def train(self):
self.sess = tf.Session()
self.sess.run(tf.global_variables_initializer())
learn_rate = self.lr
for i in range(self.iter_num + 1):
att_batch, img_batch, lab_batch = data_iterator(
self.sess, self.train_data, self.selected_num, self.ave_img_pl,
self.ave_att_pl, self.Weights_encoder)
dataset = LoadDataset(att_batch, img_batch, lab_batch)
next_batch = dataset.get_batch
con_att_batch, con_img_batch, con_lab_batch = next_batch(
self.batch_size)
self.sess.run(self.optimizer,
feed_dict={
self.att_pl: att_batch,
self.img_pl: img_batch,
self.con_img_pl: con_img_batch,
self.con_att_pl: con_att_batch,
self.similarity_float: con_lab_batch,
self.lr_pl: learn_rate
})
if i >= 1000:
learn_rate = 5e-5
if i % 200 == 0:
print('the %d-th iter' % i)
self.test()
def validate(opt):
param = _param()
dataset = LoadDataset(opt)
param.X_dim = dataset.feature_dim
# data_layer = FeatDataLayer(dataset.labels_train, dataset.pfc_feat_data_train, opt)
# initialize model
netGs = []
checkpoint = torch.load(opt.resume)
parts = 7
for part in range(parts):
netG = _netG(dataset.text_dim, 512).cuda()
netG.load_state_dict(checkpoint['state_dict_G' + str(part)])
netGs.append(netG)
train_classifier(opt, param, dataset, netGs)
def train(creative_weight=1000, model_num=1, is_val=True):
param = _param()
if opt.dataset == 'CUB':
dataset = LoadDataset(opt, main_dir, is_val)
exp_info = 'CUB_EASY' if opt.splitmode == 'easy' else 'CUB_HARD'
elif opt.dataset == 'NAB':
dataset = LoadDataset_NAB(opt, main_dir, is_val)
exp_info = 'NAB_EASY' if opt.splitmode == 'easy' else 'NAB_HARD'
else:
print('No Dataset with that name')
sys.exit(0)
param.X_dim = dataset.feature_dim
opt.Creative_weight = creative_weight
data_layer = FeatDataLayer(dataset.labels_train,
dataset.pfc_feat_data_train, opt)
result = Result()
ones = Variable(torch.Tensor(1, 1))
ones.data.fill_(1.0)
netG = _netG(dataset.text_dim, dataset.feature_dim).cuda()
netG.apply(weights_init)
if model_num == 6:
netD = _netD(dataset.train_cls_num + 1, dataset.feature_dim).cuda()
else:
netD = _netD(dataset.train_cls_num, dataset.feature_dim).cuda()
netD.apply(weights_init)
if model_num == 2:
log_SM_ab = Scale(2)
log_SM_ab = nn.DataParallel(log_SM_ab).cuda()
elif model_num == 4 or model_num == 5:
log_SM_ab = Scale(1)
log_SM_ab = nn.DataParallel(log_SM_ab).cuda()
exp_params = 'Model_{}_CAN{}_Eu{}_Rls{}_RWz{}_{}'.format(
model_num, opt.Creative_weight, opt.CENT_LAMBDA, opt.REG_W_LAMBDA,
opt.REG_Wz_LAMBDA, opt.exp_name)
out_subdir = main_dir + 'out/{:s}/{:s}'.format(exp_info, exp_params)
if not os.path.exists(out_subdir):
os.makedirs(out_subdir)
log_dir = out_subdir + '/log_{:s}.txt'.format(exp_info)
with open(log_dir, 'a') as f:
f.write('Training Start:')
f.write(strftime("%a, %d %b %Y %H:%M:%S +0000", gmtime()) + '\n')
start_step = 0
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
netG.load_state_dict(checkpoint['state_dict_G'])
netD.load_state_dict(checkpoint['state_dict_D'])
start_step = checkpoint['it']
print(checkpoint['log'])
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
if model_num == 2 or model_num == 4 or model_num == 5:
nets = [netG, netD, log_SM_ab]
else:
nets = [netG, netD]
tr_cls_centroid = Variable(
torch.from_numpy(dataset.tr_cls_centroid.astype('float32'))).cuda()
optimizerD = optim.Adam(netD.parameters(), lr=opt.lr, betas=(0.5, 0.9))
optimizerG = optim.Adam(netG.parameters(), lr=opt.lr, betas=(0.5, 0.9))
if model_num == 2 or model_num == 4 or model_num == 5:
optimizer_SM_ab = optim.Adam(log_SM_ab.parameters(),
lr=opt.lr,
betas=(0.5, 0.999))
for it in tqdm(range(start_step, 3000 + 1)):
# Creative Loss
blobs = data_layer.forward()
labels = blobs['labels'].astype(int)
new_class_labels = Variable(
torch.from_numpy(np.ones_like(labels) *
dataset.train_cls_num)).cuda()
text_feat_1 = np.array(
[dataset.train_text_feature[i, :] for i in labels])
text_feat_2 = np.array(
[dataset.train_text_feature[i, :] for i in labels])
np.random.shuffle(
text_feat_1
) # Shuffle both features to guarantee different permutations
np.random.shuffle(text_feat_2)
alpha = (np.random.random(len(labels)) * (.8 - .2)) + .2
text_feat_mean = np.multiply(alpha, text_feat_1.transpose())
text_feat_mean += np.multiply(1. - alpha, text_feat_2.transpose())
text_feat_mean = text_feat_mean.transpose()
text_feat_mean = normalize(text_feat_mean, norm='l2', axis=1)
text_feat_Creative = Variable(
torch.from_numpy(text_feat_mean.astype('float32'))).cuda()
z_creative = Variable(torch.randn(opt.batchsize, param.z_dim)).cuda()
G_creative_sample = netG(z_creative, text_feat_Creative)
""" Discriminator """
for _ in range(5):
blobs = data_layer.forward()
feat_data = blobs['data'] # image data
labels = blobs['labels'].astype(int) # class labels
text_feat = np.array(
[dataset.train_text_feature[i, :] for i in labels])
text_feat = Variable(torch.from_numpy(
text_feat.astype('float32'))).cuda()
X = Variable(torch.from_numpy(feat_data)).cuda()
y_true = Variable(torch.from_numpy(labels.astype('int'))).cuda()
z = Variable(torch.randn(opt.batchsize, param.z_dim)).cuda()
# GAN's D loss
D_real, C_real = netD(X)
D_loss_real = torch.mean(D_real)
C_loss_real = F.cross_entropy(C_real, y_true)
DC_loss = -D_loss_real + C_loss_real
DC_loss.backward()
# GAN's D loss
G_sample = netG(z, text_feat).detach()
D_fake, C_fake = netD(G_sample)
D_loss_fake = torch.mean(D_fake)
C_loss_fake = F.cross_entropy(C_fake, y_true)
DC_loss = D_loss_fake + C_loss_fake
DC_loss.backward()
# train with gradient penalty (WGAN_GP)
grad_penalty = calc_gradient_penalty(netD, X.data, G_sample.data)
grad_penalty.backward()
Wasserstein_D = D_loss_real - D_loss_fake
optimizerD.step()
reset_grad(nets)
""" Generator """
for _ in range(1):
blobs = data_layer.forward()
feat_data = blobs['data'] # image data
labels = blobs['labels'].astype(int) # class labels
text_feat = np.array(
[dataset.train_text_feature[i, :] for i in labels])
text_feat = Variable(torch.from_numpy(
text_feat.astype('float32'))).cuda()
X = Variable(torch.from_numpy(feat_data)).cuda()
y_true = Variable(torch.from_numpy(labels.astype('int'))).cuda()
z = Variable(torch.randn(opt.batchsize, param.z_dim)).cuda()
G_sample = netG(z, text_feat)
D_fake, C_fake = netD(G_sample)
_, C_real = netD(X)
# GAN's G loss
G_loss = torch.mean(D_fake)
# Auxiliary classification loss
C_loss = (F.cross_entropy(C_real, y_true) +
F.cross_entropy(C_fake, y_true)) / 2
GC_loss = -G_loss + C_loss
# Centroid loss
Euclidean_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.REG_W_LAMBDA != 0:
for i in range(dataset.train_cls_num):
sample_idx = (y_true == i).data.nonzero().squeeze()
if sample_idx.numel() == 0:
Euclidean_loss += 0.0
else:
G_sample_cls = G_sample[sample_idx, :]
Euclidean_loss += (
G_sample_cls.mean(dim=0) -
tr_cls_centroid[i]).pow(2).sum().sqrt()
Euclidean_loss *= 1.0 / dataset.train_cls_num * opt.CENT_LAMBDA
# ||W||_2 regularization
reg_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.REG_W_LAMBDA != 0:
for name, p in netG.named_parameters():
if 'weight' in name:
reg_loss += p.pow(2).sum()
reg_loss.mul_(opt.REG_W_LAMBDA)
# ||W_z||21 regularization, make W_z sparse
reg_Wz_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.REG_Wz_LAMBDA != 0:
Wz = netG.rdc_text.weight
reg_Wz_loss = Wz.pow(2).sum(dim=0).sqrt().sum().mul(
opt.REG_Wz_LAMBDA)
# D(C| GX_fake)) + Classify GX_fake as real
D_creative_fake, C_creative_fake = netD(G_creative_sample)
if model_num == 1: # KL Divergence
G_fake_C = F.log_softmax(C_creative_fake)
else:
G_fake_C = F.softmax(C_creative_fake)
if model_num == 1: # KL Divergence
entropy_GX_fake = (G_fake_C / G_fake_C.data.size(1)).mean()
elif model_num == 2: # SM Divergence
q_shape = Variable(
torch.FloatTensor(G_fake_C.data.size(0),
G_fake_C.data.size(1))).cuda()
q_shape.data.fill_(1.0 / G_fake_C.data.size(1))
SM_ab = F.sigmoid(log_SM_ab(ones))
SM_a = 0.2 + torch.div(SM_ab[0][0], 1.6666666666666667).cuda()
SM_b = 0.2 + torch.div(SM_ab[0][1], 1.6666666666666667).cuda()
pow_a_b = torch.div(1 - SM_a, 1 - SM_b)
alpha_term = (torch.pow(G_fake_C + 1e-5, SM_a) *
torch.pow(q_shape, 1 - SM_a)).sum(1)
entropy_GX_fake_vec = torch.div(
torch.pow(alpha_term, pow_a_b) - 1, SM_b - 1)
elif model_num == 3: # Bachatera Divergence
q_shape = Variable(
torch.FloatTensor(G_fake_C.data.size(0),
G_fake_C.data.size(1))).cuda()
q_shape.data.fill_(1.0 / G_fake_C.data.size(1))
SM_a = Variable(torch.FloatTensor(1, 1)).cuda()
SM_a.data.fill_(opt.SM_Alpha)
SM_b = Variable(torch.FloatTensor(1, 1)).cuda()
SM_b.data.fill_(opt.SM_Alpha)
pow_a_b = torch.div(1 - SM_a, 1 - SM_b)
alpha_term = (torch.pow(G_fake_C + 1e-5, SM_a) *
torch.pow(q_shape, 1 - SM_a)).sum(1)
entropy_GX_fake_vec = -torch.div(
torch.pow(alpha_term, pow_a_b) - 1, SM_b - 1)
elif model_num == 4: # Tsallis Divergence
q_shape = Variable(
torch.FloatTensor(G_fake_C.data.size(0),
G_fake_C.data.size(1))).cuda()
q_shape.data.fill_(1.0 / G_fake_C.data.size(1))
SM_ab = F.sigmoid(log_SM_ab(ones))
SM_a = 0.2 + torch.div(SM_ab[0][0], 1.6666666666666667).cuda()
SM_b = SM_a
pow_a_b = torch.div(1 - SM_a, 1 - SM_b)
alpha_term = (torch.pow(G_fake_C + 1e-5, SM_a) *
torch.pow(q_shape, 1 - SM_a)).sum(1)
entropy_GX_fake_vec = -torch.div(
torch.pow(alpha_term, pow_a_b) - 1, SM_b - 1)
elif model_num == 5: # Renyi Divergence
q_shape = Variable(
torch.FloatTensor(G_fake_C.data.size(0),
G_fake_C.data.size(1))).cuda()
q_shape.data.fill_(1.0 / G_fake_C.data.size(1))
SM_ab = F.sigmoid(log_SM_ab(ones))
SM_a = 0.2 + torch.div(SM_ab[0][0], 1.6666666666666667).cuda()
SM_b = Variable(torch.FloatTensor(1, 1)).cuda()
SM_b.data.fill_(opt.SM_Beta)
pow_a_b = torch.div(1 - SM_a, 1 - SM_b)
alpha_term = (torch.pow(G_fake_C + 1e-5, SM_a) *
torch.pow(q_shape, 1 - SM_a)).sum(1)
entropy_GX_fake_vec = -torch.div(
torch.pow(alpha_term, pow_a_b) - 1, SM_b - 1)
if model_num == 6:
loss_creative = F.cross_entropy(C_creative_fake,
new_class_labels)
else:
if model_num != 1:
# Normalize SM-Divergence & Report mean
min_e, max_e = torch.min(entropy_GX_fake_vec), torch.max(
entropy_GX_fake_vec)
entropy_GX_fake_vec = (entropy_GX_fake_vec -
min_e) / (max_e - min_e)
