def train_binary_predictor():
clf = classifier(vocab_size, embedding_size, hidden_size, value_size)
optimizer = opt.Adam(clf.parameters(), lr=1e-3)
num_epoch = 20
for i in range(num_epoch):
print('epoch {}/{}'.format(i + 1, num_epoch))
shuffle_indices = np.random.permutation(np.arange(data_size))
# shuffle_indices = np.arange(data_size)
txt_ = ref[shuffle_indices]
lengths_ = ref_lengths[shuffle_indices]
tgt_ = binary_representation[shuffle_indices]
for j in range(iters):
start = j * batch_size
end = min(data_size, (j + 1) * batch_size)
y = clf.forward(torch.LongTensor(txt_[start:end]), torch.LongTensor(lengths_[start:end]))
loss = -torch.sum(torch.mul(torch.log(y), torch.LongTensor(tgt_[start:end]))) \
- torch.sum(torch.mul(torch.log(1-y), torch.LongTensor(1 - tgt_[start:end])))
optimizer.zero_grad()
loss.backward()
optimizer.step()
print(loss)
torch.save(clf.state_dict(), 'checkpoint/' + str(101+i) + '-parameter.pkl')
def classifierA():
if request.method == "POST":
url = request.form.get("url")
out = classifier(url)
return render_template("index.html",out=out)
return render_template("index.html",out="")
def object_recognition_classifier(clf, data_train, data_test, label_train,
label_test, num_of_features):
"""
Train a classifier and test it based on provided data
:param clf:
:param data_train:
:param data_test:
:param label_train:
:param label_test:
:param num_of_features:
:return: accuracy, prediction
"""
train_cats_data = data_train.reshape(-1, num_of_features)
train_cats_label = label_train.reshape(-1, 1).flatten()
test_cats_data = data_test.reshape(-1, num_of_features)
test_cats_label = label_test.reshape(-1, 1).flatten()
y_acc, y_pred = classifier(clf, train_cats_data, test_cats_data,
train_cats_label, test_cats_label)
return y_acc, y_pred
def main():
use_cuda = True
batch_size=64
lr=0.1
test_batch_size=128
epochs=20
device = torch.device("cuda" if use_cuda else "cpu")
train_kwargs = {'batch_size': batch_size}
test_kwargs = {'batch_size': test_batch_size}
if use_cuda:
cuda_kwargs = {'num_workers': 1,
'pin_memory': True,
'shuffle': True}
train_kwargs.update(cuda_kwargs)
test_kwargs.update(cuda_kwargs)
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Resize((32,32))
])
dataset1 = Mnist_m('mnist_m',train=True,transform=transform)
dataset2=Mnist_m('mnist_m',train=False,transform=transform)
datasets=[dataset1,dataset2]
dataset_test = ConcatDataset(datasets)
test_loader = torch.utils.data.DataLoader(dataset_test, **test_kwargs)
model = classifier().cuda()#.to(device)
model.load_state_dict(torch.load('classifier_complete.pt'))
test(model, device, test_loader)
def train(epoch, dataloader):
loss = []
f1score = []
for i in range(epoch):
y_pred = []
y_true = []
for batch_idx, (data, target) in enumerate(dataloader):
data, target = Variable(data), Variable(target)
data = data.type(torch.cuda.FloatTensor)
target = target.type(torch.cuda.FloatTensor)
optimizer.zero_grad()
output = classifier(data)
loss = F.binary_cross_entropy(output, target)
loss.backward()
optimizer.step()
output = output.cpu().detach().numpy()
y_pred.append(output)
target = target.cpu().numpy()
y_true.append(target)
if batch_idx % 100 == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
i, batch_idx * len(data), len(dataloader.dataset),
100. * batch_idx / len(dataloader), loss.item()))
y_pred = get_pred(y_pred)
f_score = get_fscore(y_true, y_pred)
loss.append(loss.item)
f1score.append(f_score)
print('Train Epoch: {} \tf1_score: {:.6f}'.format(epoch, f_score))
return loss, f1score
def classify(image, top_k, k_patches, ckpt_path, imagenet_path):
wnids, words = tu.load_imagenet_meta(
os.path.join(imagenet_path, 'data/meta.mat'))
image_patches = tu.read_k_patches(image, k_patches)
x = tf.placeholder(tf.float32, [None, 224, 224, 3])
_, pred = model.classifier(x, dropout=1.0)
avg_prediction = tf.div(tf.reduce_sum(pred, 0), k_patches)
scores, indexes = tf.nn.top_k(avg_prediction, k=top_k)
saver = tf.train.Saver()
with tf.Session(config=tf.ConfigProto()) as sess:
saver.restore(sess, os.path.join(ckpt_path, 'cnn.ckpt'))
s, i = sess.run([scores, indexes], feed_dict={x: image_patches})
s, i = np.squeeze(s), np.squeeze(i)
print('AlexNet saw:')
for idx in range(top_k):
print('{} - score: {}'.format(words[i[idx]], s[idx]))
def object_based_5_fold_cross_validation(clf, data_train, data_test, labels, num_of_features):
"""
Perform object based 5 fold cross validation and return mean accuracy
:param clf: classifier
:param data_train: Training dataset
:param data_test: Testing dataset
:param labels: True labels
:param num_of_features: Number of features of the robot
:return: mean accuracy of 5 fold validation
"""
tts = train_test_splits(OBJECTS_PER_CATEGORY)
my_acc = []
for a_fold in sorted(tts):
train_cats_index = tts[a_fold]["train"]
test_cats_index = tts[a_fold]["test"]
train_cats_data = data_train[:, train_cats_index]
train_cats_label = labels[:, train_cats_index]
train_cats_data = train_cats_data.reshape(-1, num_of_features)
train_cats_label = train_cats_label.reshape(-1, 1).flatten()
test_cats_data = data_test[:, test_cats_index]
test_cats_label = labels[:, test_cats_index]
test_cats_data = test_cats_data.reshape(-1, num_of_features)
test_cats_label = test_cats_label.reshape(-1, 1).flatten()
y_acc, y_pred = classifier(clf, train_cats_data, test_cats_data, train_cats_label, test_cats_label)
my_acc.append(y_acc)
return np.mean(my_acc)
def main():
