Fix: preload CLIP model during Docker build

Dockerfile

@@ -10,6 +10,9 @@ RUN apt-get update && apt-get install -y wget
 WORKDIR /app
 COPY . .
 
+RUN mkdir -p /models/clip && \
+    python3 -c "from transformers import CLIPModel; CLIPModel.from_pretrained('openai/clip-vit-base-patch32').save_pretrained('/models/clip')"
+
 RUN mkdir -p vit_captioning/artifacts && \
     wget https://huggingface.co/datasets/ClemSummer/clip-checkpoints/resolve/main/CLIPEncoder_40epochs_unfreeze12.pth \
     -O vit_captioning/artifacts/CLIPEncoder_40epochs_unfreeze12.pth

vit_captioning/generate.py

@@ -40,7 +40,8 @@ class CaptionGenerator:
             self.encoder_dim = 512
             #self.processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
             #HF needs all model downloads to a special read-write cache dir
-            self.processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32", cache_dir="/tmp")
+            #self.processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32", cache_dir="/tmp")
+            self.processor = CLIPProcessor.from_pretrained("/models/clip")
         else:
             raise ValueError("Unknown model type")
 

vit_captioning/models/encoder.py

@@ -35,7 +35,8 @@ class CLIPEncoder(nn.Module):
         super(CLIPEncoder, self).__init__()
         #self.clip = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
         #HF needs all model downloads to a special read-write cache dir
-        self.clip = CLIPModel.from_pretrained("openai/clip-vit-base-patch32", cache_dir="/tmp")
+        #self.clip = CLIPModel.from_pretrained("openai/clip-vit-base-patch32", cache_dir="/tmp")
+        self.clip = CLIPModel.from_pretrained("/models/clip")
 
     def forward(self, pixel_values):
         # ✅ Directly get the pooled image features (already the final representation)