Spaces:
Paused
Paused
Update app.py
Browse files
app.py
CHANGED
@@ -46,7 +46,7 @@ def load_model(model_name):
|
|
46 |
except Exception as e:
|
47 |
return f"Erreur lors du chargement du modèle : {str(e)}"
|
48 |
|
49 |
-
def
|
50 |
global model, tokenizer
|
51 |
|
52 |
if model is None or tokenizer is None:
|
@@ -56,39 +56,48 @@ def generate_text(input_text, temperature, top_p, top_k):
|
|
56 |
|
57 |
try:
|
58 |
with torch.no_grad():
|
59 |
-
outputs = model
|
60 |
-
**inputs,
|
61 |
-
max_new_tokens=50,
|
62 |
-
temperature=temperature,
|
63 |
-
top_p=top_p,
|
64 |
-
top_k=top_k,
|
65 |
-
output_attentions=True,
|
66 |
-
return_dict_in_generate=True,
|
67 |
-
output_scores=True
|
68 |
-
)
|
69 |
|
70 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
71 |
|
72 |
-
if hasattr(outputs, '
|
73 |
-
last_token_logits = outputs.scores[-1][0]
|
74 |
-
probabilities = torch.nn.functional.softmax(last_token_logits, dim=-1)
|
75 |
-
top_k = 5
|
76 |
-
top_probs, top_indices = torch.topk(probabilities, top_k)
|
77 |
-
top_words = [tokenizer.decode([idx.item()]) for idx in top_indices]
|
78 |
-
prob_data = {word: prob.item() for word, prob in zip(top_words, top_probs)}
|
79 |
-
prob_plot = plot_probabilities(prob_data)
|
80 |
-
else:
|
81 |
-
prob_plot = None
|
82 |
-
|
83 |
-
if hasattr(outputs, 'attentions') and outputs.attentions:
|
84 |
attention_data = torch.mean(torch.stack(outputs.attentions), dim=(0, 1)).cpu().numpy()
|
85 |
attention_plot = plot_attention(attention_data, tokenizer.convert_ids_to_tokens(inputs["input_ids"][0]))
|
86 |
else:
|
87 |
attention_plot = None
|
88 |
|
89 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
90 |
except Exception as e:
|
91 |
-
return f"Erreur lors de la génération : {str(e)}"
|
92 |
|
93 |
def plot_attention(attention, tokens):
|
94 |
fig, ax = plt.subplots(figsize=(10, 10))
|
@@ -119,10 +128,10 @@ def reset():
|
|
119 |
global model, tokenizer
|
120 |
model = None
|
121 |
tokenizer = None
|
122 |
-
return "", 1.0, 1.0, 50, None, None, None
|
123 |
|
124 |
with gr.Blocks() as demo:
|
125 |
-
gr.Markdown("#
|
126 |
|
127 |
with gr.Accordion("Sélection du modèle"):
|
128 |
model_dropdown = gr.Dropdown(choices=models, label="Choisissez un modèle")
|
@@ -135,22 +144,28 @@ with gr.Blocks() as demo:
|
|
135 |
top_k = gr.Slider(1, 100, value=50, step=1, label="Top-k")
|
136 |
|
137 |
input_text = gr.Textbox(label="Texte d'entrée", lines=3)
|
138 |
-
|
|
|
139 |
|
140 |
-
|
141 |
|
142 |
with gr.Row():
|
143 |
attention_plot = gr.Plot(label="Visualisation de l'attention")
|
144 |
prob_plot = gr.Plot(label="Probabilités des tokens suivants")
|
145 |
|
|
|
|
|
146 |
reset_button = gr.Button("Réinitialiser")
|
147 |
|
148 |
load_button.click(load_model, inputs=[model_dropdown], outputs=[load_output])
|
|
|
|
|
|
|
149 |
generate_button.click(generate_text,
|
150 |
inputs=[input_text, temperature, top_p, top_k],
|
151 |
-
outputs=[
|
152 |
reset_button.click(reset,
|
153 |
-
outputs=[input_text, temperature, top_p, top_k,
|
154 |
|
155 |
if __name__ == "__main__":
|
156 |
demo.launch()
|
|
|
46 |
except Exception as e:
|
47 |
return f"Erreur lors du chargement du modèle : {str(e)}"
|
48 |
|
49 |
+
def analyze_next_token(input_text, temperature, top_p, top_k):
|
50 |
global model, tokenizer
|
51 |
|
52 |
if model is None or tokenizer is None:
|
|
|
56 |
|
57 |
try:
|
58 |
with torch.no_grad():
|
59 |
+
outputs = model(**inputs)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
60 |
|
61 |
+
last_token_logits = outputs.logits[0, -1, :]
|
