def main():
train_list = make_datapath_list(phase="train")
val_list = make_datapath_list(phase="val")
# Dataset
train_dataset = MyDataset(train_list,
transform=ImageTransform(resize, mean, std),
phase="train")
val_dataset = MyDataset(val_list,
transform=ImageTransform(resize, mean, std),
phase="val")
train_dataloader = torch.utils.data.DataLoader(train_dataset,
batch_size=4,
shuffle=True)
val_dataloader = torch.utils.data.DataLoader(val_dataset,
batch_size=4,
shuffle=False)
dataloader_dict = {"train": train_dataloader, "val": val_dataloader}
# Network
use_pretrained = "true"
net = models.vgg16(pretrained=use_pretrained)
net.classifier[6] = nn.Linear(in_features=4096, out_features=2)
# Loss
criterior = nn.CrossEntropyLoss()
# Optimizer
# param_to_update = []
# update_param_name = ["classifier.6.weight", "classifier.6.bias"]
# for name, param in net.named_parameters():
# if name in update_param_name:
# param.requires_grad = True
# param_to_update.append(param)
# print(name)
# else:
# param.requires_grad = False
params1, params2, params3 = param_to_update(net)
optimizer = optim.SGD([
{
'params': params1,
'lr': 1e-4
},
{
'params': params2,
'lr': 5e-4
},
{
'params': params3,
'lr': 1e-3
},
],
momentum=0.9)
train_model(net, dataloader_dict, criterior, optimizer, num_epochs)
def main():
train_list = make_datapath_list("train")
val_list = make_datapath_list("val")
# dataset
train_dataset = MyDataset(train_list,
transform=ImageTransform(resize, mean, std),
phase="train")
val_dataset = MyDataset(val_list,
transform=ImageTransform(resize, mean, std),
phase="val")
# dataloader
train_dataloader = torch.utils.data.DataLoader(train_dataset,
batch_size,
shuffle=True)
val_dataloader = torch.utils.data.DataLoader(val_dataset,
batch_size,
shuffle=False)
dataloader_dict = {"train": train_dataloader, "val": val_dataloader}
# network
use_pretrained = True
net = models.vgg16(pretrained=use_pretrained)
net.classifier[6] = nn.Linear(in_features=4096, out_features=2)
# loss
criterior = nn.CrossEntropyLoss()
# optimizer
params1, params2, params3 = params_to_update(net)
optimizer = optim.SGD([
{
'params': params1,
'lr': 1e-4
},
{
'params': params2,
'lr': 5e-4
},
{
'params': params3,
'lr': 1e-3
},
],
momentum=0.9)
# training
train_model(net, dataloader_dict, criterior, optimizer, num_epochs)
def create_menu_bar(self, root):
self.file_menu.add_command(label = "Open audio file", command = self.open_audio_file)
self.menu_bar.add_cascade(label="File", menu=self.file_menu)
self.ds_menu.add_command(label = "Generate graphics", command = lambda: spectogram.create_data_set_graphs())
self.ds_menu.add_command(label = "Graphics augmentation", command = lambda: ImageTransform.gen_dataset_augmens())
self.menu_bar.add_cascade(label = "Data-Set", menu=self.ds_menu)
self.nn_menu.add_command(label = "Train", command = lambda: NeuralNetwork.create_and_train_nn())
self.nn_menu.add_command(label = "Load last model weights", command=lambda : NeuralNetwork.load_model_weights())
self.menu_bar.add_cascade(label = "Neural Network", menu=self.nn_menu)
def plotstft(audiopath, generatefig=True, binsize=2**10, plotpath=None, colormap="jet"): #colormap="jet"
samplerate, samples = wav.read(audiopath)
s = stft(samples, binsize)
audio_path_split = audiopath.split('/')
if len(audio_path_split)-1 > 0:
plotpath = audio_path_split[0] + "/" + audio_path_split[1] + "/graphs/" + audio_path_split[2] # dodaj folder graphs u putanju za cuvanje grafika
plotpath = plotpath.replace('.wav', '.png') # zameni ekstenziju fajla na .png
plotpath = audiopath.replace('.wav', '.png')
sshow, freq = logscale_spec(s, factor=80.0, sr=samplerate)
ims = 20.*np.log10(np.abs(sshow)/10e-6) # amplitude to decibel
timebins, freqbins = np.shape(ims)
fig = plt.figure(figsize=(8, 4.25))
plt.imshow(np.transpose(ims), origin="lower", aspect="auto", cmap=colormap, interpolation="none")
plt.colorbar()
plt.xlabel("Time [s]")
plt.ylabel("Frequency dB[Hz]")
plt.xlim([0, timebins-1])
plt.ylim([0, freqbins])
xlocs = np.float32(np.linspace(0, timebins-1, 10))
plt.xticks(xlocs, ["%.02f" % l for l in ((xlocs*len(samples)/timebins)+(0.5*binsize))/samplerate])
ylocs = np.int16(np.round(np.linspace(0, freqbins-1, 20)))
plt.yticks(ylocs, ["%.02f" % freq[i] for i in ylocs])
#plt.clf()
fig.canvas.draw() # bitno!!! formira model grafika tj samu matricu grafika, ali je ne prikazuje korisniku!
