Spaces:
Running
on
Zero
Running
on
Zero
Commit
·
743ec89
1
Parent(s):
2812121
add app file and mosaic file
Browse files- gradio_app.py +94 -0
- mosaic.py +344 -0
gradio_app.py
ADDED
@@ -0,0 +1,94 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from mosaic import Mosaic # adjust import as needed
|
3 |
+
|
4 |
+
# Maximum number of model textboxes
|
5 |
+
MAX_MODELS = 10
|
6 |
+
|
7 |
+
def update_textboxes(n_visible):
|
8 |
+
"""
|
9 |
+
Given the current visible count, increments it by 1 (up to MAX_MODELS)
|
10 |
+
and returns updated visibility settings for all model textboxes.
|
11 |
+
"""
|
12 |
+
if n_visible < MAX_MODELS:
|
13 |
+
n_visible += 1
|
14 |
+
# Create a list of update objects for each textbox: visible if its index is less than n_visible.
|
15 |
+
updates = []
|
16 |
+
for i in range(MAX_MODELS):
|
17 |
+
if i < n_visible:
|
18 |
+
updates.append(gr.update(visible=True))
|
19 |
+
else:
|
20 |
+
updates.append(gr.update(visible=False))
|
21 |
+
return n_visible, *updates
|
22 |
+
|
23 |
+
def run_scoring(input_text, model1, model2, model3, model4, model5, model6, model7, model8, model9, model10, threshold_choice, custom_threshold):
|
24 |
+
"""
|
25 |
+
Collect all non-empty model paths, instantiate Mosaic, compute the score,
|
26 |
+
and return a message based on the threshold.
|
27 |
+
"""
|
28 |
+
model_paths = []
|
29 |
+
for m in [model1, model2, model3, model4, model5, model6, model7, model8, model9, model10]:
|
30 |
+
if m.strip() != "":
|
31 |
+
model_paths.append(m.strip())
|
32 |
+
if len(model_paths) < 2:
|
33 |
+
return "Please enter at least two model paths.", None, None
|
34 |
+
# Choose threshold value
|
35 |
+
if threshold_choice == "default":
|
36 |
+
threshold = 0.0
|
37 |
+
elif threshold_choice == "raid":
|
38 |
+
threshold = 0.23
|
39 |
+
elif threshold_choice == "custom":
|
40 |
+
threshold = custom_threshold
|
41 |
+
else:
|
42 |
+
threshold = 0.0
|
43 |
+
# Instantiate the Mosaic class with the selected model paths.
|
44 |
+
mosaic_instance = Mosaic(model_name_or_paths=model_paths, one_model_mode=False)
|
45 |
+
final_score = mosaic_instance.compute_end_score(input_text)
|
46 |
+
if final_score < threshold:
|
47 |
+
result_message = "This text was probably generated."
|
48 |
+
else:
|
49 |
+
result_message = "This text is likely human-generated."
|
50 |
+
return result_message, final_score, threshold
|
51 |
+
|
52 |
+
with gr.Blocks() as demo:
|
53 |
+
gr.Markdown("# MOSAIC Scoring App")
|
54 |
+
with gr.Row():
|
55 |
+
input_text = gr.Textbox(lines=10, placeholder="Enter text here...", label="Input Text")
|
56 |
+
with gr.Column():
|
57 |
+
gr.Markdown("### Model Paths (at least 2 required)")
|
58 |
+
gr.Markdown("Order matters for model 1 only, the Reference model. Please use the one with the best perplexity on human texts. (The largest LLM if applicable.) GPT2 models are enough to detect easy prompts from chatgpt.")