in_arg = get_input_args()
data_dir = in_arg.data_dir
save_dir = in_arg.save_dir
arch = in_arg.arch
learning_rate = in_arg.learning_rate
hidden_units = in_arg.hidden_units
epochs = in_arg.epochs
#processing_unit = in_arg.gpu
if torch.cuda.is_available() and in_arg.gpu == 'gpu':
print('GPU will be used')
processing_unit = 'gpu'
elif torch.cuda.is_available() == False:
print('CPU will be used')
processing_unit = 'cpu'
print(in_arg)
training_dataloaders, validation_dataloaders, testing_dataloaders, class_to_idx = load_datas(
data_dir)
pre_model = pretrained_model(arch)
model = classifier(pre_model, hidden_units)
after_train_model = train_model(model, training_dataloaders,
validation_dataloaders, learning_rate,
epochs, processing_unit)
valid_model(after_train_model, testing_dataloaders, processing_unit)
save_checkpoint(model, save_dir, class_to_idx)
def upload_file():
if request.method == 'POST':
f1 = request.form['text1']
f2 = request.form['text2']
df, result = form_dataset(f1, f2)
df.both = df.both.apply(mystem_combined)
from model import obrabotka
obrabotka(df.both)
from model import features
geo_class = features(df, result)
from model import classifier
answers, predictions = classifier(geo_class)
if answers[0] == 1:
whether_dup = u'являются'
proba = round(predictions[0][1] * 100, 2)
else:
whether_dup = u'не являются'
proba = round(predictions[0][0] * 100, 2)
visible = True
return render_template('index.html',
visible=visible,
proba=proba,
whether_dup=whether_dup,
f1=f1,
f2=f2)
def findFractureClassifier(params):
with open(os.path.join(params.median, "ribs.json"), "r") as file:
ribs = json.load(file)
dataset = data.ClassifierTestDataset(params.processed, ribs, params)
# classifier = model.classifier(False, os.path.join(
# params.model_path, "resnet34-333f7ec4.pth"))
classifier = model.classifier(False)
classifier = torch.nn.DataParallel(classifier, device_ids=[0])
if params.useGPU:
classifier = classifier.cuda()
d = torch.load(os.path.join(params.model_path, "classifier.pt"))
# d1 = torch.load()
# d.update(d1)
classifier.load_state_dict(d)
classifier.eval()
detected = []
cnt = 0
timer = Timer()
for batch, centers, imgIDs in torch.utils.data.DataLoader(
dataset, batch_size=params.batchSize, pin_memory=True,
num_workers=0):
cnt += 1
if params.useGPU:
batch = batch.cuda()
with torch.no_grad():
output = classifier(batch)
output = torch.nn.functional.softmax(output, 1)
output = output.cpu().numpy()
centers = centers.numpy()
imgIDs = imgIDs.numpy()
for i in range(output.shape[0]):
out = output[i]
if out[1] > params.detectThreshold:
detected.append((centers[i], imgIDs[i], out[1]))
if cnt % 100 == 0:
print(f"Batch {cnt} {timer()}")
with open(params.anno_path, "r") as file:
anno = json.load(file)
anno["annotations"] = []
regionSize = params.detectRegionSize
for i, d in enumerate(detected):
center, imgID, score = d
anno["annotations"].append({
"bbox": [
float(center[0]) - regionSize / 2,
float(center[1]) - regionSize / 2, regionSize, regionSize
],
"id":
i,
"image_id":
int(imgID),
"score":
float(score)
})
with open(os.path.join(params.median, "detection.json"), "w") as file:
json.dump(anno, file, indent=4)
def build_model(self):
input_width = self.config['model']['input_width']
input_height = self.config['model']['input_height']
class_num = self.config['model']['class_num']
backbone = self.config['model']['backbone']
train_base = self.config['train']['train_base']
return classifier(class_num, input_width, input_height, backbone,
train_base)
def inference(hypes, image, phase):
fex = feature_extractorl('vgg16.npy')
feature = fex.build(image)
string_length = hypes['arch']['maxstr']
logits = hypes['arch']['dictlen']
char_list = [
tf.expand_dims(
classifier('chcnt_' + str(i), logits).build(feature, phase), 1)
for i in range(string_length)
]
return tf.concat(char_list, 1)
def classifierA():
if request.method == "POST":
Clump = int(request.form.get("Clump"))
UnifSize = int(request.form.get("UnifSize"))
UnifShape = int(request.form.get("UnifShape"))
MargAdh = int(request.form.get("MargAdh"))
SingEpiSize = int(request.form.get("SingEpiSize"))
BareNuc = int(request.form.get("BareNuc"))
BlandChrom = int(request.form.get("BlandChrom"))
NormNucl = int(request.form.get("NormNucl"))
Mit = int(request.form.get("Mit"))
out = classifier(Clump, UnifSize, UnifShape, MargAdh, SingEpiSize,
BareNuc, BlandChrom, NormNucl, Mit)
return render_template("login.html",
out="The Class Of the Tumor Is:" + out)
return render_template("login.html", out="")
def test(
top_k,
k_patches,
display_step,
imagenet_path,
ckpt_path):
test_images = sorted(os.listdir(os.path.join(imagenet_path, 'ILSVRC2012_img_val')))
test_labels = tu.read_test_labels(os.path.join(imagenet_path, 'data/ILSVRC2012_validation_ground_truth.txt'))
test_examples = len(test_images)
x = tf.placeholder(tf.float32, [None, 224, 224, 3])
y = tf.placeholder(tf.float32, [None, 1000])
_, pred = model.classifier(x, 1.0)
avg_prediction = tf.div(tf.reduce_sum(pred, 0), k_patches)
top1_correct = tf.equal(tf.argmax(avg_prediction, 0), tf.argmax(y, 1))
top1_accuracy = tf.reduce_mean(tf.cast(top1_correct, tf.float32))
topk_correct = tf.nn.in_top_k(tf.stack([avg_prediction]), tf.argmax(y, 1), k=top_k)
topk_accuracy = tf.reduce_mean(tf.cast(topk_correct, tf.float32))
saver = tf.train.Saver()
with tf.Session(config=tf.ConfigProto()) as sess:
saver.restore(sess, os.path.join(ckpt_path, 'cnn.ckpt'))
total_top1_accuracy = 0.
