File size: 18,005 Bytes
e56ce5e
 
 
 
0242b2e
e56ce5e
0242b2e
1e9f08c
0242b2e
 
50cfa80
e56ce5e
0242b2e
 
e56ce5e
 
0242b2e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
44fc84f
 
 
 
 
 
 
 
 
 
 
 
0242b2e
e56ce5e
 
0242b2e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
43f46f9
942d759
0242b2e
e56ce5e
 
 
 
 
 
43f46f9
942d759
e56ce5e
 
 
 
 
 
 
 
0242b2e
 
 
 
 
70d91d0
4ab3b97
0242b2e
 
 
 
 
 
 
e56ce5e
0242b2e
 
 
 
 
e56ce5e
0242b2e
 
e56ce5e
0242b2e
e56ce5e
0242b2e
 
e56ce5e
 
 
 
 
0242b2e
 
e56ce5e
f050e78
 
e56ce5e
 
1b9fec9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e56ce5e
 
 
44fc84f
e56ce5e
 
 
942d759
 
1e9f08c
942d759
 
e56ce5e
 
 
 
 
 
0242b2e
e56ce5e
 
042b8a4
e56ce5e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
43f46f9
 
 
 
 
e56ce5e
 
43f46f9
 
 
 
e56ce5e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0242b2e
1b9fec9
e56ce5e
 
 
 
 
 
 
6ad0c29
e56ce5e
 
 
 
 
44fc84f
 
e56ce5e
 
6ad0c29
e56ce5e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
102b976
1e9f08c
e56ce5e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
import time
import streamlit as st
import string
from io import StringIO 
import pdb
import json
from twc_embeddings import HFModel,SimCSEModel,SGPTModel
import torch


MAX_INPUT = 100


from transformers import BertTokenizer, BertForMaskedLM

model_names = [

            {   "name":"sentence-transformers/all-MiniLM-L6-v2", 
                "model":"sentence-transformers/all-MiniLM-L6-v2",
                "fork_url":"https://github.com/taskswithcode/sentence_similarity_hf_model",
                "orig_author_url":"https://github.com/UKPLab",
                "orig_author":"Ubiquitous Knowledge Processing Lab",
                "sota_info": {   
                                 "task":"Over 3.8  million downloads from huggingface",
                                 "sota_link":"https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2"
                            },
                "paper_url":"https://arxiv.org/abs/1908.10084",
                "mark":True,
                "class":"HFModel"},
            {   "name":"sentence-transformers/paraphrase-MiniLM-L6-v2", 
                "model":"sentence-transformers/paraphrase-MiniLM-L6-v2",
                "fork_url":"https://github.com/taskswithcode/sentence_similarity_hf_model",
                "orig_author_url":"https://github.com/UKPLab",
                "orig_author":"Ubiquitous Knowledge Processing Lab",
                "sota_info": {   
                                 "task":"Over 2.4 million downloads from huggingface",
                                 "sota_link":"https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2"
                            },
                "paper_url":"https://arxiv.org/abs/1908.10084",
                "mark":True,
                "class":"HFModel"},
            {   "name":"sentence-transformers/bert-base-nli-mean-tokens", 
                "model":"sentence-transformers/bert-base-nli-mean-tokens",
                "fork_url":"https://github.com/taskswithcode/sentence_similarity_hf_model",
                "orig_author_url":"https://github.com/UKPLab",
                "orig_author":"Ubiquitous Knowledge Processing Lab",
                "sota_info": {   
                                 "task":"Over 700,000 downloads from huggingface",
                                 "sota_link":"https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2"
                            },
                "paper_url":"https://arxiv.org/abs/1908.10084",
                "mark":True,
                "class":"HFModel"},
            {   "name":"sentence-transformers/all-mpnet-base-v2", 
                "model":"sentence-transformers/all-mpnet-base-v2",
                "fork_url":"https://github.com/taskswithcode/sentence_similarity_hf_model",
                "orig_author_url":"https://github.com/UKPLab",
                "orig_author":"Ubiquitous Knowledge Processing Lab",
                "sota_info": {   
                                 "task":"Over 500,000 downloads from huggingface",
                                 "sota_link":"https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2"
                            },
                "paper_url":"https://arxiv.org/abs/1908.10084",
                "mark":True,
                "class":"HFModel"},
            {   "name":"sentence-transformers/all-MiniLM-L12-v2",
                "model":"sentence-transformers/all-MiniLM-L12-v2",
                "fork_url":"https://github.com/taskswithcode/sentence_similarity_hf_model",
                "orig_author_url":"https://github.com/UKPLab",
                "orig_author":"Ubiquitous Knowledge Processing Lab",
                "sota_info": {   
                                 "task":"Over 500,000 downloads from huggingface",
                                 "sota_link":"https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2"
                            },
                "paper_url":"https://arxiv.org/abs/1908.10084",
                "mark":True,
                "class":"HFModel"},

