def predict(dataset, network, models, dic, net_name):
#测试数据
#valdata = dataset.
val_data = dataset.data_generateor(
'pred',
rand_h=dic['rand_h'],
rand_h_num=dic['rand_h_num'],
height=dic['height'],
pooling_stride=dic['pooling_stride'],
batch=Config.pred_batch)
val_data_sum = []
pred_sum = []
for i in range(2000 // Config.pred_batch):
val_data_i = next(val_data)
#print(np.shape(val_data[0]))
pred = models.pred(network, val_data=val_data_i[0])
val_data_sum.extend(val_data_i[1])
pred_sum.extend(pred)
#保存数据
save_data(
file_dir='saved_results/' + net_name + '/',
filename='predict',
data=pred_sum)
save_data(
file_dir='saved_results/' + net_name + '/',
filename='real_data',
data=val_data_sum)
#画图
print(np.shape(pred_sum), np.shape(val_data_sum))
plot_fig(val_data_sum, pred_sum, 'real', 'pred')
def train(self, model, train_gen, val_gen, net_name, eps, is_graph=True):
'''
训练数据
'''
assert self.mode == "training", "Create model in training mode."
# Callbacks
callbacks = [
keras.callbacks.TensorBoard(
log_dir=self.log_dir,
histogram_freq=0,
write_graph=True,
write_images=False),
keras.callbacks.ModelCheckpoint(
self.checkpoint_path, verbose=0, save_weights_only=True),
]
log("\nStarting at epoch {}.\n".format(self.epoch))
log("Checkpoint Path: {}".format(self.checkpoint_path))
model.compile(loss=Config.loss, optimizer=Config.optimizer)
# if os.name is 'nt':
# workers = 0
# else:
# workers = multiprocessing.cpu_count()
history = History()
history = model.fit_generator(
train_gen,
initial_epoch=self.epoch,
epochs=eps,
steps_per_epoch=self.STEPS_PER_EPOCH,
callbacks=callbacks,
validation_data=val_gen,
validation_steps=self.VALIDATION_STEPS,
#max_queue_size=100,
#workers=0,
#use_multiprocessing=False,
)
#model.save('save/train_model.h5')
save_data(
file_dir='saved_results/' + net_name + '/',
filename='val_loss',
data=history.history["val_loss"])
save_data(
file_dir='saved_results/' + net_name + '/',
filename='train_loss',
data=history.history["loss"])
if is_graph:
plot_fig(history.history["loss"], history.history["val_loss"],
'train_loss', 'val_loss')
def test_model(model,
labels,
data_name='sk_eigenjoint_nor_528',
valid_segment_idx=650):
data = data_dir + '/' + data_name
file_paths = glob(data + "/*.npy")
losses = []
ground_ys = []
pred_ys = []
for index, file_path in enumerate(file_paths):
if index >= valid_segment_idx:
print 'Predict ' + file_path
valid_x = np.load(file_path)
valid_x = np.reshape(valid_x,
newshape=(1, valid_x.shape[0],
valid_x.shape[1]))
valid_y = labels[index]
# valid_y = valid_y[:-1]
pred_y = model.predict_on_batch(valid_x)
file_name = str(index - valid_segment_idx)
plot_fig(pred_y, valid_y, file_name, save_flag=True)
pred_y = np.argmax(pred_y, axis=2)
pred_y = pred_y.ravel()
ground_ys.append(valid_y)
pred_ys.append(pred_y)
# pred_y = clear_pred(pred_y)
loss = eval_jaccard(valid_y, pred_y)
losses.append(loss)
ground_ys = np.concatenate(ground_ys)
pred_ys = np.concatenate(pred_ys)
print(classification_report(ground_ys, pred_ys))
cnf_matrix = confusion_matrix(ground_ys, pred_ys)
np.set_printoptions(precision=2)
cnf_matrix = cnf_matrix.astype('float') / cnf_matrix.sum(
axis=1)[:, np.newaxis]
plt.figure()
plot_confusion_matrix(cnf_matrix,
classes=range(21),
normalize=True,
title='Normalized confusion matrix')
plt.show()
# print cnf_matrix
return losses, cnf_matrix
def train(config, model, sde, train_loader):
ema_model = copy.deepcopy(model)
optim = torch.optim.Adam(model.parameters(), lr=1e-3)
