optimizing app

app.py

@@ -14,6 +14,25 @@ _plotly_config={'displayModeBar': False}
 from sentiment_clf_helper import classify_sentiment,create_onnx_model_sentiment,classify_sentiment_onnx
 from zeroshot_clf_helper import zero_shot_classification,create_onnx_model_zs,zero_shot_classification_onnx
 
+import yaml
+def read_yaml(file_path):
+    with open(file_path, "r") as f:
+        return yaml.safe_load(f)
+
+config = read_yaml('config.yaml')
+
+sent_chkpt=config['SENTIMENT_CLF']['sent_chkpt']
+sent_mdl_dir=config['SENTIMENT_CLF']['sent_mdl_dir']
+sent_onnx_mdl_dir=config['SENTIMENT_CLF']['sent_onnx_mdl_dir']
+sent_onnx_mdl_name=config['SENTIMENT_CLF']['sent_onnx_mdl_name']
+sent_onnx_quant_mdl_name=config['SENTIMENT_CLF']['sent_onnx_quant_mdl_name']
+
+zs_chkpt=config['ZEROSHOT_CLF']['zs_chkpt']
+zs_mdl_dir=config['ZEROSHOT_CLF']['zs_mdl_dir']
+zs_onnx_mdl_dir=config['ZEROSHOT_CLF']['zs_onnx_mdl_dir']
+zs_onnx_mdl_name=config['ZEROSHOT_CLF']['zs_onnx_mdl_name']
+zs_onnx_quant_mdl_name=config['ZEROSHOT_CLF']['zs_onnx_quant_mdl_name']
+
 
 st.set_page_config(  # Alternate names: setup_page, page, layout
     layout="wide",  # Can be "centered" or "wide". In the future also "dashboard", etc.
@@ -97,24 +116,27 @@ if select_task=='README':
 
 ############### Pre-Download & instantiate objects for sentiment analysis *********************** START **********************
 
-sent_chkpt = "distilbert-base-uncased-finetuned-sst-2-english"
-sent_model_dir="sentiment_model_dir"
-#create model/token dir for sentiment classification
-create_model_dir(chkpt=sent_chkpt, model_dir=sent_model_dir)
+# #create model/token dir for sentiment classification for faster inference
+# create_model_dir(chkpt=sent_chkpt, model_dir=sent_mdl_dir)
 
 
 @st.cache(allow_output_mutation=True, suppress_st_warning=True, max_entries=None, ttl=None)
-def sentiment_task_selected(task,sent_model_dir=sent_model_dir):
+def sentiment_task_selected(task,
+                            sent_chkpt=sent_chkpt,
+                            sent_mdl_dir=sent_mdl_dir,
+                            sent_onnx_mdl_dir=sent_onnx_mdl_dir,
+                            sent_onnx_mdl_name=sent_onnx_mdl_name,
+                            sent_onnx_quant_mdl_name=sent_onnx_quant_mdl_name):
     #model & tokenizer initialization for normal sentiment classification
-    model_sentiment=AutoModelForSequenceClassification.from_pretrained(sent_model_dir)
-    tokenizer_sentiment=AutoTokenizer.from_pretrained(sent_model_dir)
+    model_sentiment=AutoModelForSequenceClassification.from_pretrained(sent_chkpt)
+    tokenizer_sentiment=AutoTokenizer.from_pretrained(sent_chkpt)
 
     # create onnx model for sentiment classification
     create_onnx_model_sentiment(_model=model_sentiment, _tokenizer=tokenizer_sentiment)
 
     #create inference session
-    sentiment_session = ort.InferenceSession("sent_clf_onnx_dir/sentiment_classifier_onnx.onnx")
-    sentiment_session_quant = ort.InferenceSession("sent_clf_onnx_dir/sentiment_classifier_onnx_quant.onnx")
+    sentiment_session = ort.InferenceSession(f"{sent_onnx_mdl_dir}/{sent_onnx_mdl_name}")
+    sentiment_session_quant = ort.InferenceSession(f"{sent_onnx_mdl_dir}/{sent_onnx_quant_mdl_name}")
 
     return model_sentiment,tokenizer_sentiment,sentiment_session,sentiment_session_quant
 