total_topk_accuracy = 0.
for i in range(test_examples):
image_patches = tu.read_k_patches(os.path.join(imagenet_path, 'ILSVRC2012_img_val', test_images[i]), k_patches)
label = test_labels[i]
top1_a, topk_a = sess.run([top1_accuracy, topk_accuracy], feed_dict={x: image_patches, y: [label]})
total_top1_accuracy += top1_a
total_topk_accuracy += topk_a
if i % display_step == 0:
print ('Examples done: {:5d}/{} ---- Top-1: {:.4f} -- Top-{}: {:.4f}'.format(i + 1, test_examples, total_top1_accuracy / (i + 1), top_k, total_topk_accuracy / (i + 1)))
print ('---- Final accuracy ----')
print ('Top-1: {:.4f} -- Top-{}: {:.4f}'.format(total_top1_accuracy / test_examples, top_k, total_topk_accuracy / test_examples))
print ('Top-1 error rate: {:.4f} -- Top-{} error rate: {:.4f}'.format(1 - (total_top1_accuracy / test_examples), top_k, 1 - (total_topk_accuracy / test_examples)))
def classify(
image,
top_k,
k_patches,
ckpt_path,
imagenet_path):
""" Procedure to classify the image given through the command line
Args:
image: path to the image to classify
top_k: integer representing the number of predictions with highest probability
to retrieve
k_patches: number of crops taken from an image and to input to the model
ckpt_path: path to model's tensorflow checkpoint
imagenet_path: path to ILSRVC12 ImageNet folder containing train images,
validation images, annotations and metadata file
"""
wnids, words = tu.load_imagenet_meta(os.path.join(imagenet_path, 'data/meta.mat'))
# taking a few crops from an image
image_patches = tu.read_k_patches(image, k_patches)
x = tf.placeholder(tf.float32, [None, 224, 224, 3])
_, pred = model.classifier(x, dropout=1.0)
# calculate the average precision through the crops
avg_prediction = tf.div(tf.reduce_sum(pred, 0), k_patches)
# retrieve top 5 scores
scores, indexes = tf.nn.top_k(avg_prediction, k=top_k)
saver = tf.train.Saver()
with tf.Session(config=tf.ConfigProto()) as sess:
saver.restore(sess, os.path.join(ckpt_path, 'cnn.ckpt'))
s, i = sess.run([scores, indexes], feed_dict={x: image_patches})
s, i = np.squeeze(s), np.squeeze(i)
print('AlexNet saw:')
for idx in range(top_k):
print ('{} - score: {}'.format(words[i[idx]], s[idx]))
def __init__(self, dataset_folder, mode, mean, label, name, quantized,
prefix, subset_size):
self.dataset = dataset_folder
self.prefix = prefix
self.ds = dataset(self.dataset, label)
self.random_pick = False
self.max_images = subset_size
self.filenames = self.ds.get_subset(max_size=self.max_images)
self.mean = mean
self.mode = mode
self.model_file = prefix + '/' + name + '/' + quantized + '_model.prototxt'
self.weights_file = prefix + '/' + name + '/' + quantized + '_weights.caffemodel'
self.quantized = quantized
self.net = classifier(self.model_file, self.weights_file, self.mean,
self.mode)
self.inj = injector(self.net, self.quantized)
def __init__(self):
super(runModel, self).__init__()
self.keep_going = False
self.load_path = "weight/91acc_9category_1024_512_256.pt" # 继续训练时，要加载的参数文件
self.batch_size = 30 # 分批训练数据、每批数据量
self.learning_rate = 0.01 # 1e-2 # 学习率
self.num_epoches = 700 # 训练次数
cfg_path = "./cfg/network.cfg"
net_block = parse_model_cfg(cfg_path)[0] # [net]
self.n_classes = net_block["n_classes"] # 类别数
# 保存用于可视化观察参数变化的列表
self.loss_list, self.lr_list, self.acc_list = [], [], [
] # 记录损失值\学习率\准确率变化
self.layer_list = [1024, 512, 256]
# self.model = MyFC(self.layer_list, 42, self.n_classes)
self.model = classifier("./cfg/network.cfg")
self.train_loader = MyDataLoader(batch_size=self.batch_size).train()
self.test_loader = MyDataLoader(batch_size=self.batch_size).test()
self.best_acc = 0 # 最高准确类
self.writer = SummaryWriter(logdir='./log') # 记录训练日志
shuffle=True,
# num_workers=config.cpu_processor,
drop_last=True)
test_dataset = custom_dataset.Custom_dataset(word_to_index,
path_csv="train_data.csv")
test_data = test_dataset.get_data()
test_loader = DataLoader(test_data,
batch_size=config.batch,
shuffle=False,
num_workers=config.cpu_processor,
drop_last=True)
# 모델 설정
device = 'cpu' #torch.device(config.gpu if torch.cuda.is_available() else 'cpu')
model = model.classifier(len(word_to_index))
model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=config.learning_rate)
loss_function = nn.CrossEntropyLoss()
# 훈련
step_list = []
loss_list = []
acc_test_list = []
acc_dev_list = []
step = 0
for i in range(config.epoch):
print("epoch = ", i)
start = time.time()
for n, (label, sent) in enumerate(train_loader):
drop_last=True)
# test_loader = DataLoader(test_data,
# batch_size=c.batch,
# shuffle=False,
# num_workers=c.cpu_processor,
# drop_last=True)
dev_loader = DataLoader(val_data,
batch_size=c.batch,
shuffle=False,
num_workers=c.cpu_processor,
drop_last=True)
# 모델 설정
model = model.classifier()
model.cuda()
optimizer = torch.optim.Adam(model.parameters(), lr=c.learning_rate)
loss_function = nn.CrossEntropyLoss()
# 훈련
step_list = []
loss_list = []
acc_test_list = []
acc_dev_list = []
step = 0
for i in range(c.epoch):
start = time.time()
for n, (q, label) in enumerate(train_loader):
optimizer.zero_grad() # 초기화
init_trl_label = keras.utils.to_categorical(init_trl_label)
init_trunl_data = np.expand_dims(init_trunl_data, axis=4)
init_trunl_label = keras.utils.to_categorical(init_trunl_label)
te_data = np.expand_dims(te_data, axis=4)
te_label = keras.utils.to_categorical(te_label)
init_trl_set = tf.data.Dataset.from_tensor_slices(
(init_trl_data,
init_trl_label)).shuffle(len(init_trl_data)).batch(opt.BATCH_SIZE)
te_set = tf.data.Dataset.from_tensor_slices(
(te_data, te_label)).batch(opt.BATCH_SIZE)
"""
Model.