entropy_GX_fake = -entropy_GX_fake_vec.mean()
loss_creative = -opt.Creative_weight * entropy_GX_fake
disc_GX_fake_real = -torch.mean(D_creative_fake)
total_loss_creative = loss_creative + disc_GX_fake_real
all_loss = GC_loss + Euclidean_loss + reg_loss + reg_Wz_loss + total_loss_creative
all_loss.backward()
if model_num == 2 or model_num == 4 or model_num == 5:
optimizer_SM_ab.step()
optimizerG.step()
reset_grad(nets)
if it % opt.disp_interval == 0 and it:
acc_real = (np.argmax(C_real.data.cpu().numpy(), axis=1)
== y_true.data.cpu().numpy()).sum() / float(
y_true.data.size()[0])
acc_fake = (np.argmax(C_fake.data.cpu().numpy(), axis=1)
== y_true.data.cpu().numpy()).sum() / float(
y_true.data.size()[0])
log_text = 'Iter-{}; rl: {:.4}%; fk: {:.4}%'.format(
it, acc_real * 100, acc_fake * 100)
with open(log_dir, 'a') as f:
f.write(log_text + '\n')
if it % opt.evl_interval == 0 and it > opt.disp_interval:
netG.eval()
cur_acc = eval_fakefeat_test(it, netG, dataset, param, result)
cur_auc = eval_fakefeat_GZSL(netG, dataset, param, out_subdir,
result)
if cur_acc > result.best_acc:
result.best_acc = cur_acc
if cur_auc > result.best_auc:
result.best_auc = cur_auc
if it % opt.save_interval:
files2remove = glob.glob(out_subdir + '/Best_model*')
for _i in files2remove:
os.remove(_i)
torch.save(
{
'it': it + 1,
'state_dict_G': netG.state_dict(),
'state_dict_D': netD.state_dict(),
'random_seed': opt.manualSeed,
'log': log_text,
}, out_subdir +
'/Best_model_AUC_{:.2f}.tar'.format(cur_auc))
netG.train()
return result
def train():
param = _param()
dataset = LoadDataset(opt)
param.X_dim = dataset.feature_dim
data_layer = FeatDataLayer(dataset.labels_train, dataset.pfc_feat_data_train, opt)
result = Result()
netG = _netG(dataset.text_dim, dataset.feature_dim).cuda()
netG.apply(weights_init)
print(netG)
netD = _netD(dataset.train_cls_num, dataset.feature_dim).cuda()
netD.apply(weights_init)
print(netD)
exp_info = 'CUB_EASY' if opt.splitmode == 'easy' else 'CUB_HARD'
exp_params = 'Eu{}_Rls{}_RWz{}'.format(opt.CENT_LAMBDA , opt.REG_W_LAMBDA, opt.REG_Wz_LAMBDA)
out_dir = 'out/{:s}'.format(exp_info)
out_subdir = 'out/{:s}/{:s}'.format(exp_info, exp_params)
if not os.path.exists('out'):
os.mkdir('out')
if not os.path.exists(out_dir):
os.mkdir(out_dir)
if not os.path.exists(out_subdir):
os.mkdir(out_subdir)
cprint(" The output dictionary is {}".format(out_subdir), 'red')
log_dir = out_subdir + '/log_{:s}.txt'.format(exp_info)
with open(log_dir, 'a') as f:
f.write('Training Start:')
f.write(strftime("%a, %d %b %Y %H:%M:%S +0000", gmtime()) + '\n')
start_step = 0
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
netG.load_state_dict(checkpoint['state_dict_G'])
netD.load_state_dict(checkpoint['state_dict_D'])
start_step = checkpoint['it']
print(checkpoint['log'])
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
nets = [netG, netD]
tr_cls_centroid = Variable(torch.from_numpy(dataset.tr_cls_centroid.astype('float32'))).cuda()
optimizerD = optim.Adam(netD.parameters(), lr=opt.lr, betas=(0.5, 0.9))
optimizerG = optim.Adam(netG.parameters(), lr=opt.lr, betas=(0.5, 0.9))
for it in range(start_step, 3000+1):
""" Discriminator """
for _ in range(5):
blobs = data_layer.forward()
feat_data = blobs['data'] # image data
labels = blobs['labels'].astype(int) # class labels
text_feat = np.array([dataset.train_text_feature[i,:] for i in labels])
text_feat = Variable(torch.from_numpy(text_feat.astype('float32'))).cuda()
X = Variable(torch.from_numpy(feat_data)).cuda()
y_true = Variable(torch.from_numpy(labels.astype('int'))).cuda()
z = Variable(torch.randn(opt.batchsize, param.z_dim)).cuda()
# GAN's D loss
D_real, C_real = netD(X)
D_loss_real = torch.mean(D_real)
C_loss_real = F.cross_entropy(C_real, y_true)
DC_loss = -D_loss_real + C_loss_real
DC_loss.backward()
# GAN's D loss
G_sample = netG(z, text_feat).detach()
D_fake, C_fake = netD(G_sample)
D_loss_fake = torch.mean(D_fake)
C_loss_fake = F.cross_entropy(C_fake, y_true)
DC_loss = D_loss_fake + C_loss_fake
DC_loss.backward()
# train with gradient penalty (WGAN_GP)
grad_penalty = calc_gradient_penalty(netD, X.data, G_sample.data)
grad_penalty.backward()
Wasserstein_D = D_loss_real - D_loss_fake
optimizerD.step()
reset_grad(nets)
""" Generator """
for _ in range(1):
blobs = data_layer.forward()
feat_data = blobs['data'] # image data
labels = blobs['labels'].astype(int) # class labels
text_feat = np.array([dataset.train_text_feature[i, :] for i in labels])
text_feat = Variable(torch.from_numpy(text_feat.astype('float32'))).cuda()
X = Variable(torch.from_numpy(feat_data)).cuda()
y_true = Variable(torch.from_numpy(labels.astype('int'))).cuda()
z = Variable(torch.randn(opt.batchsize, param.z_dim)).cuda()
G_sample = netG(z, text_feat)
D_fake, C_fake = netD(G_sample)
_, C_real = netD(X)
# GAN's G loss
G_loss = torch.mean(D_fake)
# Auxiliary classification loss
C_loss = (F.cross_entropy(C_real, y_true) + F.cross_entropy(C_fake, y_true))/2
GC_loss = -G_loss + C_loss
# Centroid loss
Euclidean_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.REG_W_LAMBDA != 0:
for i in range(dataset.train_cls_num):
sample_idx = (y_true == i).data.nonzero().squeeze()
if sample_idx.numel() == 0:
Euclidean_loss += 0.0
else:
G_sample_cls = G_sample[sample_idx, :]
Euclidean_loss += (G_sample_cls.mean(dim=0) - tr_cls_centroid[i]).pow(2).sum().sqrt()
Euclidean_loss *= 1.0/dataset.train_cls_num * opt.CENT_LAMBDA
# ||W||_2 regularization
reg_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.REG_W_LAMBDA != 0:
for name, p in netG.named_parameters():
if 'weight' in name:
reg_loss += p.pow(2).sum()
reg_loss.mul_(opt.REG_W_LAMBDA)
# ||W_z||21 regularization, make W_z sparse
reg_Wz_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.REG_Wz_LAMBDA != 0:
Wz = netG.rdc_text.weight
reg_Wz_loss = Wz.pow(2).sum(dim=0).sqrt().sum().mul(opt.REG_Wz_LAMBDA)
all_loss = GC_loss + Euclidean_loss + reg_loss + reg_Wz_loss
all_loss.backward()
optimizerG.step()
reset_grad(nets)
if it % opt.disp_interval == 0 and it:
acc_real = (np.argmax(C_real.data.cpu().numpy(), axis=1) == y_true.data.cpu().numpy()).sum() / float(y_true.data.size()[0])
acc_fake = (np.argmax(C_fake.data.cpu().numpy(), axis=1) == y_true.data.cpu().numpy()).sum() / float(y_true.data.size()[0])
log_text = 'Iter-{}; Was_D: {:.4}; Euc_ls: {:.4}; reg_ls: {:.4}; Wz_ls: {:.4}; G_loss: {:.4}; D_loss_real: {:.4};' \
' D_loss_fake: {:.4}; rl: {:.4}%; fk: {:.4}%'\
.format(it, Wasserstein_D.data[0], Euclidean_loss.data[0], reg_loss.data[0],reg_Wz_loss.data[0],
G_loss.data[0], D_loss_real.data[0], D_loss_fake.data[0], acc_real * 100, acc_fake * 100)
print(log_text)
with open(log_dir, 'a') as f:
f.write(log_text+'\n')
if it % opt.evl_interval == 0 and it >= 100:
netG.eval()
eval_fakefeat_test(it, netG, dataset, param, result)
if result.save_model:
files2remove = glob.glob(out_subdir + '/Best_model*')
for _i in files2remove:
os.remove(_i)
torch.save({
'it': it + 1,
'state_dict_G': netG.state_dict(),
'state_dict_D': netD.state_dict(),
'random_seed': opt.manualSeed,
'log': log_text,
}, out_subdir + '/Best_model_Acc_{:.2f}.tar'.format(result.acc_list[-1]))
netG.train()
if it % opt.save_interval == 0 and it:
torch.save({
'it': it + 1,
'state_dict_G': netG.state_dict(),
'state_dict_D': netD.state_dict(),
'random_seed': opt.manualSeed,
'log': log_text,
}, out_subdir + '/Iter_{:d}.tar'.format(it))
cprint('Save model to ' + out_subdir + '/Iter_{:d}.tar'.format(it), 'red')
import torch.nn as nn
import torch.nn.functional as F
from torch import optim
from torch.autograd import Variable
from torch.utils.data import DataLoader
from datetime import datetime
from dataset import LoadDataset
from evalution_segmentaion import eval_semantic_segmentation
from Models import FCN
import cfg
device = t.device('cuda') if t.cuda.is_available() else t.device('cpu')
num_class = cfg.DATASET[1]
Load_train = LoadDataset([cfg.TRAIN_ROOT, cfg.TRAIN_LABEL], cfg.crop_size)
Load_val = LoadDataset([cfg.VAL_ROOT, cfg.VAL_LABEL], cfg.crop_size)
train_data = DataLoader(Load_train, batch_size=cfg.BATCH_SIZE, shuffle=True, num_workers=1)
val_data = DataLoader(Load_val, batch_size=cfg.BATCH_SIZE, shuffle=True, num_workers=1)
fcn = FCN.FCN(num_class)
fcn = fcn.to(device)
criterion = nn.NLLLoss().to(device)
optimizer = optim.Adam(fcn.parameters(), lr=1e-4)
def train(model):
best = [0]
net = model.train()
def train():
param = _param()
dataset = LoadDataset(opt)
param.X_dim = dataset.feature_dim
data_layer = FeatDataLayer(dataset.labels_train,
dataset.pfc_feat_data_train, opt)
# initialize model
netGs = []
netDs = []
parts = 6 #if opt.dataset == "CUB2011" else 6
for part in range(parts):
netGs.append(_netG(dataset.text_dim, 512).cuda().apply(weights_init))
netDs.append(
_netD(dataset.train_cls_num, 512).cuda().apply(weights_init))
exp_info = 'CUB_EASY' if opt.splitmode == 'easy' else 'CUB_HARD'
exp_params = 'Eu{}_Rls{}_RWz{}'.format(opt.CENT_LAMBDA, opt.REG_W_LAMBDA,
opt.REG_Wz_LAMBDA)
out_dir = 'out/{:s}'.format(exp_info)
out_subdir = 'out/{:s}/{:s}'.format(exp_info, exp_params)
if not os.path.exists('out'):
os.mkdir('out')
if not os.path.exists(out_dir):
os.mkdir(out_dir)
if not os.path.exists(out_subdir):
os.mkdir(out_subdir)
cprint(" The output dictionary is {}".format(out_subdir), 'red')
log_dir = out_subdir + '/log_{:s}.txt'.format(exp_info)
with open(log_dir, 'a') as f:
f.write('Training Start:')
f.write(strftime("%a, %d %b %Y %H:%M:%S +0000", gmtime()) + '\n')
start_step = 0
part_cls_centrild = torch.from_numpy(
dataset.part_cls_centrild.astype('float32')).cuda()
# initialize optimizers
optimizerGs = []
optimizerDs = []
for netG in netGs:
optimizerGs.append(
optim.Adam(netG.parameters(), lr=opt.lr, betas=(0.5, 0.9)))
for netD in netDs:
optimizerDs.append(
optim.Adam(netD.parameters(), lr=opt.lr, betas=(0.5, 0.9)))
for it in range(start_step, 3000 + 1):
""" Discriminator """
for _ in range(5):
blobs = data_layer.forward()
feat_data = blobs['data'] # image data
labels = blobs['labels'].astype(int) # class labels
text_feat = np.array(
