def saveImg(img_folder):
test_img_file = 'data/t10k-images.idx3-ubyte'
test_label_file = 'data/t10k-labels.idx1-ubyte'
test_img = loadImg(test_img_file)
test_label = loadLabel(test_label_file)
label = []
for i in range(10):
img = test_img[i][0] * 255
mpimg.imsave('%s/%d.png' % (img_folder, i), img, format='png')
label.append(test_label[i])
return label
right = 0
total_loss = 0
for i in range(len(y)):
network.img_placeholder.feed(X[i])
network.label_placeholder.feed(y[i])
loss, result = network.forword()
total_loss += loss
if y[i][result] == 1: right += 1
print 'Accuracy: %.2f' % (right * 1.0 / len(y))
print 'Avarage loss: %f' % (total_loss / len(y))
if __name__ == '__main__':
if len(sys.argv) < 2:
print 'model folder is needed'
exit()
else:
model_folder = sys.argv[1]
test_img_file = 'data/t10k-images.idx3-ubyte'
test_label_file = 'data/t10k-labels.idx1-ubyte'
config_file = '%s/config.txt' % model_folder
test_img = loadImg(test_img_file)
test_label = loadLabel(test_label_file)
network = NetworkConstructor(test_img[0].shape, test_label[0].shape)
network.initNetwork(config_file, model_folder)
test(network, test_img, test_label)
def main(datasetName, n_sample_per_class, run, encoderPath=None):
# 加载数据和标签
info = DatasetInfo.info[datasetName]
data_path = "./data/{}/{}.mat".format(datasetName, datasetName)
label_path = './trainTestSplit/{}/sample{}_run{}.mat'.format(
datasetName, n_sample_per_class, run)
isExists(data_path)
data = loadmat(data_path)[info['data_key']]
bands = data.shape[2]
isExists(label_path)
trainLabel, testLabel = loadLabel(label_path)
res = torch.zeros((3, EPOCHS))
# 数据转换
data, trainLabel, testLabel = data.astype(np.float32), trainLabel.astype(
np.int), testLabel.astype(np.int)
nc = int(np.max(trainLabel))
trainDataset = HSIDatasetV1(data, trainLabel, patchsz=42)
testDataset = HSIDatasetV1(data, testLabel, patchsz=42)
trainLoader = DataLoader(trainDataset,
batch_size=BATCHSZ,
shuffle=True,
num_workers=NUM_WORKERS)
testLoader = DataLoader(testDataset,
batch_size=128,
shuffle=True,
num_workers=NUM_WORKERS)
UNITS[0] = bands
model = SSDL(UNITS, nc)
model.apply(weight_init)
# 加载编码器的预训练参数
if encoderPath is not None:
isExists(encoderPath)
raw_dict = torch.load(encoderPath, map_location=torch.device('cpu'))
target_dict = model.encoder.state_dict()
keys = list(zip(raw_dict.keys(), target_dict.keys()))
for raw_key, target_key in keys:
target_dict[target_key] = raw_dict[raw_key]
model.encoder.load_state_dict(target_dict)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=LR)
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=10, gamma=0.8)
for epoch in range(EPOCHS):
print('*' * 5 + 'Epoch:{}'.format(epoch) + '*' * 5)
model, trainLoss = train(model,
criterion=criterion,
optimizer=optimizer,
dataLoader=trainLoader)
acc, evalLoss = test(model, criterion=criterion, dataLoader=testLoader)
print('epoch:{} trainLoss:{:.8f} evalLoss:{:.8f} acc:{:.4f}'.format(
epoch, trainLoss, evalLoss, acc))
print('*' * 18)
res[0][epoch], res[1][epoch], res[2][epoch] = trainLoss, evalLoss, acc
if epoch % 5 == 0:
torch.save(
model.state_dict(),
os.path.join(
ROOT, 'ssdl_sample{}_run{}_epoch{}.pkl'.format(
n_sample_per_class, run, epoch)))
scheduler.step()
tmp = res.numpy()
savemat(os.path.join(ROOT, 'res.mat'), {
'trainLoss': tmp[0],
'evalLoss': tmp[1],
'acc': tmp[2]
})
return res
network.label_placeholder.feed(y[i])
loss, result = network.forword()
if result == True: right += 1
print 'Accuracy: %.2f' % (right * 1.0 / len(y))
if __name__ == '__main__':
if len(sys.argv) < 2:
print 'model folder is needed'
exit()
else:
model_folder = sys.argv[1]
if len(sys.argv) > 2:
batch_size = int(sys.argv[2])
init_learning_rate = float(sys.argv[3])
max_epoch = int(sys.argv[4])
train_img_file = 'data/train-images.idx3-ubyte'
train_label_file = 'data/train-labels.idx1-ubyte'
test_img_file = 'data/t10k-images.idx3-ubyte'
test_label_file = 'data/t10k-labels.idx1-ubyte'
config_file = '%s/config.txt' % model_folder
train_img = loadImg(train_img_file)
train_label = loadLabel(train_label_file)
network = NetworkConstructor(train_img[0].shape, train_label[0].shape)
network.initNetwork(config_file, model_folder)
train(network, train_img, train_label, model_folder)
def main(datasetName, n_sample_per_class, run, preTrainedModel=None):
# 加载数据和标签
info = DatasetInfo.info[datasetName]
data_path = "./data/{}/{}.mat".format(datasetName, datasetName)
