CQH/engine.py at main · Greekatz/CQH · 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

import os
import logging
from collections import OrderedDict
from evaluation import get_db_codes_and_targets
from tqdm import tqdm
import numpy as np
import torch
from network import HyperPQ
import tensorboard
from hierarchical_funcs import compute_features, hier_clus
from loss import *
import time

def train(datahub, model: HyperPQ, loss_fn1, loss_fn2: ProtoLoss,  optimizer, lr_scheduler, config,
          compute_err=True, evaluator=None, monitor=None, writer=None, save_quant_error=None,
          prot_loss_weight=1.0, neighbor_loss_weight=0.1, aug_loss_weight=1.0, start_epoch=0):
    model = model.to(config.device)

    all_quant_error = []

    cluster_result = None
    im2cluster = None
    epoch = start_epoch - 1
    for epoch in range(start_epoch, config.epoch_num):

        if epoch > config.warmup_epoch and (epoch - config.warmup_epoch) % config.clus_interval == 0:
            logging.info('Current epoch is :{}'.format(epoch))
            start = time.time()
            features = compute_features(datahub.clus_loader, model).cpu().numpy()

            cluster_result = {'im2cluster': [], 'centroids': [], 'density': []}
            for num_cluster in config.num_clus_list:
                cluster_result['im2cluster'].append(torch.zeros(len(datahub.clus_loader.dataset),dtype=torch.long).cuda())
                cluster_result['centroids'].append(torch.zeros(int(num_cluster), config.feat_dim).cuda())
                cluster_result['density'].append(torch.zeros(int(num_cluster)).cuda())

            if config.clus_mode == "hier_clus":
                cluster_result = hier_clus(x=features, num_clus_list=config.num_clus_list)
            else:
                raise NotImplementedError("")
            clus_end = time.time()
            logging.info("Hierarchical Clustering takes {}s".format(clus_end - start))

            if cluster_result is not None:
                im2cluster = cluster_result['im2cluster']

        epoch_loss, epoch_aug_inst_loss, epoch_prot_loss, epoch_quant_err, epoch_neighbor_loss, batch_num = 0, 0, 0, 0, 0, len(datahub.train_loader)

        for i, (train_data, index, _) in enumerate(tqdm(datahub.train_loader, desc="epoch %d" % epoch)):
            global_step = i + epoch * batch_num
            view1_data = train_data[0].to(config.device)
            view2_data = train_data[1].to(config.device)


            if lr_scheduler is not None:
                curr_lr = lr_scheduler.step()
                if writer is not None:
                    writer.add_scalar('lr', curr_lr, global_step)
            optimizer.zero_grad()


            view1_feats, view1_hat_feats, view1_soft_codes, view1_err = model(view1_data)
            view2_feats, view2_hat_feats, view2_soft_codes, view2_err = model(view2_data)


            if im2cluster is not None:
                tmp = [i_im2cluster[index] for i_im2cluster in im2cluster]
                aug_inst_loss, neighbor_inst_loss = loss_fn1(view1_hat_feats, view2_hat_feats, view1_feats, view2_feats, model.hyper_pq_head.neg_curvs, im2cluster=tmp)
            else:
                aug_inst_loss, neighbor_inst_loss = loss_fn1(view1_hat_feats, view2_hat_feats, view1_feats, view2_feats, model.hyper_pq_head.neg_curvs, im2cluster=None)


            if writer is not None:
                writer.add_scalar('loss/hyper_aug_inst_loss', aug_inst_loss.item(), global_step)
                writer.add_scalar('loss/hyper_neighbor_inst_loss', neighbor_inst_loss.item(), global_step)


            if cluster_result is not None:
                proto_loss_list = loss_fn2(view1_hat_feats=view1_hat_feats, neg_curvs=model.hyper_pq_head.neg_curvs,
                                           tangent_to_hyper_func=model.hyper_pq_head.tangent_to_hyper,
                                           clus_mode=config.clus_mode, index=index, cluster_result=cluster_result)
                proto_loss = sum(proto_loss_list) / float(len(proto_loss_list))
                if writer is not None:
                    for step, cur_loss in enumerate(proto_loss_list):
                        writer.add_scalar("'loss/hyper_proto_loss_{}".format(step), cur_loss.item(), global_step)

