FactFlip/trainer.py at master · softlab-unimore/FactFlip · 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
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
import torch
import time

import numpy as np
import pandas as pd
import torch.nn as nn

from tqdm import tqdm

from utils import output_metrics
from sklearn.decomposition import PCA
from sklearn.metrics.pairwise import cosine_similarity
from transformers import AutoTokenizer

import torch.nn.functional as F

def max_pos(l):
    pos = 0
    max_value = -float('inf')
    for i, el in enumerate(l):
        if el > max_value:
            pos = i
            max_value = el

    return pos

class Trainer:
    def __init__(self, config, device):
        self.config = config
        self.device = device

    def train(self, epoch, model, loss_fn, optimizer, train_loader):
        epoch_start_time = time.time()
        model.train()
        tr_loss = 0

        for batch in tqdm(train_loader, desc='Iteration'):
            batch = tuple(t.to(self.device) if not isinstance(t, list) and not isinstance(t, str) else t for t in batch)
            #batch = tuple(t if not isinstance(t, list) and not isinstance(t, str) else t for t in batch)
            claim, evidence, ids_sent1, segs_sent1, att_mask_sent1, labels = batch

            out = model(ids_sent1, segs_sent1, att_mask_sent1)
            if isinstance(labels, list):
                labels = torch.tensor(np.array(labels)).to(self.device)
            loss = loss_fn(out, labels.float())

            tr_loss += loss.item()

            loss.backward()
            torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
            optimizer.step()
            optimizer.zero_grad()

        timing = time.time() - epoch_start_time
        cur_lr = optimizer.param_groups[0]["lr"]
        print(f"Timing: {timing}, Epoch: {epoch + 1}, training loss: {tr_loss}, current learning rate {cur_lr}")

    def val(self, model, val_loader, return_preds=False):
        model.eval()

        loss_fn = nn.CrossEntropyLoss()

        val_loss = 0
        val_preds = []
        val_labels = []
        outs = []
        for batch in tqdm(val_loader):
            batch = tuple(t.to(self.device) if not isinstance(t, list) and not isinstance(t, str) else t for t in batch)
            claim, evidence, ids_sent1, segs_sent1, att_mask_sent1, labels = batch

            with torch.no_grad():
                out = model(ids_sent1, segs_sent1, att_mask_sent1)
                if out.shape[-1] > 3:
                    # hardcoded initial ids for "support", "refute", "not enough information" for Qwen tokenizer
                    if labels.shape[-1] != 2:
                        out.data = out.data[:, torch.tensor([1824, 83177, 537])] #refute: 83177, contrast: 12872, negate: 71219
                    else:
                        out.data = out.data[:, torch.tensor([1824, 83177])]

                preds = torch.max(out.data, 1)[1].cpu().numpy().tolist()
                print(preds, end=" ")

                loss = loss_fn(out, labels.float())
                val_loss += loss.item()

                outs.extend([out.data.cpu().numpy().tolist()[i] for i in range(len(out)) if torch.max(out.data, 1)[1].cpu().numpy().tolist()[i] == torch.max(labels,1)[1].cpu().numpy().tolist()[i]])

                labels = labels.cpu().numpy().tolist()

                val_labels.extend(labels)
                if len(labels[0]) != 2:
                    for pred in preds:
                        if pred == 0:
                            val_preds.append([1,0,0])
                        elif pred == 1:
                            val_preds.append([0,1,0])
                        else:
                            val_preds.append([0,0,1])
                else:
                    val_preds.extend([[1,0] if pred == 0 else [0,1] for pred in preds])

        print(f"val loss: {val_loss}")

        val_acc, val_prec, val_recall, val_f1 = output_metrics(val_labels, val_preds)

        if return_preds:
            return val_acc, val_prec, val_recall, val_f1, val_preds, val_labels
        return val_acc, val_prec, val_recall, val_f1

class AntonymTrainer:
    def __init__(self, config, device):
        self.config = config
        self.device = device

