Spaces:
Running
on
Zero
Running
on
Zero
Add confidence based noising
Browse files
app.py
CHANGED
@@ -73,6 +73,21 @@ def noisify_answer(input_ids, answer_start, threshold=1.0, eot_weight=1.0):
|
|
73 |
noised[idx] = val
|
74 |
return noised
|
75 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
76 |
@spaces.GPU
|
77 |
def generate_diffusion_text(input_ids, answer_start):
|
78 |
with torch.no_grad():
|
@@ -81,18 +96,21 @@ def generate_diffusion_text(input_ids, answer_start):
|
|
81 |
probs = torch.nn.functional.softmax(logits, dim=-1).squeeze()
|
82 |
probs = torch.clamp(probs, min=1e-8, max=1.0)
|
83 |
sampled = torch.multinomial(probs, num_samples=1).squeeze().tolist()
|
84 |
-
|
|
|
|
|
|
|
85 |
|
86 |
# --- Inference Wrapper ---
|
87 |
|
88 |
|
89 |
-
|
|
|
90 |
placeholder = "What do you know about the city of New York?"
|
91 |
if question.strip() == "":
|
92 |
question = placeholder
|
93 |
-
|
94 |
-
print('started generation')
|
95 |
|
|
|
96 |
prompt = f"User: {question}\nAssistant:"
|
97 |
input_ids = tokenizer.encode(prompt, add_special_tokens=False)
|
98 |
answer_start = find_answer_start(input_ids, assistant_marker_ids)
|
@@ -112,7 +130,7 @@ def diffusion_chat(question, eot_weight, max_it, sharpness):
|
|
112 |
|
113 |
for i in range(max_it):
|
114 |
print('Generating output')
|
115 |
-
generated_tokens = generate_diffusion_text(current_tokens, answer_start)
|
116 |
current_tokens = generated_tokens
|
117 |
|
118 |
decoded_ids = current_tokens[answer_start:]
|
@@ -141,7 +159,11 @@ def diffusion_chat(question, eot_weight, max_it, sharpness):
|
|
141 |
break
|
142 |
|
143 |
threshold = get_noising_schedule(i, max_it, sharpness=sharpness)
|
144 |
-
|
|
|
|
|
|
|
|
|
145 |
time.sleep(0.01)
|
146 |
|
147 |
final_tokens = tokenizer.convert_ids_to_tokens(current_tokens[answer_start:])
|
@@ -162,7 +184,8 @@ demo = gr.Interface(
|
|
162 |
gr.Textbox(label="User Question", lines=2, placeholder="What do you know about the city of New York?"),
|
163 |
gr.Slider(0, 1, value=0.4, step=0.05, label="↓ = longer answers (EOT weight)"),
|
164 |
gr.Slider(1, 512, value=64, step=1, label="↑ = more iterations"),
|
165 |
-
gr.Slider(1.0, 20.0, value=5.0, step=0.5, label="↓ = more noising (sharpness)")
|
|
|
166 |
],
|
167 |
outputs=[gr.HTML(label="Diffusion Output")],
|
168 |
title="Diffusion Language Model Chat",
|
|
|
73 |
noised[idx] = val
|
74 |
return noised
|
75 |
|
76 |
+
# Add new noising function
|
77 |
+
def confidence_guided_noising(input_ids, answer_start, confidences, eot_weight):
|
78 |
+
noised = input_ids.copy()
|
79 |
+
mixed_probs = token_probabilities.copy()
|
80 |
+
mixed_probs[eot_token_id] *= eot_weight
|
81 |
+
mixed_probs /= mixed_probs.sum()
|
82 |
+
|
83 |
+
for i, conf in enumerate(confidences[answer_start:]):
|
84 |
+
p_noise = 1.0 - conf
|
85 |
+
if rng.random() < p_noise:
|
86 |
+
idx = answer_start + i
|
87 |
+
noised[idx] = rng.choice(np.arange(vocab_size), p=mixed_probs)
|
88 |
+
|
89 |
+
return noised
|
90 |
+
|
91 |
@spaces.GPU
|
92 |
def generate_diffusion_text(input_ids, answer_start):
|
93 |
with torch.no_grad():
|
|
|
96 |
probs = torch.nn.functional.softmax(logits, dim=-1).squeeze()
|
97 |
probs = torch.clamp(probs, min=1e-8, max=1.0)
|
98 |
sampled = torch.multinomial(probs, num_samples=1).squeeze().tolist()
|
99 |
+
|
100 |
+
# Extract confidence of selected tokens
|
101 |
+
conf = probs[range(len(sampled)), sampled].cpu().numpy()
|
102 |
+
return sampled, conf # ✅ NEW: Return confidence
|
103 |
|
104 |
# --- Inference Wrapper ---
|
105 |
|
106 |
|
107 |
+
# Modify diffusion_chat to use new noising conditionally
|
108 |
+
def diffusion_chat(question, eot_weight, max_it, sharpness, use_confidence_noising):
|
109 |
placeholder = "What do you know about the city of New York?"
|
110 |
if question.strip() == "":
|
111 |
question = placeholder
|
|
|
|
|
112 |
|
113 |
+
print('started generation')
|
114 |
prompt = f"User: {question}\nAssistant:"
|
115 |
input_ids = tokenizer.encode(prompt, add_special_tokens=False)
|
116 |
answer_start = find_answer_start(input_ids, assistant_marker_ids)
|
|
|
130 |
|
131 |
for i in range(max_it):
|
132 |
print('Generating output')
|
133 |
+
generated_tokens, confidences = generate_diffusion_text(current_tokens, answer_start)
|
134 |
current_tokens = generated_tokens
|
135 |
|
136 |
decoded_ids = current_tokens[answer_start:]
|
|
|
159 |
break
|
160 |
|
161 |
threshold = get_noising_schedule(i, max_it, sharpness=sharpness)
|
162 |
+
if use_confidence_noising:
|
163 |
+
current_tokens = confidence_guided_noising(generated_tokens, answer_start, confidences, eot_weight)
|
164 |
+
else:
|
165 |
+
current_tokens = noisify_answer(generated_tokens, answer_start, threshold=threshold, eot_weight=eot_weight)
|
166 |
+
|
167 |
time.sleep(0.01)
|
168 |
|
169 |
final_tokens = tokenizer.convert_ids_to_tokens(current_tokens[answer_start:])
|
|
|
184 |
gr.Textbox(label="User Question", lines=2, placeholder="What do you know about the city of New York?"),
|
185 |
gr.Slider(0, 1, value=0.4, step=0.05, label="↓ = longer answers (EOT weight)"),
|
186 |
gr.Slider(1, 512, value=64, step=1, label="↑ = more iterations"),
|
187 |
+
gr.Slider(1.0, 20.0, value=5.0, step=0.5, label="↓ = more noising (sharpness)"),
|
188 |
+
gr.Checkbox(value=False, label="Use confidence-guided noising") # ✅ NEW
|
189 |
],
|
190 |
outputs=[gr.HTML(label="Diffusion Output")],
|
191 |
title="Diffusion Language Model Chat",
|