File size: 9,893 Bytes
4014562
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
PROJECT_PATH = 'cleaned_code'

import os
import sys
sys.path.append(PROJECT_PATH)
import numpy as np
import pickle
import h5py
from tqdm import tqdm
from transformers import AutoTokenizer
from scipy.special import expit 
import torch
from typing import Optional
import json
from src import BertForSemanticEmbedding, getLabelModel
from src import DataTrainingArguments, ModelArguments, CustomTrainingArguments, read_yaml_config
from src import dataset_classification_type
from src import SemSupDataset
from transformers import AutoConfig, HfArgumentParser, AutoTokenizer
import torch
import json
from tqdm import tqdm


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


def compute_tok_score_cart(doc_reps, doc_input_ids, qry_reps, qry_input_ids, qry_attention_mask):
    qry_input_ids = qry_input_ids.unsqueeze(2).unsqueeze(3)  # Q * LQ * 1 * 1
    doc_input_ids = doc_input_ids.unsqueeze(0).unsqueeze(1)  # 1 * 1 * D * LD
    exact_match = doc_input_ids == qry_input_ids  # Q * LQ * D * LD
    exact_match = exact_match.float()
    scores_no_masking = torch.matmul(
        qry_reps.view(-1, 16),  # (Q * LQ) * d
        doc_reps.view(-1, 16).transpose(0, 1)  # d * (D * LD)
    )
    scores_no_masking = scores_no_masking.view(
        *qry_reps.shape[:2], *doc_reps.shape[:2])  # Q * LQ * D * LD
    scores, _ = (scores_no_masking * exact_match).max(dim=3)  # Q * LQ * D
    tok_scores = (scores * qry_attention_mask.reshape(-1, qry_attention_mask.shape[-1]).unsqueeze(2))[:, 1:].sum(1)
    
    return tok_scores


def coil_fast_eval_forward(
    input_ids: Optional[torch.Tensor] = None,
    doc_reps = None,
    logits: Optional[torch.Tensor] = None,
    desc_input_ids = None,
    desc_attention_mask = None,
    lab_reps = None,
    label_embeddings = None
):
    tok_scores = compute_tok_score_cart(
            doc_reps, input_ids,
            lab_reps, desc_input_ids.reshape(-1, desc_input_ids.shape[-1]), desc_attention_mask
    )
    logits = (logits.unsqueeze(0) @ label_embeddings.T)
    new_tok_scores = torch.zeros(logits.shape, device = logits.device)
    for i in range(tok_scores.shape[1]):
        stride = tok_scores.shape[0]//tok_scores.shape[1]
        new_tok_scores[i] = tok_scores[i*stride: i*stride + stride ,i]
    return (logits + new_tok_scores).squeeze()



class DemoModel:


    def __init__(self, ):
        self.label_list = [x.strip() for x in open(f'{PROJECT_PATH}/datasets/Amzn13K/all_labels.txt')]
        unseen_label_list = [x.strip() for x in open(f'{PROJECT_PATH}/datasets/Amzn13K/unseen_labels_split6500_2.txt')]
        num_labels = len(self.label_list)
        self.label_list.sort() # For consistency
        l2i = {v: i for i, v in enumerate(self.label_list)}
        unseen_label_indexes = [l2i[x] for x in unseen_label_list]

        self.coil_cluster_map = json.load(open(f'{PROJECT_PATH}/bert_coil_map_dict_lemma255K_isotropic.json'))  



        all_lab_reps1, all_label_embeddings1, all_desc_input_ids_orig1, all_desc_input_ids1, all_desc_attention_mask1 = pickle.load(open(f'{PROJECT_PATH}/precomputed/Amzn13K/amzn_base_labels_data1_1.pkl','rb'))
        all_lab_reps2, all_label_embeddings2, all_desc_input_ids_orig2, all_desc_input_ids2, all_desc_attention_mask2 = pickle.load(open(f'{PROJECT_PATH}/precomputed/Amzn13K/amzn_base_labels_data1_2.pkl','rb'))
        all_lab_reps3, all_label_embeddings3, all_desc_input_ids_orig3, all_desc_input_ids3, all_desc_attention_mask3 = pickle.load(open(f'{PROJECT_PATH}/precomputed/Amzn13K/amzn_base_labels_data1_3.pkl','rb'))
        all_lab_reps4, all_label_embeddings4, all_desc_input_ids_orig4, all_desc_input_ids4, all_desc_attention_mask4 = pickle.load(open(f'{PROJECT_PATH}/precomputed/Amzn13K/amzn_base_labels_data1_4.pkl','rb'))
        all_lab_reps5, all_label_embeddings5, all_desc_input_ids_orig5, all_desc_input_ids5, all_desc_attention_mask5 = pickle.load(open(f'{PROJECT_PATH}/precomputed/Amzn13K/amzn_base_labels_data1_5.pkl','rb'))


        self.all_lab_reps = [all_lab_reps1.to(device), all_lab_reps2.to(device), all_lab_reps3.to(device), all_lab_reps4.to(device), all_lab_reps5.to(device)]
        self.all_label_embeddings = [all_label_embeddings1.to(device), all_label_embeddings2.to(device), all_label_embeddings3.to(device), all_label_embeddings4.to(device), all_label_embeddings5.to(device)]
        self.all_desc_input_ids_orig = [all_desc_input_ids_orig1.to(device), all_desc_input_ids_orig2.to(device), all_desc_input_ids_orig3.to(device), all_desc_input_ids_orig4.to(device), all_desc_input_ids_orig5.to(device)]
        self.all_desc_input_ids = [all_desc_input_ids1.to(device), all_desc_input_ids2.to(device), all_desc_input_ids3.to(device), all_desc_input_ids4.to(device), all_desc_input_ids5.to(device)]
        self.all_desc_attention_mask = [all_desc_attention_mask1.to(device), all_desc_attention_mask2.to(device), all_desc_attention_mask3.to(device), all_desc_attention_mask4.to(device), all_desc_attention_mask5.to(device)]

