HF needs all model downloads to a special read-write cache dir, also affects encoder.py

vit_captioning/models/encoder.py

@@ -12,7 +12,10 @@ class ViTEncoder(nn.Module):
         super(ViTEncoder, self).__init__()
 
         #weights = ViT_B_16_Weights.DEFAULT
-        self.vit = ViTModel.from_pretrained('google/vit-base-patch16-224-in21k')
+
+        #self.vit = ViTModel.from_pretrained('google/vit-base-patch16-224-in21k')
+        #HF needs all model downloads to a special read-write cache dir
+        self.vit = ViTModel.from_pretrained('google/vit-base-patch16-224-in21k', cache_dir="/tmp")
 
     def forward(self, pixel_values):
 
@@ -30,7 +33,9 @@ from transformers import CLIPModel
 class CLIPEncoder(nn.Module):
     def __init__(self):
         super(CLIPEncoder, self).__init__()
-        self.clip = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+        #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")
 
     def forward(self, pixel_values):
         # ✅ Directly get the pooled image features (already the final representation)