HE/train.py at master · morgan76/HE · 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
191
192
193
# -*- coding: utf-8 -*-
import sys

import numpy as np
import pandas as pd
import tqdm
import os
import torch
import torch.optim as optim
from torch.utils.data import DataLoader

from configuration import config as experiment_config
from torch.autograd import Variable
from data import CondTripletDset
from model.CS_Tripletnet import CS_Tripletnet
from model.csn import ConditionalSimNet
from model.models import EmbedNet
from parser import get_parser
from eval_segmentation import eval_segmentation
import utils
from losses import TripletLoss_margins


def train_epoch(model, loss_func, device, train_loader, optimizer, disable=False):
    model.train()
    train_losses = []

    for batch_idx, (data1, data2, data3, c) in enumerate(tqdm.tqdm(train_loader,
                                                        disable=disable)):
        data1, data2, data3, c = data1.to(device), data2.to(device), data3.to(device), c.to(device)
        data1, data2, data3, c = Variable(data1), Variable(data2), Variable(data3), Variable(c)
        embedded_x, embedded_y, embedded_z = model(data1, data2, data3, c)
        loss = loss_func(embedded_x, embedded_z, embedded_y, c)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        train_losses.append(loss.item())

    return np.mean(train_losses)


def validate_using_triplets(model, loss_func, device, valid_loader, epoch,
                              disable=False):
    model.eval()
    with torch.no_grad():
        loss_valid = []
        for batch_idx, (data1, data2, data3, c) in enumerate(tqdm.tqdm(valid_loader, disable=disable)):
            data1, data2, data3, c = data1.to(device), data2.to(device), data3.to(device), c.to(device)
            data1, data2, data3, c = Variable(data1), Variable(data2), Variable(data3), Variable(c)
            embedded_x, embedded_y, embedded_z = model(data1, data2, data3, c)
            valid_loss =  loss_func(embedded_x, embedded_z, embedded_y, c).data.item()
            loss_valid.append(valid_loss)
    return np.mean(loss_valid)

def validate_using_segmentation(tracklist, embedding_net, device, config,
                                  epoch, disable=False, name='valid'):
    results_window_1 = []
    results_window_3 = []
    recalls_3 = []
    precision_3 = []
    embedding_net.eval()
    with torch.no_grad():
        for idx, track in enumerate(tqdm.tqdm(tracklist, disable=disable)):
            #try:
            results1, results2, results3 = eval_segmentation(track, embedding_net, config, device, experiment_config.feat_id, return_data=False)
            results_window_1.append(results1['F1'])
            results_window_3.append(results3['F3'])
            recalls_3.append(results3['R3'])
            precision_3.append(results3['P3'])
            #except:
            #    pass
    F1 = np.mean(results_window_1)
    F3 = np.mean(results_window_3)
    R3 = np.mean(recalls_3)
    P3 = np.mean(precision_3)
    return F1, F3, R3, P3

def train_model(exp_config):
    # Configuration
    print(exp_config)

    use_cuda = torch.cuda.is_available() and not exp_config.no_cuda
    print('Using GPU:', use_cuda)
    # Create 'models' folder if it does not exist
    exp_config.checkpoints_dir.mkdir(exist_ok=True, parents=True)
    exp_config.models_dir.mkdir(exist_ok=True, parents=True)

    # Seed for reproductible experiments
    torch.manual_seed(exp_config.seed)
    torch.cuda.manual_seed(exp_config.seed)
    np.random.seed(exp_config.seed)

    # Dataset
    train_data = CondTripletDset(exp_config, split ='train')
    valid_data = CondTripletDset(exp_config, split ='valid')

    dataloader_kwargs = {'num_workers': exp_config.nb_workers,
                        'pin_memory': True} if use_cuda else {}

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

    # Data loaders
    train_loader = DataLoader(train_data, batch_size=exp_config.batch_size, shuffle=False, **dataloader_kwargs)
    valid_loader = DataLoader(valid_data, batch_size=exp_config.batch_size, shuffle=False, **dataloader_kwargs)


    # Loss function
    loss_func = TripletLoss_margins()

    # Network
    model_path = exp_config.models_dir.joinpath(exp_config.model_name+'.pt')

    if not exp_config.resume:
        print('New model {}.pt will be created'.format(exp_config.model_name))


        model = EmbedNet(exp_config).to(device)
        csn_model = ConditionalSimNet(model, n_conditions=exp_config.n_conditions,
        embedding_size=128, learnedmask=exp_config.learnedmask, prein=exp_config.prein)
        global mask_var
        mask_var = csn_model.masks.weight
        embedding_net = CS_Tripletnet(csn_model)
        embedding_net.to(device)

    else:
        model_temp = EmbedNet(exp_config)
        csn_model_temp = ConditionalSimNet(model_temp, n_conditions=exp_config.n_conditions,
        embedding_size=128, learnedmask=exp_config.learnedmask, prein=exp_config.prein)
        embedding_net = CS_Tripletnet(csn_model_temp)
        model_path = os.path.join(exp_config.models_dir, exp_config.model_name + "." + 'pt')
        embedding_net.load_state_dict(torch.load(model_path)['state_dict'])
        embedding_net.to(device)


    # Optimizer
    parameters = filter(lambda p: p.requires_grad, embedding_net.parameters())
    optimizer = optim.RMSprop(parameters, lr=exp_config.learning_rate, alpha=0.99, eps=1e-08, weight_decay=0, momentum=0, centered=False)


    t = tqdm.trange(1, exp_config.epochs + 1, disable=exp_config.quiet, file=sys.stdout)

    best_F3 = 0
    patience_max = exp_config.patience
    patience = 0

    # Training
    for epoch in t:
        t.set_description('Training Epoch')
        # Training
        train_loss = train_epoch(embedding_net, loss_func, device,
                    train_loader, optimizer, disable=exp_config.quiet)

        # Validate using a fixed set of triplets of the valid part
        valid_loss = validate_using_triplets(embedding_net, loss_func, device,
                                valid_loader, epoch, disable=exp_config.quiet)

        F1, F3, R3, P3 = validate_using_segmentation(valid_data.tracklist, embedding_net,
                                    device, exp_config, epoch,
                                    disable=exp_config.quiet)

        print("Epoch {} Training Loss = {}".format(epoch, train_loss), "Valid Loss = {}".format(valid_loss))

        if (F3 > best_F3) :

            print("\nEpoch {}: best model yet, with a F3 score of {} for a window"
                " of 3 seconds (best was {})\n".format(epoch, F3, best_F3,))

            utils.save_model(exp_config.models_dir, exp_config, epoch,
                        embedding_net, optimizer)

            best_F3 = F3
            patience = 0

        else:

            print("\nEpoch {}: Current F3={}, whereas best is {}\n".format(epoch, F3, best_F3))
            patience += 1
            if patience >= patience_max:
                optimizer.param_groups[0]['lr'] *= 0.5
                print('Changing learning rate: lr =', optimizer.param_groups[0]['lr'])
                patience = 0

def update_config_with_args(args):
    variables = vars(args)
    for var in variables:
        if variables[var] is not None:
            setattr(experiment_config, var, variables[var])


if __name__ == '__main__':
    parser = get_parser()
    args, _ = parser.parse_known_args()
    update_config_with_args(args)