Computes the GRU cell back-propagation for 1 time step.
tf.raw_ops.GRUBlockCellGrad(
x, h_prev, w_ru, w_c, b_ru, b_c, r, u, c, d_h, name=None
)
Args x: Input to the GRU cell. h_prev: State input from the previous GRU cell. w_ru: Weight matrix for the reset and update gate. w_c: Weight matrix for the cell connection gate. b_ru: Bias vector for the reset and update gate. b_c: Bias vector for the cell connection gate. r: Output of the reset gate. u: Output of the update gate. c: Output of the cell connection gate. d_h: Gradients of the h_new wrt to objective function.
Returns d_x: Gradients of the x wrt to objective function. d_h_prev: Gradients of the h wrt to objective function. d_c_bar Gradients of the c_bar wrt to objective function. d_r_bar_u_bar Gradients of the r_bar & u_bar wrt to objective function.
This kernel op implements the following mathematical equations:
Note on notation of the variables:
Concatenation of a and b is represented by a_b Element-wise dot product of a and b is represented by ab Element-wise dot product is represented by \circ Matrix multiplication is represented by *
Additional notes for clarity:
w_ru
can be segmented into 4 different matrices.
w_ru = [w_r_x w_u_x
w_r_h_prev w_u_h_prev]
Similarly, w_c
can be segmented into 2 different matrices.
w_c = [w_c_x w_c_h_prevr]
Same goes for biases.
b_ru = [b_ru_x b_ru_h]
b_c = [b_c_x b_c_h]
Another note on notation:
d_x = d_x_component_1 + d_x_component_2
where d_x_component_1 = d_r_bar * w_r_x^T + d_u_bar * w_r_x^T
and d_x_component_2 = d_c_bar * w_c_x^T
d_h_prev = d_h_prev_component_1 + d_h_prevr \circ r + d_h \circ u
where d_h_prev_componenet_1 = d_r_bar * w_r_h_prev^T + d_u_bar * w_r_h_prev^T
Mathematics behind the Gradients below:
d_c_bar = d_h \circ (1-u) \circ (1-c \circ c)
d_u_bar = d_h \circ (h-c) \circ u \circ (1-u)
d_r_bar_u_bar = [d_r_bar d_u_bar]
[d_x_component_1 d_h_prev_component_1] = d_r_bar_u_bar * w_ru^T
[d_x_component_2 d_h_prevr] = d_c_bar * w_c^T
d_x = d_x_component_1 + d_x_component_2
d_h_prev = d_h_prev_component_1 + d_h_prevr \circ r + u
Below calculation is performed in the python wrapper for the Gradients (not in the gradient kernel.)
d_w_ru = x_h_prevr^T * d_c_bar
d_w_c = x_h_prev^T * d_r_bar_u_bar
d_b_ru = sum of d_r_bar_u_bar along axis = 0
d_b_c = sum of d_c_bar along axis = 0