Ktorch: PyTorch Backend for Keras

Still WIP

---

Internal topology library

Imperative

from ktorch import *
import numpy as np

a = Variable(np.zeros((2, 3, 4)))
b = Variable(np.ones((3, 4)))
c = a + 0.2 + b * 0.3
print c
'''
<ktorch.graph.tensor.Tensor object at 0x0000000003E82DA0>
'''
print c.value
'''
[[[ 0.5  0.5  0.5  0.5]
  [ 0.5  0.5  0.5  0.5]
  [ 0.5  0.5  0.5  0.5]]

 [[ 0.5  0.5  0.5  0.5]
  [ 0.5  0.5  0.5  0.5]
  [ 0.5  0.5  0.5  0.5]]]
'''

Symbolic

from ktorch import *
import numpy as np

a = Tensor()
b = Tensor()
c = a + 0.2 + b * 0.3
f = Function([a, b], c)

x = np.zeros((2, 3, 4))
y = np.ones((3, 4))

print f([x, y])[0]  # Function returns a list

'''
[[[ 0.5  0.5  0.5  0.5]
  [ 0.5  0.5  0.5  0.5]
  [ 0.5  0.5  0.5  0.5]]

 [[ 0.5  0.5  0.5  0.5]
  [ 0.5  0.5  0.5  0.5]
  [ 0.5  0.5  0.5  0.5]]]
'''

Note that evaluation is greedy. The value of a tensor is computed the instant all the information required to compute it is available. The value will be cached in the .value attribute of the tensor. You can explicitly set the value for an input tensor using the .set_value() method, and all the tensors in the graph depending on that input will be updated in real time.

from ktorch import *
import numpy as np

a = Tensor()
b = Tensor()
c = Tensor()
d = a + b * c
print d.value
'''
AttributeError: 'Tensor' object has no attribute 'value'
'''
#Obviously, because we haven't set values for a, b and c
a.set_value(5)
b.set_value(3)
c.set_value(2)
print d.value
'''
11
'''
'''
Change the value for any of the inputs, and value of d will be automatically updated:
'''
c.set_value(4)
print d.value
'''
17
'''

Working with Keras

• Use ktorch branch of my fork

git clone http://www.github.com/farizrahman4u/keras.git
cd keras
git checkout ktorch
python setup.py install

• Make sure import keras prints Using Torch backend

• This allows using if statements, loops etc in custom layers (order of imports is important):

from keras.layers import *
from ktorch import dynamic_graph
dynamic_graph.initialize(globals())