Spaces:
Sleeping
Sleeping
Pclanglais
commited on
Commit
•
0dfb412
1
Parent(s):
8252e5b
Update app.py
Browse files
app.py
CHANGED
@@ -1,5 +1,6 @@
|
|
|
|
1 |
import re
|
2 |
-
from transformers import
|
3 |
from vllm import LLM, SamplingParams
|
4 |
import torch
|
5 |
import gradio as gr
|
@@ -7,169 +8,82 @@ import json
|
|
7 |
import os
|
8 |
import shutil
|
9 |
import requests
|
10 |
-
import
|
11 |
import pandas as pd
|
12 |
-
from
|
13 |
-
from
|
14 |
-
from sklearn.metrics.pairwise import cosine_similarity
|
15 |
-
|
16 |
-
from transformers import AutoModelForSequenceClassification
|
17 |
|
|
|
18 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
19 |
|
20 |
-
|
21 |
-
embedding_model = BGEM3FlagModel('BAAI/bge-m3',
|
22 |
-
use_fp16=True) # Setting use_fp16 to True speeds up computation with a slight performance degradation
|
23 |
-
|
24 |
-
embeddings = np.load("embeddings_albert_tchap.npy")
|
25 |
-
embeddings_data = pd.read_json("embeddings_albert_tchap.json")
|
26 |
-
embeddings_text = embeddings_data["text_with_context"].tolist()
|
27 |
-
|
28 |
-
#Importing the classifier/router (deberta)
|
29 |
-
classifier_model = AutoModelForSequenceClassification.from_pretrained("AgentPublic/chatrag-deberta")
|
30 |
-
classifier_tokenizer = AutoTokenizer.from_pretrained("AgentPublic/chatrag-deberta")
|
31 |
-
|
32 |
-
#Importing the actual generative LLM (llama-based)
|
33 |
-
model_name = "Pclanglais/Tchap"
|
34 |
-
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
35 |
-
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16)
|
36 |
-
model = model.to('cuda:0')
|
37 |
-
|
38 |
-
system_prompt = "<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\nTu es Albert, l'agent conversationnel des services publics qui peut décrire des documents de référence ou aider à des tâches de rédaction<|eot_id|>"
|
39 |
-
source_text = "Les sources utilisées par Albert-Tchap vont apparaître ici'"
|
40 |
-
|
41 |
-
|
42 |
-
#Function to guess whether we use the RAG or not.
|
43 |
-
def classification_chatrag(query):
|
44 |
-
print(query)
|
45 |
-
encoding = classifier_tokenizer(query, return_tensors="pt")
|
46 |
-
encoding = {k: v.to(classifier_model.device) for k,v in encoding.items()}
|
47 |
-
|
48 |
-
outputs = classifier_model(**encoding)
|
49 |
-
|
50 |
-
logits = outputs.logits
|
51 |
-
logits.shape
|
52 |
-
|
53 |
-
# apply sigmoid + threshold
|
54 |
-
sigmoid = torch.nn.Sigmoid()
|
55 |
-
probs = sigmoid(logits.squeeze().cpu())
|
56 |
-
predictions = np.zeros(probs.shape)
|
57 |
-
|
58 |
-
# Extract the float value from the tensor
|
59 |
-
float_value = round(probs.item()*100)
|
60 |
-
|
61 |
-
print(float_value)
|
62 |
-
|
63 |
-
if float_value > 50:
|
64 |
-
status = True
|
65 |
-
print("We activate RAG")
|
66 |
-
else:
|
67 |
-
status = False
|
68 |
-
print("We remove RAG")
|
69 |
-
return status
|
70 |
-
|
71 |
-
#Vector search over the database
|
72 |
-
def vector_search(sentence_query):
|
73 |
-
|
74 |
-
query_embedding = embedding_model.encode(sentence_query,
|
75 |
-
batch_size=12,
|
76 |
-
max_length=256, # If you don't need such a long length, you can set a smaller value to speed up the encoding process.
