itda/interp.py at main · pleask/itda · GitHub

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
"""
Code for manual inspection of ITDA latent activations.
"""
# %%

import os
from typing import Optional

import torch
import torch.nn as nn
import torch.nn.functional as F
import yaml
from datasets import load_dataset
from dictionary_learning.dictionary import Dictionary
from dictionary_learning.trainers.trainer import SAETrainer
from example_saes.train import ITDA, model_configs
from IPython.display import HTML, display
from tqdm import tqdm
from transformer_lens import HookedTransformer
from transformers import AutoModel, AutoTokenizer


def highlight_string(tokens, idx, tokenizer, crop=-1):
    str_ = ""

    if crop != -1:
        start_idx = max(0, idx - crop)
        end_idx = min(len(tokens), idx + crop + 1)
        tokens = tokens[start_idx:end_idx]
        idx -= start_idx

    for i, token in enumerate(tokens):
        token_str = tokenizer.decode([token]).replace("\n", "")
        if i == idx:
            str_ += f'<span style="background-color: darkblue; font-weight: bold;">{token_str}</span>'
        else:
            str_ += f"{token_str}"

    return HTML(str_)


# %%

if __name__ == "__main__":
    torch.backends.cudnn.allow_tf32 = True
    torch.backends.cuda.matmul.allow_tf32 = True
    device = "cuda" if torch.cuda.is_available() else "cpu"
    torch.set_grad_enabled(False)

    run = "f87no63b"
    itda = ITDA.from_pretrained(f"artifacts/runs/{run}")

    dataset_name = "NeelNanda/pile-10k"
    seq_len = 128
    batch_size = 256
    max_steps = int(10_000 / batch_size + 1)

    model_name = "EleutherAI/pythia-70m-deduped"
    model = HookedTransformer.from_pretrained(model_name, device=device)
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token

    layer = model_configs[model_name]["layer"]

# %%

if __name__ == "__main__":
    latents = [1619]

    dataset = load_dataset(dataset_name, split="train", streaming=True)
    data_stream = (item["text"] for item in dataset)

    latent_activations = []
    all_tokens = []
    progress = tqdm(range(max_steps), desc="Getting activations", unit="step")
    for step in progress:
        batch = []
        for _ in range(batch_size):
            try:
                batch.append(next(data_stream))
            except StopIteration:
                break
        if not batch:
            print("Data stream exhausted.")
            break

        tokens = tokenizer(
            batch,
            padding="max_length",
            truncation=True,
            max_length=seq_len,
            return_tensors="pt",
        )
        tokens = {k: v[:, :seq_len].to(device) for k, v in tokens.items()}
        all_tokens.append(tokens["input_ids"].cpu())
        _, cache = model.run_with_cache(
            tokens["input_ids"],
            stop_at_layer=layer + 1,
            names_filter=[f"blocks.{layer}.hook_resid_post"],
        )
        model_activations = cache[f"blocks.{layer}.hook_resid_post"]

        itda_activations = itda.encode(model_activations)
        latent_activations.append(itda_activations[:, :, latents].cpu())

        # clear the cuda cache
        torch.cuda.empty_cache()
    latent_activations = torch.cat(latent_activations, dim=0)
    all_tokens = torch.cat(all_tokens, dim=0)

# %%

if __name__ == "__main__":
    latent_idx = 0

    values, indices = torch.topk(latent_activations[:, :, latent_idx].flatten(), 10)
    if values.allclose(torch.zeros_like(values)):
        print("No activations found.")
    else:
        rows, cols = (
            torch.div(
                indices,
                latent_activations[:, :, latent_idx].shape[1],
                rounding_mode="floor",
            ),
            indices % latent_activations[:, :, latent_idx].shape[1],
        )
        top_10_indices = torch.stack((rows, cols), dim=1)

        for seq_idx, tok_idx in top_10_indices:
            display(highlight_string(all_tokens[seq_idx], tok_idx, tokenizer, crop=10))