def backward(self):
N = self.num_output
K = np.prod(self.bottom.shape[1:])
M = self.bottom.shape[0]
sgemm(True, False, 1.0, self.top_diff, 0, N, self.bottom, 0, K, 0.0,
self.weights_diff, 0, K, N, K, M)
sgemv(True, M, N, 1.0, self.top_diff, 0, N, self.bias_multiplier, 0, 1,
0.0, self.bias_diff, 0, 1)
sgemm(False, False, 1.0, self.top_diff, 0, N, self.weights, 0,
K, 0.0, self.bottom_diff, 0, K, M, K, N)
def launch(self, symbol_table, wait_for):
queue = queues[0]
bottom = symbol_table[sources[0]]
bot_offset = np.prod(bottom.shape[1:])
top_diff = symbol_table[sources[1]]
top_offset = np.prod(top_diff.shape[1:])
weights = symbol_table[sources[2]]
bottom_diff = symbol_table[sinks[0]]
bottom_diff.fill(0)
bottom_diff.sync_ocl()
weights_diff = symbol_table[sinks[1]]
weights_diff.fill(0)
weights_diff.sync_ocl()
bias_diff = symbol_table[sinks[2]]
bias_diff.fill(0)
bias_diff.sync_ocl()
for i in range(bottom.shape[0]):
n = np.prod(top_diff.shape[2:])
sgemv(False, top_diff.shape[1],
n, 1.0, top_diff, i *
top_offset, n, bias_multiplier, 0, 1, 1.0,
bias_diff, 0, 1)
im2col(bottom.ocl_buf, col_data.ocl_buf, i
* bot_offset).on(queue, im2col_global_size)
m = top_diff.shape[1]
n = col_data.shape[0]
k = col_data.shape[1]
sgemm(False, True, 1.0, top_diff, i *
top_offset, k, col_data, 0, k, 1.0,
weights_diff, 0, n, m, n, k)
m = weights.shape[1]
n = col_data.shape[1]
k = weights.shape[0]
sgemm(True, False, 1.0, weights, 0, m,
top_diff, i * top_offset, n, 0.0,
col_data, 0, n,
m, n, k)
col2im(col_data.ocl_buf,
bottom_diff.ocl_buf, i *
bot_offset).on(queue, col2im_global_size)
def launch(self, symbol_table, wait_for):
queue = queues[0]
bottom = symbol_table[sources[0]]
bot_offset = np.prod(bottom.shape[1:])
top_diff = symbol_table[sources[1]]
top_offset = np.prod(top_diff.shape[1:])
weights = symbol_table[sources[2]]
bottom_diff = symbol_table[sinks[0]]
bottom_diff.fill(0)
bottom_diff.sync_ocl()
weights_diff = symbol_table[sinks[1]]
weights_diff.fill(0)
weights_diff.sync_ocl()
bias_diff = symbol_table[sinks[2]]
bias_diff.fill(0)
bias_diff.sync_ocl()
for i in range(bottom.shape[0]):
n = np.prod(top_diff.shape[2:])
sgemv(False, top_diff.shape[1], n, 1.0, top_diff,
i * top_offset, n, bias_multiplier, 0, 1, 1.0,
bias_diff, 0, 1)
im2col(bottom.ocl_buf, col_data.ocl_buf,
i * bot_offset).on(queue, im2col_global_size)
m = top_diff.shape[1]
n = col_data.shape[0]
k = col_data.shape[1]
sgemm(False, True, 1.0, top_diff, i * top_offset, k,
col_data, 0, k, 1.0, weights_diff, 0, n, m, n, k)
m = weights.shape[1]
n = col_data.shape[1]
k = weights.shape[0]
sgemm(True, False, 1.0, weights, 0, m, top_diff,
i * top_offset, n, 0.0, col_data, 0, n, m, n, k)
col2im(col_data.ocl_buf, bottom_diff.ocl_buf,
i * bot_offset).on(queue, col2im_global_size)