|
59 |
+
# State to keep track of the number of visible textboxes (starting with 2)
|
60 |
+
n_models_state = gr.State(2)
|
61 |
+
# Create 10 textboxes. We'll name them model1, model2, ..., model10.
|
62 |
+
model1 = gr.Textbox(value="openai-community/gpt2-large", label="Model 1 Path ", visible=True)
|
63 |
+
model2 = gr.Textbox(value="openai-community/gpt2-medium", label="Model 2 Path", visible=True)
|
64 |
+
model3 = gr.Textbox(value="", label="Model 3 Path", visible=False)
|
65 |
+
model4 = gr.Textbox(value="", label="Model 4 Path", visible=False)
|
66 |
+
model5 = gr.Textbox(value="", label="Model 5 Path", visible=False)
|
67 |
+
model6 = gr.Textbox(value="", label="Model 6 Path", visible=False)
|
68 |
+
model7 = gr.Textbox(value="", label="Model 7 Path", visible=False)
|
69 |
+
model8 = gr.Textbox(value="", label="Model 8 Path", visible=False)
|
70 |
+
model9 = gr.Textbox(value="", label="Model 9 Path", visible=False)
|
71 |
+
model10 = gr.Textbox(value="", label="Model 10 Path", visible=False)
|
72 |
+
# Add a plus button to reveal one more textbox.
|
73 |
+
plus_button = gr.Button("+", elem_id="plus_button")
|
74 |
+
# When plus_button is clicked, update n_models_state and all model textboxes.
|
75 |
+
plus_button.click(
|
76 |
+
fn=update_textboxes,
|
77 |
+
inputs=n_models_state,
|
78 |
+
outputs=[n_models_state, model1, model2, model3, model4, model5, model6, model7, model8, model9, model10]
|
79 |
+
)
|
80 |
+
with gr.Row():
|
81 |
+
threshold_choice = gr.Radio(choices=["default", "raid", "custom"], value="default", label="Threshold Choice")
|
82 |
+
custom_threshold = gr.Number(value=0.0, label="Custom Threshold (if 'custom' selected)")
|
83 |
+
with gr.Row():
|
84 |
+
output_message = gr.Textbox(label="Result Message")
|
85 |
+
output_score = gr.Number(label="Final Score")
|
86 |
+
output_threshold = gr.Number(label="Threshold Used")
|
87 |
+
run_button = gr.Button("Run Scoring")
|
88 |
+
run_button.click(
|
89 |
+
fn=run_scoring,
|
90 |
+
inputs=[input_text, model1, model2, model3, model4, model5, model6, model7, model8, model9, model10, threshold_choice, custom_threshold],
|
91 |
+
outputs=[output_message, output_score, output_threshold]
|
92 |
+
)
|
93 |
+
|
94 |
+
demo.launch()
|
mosaic.py
ADDED
@@ -0,0 +1,344 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import List, Optional
|
2 |
+
import numpy as np
|
3 |
+
import torch
|
4 |
+
import transformers
|
5 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
6 |
+
import torch.nn.functional as F
|
7 |
+
|
8 |
+
torch.set_grad_enabled(False)
|
9 |
+
|
10 |
+
def apply_top_p_with_epsilon(logits: torch.Tensor, top_p: float, epsilon: float = 1e-10) -> torch.Tensor:
|
11 |
+
"""
|
12 |
+
Applies a top-p (nucleus) filtering to logits but, instead of setting
|
13 |
+
the logits of non-selected tokens to -inf (which would result in zero probability),
|
14 |
+
sets them to log(epsilon), so that the support remains the same.
|
15 |
+
|
16 |
+
Parameters:
|
17 |
+
logits: Tensor of shape (batch, seq_len, vocab_size)
|
18 |
+
top_p: The nucleus threshold (e.g. 0.7, 0.8, etc.)
|
19 |
+
epsilon: The small value to assign to tokens not selected.
|
20 |
+
|
21 |
+
Returns:
|
22 |
+
new_logits: Tensor with the same shape as logits.