"""
'''create classifier'''
classifier = classifier()
optim = keras.optimizers.Adam(lr=opt.LR, decay=opt.DECAY)
classifier.build(input_shape=(opt.BATCH_SIZE, opt.WINDOW_SIZE, opt.WINDOW_SIZE,
opt.CHANNEL, 1))
classifier.summary()
'''create feature generator & feature discriminator'''
fea_g = generator()
fea_g.build(input_shape=(opt.BATCH_SIZE, opt.DIM_Z))
fea_g.summary()
fea_d = discriminator()
fea_d.build(input_shape=(opt.BATCH_SIZE, 17 * 17 * 64))
fea_d.summary()
d_loss, g_loss = dcgan_loss()
fea_g_optim = keras.optimizers.Adam(learning_rate=opt.GAN_LR, beta_1=0.5)
def __init__(self, args):
# Misc
use_cuda = args.cuda and torch.cuda.is_available()
self.device = 'cuda' if use_cuda else 'cpu'
self.name = args.name
self.max_iter = int(args.max_iter)
self.print_iter = args.print_iter
self.global_iter = 0
self.global_iter_cls = 0
self.pbar = tqdm(total=self.max_iter)
self.pbar_cls = tqdm(total=self.max_iter)
# Data
self.dset_dir = args.dset_dir
self.dataset = args.dataset
self.batch_size = args.batch_size
self.eval_batch_size = args.eval_batch_size
self.data_loader = return_data(args, 0)
self.data_loader_eval = return_data(args, 2)
# Networks & Optimizers
self.z_dim = args.z_dim
self.gamma = args.gamma
self.beta = args.beta
self.lr_VAE = args.lr_VAE
self.beta1_VAE = args.beta1_VAE
self.beta2_VAE = args.beta2_VAE
self.lr_D = args.lr_D
self.beta1_D = args.beta1_D
self.beta2_D = args.beta2_D
self.alpha = args.alpha
self.beta = args.beta
self.grl = args.grl
self.lr_cls = args.lr_cls
self.beta1_cls = args.beta1_D
self.beta2_cls = args.beta2_D
if args.dataset == 'dsprites':
self.VAE = FactorVAE1(self.z_dim).to(self.device)
self.nc = 1
else:
self.VAE = FactorVAE2(self.z_dim).to(self.device)
self.nc = 3
self.optim_VAE = optim.Adam(self.VAE.parameters(),
lr=self.lr_VAE,
betas=(self.beta1_VAE, self.beta2_VAE))
self.pacls = classifier(30, 2).cuda()
self.revcls = classifier(30, 2).cuda()
self.tcls = classifier(30, 2).cuda()
self.trevcls = classifier(30, 2).cuda()
self.targetcls = classifier(59, 2).cuda()
self.pa_target = classifier(30, 2).cuda()
self.target_pa = paclassifier(1, 1).cuda()
self.pa_pa = classifier(30, 2).cuda()
self.D = Discriminator(self.z_dim).to(self.device)
self.optim_D = optim.Adam(self.D.parameters(),
lr=self.lr_D,
betas=(self.beta1_D, self.beta2_D))
self.optim_pacls = optim.Adam(self.pacls.parameters(), lr=self.lr_D)
self.optim_revcls = optim.Adam(self.revcls.parameters(), lr=self.lr_D)
self.optim_tcls = optim.Adam(self.tcls.parameters(), lr=self.lr_D)
self.optim_trevcls = optim.Adam(self.trevcls.parameters(),
lr=self.lr_D)
self.optim_cls = optim.Adam(self.targetcls.parameters(),
lr=self.lr_cls)
self.optim_pa_target = optim.Adam(self.pa_target.parameters(),
lr=self.lr_cls)
self.optim_target_pa = optim.Adam(self.target_pa.parameters(),
lr=self.lr_cls)
self.optim_pa_pa = optim.Adam(self.pa_pa.parameters(), lr=self.lr_cls)
self.nets = [
self.VAE, self.D, self.pacls, self.targetcls, self.revcls,
self.pa_target, self.tcls, self.trevcls
]
# Visdom
self.viz_on = args.viz_on
self.win_id = dict(D_z='win_D_z',
recon='win_recon',
kld='win_kld',
acc='win_acc')
self.line_gather = DataGather('iter', 'soft_D_z', 'soft_D_z_pperm',
'recon', 'kld', 'acc')
self.image_gather = DataGather('true', 'recon')
if self.viz_on:
self.viz_port = args.viz_port
self.viz = visdom.Visdom(port=self.viz_port)
self.viz_ll_iter = args.viz_ll_iter
self.viz_la_iter = args.viz_la_iter
self.viz_ra_iter = args.viz_ra_iter
self.viz_ta_iter = args.viz_ta_iter
if not self.viz.win_exists(env=self.name + '/lines',
win=self.win_id['D_z']):
self.viz_init()
# Checkpoint
self.ckpt_dir = os.path.join(args.ckpt_dir, args.name)
self.ckpt_save_iter = args.ckpt_save_iter
mkdirs(self.ckpt_dir + "/cls")
mkdirs(self.ckpt_dir + "/vae")
if args.ckpt_load:
self.load_checkpoint(args.ckpt_load)
# Output(latent traverse GIF)
self.output_dir = os.path.join(args.output_dir, args.name)
self.output_save = args.output_save
mkdirs(self.output_dir)
def inference_from_embeddings(embeddings, threshold):
