = nn.Conv2d(n_in, n_out, 1, bias=False) if n_in != n_out else nn.Identity()
+        self.fuse = nn.ReLU()
+    def forward(self, x):
+        return self.fuse(self.conv(x) + self.skip(x))
+def E(latent_channels=4):
+    return nn.Sequential(
+        C(3, 64), B(64, 64),
+        C(64, 64, stride=2, bias=False), B(64, 64), B(64, 64), B(64, 64),
+        C(64, 64, stride=2, bias=False), B(64, 64), B(64, 64), B(64, 64),
+        C(64, 64, stride=2, bias=False), B(64, 64), B(64, 64), B(64, 64),
+        C(64, latent_channels),
+    )
+def D(latent_channels=16):
+    return nn.Sequential(
+        Clamp(), 
+        C(latent_channels, 48),nn.ReLU(),B(48, 48), B(48, 48),
+        nn.Upsample(scale_factor=2), C(48, 48, bias=False),B(48, 48), B(48, 48), 
+        nn.Upsample(scale_factor=2), C(48, 48, bias=False),B(48, 48),
+        nn.Upsample(scale_factor=2), C(48, 48, bias=False),B(48, 48), 
+        C(48, 3), 
+    )    
+class M(nn.Module):
+    lm, ls = 3, 0.5
+    def __init__(s, ep="encoder.pth", dp="decoder.pth", lc=None):
+        super().__init__()
+        if lc is None: lc = s.glc(str(ep))
+        s.e, s.d = E(lc), D(lc)
+        def f(sd, mod, pfx):
+            f_sd = {k.strip(pfx): v for k, v in sd.items() if k.strip(pfx) in mod.state_dict() and v.size() == mod.state_dict()[k.strip(pfx)].size()}
+            mod.load_state_dict(f_sd, strict=False)
+        if ep: f(t.load(ep, map_location="cpu", weights_only=True), s.e, "encoder.")
+        if dp: f(t.load(dp, map_location="cpu", weights_only=True), s.d, "decoder.")
+        s.e.requires_grad_(False)
+        s.d.requires_grad_(False)
+    def glc(s, ep): return 16 if "taef1" in ep or "taesd3" in ep else 4
+    @staticmethod
+    def sl(x): return x.div(2 * M.lm).add(M.ls).clamp(0, 1)
+    @staticmethod
+    def ul(x): return x.sub(M.ls).mul(2 * M.lm)
+    def forward(s, x, rl=False):
+        l, o = s.e(x), s.d(s.e(x))
+        return (o.clamp(0, 1), l) if rl else o.clamp(0, 1)
+def filter_state_dict(model, state_dict_path):
+    global E
+    state_dict = t.load(state_dict_path, map_location="cpu", weights_only=True)
+    prefix = 'encoder.' if type(model) == E else 'decoder.'
+    return {k.strip(prefix): v for k, v in state_dict.items() if k.strip(prefix) in model.state_dict() and v.size() == model.state_dict()[k.strip(prefix)].size()}
+def _load(model, name, dtype=t.bfloat16):
+    model = E(16) if name=="E" else D(16)
+    model.load_state_dict(filter_state_dict(model, e_sd_pt if name=="E" else d_sd_pt), strict=False)
+    model.requires_grad_(False).to(dtype=dtype)