@@ -123,26 +145,31 @@ def sentiment_task_selected(task,sent_model_dir=sent_model_dir):
 
 ############### Pre-Download & instantiate objects for Zero shot clf *********************** START **********************
 
-zs_chkpt = "valhalla/distilbart-mnli-12-1"
-zs_model_dir = "zs_model_dir"
-# create model/token dir for zeroshot clf
-create_model_dir(chkpt=zs_chkpt, model_dir=zs_model_dir)
+# # create model/token dir for zeroshot clf
+# create_model_dir(chkpt=zs_chkpt, model_dir=zs_mdl_dir)
 
 
 @st.cache(allow_output_mutation=True, suppress_st_warning=True, max_entries=None, ttl=None)
-def zs_task_selected(task, zs_model_dir=zs_model_dir,onnx_dir='zeroshot_onnx_dir'):
-    #model & tokenizer initialization for normal ZS classification
-    model_zs=AutoModelForSequenceClassification.from_pretrained(zs_model_dir)
-    tokenizer_zs=AutoTokenizer.from_pretrained(zs_model_dir)
-
-    # ceate onnx model for zeroshot
+def zs_task_selected(task,
+                     zs_chkpt=zs_chkpt ,
+                     zs_mdl_dir=zs_mdl_dir,
+                     zs_onnx_mdl_dir=zs_onnx_mdl_dir,
+                     zs_onnx_mdl_name=zs_onnx_mdl_name,
+                     zs_onnx_quant_mdl_name=zs_onnx_quant_mdl_name):
+
+    ##model & tokenizer initialization for normal ZS classification
+    # model_zs=AutoModelForSequenceClassification.from_pretrained(zs_chkpt)
+    # we just need tokenizer for inference and not model since onnx model is already saved
+    tokenizer_zs=AutoTokenizer.from_pretrained(zs_chkpt)
+
+    # create onnx model for zeroshot
     create_onnx_model_zs()
 
     #create inference session from onnx model
-    zs_session = ort.InferenceSession(f"{onnx_dir}/model.onnx")
-    zs_session_quant = ort.InferenceSession(f"{onnx_dir}/model_quant.onnx")
+    zs_session = ort.InferenceSession(f"{zs_onnx_mdl_dir}/{zs_onnx_mdl_name}")
+    zs_session_quant = ort.InferenceSession(f"{zs_onnx_mdl_dir}/{zs_onnx_quant_mdl_name}")
 
-    return model_zs,tokenizer_zs,zs_session,zs_session_quant
+    return tokenizer_zs,zs_session,zs_session_quant
 
 ############## Pre-Download & instantiate objects for Zero shot analysis ********************* END **********************************
 
@@ -256,7 +283,7 @@ if select_task == 'Detect Sentiment':
 if select_task=='Zero Shot Classification':
 
     t1=time.time()
-    model_zs,tokenizer_zs,zs_session,zs_session_quant = zs_task_selected(task=select_task)
+    tokenizer_zs,zs_session,zs_session_quant = zs_task_selected(task=select_task)
     t2 = time.time()
     st.write(f"Total time to load Model is {(t2-t1)*1000:.1f} ms")
 
@@ -267,29 +294,16 @@ if select_task=='Zero Shot Classification':
     c1,c2,c3,c4=st.columns(4)
 
     with c1:
-        response1=st.button("Normal runtime")
+        response1=st.button("ONNX runtime")
     with c2:
-        response2=st.button("ONNX runtime")
-    with c3:
-        response3=st.button("ONNX runtime with Quantization")
-    with c4:
-        response4 = st.button("Simulate 10 runs each runtime")
+        response2=st.button("ONNX runtime Quantized")
+    # with c3:
+    #     response3=st.button("ONNX runtime with Quantization")
+    # with c4:
+    #     response4 = st.button("Simulate 10 runs each runtime")
 