[dataset.train_text_feature[i, :] for i in labels])
text_feat = torch.from_numpy(text_feat.astype('float32')).cuda()
X = torch.from_numpy(feat_data).cuda()
y_true = torch.from_numpy(labels.astype('int')).cuda()
z = torch.randn(opt.batchsize, param.z_dim).cuda()
for part in range(parts):
z = torch.randn(opt.batchsize, param.z_dim).cuda()
D_real, C_real = netDs[part](X[:, part * 512:(part + 1) * 512])
D_loss_real = torch.mean(D_real)
C_loss_real = F.cross_entropy(C_real, y_true)
DC_loss = -D_loss_real + C_loss_real
DC_loss.backward()
G_sample = netGs[part](z, text_feat)
D_fake, C_fake = netDs[part](G_sample)
D_loss_fake = torch.mean(D_fake)
C_loss_fake = F.cross_entropy(C_fake, y_true)
DC_loss = D_loss_fake + C_loss_fake
DC_loss.backward()
grad_penalty = calc_gradient_penalty(
netDs[part], X.data[:, part * 512:(part + 1) * 512],
G_sample.data)
grad_penalty.backward()
Wasserstein_D = D_loss_real - D_loss_fake
# writer.add_scalar("Wasserstein_D"+str(part), Wasserstein_D.item(), it)
optimizerDs[part].step()
netGs[part].zero_grad()
netDs[part].zero_grad()
""" Generator """
for _ in range(1):
blobs = data_layer.forward()
feat_data = blobs['data'] # image data
labels = blobs['labels'].astype(int) # class labels
text_feat = np.array(
[dataset.train_text_feature[i, :] for i in labels])
text_feat = torch.from_numpy(text_feat.astype('float32')).cuda()
X = torch.from_numpy(feat_data).cuda()
y_true = torch.from_numpy(labels.astype('int')).cuda()
for part in range(parts):
z = torch.randn(opt.batchsize, param.z_dim).cuda()
G_sample = netGs[part](z, text_feat)
# G_sample_all[:, part*512:(part+1)*512] = G_sample
D_fake, C_fake = netDs[part](G_sample)
_, C_real = netDs[part](X[:, part * 512:(part + 1) * 512])
G_loss = torch.mean(D_fake)
C_loss = (F.cross_entropy(C_real, y_true) +
F.cross_entropy(C_fake, y_true)) / 2
GC_loss = -G_loss + C_loss
# writer.add_scalar("GC_loss"+str(part), GC_loss.item(), it)
Euclidean_loss = torch.tensor([0.0]).cuda()
if opt.REG_W_LAMBDA != 0:
for i in range(dataset.train_cls_num):
sample_idx = (y_true == i).data.nonzero().squeeze()
if sample_idx.numel() == 0:
Euclidean_loss += 0.0
else:
G_sample_cls = G_sample[sample_idx, :]
Euclidean_loss += (G_sample_cls.mean(dim=0) -
part_cls_centrild[i][part]
).pow(2).sum().sqrt()
Euclidean_loss *= 1.0 / dataset.train_cls_num * opt.CENT_LAMBDA
# ||W||_2 regularization
reg_loss = torch.Tensor([0.0]).cuda()
if opt.REG_W_LAMBDA != 0:
for name, p in netGs[part].named_parameters():
if 'weight' in name:
reg_loss += p.pow(2).sum()
reg_loss.mul_(opt.REG_W_LAMBDA)
# writer.add_scalar("reg_loss"+str(part), reg_loss.item(), it)
# ||W_z||21 regularization, make W_z sparse
reg_Wz_loss = torch.Tensor([0.0]).cuda()
if opt.REG_Wz_LAMBDA != 0:
Wz = netGs[part].rdc_text.weight
reg_Wz_loss = reg_Wz_loss + Wz.pow(2).sum(
dim=0).sqrt().sum().mul(opt.REG_Wz_LAMBDA)
# writer.add_scalar("reg_Wz_loss"+str(part), reg_Wz_loss.item(), it)
all_loss = GC_loss + Euclidean_loss + reg_loss + reg_Wz_loss
all_loss.backward()
optimizerGs[part].step()
if it % opt.evl_interval == 0 and it >= 1000:
print(it)
for part in range(parts):
netGs[part].eval()
train_classifier(opt, param, dataset, netGs)
for part in range(parts):
netGs[part].train()
def train(self):
# 新建保存 checkpoint 的文件夹
if not os.path.exists(self.checkpoint_save):
os.makedirs(self.checkpoint_save)
for epoch in range(epochs):
print("=====This is Epoch =====", epoch)
'''
if epoch < 2:
for idx, p in enumerate(self.model.parameters()):
if self.flag[idx] == 0:
p.requires_grad = False
else:
for idx, p in enumerate(self.model.parameters()):
p.requires_grad = True
fixed = 1
'''
if epoch % decay == 0:
for param_group in self.optimizer.param_groups:
param_group['lr'] = self.lr * 0.1
self.lr = param_group['lr']
print("learning rate : {}".format(
self.optimizer.state_dict()['param_groups'][0]['lr']))
learning_rate = str(
self.optimizer.state_dict()['param_groups'][0]['lr'])
dir_num = 0
for dir in self.train_list:
i = 0
train_set = LoadDataset(dir)
train_loader = DataLoader(train_set,
batch_size=batch_size,
num_workers=0,
shuffle=True)
for idx, train_data in enumerate(train_loader):
audio, gt = train_data
# print('Train : ', audio.shape, gt.shape)
# exit(0)
# torch.Size([bs, 3, f_num, 256, 320])
# torch.Size([bs, 3, f_num, 256, 320])
# torch.Size([bs, 1, 256, 320])
# image = Variable(image).cuda()
audio = Variable(audio).cuda()
gt = Variable(gt).cuda()
det = self.model(audio)
# print('Result : ', det.shape, gt.shape)
# exit(0)
# loss1 = obj_func(att, gt)
loss2 = obj_func(det, gt)
loss = loss2
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
self.loss_dict.append(loss.item())
i += 1
print(
'Epoch: [{}], Root_folder: {}, Frames: {}, Enumerate: {}\n=====Loss: {:.6f}=====\n'
.format(epoch, dir, len(train_set), i, loss.item()))
'''
dir_num += 1
if dir_num == 500:
break
'''
if epoch % save_hop == 0:
torch.save(
self.model.state_dict(),
os.path.join(
self.checkpoint_save, 'model_%s_%s_%s_%d_%s_%d.pth' %
(checkpoint_name, self.optimizer_name, learning_rate,
batch_size, self.func_name, epoch)))
end_time = str(datetime.now())[11:13] + str(datetime.now())[14:16]
save_name = './charts/' + self.optimizer_name + '_' + str(
lr) + '_' + self.func_name + '_' + end_time + '.png'
plt.title(self.optimizer_name + '_' + str(lr) + '_' + str(epochs) +
'_' + str(decay) + '_' + str(fixed) + end_time)
plt.xlabel('batch')
plt.ylabel('loss')
plt.plot(self.loss_dict)
plt.savefig(save_name)
plt.show()
def train():
param = _param()
dataset = LoadDataset(opt)
param.X_dim = dataset.feature_dim
data_layer = FeatDataLayer(dataset.labels_train,
dataset.pfc_feat_data_train, opt)
result = Result()
result_gzsl = Result()
netG = _netG(dataset.text_dim, dataset.feature_dim).cuda()
netG.apply(weights_init)
print(netG)
netD = _netD(dataset.train_cls_num, dataset.feature_dim).cuda()
netD.apply(weights_init)
print(netD)
exp_info = 'CUB_EASY' if opt.splitmode == 'easy' else 'CUB_HARD'
exp_params = 'Eu{}_Rls{}_RWz{}'.format(opt.CENT_LAMBDA, opt.REG_W_LAMBDA,
opt.REG_Wz_LAMBDA)
train_dic = {}
for i in range(len(dataset.labels_train)):
try:
train_dic[dataset.labels_train[i]].append(
dataset.pfc_feat_data_train[i])
except:
train_dic[dataset.labels_train[i]] = [
dataset.pfc_feat_data_train[i]
]
out_dir = 'out/{:s}'.format(exp_info)
out_subdir = 'out/{:s}/{:s}'.format(exp_info, exp_params)
if not os.path.exists('out'):
os.mkdir('out')
if not os.path.exists(out_dir):
os.mkdir(out_dir)
if not os.path.exists(out_subdir):
os.mkdir(out_subdir)
cprint(" The output dictionary is {}".format(out_subdir), 'red')
log_dir = out_subdir + '/log_{:s}.txt'.format(exp_info)
with open(log_dir, 'a') as f:
f.write('Training Start:')
f.write(strftime("%a, %d %b %Y %H:%M:%S +0000", gmtime()) + '\n')
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
netG.load_state_dict(checkpoint['state_dict_G'])
netD.load_state_dict(checkpoint['state_dict_D'])
start_step = checkpoint['it']
print(checkpoint['log'])
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
nets = [netG, netD]
tr_cls_centroid = Variable(
torch.from_numpy(dataset.tr_cls_centroid.astype('float32'))).cuda()
optimizerD = optim.Adam(netD.parameters(), lr=opt.lr, betas=(0.5, 0.9))
optimizerG = optim.Adam(netG.parameters(), lr=opt.lr, betas=(0.5, 0.9))
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
unsupervisedData = UnsupervisedData(dataset.test_text_feature,
dataset.labels_test,
dataset.pfc_feat_data_test,
dataset.train_cls_num)
first = True if opt.resume != None else False
class_increment = False
while True:
if not first:
start_step = 0
class_increment = False
for it in range(start_step, 3000 + 1):
""" Discriminator """
for _ in range(5):
blobs = data_layer.forward()
feat_data = blobs['data'] # image data
labels = blobs['labels'].astype(int) # class labels
text_feat = np.array(
[dataset.train_text_feature[i, :] for i in labels])
text_feat = Variable(
torch.from_numpy(text_feat.astype('float32'))).cuda()
np.unique(labels)
X = Variable(torch.from_numpy(feat_data)).cuda()
y_true = Variable(torch.from_numpy(
labels.astype('int'))).cuda()
z = Variable(torch.randn(opt.batchsize,
param.z_dim)).cuda()
y_true = y_true.to(device=device, dtype=torch.long)
# GAN's D loss
D_real, C_real = netD(X)
D_loss_real = torch.mean(D_real)
# print(C_real)
# print(y_true)
C_loss_real = F.cross_entropy(C_real, y_true)
DC_loss = -D_loss_real + C_loss_real
DC_loss.backward()
# GAN's D loss
G_sample = netG(z, text_feat).detach()
D_fake, C_fake = netD(G_sample)
D_loss_fake = torch.mean(D_fake)
C_loss_fake = F.cross_entropy(C_fake, y_true)
DC_loss = D_loss_fake + C_loss_fake
DC_loss.backward()
# train with gradient penalty (WGAN_GP)
grad_penalty = calc_gradient_penalty(
netD, X.data, G_sample.data)
grad_penalty.backward()
Wasserstein_D = D_loss_real - D_loss_fake
optimizerD.step()
reset_grad(nets)
""" Generator """
for _ in range(1):
blobs = data_layer.forward()
feat_data = blobs['data'] # image data, 最小批的图片数据
labels = blobs['labels'].astype(
int) # class labels, 图片对应的标签
text_feat = np.array(
[dataset.train_text_feature[i, :] for i in labels])
text_feat = Variable(
torch.from_numpy(
text_feat.astype('float32'))).cuda() # 获取对应的文本
anchor_text_feat = Variable(
torch.from_numpy(
dataset.train_text_feature.astype(
'float32'))).cuda()
X = Variable(torch.from_numpy(feat_data)).cuda()
y_true = Variable(torch.from_numpy(
labels.astype('int'))).cuda()
y_true = y_true.to(device=device, dtype=torch.long)
z = Variable(torch.randn(opt.batchsize,
param.z_dim)).cuda()
anchor_z = Variable(
torch.randn(len(dataset.train_text_feature),
param.z_dim)).cuda()
G_sample = netG(z, text_feat)
D_fake, C_fake = netD(G_sample)
_, C_real = netD(X)
# GAN's G loss
G_loss = torch.mean(D_fake)
# Auxiliary classification loss
C_loss = (F.cross_entropy(C_real, y_true) +
F.cross_entropy(C_fake, y_true)) / 2
GC_loss = -G_loss + C_loss
# Centroid loss
Euclidean_loss_1 = Variable(torch.Tensor([0.0])).cuda()
Euclidean_loss_2 = Variable(torch.Tensor([0.0])).cuda()
if opt.CENT_LAMBDA != 0:
for i in range(dataset.train_cls_num):
sample_idx = (y_true == i).data.nonzero().squeeze()
