SAE-TM/train_sae_cached.py at main · ExplainableML/SAE-TM · 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
133
134
135
import logging
import torch
import numpy as np
import argparse
import os

from torch.utils.data import IterableDataset, DataLoader
from libraries.dictionary_learning.training import trainSAE
from libraries.dictionary_learning.trainers import StandardTrainer
from libraries.dictionary_learning.trainers import JumpReluTrainer
from libraries.dictionary_learning.trainers import BatchTopKTrainer
from libraries.dictionary_learning.trainers import TopKTrainer

logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')

sparse_autoencoder_trainers = {
    "StandardTrainer": StandardTrainer,
    "JumpReluTrainer": JumpReluTrainer,
    "BatchTopKTrainer": BatchTopKTrainer,
    "TopKTrainer": TopKTrainer
}


class CachedEmbeddingsDataset(IterableDataset):
    def __init__(self, path_to_embeddings: str):
        self.path_to_embeddings = path_to_embeddings
        if os.path.isdir(self.path_to_embeddings):
            self.embedding_files = os.listdir(self.path_to_embeddings)
            self.embedding_files = [file for file in self.embedding_files if file.endswith(".pt")]
        elif os.path.isfile(self.path_to_embeddings):
            self.embedding_files = [os.path.basename(self.path_to_embeddings)]
            self.path_to_embeddings = os.path.dirname(self.path_to_embeddings)
        else:
            raise ValueError(f"Path {path_to_embeddings} is neither a directory nor a file.")

    def __iter__(self):
        # Get worker info
        worker_info = torch.utils.data.get_worker_info()
        if worker_info is not None:
            embedding_files = self.embedding_files[worker_info.id::worker_info.num_workers]
            generator = np.random.default_rng(seed=worker_info.id)
        else:
            embedding_files = self.embedding_files
            generator = np.random.default_rng(seed=42)

        # Cycle over embedding files indefinitely so that training can run
        # for the full number of requested steps (trainSAE breaks at step >= steps)
        while True:
            if len(embedding_files) > 1:
                generator.shuffle(embedding_files)
            for file in embedding_files:
                embeddings = torch.load(os.path.join(self.path_to_embeddings, file), map_location="cpu").to(dtype=torch.float8_e4m3fn)
                permuted_indices = generator.permutation(len(embeddings))
                for idx in permuted_indices:
                    yield embeddings[idx].to(dtype=torch.float32)

    @staticmethod
    def collate_fn(batch):
        return torch.stack(batch, dim=0)


if __name__ == "__main__":
    # Parse arguments
    parser = argparse.ArgumentParser()
    parser.add_argument("--trainer", type=str, default='StandardTrainer')
    parser.add_argument("--expansion-factor", type=int, default=16)
    parser.add_argument("--lr", type=float, default=1e-3)
    parser.add_argument("--top-k", type=int, default=32)
    parser.add_argument("--l1-penalty", type=float, default=1e-1)
    parser.add_argument("--batch-size", type=int, default=1024)
    parser.add_argument("--num-workers", type=int, default=1)
    parser.add_argument("--steps", type=int, default=200)
    parser.add_argument("--warmup-ratio", type=float, default=0.1)
    parser.add_argument("--save-path", type=str, default='results/trained_models/')
    parser.add_argument("--save-interval", type=int, default=100)
    parser.add_argument("--seed", type=int, default=112233)
    parser.add_argument("--path-to-embeddings", type=str, default="data/embedding_datasets/")
    args = parser.parse_args()

    # Set seed
    torch.manual_seed(args.seed)
    torch.cuda.manual_seed(args.seed)
    np.random.seed(args.seed)

    # Load dataset
    dataset = CachedEmbeddingsDataset(path_to_embeddings=args.path_to_embeddings)
    dataloader = DataLoader(
        dataset, batch_size=args.batch_size, num_workers=args.num_workers, shuffle=False, collate_fn=CachedEmbeddingsDataset.collate_fn
    )

    embedding_dim = next(iter(dataset)).shape[-1]

    # Make save path
    experiment_name = f"trainer={args.trainer}_expansion_factor={args.expansion_factor}"
    hparams_str = f"lr={args.lr}_top_k={args.top_k}_l1_penalty={args.l1_penalty}_batch_size={args.batch_size}_steps={args.steps}_warmup_ratio={args.warmup_ratio}"
    save_path = os.path.join(args.save_path, experiment_name, hparams_str)
    os.makedirs(save_path, exist_ok=True)

    # Get device
    device = 'cuda' if torch.cuda.is_available() else 'cpu'

    # Train SAE
    steps = args.steps
    warmup_steps = args.warmup_ratio * steps
    save_steps = list(range(0, steps, args.save_interval)) + [steps-1]
    trainer = sparse_autoencoder_trainers[args.trainer]

    trainer_cfg = {
        "trainer": trainer,
        "activation_dim": embedding_dim,
        "dict_size": args.expansion_factor * embedding_dim,
        "lr": args.lr,
        "device": device,
        "steps": steps,
        "lm_name": "cached_embeddings",
        "layer": "embedding",
        "warmup_steps": warmup_steps,
    }

    if args.trainer in ["StandardTrainer", "JumpReluTrainer"]:
        trainer_cfg["sparsity_warmup_steps"] = warmup_steps
        trainer_cfg["l1_penalty"] = args.l1_penalty
    elif args.trainer in ["BatchTopKTrainer", "TopKTrainer"]:
        trainer_cfg["k"] = args.top_k

    trainSAE(
        data=dataloader,
        trainer_configs=[trainer_cfg],
        steps=steps,
        save_steps=save_steps,
        save_dir=save_path,
        verbose=True,
        log_steps=200,
        device=device
    )