        ARGS_FILE = f'{PROJECT_PATH}/configs/ablation_amzn_eda.yml'

        parser = HfArgumentParser((ModelArguments, DataTrainingArguments, CustomTrainingArguments))
        self.model_args, self.data_args, self.training_args = parser.parse_dict(read_yaml_config(ARGS_FILE, output_dir = 'demo_tmp',  extra_args = {}))

        config = AutoConfig.from_pretrained(
            self.model_args.config_name if self.model_args.config_name else self.model_args.model_name_or_path,
            finetuning_task=self.data_args.task_name,
            cache_dir=self.model_args.cache_dir,
            revision=self.model_args.model_revision,
            use_auth_token=True if self.model_args.use_auth_token else None,
        )

        config.model_name_or_path = self.model_args.model_name_or_path
        config.problem_type = dataset_classification_type[self.data_args.task_name]
        config.negative_sampling = self.model_args.negative_sampling
        config.semsup = self.model_args.semsup
        config.encoder_model_type = self.model_args.encoder_model_type
        config.arch_type = self.model_args.arch_type
        config.coil = self.model_args.coil
        config.token_dim = self.model_args.token_dim
        config.colbert = self.model_args.colbert

        label_model, label_tokenizer = getLabelModel(self.data_args, self.model_args)
        config.label_hidden_size = label_model.config.hidden_size
        model = BertForSemanticEmbedding(config)
        model.label_model = label_model
        model.label_tokenizer = label_tokenizer
        model.config.label2id = {l: i for i, l in enumerate(self.label_list)}
        model.config.id2label = {id: label for label, id in config.label2id.items()}

        self.tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased')

        model.to(device)
        model.eval()
        torch.set_grad_enabled(False)

        model.load_state_dict(torch.load(f'{PROJECT_PATH}/ckpt/Amzn13K/amzn_main_model.bin', map_location = device))
        self.model = model


        self.extracted_descs = [self.extract_descriptions(adi) for adi in self.all_desc_input_ids_orig]
        tot_len = len(self.all_desc_input_ids_orig)
        for i in range(len(self.all_desc_input_ids_orig[0])):
            for j in range(tot_len):
                if self.extracted_descs[j][i] == "":
                    for k in range(tot_len):
                        if self.extracted_descs[k][i] != '':
                            self.extracted_descs[j][i] = self.extracted_descs[k][i]
                            break
        
            


    def extract_descriptions(self, input_ids):
        descs = self.tokenizer.batch_decode(input_ids, skip_special_tokens = True)
        new_descs = []
        for desc in descs:
            a =  desc.find('description is')
            if a == -1:
                # There is no description to use, lets go with empty
                new_descs.append("")
                continue
            b = min([desc.find(x, a) if desc.find(x, a) !=-1 else 99999999999 for x in ['label is','parents are','children are']])
            if b == 99999999999:
                new_descs.append(desc[a:].strip())
            else:
                new_descs.append(desc[a:b].strip())
        return new_descs

    def classify(self, text, unseen_labels = None):
        self.model.eval()
        with torch.no_grad():
            item = self.tokenizer(text, padding='max_length', max_length=self.data_args.max_seq_length, truncation=True)
            item = {k:torch.tensor(v, device = device).unsqueeze(0) for k,v in item.items()}

            outputs_doc, logits = self.model.forward_input_encoder(**item)
            doc_reps = self.model.tok_proj(outputs_doc.last_hidden_state)

            input_ids = torch.tensor([self.coil_cluster_map[str(x.item())] for x in item['input_ids'][0]]).to(device).unsqueeze(0)
            all_logits = []
            descriptions = []
            for adi, ada, alr, ale in zip(self.all_desc_input_ids, self.all_desc_attention_mask, self.all_lab_reps, self.all_label_embeddings):
                all_logits.append(coil_fast_eval_forward(input_ids, doc_reps, logits, adi, ada, alr, ale))

            final_logits = sum([expit(x.cpu()) for x in all_logits]) / len(all_logits)
            max_indices = torch.argmax(torch.stack(all_logits), dim=0).cpu().tolist()
            # from pdb import set_trace as bp
            # bp()

            outs = torch.topk(final_logits, k =  50)
            preds_dic = dict()
            descs_dic = dict()
            for i,v in zip(outs.indices, outs.values):
                preds_dic[self.label_list[i]] = v.item()
                print(self.extracted_descs[max_indices[i]][i])
                descs_dic[self.label_list[i]] = self.extracted_descs[max_indices[i]][i]
            return preds_dic, descs_dic

if __name__ == '__main__':
    model = DemoModel()
    model.classify('Hello')