File size: 6,776 Bytes
1b3a7eb
 
 
 
 
 
 
 
 
 
 
 
85fa1f8
1b3a7eb
 
85fa1f8
 
 
1b3a7eb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
85fa1f8
1b3a7eb
 
 
85fa1f8
 
1b3a7eb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
85fa1f8
1b3a7eb
 
 
 
85fa1f8
 
 
1b3a7eb
 
 
 
 
 
 
 
85fa1f8
1b3a7eb
 
85fa1f8
 
1b3a7eb
 
 
 
 
 
 
 
85fa1f8
1b3a7eb
 
 
85fa1f8
1b3a7eb
 
 
85fa1f8
 
 
1b3a7eb
 
 
 
 
 
 
 
 
 
 
 
 
85fa1f8
 
1b3a7eb
 
85fa1f8
1b3a7eb
 
 
 
 
 
 
 
 
 
 
 
 
 
85fa1f8
1b3a7eb
 
 
85fa1f8
1b3a7eb
85fa1f8
1b3a7eb
85fa1f8
1b3a7eb
 
 
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
import torch
from transformers import GPT2LMHeadModel, GPT2Tokenizer
import gradio as gr
import pandas as pd
from collections import Counter, defaultdict
import os
from huggingface_hub import login

# Get the token from the environment variable
api_token = os.getenv('HF_TOKEN')

# Load pre-trained model and tokenizer
model_name = "gpt2-large"
model = GPT2LMHeadModel.from_pretrained(model_name)
tokenizer = GPT2Tokenizer.from_pretrained(model_name)

device = torch.device("mps") if torch.has_mps else torch.device("cpu")
model.to(device)
model.eval()

def create_ngrams(tokens, n):
    return [tuple(tokens[i:i+n]) for i in range(len(tokens)-n+1)]

def calculate_probabilities(four_gram_counts, three_gram_counts):
    probabilities = defaultdict(lambda: defaultdict(float))
    for four_gram, count in four_gram_counts.items():
        three_gram = four_gram[:-1]
        probabilities[three_gram][four_gram[-1]] = count / three_gram_counts[three_gram]
    return probabilities

def kneser_ney_smoothing(ngram_counts, lower_order_counts, discount=0.75):
    continuation_counts = Counter()
    lower_counts = Counter()

    for ngram in ngram_counts:
        lower_counts[ngram[1:]] += 1
        continuation_counts[ngram[1:]] += 1

    def continuation_probability(word):
        return continuation_counts[word] / sum(continuation_counts.values())

    probabilities = defaultdict(lambda: defaultdict(float))
    for ngram, count in ngram_counts.items():
        lower_ngram = ngram[:-1]
        discounted_count = max(count - discount, 0)
        lambda_factor = (discount / lower_order_counts[lower_ngram]) * len(continuation_counts)
        probabilities[lower_ngram][ngram[-1]] = (discounted_count / lower_order_counts[lower_ngram]) + lambda_factor * continuation_probability(ngram[-1])

    return probabilities

def generate_text_with_probs(initial_context, top_p, max_length, top_k):
    input_ids = tokenizer.encode(initial_context, return_tensors="pt").to(device)
    generated_text = initial_context
    token_tables = []

    token_no = 1

    with torch.no_grad():
        for _ in range(max_length):
            outputs = model(input_ids=input_ids)
            next_token_logits = outputs.logits[:, -1, :]

            # Apply top-p (nucleus) sampling
            sorted_logits, sorted_indices = torch.sort(next_token_logits, descending=True)
            cumulative_probs = torch.cumsum(torch.softmax(sorted_logits, dim=-1), dim=-1)
            sorted_indices_to_remove = cumulative_probs > top_p
            sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
            sorted_indices_to_remove[..., 0] = 0

            indices_to_remove = sorted_indices[sorted_indices_to_remove]
            next_token_logits[:, indices_to_remove] = -float('Inf')
            probabilities = torch.softmax(next_token_logits, dim=-1)