if not(generatefig):
plt.show()
""" -temp- deo samo za prikaz sta ce ici u obucavanje mreze... posle obrisati.. """
# odlicno radi...
img_data = ImageTransform.fig2data(fig)
img_data = ImageTransform.transform(img_data)
plt.imshow(img_data, 'gray')
plt.figure()
img_data = ImageTransform.image_bin(img_data)
img_data = ImageTransform.invert(img_data)
img_data = ImageTransform.remove_noise(img_data) # zatvaranje 1.dilate 2.erode
img_data = ImageTransform.resize_graph(img_data, 70, 33) #org 350x165, 350%5=70, 165%5=33, odrzane proporcije
cv2.imwrite("test.png", img_data)
plt.imshow(img_data, 'gray')
plt.show()
else:
img_data = prepare_fig_to_img(fig) #za formiranje grafika u data-set-u ... TODO: napraviti zasebnu fun..
cv2.imwrite(plotpath, img_data)
#plt.close(fig) # sprecava memory leak - curenje memorije
return fig # vrati matlabov plot obj(numpy array)
def create_menu_bar(self, root):
self.file_menu.add_command(label="Open audio file",
command=self.open_audio_file)
self.menu_bar.add_cascade(label="File", menu=self.file_menu)
self.ds_menu.add_command(
label="Generate graphics",
command=lambda: spectogram.create_data_set_graphs())
self.ds_menu.add_command(
label="Graphics augmentation",
command=lambda: ImageTransform.gen_dataset_augmens())
self.menu_bar.add_cascade(label="Data-Set", menu=self.ds_menu)
self.nn_menu.add_command(
label="Train", command=lambda: NeuralNetwork.create_and_train_nn())
self.nn_menu.add_command(
label="Load last model weights",
command=lambda: NeuralNetwork.load_model_weights())
self.menu_bar.add_cascade(label="Neural Network", menu=self.nn_menu)
def main():
# 画像の読み込みと表示
img_path = './data/9497/resize-070327.jpg'
img = Image.open(img_path)
plt.imshow(img)
plt.show()
size = (224, 224)
mean = (0.485, 0.456, 0.406)
std = (0.229, 0.224, 0.225)
transform = ImageTransform(size)
img_transform = transform(img, phase='train')
print(img_transform.shape)
img_transformed = img_transform.numpy().transpose((1, 2, 0))
img_transformed = np.clip(img_transformed, 0, 1)
plt.imshow(img_transformed)
plt.show()
def predict(img):
# prepare network
use_pretrained = True
net = models.vgg16(pretrained=use_pretrained)
net.classifier[6] = nn.Linear(in_features=4096, out_features=2)
net.eval()
# prepare model
model = load_model(net, save_path)
# prepare input img
transform = ImageTransform(resize, mean, std)
img = transform(img, phase="test")
img = img.unsqueeze_(
0) # (chan, height, width) -> (1, chan, height, width)
# predict
output = model(img)
response = predictor.predict_max(output)
return response
def prepare_fig_to_img(graph_fig):
"""
@brief
Ulaz: matlabov grafik objekat
Matlabova figura postaje slika, nad slikom se vrsi
1. crop-ovanje
2. grayscale
3. binarizacija
4. uklanjanje suma
5. resize
Izlaz: slika spremna za obucavanje mreze (numpy matrica)
"""
img_data = ImageTransform.fig2data(graph_fig)
img_data = ImageTransform.transform(img_data)
img_data = ImageTransform.image_bin(img_data)
img_data = ImageTransform.invert(img_data)
img_data = ImageTransform.remove_noise(img_data) # zatvaranje 1.dilate 2.erode
img_data = ImageTransform.resize_graph(img_data, 70, 33) #org 350x165, 350%5=70, 165%5=33, odrzane proporcije
return img_data
def prepare_fig_to_img(graph_fig):
"""
@brief
Ulaz: matlabov grafik objekat
Matlabova figura postaje slika, nad slikom se vrsi
1. crop-ovanje
2. grayscale
3. binarizacija
4. uklanjanje suma
5. resize