    def val(self, model, train_loader, **kwargs):
        concept_vectors = {}

        model.eval() # we put it to eval mode because we do not update the gradient
        with torch.no_grad():
            w = model.linear_layer.weight

            for batch in tqdm(train_loader, desc='Training...'):
                batch = tuple(t.to(self.device) if not isinstance(t, list) else t for t in batch)
                sent1, sent2, ids_sent1, segs_sent1, att_mask_sent1, ids_sent2, segs_sent2, att_mask_sent2 = batch
                out, out_sent1, out_sent2 = model.compute_concept_vector(ids_sent1, segs_sent1, att_mask_sent1, ids_sent2, segs_sent2, att_mask_sent2)

                out_normalized = out

                for i in range(len(w)):
                    w_normalized = w[i]
                    cos_sim = torch.matmul(out_normalized, w_normalized.unsqueeze(-1)).squeeze(-1)
                    for j in range(len(out)):
                        if (sent1[j], sent2[j]) not in concept_vectors.keys():
                            concept_vectors[(sent1[j], sent2[j])] = {}
                        if i == 0:
                            concept_vectors[(sent1[j], sent2[j])]["support"] = cos_sim[j].item()
                        elif i == 1:
                            concept_vectors[(sent1[j], sent2[j])]["refute"] = cos_sim[j].item()
                        else:
                            concept_vectors[(sent1[j], sent2[j])]["nei"] = cos_sim[j].item()

        return concept_vectors

    def val_potency(self, model, train_loader, num_labels=2, **kwargs):
        concept_vectors = {}

        model.eval()  # we put it to eval mode because we do not update the gradient
        with torch.no_grad():
            try:
                # roberta
                w = model.linear_layer.weight
            except:
                # qwen
                w = model.plm.lm_head.weight
                if num_labels != 2:
                    w = w[torch.tensor([1824, 83177, 537])] # for qwen2.5
                else:
                    w = w[torch.tensor([1824, 83177])] # for qwen2.5

            for batch in tqdm(train_loader, desc='Training...'):
                batch = tuple(t.to(self.device) if not isinstance(t, list) else t for t in batch)
                #batch = tuple(t if not isinstance(t, list) else t for t in batch)
                # words are provided in pairs following the antonym setup
                # in this case, words are not evaluated wrt each other, so we get the out_sent1 and out_sent2 to evaluate them separately
                # roberta
                sent1, sent2, ids_sent1, segs_sent1, att_mask_sent1, ids_sent2, segs_sent2, att_mask_sent2 = batch
                out, out_sent1, out_sent2 = model.compute_concept_vector(ids_sent1, segs_sent1, att_mask_sent1,
                                                                         ids_sent2, segs_sent2, att_mask_sent2)
                out_sent1 = out_sent1.to(w.device)
                out_sent2 = out_sent2.to(w.device)
                for i in range(len(w)):
                    cos_sim1 = torch.matmul(out_sent1, w[i].unsqueeze(-1)).squeeze(-1)
                    cos_sim2 = torch.matmul(out_sent2, w[i].unsqueeze(-1)).squeeze(-1)
                    for j in range(len(out_sent1)):
                        if sent1[j] not in concept_vectors.keys():
                            concept_vectors[sent1[j]] = {}
                        if sent2[j] not in concept_vectors.keys():
                            concept_vectors[sent2[j]] = {}

                        if i == 0:
                            concept_vectors[sent1[j]]["support"] = cos_sim1[j].item()
                            concept_vectors[sent2[j]]["support"] = cos_sim2[j].item()
                        elif i == 1:
                            concept_vectors[sent1[j]]["refute"] = cos_sim1[j].item()
                            concept_vectors[sent2[j]]["refute"] = cos_sim2[j].item()
                        else:
                            concept_vectors[sent1[j]]["nei"] = cos_sim1[j].item()
                            concept_vectors[sent2[j]]["nei"] = cos_sim2[j].item()

        return concept_vectors