|
77 |
-
)['dense_vecs']
|
78 |
-
|
79 |
-
# Reshape the query embedding to fit the cosine_similarity function requirements
|
80 |
-
query_embedding_reshaped = query_embedding.reshape(1, -1)
|
81 |
-
|
82 |
-
# Compute cosine similarities
|
83 |
-
similarities = cosine_similarity(query_embedding_reshaped, embeddings)
|
84 |
-
|
85 |
-
# Find the index of the closest document (highest similarity)
|
86 |
-
closest_doc_index = np.argmax(similarities)
|
87 |
-
|
88 |
-
# Closest document's embedding
|
89 |
-
closest_doc_embedding = embeddings_text[closest_doc_index]
|
90 |
-
|
91 |
-
return closest_doc_embedding
|
92 |
-
|
93 |
-
|
94 |
-
class StopOnTokens(StoppingCriteria):
|
95 |
-
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
|
96 |
-
stop_ids = [29, 0]
|
97 |
-
for stop_id in stop_ids:
|
98 |
-
if input_ids[0][-1] == stop_id:
|
99 |
-
return True
|
100 |
-
return False
|
101 |
-
|
102 |
-
|
103 |
-
def predict(history_transformer_format):
|
104 |
-
|
105 |
-
print(history_transformer_format)
|
106 |
-
stop = StopOnTokens()
|
107 |
-
|
108 |
-
messages = []
|
109 |
-
id_message = 1
|
110 |
-
total_message = len(history_transformer_format)
|
111 |
-
for item in history_transformer_format:
|
112 |
|
113 |
-
|
114 |
-
if id_message == total_message:
|
115 |
-
if assess_rag:
|
116 |
-
question = "<|start_header_id|>user<|end_header_id|>\n\n"+ item[0] + "\n\n### Source ###\n" + source_text
|
117 |
-
else:
|
118 |
-
question = "<|start_header_id|>user<|end_header_id|>\n\n"+ item[0]
|
119 |
-
else:
|
120 |
-
question = "<|start_header_id|>user<|end_header_id|>\n\n"+ item[0]
|
121 |
-
answer = "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"+item[1]
|
122 |
-
result = "".join([question, answer])
|
123 |
-
messages.append(result)
|
124 |
-
id_message = id_message + 1
|
125 |
|
126 |
-
messages = "".join(messages)
|
127 |
-
|
128 |
-
print(messages)
|
129 |
-
|
130 |
-
messages = system_prompt + messages
|
131 |
-
|
132 |
-
print(messages)
|
133 |
-
|
134 |
-
model_inputs = tokenizer([messages], return_tensors="pt").to("cuda")
|
135 |
-
streamer = TextIteratorStreamer(tokenizer, timeout=10., skip_prompt=True, skip_special_tokens=True)
|
136 |
-
generate_kwargs = dict(
|
137 |
-
model_inputs,
|
138 |
-
streamer=streamer,
|
139 |
-
max_new_tokens=1024,
|
140 |
-
do_sample=False,
|
141 |
-
top_p=0.95,
|
142 |
-
temperature=0.4,
|
143 |
-
stopping_criteria=StoppingCriteriaList([stop])
|
144 |
-
)
|
145 |
-
t = Thread(target=model.generate, kwargs=generate_kwargs)
|
146 |
-
t.start()
|
147 |
-
|
148 |
-
history_transformer_format[-1][1] = ""
|
149 |
-
for new_token in streamer:
|
150 |
-
if new_token != '<':
|
151 |
-
history_transformer_format[-1][1] += new_token
|
152 |
-
yield history_transformer_format
|
153 |
-
|
154 |
-
def user(message, history):
|
155 |
-
global source_text
|
156 |
-
global assess_rag
|
157 |
-
#For now, we only query the vector database once, at the start.
|
158 |
-
if len(history) == 0:
|
159 |
-
assess_rag = classification_chatrag(message)
|
160 |
-
if assess_rag:
|
161 |
-
source_text = vector_search(message)
|
162 |
-
else:
|
163 |
-
source_text = "Albert-Tchap n'utilise pas de sources comme votre requête n'a pas l'air d'en recueillir."
|
164 |
|
165 |
-
|
166 |
-
|
167 |
-
|
168 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
169 |
|
170 |
# Define the Gradio interface
|
171 |
-
title = "
|
172 |
-
description = "
|
173 |
examples = [
|
174 |
[
|
175 |
"Qui peut bénéficier de l'AIP?", # user_message
|
@@ -177,26 +91,26 @@ examples = [
|
|
177 |
]
|
178 |
]
|
179 |
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
|
188 |
-
|
189 |
-
|
190 |
-
|
191 |
-
gr.HTML("<h2>Source utilisée</2>")
|
192 |
-
user_output = gr.HTML() # To display the user's message
|
193 |
|
194 |
-
|
195 |
-
predict, chatbot, chatbot
|
196 |
-
)
|
197 |
-
|
198 |
-
clear.click(lambda: None, None, chatbot, queue=False)
|
199 |
|
|
|
|
|
|
|
|
|
|
|
|
|
200 |
|
201 |
-
|
202 |
-
demo.launch()
|
|
|
1 |
+
import transformers
|
2 |
import re
|
3 |
+
from transformers import AutoConfig, AutoTokenizer, AutoModel, AutoModelForCausalLM