-    if any([response1,response2,response3,response4]):
+    if any([response1,response2]):
         if response1:
-            start=time.time()
-            df_output = zero_shot_classification(input_texts, input_lables,model=model_zs,tokenizer=tokenizer_zs)
-            end=time.time()
-            st.write("")
-            st.write(f"Time taken for computation {(end-start)*1000:.1f} ms")
-            fig = px.bar(x='Probability',
-                         y='labels',
-                         text='Probability',
-                         data_frame=df_output,
-                         title='Zero Shot Normalized Probabilities')
-
-            st.plotly_chart(fig, config=_plotly_config)
-        elif response2:
             start = time.time()
             df_output=zero_shot_classification_onnx(premise=input_texts,labels=input_lables,_session=zs_session,_tokenizer=tokenizer_zs)
             end=time.time()
@@ -303,7 +317,7 @@ if select_task=='Zero Shot Classification':
                          title='Zero Shot Normalized Probabilities')
 
             st.plotly_chart(fig,config=_plotly_config)
-        elif response3:
+        elif response2:
             start = time.time()
             df_output = zero_shot_classification_onnx(premise=input_texts, labels=input_lables, _session=zs_session_quant,
                                                    _tokenizer=tokenizer_zs)
@@ -317,53 +331,6 @@ if select_task=='Zero Shot Classification':
                          title='Zero Shot Normalized Probabilities')
 
             st.plotly_chart(fig, config=_plotly_config)
-        elif response4:
-            normal_runtime = []
-            for i in range(100):
-                start = time.time()
-                _ = zero_shot_classification(input_texts, input_lables,model=model_zs,tokenizer=tokenizer_zs)
-                end = time.time()
-                t = (end - start) * 1000
-                normal_runtime.append(t)
-            normal_runtime = np.clip(normal_runtime, 50, 400)
-
-            onnx_runtime = []
-            for i in range(100):
-                start = time.time()
-                _ = zero_shot_classification_onnx(premise=input_texts, labels=input_lables, _session=zs_session,
-                                                   _tokenizer=tokenizer_zs)
-                end = time.time()
-                t = (end - start) * 1000
-                onnx_runtime.append(t)
-            onnx_runtime = np.clip(onnx_runtime, 50, 200)
-
-            onnx_runtime_quant = []
-            for i in range(100):
-                start = time.time()
-                _ = zero_shot_classification_onnx(premise=input_texts, labels=input_lables, _session=zs_session_quant,
-                                                   _tokenizer=tokenizer_zs)
-                end = time.time()
-
-                t = (end - start) * 1000
-                onnx_runtime_quant.append(t)
-            onnx_runtime_quant = np.clip(onnx_runtime_quant, 50, 200)
-
-            temp_df = pd.DataFrame({'Normal Runtime (ms)': normal_runtime,
-                                    'ONNX Runtime (ms)': onnx_runtime,
-                                    'ONNX Quant Runtime (ms)': onnx_runtime_quant})
-
-            from plotly.subplots import make_subplots
-
-            fig = make_subplots(rows=1, cols=3, start_cell="bottom-left",
-                                subplot_titles=['Normal Runtime', 'ONNX Runtime', 'ONNX Runtime with Quantization'])
-
-            fig.add_trace(go.Histogram(x=temp_df['Normal Runtime (ms)']), row=1, col=1)
-            fig.add_trace(go.Histogram(x=temp_df['ONNX Runtime (ms)']), row=1, col=2)
-            fig.add_trace(go.Histogram(x=temp_df['ONNX Quant Runtime (ms)']), row=1, col=3)
-            fig.update_layout(height=400, width=1000,
-                              title_text="10 Simulations of different Runtimes",
-                              showlegend=False)
-            st.plotly_chart(fig, config=_plotly_config)
         else:
             pass