HELM/h_scores.py at main · johnsonandjohnson/HELM · 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
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
162
163
164
165
166
167
168
169
170
171
172
173
174
from functools import partial
import torch
import numpy as np
from torch.nn import functional as F
import pandas as pd
from torchmetrics import Accuracy, MetricCollection
from torchmetrics.text import Perplexity
from dataset import DataAntiBody
from utils import encode_sequence_codon, CODON_TO_ID_3, ID_TO_CODON_3
from lightening_module_hscore import AntibodyLLMLightening
from utils import AA_TO_ID
from create_tree import all_tokens, get_classes
from lightning import Trainer
import pandas as pd
import numpy as np
import torch
from torch.utils.data import Dataset
import pandas as pd
from utils import encode_sequence, AA_TO_ID
from torch.nn import functional as F
import os
import argparse
import os
import torch
from torch.utils.data import Dataset
import pandas as pd
import numpy as np
import yaml

from masking import multi_token_masking
from utils import encode_sequence, AA_TO_ID
from utils import encode_sequence_codon
import torch.nn.functional as F

np.random.seed(42)
noise = np.random.rand(444)

def single_token_masking(ids, pos_ids, mask_ratio, token_to_id, noise, tree=None):
    """Mask a sequence of tokens."""
    assert 0 <= mask_ratio <= 1, "Mask ratio must be between 0 and 1."
    cls_token = token_to_id["<cls>"] if tree is None else tree.index("<cls>")
    eos_token = token_to_id["<eos>"] if tree is None else tree.index("<eos>")
    pad_token = token_to_id["<pad>"] if tree is None else tree.index("<pad>")
    mask_token = token_to_id["<mask>"] if tree is None else tree.index("<mask>")
    ids = np.array(ids)
    special_tokens = (ids == cls_token) | (ids == eos_token) | (ids == pad_token) | (ids == mask_token)
    noise = noise[:len(ids)]
    mask = (noise < np.array([mask_ratio])) & (~special_tokens)
    ids[mask] = mask_token
    return ids, pos_ids

class DataAntiBody(Dataset):
    def __init__(self, path, file, tokenizer, token_to_id, position_ids, masking_strategy, masking_ratio, max_seq_len, mode, extension="txt", column_name=None, tree=None, num_samples=-1):
        p = os.path.join(path, file)
        if extension == "csv":
            assert column_name is not None, "Column name must be provided for csv files."
            df = pd.read_csv(p, skiprows=0)
            if num_samples == -1:
                self.data = df[column_name].values
            else:
                self.data = df[column_name].values[0:num_samples]

        if extension == "txt":
            with open(p, "r") as f:
                if num_samples == -1:
                    self.data = f.readlines()
                else:
                    self.data = f.readlines()[0:num_samples]

        self.data = tokenizer(self.data)
        self.position_ids = position_ids
        self.masking_strategy = masking_strategy
        self.masking_ratio = masking_ratio
        self.max_seq_len = max_seq_len
        self.mode = mode
        self.token_to_id = token_to_id
        self.pad_token = self.token_to_id["<pad>"] if tree is None else tree.index("<pad>")
        self.lens = [len(d) for d in self.data]
        self.tree = tree
        print(max(self.lens))

    def __len__(self):
        return len(self.data)

    def __getitem__(self, idx):
        if self.mode == "mlm":
            sequences = self.data[idx].copy()
            pos_ids = np.arange(len(self.data[idx]))
            if "single" in self.masking_strategy:
                sequences, pos_ids = single_token_masking(sequences, pos_ids, self.masking_ratio, self.token_to_id, noise, self.tree)
            if "film" in self.masking_strategy:
                sequences, pos_ids = multi_token_masking(sequences, pos_ids, 1, mask_ratio=0.15, token_to_id=self.token_to_id)

            seq_size = sequences.shape[0]

            if seq_size >= self.max_seq_len:
                sequences = sequences[:self.max_seq_len]
                pos_ids = pos_ids[:self.max_seq_len]
                target = self.data[idx][:self.max_seq_len]