loss_fn = get_sde_loss_fn(sde,
train=True,
reduce_mean=True,
continuous=True,
likelihood_weighting=False)
ema = EMA(config.train.ema_decay)
num = 0
for epoch in range(config.train.epochs):
with tqdm(train_loader) as it:
for data, _ in it:
num += 1
optim.zero_grad()
loss = loss_fn(model, data.cuda())
loss.backward()
optim.step()
it.set_postfix(ordered_dict={
'train loss': loss.item(),
'epoch': epoch
},
refresh=False)
if num % config.train.update_ema_every == 0:
if num < 1000:
ema_model.load_state_dict(model.state_dict())
else:
ema.update_model_average(ema_model, model)
# sample
predictor = get_predictor(config.sampling.predictor)
corrector = get_corrector(config.sampling.corrector)
sampling_fn = get_pc_sampler(sde,
config.shape,
predictor=predictor,
corrector=corrector,
inverse_scaler=lambda x: x,
snr=config.sampling.snr,
eps=config.sampling.eps)
samples, n = sampling_fn(model)
samples = samples.detach().cpu().numpy()
plot_fig(samples)
# plt.show()
return sde
def histogramme(self):
histo = histogramme_valeur(self.resultat)
histo_lum = histogramme_luminance(self.resultat)
ddp_ = ddp(self.resultat)
res_ddp = ddp_cumule(self.resultat)
# print(pt_to_pt(self.resultat, histogram_equalization, res_ddp, 0, 255, 0, 255))
plot_fig(histo,
self.config.figure_param,
2,
2,
1,
title="Histogramme de Valeur",
clear=True)
plot_fig(histo_lum,
self.config.figure_param,
2,
2,
2,
title="Histogramme de Luminance")
plot_fig(ddp_, self.config.figure_param, 2, 2, 3, title="PDF")
plot_fig(res_ddp, self.config.figure_param, 2, 2, 4, title="CDF")
plt.show()
hypotheses, yy = get_hypotheses(num_eval_batches, num_eval_samples,
sess, y_hat, m.idx2token)
logging.info("# write results")
model_output = "barrages_E%02dL%.2f" % (epoch, _loss)
if not os.path.exists(hp.evaldir): os.makedirs(hp.evaldir)
translation = os.path.join(hp.evaldir, model_output)
with open(translation, 'w', encoding="utf-8") as fout:
fout.write("\n".join(hypotheses))
logging.info(hypotheses[1:10])
logging.info("# calc bleu score and append it to translation")
# calc_bleu(hp.eval_y2, translation)
bleu_score.append(
calc_belu_nltk(hp.bpe_model, hp.eval_y2, hypotheses))
cur_epoch = list(range(len(bleu_score)))
plot_fig(cur_epoch, bleu_score, "BLEU-2", hp.belu_fig)
logging.info("# save models")
ckpt_name = os.path.join(hp.logdir, model_output)
saver.save(sess, ckpt_name, global_step=_gs)
logging.info(
"after training of {} epochs, {} has been saved.".format(
epoch, ckpt_name))
logging.info("# fall back to train mode")
sess.run(train_init_op)
summary_writer.close()
logging.info("# Done")
num_bits=8,
loss_type='hybrid',
hybrid_coeff=0.001,
model_prediction='x_start',
model_output='logistic_pars').cuda()
optim = torch.optim.Adam(md.parameters(), lr=1e-3)
train_loader, test_loader, val_loader = get_mnist_loaders()
num = 0
for epoch in range(epochs):
with tqdm(train_loader) as it:
for x, label in it:
num += 1
optim.zero_grad()
x *= 255
x = x.long().cuda()
loss = torch.sum(diffusion.training_losses(md, x))
loss.backward()
optim.step()
it.set_postfix(ordered_dict={
'train_loss': loss.item(),
'epoch': epoch
},
refresh=False)
shape = (36, 1, 28, 28)
samples = diffusion.p_sample_loop(md, shape, num_timesteps=None)
plot_fig(samples.detach().cpu().numpy(), show=True)
plt.show()
if num % update_ema_every == 0:
if num < 1000:
ema_model.load_state_dict(model.state_dict())
else:
ema.update_model_average(ema_model, model)
# sample images
shape = (36, 1, 28, 28)
l = 5
cond = (torch.ones(36) * l).cuda()
eta = 0.
