feat(api): switch to codeformer lib that works with torch 2.x

api/onnx_web/chain/correct_codeformer.py

@@ -1,7 +1,10 @@
 from logging import getLogger
+from os import path
 from typing import Optional
 
+import torch
 from PIL import Image
+from torchvision.transforms.functional import normalize
 
 from ..params import ImageParams, StageParams, UpscaleParams
 from ..server import ServerContext
@@ -16,7 +19,7 @@ class CorrectCodeformerStage(BaseStage):
     def run(
         self,
         worker: WorkerContext,
-        _server: ServerContext,
+        server: ServerContext,
         _stage: StageParams,
         _params: ImageParams,
         sources: StageResult,
@@ -27,10 +30,88 @@ def run(
     ) -> StageResult:
         # must be within the load function for patch to take effect
         # TODO: rewrite and remove
-        from codeformer import CodeFormer
+        from codeformer.basicsr.utils import img2tensor, tensor2img
+        from codeformer.basicsr.utils.registry import ARCH_REGISTRY
+        from codeformer.facelib.utils.face_restoration_helper import FaceRestoreHelper
 
         upscale = upscale.with_args(**kwargs)
-
         device = worker.get_device()
-        pipe = CodeFormer(upscale=upscale.face_outscale).to(device.torch_str())
-        return StageResult(images=[pipe(source) for source in sources.as_image()])
+
+        net = ARCH_REGISTRY.get("CodeFormer")(
+            dim_embd=512,
+            codebook_size=1024,
+            n_head=8,
+            n_layers=9,
+            connect_list=["32", "64", "128", "256"],
+        ).to(device.torch_str())
+
+        ckpt_path = path.join(server.cache_path, "correction-codeformer.pth")
+        checkpoint = torch.load(ckpt_path)["params_ema"]
+        net.load_state_dict(checkpoint)
+        net.eval()
+
+        det_model = "retinaface_resnet50"
+        face_helper = FaceRestoreHelper(
+            upscale.face_outscale,
+            face_size=512,
+            crop_ratio=(1, 1),
+            det_model=det_model,
+            save_ext="png",
+            use_parse=True,
+            device=device.torch_str(),
+        )
+
+        results = []
+        for img in sources.as_image():
+            # clean all the intermediate results to process the next image
+            face_helper.clean_all()
+            face_helper.read_image(img)
+
+            # get face landmarks for each face
+            num_det_faces = face_helper.get_face_landmarks_5(
+                only_center_face=False, resize=640, eye_dist_threshold=5
+            )
+            logger.debug("detected %s faces", num_det_faces)
+
+            # align and warp each face
+            face_helper.align_warp_face()
+
+            # face restoration for each cropped face
+            for cropped_face in face_helper.cropped_faces:
+                # prepare data
+                cropped_face_t = img2tensor(
+                    cropped_face / 255.0, bgr2rgb=True, float32=True
+                )
+                normalize(
+                    cropped_face_t, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True
+                )
+                cropped_face_t = cropped_face_t.unsqueeze(0).to(device.torch_str())
+
+                try:
+                    with torch.no_grad():
+                        output = net(
+                            cropped_face_t, w=upscale.face_strength, adain=True
+                        )[0]
+                        restored_face = tensor2img(
+                            output, rgb2bgr=True, min_max=(-1, 1)
+                        )
+
+                    del output
+                except Exception:
+                    logger.exception("inference failed for CodeFormer")
+                    restored_face = tensor2img(
+                        cropped_face_t, rgb2bgr=True, min_max=(-1, 1)
+                    )
+
+                restored_face = restored_face.astype("uint8")
+                face_helper.add_restored_face(restored_face, cropped_face)
+
+            # paste_back
+            face_helper.get_inverse_affine(None)
+
+            # paste each restored face to the input image
+            results.append(
+                face_helper.paste_faces_to_input_image(upsample_img=img, draw_box=False)
+            )
+
+        return StageResult.from_images(results)

api/onnx_web/server/hacks.py

@@ -150,6 +150,7 @@ def apply_patch_basicsr(server: ServerContext):
 def apply_patch_codeformer(server: ServerContext):
     logger.debug("patching CodeFormer module")
     try:
+        import codeformer.basicsr.utils  # download_util
         import codeformer.facelib.utils.misc
 
         codeformer.facelib.utils.misc.download_pretrained_models = patch_not_impl

api/requirements/base.txt

@@ -18,11 +18,12 @@ onnx==1.15.0
 optimum==1.16.0
 safetensors==0.4.1
 timm==0.9.12
+torchsde==0.2.6
 transformers==4.36.1
 
 #### Upscaling and face correction
 basicsr==1.4.2
-codeformer-perceptor==0.1.2
+# codeformer-perceptor==0.1.2
 facexlib==0.2.5
 gfpgan==1.3.8
 realesrgan==0.3.0