Izlaz: slika spremna za obucavanje mreze (numpy matrica)
"""
img_data = ImageTransform.fig2data(graph_fig)
img_data = ImageTransform.transform(img_data)
img_data = ImageTransform.image_bin(img_data)
img_data = ImageTransform.invert(img_data)
img_data = ImageTransform.remove_noise(
img_data) # zatvaranje 1.dilate 2.erode
img_data = ImageTransform.resize_graph(
img_data, 70, 33) #org 350x165, 350%5=70, 165%5=33, odrzane proporcije
return img_data
img_mean = (0.4914, 0.4822, 0.4465)
img_std = (0.2023, 0.1994, 0.2010)
# Transforms
transform_train = [
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(img_mean, img_std),
]
transform_test = [
transforms.ToTensor(),
transforms.Normalize(img_mean,img_std),
]
t_transform_train = ImageTransform.transform(transform_train)
t_transform_test = ImageTransform.transform(transform_test)
# Dataset and DataLoader arguments
dataset_name = torchvision.datasets.CIFAR10
trainSet_dict = dict(root='./data', train=True, download=True, transform=t_transform_train)
trainLoad_dict = dict(batch_size=32, shuffle=True, num_workers=4)
testSet_dict = dict(root='./data', train=False, download=True, transform=t_transform_test)
testLoad_dict = dict(batch_size=32, shuffle=False, num_workers=4)
IMAGE_PATH = "images/"
MODEL_PATH = "model/"
def main():
# Device
SEED = 1
cuda = torch.cuda.is_available()
def plotstft(audiopath,
generatefig=True,
binsize=2**10,
plotpath=None,
colormap="jet"): #colormap="jet"
samplerate, samples = wav.read(audiopath)
s = stft(samples, binsize)
audio_path_split = audiopath.split('/')
if len(audio_path_split) - 1 > 0:
plotpath = audio_path_split[0] + "/" + audio_path_split[
1] + "/graphs/" + audio_path_split[
2] # dodaj folder graphs u putanju za cuvanje grafika
plotpath = plotpath.replace('.wav',
'.png') # zameni ekstenziju fajla na .png
plotpath = audiopath.replace('.wav', '.png')
sshow, freq = logscale_spec(s, factor=80.0, sr=samplerate)
ims = 20. * np.log10(np.abs(sshow) / 10e-6) # amplitude to decibel
timebins, freqbins = np.shape(ims)
fig = plt.figure(figsize=(8, 4.25))
plt.imshow(np.transpose(ims),
origin="lower",
aspect="auto",
cmap=colormap,
interpolation="none")
plt.colorbar()
plt.xlabel("Time [s]")
plt.ylabel("Frequency dB[Hz]")
plt.xlim([0, timebins - 1])
plt.ylim([0, freqbins])
xlocs = np.float32(np.linspace(0, timebins - 1, 10))
plt.xticks(xlocs, [
"%.02f" % l for l in ((xlocs * len(samples) / timebins) +
(0.5 * binsize)) / samplerate
])
ylocs = np.int16(np.round(np.linspace(0, freqbins - 1, 20)))
plt.yticks(ylocs, ["%.02f" % freq[i] for i in ylocs])
#plt.clf()
fig.canvas.draw(
) # bitno!!! formira model grafika tj samu matricu grafika, ali je ne prikazuje korisniku!
if not (generatefig):
plt.show()
""" -temp- deo samo za prikaz sta ce ici u obucavanje mreze... posle obrisati.. """
# odlicno radi...
img_data = ImageTransform.fig2data(fig)
img_data = ImageTransform.transform(img_data)
plt.imshow(img_data, 'gray')
plt.figure()
img_data = ImageTransform.image_bin(img_data)
img_data = ImageTransform.invert(img_data)
img_data = ImageTransform.remove_noise(
img_data) # zatvaranje 1.dilate 2.erode
img_data = ImageTransform.resize_graph(
img_data, 70,
33) #org 350x165, 350%5=70, 165%5=33, odrzane proporcije
cv2.imwrite("test.png", img_data)
plt.imshow(img_data, 'gray')
plt.show()
else:
img_data = prepare_fig_to_img(
fig
) #za formiranje grafika u data-set-u ... TODO: napraviti zasebnu fun..