62 |
+
probabilities = torch.nn.functional.softmax(last_token_logits, dim=-1)
|
63 |
+
top_k = 5
|
64 |
+
top_probs, top_indices = torch.topk(probabilities, top_k)
|
65 |
+
top_words = [tokenizer.decode([idx.item()]) for idx in top_indices]
|
66 |
+
prob_data = {word: prob.item() for word, prob in zip(top_words, top_probs)}
|
67 |
+
prob_plot = plot_probabilities(prob_data)
|
68 |
|
69 |
+
if hasattr(outputs, 'attentions') and outputs.attentions is not None:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
70 |
attention_data = torch.mean(torch.stack(outputs.attentions), dim=(0, 1)).cpu().numpy()
|
71 |
attention_plot = plot_attention(attention_data, tokenizer.convert_ids_to_tokens(inputs["input_ids"][0]))
|
72 |
else:
|
73 |
attention_plot = None
|
74 |
|
75 |
+
return "\n".join([f"{word}: {prob:.4f}" for word, prob in prob_data.items()]), attention_plot, prob_plot
|
76 |
+
except Exception as e:
|
77 |
+
return f"Erreur lors de l'analyse : {str(e)}", None, None
|
78 |
+
|
79 |
+
def generate_text(input_text, temperature, top_p, top_k):
|
80 |
+
global model, tokenizer
|
81 |
+
|
82 |
+
if model is None or tokenizer is None:
|
83 |
+
return "Veuillez d'abord charger un modèle."
|
84 |
+
|
85 |
+
inputs = tokenizer(input_text, return_tensors="pt", padding=True, truncation=True, max_length=512).to(model.device)
|
86 |
+
|
87 |
+
try:
|
88 |
+
with torch.no_grad():
|
89 |
+
outputs = model.generate(
|
90 |
+
**inputs,
|
91 |
+
max_new_tokens=50,
|
92 |
+
temperature=temperature,
|
93 |
+
top_p=top_p,
|
94 |
+
top_k=top_k
|
95 |
+
)
|
96 |
+
|
97 |
+
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
98 |
+
return generated_text
|
99 |
except Exception as e:
|
100 |
+
return f"Erreur lors de la génération : {str(e)}"
|
101 |
|
102 |
def plot_attention(attention, tokens):
|
103 |
fig, ax = plt.subplots(figsize=(10, 10))
|
|
|
128 |
global model, tokenizer
|
129 |
model = None
|
130 |
tokenizer = None
|
131 |
+
return "", 1.0, 1.0, 50, None, None, None, None
|
132 |
|
133 |
with gr.Blocks() as demo:
|
134 |
+
gr.Markdown("# Analyse et génération de texte")
|
135 |
|
136 |
with gr.Accordion("Sélection du modèle"):
|
137 |
model_dropdown = gr.Dropdown(choices=models, label="Choisissez un modèle")
|
|
|
144 |
top_k = gr.Slider(1, 100, value=50, step=1, label="Top-k")
|
145 |
|
146 |
input_text = gr.Textbox(label="Texte d'entrée", lines=3)
|
147 |
+
analyze_button = gr.Button("Analyser le prochain token")
|
148 |
+
generate_button = gr.Button("Générer la suite du texte")
|
149 |
|
150 |
+
next_token_probs = gr.Textbox(label="Probabilités du prochain token")
|
151 |
|
152 |
with gr.Row():
|
153 |
attention_plot = gr.Plot(label="Visualisation de l'attention")
|
154 |
prob_plot = gr.Plot(label="Probabilités des tokens suivants")
|
155 |
|
156 |
+
generated_text = gr.Textbox(label="Texte généré", lines=5)
|
157 |
+
|
158 |
reset_button = gr.Button("Réinitialiser")
|
159 |
|
160 |
load_button.click(load_model, inputs=[model_dropdown], outputs=[load_output])
|
161 |
+
analyze_button.click(analyze_next_token,
|
162 |
+
inputs=[input_text, temperature, top_p, top_k],
|
163 |
+
outputs=[next_token_probs, attention_plot, prob_plot])
|
164 |
generate_button.click(generate_text,
|
165 |
inputs=[input_text, temperature, top_p, top_k],
|
166 |
+
outputs=[generated_text])
|
167 |
reset_button.click(reset,
|
168 |
+
outputs=[input_text, temperature, top_p, top_k, next_token_probs, attention_plot, prob_plot, generated_text])
|
169 |
|
170 |
if __name__ == "__main__":
|
171 |
demo.launch()
|