                loss = aug_loss_weight * aug_inst_loss +  prot_loss_weight * proto_loss + neighbor_loss_weight * neighbor_inst_loss

            else:
                proto_loss = 0.
                loss = aug_inst_loss + neighbor_inst_loss
            if writer is not None:
                writer.add_scalar('loss/all_loss', loss.item(), global_step)

            if compute_err:
                quant_err = (view1_err + view2_err) / 2
                epoch_quant_err += quant_err.item()
                all_quant_error.append(quant_err.item())
                if writer is not None:
                    writer.add_scalar('quant_err', quant_err.item(), global_step)

            epoch_loss += loss.item()
            epoch_aug_inst_loss += aug_inst_loss.item()
            epoch_neighbor_loss += neighbor_inst_loss.item()
            if hasattr(proto_loss, 'item'):
                # proto_loss is a tensor
                proto_loss_val = proto_loss.item()
                if proto_loss_val > 0:
                    epoch_prot_loss += proto_loss_val
            elif isinstance(proto_loss, (int, float)) and proto_loss > 0:
                # proto_loss is a scalar (int/float)
                epoch_prot_loss += float(proto_loss)

            loss.backward()
            optimizer.step()

            # keep the neg_curvs > 0
            model.hyper_pq_head.neg_curvs.data.clamp_(min=0.01, max=10)

        logging.info("epoch %d: avg loss=%f,  avg aug inst loss=%f, avg neighbor inst loss=%f, avg proto loss=%f, avg quantization error=%f" %
                     (epoch, epoch_loss / batch_num, epoch_aug_inst_loss / batch_num, epoch_neighbor_loss / batch_num, epoch_prot_loss / batch_num, epoch_quant_err / batch_num))
        if model.hyper_pq_head.alpha_scaler is not None:
            logging.info("epoch {}, --alpha_scaler {}".format(epoch, model.hyper_pq_head.alpha_scaler))

        if evaluator is not None:
            if (epoch+1) % config.eval_interval == 0:
                logging.info("begin to evaluate model")

                hyper_codebooks = model.hyper_codebooks()
                evaluator.set_codebooks(codebooks=hyper_codebooks)
                db_codes, db_targets = get_db_codes_and_targets(datahub.database_loader,
                                                                model, device=config.device)

                evaluator.set_db_codes(db_codes=db_codes)
                evaluator.set_db_targets(db_targets=db_targets)
                logging.info("compute mAP")

                val_mAP = evaluator.MAP(datahub.test_loader, model, topK=config.topK)
                logging.info("val mAP=%f" % val_mAP)
                if writer is not None:
                    writer.add_scalar("val_mAP", val_mAP, epoch)
                if monitor:
                    is_break, is_lose_patience = monitor.update(val_mAP, epoch=epoch, model=model, optimizer=optimizer, scheduler=lr_scheduler)
                    # Save PR curve when we find a new best model
                    if monitor.best_epoch == epoch and hasattr(config, 'save_pr_curve') and config.save_pr_curve:
                        logging.info("New best model found! Saving PR curve...")
                        bits = 16 if config.M == 2 else 32 if config.M == 4 else 64 if config.M == 8 else config.M * 8
                        dataset_name = config.dataset.lower()
                        pr_filename = f"{dataset_name}_{bits}bits_epoch{epoch}_pr_curve.txt"
                        evaluator.compute_and_save_pr_curve(datahub.test_loader, model, pr_filename, topK=config.topK)
                        logging.info(f"PR curve saved as {pr_filename}")
                    if is_break:
                        logging.info("early stop")
                        break


    # epoch is now properly initialized before the loop
    logging.info("finish trainning at epoch %d" % epoch)
    if monitor is not None:
        logging.info("Best Map: {}".format(monitor.best_value))


    if save_quant_error is not None:
        all_quant_error = np.array(all_quant_error)
        np.save(save_quant_error, all_quant_error)


def test(datahub, model, config, evaluator, writer=None):
    '''evaluator must be loaded with correct codebook, db_codes and db_targets'''
    logging.info("compute mAP")
    model = model.to(config.device)
    test_mAP = evaluator.MAP(datahub.test_loader, model, topK=config.topK)
    logging.info("test mAP=%f" % test_mAP)

    logging.info("finish testing")