            {   "name":"SGPT-125M", 
                "model":"Muennighoff/SGPT-125M-weightedmean-nli-bitfit",
                "fork_url":"https://github.com/taskswithcode/sgpt",
                "orig_author_url":"https://github.com/Muennighoff",
                "orig_author":"Niklas Muennighoff",
                "sota_info": {   
                                 "task":"#1 in multiple information retrieval & search tasks(smaller variant)",
                                 "sota_link":"https://paperswithcode.com/paper/sgpt-gpt-sentence-embeddings-for-semantic",
                            },
                "paper_url":"https://arxiv.org/abs/2202.08904v5",
                "mark":True,
                "class":"SGPTModel"},
            {   "name":"SGPT-1.3B",
                "model": "Muennighoff/SGPT-1.3B-weightedmean-msmarco-specb-bitfit",
                "fork_url":"https://github.com/taskswithcode/sgpt",
                "orig_author_url":"https://github.com/Muennighoff",
                "orig_author":"Niklas Muennighoff",
                "sota_info": {   
                                 "task":"#1 in multiple information retrieval & search tasks(smaller variant)",
                                 "sota_link":"https://paperswithcode.com/paper/sgpt-gpt-sentence-embeddings-for-semantic",
                            },
                "paper_url":"https://arxiv.org/abs/2202.08904v5",
                "Note":"If this large model takes too long or fails to load , try this ",
                "alt_url":"http://www.taskswithcode.com/sentence_similarity/",
                "mark":True,
                "class":"SGPTModel"},
            {   "name":"SGPT-5.8B",
                "model": "Muennighoff/SGPT-5.8B-weightedmean-msmarco-specb-bitfit" ,
                "fork_url":"https://github.com/taskswithcode/sgpt",
                "orig_author_url":"https://github.com/Muennighoff",
                "orig_author":"Niklas Muennighoff",
                "Note":"If this large model takes too long or fails to load , try this ",
                "alt_url":"http://www.taskswithcode.com/sentence_similarity/",
                "sota_info": {   
                                 "task":"#1 in multiple information retrieval & search tasks",
                                 "sota_link":"https://paperswithcode.com/paper/sgpt-gpt-sentence-embeddings-for-semantic",
                            },
                "paper_url":"https://arxiv.org/abs/2202.08904v5",
                "mark":True,
                "class":"SGPTModel"},

            {   "name":"SIMCSE-large" ,
                "model":"princeton-nlp/sup-simcse-roberta-large",
                "fork_url":"https://github.com/taskswithcode/SimCSE",
                "orig_author_url":"https://github.com/princeton-nlp",
                "orig_author":"Princeton Natural Language Processing",
                "Note":"If this large model takes too long or fails to load , try this ",
                "alt_url":"http://www.taskswithcode.com/sentence_similarity/",
                "sota_info": {   
                                 "task":"Within top 10 in multiple semantic textual similarity tasks",
                                 "sota_link":"https://paperswithcode.com/paper/simcse-simple-contrastive-learning-of"
                            },
                "paper_url":"https://arxiv.org/abs/2104.08821v4",
                "mark":True,
                "class":"SimCSEModel","sota_link":"https://paperswithcode.com/sota/semantic-textual-similarity-on-sick"},

            {  "name":"SIMCSE-base" ,
                "model":"princeton-nlp/sup-simcse-roberta-base",
                "fork_url":"https://github.com/taskswithcode/SimCSE",
                "orig_author_url":"https://github.com/princeton-nlp",
                "orig_author":"Princeton Natural Language Processing",
                "sota_info": {   
                                 "task":"Within top 10 in multiple semantic textual similarity tasks(smaller variant)",
                                 "sota_link":"https://paperswithcode.com/paper/simcse-simple-contrastive-learning-of"
                            },
                "paper_url":"https://arxiv.org/abs/2104.08821v4",
                "mark":True,
                "class":"SimCSEModel","sota_link":"https://paperswithcode.com/sota/semantic-textual-similarity-on-sick"},


