Save and Load

In this notebook, we will

train a LeNet and save it
load the model from the file
test the loaded model



In [1]:

    
from tfs.models import LeNet
from tfs.dataset import Mnist
net = LeNet()
dataset = Mnist()

modify the network:
- L1 regularizer
- SGD optimizer



In [2]:

    
from tfs.core.optimizer import GradientDecentOptimizer
from tfs.core.regularizers import L1

net.optimizer = GradientDecentOptimizer(net)
net.regularizer = L1(net,l1=0.001)

net.build()









    Out[2]:





<tf.Tensor 'prob:0' shape=(?, 10) dtype=float32>



In [3]:

    
net.fit(dataset,batch_size=200,n_epoch=1,max_step=100)









    



step 10. loss 16.249886, score:0.403800
step 20. loss 15.966318, score:0.514600
step 30. loss 15.660644, score:0.597500
step 40. loss 15.390107, score:0.671600
step 50. loss 15.192878, score:0.712800
step 60. loss 15.279219, score:0.747200
step 70. loss 15.208706, score:0.776600
step 80. loss 15.325734, score:0.794400
step 90. loss 15.101788, score:0.813800
step 100. loss 15.025336, score:0.827500






    Out[3]:





<tfs.models.lenet.LeNet at 0x103f28c90>



In [4]:

    
net.save('lenet_epoch_1')



In [5]:

    
!ls ./









    



0.Train-LeNet.ipynb           4.Custom-Initializer.ipynb
1.Save-and-load.ipynb         lenet_epoch_1.model
2.Visualize-LeNet.ipynb       lenet_epoch_1.modeldef
3.Define-Custom-Network.ipynb

load the model



In [6]:

    
from tfs.network import Network
net2 = Network()
net2.load('lenet_epoch_1')



In [7]:

    
print net2









    



Name:conv1     	Type:Conv2d(knum=20,ksize=[5, 5],strides=[1, 1],padding=VALID,activation=None)
Name:pool1     	Type:MaxPool(ksize=[2, 2],strides=[2, 2])
Name:conv2     	Type:Conv2d(knum=50,ksize=[5, 5],strides=[1, 1],padding=VALID,activation=relu)
Name:pool2     	Type:MaxPool(ksize=[2, 2],strides=[2, 2])
Name:ip1       	Type:FullyConnect(outdim=500,activation=relu)
Name:ip2       	Type:FullyConnect(outdim=10,activation=None)
Name:prob      	Type:Softmax()



In [8]:

    
print net2.optimizer









    



GradientDecentOptimizer
-----param-----
learning_rate=0.001,print_names=['learning_rate']
---------------



In [9]:

    
print net2.initializer









    



DefaultInitializer
-----param-----
print_names=[]
-----nodes-----
conv1
    conv1/weights:0     xavier(seed=None,uniform=True,mode=FAN_AVG,factor=1.0)
    conv1/biases:0      constant(val=0.1)
pool1
conv2
    conv2/biases:0      constant(val=0.1)
    conv2/weights:0     xavier(seed=None,uniform=True,mode=FAN_AVG,factor=1.0)
pool2
ip1
    ip1/weights:0       xavier(seed=None,uniform=True,mode=FAN_AVG,factor=1.0)
    ip1/biases:0        constant(val=0.1)
ip2
    ip2/biases:0        constant(val=0.1)
    ip2/weights:0       xavier(seed=None,uniform=True,mode=FAN_AVG,factor=1.0)
prob



In [10]:

    
print net2.losser









    



CrossEntropyByLogitLabel (ip2)
-----param-----
----------------



In [11]:

    
print 'accuracy',net2.score(dataset.test)









    



accuracy 0.8275

fine-tune the loaded model



In [12]:

    
net2.fit(dataset,batch_size=200,n_epoch=1,max_step=100)









    



step 10. loss 14.938343, score:0.841300
step 20. loss 14.967062, score:0.850700
step 30. loss 14.971493, score:0.854300
step 40. loss 14.915509, score:0.864300
step 50. loss 14.754118, score:0.864100
step 60. loss 14.821177, score:0.871000
step 70. loss 14.832127, score:0.882600
step 80. loss 15.034270, score:0.875900
step 90. loss 14.842802, score:0.886700
step 100. loss 14.856237, score:0.889400






    Out[12]:





<tfs.network.base.Network at 0x1187b7590>



In [13]:

    
net2.score(dataset.test)









    Out[13]:





0.88939999999999997



In [ ]: