File size: 12,673 Bytes
6887a13
 
 
 
 
 
 
 
 
f07c563
 
6887a13
 
 
 
f07c563
 
6887a13
 
 
 
 
 
 
 
 
f07c563
6887a13
227e6a4
f07c563
 
 
 
 
1e39a5b
 
 
 
857351f
1e39a5b
 
 
227e6a4
1e39a5b
7d95793
 
1e39a5b
7d95793
 
f4c79c8
 
 
 
 
 
1e39a5b
f4c79c8
 
 
 
1e39a5b
 
f07c563
 
 
 
 
 
 
 
227e6a4
f07c563
 
 
 
 
 
 
1e39a5b
 
 
f07c563
227e6a4
 
 
 
f07c563
 
 
 
 
 
 
 
 
 
 
 
 
227e6a4
 
 
 
 
 
 
 
 
 
 
f07c563
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
227e6a4
 
 
1bdce4c
227e6a4
f803062
 
227e6a4
 
 
f803062
 
227e6a4
 
 
 
 
 
f07c563
 
 
 
227e6a4
6887a13
 
 
 
 
 
 
 
f07c563
 
 
1e39a5b
6887a13
1bacfa9
6887a13
1e39a5b
 
f07c563
 
 
 
 
227e6a4
 
 
 
 
 
 
 
 
7d95793
 
227e6a4
 
 
 
 
 
 
 
 
7d95793
 
227e6a4
 
f07c563
 
 
 
 
227e6a4
 
f07c563
 
 
1bacfa9
227e6a4
 
1bacfa9
227e6a4
 
1bacfa9
f07c563
227e6a4
 
1bacfa9
 
227e6a4
f07c563
 
 
 
 
 
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
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
from typing import Dict, List, Any
import os
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from transformers import PreTrainedTokenizerFast
from transformers import GenerationConfig
import transformers 
import pandas as pd
import time
import numpy as np
from precious3_gpt_multi_modal import Custom_MPTForCausalLM


class EndpointHandler:
    def __init__(self, path=""):

        self.path = path
        # load model and processor from path
        self.model = Custom_MPTForCausalLM.from_pretrained(path, torch_dtype=torch.bfloat16).to('cuda')
        self.tokenizer = PreTrainedTokenizerFast(tokenizer_file = os.path.join(path, "tokenizer.json"), unk_token="[UNK]",
            pad_token="[PAD]",
            eos_token="[EOS]",
            bos_token="[BOS]")
        self.model.config.pad_token_id = self.tokenizer.pad_token_id
        self.model.config.bos_token_id = self.tokenizer.bos_token_id
        self.model.config.eos_token_id = self.tokenizer.eos_token_id
        unique_entities_p3 = pd.read_csv(os.path.join(path, 'p3_entities_with_type.csv'))
        self.unique_compounds_p3 = [i.strip() for i in unique_entities_p3[unique_entities_p3.type=='compound'].entity.to_list()]
        self.unique_genes_p3 = [i.strip() for i in unique_entities_p3[unique_entities_p3.type=='gene'].entity.to_list()]
        
        self.emb_gpt_genes = pd.read_pickle(os.path.join(self.path, 'multi-modal-data/emb_gpt_genes.pickle'))
        self.emb_hgt_genes = pd.read_pickle(os.path.join(self.path, 'multi-modal-data/emb_hgt_genes.pickle'))


    def create_prompt(self, prompt_config):

        prompt = "[BOS]"

        multi_modal_prefix = '<modality0><modality1><modality2><modality3>'*3

        for k, v in prompt_config.items():
            if k=='instruction':
                prompt+=f'<{v}>' if isinstance(v, str) else "".join([f'<{v_i}>' for v_i in v])
            elif k=='up':
                if v:
                    prompt+=f'{multi_modal_prefix}<{k}>{v}</{k}>' if isinstance(v, str) else f'{multi_modal_prefix}<{k}>{" ".join(v)} </{k}>'
            elif k=='down':
                if v:
                    prompt+=f'{multi_modal_prefix}<{k}>{v}</{k}>' if isinstance(v, str) else f'{multi_modal_prefix}<{k}>{" ".join(v)} </{k}>'
            elif k=='age':
                if isinstance(v, int):
                    if prompt_config['species'].strip() == 'human':
                        prompt+=f'<{k}_individ>{v} </{k}_individ>'
                    elif prompt_config['species'].strip() == 'macaque':
                        prompt+=f'<{k}_individ>Macaca-{int(v/20)} </{k}_individ>'
            else:
                if v:
                    prompt+=f'<{k}>{v.strip()} </{k}>' if isinstance(v, str) else f'<{k}>{" ".join(v)} </{k}>'
                else:
                    prompt+=f'<{k}></{k}>'
        return prompt

    def custom_generate(self,
                        input_ids, 
                        acc_embs_up_kg_mean, 
                        acc_embs_down_kg_mean, 
                        acc_embs_up_txt_mean, 
                        acc_embs_down_txt_mean,
                        device, 
                        max_new_tokens,
                        mode, 
                        temperature=0.8, 
                        top_p=0.2, top_k=3550, 
                        n_next_tokens=50, num_return_sequences=1):
        torch.manual_seed(137)