try:
eq_idx_len = sample_idx.shape[0]
except:
eq_idx_len = 0
if sample_idx.numel() == 0:
Euclidean_loss_1 += 0.0
else:
G_sample_cls = G_sample[sample_idx, :]
Euclidean_loss_1 += (
G_sample_cls.mean(dim=0) -
tr_cls_centroid[i]).pow(2).sum().sqrt()
sample_idx = (y_true != i).data.nonzero().squeeze()
try:
sample_idx = random.sample(
sample_idx, eq_idx_len)
except:
pass
if eq_idx_len == 0:
Euclidean_loss_2 += 0.0
else:
G_sample_cls = G_sample[sample_idx, :]
Euclidean_loss_2 += (
G_sample_cls.mean(dim=0) -
tr_cls_centroid[i]).pow(2).sum().sqrt()
Euclidean_loss_1 *= 1.0 / dataset.train_cls_num * opt.CENT_LAMBDA
Euclidean_loss_2 *= 1.0 / dataset.train_cls_num * opt.CENT_LAMBDA
Euclidean_loss = Euclidean_loss_1 - Euclidean_loss_2
# ||W||_2 regularization
reg_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.REG_W_LAMBDA != 0:
for name, p in netG.named_parameters():
if 'weight' in name:
reg_loss += p.pow(2).sum()
reg_loss.mul_(opt.REG_W_LAMBDA)
# ||W_z||21 regularization, make W_z sparse
reg_Wz_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.REG_Wz_LAMBDA != 0:
Wz = netG.rdc_text.weight
reg_Wz_loss = Wz.pow(2).sum(dim=0).sqrt().sum().mul(
opt.REG_Wz_LAMBDA)
anchor = netG(anchor_z, anchor_text_feat)
triplet_loss = cal_triplets_loss(anchor, train_dic,
opt.margin)
all_loss = GC_loss + Euclidean_loss + reg_loss + reg_Wz_loss + triplet_loss
all_loss.backward()
optimizerG.step()
reset_grad(nets)
if it % opt.disp_interval == 0 and it:
acc_real = (np.argmax(C_real.data.cpu().numpy(), axis=1)
== y_true.data.cpu().numpy()).sum() / float(
y_true.data.size()[0])
acc_fake = (np.argmax(C_fake.data.cpu().numpy(), axis=1)
== y_true.data.cpu().numpy()).sum() / float(
y_true.data.size()[0])
log_text = 'Iter-{}; Was_D: {:.4}; Euc_triplet_ls: {:.4}; reg_ls: {:.4}; Wz_ls: {:.4}; G_loss: {:.4}; D_loss_real: {:.4};' \
' D_loss_fake: {:.4}; rl: {:.4}%; fk: {:.4}%'\
.format(it,
Wasserstein_D.item(),
Euclidean_loss.item()+triplet_loss.item(),
reg_loss.item(),
reg_Wz_loss.item(),
G_loss.item(),
D_loss_real.item(),
D_loss_fake.item(),
acc_real * 100, acc_fake * 100)
print(log_text)
with open(log_dir, 'a') as f:
f.write(log_text + '\n')
if it % opt.evl_interval == 0 and it >= 100:
netG.eval()
eval_fakefeat_test(it, netG, dataset, param, result)
eval_fakefeat_GZSL(it, netG, dataset, param, result_gzsl)
if result.save_model:
files2remove = glob.glob(out_subdir + '/Best_model*')
for _i in files2remove:
os.remove(_i)
torch.save(
{
'it': it + 1,
'state_dict_G': netG.state_dict(),
'state_dict_D': netD.state_dict(),
'random_seed': opt.manualSeed,
'log': log_text,
},
out_subdir + '/Best_model_Acc_{:.2f}.tar'.format(
result.acc_list[-1]))
netG.train()
if it % opt.save_interval == 0 and it:
torch.save(
{
'it': it + 1,
'state_dict_G': netG.state_dict(),
'state_dict_D': netD.state_dict(),
'random_seed': opt.manualSeed,
'log': log_text,
}, out_subdir + '/Iter_{:d}.tar'.format(it))
cprint(
'Save model to ' + out_subdir +
'/Iter_{:d}.tar'.format(it), 'red')
first = False
# semi-supervised
text_feat = Variable(
torch.from_numpy(
unsupervisedData.text_feature.astype('float32'))).cuda()
z = Variable(torch.randn(text_feat.shape[0], param.z_dim)).cuda()
text_feat = netG(z, text_feat).data.cpu().numpy()
model = KNeighborsClassifier(50)
model.fit(text_feat, unsupervisedData.labels)
y_pro = model.predict_proba(unsupervisedData.image_feature)
y = model.predict(unsupervisedData.image_feature)
probabilities = y_pro[:, np.argsort(y_pro)[::, -1][0]]
selectedHighConvinceIndex = list(
np.where(probabilities >= opt.confidence)[0])
selectedHighConvinceIndex_y = y[selectedHighConvinceIndex]
print("select high confidence number : " +
str(len(selectedHighConvinceIndex)))
for i, label in enumerate(selectedHighConvinceIndex_y):
if label in unsupervisedData.unsupervised_label_mapping:
label = unsupervisedData.unsupervised_label_mapping[label]
insertIndex = np.where(dataset.labels_train == label)[0][0]
dataset.labels_train = np.insert(dataset.labels_train,
insertIndex,
values=label,
axis=0)
dataset.pfc_feat_data_train = np.insert(
dataset.pfc_feat_data_train,
insertIndex,
values=unsupervisedData.image_feature[
selectedHighConvinceIndex[i]],
axis=0)
train_dic[label].append(unsupervisedData.image_feature[
selectedHighConvinceIndex[i]])
else:
unsupervisedData.unsupervised_label_mapping[
label] = unsupervisedData.label_index
unsupervisedData.label_index += 1
label = unsupervisedData.unsupervised_label_mapping[label]
dataset.labels_train = np.hstack(
[dataset.labels_train, [label]])
dataset.pfc_feat_data_train = np.vstack([
dataset.pfc_feat_data_train,
[
unsupervisedData.image_feature[
selectedHighConvinceIndex[i]]
]
])
dataset.train_text_feature = np.vstack([
dataset.train_text_feature,
[
unsupervisedData.text_feature[
selectedHighConvinceIndex[i]]
]
])
train_dic[label] = [
unsupervisedData.image_feature[
selectedHighConvinceIndex[i]]
]
dataset.train_cls_num += 1
class_increment = True
unsupervisedData.text_feature = np.delete(
unsupervisedData.text_feature, selectedHighConvinceIndex, axis=0)
unsupervisedData.image_feature = np.delete(
unsupervisedData.image_feature, selectedHighConvinceIndex, axis=0)
unsupervisedData.labels = np.delete(unsupervisedData.labels,
selectedHighConvinceIndex,
axis=0)
dataset.tr_cls_centroid = np.zeros(
[dataset.train_cls_num,
dataset.pfc_feat_data_train.shape[1]]).astype(np.float32)
for i in range(dataset.train_cls_num):
dataset.tr_cls_centroid[i] = np.mean(
dataset.pfc_feat_data_train[dataset.labels_train == i], axis=0)
tr_cls_centroid = Variable(
torch.from_numpy(
dataset.tr_cls_centroid.astype('float32'))).cuda()
if class_increment:
del netD
netD = _netD(dataset.train_cls_num, dataset.feature_dim).cuda()
netD.apply(weights_init)
netG = _netG(dataset.text_dim, dataset.feature_dim).cuda()
netG.apply(weights_init)
optimizerD = optim.Adam(netD.parameters(),
lr=opt.lr,
betas=(0.5, 0.9))
optimizerG = optim.Adam(netG.parameters(),
lr=opt.lr,
betas=(0.5, 0.9))
nets = [netG, netD]
print(netG)
print(netD)
data_layer = FeatDataLayer(dataset.labels_train,
dataset.pfc_feat_data_train, opt)
# model = UNet()
# serializers.load_npz(C.PATH_MODEL / "unet.model", model)
run_num = args.runnum
data_len = args.datalength
epoch = args.epoch
lr = args.learningrate
shuffle = True if args.shuffle == 1 else False
gpu_id = args.gpuid
for i in range(1, 1 + run_num):
print("{} / {} loop".format(i, run_num))
# データロード
X, Y = LoadDataset(length=data_len,
offset=(i - 1) * data_len,
shuffle=shuffle)
print("Dataset loaded.")
# 学習
model, _ = \
TrainUNet(
X=X,
Y=Y,
model_=model,
optimizer_=optimizer,
epoch=epoch,
earlystop=False,
alpha=lr,
loop=i,
gpu_id=gpu_id)
parser.add_argument('--pretrained', action='store_false', default=True)
parser.add_argument('--load', default='', type=str)
parser.add_argument('--name', default='', type=str)
parser.add_argument('--start', type=int, default=0)
params = parser.parse_args()
data_dir = params.data_path
if params.name == '':
params.name = data_dir.split('/')[-1]
if params.name == '':
params.name = data_dir.split('/')[-2]
dataloaders = {
x: LoadDataset(x, data_dir, batch_size=10, n_jobs=6)
for x in ['train', 'val']
}
os.makedirs(os.path.join('log', params.name), exist_ok=True)
logger = SummaryWriter(os.path.join('log', params.name))
device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
os.makedirs(os.path.join('ckpt', params.name), exist_ok=True)
ckpt_dir = os.path.join('ckpt', params.name)
def train_model(model,
criterion,
optimizer,
def train():
start_time = time.time()
param = _param()
dataset = LoadDataset(opt)
param.X_dim = dataset.feature_dim
data_layer = FeatDataLayer(dataset.labels_train, dataset.pfc_feat_data_train, dataset.seen_label_mapping, opt)
result = Result()
result_gzsl = Result()
netG = _netG(dataset.text_dim, dataset.feature_dim).cuda()
netG.apply(weights_init)
print(netG)
netD = _netD(dataset.train_cls_num + dataset.test_cls_num, dataset.feature_dim).cuda()
netD.apply(weights_init)
print(netD)
exp_info = 'CUB_EASY' if opt.splitmode == 'easy' else 'CUB_HARD'
exp_params = 'Eu{}_Rls{}_RWz{}'.format(opt.CENT_LAMBDA , opt.REG_W_LAMBDA, opt.REG_Wz_LAMBDA)
out_dir = 'out_' + str(opt.epsilon) + '/{:s}'.format(exp_info)
out_subdir = 'out_' + str(opt.epsilon) + '/{:s}/{:s}'.format(exp_info, exp_params)
if not os.path.exists('out_' + str(opt.epsilon) ):
os.mkdir('out_' + str(opt.epsilon))
if not os.path.exists(out_dir):
os.mkdir(out_dir)
if not os.path.exists(out_subdir):
os.mkdir(out_subdir)
cprint(" The output dictionary is {}".format(out_subdir), 'red')
log_dir = out_subdir + '/log_{:s}.txt'.format(exp_info)
with open(log_dir, 'a') as f:
f.write('Training Start:')
f.write(strftime("%a, %d %b %Y %H:%M:%S +0000", gmtime()) + '\n')
f.write("Running Parameter Logs")
f.write(runing_parameters_logs)
start_step = 0
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
netG.load_state_dict(checkpoint['state_dict_G'])
netD.load_state_dict(checkpoint['state_dict_D'])
start_step = checkpoint['it']
print(checkpoint['log'])
log_text = checkpoint['log']
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
nets = [netG, netD]
tr_cls_centroid = Variable(torch.from_numpy(dataset.tr_cls_centroid.astype('float32'))).cuda()
optimizerD = optim.Adam(netD.parameters(), lr=opt.lr, betas=(0.5, 0.9))
optimizerG = optim.Adam(netG.parameters(), lr=opt.lr, betas=(0.5, 0.9))
for it in range(start_step, 5000+1):
if it > opt.mode_change:
train_text = Variable(torch.from_numpy(dataset.train_text_feature.astype('float32'))).cuda()
test_text = Variable(torch.from_numpy(dataset.test_text_feature.astype('float32'))).cuda()
z_train = Variable(torch.randn(dataset.train_cls_num, param.z_dim)).cuda()
z_test = Variable(torch.randn(dataset.test_cls_num, param.z_dim)).cuda()
_, train_text_feature = netG(z_train, train_text)
_, test_text_feature = netG(z_test, test_text)
dataset.semantic_similarity_check(opt.Knn, train_text_feature.data.cpu().numpy(), test_text_feature.data.cpu().numpy())
""" Discriminator """
for _ in range(5):
blobs = data_layer.forward()
feat_data = blobs['data'] # image data
