File size: 8,060 Bytes
f831146 |
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 |
import time
import numpy as np
import torch
import tqdm
from torch import optim
import torch.nn.functional as F
from torch.utils.data import DataLoader
from utils.pt_util import restore_model, restore_objects, save_model, save_objects
from data_proc.triplet_loss_dataset import FBanksTripletDataset
from models.triplet_loss_model import FBankTripletLossNet
def _get_cosine_distance(a, b):
return 1 - F.cosine_similarity(a, b)
def train(model, device, train_loader, optimizer, epoch, log_interval):
model.train()
losses = []
positive_accuracy = 0
negative_accuracy = 0
postitive_distances = []
negative_distances = []
for batch_idx, ((ax, ay), (px, py), (nx, ny)) in enumerate(tqdm.tqdm(train_loader)):
ax, px, nx = ax.to(device), px.to(device), nx.to(device)
optimizer.zero_grad()
a_out, p_out, n_out = model(ax, px, nx)
loss = model.loss(a_out, p_out, n_out)
losses.append(loss.item())
with torch.no_grad():
p_distance = _get_cosine_distance(a_out, p_out)
postitive_distances.append(torch.mean(p_distance).item())
n_distance = _get_cosine_distance(a_out, n_out)
negative_distances.append(torch.mean(n_distance).item())
positive_distance_mean = np.mean(postitive_distances)
negative_distance_mean = np.mean(negative_distances)
positive_std = np.std(postitive_distances)
threshold = positive_distance_mean + 3 * positive_std
positive_results = p_distance < threshold
positive_accuracy += torch.sum(positive_results).item()
negative_results = n_distance >= threshold
negative_accuracy += torch.sum(negative_results).item()
loss.backward()
optimizer.step()
if batch_idx % log_interval == 0:
print('{} Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
time.ctime(time.time()),
epoch, batch_idx * len(ax), len(train_loader.dataset),
100. * batch_idx / len(train_loader), loss.item()))
positive_distance_mean = np.mean(postitive_distances)
negative_distance_mean = np.mean(negative_distances)
print('Train Set: positive_distance_mean: {}, negative_distance_mean: {}, std: {}, threshold: {}'.format(
positive_distance_mean, negative_distance_mean, positive_std, threshold))
positive_accuracy_mean = 100. * positive_accuracy / len(train_loader.dataset)
negative_accuracy_mean = 100. * negative_accuracy / len(train_loader.dataset)
return np.mean(losses), positive_accuracy_mean, negative_accuracy_mean
def test(model, device, test_loader, log_interval=None):
model.eval()
losses = []
positive_accuracy = 0
negative_accuracy = 0
postitive_distances = []
negative_distances = []
with torch.no_grad():
for batch_idx, ((ax, ay), (px, py), (nx, ny)) in enumerate(tqdm.tqdm(test_loader)):
ax, px, nx = ax.to(device), px.to(device), nx.to(device)
a_out, p_out, n_out = model(ax, px, nx)
test_loss_on = model.loss(a_out, p_out, n_out, reduction='mean').item()
losses.append(test_loss_on)
p_distance = _get_cosine_distance(a_out, p_out)
postitive_distances.append(torch.mean(p_distance).item())
n_distance = _get_cosine_distance(a_out, n_out)
negative_distances.append(torch.mean(n_distance).item())
positive_distance_mean = np.mean(postitive_distances)
negative_distance_mean = np.mean(negative_distances)
positive_std = np.std(postitive_distances)
threshold = positive_distance_mean + 3 * positive_std
# experiment with this threshold distance to play with accuracy numbers
positive_results = p_distance < threshold
positive_accuracy += torch.sum(positive_results).item()
negative_results = n_distance >= threshold
negative_accuracy += torch.sum(negative_results).item()
if log_interval is not None and batch_idx % log_interval == 0:
print('{} Test: [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
time.ctime(time.time()),
batch_idx * len(ax), len(test_loader.dataset),
100. * batch_idx / len(test_loader), test_loss_on))
test_loss = np.mean(losses)
positive_accuracy_mean = 100. * positive_accuracy / len(test_loader.dataset)
negative_accuracy_mean = 100. * negative_accuracy / len(test_loader.dataset)
positive_distance_mean = np.mean(postitive_distances)
negative_distance_mean = np.mean(negative_distances)
print('Test Set: positive_distance_mean: {}, negative_distance_mean: {}, std: {}, threshold: {}'.format(
positive_distance_mean, negative_distance_mean, positive_std, threshold))
print(
'\nTest set: Average loss: {:.4f}, Positive Accuracy: {}/{} ({:.0f}%), Negative Accuracy: {}/{} ({:.0f}%)\n'.format(
test_loss, positive_accuracy, len(test_loader.dataset), positive_accuracy_mean, negative_accuracy,
len(test_loader.dataset), negative_accuracy_mean))
return test_loss, positive_accuracy_mean, negative_accuracy_mean
def main():
model_path = 'siamese_fbanks_saved/'
use_cuda = True
device = torch.device("cuda" if use_cuda else "cpu")
print('using device', device)
import multiprocessing
print('num cpus:', multiprocessing.cpu_count())
kwargs = {'num_workers': multiprocessing.cpu_count(),
'pin_memory': True} if use_cuda else {}
train_dataset = FBanksTripletDataset('fbanks_train')
train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True, **kwargs)
test_dataset = FBanksTripletDataset('fbanks_test')
test_loader = DataLoader(test_dataset, batch_size=32, shuffle=True, **kwargs)
model = FBankTripletLossNet(margin=0.2).to(device)
model = restore_model(model, model_path)
last_epoch, max_accuracy, train_losses, test_losses, train_positive_accuracies, train_negative_accuracies, \
test_positive_accuracies, test_negative_accuracies = restore_objects(model_path, (0, 0, [], [], [], [], [], []))
start = last_epoch + 1 if max_accuracy > 0 else 0
optimizer = optim.Adam(model.parameters(), lr=0.0005)
for epoch in range(start, start + 20):
train_loss, train_positive_accuracy, train_negative_accuracy = train(model, device, train_loader, optimizer,
epoch, 500)
test_loss, test_positive_accuracy, test_negative_accuracy = test(model, device, test_loader)
print('After epoch: {}, train loss is : {}, test loss is: {}, '
'train positive accuracy: {}, train negative accuracy: {}'
'tes positive accuracy: {}, and test negative accuracy: {} '
.format(epoch, train_loss, test_loss, train_positive_accuracy, train_negative_accuracy,
test_positive_accuracy, test_negative_accuracy))
train_losses.append(train_loss)
test_losses.append(test_loss)
train_positive_accuracies.append(train_positive_accuracy)
test_positive_accuracies.append(test_positive_accuracy)
train_negative_accuracies.append(train_negative_accuracy)
test_negative_accuracies.append(test_negative_accuracy)
test_accuracy = (test_positive_accuracy + test_negative_accuracy) / 2
if test_accuracy > max_accuracy:
max_accuracy = test_accuracy
save_model(model, epoch, model_path)
save_objects((epoch, max_accuracy, train_losses, test_losses, train_positive_accuracies,
train_negative_accuracies, test_positive_accuracies, test_negative_accuracies),
epoch, model_path)
print('saved epoch: {} as checkpoint'.format(epoch))
if __name__ == '__main__':
main() |