        # Set parameters
        # temperature - Higher value for more randomness, lower for more control
        # top_p - Probability threshold for nucleus sampling (aka top-p sampling)
        # top_k - Ignore logits below the top-k value to reduce randomness (if non-zero)
        # n_next_tokens - Number of top next tokens when predicting compounds

        modality0_emb = torch.unsqueeze(torch.from_numpy(acc_embs_up_kg_mean), 0).to(device) if isinstance(acc_embs_up_kg_mean, np.ndarray) else None
        modality1_emb = torch.unsqueeze(torch.from_numpy(acc_embs_down_kg_mean), 0).to(device) if isinstance(acc_embs_down_kg_mean, np.ndarray) else None
        modality2_emb = torch.unsqueeze(torch.from_numpy(acc_embs_up_txt_mean), 0).to(device) if isinstance(acc_embs_up_txt_mean, np.ndarray) else None
        modality3_emb = torch.unsqueeze(torch.from_numpy(acc_embs_down_txt_mean), 0).to(device) if isinstance(acc_embs_down_txt_mean, np.ndarray) else None


        # Generate sequences
        outputs = []
        next_token_compounds = [] 

        for _ in range(num_return_sequences):
            start_time = time.time()
            generated_sequence = []
            current_token = input_ids.clone()

            for _ in range(max_new_tokens):  # Maximum length of generated sequence
                # Forward pass through the model
                logits = self.model.forward(
                    input_ids=current_token, 
                    modality0_emb=modality0_emb,
                    modality0_token_id=self.tokenizer.encode('<modality0>')[0], # 62191, 
                    modality1_emb=modality1_emb,
                    modality1_token_id=self.tokenizer.encode('<modality1>')[0], # 62192,
                    modality2_emb=modality2_emb,
                    modality2_token_id=self.tokenizer.encode('<modality2>')[0], # 62193, 
                    modality3_emb=modality3_emb,
                    modality3_token_id=self.tokenizer.encode('<modality3>')[0], # 62194
                )[0]

                # Apply temperature to logits
                if temperature != 1.0:
                    logits = logits / temperature

                # Apply top-p sampling (nucleus sampling)
                sorted_logits, sorted_indices = torch.sort(logits, descending=True)
                cumulative_probs = torch.cumsum(torch.softmax(sorted_logits, dim=-1), dim=-1)
                sorted_indices_to_remove = cumulative_probs > top_p

                if top_k > 0:
                    sorted_indices_to_remove[..., top_k:] = 1

                # Set the logit values of the removed indices to a very small negative value
                inf_tensor = torch.tensor(float("-inf")).type(torch.bfloat16).to(logits.device)

                logits = logits.where(sorted_indices_to_remove, inf_tensor)


                # Sample the next token
                if current_token[0][-1] == self.tokenizer.encode('<drug>')[0]:
                    next_token_compounds.append(torch.topk(torch.softmax(logits, dim=-1)[0][len(current_token[0])-1, :].flatten(), n_next_tokens).indices)

                next_token = torch.multinomial(torch.softmax(logits, dim=-1)[0], num_samples=1)[len(current_token[0])-1, :].unsqueeze(0)


                # Append the sampled token to the generated sequence
                generated_sequence.append(next_token.item())

                # Stop generation if an end token is generated
                if next_token == self.tokenizer.eos_token_id:
                    break

                # Prepare input for the next iteration
                current_token = torch.cat((current_token, next_token), dim=-1)
            print(time.time()-start_time)
            outputs.append(generated_sequence)
        
