def __init__(self, batch_size, iterations):
self.batch_size = batch_size
self.iter = iterations
self.len = self.batch_size * self.iter + self.batch_size
dh = MNISTDataHandler("../DIRNet-Keras/MNIST_data", is_train=True)
batch_x, batch_y = dh.sample_pair(self.batch_size)
self.xy_batch = np.concatenate([batch_x, batch_y], axis=1)
for i in range(self.iter):
batch_x, batch_y = dh.sample_pair(self.batch_size)
self.xy_batch = np.concatenate(
[self.xy_batch,
np.concatenate([batch_x, batch_y], axis=1)],
axis=0)
print("length-->", self.xy_batch.shape[0])
def main():
sess = tf.Session()
config = get_config(is_train=False)
mkdir(config.result_dir)
reg = DIRNet(sess, config, "DIRNet", is_train=False)
reg.restore(config.ckpt_dir)
dh = MNISTDataHandler("MNIST_data", is_train=False)
for i in range(1): #10
result_i_dir = config.result_dir + "/{}".format(i)
mkdir(result_i_dir)
batch_x, batch_y = dh.sample_pair(config.batch_size, i)
reg.deploy(result_i_dir, batch_x, batch_y)
def main():
sess = tf.Session()
config = get_config(is_train=True)
mkdir(config.tmp_dir)
mkdir(config.ckpt_dir)
q = 100
reg = DIRNet(sess, config, "DIRNet", is_train=True)
dh = MNISTDataHandler("MNIST_data", is_train=True)
for i in range(config.iteration):
batch_x, batch_y = dh.sample_pair(config.batch_size)
loss = reg.fit(batch_x, batch_y)
print("iter {:>6d} : {}".format(i + 1, loss))
if (i + 1) % 1000 == 0:
reg.deploy(config.tmp_dir, batch_x, batch_y)
reg.save(config.ckpt_dir)
from utils import get_logger
logger = get_logger()
model = DIRNetDeform(logger, device='cpu')
resultDir = 'imgs'
os.makedirs(resultDir, exist_ok=True)
# set your checkpoint here
checkpointFileName = 'dirConvnet_model-zncc_SSD_unet-2_gpu_deformConv_affine.ckpt'
# checkpointFileName = 'dirConvnet_model-zncc_unet_deformConv_affine-mnist_gpu.ckpt'
#
checkpoint = torch.load(checkpointFileName, map_location={'cuda:0': 'cpu'})
model.load_state_dict(checkpoint['model_state_dict'])
shape = [28, 28]
dh = MNISTDataHandler("../DIRNet-Keras/MNIST_data", is_train=False)
# dh = FashionMNISTDataHandler("../fashionMnist", is_train=False)
Iwarped_mean = 0
targetImage_mean = 0
sourceImage_mean = 0
length = 25
for i in range(length):
batch_x, batch_y = dh.sample_pair(1)
input2model = torch.FloatTensor(
np.concatenate([batch_x, batch_y], axis=1))
model.eval()
Iwarped = model(input2model)
def main():
config = get_config(is_train=True)
mkdir(config.tmp_dir)
mkdir(config.ckpt_dir)
model = DIRNet(config)
transform = tf.Compose([tf.Resize([16, 16]), tf.ToTensor()])
train_loader = DataLoader(MNIST('mnist',
train=True,
download=True,
transform=transform),
batch_size=train_batch)
test_loader = DataLoader(MNIST('mnist',
train=False,
download=True,
transform=transform),
batch_size=test_batch)
for batch, (data, label) in enumerate(train_loader):
if batch == 0:
num_images = 16000
# num_images = 300
digit_0 = data.index_select(0,
label.eq(0).nonzero().squeeze())[:3000]
digit_1 = data.index_select(0,
label.eq(1).nonzero().squeeze())[:3000]
digit_2 = data.index_select(0,
label.eq(2).nonzero().squeeze())[:3000]
digit_3 = data.index_select(0,
label.eq(3).nonzero().squeeze())[:3000]
digit_4 = data.index_select(0,
label.eq(4).nonzero().squeeze())[:3000]
digit_5 = data.index_select(0,
label.eq(5).nonzero().squeeze())[:3000]
digit_6 = data.index_select(0,
label.eq(6).nonzero().squeeze())[:3000]
digit_7 = data.index_select(0,
label.eq(7).nonzero().squeeze())[:3000]
digit_8 = data.index_select(0,
label.eq(8).nonzero().squeeze())[:3000]
digit_9 = data.index_select(0,
label.eq(9).nonzero().squeeze())[:3000]
digit = torch.stack([
digit_0, digit_1, digit_2, digit_3, digit_4, digit_5, digit_6, digit_7,
digit_8, digit_9
],
dim=0)
for batch, (data, label) in enumerate(test_loader):
if batch == 0:
num_images = 16000
# num_images = 300
digit_0_t = data.index_select(
0,
label.eq(0).nonzero().squeeze())[:500]
digit_1_t = data.index_select(
0,
label.eq(1).nonzero().squeeze())[:500]
digit_2_t = data.index_select(
0,
label.eq(2).nonzero().squeeze())[:500]
digit_3_t = data.index_select(
0,
label.eq(3).nonzero().squeeze())[:500]
digit_4_t = data.index_select(
0,
label.eq(4).nonzero().squeeze())[:500]
digit_5_t = data.index_select(
0,
label.eq(5).nonzero().squeeze())[:500]
digit_6_t = data.index_select(
0,
label.eq(6).nonzero().squeeze())[:500]
digit_7_t = data.index_select(
0,
label.eq(7).nonzero().squeeze())[:500]
digit_8_t = data.index_select(
0,
label.eq(8).nonzero().squeeze())[:500]
digit_9_t = data.index_select(
0,
label.eq(9).nonzero().squeeze())[:500]
digit_t = torch.stack([
digit_0_t, digit_1_t, digit_2_t, digit_3_t, digit_4_t, digit_5_t,
digit_6_t, digit_7_t, digit_8_t, digit_9_t
],
dim=0)
optim = torch.optim.Adam(model.parameters(), lr=config.lr)
scheduler = StepLR(optim, step_size=200, gamma=0.5)
train_pr = MNISTDataHandler(digit)
test_pr = MNISTDataHandler(digit_t)
total_loss = 0
for i in range(config.iteration):
batch_x, batch_y = train_pr.sample_pair(config.batch_size)
optim.zero_grad()
_, loss = model(batch_x, batch_y)
loss.backward()
optim.step()
scheduler.step()
total_loss += loss
if (i + 1) % 100 == 0:
print("iter {:>6d} : {}".format(i + 1, total_loss))
total_loss = 0
batch_x, batch_y = test_pr.sample_pair(config.batch_size)
model.deploy(config.tmp_dir, batch_x, batch_y)
ana = BaseStructure()
#ana.addBranchEntropy([0,1,2,3,4])
#ana.addBranchEntropy([0,1,2,3,4])
#ana.addBranchEntropy([0,1,2,3,4])
#ana.addBranchEntropy([0,1,2,3,4])
#ana.addBranchEntropy([0,1,2,4,4])
ana.agCizdir()
return 0
if __name__ == "__main__":
dh = MNISTDataHandler("MNIST_data", is_train=True)
main()
# TODO
# 1. encode image forinput
# 2. build tree from input
# 3. display tree
# 4. clasify
# 5. deeper thinking
# 6.