def forward(self):
N = self.num_output
K = np.prod(self.bottom.shape[1:])
M = self.bottom.shape[0]
sgemm(False, True, 1.0, self.bottom, 0, K, self.weights, 0, K, 0.0,
self.top, 0, N, M, N, K)
sgemm(False, False, 1.0, self.bias_multiplier, 0, 1, self.bias, 0, N,
1.0, self.top, 0, N, M, N, 1)
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):
bottom = symbol_table[sources[0].name]
weights = symbol_table[sources[1].name]
bias = symbol_table[sources[2].name]
top = symbol_table[sinks[0].name]
evt = sgemm(False, True, 1.0, bottom, 0, K, weights,
0, K, 0.0, top, 0, N, M, N, K, wait_for=wait_for)
evt = sgemm(False, False, 1.0, bias_multiplier, 0,
1, bias, 0, N, 1.0, top, 0, N, M, N,
1, wait_for=evt)
return [evt]
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):
bottom = symbol_table[sources[0].name]
weights = symbol_table[sources[1].name]
bias = symbol_table[sources[2].name]
top = symbol_table[sinks[0].name]
evt = sgemm(False,
True,
1.0,
bottom,
0,
K,
weights,
0,
K,
0.0,
top,
0,
N,
M,
N,
K,
wait_for=wait_for)
evt = sgemm(False,
False,
1.0,
bias_multiplier,
0,
1,
bias,
0,
N,
1.0,
top,
0,
N,
M,
N,
1,
wait_for=evt)
return [evt]
def launch(self, symbol_table, wait_for):
bottom = symbol_table[sources[0].name]
bot_offset = np.prod(bottom.shape[1:])
weights = symbol_table[sources[1].name]
bias = symbol_table[sources[2].name]
top = symbol_table[sinks[0].name]
top_offset = np.prod(top.shape[1:])
m = weights.shape[0]
n = np.prod(top.shape[2:])
k = np.prod(weights.shape[1:])
# cl.clFinish(queues[0])
evts = []
if is_1x1:
for i in range(bottom.shape[0]):
evt = sgemm(False, False, 1.0, weights, 0, k,
bottom, i * bot_offset, n, 0.0,
top, i * top_offset, n, m, n,
k, queues[i % len(queues)], wait_for=wait_for)
evt = sgemm(False, False, 1.0, bias, 0, 1,
bias_multiplier, 0, n, 1.0, top, i *
top_offset, n, m, n, 1, queues[i % len(queues)], wait_for=evt)
evts.append(evt)
else:
for i in range(bottom.shape[0]):
evt = im2col(bottom.ocl_buf,
col_datas[i % len(queues)].ocl_buf,
i * bot_offset
).on(queues[i % len(queues)], (padded, ),
wait_for=wait_for)
evt = sgemm(False, False, 1.0, weights, 0, k,
col_datas[i % len(queues)],
0, n, 0.0, top, i * top_offset, n, m, n,
k, queues[i % len(queues)], wait_for=evt)
evt = sgemm(False, False, 1.0, bias, 0, 1,
bias_multiplier, 0, n, 1.0, top, i *
top_offset, n, m, n, 1, queues[i % len(queues)], wait_for=evt)
evts.append(evt)
return evts
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 forward(data):
global fc6, fc7, fc8
conv1 = ConvForward(data, conv1_filters, conv1_bias, kernel_size=(11, 11), padding=(0, 0), stride=(4, 4))
conv1 = ReluForward(conv1)
norm1, norm1_scale = LrnForward(conv1, alpha=alpha, beta=beta, local_size=local_size, k=1)
pool1, pool1_mask = PoolForward(norm1, kernel_size=(3, 3), padding=(0, 0), stride=(2, 2))
conv2 = ConvForward(pool1, conv2_filters, conv2_bias, kernel_size=(5, 5), padding=(2, 2), stride=(1, 1))
conv2 = ReluForward(conv2)
