cdiff/experiment.py at main · networkslab/cdiff · 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
import math
import torch
from cross_diffusion_utils.experiment import DiffusionExperiment, add_exp_args
import numpy as np
# Metric
from metrics import get_distances_diffusion, time_rmse_tensor, type_rmse_diffusion, sMape_metric
from scipy.special import boxcox, inv_boxcox

import tpp_utils_seq2seq.dataset_seq2seq.Constants as Constants


class Experiment(DiffusionExperiment):
    def train_fn(self, epoch):

        self.model.train()
        loss_sum = 0.0
        dt_loss_sum = 0.0
        type_loss_sum = 0.0
        loss_count = 0
        loss_moving = None

        for iteration, batch in enumerate(self.train_loader):

            history_times = batch.history_times
            hist_e = batch.history_types.long()
            hist_x = batch.history_dt
            tgt_e = batch.target_types.long()
            tgt_x = batch.target_dt
            loss, dt_loss, type_loss = self.model(hist_x, hist_e, tgt_x, tgt_e, history_times)
            loss.backward()
            torch.nn.utils.clip_grad_norm_(self.model.parameters(), 1.)

            self.ema.update(self.model)

            if (iteration + 1) % self.args.update_freq == 0:
                self.optimizer.step()
                self.optimizer.zero_grad()
                # if self.scheduler_iter: self.scheduler_iter.step()
                self.scheduler_epoch.step()

            loss_sum += loss.detach().cpu().item() * len(tgt_x)
            dt_loss_sum += dt_loss.detach().cpu().item() * len(tgt_x)
            type_loss_sum += type_loss.detach().cpu().item() * len(tgt_x)
            loss_count += len(tgt_x)

            if loss_moving is None:
                loss_moving = loss.detach().cpu().item()
            else:
                loss_moving = .99 * loss_moving + .01 * loss.detach().cpu().item()

            if self.args.debug and loss_count > self.args.debug:
                break
            print('Training. Epoch: {}/{}, Datapoint: {}/{}, Total loss : {:.3f}, '
                  'Total time loss: {:.3f}, Total type loss: {:.3f}'.format(epoch + 1, self.args.epochs,
                                                                            loss_count,
                                                                            len(self.train_loader.dataset),
                                                                            loss_sum / loss_count,
                                                                            dt_loss_sum / loss_count,
                                                                            type_loss_sum / loss_count), end='\r')
        print('')
        # if self.scheduler_epoch: self.scheduler_epoch.step()
        return {'total_loss': loss_sum / loss_count, 'dt_loss': dt_loss_sum / loss_count,
                'type_loss': type_loss_sum / loss_count}

    def eval_fn(self, epoch):
        self.model.eval()
        print()
        with torch.no_grad():
            loss_sum = 0.0
            dt_loss_sum = 0.0
            type_loss_sum = 0.0
            loss_count = 0

            batch_count = 0
            total_distances_wo_filter = []
            total_rmse_types_wo_filter = []
            total_smape = []

            for iteration, batch in enumerate(self.eval_loader):
                history_times = batch.history_times
                hist_e = batch.history_types.long()
                hist_x = batch.history_dt
                tgt_e = batch.target_types.long()
                tgt_x = batch.target_dt
                unnormed_target_dt = batch.unnormed_target_dt

                batch_count += 1

                pred_e, pred_x = self.model.sample(hist_x, hist_e, self.args.tgt_len, history_times)

                if self.args.boxcox:
                    # https://stats.stackexchange.com/questions/541748/simple-problem-with-box-cox-transformation-in-a-time-series-model
                    # Why need clamp, this website gives the answer

