def decoder_body(time, old_state, output_ta_t, attention_tracker):
if feedback:
def from_previous():
prev_1 = tf.matmul(old_state, W_out) + b_out
return tf.gather(embeddings, tf.argmax(prev_1, 1))
x_t = tf.cond(tf.greater(time, 0), from_previous,
lambda: input_ta.read(0))
else:
x_t = input_ta.read(time)
# attention
part2 = tf.matmul(old_state, W_a) + b_a
part2 = tf.expand_dims(part2, 1)
john = part1 + part2
e = tf.reduce_sum(v_a * tf.tanh(john), [2])
alpha = tf.nn.softmax(e)
alpha = tf.to_float(mask(attention_lengths)) * alpha
alpha = alpha / tf.reduce_sum(alpha, [1], keep_dims=True)
attention_tracker = attention_tracker.write(time, alpha)
c = tf.reduce_sum(
tf.expand_dims(alpha, 2) * tf.squeeze(hidden), [1])
# GRU
con = tf.concat(1, [x_t, old_state, c])
z = tf.sigmoid(tf.matmul(con, W_z) + b_z)
r = tf.sigmoid(tf.matmul(con, W_r) + b_r)
con = tf.concat(1, [x_t, r * old_state, c])
h = tf.tanh(tf.matmul(con, W_h) + b_h)
new_state = (1 - z) * h + z * old_state
output_ta_t = output_ta_t.write(time, new_state)
return (time + 1, new_state, output_ta_t, attention_tracker)
def decoder_body(time, old_state, output_ta_t, attention_tracker):
if feedback:
def from_previous():
prev_1 = tf.matmul(old_state, W_out) + b_out
return tf.gather(embeddings, tf.argmax(prev_1, 1))
x_t = tf.cond(tf.greater(time, 0), from_previous, lambda: input_ta.read(0))
else:
x_t = input_ta.read(time)
# attention
part2 = tf.matmul(old_state, W_a) + b_a
part2 = tf.expand_dims(part2, 1)
john = part1 + part2
e = tf.reduce_sum(v_a * tf.tanh(john), [2])
alpha = tf.nn.softmax(e)
alpha = tf.to_float(mask(attention_lengths)) * alpha
alpha = alpha / tf.reduce_sum(alpha, [1], keep_dims=True)
attention_tracker = attention_tracker.write(time, alpha)
c = tf.reduce_sum(tf.expand_dims(alpha, 2) * tf.squeeze(hidden), [1])
# GRU
con = tf.concat(1, [x_t, old_state, c])
z = tf.sigmoid(tf.matmul(con, W_z) + b_z)
r = tf.sigmoid(tf.matmul(con, W_r) + b_r)
con = tf.concat(1, [x_t, r*old_state, c])
h = tf.tanh(tf.matmul(con, W_h) + b_h)
new_state = (1-z)*h + z*old_state
output_ta_t = output_ta_t.write(time, new_state)
return (time + 1, new_state, output_ta_t, attention_tracker)
def decoder_body(time, old_state, output_ta_t, attention_tracker):
if feedback:
def from_previous():
prev_1 = tf.matmul(old_state, W_out) + b_out
return tf.gather(embeddings, tf.argmax(prev_1, 1))
x_t = tf.cond(tf.greater(time, 0), from_previous, lambda: input_ta.read(0))
else:
x_t = input_ta.read(time)
# attention
part1_ex = tf.expand_dims(old_state, 1)
#part1_ex = tf.Print(part1_ex, [tf.shape(part1_ex)], summarize=1000)
part1_tiled = tf.tile(part1_ex, tf.pack([1, attn_len, 1]))
#part1_tiled = tf.Print(part1_tiled, [tf.shape(part1_tiled)], summarize=1000)
dot_input = tf.concat(2, [part1_tiled, attention_input])
#dot_input = tf.Print(dot_input, [tf.shape(dot_input)], summarize=1000)
dot_reshape = tf.reshape(dot_input, tf.pack([-1, attn_len, 1, attention_dims*2]))
#dot_reshape = tf.Print(dot_reshape, [tf.shape(dot_reshape)], summarize=1000)
john = tf.nn.conv2d(dot_reshape, U_a, [1, 1, 1, 1], "SAME")
#john = tf.Print(john, [tf.shape(john)], summarize=1000)
john = tf.squeeze(john, [2]) # squeeze out the third dimension
#john = tf.Print(john, [tf.shape(john)], summarize=1000)
a = v_a * tf.tanh(john)
#a = tf.Print(a, [tf.shape(a)], summarize=1000)
e = tf.reduce_sum(a, [2])
#e = tf.Print(e, [tf.shape(e)], summarize=1000)
alpha = tf.nn.softmax(e)
alpha = tf.to_float(mask(attention_lengths)) * alpha
alpha = alpha / tf.reduce_sum(alpha, [1], keep_dims=True)
#alpha = tf.Print(alpha, [tf.shape(alpha)], summarize=1000)
attention_tracker = attention_tracker.write(time, alpha)
c = tf.reduce_sum(tf.expand_dims(alpha, 2) * attention_input, [1])
# GRU
con = tf.concat(1, [x_t, old_state, c])
z = tf.sigmoid(tf.matmul(con, W_z) + b_z)
r = tf.sigmoid(tf.matmul(con, W_r) + b_r)
con = tf.concat(1, [x_t, r*old_state, c])
h = tf.tanh(tf.matmul(con, W_h) + b_h)
new_state = (1-z)*h + z*old_state
output_ta_t = output_ta_t.write(time, new_state)
return (time + 1, new_state, output_ta_t, attention_tracker)
