fix: correct SEA-RAFT imports (core.raft, core.utils.utils) + huggingface_hub>=0.20

Dockerfile

@@ -26,7 +26,7 @@ RUN pip install --no-cache-dir \
 # SEA-RAFT (dense optical flow — ECCV 2024 Best Paper Candidate)
 # Clone repo so we can import the model architecture
 RUN git clone --depth 1 https://github.com/princeton-vl/SEA-RAFT.git /app/SEA-RAFT && \
-    pip install --no-cache-dir einops scipy huggingface_hub
+    pip install --no-cache-dir einops scipy "huggingface_hub>=0.20"
 
 # Patch lightglue __init__.py so optional models (DISK, ALIKED)
 # don't crash when their kornia deps are incompatible

app.py

@@ -99,19 +99,21 @@ def get_sea_raft_model():
         try:
             import sys
             import os
+            import argparse
+
             # SEA-RAFT is cloned into /app/SEA-RAFT during Docker build
-            sea_raft_dir = os.path.join(os.path.dirname(__file__), "SEA-RAFT")
+            sea_raft_dir = os.environ.get("SEA_RAFT_DIR", "/app/SEA-RAFT")
             if os.path.isdir(sea_raft_dir) and sea_raft_dir not in sys.path:
                 sys.path.insert(0, sea_raft_dir)
+                print(f"[init] Added SEA-RAFT to path: {sea_raft_dir}")
 
-            from sea_raft import RAFT
-            from utils.utils import InputPadder
-            import argparse
+            from core.raft import RAFT
+            from core.utils.utils import InputPadder
 
             print("[init] Loading SEA-RAFT model (Small)...")
             t0 = time.time()
 
-            # Load pre-trained model from HuggingFace
+            # Model args for SEA-RAFT Small variant
             args = argparse.Namespace(
                 mixed_precision=False,
                 num_heads=1,
@@ -123,6 +125,7 @@ def get_sea_raft_model():
                 iters=12,
             )
 
+            # Load pre-trained from HuggingFace hub
             model = RAFT.from_pretrained("MemorySlices/Tartan-C-T-TSKH-spring540x960-S", args=args)
             device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
             model = model.to(device).eval()