dynamically add choices

app.py

@@ -6,6 +6,8 @@ import pdb
 from types import SimpleNamespace
 import pandas as pd
 import plotly.express as px
+import matplotlib.pyplot as plt
+import numpy as np
 # os.system('git clone https://github.com/EleutherAI/lm-evaluation-harness')
 # os.system('cd lm-evaluation-harness')
 # os.system('pip install -e .')
@@ -95,7 +97,52 @@ def plotly_plot():#(df, x, y, color,title, x_title, y_title):
 #     fig.update_yaxes(title_text=y_title)
 #     return fig
 
+def show_attention_plot(model_name,texts):
+    # 初始化分词器和模型，确保在模型配置中设置 output_attentions=True
+    args=SimpleNamespace(texts=texts,model=model_name)
+    print(f'L60,text:{texts}')
+    rtn_dic=run_get_loss(args)
+    # print(rtn_dic)
+    # pdb.set_trace()
+    # {'logit':logit,'input_ids':input_chunk,'tokenizer':tokenizer,'neg_log_prob_temp':neg_log_prob_temp}
+    # ids, loss =rtn_dic['input_ids'],rtn_dic['loss']#= get_ids_loss(text, tokenizer, model)
+    # notice here is numpy ndarray
+    tokenizer, model = rtn_dic['tokenizer'],rtn_dic['model'] 
+    text = "Here is some text to encode"
+
+    # 使用分词器处理输入文本
+    inputs = tokenizer(text, return_tensors="pt")
+    # 进行前向传播，获取输出
+    outputs = model(**inputs, output_attentions=True)
+
+    # 检查是否成功获得了 attentions
+    if "attentions" in outputs:
+        last_layer_attentions = outputs.attentions[-1]  # 获取最后一层的 attention 矩阵
+        print("Successfully retrieved the attention matrix:", last_layer_attentions.shape)
+    else:
+        pdb.set_trace()
+        print("Attention matrix not found in outputs.")
 
+    # 假设 last_layer_attentions 是我们从模型中提取的注意力矩阵
+    # last_layer_attentions 的形状应该是 [batch_size, num_heads, seq_length, seq_length]
+    # 为了简化，我们这里只查看第一个样本、第一个头的注意力矩阵
+    attention_matrix = last_layer_attentions[0, 0].detach().numpy()
+
+    # 使用 matplotlib 绘制热图
+    plt.figure(figsize=(10, 8))
+    plt.imshow(attention_matrix, cmap='viridis')
+
+    # 添加标题和标签以提高可读性
+    plt.title('Attention Matrix Visualization')
+    plt.xlabel('Tokens in Sequence')
+    plt.ylabel('Tokens in Sequence')
+
+    # 添加颜色条
+    plt.colorbar()
+
+    # 保存图表到文件
+    # plt.savefig('/223040239/medbase/attention_matrix_visualization.png')
+    return plt
 
 
 def color_pipeline(texts=["Hi","FreshEval","!"],  model=None):
@@ -151,6 +198,11 @@ with gr.Blocks() as demo:
 
         # markdown
         gr.Markdown('### How to use this app')
+
+
+        attention_plot=gr.Plot(label='attention plot')
+        see_attention_button = gr.Button("see attention").click(show_attention_plot,inputs=[model_input, text_input],outputs=[attention_plot])
+        
         
 
 

get_loss/get_loss_hf.py

@@ -129,10 +129,12 @@ def load_hf_model(path, cache_path):
                                                         device_map=device,
                                                         trust_remote_code=True,
                                                         cache_dir=cache_path).eval()
+        # output_attentions=True)
     else:
         hf_model = AutoModelForCausalLM.from_pretrained(path,
                                                         device_map=device,
-                                                        trust_remote_code=True).eval()
+                                                        trust_remote_code=True,
+                                                        output_attentions=True).eval()
     hf_tokenizer = AutoTokenizer.from_pretrained(path)
 
     print_model_parameters_in_billions(hf_model)
@@ -227,7 +229,8 @@ def eval_hf_model(model, tokenizer, texts, chunk_size):
 
     # print(f'log probability sum: {sum(data) / len(data):.2f}')
     # print(f'avg tokens: {sum(token_length_list) / len(token_length_list):.0f}')
-        rtn_dic={'logit':logit.cpu().numpy(),'input_ids':input_chunk.cpu().numpy()[0],'loss':loss,'tokenizer':tokenizer,'neg_log_prob_temp':neg_log_prob_temp}
+        rtn_dic={'logit':logit.cpu().numpy(),'input_ids':input_chunk.cpu().numpy()[0],
+                 'loss':loss,'tokenizer':tokenizer,'neg_log_prob_temp':neg_log_prob_temp,'model':model}
         return rtn_dic
 
 

gradio_samples/add_components.py

@@ -0,0 +1,68 @@
+
+# def words():
+#     sentence = "A test of Gradio"
+#     words = sentence.split()
+#     update_show = [gr.Button.update(visible=True, value=w) for w in words]
+#     update_hide = [gr.Button.update(visible=False, value="") for _ in range(10-len(words))]
+#     return update_show + update_hide 
+
+
+# import gradio as gr
+
+# def words():
+#     sentence = "A test of Gradio"
+#     words = sentence.split()
+#     update_show = [gr.Button(visible=True, value=w) for w in words]
+#     update_hide = [gr.Button(visible=False, value="") for _ in range(10-len(words))]
+#     return update_show + update_hide 
+
+# btn_list = []
+    
+# with gr.Blocks() as demo:
+#     with gr.Tab():
+#         for i in range(10):
+#             btn = gr.Button(visible=False)
+#             btn_list.append(btn)
+#     b = gr.Button("Run")
+#     b.click(words, None, btn_list)
+        
+# demo.launch()
+
+
+
+import gradio as gr
+
+def words():
+    sentence = "A test of Gradio"
+    words = sentence.split()
+    update_show = [gr.Textbox(visible=True, value=w,interactive=True) for w in words]
+    update_hide = [gr.Textbox(visible=False, value="",interactive=True) for _ in range(10-len(words))]
+    return update_show + update_hide 
+
+    
+def get_text_content(*btn_list):
+    # make all the input as a list
+    # merge_list 
+    # rtn =[w.value for w in btn_list if w.visible]
+
+    rtn =' '.join([w for w in btn_list ])
+
+    print(rtn)
+    return  rtn
+
+
+btn_list = []
+    
+with gr.Blocks() as demo:
+    with gr.Row():
+        for i in range(10):
+            btn = gr.Textbox(visible=False)
+            btn_list.append(btn)
+    b = gr.Button("Run")
+    b.click(words, None, btn_list)
+
+    b = gr.Button("Get Text Content")
+    output = gr.Textbox()
+    b.click(get_text_content, btn_list, output)
+        
+demo.launch(debug=True)

plot.py → gradio_samples/plot.py

@@ -1,3 +1,4 @@
+# seem not working here...
 import altair
 
 import gradio as gr
@@ -20,7 +21,7 @@ def outbreak(plot_type, r, month, countries, social_distancing):
     # df = pd.DataFrame({"day": x})
     # for country in countries:
     #     df[country] = x ** (r) * (pop_count[country] + 1)
-    df=pd.read_csv('./data/tmp.csv')
+    df=pd.read_csv('../data/ob.csv')
     print(df.head())
     # pdb.set_trace()