label_path = './trainTestSplit/{}/sample{}_run{}.mat'.format(
datasetName, n_sample_per_class, run)
isExists(data_path)
data = loadmat(data_path)[info['data_key']]
bands = data.shape[2]
isExists(label_path)
trainLabel, testLabel = loadLabel(label_path)
res = torch.zeros((3, EPOCHS))
# 数据转换
data, trainLabel, testLabel = data.astype(np.float32), trainLabel.astype(
np.int), testLabel.astype(np.int)
nc = int(np.max(trainLabel))
trainDataset = HSIDatasetV1(data, trainLabel, patchsz=5, n_components=4)
testDataset = HSIDatasetV1(data, testLabel, patchsz=5, n_components=4)
trainLoader = DataLoader(trainDataset,
batch_size=BATCHSZ,
shuffle=True,
num_workers=NUM_WORKERS)
testLoader = DataLoader(testDataset,
batch_size=256,
shuffle=True,
num_workers=NUM_WORKERS)
# 定义编码模型
encoder = Encoder(info['band'] + 5 * 5 * 4, info['hidden_size'])
# 读取预训练模型参数
assert isExists(preTrainedModel), '预训练模型路径不存在'
preTrainedModelDict = torch.load(preTrainedModel,
map_location=torch.device('cpu'))
# 移除b_decoder预训练参数
preTrainedModelDict.pop('b_decoder')
# 加载预训练参数
encoder.load_state_dict(preTrainedModelDict)
# 定义逻辑回归模型
classifier = nn.Linear(info['hidden_size'], nc)
# 整体
model = nn.Sequential(encoder, classifier)
# 损失函数，优化器，学习率下架管理器
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=LR)
scheduler = optim.lr_scheduler.StepLR(optimizer,
step_size=EPOCHS // 2,
gamma=0.1)
for epoch in range(EPOCHS):
print('*' * 5 + 'Epoch:{}'.format(epoch) + '*' * 5)
model, trainLoss = train(model,
criterion=criterion,
optimizer=optimizer,
dataLoader=trainLoader,
device=DEVICE)
acc, evalLoss = test(model,
criterion=criterion,
dataLoader=testLoader,
device=DEVICE)
print('epoch:{} trainLoss:{:.8f} evalLoss:{:.8f} acc:{:.4f}'.format(
epoch, trainLoss, evalLoss, acc))
print('*' * 18)
res[0][epoch], res[1][epoch], res[2][epoch] = trainLoss, evalLoss, acc
if epoch % 5 == 0:
torch.save(
model.state_dict(),
os.path.join(
ROOT, '{}_sample{}_run{}_epoch{}.pkl'.format(
MODEL_NAME, n_sample_per_class, run, epoch)))
scheduler.step()
tmp = res.numpy()
savemat(os.path.join(ROOT, 'res.mat'), {
'trainLoss': tmp[0],
'evalLoss': tmp[1],
'acc': tmp[2]
})
return res
def main(datasetName, n_sample_per_class, run, encoderPath=None):
# 加载数据和标签
info = DatasetInfo.info[datasetName]
data_path = "./data/{}/{}.mat".format(datasetName, datasetName)
label_path = './trainTestSplit/{}/sample{}_run{}.mat'.format(
datasetName, n_sample_per_class, run)
isExists(data_path)
data = loadmat(data_path)[info['data_key']]
bands = data.shape[2]
isExists(label_path)
trainLabel, testLabel = loadLabel(label_path)
# 数据转换
data, trainLabel, testLabel = data.astype(np.float32), trainLabel.astype(
np.int), testLabel.astype(np.int)
nc = int(np.max(trainLabel))
trainDataset = HSIDatasetV1(data, trainLabel, patchsz=42)
testDataset = HSIDatasetV1(data, testLabel, patchsz=42)
trainLoader = DataLoader(trainDataset,
batch_size=BATCHSZ,
shuffle=True,
num_workers=NUM_WORKERS)
testLoader = DataLoader(testDataset,
batch_size=BATCHSZ,
shuffle=True,
num_workers=NUM_WORKERS)
UNITS[0] = bands
model = SSDL(UNITS, nc)
model.apply(weight_init)
# 加载编码器的预训练参数
if encoderPath is not None:
isExists(encoderPath)
raw_dict = torch.load(encoderPath, map_location=torch.device('cpu'))
target_dict = model.encoder.state_dict()
keys = list(zip(raw_dict.keys(), target_dict.keys()))
for raw_key, target_key in keys:
target_dict[target_key] = raw_dict[raw_key]
model.encoder.load_state_dict(target_dict)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=LR, weight_decay=1e-2)
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=10, gamma=0.8)
for epoch in range(EPOCHS):
print('*' * 5 + 'Epoch:{}'.format(epoch) + '*' * 5)
model, trainLoss = train(model,
criterion=criterion,
optimizer=optimizer,
dataLoader=trainLoader)
acc, evalLoss = test(model, criterion=criterion, dataLoader=testLoader)
viz.line([[trainLoss, evalLoss]], [epoch],
win='train&test loss',
update='append')
viz.line([acc], [epoch], win='accuracy', update='append')
print('epoch:{} trainLoss:{:.8f} evalLoss:{:.8f} acc:{:.4f}'.format(
epoch, trainLoss, evalLoss, acc))
print('*' * 18)
# if os.path.isdir('./ssdl/{}/ssdl_sample{}_run{}_epoch{}'.format(datasetName, n_sample_per_class, run, epoch))
if not os.path.isdir('./ssdl/{}'.format(datasetName)):
os.makedirs('./ssdl/{}'.format(datasetName))
torch.save(
model.state_dict(),
'./ssdl/{}/ssdl_sample{}_run{}_epoch{}.pkl'.format(
datasetName, n_sample_per_class, run, epoch))
scheduler.step()