            next_token = torch.multinomial(probabilities, num_samples=1)
            next_token_prob = probabilities[0, next_token].item()
            next_token_text = tokenizer.decode(next_token.item())

            top_tokens = sorted_indices[0, :top_k]
            top_probs = probabilities[0, top_tokens]
            top_token_probs = [(tokenizer.decode([token.item()]), prob.item()) for token, prob in zip(top_tokens, top_probs)]

            df = pd.DataFrame(top_token_probs, columns=["Token", "Probability"])
            df.index = df.index + 1
            token_tables.append((f"{token_no}>> Next token: {next_token_text} (Probability: {next_token_prob:.8f})", df))
            token_no+=1

            input_ids = torch.cat([input_ids, next_token], dim=-1)

            if next_token.item() == tokenizer.eos_token_id:
                break

    generated_text = tokenizer.decode(input_ids[0], skip_special_tokens=True)

    return generated_text[len(initial_context):], token_tables

def predict_next_token_ngram(input_text, context_text, max_length):

    ip = input_text
    context_tokens = tokenizer.tokenize(context_text)
    four_grams = create_ngrams(context_tokens, 4)
    four_gram_counts = Counter(four_grams)
    three_gram_counts = Counter([gram[:-1] for gram in four_grams])
    probabilities = calculate_probabilities(four_gram_counts, three_gram_counts)
    probs = kneser_ney_smoothing(four_gram_counts, three_gram_counts)

    input_tokens = tokenizer.tokenize(input_text)
    generated_tokens = input_tokens.copy()
    generated_text = input_text
    token_tables = []

    if len(input_tokens) >= (max_length + len(generated_tokens)):
        generated_text = tokenizer.convert_tokens_to_string(input_tokens)
        return generated_text, token_tables

    token_no = 1

    while len(input_tokens) < (max_length + len(generated_tokens)):
        input_3_gram = tuple(input_tokens[-3:])
        next_token_probs = probs.get(input_3_gram, {})
        if not next_token_probs:
            break
        next_token = max(next_token_probs, key=next_token_probs.get)
        input_tokens.append(next_token)

        top_k = 4
        top_k_tokens = sorted(next_token_probs.items(), key=lambda x: x[1], reverse=True)[:top_k]
        top_k_tokens_df = pd.DataFrame(top_k_tokens, columns=["Token", "Probability"])
        top_k_tokens_df.index = top_k_tokens_df.index + 1  # Add numbering to the DataFrame
        top_k_tokens_df["Token"] = top_k_tokens_df["Token"].apply(lambda x: tokenizer.convert_tokens_to_string([x]))

        token_tables.append((f"{token_no}>> Next token: {next_token}", top_k_tokens_df))
        token_no+=1

    generated_text = tokenizer.convert_tokens_to_string(input_tokens)
    return generated_text[len(ip):], token_tables

def combined_model_predictions(context_text, initial_context, top_p, max_length, top_k):
    generated_text, token_tables = generate_text_with_probs(initial_context, top_p, max_length, top_k)
    ngram_generated_text, ngram_token_tables = predict_next_token_ngram(initial_context, context_text, max_length)

    return generated_text, token_tables, ngram_generated_text, ngram_token_tables

iface = gr.Interface(
    fn=combined_model_predictions,
    inputs=[
        gr.Textbox(lines=4, placeholder="Enter context for N-gram model..."),
        gr.Textbox(lines=2, placeholder="Enter initial context here..."),
        gr.Slider(0, 1, step=0.01, value=0.9, label="Top-p (nucleus) sampling"),
        gr.Slider(1, 100, step=1, value=50, label="Max length"),
        gr.Slider(1, 50, step=1, value=10, label="Top-k"),
    ],
    outputs=[
        gr.Textbox(label="Generated Text"),
        gr.Dataframe(label="LLM Token Probabilities"),
        gr.Textbox(label="N-gram Generated Text"),
        gr.Dataframe(label="N-gram Token Predictions"),
    ],
    title="Next Token Visualizer (GPT-2-large - 812M param.)"
)

iface.launch()