            else:
                sequences = F.pad(torch.tensor(sequences), (0, self.max_seq_len - len(self.data[idx])), 'constant',
                                self.pad_token).squeeze(0)
                pos_ids = torch.concatenate([torch.tensor(pos_ids), torch.arange(seq_size, sequences.shape[0])])
                target = F.pad(torch.tensor(self.data[idx]), (0, self.max_seq_len - len(self.data[idx])), 'constant',
                            self.pad_token).squeeze(0)

            return torch.tensor(sequences), torch.tensor(pos_ids), torch.tensor(target)

        elif self.mode == "next_token_pred":
            data = self.data[idx]
            data = F.pad(torch.tensor(data), (0, self.max_seq_len - len(self.data[idx])), 'constant', self.pad_token).squeeze(0)
            pos_ids = torch.arange(len(data) - 1)
            return torch.tensor(data[:-1]), torch.tensor(pos_ids), torch.tensor(data[1:])


mode = "next_token_pred"
model = "xe"
path = "./downstream/gpt_clm/last.ckpt"
tree = get_classes(all_tokens)[0]
from hxe import HierarchicalCrossEntropyLoss
from create_tree import all_tokens,  get_weighting, get_classes
weights = get_weighting(all_tokens, "exponential", value=0.2)
classes = get_classes(all_tokens)[0]
if model == "xe":
    criterion = torch.nn.CrossEntropyLoss()
    tokenizer = partial(encode_sequence_codon, tree=None)
    tree = None
elif model == "hxe":
    criterion = HierarchicalCrossEntropyLoss(all_tokens, classes, weights)
    CODON_TO_ID = CODON_TO_ID_3
    tree = classes
    tokenizer = partial(encode_sequence_codon, tree=tree)

from models.transformer import TransformerConfig, Transformer
model_config = TransformerConfig()
model_config.model = "gpt2"
model_config.amp = True
model_config.token_to_id = CODON_TO_ID_3
model_config.bidirectional = True
model_config.vocab_size = 70
model_config.d_model = 640
model_config.d_intermediate = 2560
model_config.n_layer = 10
model_config.pad_vocab_size_multiple = 1
model_config.distillation = False
model_config.max_position_embeddings = 1024
model_config.pos_embedding = "absolute"
model_config.pos_embedding_apply = "add"
model_config.pos_embedding_dim = None
model_config.distillation_mode = None
model_config.num_heads = 8
model_config.attn_dropout = 0.0
model_config.tree = tree if model == "hxe" else None
model_config.attn_layer_idx = [i for i in range(10)]
model_config.tie_embeddings = False
model = Transformer(config=model_config)
label_metrics = MetricCollection(
    [Accuracy(task="multiclass", num_classes=model.vocab_size)])
logits_metrics = MetricCollection([Perplexity()])
lm = AntibodyLLMLightening.load_from_checkpoint(path, model=model, train_mode=mode, distill_mode=False, distill_type=None, distill_alpha=None, distill_temp=None, teacher_model=None, criterion=criterion, optimizer=None, scheduler=None, label_based_metrics=label_metrics, logits_based_metrics=logits_metrics, token_to_id=CODON_TO_ID_3, tree=tree)
t = Trainer(accelerator="gpu", precision="bf16-mixed")
data = DataAntiBody("./downstream/", "oas_codonized_v2_test_split.csv", tokenizer, CODON_TO_ID_3, 0, ["single"], 0.15, 444, mode, "csv", "codonized_all", tree, 1000)
dataloader = torch.utils.data.DataLoader(data, batch_size=32, shuffle=False, num_workers=8)
a = t.predict(lm, dataloaders=dataloader)
output = [b[0] for b in a]
output = np.concatenate(output, axis=0)
target = [b[1] for b in a]
target = np.concatenate(target, axis=0)

import pickle
with open("./downstream/xe.pkl", "wb") as f:
    pickle.dump((output, target), f)