├── FSMNCell.py
└── README.md


/FSMNCell.py:
--------------------------------------------------------------------------------
 1 | import tensorflow as tf
 2 | 
 3 | class DFSMN(object):
 4 |     def __init__(self, memory_size_left, memory_size_right, stride_l, stride_r,
 5 |                  input_size, output_size, dtype=tf.float32):
 6 |         self._memory_size_left = memory_size_left
 7 |         self._memory_size_right = memory_size_right
 8 |         self._memory_size = memory_size_left + memory_size_right + 1
 9 |         self._stride_l = stride_l
10 |         self._stride_r = stride_r
11 |         self._input_size = input_size
12 |         self._output_size = output_size
13 |         self._dtype = dtype
14 |         self._build_graph()
15 | 
16 |     def _build_graph(self):
17 |         self._W = tf.get_variable("dfsmnn_w", [self._input_size, self._output_size],
18 |                                    initializer=tf.truncated_normal_initializer(stddev=5e-2, dtype=self._dtype))
19 |         self._bias = tf.get_variable("dfsmnn_bias", [self._output_size],
20 |                                     initializer=tf.constant_initializer(0.0, dtype=self._dtype))
21 |         self._memory_weights =  tf.get_variable("memory_weights", [self._input_size, self._memory_size],
22 |                                                 initializer=tf.constant_initializer(1.0, dtype=self._dtype))
23 | 
24 |     def __call__(self, skip_con, input_data):
25 |         size = input_data.get_shape()
26 |         batch_size = size[0].value
27 |         num_steps = size[1].value
28 | 
29 |         def skip_connection_func(input_skip):
30 |             # Need to implement the transform
31 |             return input_skip
32 | 
33 |         # Construct memory matrix
34 |         memory_matrix = []
35 |         for step in range(num_steps):
36 |             left_num = tf.maximum(0, step - self._memory_size_left)
37 |             right_num = tf.maximum(0, num_steps - step - self._memory_size_right - 1)
38 |             weight_start = tf.minimum(-self._memory_size_left + step - 1, -1)
39 |             weight_middle = -self._memory_size_left-1
40 |             weight_end = tf.maximum(step - self._memory_size_left - num_steps, -self._memory_size)
41 |             mem_l = self._memory_weights[weight_start:weight_middle:-1]
42 |             mem_r = self._memory_weights[weight_middle-1:weight_end:-1]
43 |             left_num = left_num - self._stride_l*len(mem_l)
44 |             right_num = right_num - self._stride_r*len(mem_r)
45 |             ele_1, ele_2 = 1, 0
46 |             while ele_1 <= len(mem_l):
47 |                 for count1 in range(self._stride_l):
48 |                     mem_l.insert(ele_1, self._input_size*[0])
49 |                 ele_1 += self._stride_l
50 |             while ele_2 < len(mem_r):
51 |                 for count2 in range(self._stride_l):
52 |                     mem_r.insert(ele_2, self._input_size*[0])
53 |                 ele_2 += self._stride_r
54 | 
55 |             mem = mem_l + self._memory_weights[self._memory_size_right] + mem_r
56 |             mem = mem[tf.maximum(len(mem_l) - step, 0): tf.minimum(num_steps - step - 1 -len(mem_r), -1)]
57 | 
58 |             # strides padding
59 |             if left_num <= 0:
60 |                 left_num = 0
61 |             if right_num <=0:
62 |                 right_num = 0
63 |             d_batch = tf.pad(mem, [[left_num, right_num], [0, 0]])
64 |             memory_matrix.append([d_batch])
65 |         memory_matrix = tf.concat(0, memory_matrix)
66 | 
67 |         # Compute the layer output
68 |         h_hatt = tf.matmul([memory_matrix] * batch_size, input_data)
69 |         p_s = skip_connection_func(skip_con)
70 |         p_hatt = p_s + h_hatt
71 |         h = tf.matmul(p_hatt, [self._W] * batch_size) + self._bias
72 |         return h, p_hatt
73 | 


--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
1 | # DFSMN
2 | Tensorflow version of DFSMN core structure
3 | https://arxiv.org/abs/1803.05030
4 | 


--------------------------------------------------------------------------------