        # Process generated up/down lists
        processed_outputs = {"up": [], "down": []}
        if mode in ['meta2diff', 'meta2diff2compound']:
            for output in outputs:
                up_split_index = output.index(self.tokenizer.convert_tokens_to_ids('</up>'))
                generated_up_raw = [i.strip() for i in self.tokenizer.convert_ids_to_tokens(output[:up_split_index])]
                generated_up = sorted(set(generated_up_raw) & set(self.unique_genes_p3), key = generated_up_raw.index)
                processed_outputs['up'].append(generated_up)
                
                down_split_index = output.index(self.tokenizer.convert_tokens_to_ids('</down>'))
                generated_down_raw = [i.strip() for i in self.tokenizer.convert_ids_to_tokens(output[up_split_index:down_split_index+1])]
                generated_down = sorted(set(generated_down_raw) & set(self.unique_genes_p3), key = generated_down_raw.index)
                processed_outputs['down'].append(generated_down)
                
        else:
            processed_outputs = outputs
        
        predicted_compounds_ids = [self.tokenizer.convert_ids_to_tokens(j) for j in next_token_compounds]
        predicted_compounds = []
        for j in predicted_compounds_ids:
            predicted_compounds.append([i.strip() for i in j])
        return processed_outputs, predicted_compounds


    def __call__(self, data: Dict[str, Any]) -> Dict[str, str]:
        """
        Args:
            data (:dict:):
                The payload with the text prompt and generation parameters.
        """
        torch.manual_seed(137)
        
        device = "cuda"
        config_data = data.pop("inputs", None)
        parameters = data.pop("parameters", None)
        mode = data.pop('mode', 'Not specified')
        
        prompt = self.create_prompt(config_data)
        
        inputs = self.tokenizer(prompt, return_tensors="pt")
        input_ids = inputs["input_ids"].to(device)

        max_new_tokens = 600 - len(input_ids[0]) 
        try:
            if set(["up", "down"]) & set(config_data.keys()):
                acc_embs_up1 = []
                acc_embs_up2 = []
                for gs in config_data['up']: 
                    try:
                        acc_embs_up1.append(self.emb_hgt_genes[self.emb_hgt_genes.gene_symbol==gs].embs.values[0])
                        acc_embs_up2.append(self.emb_gpt_genes[self.emb_gpt_genes.gene_symbol==gs].embs.values[0])
                    except Exception as e: 
                        pass
                acc_embs_up1_mean = np.array(acc_embs_up1).mean(0) if acc_embs_up1 else None
                acc_embs_up2_mean = np.array(acc_embs_up2).mean(0) if acc_embs_up2 else None

                acc_embs_down1 = []
                acc_embs_down2 = []
                for gs in config_data['down']:
                    try:
                        acc_embs_down1.append(self.emb_hgt_genes[self.emb_hgt_genes.gene_symbol==gs].embs.values[0])
                        acc_embs_down2.append(self.emb_gpt_genes[self.emb_gpt_genes.gene_symbol==gs].embs.values[0])
                    except Exception as e: 
                        pass
                acc_embs_down1_mean = np.array(acc_embs_down1).mean(0) if acc_embs_down1 else None
                acc_embs_down2_mean = np.array(acc_embs_down2).mean(0) if acc_embs_down2 else None
            else:
                acc_embs_up1_mean, acc_embs_up2_mean, acc_embs_down1_mean, acc_embs_down2_mean = None, None, None, None

            generated_sequence, raw_next_token_generation = self.custom_generate(input_ids = input_ids, 
                                                             acc_embs_up_kg_mean=acc_embs_up1_mean,
                                                             acc_embs_down_kg_mean=acc_embs_down1_mean, 
                                                             acc_embs_up_txt_mean=acc_embs_up2_mean,
                                                             acc_embs_down_txt_mean=acc_embs_down2_mean, max_new_tokens=max_new_tokens, mode=mode,
                                                             device=device, **parameters)
            next_token_generation = [sorted(set(i) & set(self.unique_compounds_p3), key = i.index) for i in raw_next_token_generation]

            if mode == "meta2diff":
                outputs = {"up": generated_sequence['up'], "down": generated_sequence['down']}
                out = {"output": outputs, "mode": mode, "message": "Done!", "input": prompt}
            elif mode == "meta2diff2compound":
                outputs = {"up": generated_sequence['up'], "down": generated_sequence['down']}
                out = {
                "output": outputs, "compounds": next_token_generation, "raw_output": raw_next_token_generation, "mode": mode, "message": "Done!", "input": prompt}
            elif mode == "diff2compound":
                outputs = generated_sequence
                out = {
                "output": outputs, "compounds": next_token_generation, "raw_output": raw_next_token_generation, "mode": mode, "message": "Done!", "input": prompt}
            else:
                out = {"message": f"Specify one of the following modes: meta2diff, meta2diff2compound, diff2compound. Your mode is: {mode}"}

        except Exception as e:
            print(e)
            outputs, next_token_generation = [None], [None]
            out = {"output": outputs, "mode": mode, 'message': f"{e}"}

        return out