labels = blobs['labels'].astype(int) # class labels
true_labels = blobs['true_labels'].astype(int)
text_feat = np.array([dataset.train_text_feature[i,:] for i in labels])
text_feat = Variable(torch.from_numpy(text_feat.astype('float32'))).cuda()
X = Variable(torch.from_numpy(feat_data)).cuda()
y_true = Variable(torch.from_numpy(true_labels.astype('int'))).cuda()
z = Variable(torch.randn(opt.batchsize, param.z_dim)).cuda()
# GAN's D loss
D_real, C_real = netD(X)
D_loss_real = torch.mean(D_real)
C_loss_real = F.cross_entropy(C_real, y_true)
DC_loss = -D_loss_real + C_loss_real
DC_loss.backward()
# GAN's D loss
G_sample, _ = netG(z, text_feat)
D_fake, C_fake = netD(G_sample)
D_loss_fake = torch.mean(D_fake)
C_loss_fake = F.cross_entropy(C_fake, y_true)
DC_loss = D_loss_fake + C_loss_fake
DC_loss.backward()
# train with gradient penalty (WGAN_GP)
grad_penalty = calc_gradient_penalty(netD, X.data, G_sample.data)
grad_penalty.backward()
Wasserstein_D = D_loss_real - D_loss_fake
optimizerD.step()
reset_grad(nets)
""" Generator """
for _ in range(1):
blobs = data_layer.forward()
feat_data = blobs['data'] # image data
labels = blobs['labels'].astype(int) # class labels
true_labels = blobs['true_labels'].astype(int) #True seen label class
text_feat = np.array([dataset.train_text_feature[i, :] for i in labels])
text_feat = Variable(torch.from_numpy(text_feat.astype('float32'))).cuda()
X = Variable(torch.from_numpy(feat_data)).cuda()
y_true = Variable(torch.from_numpy(true_labels.astype('int'))).cuda()
y_dummy = Variable(torch.from_numpy(labels.astype('int'))).cuda()
z = Variable(torch.randn(opt.batchsize, param.z_dim)).cuda()
G_sample, _ = netG(z, text_feat)
D_fake, C_fake = netD(G_sample)
_, C_real = netD(X)
# GAN's G loss
G_loss = torch.mean(D_fake)
# Auxiliary classification loss
C_loss = (F.cross_entropy(C_real, y_true) + F.cross_entropy(C_fake, y_true))/2
GC_loss = -G_loss + C_loss
# Centroid loss
Euclidean_loss = Variable(torch.Tensor([0.0])).cuda()
Correlation_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.CENT_LAMBDA != 0:
for i in range(dataset.train_cls_num):
sample_idx = (y_dummy == i).data.nonzero().squeeze()
if sample_idx.numel() == 0:
Euclidean_loss += 0.0
else:
G_sample_cls = G_sample[sample_idx, :]
if sample_idx.numel() != 1:
generated_mean = G_sample_cls.mean(dim=0)
else:
generated_mean = G_sample_cls
Euclidean_loss += (generated_mean - tr_cls_centroid[i]).pow(2).sum().sqrt()
for n in range(dataset.Neighbours):
Neighbor_correlation = cosine_similarity(generated_mean.data.cpu().numpy().reshape((1, dataset.feature_dim)),
tr_cls_centroid[dataset.idx_mat[i,n]].data.cpu().numpy().reshape((1, dataset.feature_dim)))
lower_limit = dataset.semantic_similarity_seen [i,n] - opt.epsilon
upper_limit = dataset.semantic_similarity_seen [i,n] + opt.epsilon
lower_limit = torch.as_tensor(lower_limit.astype('float'))
upper_limit = torch.as_tensor(upper_limit.astype('float'))
corr = torch.as_tensor(Neighbor_correlation[0][0].astype('float'))
margin = (torch.max(corr- corr, corr - upper_limit))**2 + (torch.max(corr- corr, lower_limit - corr ))**2
Correlation_loss += margin
Euclidean_loss *= 1.0/dataset.train_cls_num * opt.CENT_LAMBDA
Correlation_loss = Correlation_loss * opt.correlation_penalty
# ||W||_2 regularization
reg_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.REG_W_LAMBDA != 0:
for name, p in netG.named_parameters():
if 'weight' in name:
reg_loss += p.pow(2).sum()
reg_loss.mul_(opt.REG_W_LAMBDA)
# ||W_z||21 regularization, make W_z sparse
reg_Wz_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.REG_Wz_LAMBDA != 0:
Wz = netG.rdc_text.weight
reg_Wz_loss = Wz.pow(2).sum(dim=0).sqrt().sum().mul(opt.REG_Wz_LAMBDA)
all_loss = GC_loss + Euclidean_loss + reg_loss + reg_Wz_loss + Correlation_loss
all_loss.backward()
optimizerG.step()
reset_grad(nets)
if (it > opt.unseen_start):
for _ in range(1):
# Zero shot Discriminator is training
zero_shot_labels = np.random.randint(dataset.test_cls_num, size = opt.zeroshotbatchsize).astype(int)
zero_shot_true_labels = np.array([dataset.unseen_label_mapping[i] for i in zero_shot_labels])
zero_text_feat = np.array([dataset.test_text_feature[i,:] for i in zero_shot_labels])
zero_text_feat = Variable(torch.from_numpy(zero_text_feat.astype('float32'))).cuda()
zero_y_true = Variable(torch.from_numpy(zero_shot_true_labels.astype('int'))).cuda()
z = Variable(torch.randn(opt.zeroshotbatchsize, param.z_dim)).cuda()
# GAN's D loss
G_sample_zero, _ = netG(z, zero_text_feat)
_, C_fake_zero = netD(G_sample_zero)
C_loss_fake_zero = F.cross_entropy(C_fake_zero, zero_y_true)
C_loss_fake_zero.backward()
optimizerD.step()
reset_grad(nets)
for _ in range(1):
# Zero shot Generator is training
zero_shot_labels = np.random.randint(dataset.test_cls_num, size = opt.zeroshotbatchsize).astype(int)
zero_shot_true_labels = np.array([dataset.unseen_label_mapping[i] for i in zero_shot_labels])
zero_text_feat = np.array([dataset.test_text_feature[i,:] for i in zero_shot_labels])
zero_text_feat = Variable(torch.from_numpy(zero_text_feat.astype('float32'))).cuda()
zero_y_true = Variable(torch.from_numpy(zero_shot_true_labels.astype('int'))).cuda()
y_dummy_zero = Variable(torch.from_numpy(zero_shot_labels.astype('int'))).cuda()
z = Variable(torch.randn(opt.zeroshotbatchsize, param.z_dim)).cuda()
# GAN's D loss
G_sample_zero, _ = netG(z, zero_text_feat)
_, C_fake_zero = netD(G_sample_zero)
C_loss_fake_zero = F.cross_entropy(C_fake_zero, zero_y_true)
Correlation_loss_zero = Variable(torch.Tensor([0.0])).cuda()
if opt.CENT_LAMBDA != 0:
for i in range(dataset.test_cls_num):
sample_idx = (y_dummy_zero == i).data.nonzero().squeeze()
if sample_idx.numel() != 0:
G_sample_cls = G_sample_zero[sample_idx, :]
if sample_idx.numel() != 1:
generated_mean = G_sample_cls.mean(dim=0)
else:
generated_mean = G_sample_cls
for n in range(dataset.Neighbours):
Neighbor_correlation = cosine_similarity(generated_mean.data.cpu().numpy().reshape((1, dataset.feature_dim)),
tr_cls_centroid[dataset.unseen_idx_mat[i,n]].data.cpu().numpy().reshape((1, dataset.feature_dim)))
lower_limit = dataset.semantic_similarity_unseen [i,n] - opt.epsilon
upper_limit = dataset.semantic_similarity_unseen [i,n] + opt.epsilon
lower_limit = torch.as_tensor(lower_limit.astype('float'))
upper_limit = torch.as_tensor(upper_limit.astype('float'))
corr = torch.as_tensor(Neighbor_correlation[0][0].astype('float'))
margin = (torch.max(corr- corr, corr - upper_limit))**2 + (torch.max(corr- corr, lower_limit - corr ))**2
Correlation_loss_zero += margin
Correlation_loss_zero = Correlation_loss_zero *opt.correlation_penalty
# ||W||_2 regularization
reg_loss_zero = Variable(torch.Tensor([0.0])).cuda()
if opt.REG_W_LAMBDA != 0:
for name, p in netG.named_parameters():
if 'weight' in name:
reg_loss_zero += p.pow(2).sum()
reg_loss_zero.mul_(opt.REG_W_LAMBDA)
# ||W_z||21 regularization, make W_z sparse
reg_Wz_loss_zero = Variable(torch.Tensor([0.0])).cuda()
if opt.REG_Wz_LAMBDA != 0:
Wz = netG.rdc_text.weight
reg_Wz_loss_zero = Wz.pow(2).sum(dim=0).sqrt().sum().mul(opt.REG_Wz_LAMBDA)
all_loss = C_loss_fake_zero + reg_loss_zero + reg_Wz_loss_zero + Correlation_loss_zero
all_loss.backward()
optimizerG.step()
reset_grad(nets)
if it % opt.disp_interval == 0 and it:
acc_real = (np.argmax(C_real.data.cpu().numpy(), axis=1) == y_true.data.cpu().numpy()).sum() / float(y_true.data.size()[0])
acc_fake = (np.argmax(C_fake.data.cpu().numpy(), axis=1) == y_true.data.cpu().numpy()).sum() / float(y_true.data.size()[0])
log_text = 'Iter-{}; Was_D: {:.4}; Euc_ls: {:.4};reg_ls: {:.4}; Wz_ls: {:.4}; G_loss: {:.4}; Correlation_loss : {:.4} ; D_loss_real: {:.4};' \
' D_loss_fake: {:.4}; rl: {:.4}%; fk: {:.4}%'.format(it, Wasserstein_D.item(), Euclidean_loss.item(), reg_loss.item(), reg_Wz_loss.item(),
G_loss.item(), Correlation_loss.item() , D_loss_real.item(), D_loss_fake.item(), acc_real * 100, acc_fake * 100)
log_text1 = ""
if it > opt.unseen_start :
acc_fake_zero = (np.argmax(C_fake_zero.data.cpu().numpy(), axis=1) == zero_y_true.data.cpu().numpy()).sum() / float(zero_y_true.data.size()[0])
log_text1 = 'Zero_Shot_Iter-{}; Correlation_loss : {:.4}; fk: {:.4}%'.format(it,
Correlation_loss_zero.item(), acc_fake_zero * 100)
'''
Here I have added .item instead of the .data[0] - Maunil
'''
print(log_text)
print (log_text1)
with open(log_dir, 'a') as f:
f.write(log_text+'\n')
f.write(log_text1+'\n')
if it % opt.evl_interval == 0 and it >=80 and log_text != None:
netG.eval() # This will start the testing process, no batch norm and drop out - It will disable them
eval_fakefeat_test(it, netG, netD, dataset, param, result)
eval_fakefeat_GZSL(it, netG, dataset, param, result_gzsl)
if result.save_model:
files2remove = glob.glob(out_subdir + '/Best_model*')
for _i in files2remove:
os.remove(_i)
torch.save({
'it': it + 1,
'state_dict_G': netG.state_dict(),
'state_dict_D': netD.state_dict(),
'random_seed': opt.manualSeed,
'log': log_text,
'Zero Shot Acc' : result.acc_list[-1],
'Generalized Zero Shot Acc' : result_gzsl.acc_list[-1]
}, out_subdir + '/Best_model_Acc_' + str(result.acc_list[-1]) + '_AUC_' + str(result_gzsl.acc_list[-1]) + '_' +'.tar')
netG.train()
if it % opt.save_interval == 0 and it:
torch.save({
'it': it + 1,
'state_dict_G': netG.state_dict(),
'state_dict_D': netD.state_dict(),
'random_seed': opt.manualSeed,
'log': log_text,
'Zero Shot Acc' : result.acc_list[-1],
'Generalized Zero Shot Acc' : result_gzsl.acc_list[-1]
}, out_subdir + '/Iter_{:d}.tar'.format(it))
cprint('Save model to ' + out_subdir + '/Iter_{:d}.tar'.format(it), 'red')
print ("########################################################")
print("--- %s Time took seconds ---" % (time.time() - start_time))
print ("########################################################")
import torch as t
import torch.nn.functional as F
from torch.autograd import Variable
from torch.utils.data import DataLoader
from eval_semantic_segmentation import eval_semantic_segmentation
from dataset import LoadDataset
from Models import FCN
import cfg