|
23 |
+
"""
|
24 |
+
# Compute probabilities from logits
|
25 |
+
probs = F.softmax(logits, dim=-1)
|
26 |
+
# Sort probabilities (descending) along the vocabulary dimension.
|
27 |
+
sorted_probs, sorted_indices = torch.sort(probs, descending=True, dim=-1)
|
28 |
+
# Compute the cumulative sum along the sorted probabilities.
|
29 |
+
cumulative_probs = torch.cumsum(sorted_probs, dim=-1)
|
30 |
+
# Create a mask: True for tokens to keep.
|
31 |
+
# We keep tokens until cumulative_probs <= top_p.
|
32 |
+
keep_mask = cumulative_probs <= top_p
|
33 |
+
|
34 |
+
# Ensure that at least one token is kept per example: if none are kept, keep the top one.
|
35 |
+
# Here we check along the vocab dimension.
|
36 |
+
no_token_kept = keep_mask.sum(dim=-1, keepdim=True) == 0
|
37 |
+
if no_token_kept.any():
|
38 |
+
# For positions where no token was kept, set the first token (highest probability) to True.
|
39 |
+
# Note: torch.scatter_ returns a modified tensor.
|
40 |
+
# We create a tensor of zeros (False) and then scatter True into the first column.
|
41 |
+
fix_mask = torch.zeros_like(keep_mask, dtype=torch.bool)
|
42 |
+
fix_mask.scatter_(-1, torch.zeros_like(keep_mask[..., :1], dtype=torch.long), True)
|
43 |
+
keep_mask = torch.where(no_token_kept, fix_mask, keep_mask)
|
44 |
+
|
45 |
+
# Now, create new logits: copy the original logits.
|
46 |
+
new_logits = logits.clone()
|
47 |
+
# For tokens that are not kept (i.e. where keep_mask is False), set their logit to log(epsilon)
|
48 |
+
new_logits[~keep_mask] = torch.log(torch.tensor(epsilon, device=logits.device, dtype=logits.dtype))
|
49 |
+
return new_logits
|
50 |
+
|
51 |
+
class Mosaic(object):
|
52 |
+
def __init__(self,
|
53 |
+
model_name_or_paths: List[str],
|
54 |
+
use_bfloat16: bool = True,
|
55 |
+
max_token_observed: int = 512,
|
56 |
+
unigram: Optional[str] = None,
|
57 |
+
custom_config : Optional[List[bool]] = None,
|
58 |
+
stupid_mode: bool = False,
|
59 |
+
one_model_mode: bool = False
|
60 |
+
) -> None:
|
61 |
+
self.models = []
|
62 |
+
for i, model_name_or_path in enumerate(model_name_or_paths):
|
63 |
+
model = AutoModelForCausalLM.from_pretrained(model_name_or_path,
|
64 |
+
device_map="auto",
|
65 |
+
trust_remote_code=True,
|
66 |
+
torch_dtype=torch.bfloat16 if use_bfloat16
|
67 |
+
else torch.float32
|
68 |
+
)
|
69 |
+
model.eval() # Set the model to evaluation mode
|
70 |
+
self.models.append(model)
|
71 |
+
print(f"Loaded model: {model_name_or_path}")
|
72 |
+
# Print the device map
|
73 |
+
#print(f"Device map for {model_name_or_path}: {model.hf_device_map}")
|
74 |
+
|
75 |
+
if stupid_mode:
|
76 |
+
self.max_iters = 0
|
77 |
+
else:
|
78 |
+
self.max_iters = 1000
|
79 |
+
|
80 |
+
self.one_model_mode = one_model_mode
|
81 |
+
|
82 |
+
self.tokenizer = AutoTokenizer.from_pretrained(model_name_or_paths[-1])
|
83 |
+
if not self.tokenizer.pad_token:
|
84 |
+
self.tokenizer.pad_token = self.tokenizer.eos_token
|
85 |
+
|
86 |
+
self.max_token_observed = max_token_observed
|
87 |
+
|
88 |
+
self.nb_models = len(self.models)
|
89 |
+
self.unigram_path = unigram
|
90 |
+
|
91 |
+
if custom_config is None:
|
92 |
+
custom_config = [False] * self.nb_models
|
93 |
+
self.custom_config = custom_config
|
94 |
+
|
95 |
+
def _tokenize(self, batch: list[str]) -> transformers.BatchEncoding:
|
96 |
+
encodings = self.tokenizer(
|
97 |
+
batch,
|
98 |
+
return_tensors="pt",
|
99 |
+
padding="longest",
|
100 |
+
truncation=True,
|
101 |
+
max_length=self.max_token_observed,
|
102 |
+
return_token_type_ids=False)
|
103 |
+
return encodings
|
104 |
+
|
105 |
+
def trim_logits(self, logits, max_length=32000):
|
106 |
+
# Check the shape of the logits tensor
|
107 |
+
if logits.shape[2] > max_length:
|
108 |
+
# Slice the tensor to keep only the first max_length elements along the last dimension
|
109 |
+
logits = logits[:, :, :max_length]
|
110 |
+
return logits
|
111 |
+
|
112 |
+
@torch.inference_mode()
|
113 |
+
def _get_logits(self, encodings: transformers.BatchEncoding) -> List[torch.Tensor]:
|
114 |
+
# If one_model_mode is active, we simulate multiple models by applying top-p with different thresholds.
|
115 |
+
if self.one_model_mode:
|
116 |
+
# Compute base logits from the single model.
|
117 |
+
model = self.models[0]
|
118 |
+
device = next(model.parameters()).device
|
119 |
+
model_encodings = encodings.to(device)
|
120 |
+
base_logits = model(**model_encodings).logits
|
121 |
+
# Optionally trim logits:
|
122 |
+
# base_logits = self.trim_logits(base_logits)
|
123 |
+
# Define the top-p thresholds (e.g., four different values)
|
124 |
+
top_p_values = [0.7, 0.8, 0.9, 0.95]
|
125 |
+
# Epsilon value for non-selected tokens (you can adjust this if needed)
|
126 |
+
epsilon = 1e-10
|
127 |
+
logits_list = []
|
128 |
+
for top_p in top_p_values:
|
129 |
+
warped_logits = apply_top_p_with_epsilon(base_logits, top_p, epsilon)
|
130 |
+
logits_list.append(warped_logits)
|
131 |
+
else:
|
132 |
+
# Normal mode: use each model in self.models.
|
133 |
+
logits_list = []
|
134 |
+
for i, model in enumerate(self.models):
|
135 |
+
device = next(model.parameters()).device
|
136 |
+
model_encodings = encodings.to(device)
|
137 |
+
logits = model(**model_encodings).logits
|
138 |
+
# Optionally trim logits:
|
139 |
+
# logits = self.trim_logits(logits)
|
140 |
+
logits_list.append(logits)
|
141 |
+
if device.type == "cuda":
|
142 |
+
torch.cuda.synchronize(device)
|
143 |
+
|
144 |
+
if self.unigram_path:
|
145 |
+
batch_size, seq_len, voc_size = logits_list[0].shape
|
146 |
+
unigram_proba = torch.load(self.unigram_path)
|
147 |
+
unigram_proba += 1e-10
|
148 |
+
unigram_logits = torch.log(unigram_proba)
|
149 |
+
# Optionally center logits if needed:
|
150 |
+
logits = logits_list[0] - logits_list[0].mean(dim=-1, keepdim=True)
|
151 |
+
expanded_unigram_logits = unigram_logits.unsqueeze(0).unsqueeze(0).expand(batch_size, seq_len, voc_size)
|
152 |
+
logits_list.append(expanded_unigram_logits)
|
153 |
+
return logits_list
|
154 |
+
|
155 |
+
def get_softmax_probabilities(self, input_text):
|
156 |
+
encodings = self._tokenize(input_text)
|
157 |
+
logits_list = self._get_logits(encodings)
|
158 |
+
probabilities_list = softmax_probabilities_all_models(logits_list)
|
159 |
+
return encodings, logits_list, probabilities_list
|
160 |
+
|
161 |
+
def compute_arimoto_torch(self, input_text, max_iters=1000):
|
162 |
+
encodings, logits_list, tensors_list = self.get_softmax_probabilities(input_text)
|
163 |
+
nb_models = len(tensors_list)
|
164 |
+
seq_len = len(encodings.input_ids[0])
|
165 |
+
voc_size = tensors_list[0].shape[-1]
|
166 |
+
|
167 |
+
device = tensors_list[0].device
|
168 |
+
# Move all tensors in tensors_list to the device of the first tensor
|
169 |
+
tensors_list = [tensor.to(device) for tensor in tensors_list]
|
170 |
+
|
171 |
+
# Stack all model predictions along a new dimension to form a (seq_len, nb_models, voc_size) tensor
|