output = classifier(embeddings)
predicted_probabilities = F.softmax(output, dim=1).detach().numpy()
return "HELPFUL" if predicted_probabilities[
0, 1] >= threshold else "NOT HELPFUL"
def main():
# Training settings
use_cuda = True
gamma = 0.7
save_model = True
batch_size = 128 #128
lr = 0.01
test_batch_size = 256
epochs = 500
device = torch.device("cuda" if use_cuda else "cpu")
train_kwargs = {'batch_size': batch_size}
test_kwargs = {'batch_size': test_batch_size}
if use_cuda:
cuda_kwargs = {'num_workers': 8, 'pin_memory': True, 'shuffle': True}
train_kwargs.update(cuda_kwargs)
test_kwargs.update(cuda_kwargs)
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Resize((32, 32)),
#transforms.Normalize((0.1307,), (0.3081,))
])
dataset1 = datasets.MNIST('../data',
train=True,
download=True,
transform=transform)
dataset2 = datasets.MNIST('../data', train=False, transform=transform)
train_loader = torch.utils.data.DataLoader(dataset1, **train_kwargs)
test_loader = torch.utils.data.DataLoader(dataset2, **test_kwargs)
model = classifier().cuda() #.to(device)
model2 = Net().cuda()
model.load_state_dict(torch.load('classifier_basic.pt'))
optimizer = optim.Adam(model2.parameters(), lr=0.001, betas=(0.9, 0.999))
#optimizer = optim.SGD(model2.parameters(),lr=0.001,momentum=0.9,nesterov=True)
#scheduler = StepLR(optimizer, step_size=1,gamma=gamma)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
mode='max',
factor=0.9,
patience=3)
for epoch in range(1, epochs + 1):
print("Epoch {}".format(epoch))
model2.train()
log_interval = 10
loss_sum = 0.0
from tqdm import tqdm
for (data, target) in tqdm(train_loader):
data, target = data.to(device), target.to(device)
optimizer.zero_grad()
output, loss_weight = (model2((data)))
#print(output[0][0])
#print(torch.unique(output[0].detach()))
cv2.imwrite(
'demo.png',
cv2.pyrUp(
cv2.pyrUp(
output.detach().cpu().numpy()[0].transpose(1, 2, 0) *
255)))
cv2.imwrite(
'demo_orig.png',
cv2.pyrUp(
cv2.pyrUp(
data.detach().cpu().numpy()[0].transpose(1, 2, 0) *
255)))
output_ = (model(output))
#print(data.size())
loss_sep = (loss_weight.mean())
loss_class = -(nn.CrossEntropyLoss()(output_, target)) * 1e+2
loss = loss_class #+loss_sep
loss_class.backward()
#print("Loss Train",float(np.abs(float(loss))))
loss_sum += np.abs(
float(loss) / (len(train_loader.dataset) // data.shape[0]))
#print("Loss_Sep {} Loss_Classification {}".format((float(loss_sep)),np.abs(float(loss_class))))
optimizer.step()
#break
#if batch_idx % log_interval == 0:
print('Epoch {} Train loss {}'.format(epoch, loss_sum))
model2.eval()
model.eval()
test_loss = 0
correct = 0
with torch.no_grad():
for data, target in test_loader:
data, target = data.to(device), target.to(device)
output, loss_weight = model2(data)
output = model(output)
test_loss += np.abs(
(nn.CrossEntropyLoss()(output, target).cpu().item()) +
(loss_weight.mean().cpu())) # sum up batch loss
pred = output.argmax(
dim=1,
keepdim=True) # get the index of the max log-probability
correct += pred.eq(target.view_as(pred)).sum().item()
test_loss /= len(test_loader.dataset)
print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.
format(test_loss, correct, len(test_loader.dataset),
100. * correct / len(test_loader.dataset)))
torch.save(model2.state_dict(), 'classifier_advanced.pt')
scheduler.step(test_loss)
if save_model:
torch.save(model.state_dict(), "mnist_cnn.pt")
batch_size=BATCH_SIZE,
pad_mode=pad_mode)
train_loader = DataLoader(train_dataset)
valid_dataset = TestDataset(texts_valid,
labels_valid,
dico,
batch_size=BATCH_SIZE,
pad_mode=pad_mode)
valid_loader = DataLoader(valid_dataset)
# Instantiate the model
model = mod.classifier(embedding_dim,
num_hidden_nodes,
num_output_nodes,
num_layers,
init=init,
glove=glove,
embed_weights=embed_weights,
bidirectional=bidirectional,
dropout=dropout)
model.to(device)
# Visualizing architecture
print(model)
#define optimizer and loss
optimizer = optim.Adagrad(
model.parameters(), lr=lr, weight_decay=L2
) #Momentum seems not to be in pytorch adagrad ...