            ]





example_file_names = {
"Machine learning terms (30+ phrases)": "small_test.txt",
"Customer feedback mixed with noise (50+ sentences)":"larger_test.txt"
}

view_count_file = "view_count.txt"

def get_views():
    ret_val = 0
    if ("view_count" not in st.session_state):
        try:
           data = int(open(view_count_file).read().strip("\n"))
        except:
           data = 0
        data += 1
        ret_val = data
        st.session_state["view_count"] = data
        with open(view_count_file,"w") as fp:
            fp.write(str(data))
    else:
        ret_val = st.session_state["view_count"]
    return "{:,}".format(ret_val)
        


def construct_model_info_for_display():
    options_arr  = []
    markdown_str = f"<div style=\"font-size:16px; color: #2f2f2f; text-align: left\"><br/><b>Models evaluated ({len(model_names)})</b></div>"
    for node in model_names:
        options_arr .append(node["name"])
        if (node["mark"] == True):
            markdown_str += f"<div style=\"font-size:16px; color: #5f5f5f; text-align: left\">&nbsp;•&nbsp;Model:&nbsp;<a href=\'{node['paper_url']}\' target='_blank'>{node['name']}</a><br/>&nbsp;&nbsp;&nbsp;&nbsp;Code released by:&nbsp;<a href=\'{node['orig_author_url']}\' target='_blank'>{node['orig_author']}</a><br/>&nbsp;&nbsp;&nbsp;&nbsp;Model info:&nbsp;<a href=\'{node['sota_info']['sota_link']}\' target='_blank'>{node['sota_info']['task']}</a></div>"
            if ("Note" in node):
                markdown_str += f"<div style=\"font-size:16px; color: #a91212; text-align: left\">&nbsp;&nbsp;&nbsp;&nbsp;{node['Note']}<a href=\'{node['alt_url']}\' target='_blank'>link</a></div>"
            markdown_str += "<div style=\"font-size:16px; color: #5f5f5f; text-align: left\"><br/></div>"
        
    markdown_str += "<div style=\"font-size:12px; color: #9f9f9f; text-align: left\"><b>Note:</b><br/>•&nbsp;Uploaded files are loaded into non-persistent memory for the duration of the computation. They are not saved</div>"
    limit = "{:,}".format(MAX_INPUT)
    markdown_str += f"<div style=\"font-size:12px; color: #9f9f9f; text-align: left\">•&nbsp;User uploaded file has a maximum limit of {limit} sentences.</div>"
    return options_arr,markdown_str


st.set_page_config(page_title='TWC - Compare popular/state-of-the-art models for Sentence Similarity task', page_icon="logo.jpg", layout='centered', initial_sidebar_state='auto',
            menu_items={
             'About': 'This app was created by taskswithcode. http://taskswithcode.com'
             
              })
col,pad = st.columns([85,15])

with col:
    st.image("long_form_logo_with_icon.png")


@st.experimental_memo
def load_model(model_name):
    try:
        ret_model = None
        for node in model_names:
            if (model_name.startswith(node["name"])):
                obj_class = globals()[node["class"]]
                ret_model = obj_class()
                ret_model.init_model(node["model"])
        assert(ret_model is not None)
    except Exception as e:
        st.error("Unable to load model:" + model_name + " " +  str(e))
        pass
    return ret_model

  
@st.experimental_memo
def cached_compute_similarity(sentences,_model,model_name,main_index):
    texts,embeddings = _model.compute_embeddings(sentences,is_file=False)
    results = _model.output_results(None,texts,embeddings,main_index)
    return results


def uncached_compute_similarity(sentences,_model,model_name,main_index):
    with st.spinner('Computing vectors for sentences'):
        texts,embeddings = _model.compute_embeddings(sentences,is_file=False)
        results = _model.output_results(None,texts,embeddings,main_index)
    #st.success("Similarity computation complete")
    return results

def get_model_info(model_name):
    for node in model_names:
        if (model_name == node["name"]):
            return node

def run_test(model_name,sentences,display_area,main_index,user_uploaded):
    display_area.text("Loading model:" + model_name)
    model_info = get_model_info(model_name)
    if ("Note" in model_info):
        fail_link = f"{model_info['Note']} [link]({model_info['alt_url']})"
        display_area.write(fail_link)
    model = load_model(model_name)
    display_area.text("Model " + model_name  + " load complete")
    try:
            if (user_uploaded):
                results = uncached_compute_similarity(sentences,model,model_name,main_index)
            else:
                display_area.text("Computing vectors for sentences")
                results = cached_compute_similarity(sentences,model,model_name,main_index)
                display_area.text("Similarity computation complete")
            return results
            
    except Exception as e:
        st.error("Some error occurred during prediction" + str(e))
        st.stop()
    return {}