172 |
+
probabilities_tensor = torch.stack([t[0] for t in tensors_list], dim=1).to(tensors_list[0].device)
|
173 |
+
|
174 |
+
# Run the Blahut-Arimoto algorithm on the entire batch
|
175 |
+
capacity, p = blahut_arimoto_torch(probabilities_tensor, max_iters=max_iters)
|
176 |
+
|
177 |
+
# Prepare the weighted sum tensor, initially zeros
|
178 |
+
weighted_sum_tensor = torch.zeros_like(tensors_list[0])
|
179 |
+
|
180 |
+
# Here, we need an additional mechanism if 'p' shapes or logic require different handling
|
181 |
+
# Assuming 'p' is now (seq_len, nb_models), apply weights to each model's output
|
182 |
+
for i in range(nb_models):
|
183 |
+
weighted_sum_tensor += p[:, i:i+1] * tensors_list[i]
|
184 |
+
|
185 |
+
return encodings, weighted_sum_tensor, tensors_list, p, logits_list
|
186 |
+
|
187 |
+
def compute_scores(self, input_text):
|
188 |
+
encodings, weighted_sum_tensor, probabilities_list, arimoto_weights, logits_list = self.compute_arimoto_torch(input_text, max_iters=self.max_iters)
|
189 |
+
log_ppl, ppl, nll = perplexity(encodings, weighted_sum_tensor)
|
190 |
+
ppl_list = perplexity_all_models(encodings, logits_list)
|
191 |
+
x_ppl_list = cross_entropy(weighted_sum_tensor, probabilities_list)
|
192 |
+
return log_ppl, x_ppl_list, arimoto_weights, nll, ppl_list
|
193 |
+
|
194 |
+
def compute_end_score(self, input_text):
|
195 |
+
encodings, weighted_sum_tensor, probabilities_list, arimoto_weights, logits_list = self.compute_arimoto_torch(input_text)
|
196 |
+
log_ppl, ppl, nll = perplexity(encodings, weighted_sum_tensor)
|
197 |
+
ppl_list = perplexity_all_models(encodings, logits_list)
|
198 |
+
x_ppl_list = cross_entropy(weighted_sum_tensor, probabilities_list)
|
199 |
+
log_ppl_value = log_ppl.item()
|
200 |
+
x_ppl_values = [x.item() for x in x_ppl_list]
|
201 |
+
final_score = log_ppl_value - x_ppl_values[0] #Ensure your "reference model" is given as first argument
|
202 |
+
return final_score
|
203 |
+
|
204 |
+
def perplexity(encodings, weighted_sum_tensor):
|
205 |
+
shifted_probabilities = weighted_sum_tensor[..., :-1, :].contiguous()
|
206 |
+
shifted_labels = encodings.input_ids[..., 1:].contiguous()
|
207 |
+
shifted_attention_mask = encodings.attention_mask[..., 1:].contiguous()
|
208 |
+
|
209 |
+
device = shifted_probabilities.device
|
210 |
+
|
211 |
+
# Ensure all tensors are moved to the same device
|
212 |
+
shifted_probabilities = shifted_probabilities.to(device)
|
213 |
+
shifted_labels = shifted_labels.to(device)
|
214 |
+
shifted_attention_mask = shifted_attention_mask.to(device)
|
215 |
+
|
216 |
+
actual_next_token_probabilities = torch.gather(shifted_probabilities, 2, shifted_labels.unsqueeze(-1)).squeeze(-1)
|
217 |
+
|
218 |
+
nll = -torch.log(actual_next_token_probabilities + 1e-12)
|
219 |
+
nll_masked = nll * shifted_attention_mask