device = t.device('cuda') if t.cuda.is_available() else t.device('cpu')
num_class = cfg.DATASET[1]
BATCH_SIZE = 4
miou_list = [0]
Load_test = LoadDataset([cfg.TEST_ROOT, cfg.TEST_LABEL], cfg.crop_size)
test_data = DataLoader(Load_test,
batch_size=BATCH_SIZE,
shuffle=False,
num_workers=4)
net = FCN.FCN(num_class)
net.eval()
net.to(device)
net.load_state_dict(t.load("./Results/weights/xxx.pth"))
train_acc = 0
train_miou = 0
train_class_acc = 0
train_mpa = 0
error = 0
def train(model_num=3,
is_val=True,
sim_func_number=None,
creative_weight=None):
param = _param(opt.z_dim)
best_model_acc_path = best_model_auc_path = best_model_hm_path = ''
if opt.dataset == 'CUB':
dataset = LoadDataset(opt, main_dir, is_val)
exp_info = 'CUB_EASY' if opt.splitmode == 'easy' else 'CUB_HARD'
opt.is_gbu = False
elif opt.dataset == 'NAB':
dataset = LoadDataset_NAB(opt, main_dir, is_val)
exp_info = 'NAB_EASY' if opt.splitmode == 'easy' else 'NAB_HARD'
opt.is_gbu = False
elif "GBU" in opt.dataset:
opt.dataset = opt.dataset.split('_')[1]
opt.is_gbu = True
exp_info = opt.dataset
dataset = LoadDataset_GBU(opt, main_dir, is_val)
else:
print('No Dataset with that name')
sys.exit(0)
param.X_dim = dataset.feature_dim
data_layer = FeatDataLayer(np.array(dataset.train_label),
np.array(dataset.train_feature), opt)
result = Result()
ones = Variable(torch.Tensor(1, 1))
ones.data.fill_(1.0)
if opt.is_gbu:
netG = _netG_att(param, dataset.text_dim, dataset.feature_dim).cuda()
else:
netG = _netG(dataset.text_dim, dataset.feature_dim).cuda()
netG.apply(weights_init)
netD = _netD(dataset.train_cls_num, dataset.feature_dim).cuda()
netD.apply(weights_init)
if model_num == 2 or model_num == 4:
log_SM_ab = Scale(2)
log_SM_ab = nn.DataParallel(log_SM_ab).cuda()
if model_num == 3 or model_num == 4:
netT = _netT(dataset.train_cls_num, dataset.feature_dim,
dataset.text_dim).cuda()
netT.apply(weights_init)
similarity_func = None
if sim_func_number == 1:
similarity_func = F.cosine_similarity
elif sim_func_number == 2:
similarity_func = F.mse_loss
exp_params = 'Model_{}_is_val_{}_sim_func_number_{}_creative_weight_{}_lr_{}_zdim_{}_{}'.format(
model_num, is_val, sim_func_number, creative_weight, opt.lr,
param.z_dim, opt.exp_name)
out_subdir = main_dir + 'out/{:s}/{:s}'.format(exp_info, exp_params)
if not os.path.exists(out_subdir):
os.makedirs(out_subdir)
log_dir = out_subdir + '/log_{:s}.txt'.format(exp_info)
log_dir_2 = out_subdir + '/log_{:s}_iterations.txt'.format(exp_info)
with open(log_dir, 'a') as f:
f.write('Training Start:')
f.write(strftime("%a, %d %b %Y %H:%M:%S +0000", gmtime()) + '\n')
start_step = 0
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
netG.load_state_dict(checkpoint['state_dict_G'])
netD.load_state_dict(checkpoint['state_dict_D'])
if model_num == 3 or model_num == 4:
netT.load_state_dict(checkpoint['state_dict_T'])
start_step = checkpoint['it']
print(checkpoint['log'])
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
if model_num == 1:
nets = [netG, netD]
elif model_num == 2:
nets = [netG, netD, log_SM_ab]
elif model_num == 3:
nets = [netG, netD, netT]
elif model_num == 4:
nets = [netG, netD, netT, log_SM_ab]
tr_cls_centroid = Variable(
torch.from_numpy(dataset.tr_cls_centroid.astype('float32'))).cuda()
optimizerD = optim.Adam(netD.parameters(), lr=opt.lr, betas=(0.5, 0.9))
optimizerG = optim.Adam(netG.parameters(), lr=opt.lr, betas=(0.5, 0.9))
if model_num == 2 or model_num == 4:
optimizer_SM_ab = optim.Adam(log_SM_ab.parameters(),
lr=opt.lr,
betas=(0.5, 0.999))
if model_num == 3 or model_num == 4:
optimizerT = optim.Adam(netT.parameters(), lr=opt.lr, betas=(0.5, 0.9))
for it in tqdm(range(start_step, 5000 + 1)):
blobs = data_layer.forward()
labels = blobs['labels'].astype(int)
new_class_labels = Variable(
torch.from_numpy(np.ones_like(labels) *
dataset.train_cls_num)).cuda()
text_feat_1 = np.array([dataset.train_att[i, :] for i in labels])
text_feat_2 = np.array([dataset.train_att[i, :] for i in labels])
np.random.shuffle(
text_feat_1
) # Shuffle both features to guarantee different permutations
np.random.shuffle(text_feat_2)
alpha = (np.random.random(len(labels)) * (.8 - .2)) + .2
text_feat_mean = np.multiply(alpha, text_feat_1.transpose())
text_feat_mean += np.multiply(1. - alpha, text_feat_2.transpose())
text_feat_mean = text_feat_mean.transpose()
text_feat_mean = normalize(text_feat_mean, norm='l2', axis=1)
text_feat_Creative = Variable(
torch.from_numpy(text_feat_mean.astype('float32'))).cuda()
# z_creative = Variable(torch.randn(opt.batchsize, param.z_dim)).cuda()
# G_creative_sample = netG(z_creative, text_feat_Creative)
if model_num == 3 or model_num == 4:
""" Text Feat Generator """
for _ in range(5):
blobs = data_layer.forward()
feat_data = blobs['data'] # image data
labels = blobs['labels'].astype(int) # class labels
text_feat = np.array([dataset.train_att[i, :] for i in labels])
text_feat_TG = Variable(
torch.from_numpy(text_feat.astype('float32'))).cuda()
X = Variable(torch.from_numpy(feat_data)).cuda()
y_true = Variable(torch.from_numpy(
labels.astype('int'))).cuda()
z = Variable(torch.randn(opt.batchsize, param.z_dim)).cuda()
# GAN's T loss
T_real = netT(X)
T_loss_real = torch.mean(similarity_func(text_feat_TG, T_real))
# GAN's T loss
G_sample = netG(z, text_feat_TG).detach()
T_fake_TG = netT(G_sample)
T_loss_fake = torch.mean(
similarity_func(text_feat_TG, T_fake_TG))
# GAN's T loss
G_sample_creative = netG(z, text_feat_Creative).detach()
T_fake_creative_TG = netT(G_sample_creative)
T_loss_fake_creative = torch.mean(
similarity_func(text_feat_Creative, T_fake_creative_TG))
T_loss = -1 * T_loss_real - T_loss_fake - T_loss_fake_creative
T_loss.backward()
optimizerT.step()
optimizerG.step()
reset_grad(nets)
""" Discriminator """
for _ in range(5):
blobs = data_layer.forward()
feat_data = blobs['data'] # image data
labels = blobs['labels'].astype(int) # class labels
text_feat = np.array([dataset.train_att[i, :] for i in labels])
text_feat = Variable(torch.from_numpy(
text_feat.astype('float32'))).cuda()
X = Variable(torch.from_numpy(feat_data)).cuda()
y_true = Variable(torch.from_numpy(labels.astype('int'))).cuda()
z = Variable(torch.randn(opt.batchsize, param.z_dim)).cuda()
# GAN's D loss
D_real, C_real = netD(X)
D_loss_real = torch.mean(D_real)
C_loss_real = F.cross_entropy(C_real, y_true)
DC_loss = -D_loss_real + C_loss_real
DC_loss.backward()
# GAN's D loss
G_sample = netG(z, text_feat).detach()
D_fake, C_fake = netD(G_sample)
D_loss_fake = torch.mean(D_fake)
C_loss_fake = F.cross_entropy(C_fake, y_true)
DC_loss = D_loss_fake + C_loss_fake
DC_loss.backward()
# train with gradient penalty (WGAN_GP)
grad_penalty = calc_gradient_penalty(netD, X.data, G_sample.data)
grad_penalty.backward()
Wasserstein_D = D_loss_real - D_loss_fake
optimizerD.step()
reset_grad(nets)
""" Generator """
for _ in range(1):
blobs = data_layer.forward()
feat_data = blobs['data'] # image data
labels = blobs['labels'].astype(int) # class labels
text_feat = np.array([dataset.train_att[i, :] for i in labels])
text_feat = Variable(torch.from_numpy(
text_feat.astype('float32'))).cuda()
X = Variable(torch.from_numpy(feat_data)).cuda()
y_true = Variable(torch.from_numpy(labels.astype('int'))).cuda()
z = Variable(torch.randn(opt.batchsize, param.z_dim)).cuda()
G_sample = netG(z, text_feat)
D_fake, C_fake = netD(G_sample)
_, C_real = netD(X)
# GAN's G loss
G_loss = torch.mean(D_fake)
# Auxiliary classification loss
C_loss = (F.cross_entropy(C_real, y_true) +
F.cross_entropy(C_fake, y_true)) / 2
# GAN's G loss creative
G_sample_creative = netG(z, text_feat_Creative).detach()
if model_num == 3 or model_num == 4:
D_creative_fake, _ = netD(G_sample_creative)
G_loss_fake_creative = torch.mean(D_creative_fake)
T_fake = netT(G_sample)
T_loss_fake = torch.mean(similarity_func(text_feat, T_fake))
T_fake_creative = netT(G_sample_creative)
T_loss_fake_creative = torch.mean(
similarity_func(text_feat_Creative, T_fake_creative))
GC_loss = -G_loss - G_loss_fake_creative + C_loss - T_loss_fake - T_loss_fake_creative
else:
GC_loss = -G_loss + C_loss
# Centroid loss
Euclidean_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.REG_W_LAMBDA != 0:
for i in range(dataset.train_cls_num):
sample_idx = (y_true == i).data.nonzero().squeeze()
if sample_idx.numel() == 0:
Euclidean_loss += 0.0
else:
G_sample_cls = G_sample[sample_idx, :]
Euclidean_loss += (
G_sample_cls.mean(dim=0) -
tr_cls_centroid[i]).pow(2).sum().sqrt()
Euclidean_loss *= 1.0 / dataset.train_cls_num * opt.CENT_LAMBDA
# ||W||_2 regularization
reg_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.REG_W_LAMBDA != 0:
for name, p in netG.named_parameters():
if 'weight' in name:
reg_loss += p.pow(2).sum()
reg_loss.mul_(opt.REG_W_LAMBDA)
# ||W_z||21 regularization, make W_z sparse
reg_Wz_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.REG_Wz_LAMBDA != 0 and not opt.is_gbu:
Wz = netG.rdc_text.weight
reg_Wz_loss = Wz.pow(2).sum(dim=0).sqrt().sum().mul(
opt.REG_Wz_LAMBDA)
if model_num == 2 or model_num == 4:
# D(C| GX_fake)) + Classify GX_fake as real
D_creative_fake, C_creative_fake = netD(G_sample_creative)
G_fake_C = F.softmax(C_creative_fake)
# SM Divergence
q_shape = Variable(
torch.FloatTensor(G_fake_C.data.size(0),
G_fake_C.data.size(1))).cuda()
q_shape.data.fill_(1.0 / G_fake_C.data.size(1))
SM_ab = F.sigmoid(log_SM_ab(ones))
SM_a = 0.2 + torch.div(SM_ab[0][0], 1.6666666666666667).cuda()
SM_b = 0.2 + torch.div(SM_ab[0][1], 1.6666666666666667).cuda()
pow_a_b = torch.div(1 - SM_a, 1 - SM_b)
alpha_term = (torch.pow(G_fake_C + 1e-5, SM_a) *
torch.pow(q_shape, 1 - SM_a)).sum(1)
entropy_GX_fake_vec = torch.div(
torch.pow(alpha_term, pow_a_b) - 1, SM_b - 1)
min_e, max_e = torch.min(entropy_GX_fake_vec), torch.max(
entropy_GX_fake_vec)
entropy_GX_fake_vec = (entropy_GX_fake_vec - min_e) / (max_e -
min_e)
entropy_GX_fake = -entropy_GX_fake_vec.mean()
loss_creative = -creative_weight * entropy_GX_fake
disc_GX_fake_real = -torch.mean(D_creative_fake)
total_loss_creative = loss_creative + disc_GX_fake_real
all_loss = GC_loss + Euclidean_loss + reg_loss + reg_Wz_loss + total_loss_creative
