Use pipeline for interpretation

app.py

@@ -1,5 +1,6 @@
 import gradio as gr
 import shap
+import transformers
 
 import os
 import re
@@ -9,6 +10,8 @@ import tempfile
 
 model = gr.load("ejschwartz/oo-method-test-model-bylibrary", src="models")
 
+model_interp = transformers.pipeline("text-classification", "ejschwartz/oo-method-test-model-bylibrary")
+
 def get_all_dis(bname, addrs=None):
 
     anafile = tempfile.NamedTemporaryFile(prefix=os.path.basename(bname) + "_", suffix=".bat_ana")
@@ -122,17 +125,15 @@ with gr.Blocks() as demo:
                 clazz: gr.Label.update(top_k)
         }
     
+    # XXX: Ideally we'd use the gr.load model, which uses the huggingface
+    # inference API.  But shap library appears to use information in the
+    # transformers pipeline, and I don't feel like figuring out how to
+    # reimplement that, so we'll just use a regular transformers pipeline here
+    # for interpretation.
     def interpretation_function(text):
 
-        def model_wrap(input):
-            print(f"model_wrap input = {input}", file=sys.stderr)
-            return model.fn(input)['confidences']
-
-        print(text, file=sys.stderr)
-        out = model_wrap(text)
-        print(out, file=sys.stderr)
-        explainer = shap.Explainer(model_wrap)
-        shap_values = explainer(text)
+        explainer = shap.Explainer(model_interp)
+        shap_values = explainer([text])
 
         # Dimensions are (batch size, text size, number of classes)
         # Since we care about positive sentiment, use index 1

requirements.txt

@@ -1,4 +1,5 @@
 gradio
 pandas
 numpy >= 1.22.4
-shap
+shap
+transformers[torch]