criterion = nn.CrossEntropyLoss()
data_transform = transforms.Compose(
[transforms.Resize((16, 16)),
transforms.ToTensor()])
s_trainset = datasets.MNIST('tmp',
download=True,
train=True,
transform=data_transform)
s_testset = datasets.MNIST('tmp', train=False, transform=data_transform)
s_trainloader = DataLoader(s_trainset, batch_size=batch_size, shuffle=True)
s_testloader = DataLoader(s_testset, batch_size=batch_size, shuffle=True)
t_trainset, t_testset = load_usps(data_per_class) #transformの指定は禁止
t_trainloader = DataLoader(t_trainset, batch_size=batch_size, shuffle=True)
t_testloader = DataLoader(t_testset, batch_size=64, shuffle=True)
net_g = Encoder()
net_h = classifier()
net_DCD = DCD()
loss_func = torch.nn.CrossEntropyLoss() #損失関数は共通
#ソースにおいてgとhを訓練
print("part 1 : initial training for g and h")
optimizer = torch.optim.Adam(list(net_g.parameters()) +
list(net_h.parameters()),
lr=0.001) #optimizerが両者を更新
net_g = net_g.to(device)
net_h = net_h.to(device)
net_DCD = net_DCD.to(device)
if not device == "cpu":
net_g = nn.DataParallel(net_g)
net_h = nn.DataParallel(net_h)
net_DCD = nn.DataParallel(net_DCD)
#define hyperparameters
size_of_vocab = tk.vocab_len()
embedding_dim = args.embedding_dim
num_hidden_nodes = args.num_hidden_nodes
num_output_nodes = 1
num_layers = args.num_layers
bidirection = True
dropout = args.dropout
if args.num_layers == 1:
dropout = 0.0
#instantiate the model
model = classifier(size_of_vocab,
embedding_dim,
num_hidden_nodes,
num_output_nodes,
num_layers,
bidirectional=bidirection,
dropout=dropout)
print(model)
def count_parameters(model):
return sum(p.numel() for p in model.parameters() if p.requires_grad)
print(f'The model has {count_parameters(model):,} trainable parameters')
def aucroc(y_scores, y_true):
def main():
np.random.seed(42)
phoneHMMs = np.load('../Assignment2/lab2_models_all.npz',
allow_pickle=True)['phoneHMMs'].item()
phones = sorted(phoneHMMs.keys())
nstates = {phone: phoneHMMs[phone]['means'].shape[0] for phone in phones}
state_list = [
ph + '_' + str(idx) for ph in phones for idx in range(nstates[ph])
]
filenames = ['Data/traindata.npz', 'Data/testdata.npz']
sets = [
'tidigits/disc_4.1.1/tidigits/train',
'tidigits/disc_4.2.1/tidigits/test'
]
for idx, file_name in enumerate(filenames):
if not os.path.isfile(file_name):
data = []
for root, dirs, files in os.walk(sets[idx]):
for file in tqdm(files):
if file.endswith('.wav'):
filename = os.path.join(root, file)
samples, samplingrate = loadAudio(filename)
lmfcc, mspec = lab1_proto.mfcc(samples)
wordTrans = list(path2info(filename)[2])
phoneTrans = words2phones(wordTrans, prondict.prondict)
targets = forcedAlignment(lmfcc, phoneHMMs, phoneTrans)
data.append({
'filename': filename,
'lmfcc': lmfcc,
'mspec': mspec,
'targets': targets
})
if file_name == 'traindata.npz':
np.savez(PATH + file_name, traindata=data)
elif file_name == 'testdata.npz':
np.savez(PATH + file_name, testdata=data)
traindata = np.load('Data/traindata.npz', allow_pickle=True)['traindata']
women_tracks = list()
men_tracks = list()
for tr in traindata:
if 'woman' in tr['filename']:
women_tracks.append(tr)
elif 'man' in tr['filename']:
men_tracks.append(tr)
men_tracks = np.array(men_tracks)
women_tracks = np.array(women_tracks)
val_size = int(len(traindata) * 0.1) # Percentage of validation data
men_pos = np.random.choice(len(men_tracks),
int(val_size / 2),
replace=False) # Randomly get men samples
women_pos = np.random.choice(len(women_tracks),
int(val_size / 2),
replace=False) # Randomly get women samples
men_val = men_tracks[men_pos] # Get validation men
women_val = women_tracks[women_pos] # Get validation women
men_tracks = np.delete(men_tracks,
men_pos) # Delete validation men from training set
women_tracks = np.delete(
women_tracks, women_pos) # Delete validation women from training set
# Get training, validation and testing
traindata = np.concatenate((men_tracks, women_tracks))
valdata = np.concatenate((men_val, women_val))
testdata = np.load('Data/testdata.npz', allow_pickle=True)['testdata']
dynamic = True
feature = 'lmfcc'
if dynamic:
if not os.path.isfile(PATH + 'dynxtraindata_' + feature + '.npz') or not \
os.path.isfile(PATH + 'dynytraindata_' + feature + '.npz'):
x, y = dynamic_features(traindata,
feature_name=feature,
dynamic=dynamic)
np.savez(PATH + 'dynxtraindata_' + feature + '.npz', traindata=x)
np.savez(PATH + 'dynytraindata_' + feature + '.npz', traindata=y)
x_train = np.load(PATH + 'dynxtraindata_' + feature + '.npz',
allow_pickle=True)['traindata']
y_train = np.load(PATH + 'dynytraindata_' + feature + '.npz',
allow_pickle=True)['traindata']
if not os.path.isfile(PATH + 'dynxvaldata_' + feature + '.npz') or not \
os.path.isfile(PATH + 'dynyvaldata_' + feature + '.npz'):
x, y = dynamic_features(valdata,
feature_name=feature,
dynamic=dynamic)
np.savez(PATH + 'dynxvaldata_' + feature + '.npz', valdata=x)
np.savez(PATH + 'dynyvaldata_' + feature + '.npz', valdata=y)
x_val = np.load(PATH + 'dynxvaldata_' + feature + '.npz',
allow_pickle=True)['valdata']
y_val = np.load(PATH + 'dynyvaldata_' + feature + '.npz',
allow_pickle=True)['valdata']
