upload

pages/summary_acc_250116.py

@@ -27,10 +27,10 @@ LONG_CAPTIONS = [
     'Jina-CLIP:jina-clip-v1:jinaai', 'Jina-CLIP:jina-clip-v2:jinaai'
 ]
 COMPOSITIONALITY = [
-    'StructuredCLIP:NegCLIP-ViT-B-32:coco-ft', 'StructuredCLIP:CE-CLIP-ViT-B-32:coco-ft',
-    'StructuredCLIP:DAC-LLM-ViT-B-32:cc3m-ft', 'StructuredCLIP:DAC-SAM-ViT-B-32:cc3m-ft',
-    'FSC-CLIP:fsc-clip-ViT-B-32:laioncoco-ft', 'FSC-CLIP:fsc-clip-ViT-B-16:laioncoco-ft',
-    'FSC-CLIP:fsc-clip-ViT-L-14:laioncoco-ft'
+    'OpenCLIP:ViT-B-32:openai', 'StructuredCLIP:NegCLIP-ViT-B-32:coco-ft',
+    'StructuredCLIP:CE-CLIP-ViT-B-32:coco-ft', 'StructuredCLIP:DAC-LLM-ViT-B-32:cc3m-ft',
+    'StructuredCLIP:DAC-SAM-ViT-B-32:cc3m-ft', 'FSC-CLIP:fsc-clip-ViT-B-32:laioncoco-ft',
+    'FSC-CLIP:fsc-clip-ViT-B-16:laioncoco-ft', 'FSC-CLIP:fsc-clip-ViT-L-14:laioncoco-ft'
 ]
 
 MODEL_GROUPS = {
@@ -133,6 +133,23 @@ def main():
     df.columns = [ast.literal_eval(col) for col in df.columns]
     for group, model_names in MODEL_GROUPS.items():
         st.markdown(f"## {group} models")
+        if group == "short_captions":
+            st.markdown(
+                "- **Length group**: 이미 short group부터, 80<(Num_tokens)<120. 중간에 문장 더해졌으면 60-70%정도 맞추고, 끝에 문장 더해졌으면 애초에 added sentence encoding 불가 -> accuracy 0%"
+            )
+            st.markdown(
+                "- **neg_target**: description의 끝 (=background)에 sentence 더해진 경우 accuracy 0%"
+            )
+            st.markdown("- **neg_type**: contradictory sentence가 모델 입장에서 맞추기 더 어려움")
+
+        if group == "long_captions":
+            st.markdown(
+                "- **Length group**: 모델의 context length에 성능 심하게 dependent함. DreamLIP: 77, CLIPS: 80, LoTLIP: 128, Recap-CLIP: 128, LongCLIP: 248, Jina-CLIP: 512"
+            )
+            st.markdown("- **neg_target**: 여전히 background level에서 sentence 더해진게 전반적으로 어려움")
+            st.markdown("- **neg_type**: contradictory sentence가 모델 입장에서 맞추기 더 어려움")
+        if group == "compositionality":
+            st.markdown("- context length 77의 한계. Hard Negative Caption으로 Fine-tuning 하면 일부 좋아짐")
         get_model_key_from_df(df, model_names)