    

def display_results(orig_sentences,main_index,results,response_info):
    main_sent = f"<div style=\"font-size:14px; color: #2f2f2f; text-align: left\">{response_info}<br/><br/></div>"
    main_sent += "<div style=\"font-size:14px; color: #6f6f6f; text-align: left\">Results sorted by cosine distance. Closest(1) to furthest(-1) away from main sentence</div>"
    main_sent += f"<div style=\"font-size:16px; color: #2f2f2f; text-align: left\"><b>Main sentence:</b>&nbsp;&nbsp;{orig_sentences[main_index]}</div>"
    body_sent = []
    download_data = {}
    for key in results:
        index = orig_sentences.index(key) + 1
        body_sent.append(f"<div style=\"font-size:16px; color: #2f2f2f; text-align: left\">{index}]&nbsp;{key}&nbsp;&nbsp;&nbsp;<b>{results[key]:.2f}</b></div>")
        download_data[key] =  f"{results[key]:.2f}" 
    main_sent = main_sent + "\n" + '\n'.join(body_sent)
    st.markdown(main_sent,unsafe_allow_html=True)
    st.session_state["download_ready"] = json.dumps(download_data,indent=4)


def init_session():
    st.session_state["download_ready"] = None    
    st.session_state["model_name"] = "ss_test"
    st.session_state["main_index"] = 1
    st.session_state["file_name"] = "default"
 
def main():
  init_session()
  st.markdown("<h5 style='text-align: center;'>Compare popular/state-of-the-art models for Sentence Similarity task</h5>", unsafe_allow_html=True)
  st.markdown(f"<div style='color: #9f9f9f; text-align: right'>views:&nbsp;{get_views()}</div>", unsafe_allow_html=True)


  try:
      
      
      with st.form('twc_form'):

        uploaded_file = st.file_uploader("Step 1. Upload text file(one sentence in a line) or choose an example text file below", type=".txt")

        selected_file_index = st.selectbox(label='Example files ',  
                    options = list(dict.keys(example_file_names)), index=0,  key = "twc_file")
        st.write("")
        options_arr,markdown_str = construct_model_info_for_display()
        selection_label = 'Step 2. Select Model'
        selected_model = st.selectbox(label=selection_label,  
                    options = options_arr, index=0,  key = "twc_model")
        st.write("")
        main_index = st.number_input('Step 3. Enter index of sentence in file to make it the main sentence',value=1,min_value = 1)
        st.write("")
        submit_button = st.form_submit_button('Run')

        
        input_status_area = st.empty()
        display_area = st.empty()
        if submit_button:
            start = time.time()
            if uploaded_file is not None:
                st.session_state["file_name"]  = uploaded_file.name
                sentences = StringIO(uploaded_file.getvalue().decode("utf-8")).read()
            else:
                st.session_state["file_name"]  = example_file_names[selected_file_index]
                sentences = open(example_file_names[selected_file_index]).read()
            sentences = sentences.split("\n")[:-1]
            if (len(sentences) < main_index):
                main_index = len(sentences)
                st.info("Selected sentence index is larger than number of sentences in file. Truncating to " + str(main_index)) 
            if (len(sentences) > MAX_INPUT):
                st.info(f"Input sentence count exceeds maximum sentence limit. First {MAX_INPUT} out of {len(sentences)} sentences chosen")
                sentences = sentences[:MAX_INPUT]
            st.session_state["model_name"] = selected_model
            st.session_state["main_index"] = main_index
            results = run_test(selected_model,sentences,display_area,main_index - 1,(uploaded_file is not None))
            display_area.empty()
            with display_area.container():
                device = 'GPU' if torch.cuda.is_available() else 'CPU'
                response_info = f"Computation time on {device}: {time.time() - start:.2f} secs for {len(sentences)} sentences"
                display_results(sentences,main_index - 1,results,response_info)
                #st.json(results)
      st.download_button(
         label="Download results as json",
         data= st.session_state["download_ready"] if st.session_state["download_ready"] != None else "",
         disabled = False if st.session_state["download_ready"] != None else True,
         file_name= (st.session_state["model_name"] + "_" +  str(st.session_state["main_index"]) + "_" + '_'.join(st.session_state["file_name"].split(".")[:-1]) + ".json").replace("/","_"),
         mime='text/json',
         key ="download" 
        )
      
      

  except Exception as e:
    st.error("Some error occurred during loading" + str(e))
    st.stop()  
	
  st.markdown(markdown_str, unsafe_allow_html=True)
  
 

if __name__ == "__main__":
   main()