if not os.path.isfile(PATH + 'dynxtestdata_' + feature + '.npz') or not \
os.path.isfile(PATH + 'dynytestdata_' + feature + '.npz'):
x, y = dynamic_features(testdata,
feature_name=feature,
dynamic=dynamic)
np.savez(PATH + 'dynxtestdata_' + feature + '.npz', testdata=x)
np.savez(PATH + 'dynytestdata_' + feature + '.npz', testdata=y)
x_test = np.load(PATH + 'dynxtestdata_' + feature + '.npz',
allow_pickle=True)['testdata']
y_test = np.load(PATH + 'dynytestdata_' + feature + '.npz',
allow_pickle=True)['testdata']
else:
if not os.path.isfile(PATH + 'xtraindata_' + feature + '.npz') or not \
os.path.isfile(PATH + 'ytraindata_' + feature + '.npz'):
x, y = dynamic_features(traindata,
feature_name=feature,
dynamic=dynamic)
np.savez(PATH + 'xtraindata_' + feature + '.npz', traindata=x)
np.savez(PATH + 'ytraindata_' + feature + '.npz', traindata=y)
x_train = np.load(PATH + 'xtraindata_' + feature + '.npz',
allow_pickle=True)['traindata']
y_train = np.load(PATH + 'ytraindata_' + feature + '.npz',
allow_pickle=True)['traindata']
if not os.path.isfile(PATH + 'xvaldata_' + feature + '.npz') or not \
os.path.isfile(PATH + 'yvaldata_' + feature + '.npz'):
x, y = dynamic_features(valdata,
feature_name=feature,
dynamic=dynamic)
np.savez(PATH + 'xvaldata_' + feature + '.npz', valdata=x)
np.savez(PATH + 'yvaldata_' + feature + '.npz', valdata=y)
x_val = np.load(PATH + 'xvaldata_' + feature + '.npz',
allow_pickle=True)['valdata']
y_val = np.load(PATH + 'yvaldata_' + feature + '.npz',
allow_pickle=True)['valdata']
if not os.path.isfile(PATH + 'xtestdata_' + feature + '.npz') or not \
os.path.isfile(PATH + 'ytestdata_' + feature + '.npz'):
x, y = dynamic_features(testdata,
feature_name=feature,
dynamic=dynamic)
np.savez(PATH + 'xtestdata_' + feature + '.npz', testdata=x)
np.savez(PATH + 'ytestdata_' + feature + '.npz', testdata=y)
x_test = np.load(PATH + 'xtestdata_' + feature + '.npz',
allow_pickle=True)['testdata']
y_test = np.load(PATH + 'ytestdata_' + feature + '.npz',
allow_pickle=True)['testdata']
# Flatten into matrix
x_train = flatten_data(x_train)
x_val = flatten_data(x_val)
x_test = flatten_data(x_test)
y_train = flatten_targets(y_train)
y_val = flatten_targets(y_val)
y_test = flatten_targets(y_test)
# Normalize data
scaler = StandardScaler()
scaler.fit(x_train)
x_train = scaler.transform(x_train).astype('float32')
x_val = scaler.transform(x_val).astype('float32')
x_test = scaler.transform(x_test).astype('float32')
output_dim = len(state_list)
y_train = target_to_index(y_train, state_list)
y_val = target_to_index(y_val, state_list)
y_test = target_to_index(y_test, state_list)
y_train = np_utils.to_categorical(y_train, output_dim)
y_val = np_utils.to_categorical(y_val, output_dim)
y_test = np_utils.to_categorical(y_test, output_dim)
# Model train
if dynamic:
if not os.path.isfile('model_' + feature + 'dynamic' + '.h5'):
classifier = model.classifier(x_train[0].shape, output_dim)
classifier.compile(loss='categorical_crossentropy',
optimizer='sgd',
metrics=['accuracy'])
classifier.fit(x_train,
y_train,
validation_data=(x_val, y_val),
batch_size=256,
epochs=100)
classifier.save('model_' + feature + 'dynamic' + '.h5')
else:
classifier = tf.keras.models.load_model('model_' + feature +
'dynamic' + '.h5')
else:
if not os.path.isfile('model_' + feature + '.h5'):
classifier = model.classifier(x_train[0].shape, output_dim)
classifier.compile(loss='categorical_crossentropy',
optimizer='sgd',
metrics=['accuracy'])
classifier.fit(x_train,
y_train,
validation_data=(x_val, y_val),
batch_size=256,
epochs=100)
classifier.save('model_' + feature + '.h5')
else:
classifier = tf.keras.models.load_model('model_' + feature + '.h5')
group = True
merge = True
y_pred = classifier.predict(x_test, batch_size=256)
if group:
if merge:
y_pred = np.argmax(y_pred, axis=1)
y_test = np.argmax(y_test, axis=1)
y_pred = np.array([state_list[item] for item in y_pred])
y_test = np.array([state_list[item] for item in y_test])
y_pred = group_phonem(y_pred)
y_test = group_phonem(y_test)
y_pred = target_to_index(
y_pred, sorted(set(group_phonem(np.array(state_list)))))
y_test = target_to_index(
y_test, sorted(set(group_phonem(np.array(state_list)))))
y_pred = np_utils.to_categorical(
y_pred, len(set(group_phonem(np.array(state_list)))))
y_test = np_utils.to_categorical(
y_test, len(set(group_phonem(np.array(state_list)))))
likelihood, y_test = merge_states(y_pred, y_test)
y_pred = np.argmax(likelihood, axis=1)
y_true = np.argmax(y_test, axis=1)
phone_error_rate = compute_per(y_pred, y_true, likelihood)
print("Phone Error Rate: " + str(phone_error_rate))
else:
y_pred = np.argmax(y_pred, axis=1)
y_true = np.argmax(y_test, axis=1)
y_pred = np.array([state_list[item] for item in y_pred])
y_true = np.array([state_list[item] for item in y_true])
y_pred = group_phonem(y_pred)
y_true = group_phonem(y_true)
confusion_mtx = confusion_matrix(y_true, y_pred)
plot_confusion_matrix(confusion_mtx,
'Phoneme',
classes=sorted(
set(group_phonem(np.array(state_list)))))
print("Total accuracy: " +
str(np.sum(y_true == y_pred) / y_true.shape[0]))
else:
if merge:
likelihood, y_test = merge_states(y_pred, y_test)
y_pred = np.argmax(likelihood, axis=1)
y_true = np.argmax(y_test, axis=1)
phone_error_rate = compute_per(y_pred, y_true, likelihood)
print("Phone Error Rate: " + str(phone_error_rate))
else:
y_pred = np.argmax(y_pred, axis=1)
y_true = np.argmax(y_test, axis=1)
confusion_mtx = confusion_matrix(y_true, y_pred)
plot_confusion_matrix(confusion_mtx, 'State', classes=state_list)
print("Total accuracy: " +
str(np.sum(y_true == y_pred) / y_true.shape[0]))
def main():
args = parse_args()
torch.manual_seed(args.manual_seed)
dataroot = join(os.getcwd(), args.data_dir)
transform = transforms.Compose([
# transforms.RandomCrop(args.img_size, padding=None, pad_if_needed=True, fill=0, padding_mode='edge'),
transforms.Resize((args.img_size, args.img_size)),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
])
src_dataset = dataset_public(
root=dataroot,
transform=transform,
train=True,
domains=args.source,
)
src_loader = torch.utils.data.DataLoader(
dataset=src_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
pin_memory=True,
drop_last=False,
)
tar_dataset = dataset_public(
root=dataroot,
transform=transform,
train=True,
domains=args.target,
)
tar_loader = torch.utils.data.DataLoader(
dataset=tar_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
pin_memory=True,
drop_last=False,
)
# models
src_CNN = resnet_345(args).cuda()
tar_CNN = resnet_345(args).cuda()
cls = classifier().cuda()
checkpoint_path = join(os.getcwd(), args.model_dir, args.model_name)
load_checkpoint(checkpoint_path, src_CNN)
load_checkpoint(checkpoint_path, tar_CNN)
load_checkpoint(checkpoint_path, cls)
fixed_pretrained(src_CNN)
fixed_pretrained(cls)
# optimizer
if args.SGD:
optimizer = optim.SGD(tar_CNN.parameters(), lr=args.lr, momentum=0.9)
elif args.Adam:
optimizer = optim.Adam(tar_CNN.parameters(), lr=args.lr)
min_len = min(len(src_loader), len(tar_loader))
best_acc = 0
stop_count = 0
err_log_path = join(os.getcwd(), 'models',
'adda_{}_err.txt'.format(args.target[0]))
err_log = open(err_log_path, 'w')
for epoch in range(args.start_epoch, args.max_epochs + 1):
src_CNN.eval()
tar_CNN.train()
cls.eval()
print('\nEpoch = {}'.format(epoch))
err_log.write('Epoch = {}, '.format(epoch))
losses, train_acc = AverageMeter(), AverageMeter()
train_pbar = tqdm(total=min_len, ncols=100, leave=True)
for i, (src_data, tar_data) in enumerate(zip(src_loader, tar_loader)):
src_imgs, _ = src_data
tar_imgs, tar_labels = tar_data
# src_imgs, src_labels = src_data
src_imgs, tar_imgs, tar_labels = src_imgs.cuda(), tar_imgs.cuda(
), tar_labels.cuda()
# src_imgs, tar_imgs, tar_labels, src_labels = src_imgs.cuda(), tar_imgs.cuda(), tar_labels.cuda(), src_labels.cuda()
tar_CNN.zero_grad()
src_feature = src_CNN(src_imgs)
tar_feature = tar_CNN(tar_imgs)
loss = F.mse_loss(src_feature, tar_feature, reduction='mean')
loss.backward()
class_output = cls(tar_feature)
pred = class_output.max(1, keepdim=True)[1]
correct = pred.eq(tar_labels.view_as(pred)).sum().item()
optimizer.step()
losses.update(loss.data.item(), args.batch_size)
train_acc.update(correct, args.batch_size)
train_pbar.update()
train_pbar.set_postfix({
'loss': '{:.4f}'.format(losses.avg),
'acc': '{:.4f}'.format(train_acc.acc),
})
train_pbar.close()
if train_acc.acc > best_acc:
best_acc = train_acc.acc
stop_count = 0
checkpoint_path = join(os.getcwd(), 'models',
'adda_{}.pth'.format(args.target[0]))
save_checkpoint(checkpoint_path, tar_CNN, cls, optimizer)
else:
stop_count += 1
err_log.write('Loss: {:.4f}, Accuracy: {:.4f}\n'.format(
losses.avg, train_acc.acc))
err_log.flush()
if stop_count == args.early_stopping: break
err_log.write('Best test_acc: {:.4f}\n'.format(best_acc))
err_log.close()
with tf.variable_scope('GAN'):
model = WGAN(h, w, c, z_dim=z_dim, gf_dim=64, df_dim=64)
g_loss, d_loss, X_gen_sb, img_diff_cost_sb, X_fake_test_sb, y1_fake_test_sb = model.loss(X_real_sb=names_records['X'],
batch_size=batch_size,
X_ph = X_ph,
z_ph = z_ph,
improved_wgan=True,
weight_regularize=True)
h_p, w_p = patch_size
with tf.variable_scope('Classifier'):
y_train_sb, y_test_sb = classifier(X_patch_ph, X_gen_patch_ph, h_p, w_p)
cls_train_cost_sb = tg.cost.entropy(y_train_sb, y_patch_ph)
cls_valid_cost_sb = tg.cost.entropy(y_test_sb, y_patch_ph)
# sess.run(init_op)
optimizer = tf.train.AdamOptimizer(cls_learning_rate)
update_ops = ops.get_collection(ops.GraphKeys.UPDATE_OPS, 'Classifier')
with ops.control_dependencies(update_ops):
train_op = optimizer.minimize(cls_train_cost_sb)
'--checkpoint',
type=str,
required=True,
help="model checkpoint")
args = parser.parse_args()
# Dataloader
test_data = MyDataset('./dataset/metadata/test.csv',
'test',
transform=test_transform)
test_loader = DataLoader(test_data,
batch_size=32,
shuffle=False,
num_workers=8)
clf = classifier()
clf.load_state_dict(torch.load(args.checkpoint))
clf = clf.cuda()
clf.eval()
f = open('test.csv', 'w')
preds = []
for i, (batch_input, target) in enumerate(test_loader):
batch_input = batch_input.cuda()
target = target.cuda()
output, _ = clf(batch_input)
pred = torch.max(output, 1)[1]
preds.extend(pred.tolist())