|
220 |
+
|
221 |
+
# Calculate the average NLL per sequence, taking into account only the valid (non-padded) tokens
|
222 |
+
average_nll = torch.sum(nll_masked, dim=1) / torch.sum(shifted_attention_mask, dim=1)
|
223 |
+
|
224 |
+
# Calculate perplexity per sequence
|
225 |
+
perplexity = torch.exp(average_nll)
|
226 |
+
return average_nll, perplexity, nll_masked
|
227 |
+
|
228 |
+
def cross_entropy(weighted_sum_tensor, probabilities_list):
|
229 |
+
device = weighted_sum_tensor.device
|
230 |
+
x_ppl_list = []
|
231 |
+
|
232 |
+
# Compute log of weighted_sum_tensor outside the loop since it doesn't depend on m2_probabilities
|
233 |
+
log_M1 = torch.log(weighted_sum_tensor).to(device)
|
234 |
+
|
235 |
+
for m2_probabilities in probabilities_list:
|
236 |
+
m2_probabilities = m2_probabilities.to(device)
|
237 |
+
# Ensure m2_probabilities is correctly shaped for batch matrix multiplication
|
238 |
+
# log_M1 shape is already (batch_size, sequence_length, vocabulary_size)
|
239 |
+
# We need m2_probabilities in shape (batch_size, vocabulary_size, sequence_length) for bmm
|
240 |
+
m2_probabilities_transposed = m2_probabilities.transpose(1, 2)
|
241 |
+
|
242 |
+
# Perform batch matrix multiplication
|
243 |
+
# Resulting shape: (batch_size, sequence_length, sequence_length)
|
244 |
+
# We sum over the vocabulary dimension, effectively computing the dot product for each sequence position
|
245 |
+
dot_products = torch.bmm(log_M1, m2_probabilities_transposed)
|
246 |
+
|
247 |
+
# Since we're interested in the diagonal (dot products of corresponding vectors), we extract it
|
248 |
+
# The diagonal for each item in the batch gives us the dot products we're interested in
|
249 |
+
# torch.diagonal doesn't support batched operations directly, so we need to workaround
|
250 |
+
dot_products_diagonal = torch.einsum('bii->bi', dot_products) # Using einsum to extract diagonals for batch
|
251 |
+
|
252 |
+
# Compute the mean of the dot_products_diagonal across the sequence dimension
|
253 |
+
# This gives us the average dot product per sequence, which is then negated
|
254 |
+
x_ppl = -torch.mean(dot_products_diagonal, dim=1)
|
255 |
+
|
256 |
+
x_ppl_list.append(x_ppl)
|
257 |
+
x_ppl_tensor = torch.stack(x_ppl_list)
|
258 |
+
return x_ppl_list #, x_ppl_tensor
|
259 |
+
|
260 |
+
def softmax_probabilities_all_models(logits_list: List[torch.Tensor]) -> List[torch.Tensor]:
|
261 |
+
"""
|
262 |
+
Calculates the softmax probabilities for the entire sequence of tokens for each model.
|
263 |
+
|
264 |
+
Parameters:
|
265 |
+
- logits_list: List[torch.Tensor]
|
266 |
+
A list containing the logits tensor for each model.
|
267 |
+
|
268 |
+
Returns:
|
269 |
+
- List[torch.Tensor]: A list of tensors, where each tensor is the softmax probabilities
|
270 |
+
for one model across the entire sequence of tokens.