res = diffusion.sample(shape, cond=cond)
res = res.detach().cpu().numpy()
res1 = diffusion.ddim_sample(shape, eta, cond=cond).detach().cpu().numpy()
plot_fig(res)
plot_fig(res1)
plt.show()
elif tp == 'mix_gau_ddpm':
epochs = 20
ema_decay = 0.9999
update_ema_every = 10
p = 0.5
model = Unet(dim=16, channels=1, dim_mults=(1, 2, 4)).cuda()
# betas = cosine_beta_schedule(1000, )
betas = np.linspace(1e-4, 1e-2, 1000)
diffusion = mix_gaussian_ddpm(model,
loss_type='l2',
betas=betas,
refresh=False)
torch.save(s_phi.state_dict(), schedule_path)
else:
s_phi.load_state_dict(torch.load(schedule_path))
if not load_res:
for i in range(len(hat_alphan)):
for j in range(len(hat_betan)):
schedule_beta = schedule.noise_schedule(alpha=hat_alphan[i],
beta=hat_betan[j],
shape=(1, 1, 28, 28))
write_to_file(schedule_beta.detach().cpu().numpy(), beta_path)
img = schedule.sample(shape=config.shape, betas=schedule_beta)
fig = plot_fig(
img.detach().cpu().numpy(), father_path + '/fig/fig_' +
str(round(hat_alphan[i].item(), 2)) + '_' +
str(round(hat_betan[j].item(), 2)) + '_' +
str(len(schedule_beta)) + '.jpg')
# plt.imsave(father_path + '/fig/fig_' + str(i) + '_' + str(j) + '.jpg', fig)
plt.close()
else:
schedule_betas = read_file(beta_path).to('cuda:0')
for i in range(schedule_betas.shape[0]):
schedule_beta = schedule_betas[i]
img = schedule.sample(shape=config.shape, betas=schedule_beta)
fig = plot_fig(
img.detach().cpu().numpy(),
father_path + '/fig/fig_' + str(i) + '_' + str(i + 10) + '.jpg')
# plt.imsave(father_path + '/fig/fig_' + str(i) + '_' + str(i+10) + '.jpg', fig)
plt.close()
def padim(category, batch_size, rd, net_type='eff', is_plot=False):
loader = MVTecADLoader()
loader.load(category=category, repeat=1, max_rot=0)
train_set = loader.train.batch(batch_size=batch_size,
drop_remainder=True).shuffle(
buffer_size=loader.num_train,
reshuffle_each_iteration=True)
test_set = loader.test.batch(batch_size=1, drop_remainder=False)
net, _shape = embedding_net(net_type=net_type)
h, w, c = _shape # height and width of layer1, channel sum of layer 1, 2, and 3, and randomly sampled dimension
out = []
for x, _, _ in train_set:
l1, l2, l3 = net(x)
_out = tf.reshape(embedding_concat(embedding_concat(l1, l2), l3),
(batch_size, h * w, c)) # (b, h x w, c)
out.append(_out.numpy())