else:
all_loss = GC_loss + Euclidean_loss + reg_loss + reg_Wz_loss
all_loss.backward()
if model_num == 2 or model_num == 4:
optimizer_SM_ab.step()
optimizerG.step()
reset_grad(nets)
if it % opt.disp_interval == 0 and it:
acc_real = (np.argmax(C_real.data.cpu().numpy(), axis=1)
== y_true.data.cpu().numpy()).sum() / float(
y_true.data.size()[0])
acc_fake = (np.argmax(C_fake.data.cpu().numpy(), axis=1)
== y_true.data.cpu().numpy()).sum() / float(
y_true.data.size()[0])
log_text = 'Iter-{}; rl: {:.4}%; fk: {:.4}%'.format(
it, acc_real * 100, acc_fake * 100)
with open(log_dir, 'a') as f:
f.write(log_text + '\n')
if it % opt.evl_interval == 0 and it > opt.disp_interval:
cur_acc = 0
cur_auc = 0
cur_hm = 0
netG.eval()
if is_val:
cur_acc = eval_fakefeat_test(netG, dataset.val_cls_num,
dataset.val_att,
dataset.val_unseen_feature,
dataset.val_unseen_label, param,
result)
if opt.is_gbu:
cur_hm, acc_S_T, acc_U_T = eval_fakefeat_test_gzsl(
netG, dataset, dataset.val_cls_num, dataset.val_att,
dataset.val_unseen_feature, dataset.val_unseen_label,
param, result)
else:
cur_auc = eval_fakefeat_GZSL(netG, dataset,
dataset.val_cls_num,
dataset.val_att,
dataset.val_unseen_feature,
dataset.val_unseen_label,
param, out_subdir, result)
else:
cur_acc = eval_fakefeat_test(netG, dataset.test_cls_num,
dataset.test_att,
dataset.test_unseen_feature,
dataset.test_unseen_label, param,
result)
if opt.is_gbu:
cur_hm, acc_S_T, acc_U_T = eval_fakefeat_test_gzsl(
netG, dataset, dataset.test_cls_num, dataset.test_att,
dataset.test_unseen_feature, dataset.test_unseen_label,
param, result)
else:
cur_auc = eval_fakefeat_GZSL(netG, dataset,
dataset.test_cls_num,
dataset.test_att,
dataset.test_unseen_feature,
dataset.test_unseen_label,
param, out_subdir, result)
if cur_acc > result.best_acc:
result.best_acc = cur_acc
files2remove = glob.glob(out_subdir + '/Best_model_ACC*')
for _i in files2remove:
os.remove(_i)
save_dict = {
'it': it + 1,
'state_dict_G': netG.state_dict(),
'state_dict_D': netD.state_dict(),
'random_seed': opt.manualSeed,
'log': log_text,
}
if model_num == 3 or model_num == 4:
save_dict.update({'state_dict_T': netT.state_dict()})
best_model_acc_path = '/Best_model_ACC_{:.2f}.tar'.format(
cur_acc)
torch.save(save_dict, out_subdir + best_model_acc_path)
if cur_auc > result.best_auc:
result.best_auc = cur_auc
files2remove = glob.glob(out_subdir + '/Best_model_AUC*')
for _i in files2remove:
os.remove(_i)
save_dict = {
'it': it + 1,
'state_dict_G': netG.state_dict(),
'state_dict_D': netD.state_dict(),
'random_seed': opt.manualSeed,
'log': log_text,
}
if model_num == 3 or model_num == 4:
save_dict.update({'state_dict_T': netT.state_dict()})
best_model_auc_path = '/Best_model_AUC_{:.2f}.tar'.format(
cur_auc)
torch.save(save_dict, out_subdir + best_model_auc_path)
if cur_hm > result.best_hm:
result.best_hm = cur_hm
result.best_acc_S_T = acc_S_T
result.best_acc_U_T = acc_U_T
files2remove = glob.glob(out_subdir + '/Best_model_HM*')
for _i in files2remove:
os.remove(_i)
save_dict = {
'it': it + 1,
'state_dict_G': netG.state_dict(),
'state_dict_D': netD.state_dict(),
'random_seed': opt.manualSeed,
'log': log_text,
}
if model_num == 3 or model_num == 4:
save_dict.update({'state_dict_T': netT.state_dict()})
best_model_hm_path = '/Best_model_HM_{:.2f}.tar'.format(cur_hm)
torch.save(save_dict, out_subdir + best_model_hm_path)
log_text_2 = 'iteration: %f, best_acc: %f, best_auc: %f, best_hm: %f' % (
it, result.best_acc, result.best_auc, result.best_hm)
with open(log_dir_2, 'a') as f:
f.write(log_text_2 + '\n')
netG.train()
if is_val:
if os.path.isfile(out_subdir + best_model_acc_path):
print("=> loading checkpoint '{}'".format(best_model_acc_path))
checkpoint = torch.load(out_subdir + best_model_acc_path)
netG.load_state_dict(checkpoint['state_dict_G'])
netD.load_state_dict(checkpoint['state_dict_D'])
if model_num == 3 or model_num == 4:
netT.load_state_dict(checkpoint['state_dict_T'])
it = checkpoint['it']
print("iteration: {}".format(it))
netG.eval()
test_acc = eval_fakefeat_test(netG, dataset.test_cls_num,
dataset.test_att,
dataset.test_unseen_feature,
dataset.test_unseen_label, param,
result)
result.test_acc = test_acc
else:
print("=> no checkpoint found at '{}'".format(out_subdir +
best_model_acc_path))
if os.path.isfile(out_subdir + best_model_auc_path):
print("=> loading checkpoint '{}'".format(best_model_auc_path))
checkpoint = torch.load(out_subdir + best_model_auc_path)
netG.load_state_dict(checkpoint['state_dict_G'])
netD.load_state_dict(checkpoint['state_dict_D'])
if model_num == 3 or model_num == 4:
netT.load_state_dict(checkpoint['state_dict_T'])
it = checkpoint['it']
print("iteration: {}".format(it))
netG.eval()
test_auc = eval_fakefeat_GZSL(netG, dataset, dataset.test_cls_num,
dataset.test_att,
dataset.test_unseen_feature,
dataset.test_unseen_label, param,
out_subdir, result)
result.test_auc = test_auc
else:
print("=> no checkpoint found at '{}'".format(out_subdir +
best_model_auc_path))
if os.path.isfile(out_subdir + best_model_hm_path):
print("=> loading checkpoint '{}'".format(best_model_hm_path))
checkpoint = torch.load(out_subdir + best_model_hm_path)
netG.load_state_dict(checkpoint['state_dict_G'])
netD.load_state_dict(checkpoint['state_dict_D'])
if model_num == 3 or model_num == 4:
netT.load_state_dict(checkpoint['state_dict_T'])
it = checkpoint['it']
print("iteration: {}".format(it))
netG.eval()
test_hm, test_acc_S_T, test_acc_U_T = eval_fakefeat_test_gzsl(
netG, dataset, dataset.test_cls_num, dataset.test_att,
dataset.test_unseen_feature, dataset.test_unseen_label, param,
result)
result.test_hm = test_hm
result.test_acc_S_T = test_acc_S_T
result.test_acc_U_T = test_acc_U_T
else:
print("=> no checkpoint found at '{}'".format(out_subdir +
best_model_hm_path))
log_text_2 = 'test_acc: %f, test_auc: %f, test_hm: %f, test_acc_S_T: %f, test_acc_U_T: %f' % (
result.test_acc, result.test_auc, result.test_hm,
result.test_acc_S_T, result.test_acc_U_T)
with open(log_dir_2, 'a') as f:
f.write(log_text_2 + '\n')
return result
def train():
param = _param()
print("load dataset origin")
dataset_origin = LoadDataset_origin(opt)
print("load dataset")
dataset = LoadDataset(opt)
param.X_dim = dataset.feature_dim
data_layer_origin = FeatDataLayer_origin(
dataset_origin.labels_train, dataset_origin.pfc_feat_data_train, opt)
data_layer = FeatDataLayer_add_FG(
dataset.labels_train, dataset.pfc_feat_data_train, opt,
dataset.train_text_feature, dataset.familyToText, dataset.genusToText,
dataset.familyLabelToBirdLabel, dataset.genusLabelToBirdLabel,
dataset.labels_origin_train)
result = Result()
result_gzsl = Result()
netG = _netG(dataset.text_dim, dataset.feature_dim).cuda()
netG.apply(weights_init)
print(netG)
netD = _netD(dataset.train_cls_num, dataset.feature_dim).cuda()
netD.apply(weights_init)
print(netD)
exp_info = 'CUB_EASY' if opt.splitmode == 'easy' else 'CUB_HARD'
exp_params = 'Eu{}_Rls{}_RWz{}'.format(opt.CENT_LAMBDA, opt.REG_W_LAMBDA,
opt.REG_Wz_LAMBDA)
out_dir = 'out/{:s}'.format(exp_info)
out_subdir = 'out/{:s}/{:s}'.format(exp_info, exp_params)
opt.out_subdir = out_subdir
if not os.path.exists('out'):
os.mkdir('out')
if not os.path.exists(out_dir):
os.mkdir(out_dir)
if not os.path.exists(out_subdir):
os.mkdir(out_subdir)
cprint(" The output dictionary is {}".format(out_subdir), 'red')
log_dir = out_subdir + '/log_{:s}'.format(exp_info)
if opt.exp_no != "":
log_dir += "_" + opt.exp_no
log_dir += ".txt"
opt.log_dir = log_dir
opt.auc_plot_dir = out_subdir + '/best_auc_plot{:s}_{:s}'.format(
opt.exp_no, exp_info)
opt.auc_solid_plot_dir = out_subdir + '/solid_auc_plot{:s}_{:s}'.format(
opt.exp_no, exp_info)
opt.history_D_loss_dir = out_subdir + '/D_loss_plot{:s}_{:s}'.format(
opt.exp_no, exp_info)
opt.history_G_loss_dir = out_subdir + '/G_loss_plot{:s}_{:s}'.format(
opt.exp_no, exp_info)
start_step = 0
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
netG.load_state_dict(checkpoint['state_dict_G'])
netD.load_state_dict(checkpoint['state_dict_D'])
start_step = checkpoint['it']
print(checkpoint['log'])
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
nets = [netG, netD]
# tr_cls_centroid = Variable(torch.from_numpy(dataset.tr_cls_centroid.astype('float32'))).cuda()
optimizerD = optim.Adam(netD.parameters(), lr=opt.lr, betas=(0.5, 0.9))
optimizerG = optim.Adam(netG.parameters(), lr=opt.lr, betas=(0.5, 0.9))
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
history_D_loss = []
history_G_loss = []
for it in range(start_step, 10000 + 1):
cur_D_loss = 0
cur_G_loss = 0
""" Discriminator """
for _ in range(5):
blobs = data_layer_origin.forward()
feat_data = blobs['data'] # image data
labels = blobs['labels'].astype(int) # class labels
text_feat = np.array(
[dataset_origin.train_text_feature[i, :] for i in labels])
text_feat = Variable(torch.from_numpy(
text_feat.astype('float32'))).cuda()
X = Variable(torch.from_numpy(feat_data)).cuda()
y_true = Variable(torch.from_numpy(labels.astype('int'))).cuda()
z = Variable(torch.randn(opt.batchsize, param.z_dim)).cuda()
y_true = y_true.to(device=device, dtype=torch.long)
# GAN's D loss
D_real, C_real = netD(X)
D_loss_real = torch.mean(D_real)
C_loss_real = F.cross_entropy(C_real, y_true)
DC_loss = -D_loss_real + C_loss_real
DC_loss.backward()
cur_D_loss += DC_loss.item()
# GAN's D loss
G_sample = netG(z, text_feat).detach()
D_fake, C_fake = netD(G_sample)
D_loss_fake = torch.mean(D_fake)
C_loss_fake = F.cross_entropy(C_fake, y_true)
DC_loss = D_loss_fake + C_loss_fake
DC_loss.backward()
cur_D_loss += DC_loss.item()
# train with gradient penalty (WGAN_GP)
grad_penalty = calc_gradient_penalty(netD, X.data, G_sample.data)
grad_penalty.backward()
Wasserstein_D = D_loss_real - D_loss_fake
optimizerD.step()
reset_grad(nets)
cur_D_loss += Wasserstein_D.item()
""" Generator """
for _ in range(1):
blobs = data_layer.forward()
feat_data = blobs['data'] # image data
labels = blobs['labels'].astype(int)
origin_labels = blobs['minibatch_origin_label'].astype(int)
text_feat = blobs['text_feat'] # text_feat