|
271 |
+
"""
|
272 |
+
softmax_fn = torch.nn.Softmax(dim=-1)
|
273 |
+
probabilities_list = []
|
274 |
+
|
275 |
+
for logits in logits_list:
|
276 |
+
# Calculate softmax probabilities across the vocabulary for each token position
|
277 |
+
softmax_probabilities = softmax_fn(logits)
|
278 |
+
probabilities_list.append(softmax_probabilities)
|
279 |
+
|
280 |
+
return probabilities_list
|
281 |
+
|
282 |
+
def perplexity_logits(encoding, logits):
|
283 |
+
# Ensure encoding tensors are moved to the same device as logits
|
284 |
+
device = logits.device
|
285 |
+
logits = torch.clamp(logits, min=-20, max=50)
|
286 |
+
|
287 |
+
encoding_input_ids = encoding.input_ids.to(device)
|
288 |
+
encoding_attention_mask = encoding.attention_mask.to(device)
|
289 |
+
|
290 |
+
ce_loss_fn = torch.nn.CrossEntropyLoss(reduction="none")
|
291 |
+
shifted_logits = logits[..., :-1, :].contiguous()
|
292 |
+
shifted_labels = encoding_input_ids[..., 1:].contiguous()
|
293 |
+
shifted_attention_mask = encoding_attention_mask[..., 1:].contiguous()
|
294 |
+
|
295 |
+
# Calculate Cross-Entropy loss
|
296 |
+
cross_entropy_loss = ce_loss_fn(shifted_logits.transpose(1, 2), shifted_labels)
|
297 |
+
# Apply attention mask
|
298 |
+
masked_ce_loss = cross_entropy_loss * shifted_attention_mask
|
299 |
+
# Calculate perplexity
|
300 |
+
ppl = masked_ce_loss.sum(1) / shifted_attention_mask.sum(1)
|
301 |
+
# Move result to CPU and convert to numpy for further processing if needed
|
302 |
+
ppl = ppl.to("cpu").float().numpy()
|
303 |
+
|
304 |
+
return ppl
|
305 |
+
|
306 |
+
def perplexity_all_models(encoding, logits_list):
|
307 |
+
ppl_list = []
|
308 |
+
for logits in logits_list:
|
309 |
+
ppl = perplexity_logits(encoding, logits)
|
310 |
+
ppl_list.append(ppl)
|
311 |
+
return ppl_list
|
312 |
+
|
313 |
+
def blahut_arimoto_torch(W, epsilon=1e-6, max_iters=1000):
|
314 |
+
"""
|
315 |
+
Batch-process Blahut-Arimoto using PyTorch for multiple sequences.
|
316 |
+
"""
|
317 |
+
seq_len, nb_models, voc_size = W.shape
|
318 |
+
p = torch.full((seq_len, nb_models), 1.0 / nb_models, device=W.device, dtype=W.dtype)
|
319 |
+
prod_exp = torch.ones((seq_len, nb_models), device=W.device, dtype=W.dtype)
|
320 |
+
|
321 |
+
for _ in range(max_iters):
|
322 |
+
# Calculate the marginal probabilities
|
323 |
+
sum_p_w = torch.bmm(p.unsqueeze(1), W).squeeze(1) # Resultant shape: (seq_len, voc_size)
|
324 |
+
|
325 |
+
# Calculate normalized probabilities
|
326 |
+
W_normalized = W / sum_p_w.unsqueeze(1) # Broadcasting to shape (seq_len, nb_models, voc_size)
|
327 |
+
|
328 |
+
# Avoid numerical issues with logarithms
|
329 |
+
W_normalized[W_normalized == 0] = torch.finfo(W.dtype).eps
|
330 |
+
log_term = torch.log(W_normalized)
|
331 |
+
log_term[torch.isnan(log_term) | torch.isinf(log_term)] = 0
|
332 |
+
|
333 |
+
# Compute product exponentials and update probabilities
|
334 |
+
prod_exp = torch.exp(torch.sum(W * log_term, axis=2)) # Sum across voc_size
|
335 |
+
p_new = (p * prod_exp) / torch.sum(p * prod_exp, dim=1, keepdim=True)
|
336 |
+
|
337 |
+
# Check convergence
|
338 |
+
if torch.max(torch.abs(p - p_new)) < epsilon:
|
339 |
+
break
|
340 |
+
p = p_new
|
341 |
+
|
342 |
+
# Compute channel capacity
|
343 |
+
capacity = torch.log(torch.sum(p * prod_exp, dim=1)) / torch.log(torch.tensor(2.0, device=W.device))
|
344 |
+
return capacity, p
|