# calculate multivariate Gaussian distribution.
out = np.concatenate(out, axis=0)
out = np.transpose(out, axes=[0, 2, 1]) # (b, c, h * w)
# RD: random dimension selecting
tmp = tf.unstack(out, axis=0)
_tmp = []
rd_indices = tf.random.shuffle(tf.range(c))[:rd]
for tensor in tmp:
_tmp.append(tf.gather(tensor, rd_indices))
out = tf.stack(_tmp, axis=0)
mu = np.mean(out, axis=0)
cov = np.zeros((rd, rd, h * w))
identity = np.identity(rd)
for idx in range(h * w):
cov[:, :, idx] = np.cov(out[:, :, idx], rowvar=False) + 0.01 * identity
train_outputs = [mu, cov]
out, gt_list, gt_mask, batch_size, test_imgs = [], [], [], 1, []
# x - data | y - mask | z - binary label
for x, y, z in test_set:
test_imgs.append(x.numpy())
gt_list.append(z.numpy())
gt_mask.append(y.numpy())
l1, l2, l3 = net(x)
_out = tf.reshape(embedding_concat(embedding_concat(l1, l2), l3),
(batch_size, h * w, c)) # (BS, h x w, c)
out.append(_out.numpy())
# calculate multivariate Gaussian distribution. skip random dimension selecting
out = np.concatenate(out, axis=0)
gt_list = np.concatenate(gt_list, axis=0)
out = np.transpose(out, axes=[0, 2, 1])
# RD
tmp = tf.unstack(out, axis=0)
_tmp = []
for tensor in tmp:
_tmp.append(tf.gather(tensor, rd_indices))
out = tf.stack(_tmp, axis=0)
b, _, _ = out.shape
dist_list = []
for idx in range(h * w):
mu = train_outputs[0][:, idx]
cov_inv = np.linalg.inv(train_outputs[1][:, :, idx])
dist = [mahalanobis(sample[:, idx], mu, cov_inv) for sample in out]
dist_list.append(dist)
dist_list = np.reshape(np.transpose(np.asarray(dist_list), axes=[1, 0]),
(b, h, w))
################
# DATA Level #
################
# upsample
score_map = tf.squeeze(
tf.image.resize(np.expand_dims(dist_list, -1), size=[h, w])).numpy()
for i in range(score_map.shape[0]):
score_map[i] = gaussian_filter(score_map[i], sigma=4)
# Normalization
max_score = score_map.max()
min_score = score_map.min()
scores = (score_map - min_score) / (max_score - min_score)
scores = -scores
# calculate image-level ROC AUC score
img_scores = scores.reshape(scores.shape[0], -1).max(axis=1)
gt_list = np.asarray(gt_list)
img_roc_auc = metrics.roc_auc_score(gt_list, img_scores)
if is_plot is True:
fpr, tpr, _ = metrics.roc_curve(gt_list, img_scores)
precision, recall, _ = metrics.precision_recall_curve(
gt_list, img_scores)
save_dir = os.path.join(os.getcwd(), 'img')
if os.path.isdir(save_dir) is False:
os.mkdir(save_dir)
draw_auc(fpr, tpr, img_roc_auc,
os.path.join(save_dir, 'AUROC-{}.png'.format(category)))
base_line = np.sum(gt_list) / len(gt_list)
draw_precision_recall(
precision, recall, base_line,
os.path.join(os.path.join(save_dir, 'PR-{}.png'.format(category))))
#################
# PATCH Level #
#################
# upsample
score_map = tf.squeeze(
tf.image.resize(np.expand_dims(dist_list, -1), size=[224,
224])).numpy()
for i in range(score_map.shape[0]):
score_map[i] = gaussian_filter(score_map[i], sigma=4)
# Normalization
max_score = score_map.max()
min_score = score_map.min()
scores = (score_map - min_score) / (max_score - min_score)
# Note that Binary mask indicates 0 for good and 1 for anomaly. It is opposite from our setting.
# scores = -scores
# calculate per-pixel level ROCAUC
gt_mask = np.asarray(gt_mask)
fp_list, tp_list, _ = metrics.roc_curve(gt_mask.flatten(),
scores.flatten())
patch_auc = metrics.auc(fp_list, tp_list)
precision, recall, threshold = metrics.precision_recall_curve(
gt_mask.flatten(), scores.flatten(), pos_label=1)
numerator = 2 * precision * recall
denominator = precision + recall
numerator[np.where(denominator == 0)] = 0
denominator[np.where(denominator == 0)] = 1
# get optimal threshold
f1_list = numerator / denominator
best_ths = threshold[np.argmax(f1_list).astype(int)]
print('[{}] image ROCAUC: {:.04f}\t pixel ROCAUC: {:.04f}'.format(
category, img_roc_auc, patch_auc))
if is_plot is True:
save_dir = os.path.join(os.getcwd(), 'img')
if os.path.isdir(save_dir) is False:
os.mkdir(save_dir)
plot_fig(test_imgs, scores, gt_mask, best_ths, save_dir, category)
return img_roc_auc, patch_auc