-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmodel.py
160 lines (149 loc) · 7.78 KB
/
model.py
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
import torch
from torch import nn
from torch.nn import functional as F
from transformers import AutoTokenizer, AutoModel, AutoModelForSeq2SeqLM, AutoModelForMaskedLM, AutoModelForCausalLM
from transformers import GenerationConfig, LlamaForCausalLM, LlamaTokenizer
from peft import PeftModel
from esm.modules import FeedForwardNetwork, NormalizedResidualBlock
from esm.multihead_attention import MultiheadAttention
def generate_square_subsequent_mask(sz):
mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
mask = mask.float().masked_fill(mask == 0, float('-inf')).masked_fill(mask == 1, float(0.0))
return mask
def create_padding_mask(seq, padding_token=0):
# Create mask with 0s at padding tokens and 1s elsewhere
mask = (seq == padding_token).transpose(0, 1)
return mask
class Text2Mol(nn.Module):
def __init__(self, TextModel, freezeEncoder, MoleculeModel) -> None:
super().__init__()
if "scibert" in TextModel:
self.encoder = AutoModel.from_pretrained('allenai/scibert_scivocab_uncased')
self.text_tokenizer = AutoTokenizer.from_pretrained("allenai/scibert_scivocab_uncased")
if TextModel == "ChemT5":
self.encoder = AutoModelForSeq2SeqLM.from_pretrained("GT4SD/multitask-text-and-chemistry-t5-base-standard").encoder
self.text_tokenizer = AutoTokenizer.from_pretrained("GT4SD/multitask-text-and-chemistry-t5-base-standard")
if "galactica" in TextModel:
TextModel = "facebook/" + TextModel
self.text_tokenizer = AutoTokenizer.from_pretrained(TextModel)
self.text_tokenizer.pad_token= "<pad>"
self.encoder = AutoModelForCausalLM.from_pretrained(TextModel)
if TextModel == "Mol-Instruction":
model = LlamaForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf", load_in_8bit=True, torch_dtype=torch.float16, device_map={"": 0},
token="") # Your huggingface token
self.encoder = PeftModel.from_pretrained(
model,
"adapter/",
torch_dtype=torch.float16,
device_map={"": 0},
)
self.text_tokenizer = LlamaTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf", token="")
self.text_tokenizer.pad_token= "<pad>"
if freezeEncoder:
self.encoder.requires_grad = False
for name, param in self.encoder.named_parameters():
param.requires_grad = False
self.encoder.eval()
else:
self.encoder.train()
self.encoder.requires_grad = True
for name, param in self.encoder.named_parameters():
param.requires_grad = True
if MoleculeModel == "MolGen":
model = AutoModelForSeq2SeqLM.from_pretrained("zjunlp/MolGen-Large")
self.MolGen = model.model.decoder
self.MolGen.lm_head = model.lm_head
elif MoleculeModel == "MolGen-7B":
self.MolGen = AutoModelForCausalLM.from_pretrained("zjunlp/MolGen-7b")
else:
self.MolGen = AutoModelForCausalLM.from_pretrained(MoleculeModel)
self.chemical_adapter_attn = NormalizedResidualBlock(
layer=MultiheadAttention(
self.MolGen.config.hidden_size,
num_heads=16,
kdim=self.encoder.config.hidden_size,
vdim=self.encoder.config.hidden_size,
add_bias_kv=True,
add_zero_attn=False,
use_rotary_embeddings=False,
),
embedding_dim=self.MolGen.config.hidden_size,
dropout=0.1
)
self.chemical_adapter_ffn = NormalizedResidualBlock(
layer=FeedForwardNetwork(
self.MolGen.config.hidden_size,
self.MolGen.config.hidden_size // 2,
activation_dropout=0.1
),
embedding_dim=self.MolGen.config.hidden_size,
dropout=0.1
)
for name, param in self.MolGen.named_parameters():
param.requires_grad = False
self.MolGen.eval()
def forward(self, batch):
text = batch['description']
selfies_token = batch['selfies']
selfies_token = batch['prev_tokens']
text_tokens = self.text_tokenizer(text, return_tensors="pt", padding=True, truncation=True).to(self.encoder.device)
text_output = self.encoder(text_tokens.input_ids, output_hidden_states=True)
text_embedding = text_output.hidden_states[-1].squeeze().transpose(0, 1)
smiles_embedding = self.MolGen(selfies_token, output_hidden_states=True).hidden_states[-1].transpose(0, 1)
dec_output = self.chemical_adapter_attn(
smiles_embedding,
key=text_embedding,
value=text_embedding,
key_padding_mask=text_tokens['attention_mask'].float(),
need_weights=False
)[0]
dec_output = self.chemical_adapter_ffn(dec_output).transpose(0, 1)
decode_smiles_logits = self.MolGen.lm_head(dec_output)
return decode_smiles_logits
def sample_ar(self, batch, temp, cls_idx, greedy):
text = batch['description']
text_tokens = self.text_tokenizer(text, return_tensors="pt", padding=True, truncation=True).to(self.encoder.device)
text_output = self.encoder(text_tokens.input_ids, output_hidden_states=True)
text_embedding = text_output.hidden_states[-1].squeeze().transpose(0, 1)
smiles_token = cls_idx * torch.ones((len(text), 1)).int().to(text_embedding.device)
smiles_embedding = self.MolGen(smiles_token, output_hidden_states=True).hidden_states[-1].transpose(0, 1) # , use_cache=True
smiles_sequence = cls_idx * torch.ones(len(text), 1).int().to(text_embedding.device)
all_logits = []
for step in range(350):
dec_output = self.chemical_adapter_attn(
smiles_embedding,
key=text_embedding,
value=text_embedding,
key_padding_mask=text_tokens['attention_mask'].float(),
need_weights=False
)[0]
dec_output = self.chemical_adapter_ffn(dec_output).transpose(0, 1)
logits = self.MolGen.lm_head(dec_output[:, -1, :])
logits = logits / temp
probs = F.softmax(logits, dim=-1)
if greedy == True:
next_token = torch.argmax(F.softmax(logits, dim=-1), dim=-1).unsqueeze(-1)
else:
next_token = torch.multinomial(probs, num_samples=1)
all_logits.append(logits)
smiles_sequence = torch.cat([smiles_sequence, next_token], dim=1)
smiles_embedding = self.MolGen(smiles_sequence, output_hidden_states=True).hidden_states[-1].transpose(0, 1)
return torch.stack(all_logits).transpose(0, 1), smiles_sequence[:, 1:]
def sample_nar(self, batch):
text = batch['description']
text_tokens = self.text_tokenizer(text, return_tensors="pt", padding=True, truncation=True).to(self.encoder.device)
text_output = self.encoder(**text_tokens)
text_embedding = text_output.last_hidden_state.squeeze().transpose(0, 1)
selfies_token = batch['prev_tokens']
smiles_key_padding_mask = create_padding_mask(selfies_token, padding_token=1).transpose(0, 1)
smiles_embedding = self.MolGen(selfies_token, output_hidden_states=True).hidden_states[0].transpose(0, 1)
dec_output = self.chemical_adapter_attn(
smiles_embedding,
key=text_embedding,
value=text_embedding,
key_padding_mask=text_tokens['attention_mask'].float(),
need_weights=False
)[0]
dec_output = self.chemical_adapter_ffn(dec_output).transpose(0, 1)
decode_smiles_output = self.MolGen.lm_head(dec_output)
return decode_smiles_output