                    pred_x = pred_x * self.args.train_bc_std + self.args.train_bc_mean
                    if self.args.train_lambda_boxcox > 0:
                        pred_x[
                            pred_x < -1 / self.args.train_lambda_boxcox] = -1 / self.args.train_lambda_boxcox + Constants.EPS * 100
                    else:
                        pred_x[
                            pred_x > -1 / self.args.train_lambda_boxcox] = -1 / self.args.train_lambda_boxcox - Constants.EPS * 100
                    pred_x = inv_boxcox(pred_x.cpu(), self.args.train_lambda_boxcox) / self.args.scale
                    pred_x[pred_x < 0] = ((self.args.min_inter_time + Constants.EPS) * 0.85).to(self.args.device)
                else:
                    pred_x = pred_x * self.args.train_ln_std + self.args.train_ln_mean
                    pred_x = torch.exp(pred_x)
                    if self.args.dataset == 'retweet':
                        pred_x = pred_x / Constants.SCALE_RETWEET
                    else:
                        pred_x = pred_x / Constants.SCALE_UNIFORM
                    pred_x[pred_x < 0] = ((self.args.min_inter_time + Constants.EPS) * 0.85).to(self.args.device)

                pred_x[pred_x < 0] = (self.args.min_inter_time * 0.85 + Constants.EPS).to(self.args.device)

                pred_e = pred_e.cpu().long()
                pred_x = pred_x.cpu()
                gt_e = tgt_e.cpu().long()
                gt_x = unnormed_target_dt.cpu() + Constants.EPS
                gt_e = gt_e
                gt_x = gt_x

                ########################################################################################################
                ############################################ OTD w/o filter ############################################
                ########################################################################################################

                filter = False
                distances_wo_filter = get_distances_diffusion(pred_x, pred_e, gt_x, gt_e, self.args.num_classes, filter,
                                                              self.args.time_range, self.args.distance_del_cost,
                                                              self.args.trans_cost)

                total_distances_wo_filter += list(np.array(distances_wo_filter))

                ##############################################################################################################
                ############################################ Type RMSE w/o filter ############################################
                ##############################################################################################################

                filter = False
                rmse_types_wo_filter = type_rmse_diffusion(pred_x, pred_e, gt_x, gt_e, self.args.num_classes, filter,
                                                           self.args.time_range)

                total_rmse_types_wo_filter += list(np.array(rmse_types_wo_filter))

                ##############################################################################################
                ############################################ sMAPE ###########################################
                ##############################################################################################

                s_ape = sMape_metric(pred_x.cpu(), gt_x.cpu())
                total_smape += list(np.array(s_ape))

                loss, dt_loss, type_loss = self.model(hist_x, hist_e, tgt_x, tgt_e, history_times)
                loss_sum += loss.detach().cpu().item() * len(tgt_e)
                dt_loss_sum += dt_loss.detach().cpu().item() * len(tgt_e)
                type_loss_sum += type_loss.detach().cpu().item() * len(tgt_e)
                loss_count += len(tgt_e)

            # Find mean to report for val
            total_distances_wo_filter = np.mean(total_distances_wo_filter)
            total_rmse_types_wo_filter = np.mean(total_rmse_types_wo_filter)
            total_smape = np.mean(total_smape)
            print('Evaluating train for N={} forecasting. Epoch: {}/{}, Datapoint: {}/{}, Total loss: {:.3f}, '
                  'Time loss: {:.3f}, Type loss: {:3f} '
                  'OTD fixed forecasting: {:.3f}, rmse_type fixed forecasting: {:.3f},'
                  'sMAPE: {:.3f}'.format(self.args.tgt_len,
                epoch + 1, self.args.epochs, loss_count, len(self.eval_loader.dataset), loss_sum / loss_count,
                dt_loss_sum / loss_count, type_loss_sum / loss_count,
                total_distances_wo_filter, total_rmse_types_wo_filter, total_smape), end='\r')
            print('')

        return {'total_loss': loss_sum / loss_count, 'dt_loss': dt_loss_sum / loss_count,
                'type_loss': type_loss_sum / loss_count,
                'otd_wo_filter': total_distances_wo_filter,
                'rmse_type_wo_filter': total_rmse_types_wo_filter,
                'smape': total_smape}