# text_feat = np.array([dataset.train_text_feature[i, :] for i in labels])
text_feat = Variable(torch.from_numpy(
text_feat.astype('float32'))).cuda()
X = Variable(torch.from_numpy(feat_data)).cuda()
y_true = Variable(torch.from_numpy(labels.astype('int'))).cuda()
y_origin_true = Variable(
torch.from_numpy(origin_labels.astype('int'))).cuda()
y_true = y_true.to(device=device, dtype=torch.long)
z = Variable(torch.randn(opt.batchsize, param.z_dim)).cuda()
G_sample = netG(z, text_feat)
D_fake, C_fake = netD(G_sample)
_, C_real = netD(X)
# GAN's G loss
G_loss = torch.mean(D_fake)
# Auxiliary classification loss
C_loss = (F.cross_entropy(C_real, y_true) +
F.cross_entropy(C_fake, y_true)) / 2
GC_loss = -G_loss + C_loss
cur_G_loss += -G_loss.item() + F.cross_entropy(C_fake,
y_true).item()
Bird_Euclidean_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.CENT_LAMBDA != 0 and opt.BIRD_CENT_LAMBDA != 0:
for i in range(dataset.train_cls_num):
sample_idx = (y_origin_true == i).data.nonzero().squeeze()
if sample_idx.numel() == 0:
Bird_Euclidean_loss += 0.0
else:
G_sample_cls = G_sample[sample_idx, :]
center = Variable(
torch.from_numpy(dataset.tr_cls_centroid[i].astype(
'float32'))).cuda()
Bird_Euclidean_loss += (G_sample_cls.mean(dim=0) -
center).pow(2).sum().sqrt()
Bird_Euclidean_loss *= 1.0 / dataset.train_cls_num
Family_Euclidean_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.CENT_LAMBDA != 0 and opt.FAMILY_CENT_LAMBDA != 0:
for i in range(dataset.familyLabelStart,
dataset.familyLabelEnd):
sample_idx = (y_origin_true == i).data.nonzero().squeeze()
if sample_idx.numel() == 0:
Family_Euclidean_loss += 0.0
else:
G_sample_cls = G_sample[sample_idx, :]
center = Variable(
torch.from_numpy(dataset.tr_cls_centroid[i].astype(
'float32'))).cuda()
Family_Euclidean_loss += (G_sample_cls.mean(dim=0) -
center).pow(2).sum().sqrt()
Family_Euclidean_loss *= 1.0 / (dataset.familyLabelEnd -
dataset.familyLabelStart)
Genus_Euclidean_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.CENT_LAMBDA != 0 and opt.GENUS_CENT_LAMBDA != 0:
for i in range(dataset.genusLabelStart, dataset.genusLabelEnd):
sample_idx = (y_origin_true == i).data.nonzero().squeeze()
if sample_idx.numel() == 0:
Genus_Euclidean_loss += 0.0
else:
G_sample_cls = G_sample[sample_idx, :]
center = Variable(
torch.from_numpy(dataset.tr_cls_centroid[i].astype(
'float32'))).cuda()
Genus_Euclidean_loss += (G_sample_cls.mean(dim=0) -
center).pow(2).sum().sqrt()
Genus_Euclidean_loss *= 1.0 / (dataset.genusLabelEnd -
dataset.genusLabelStart)
Euclidean_loss = opt.CENT_LAMBDA * (
opt.BIRD_CENT_LAMBDA * Bird_Euclidean_loss +
opt.FAMILY_CENT_LAMBDA * Family_Euclidean_loss +
opt.GENUS_CENT_LAMBDA * Genus_Euclidean_loss)
# ||W||_2 regularization
reg_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.REG_W_LAMBDA != 0:
for name, p in netG.named_parameters():
if 'weight' in name:
reg_loss += p.pow(2).sum()
reg_loss.mul_(opt.REG_W_LAMBDA)
# ||W_z||21 regularization, make W_z sparse
reg_Wz_loss = Variable(torch.Tensor([0.0])).cuda()
if opt.REG_Wz_LAMBDA != 0:
Wz = netG.rdc_text.weight
reg_Wz_loss = Wz.pow(2).sum(dim=0).sqrt().sum().mul(
opt.REG_Wz_LAMBDA)
all_loss = GC_loss + Euclidean_loss + reg_loss + reg_Wz_loss
all_loss.backward()
optimizerG.step()
reset_grad(nets)
cur_G_loss += Euclidean_loss.item()
history_D_loss.append(cur_D_loss)
history_G_loss.append(cur_G_loss)
print("Iter-" + str(it + 1) + "; G-loss: " + str(cur_G_loss) +
"; D-loss: " + str(cur_D_loss))
if it % opt.disp_interval == 0 and it:
acc_real = (np.argmax(C_real.data.cpu().numpy(), axis=1)
== y_true.data.cpu().numpy()).sum() / float(
y_true.data.size()[0])
acc_fake = (np.argmax(C_fake.data.cpu().numpy(), axis=1)
== y_true.data.cpu().numpy()).sum() / float(
y_true.data.size()[0])
log_text = 'Iter-{}; Was_D: {:.4}; Euc_ls: {:.4}; Bird_Euc_ls: {:.4}; family_Euc_ls: {:.4}; ' \
'Genus_Euc_ls: {:.4}; reg_ls: {:.4}; Wz_ls: {:.4}; G_loss: {:.4}; D_loss_real: {:.4};' \
' D_loss_fake: {:.4}; rl: {:.4}%; fk: {:.4}%' \
.format(it,
Wasserstein_D.item(),
Euclidean_loss.item(),
Bird_Euclidean_loss.item(),
Family_Euclidean_loss.item(),
Genus_Euclidean_loss.item(),
reg_loss.item(),
reg_Wz_loss.item(),
G_loss.item(),
D_loss_real.item(),
D_loss_fake.item(),
acc_real * 100, acc_fake * 100)
print(log_text)
if it % opt.evl_interval == 0 and it >= 100:
netG.eval()
eval_fakefeat_test(it, netG, dataset_origin, param, result)
eval_fakefeat_GZSL(it, netG, dataset_origin, param, result_gzsl)
if result.save_model:
files2remove = glob.glob(
out_subdir + '/Best_model{}_Acc*'.format(opt.exp_no))
for _i in files2remove:
os.remove(_i)
torch.save(
{
'it': it + 1,
'state_dict_G': netG.state_dict(),
'state_dict_D': netD.state_dict(),
'random_seed': opt.manualSeed,
'log': log_text,
}, out_subdir + '/Best_model{}_Acc_{:.2f}.tar'.format(
opt.exp_no, result.acc_list[-1]))
if result_gzsl.save_model:
files2remove = glob.glob(
out_subdir + '/Best_model{}_Auc*'.format(opt.exp_no))
for _i in files2remove:
os.remove(_i)
torch.save(
{
'it': it + 1,
'state_dict_G': netG.state_dict(),
'state_dict_D': netD.state_dict(),
'random_seed': opt.manualSeed,
'log': log_text,
}, out_subdir + '/Best_model{}_Auc_{:.2f}.tar'.format(
opt.exp_no, result_gzsl.best_auc * 100))
netG.train()
if it % opt.save_interval == 0 and it:
torch.save(
{
'it': it + 1,
'state_dict_G': netG.state_dict(),
'state_dict_D': netD.state_dict(),
'random_seed': opt.manualSeed,
'log': log_text,
}, out_subdir + '/Iter_{:d}.tar'.format(it))
cprint('Save model to ' + out_subdir + '/Iter_{:d}.tar'.format(it),
'red')
print("Reproduce CUB {}".format(opt.splitmode))
print("Accuracy is {:.4}%, and Generalized AUC is {:.4}%".format(
result.best_acc, result_gzsl.best_auc * 100))
np.savetxt(opt.history_D_loss_dir + '.txt', history_D_loss, fmt='%.015f')
np.savetxt(opt.history_G_loss_dir + '.txt', history_G_loss, fmt='%.015f')
def train(opt):
param = _param()
dataset = LoadDataset(opt)
param.X_dim = dataset.feature_dim
data_layer = FeatDataLayer(dataset.labels_train,
dataset.pfc_feat_data_train, opt)
# initialize model
netGs = []
netDs = []
parts = 7 if opt.dataset == "CUB2011" else 6
for part in range(parts):
netGs.append(_netG(dataset.text_dim, 512).cuda().apply(weights_init))
netDs.append(
_netD(dataset.train_cls_num, 512).cuda().apply(weights_init))
start_step = 0
part_cls_centrild = torch.from_numpy(
dataset.part_cls_centrild.astype('float32')).cuda()
# initialize optimizers
optimizerGs = []
optimizerDs = []
for netG in netGs:
optimizerGs.append(
optim.Adam(netG.parameters(), lr=opt.lr, betas=(0.5, 0.9)))
for netD in netDs:
optimizerDs.append(
optim.Adam(netD.parameters(), lr=opt.lr, betas=(0.5, 0.9)))
for it in range(start_step, 3000 + 1):
""" Discriminator """
for _ in range(5):
blobs = data_layer.forward()
feat_data = blobs['data'] # image data
labels = blobs['labels'].astype(int) # class labels
text_feat = np.array(
[dataset.train_text_feature[i, :] for i in labels])
text_feat = torch.from_numpy(text_feat.astype('float32')).cuda()
X = torch.from_numpy(feat_data).cuda()
y_true = torch.from_numpy(labels.astype('int')).cuda()
for part in range(parts):
z = torch.randn(opt.batchsize, param.z_dim).cuda()
D_real, C_real = netDs[part](X[:, part * 512:(part + 1) * 512])
D_loss_real = torch.mean(D_real)
C_loss_real = F.cross_entropy(C_real, y_true)
DC_loss = -D_loss_real + C_loss_real
DC_loss.backward()
G_sample = netGs[part](z, text_feat)
D_fake, C_fake = netDs[part](G_sample)
D_loss_fake = torch.mean(D_fake)
C_loss_fake = F.cross_entropy(C_fake, y_true)
DC_loss = D_loss_fake + C_loss_fake
DC_loss.backward()
grad_penalty = calc_gradient_penalty(
opt.batchsize, netDs[part],
X.data[:, part * 512:(part + 1) * 512], G_sample.data,
opt.GP_LAMBDA)
grad_penalty.backward()
Wasserstein_D = D_loss_real - D_loss_fake
# writer.add_scalar("Wasserstein_D"+str(part), Wasserstein_D.item(), it)
optimizerDs[part].step()
netGs[part].zero_grad()
netDs[part].zero_grad()
""" Generator """
for _ in range(1):
blobs = data_layer.forward()
feat_data = blobs['data'] # image data
labels = blobs['labels'].astype(int) # class labels
text_feat = np.array(
[dataset.train_text_feature[i, :] for i in labels])
text_feat = torch.from_numpy(text_feat.astype('float32')).cuda()
X = torch.from_numpy(feat_data).cuda()
y_true = torch.from_numpy(labels.astype('int')).cuda()
for part in range(parts):
z = torch.randn(opt.batchsize, param.z_dim).cuda()
G_sample = netGs[part](z, text_feat)
# G_sample_all[:, part*512:(part+1)*512] = G_sample
D_fake, C_fake = netDs[part](G_sample)
_, C_real = netDs[part](X[:, part * 512:(part + 1) * 512])
G_loss = torch.mean(D_fake)
C_loss = (F.cross_entropy(C_real, y_true) +
F.cross_entropy(C_fake, y_true)) / 2
GC_loss = -G_loss + C_loss
# writer.add_scalar("GC_loss"+str(part), GC_loss.item(), it)
Euclidean_loss = torch.tensor([0.0]).cuda()
if opt.REG_W_LAMBDA != 0:
for i in range(dataset.train_cls_num):
sample_idx = (y_true == i).data.nonzero().squeeze()
if sample_idx.numel() == 0:
Euclidean_loss += 0.0
else:
G_sample_cls = G_sample[sample_idx, :]
Euclidean_loss += (G_sample_cls.mean(dim=0) -
part_cls_centrild[i][part]
).pow(2).sum().sqrt()
Euclidean_loss *= 1.0 / dataset.train_cls_num * opt.CENT_LAMBDA
# ||W||_2 regularization
reg_loss = torch.Tensor([0.0]).cuda()
if opt.REG_W_LAMBDA != 0:
for name, p in netGs[part].named_parameters():
if 'weight' in name:
reg_loss += p.pow(2).sum()
reg_loss.mul_(opt.REG_W_LAMBDA)
# writer.add_scalar("reg_loss"+str(part), reg_loss.item(), it)
# ||W_z||21 regularization, make W_z sparse
reg_Wz_loss = torch.Tensor([0.0]).cuda()
if opt.REG_Wz_LAMBDA != 0:
Wz = netGs[part].rdc_text.weight
reg_Wz_loss = reg_Wz_loss + Wz.pow(2).sum(
dim=0).sqrt().sum().mul(opt.REG_Wz_LAMBDA)
# writer.add_scalar("reg_Wz_loss"+str(part), reg_Wz_loss.item(), it)
all_loss = GC_loss + Euclidean_loss + reg_loss + reg_Wz_loss
all_loss.backward()
optimizerGs[part].step()
if it % opt.evl_interval == 0 and it > 500:
print(it)
for part in range(parts):
netGs[part].eval()
train_classifier(opt, param, dataset, netGs)
for part in range(parts):
netGs[part].train()