cv2.imwrite(plotpath, img_data)
#plt.close(fig) # sprecava memory leak - curenje memorije
return fig # vrati matlabov plot obj(numpy array)
df_labels_idx = df_train.loc[df_train.duplicated(["labels", "labels_n"])==False]\
[["labels_n", "labels"]].set_index("labels_n").sort_index()
# Load Image
train_list = make_datapath_list(phase="train")
print(f"train data length : {len(train_list)}")
val_list = make_datapath_list(phase="val")
print(f"validation data length : {len(val_list)}")
test_list = make_datapath_list(phase="test")
print(f"test data length : {len(test_list)}")
# Create Dataset
train_dataset = PlantDataset(df_labels_idx,
df_train,
train_list,
transform=ImageTransform(size, mean, std),
phase='train')
val_dataset = PlantDataset(df_labels_idx,
df_train,
val_list,
transform=ImageTransform(size, mean, std),
phase='val')
test_dataset = PlantDataset(df_labels_idx,
df_train,
test_list,
transform=ImageTransform(size, mean, std),
phase='test')
index = 0
print("【train dataset】")
def main(cfg):
# 病理画像のデータのリストを取得
data_dir = make_data_path_list()
# データセットの作成
dataset = PathologicalImage(file_list=data_dir,
transform=ImageTransform(cfg.size),
num_pixels=cfg.num_pixels)
# 学習用データの枚数を取得
train_size = int(len(dataset) * cfg.rate)
# 検証用のデータの枚数を取得
val_size = len(dataset) - train_size
# データセットの分割
train_dataset, val_dataset = data.random_split(dataset,
[train_size, val_size])
# 動作確認
print("入力画像サイズ：" + str(train_dataset.__getitem__(0)[0].size()))
print("教師データサイズ：" + str(train_dataset.__getitem__(0)[1].shape))
# 学習用のDataLoaderを作成
train_dataloader = torch.utils.data.DataLoader(train_dataset,
batch_size=cfg.batch_size,
shuffle=True)
# 検証用のDataLoaderを作成
val_dataloader = torch.utils.data.DataLoader(val_dataset,
batch_size=cfg.batch_size,
shuffle=False)
# 動作確認
batch_iterator = iter(train_dataloader)
inputs, labels = next(batch_iterator)
log.info("-----Image and label shape of dataloader-----")
log.info("入力データ：" + str(inputs.size()))
log.info("入力ラベル：" + str(labels.shape))
# GPU初期設定
# ネットワークモデル(自作FCNs)をimport
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# ネットワークをGPUへ
net = FCNs()
net.to(device)
torch.backends.cudnn.benchmark = True
log.info("-----Constitution of network-----")
log.info(net)
# 損失関数の設定
criterion = nn.MSELoss()
# 最適化手法の設定
optimizer = optim.SGD(net.parameters(),
lr=cfg.SGD.lr,
momentum=cfg.SGD.momentum)
log.info("-----Details of optimizer function-----")
log.info(optimizer)
# 損失値を保持するリスト
train_loss = []
val_loss = []
# 学習
for epoch in range(cfg.num_epochs):
log.info("Epoch {} / {} ".format(epoch + 1, cfg.num_epochs))
log.info("----------")
# 学習
train_history = train_model(net, train_dataloader, criterion,
optimizer)
# 学習したlossのリストを作成
train_loss.append(train_history)
# 検証
val_history = val_model(net, val_dataloader, criterion)
# 検証したlossのリスト作成
val_loss.append(val_history)
# テストと出力値保存
test_history = test_model(net, val_dataloader, criterion)
# figインスタンスとaxインスタンスを作成
fig_loss, ax_loss = plt.subplots(figsize=(10, 10))
ax_loss.plot(range(1, cfg.num_epochs + 1, 1),
train_loss,
label="train_loss")
ax_loss.plot(range(1, cfg.num_epochs + 1, 1), val_loss, label="val_loss")
ax_loss.set_xlabel("epoch")
ax_loss.legend()
fig_loss.savefig("loss.png")
# パラメータの保存
save_path = './pathological.pth'
torch.save(net.state_dict(), save_path)