|
4 |
from vllm import LLM, SamplingParams
|
5 |
import torch
|
6 |
import gradio as gr
|
|
|
8 |
import os
|
9 |
import shutil
|
10 |
import requests
|
11 |
+
import chromadb
|
12 |
import pandas as pd
|
13 |
+
from chromadb.config import Settings
|
14 |
+
from chromadb.utils import embedding_functions
|
|
|
|
|
|
|
15 |
|
16 |
+
# Define the device
|
17 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
18 |
|
19 |
+
model_name = "PleIAs/OCRonos"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
20 |
|
21 |
+
llm = LLM(model_name, max_model_len=8128)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
|
24 |
+
#CSS for references formatting
|
25 |
+
css = """
|
26 |
+
.generation {
|
27 |
+
margin-left:2em;
|
28 |
+
margin-right:2em;
|
29 |
+
size:1.2em;
|
30 |
+
}
|
31 |
+
:target {
|
32 |
+
background-color: #CCF3DF; /* Change the text color to red */
|
33 |
+
}
|
34 |
+
.source {
|
35 |
+
float:left;
|
36 |
+
max-width:17%;
|
37 |
+
margin-left:2%;
|
38 |
+
}
|
39 |
+
.tooltip {
|
40 |
+
position: relative;
|
41 |
+
cursor: pointer;
|
42 |
+
font-variant-position: super;
|
43 |
+
color: #97999b;
|
44 |
+
}
|
45 |
+
|
46 |
+
.tooltip:hover::after {
|
47 |
+
content: attr(data-text);
|
48 |
+
position: absolute;
|
49 |
+
left: 0;
|
50 |
+
top: 120%; /* Adjust this value as needed to control the vertical spacing between the text and the tooltip */
|
51 |
+
white-space: pre-wrap; /* Allows the text to wrap */
|
52 |
+
width: 500px; /* Sets a fixed maximum width for the tooltip */
|
53 |
+
max-width: 500px; /* Ensures the tooltip does not exceed the maximum width */
|
54 |
+
z-index: 1;
|
55 |
+
background-color: #f9f9f9;
|
56 |
+
color: #000;
|
57 |
+
border: 1px solid #ddd;
|
58 |
+
border-radius: 5px;
|
59 |
+
padding: 5px;
|
60 |
+
display: block;
|
61 |
+
box-shadow: 0 4px 8px rgba(0,0,0,0.1); /* Optional: Adds a subtle shadow for better visibility */
|
62 |
+
}"""
|
63 |
+
|
64 |
+
#Curtesy of chatgpt
|
65 |
+
|
66 |
+
# Class to encapsulate the Falcon chatbot
|
67 |
+
class MistralChatBot:
|
68 |
+
def __init__(self, system_prompt="Le dialogue suivant est une conversation"):
|
69 |
+
self.system_prompt = system_prompt
|
70 |
+
|
71 |
+
def predict(self, user_message):
|
72 |
+
sampling_params = SamplingParams(temperature=0.9, top_p=0.95, max_tokens=4000, presence_penalty=0, stop=["#END#"])
|
73 |
+
detailed_prompt = correction = f"### TEXT ###\n{user_message}\n\n### CORRECTION ###\n"
|
74 |
+
print(detailed_prompt)
|
75 |
+
prompts = [detailed_prompt]
|
76 |
+
outputs = llm.generate(prompts, sampling_params, use_tqdm = False)
|
77 |
+
generated_text = outputs[0].outputs[0].text
|
78 |
+
generated_text = '<h2 style="text-align:center">Réponse</h3>\n<div class="generation">' + generated_text + "</div>"
|
79 |
+
return generated_text
|
80 |
+
|
81 |
+
# Create the Falcon chatbot instance
|
82 |
+
mistral_bot = MistralChatBot()
|
83 |
|
84 |
# Define the Gradio interface
|
85 |
+
title = "Correction d'OCR"
|
86 |
+
description = "Un outil expérimental de correction d'OCR basé sur des modèles de langue"
|
87 |
examples = [
|
88 |
[
|
89 |
"Qui peut bénéficier de l'AIP?", # user_message
|
|
|
91 |
]
|
92 |
]
|
93 |
|
94 |
+
additional_inputs=[
|
95 |
+
gr.Slider(
|
96 |
+
label="Température",
|
97 |
+
value=0.2, # Default value
|
98 |
+
minimum=0.05,
|
99 |
+
maximum=1.0,
|
100 |
+
step=0.05,
|
101 |
+
interactive=True,
|
102 |
+
info="Des valeurs plus élevées donne plus de créativité, mais aussi d'étrangeté",
|
103 |
+
),
|
104 |
+
]
|
|
|
|
|
105 |
|
106 |
+
demo = gr.Blocks()
|
|
|
|
|
|
|
|
|
107 |
|
108 |
+
with gr.Blocks(theme='JohnSmith9982/small_and_pretty', css=css) as demo:
|
109 |
+
gr.HTML("""<h1 style="text-align:center">Correction d'OCR</h1>""")
|
110 |
+
text_input = gr.Textbox(label="Votre texte.", type="text", lines=1)
|
111 |
+
text_button = gr.Button("Corriger l'OCR")
|
112 |
+
text_output = gr.HTML(label="Le texte corrigé")
|
113 |
+
text_button.click(mistral_bot.predict, inputs=text_input, outputs=[text_output])
|
114 |
|
115 |
+
if __name__ == "__main__":
|
116 |
+
demo.queue().launch()
|