norm2, norm2_scale = LrnForward(conv2, alpha=alpha, beta=beta, local_size=local_size, k=1)
pool2, pool2_mask = PoolForward(norm2, kernel_size=(3, 3), padding=(0, 0), stride=(2, 2))
conv3 = ConvForward(pool2, conv3_filters, conv3_bias, kernel_size=(3, 3), padding=(1, 1), stride=(1, 1))
conv3 = ReluForward(conv3)
conv4 = ConvForward(conv3, conv4_filters, conv4_bias, kernel_size=(3, 3), padding=(1, 1), stride=(1, 1))
conv4 = ReluForward(conv4)
conv5 = ConvForward(conv4, conv5_filters, conv5_bias, kernel_size=(3, 3), padding=(1, 1), stride=(1, 1))
conv5 = ReluForward(conv5)
pool5, pool5_mask = PoolForward(conv5, kernel_size=(3, 3), padding=(0, 0), stride=(2, 2))
N = fc6.shape[1]
K = np.prod(pool5.shape[1:])
M = pool5.shape[0]
sgemm(False, True, 1.0, pool5, 0, K, fc6_filters, 0, K, 0.0, fc6, 0, N, M, N, K)
sgemm(False, False, 1.0, fc6_bias_multiplier, 0, 1, fc6_bias, 0, N, 1.0, fc6, 0, N, M, N, 1)
fc6 = ReluForward(fc6)
N = fc7.shape[1]
K = np.prod(fc6.shape[1:])
M = fc6.shape[0]
sgemm(False, True, 1.0, fc6, 0, K, fc7_filters, 0, K, 0.0, fc7, 0, N, M, N, K)
sgemm(False, False, 1.0, fc7_bias_multiplier, 0, 1, fc7_bias, 0, N, 1.0, fc7, 0, N, M, N, 1)
fc7 = ReluForward(fc7)
N = fc8.shape[1]
K = np.prod(fc7.shape[1:])
M = fc7.shape[0]
sgemm(False, True, 1.0, fc7, 0, K, fc8_filters, 0, K, 0.0, fc8, 0, N, M, N, K)
sgemm(False, False, 1.0, fc8_bias_multiplier, 0, 1, fc8_bias, 0, N, 1.0, fc8, 0, N, M, N, 1)
prob = SoftmaxForward(fc8)
return prob
def forward(data):
global fc6, fc7, fc8
conv1 = ConvForward(data,
conv1_filters,
conv1_bias,
kernel_size=(11, 11),
padding=(0, 0),
stride=(4, 4))
conv1 = ReluForward(conv1)
norm1, norm1_scale = LrnForward(conv1,
alpha=alpha,
beta=beta,
local_size=local_size,
k=1)
pool1, pool1_mask = PoolForward(norm1,
kernel_size=(3, 3),
padding=(0, 0),
stride=(2, 2))
conv2 = ConvForward(pool1,
conv2_filters,
conv2_bias,
kernel_size=(5, 5),
padding=(2, 2),
stride=(1, 1))
conv2 = ReluForward(conv2)
norm2, norm2_scale = LrnForward(conv2,
alpha=alpha,
beta=beta,
local_size=local_size,
k=1)
pool2, pool2_mask = PoolForward(norm2,
kernel_size=(3, 3),
padding=(0, 0),
stride=(2, 2))
conv3 = ConvForward(pool2,
conv3_filters,
conv3_bias,
kernel_size=(3, 3),
padding=(1, 1),
stride=(1, 1))
conv3 = ReluForward(conv3)
conv4 = ConvForward(conv3,
conv4_filters,
conv4_bias,
kernel_size=(3, 3),
padding=(1, 1),
stride=(1, 1))
conv4 = ReluForward(conv4)
conv5 = ConvForward(conv4,
conv5_filters,
conv5_bias,
kernel_size=(3, 3),
padding=(1, 1),
stride=(1, 1))
conv5 = ReluForward(conv5)
pool5, pool5_mask = PoolForward(conv5,
kernel_size=(3, 3),
padding=(0, 0),
stride=(2, 2))
N = fc6.shape[1]
K = np.prod(pool5.shape[1:])
M = pool5.shape[0]
sgemm(False, True, 1.0, pool5, 0, K, fc6_filters, 0, K, 0.0, fc6, 0, N, M,
N, K)
sgemm(False, False, 1.0, fc6_bias_multiplier, 0, 1, fc6_bias, 0, N, 1.0,
fc6, 0, N, M, N, 1)
fc6 = ReluForward(fc6)
N = fc7.shape[1]
K = np.prod(fc6.shape[1:])
M = fc6.shape[0]
sgemm(False, True, 1.0, fc6, 0, K, fc7_filters, 0, K, 0.0, fc7, 0, N, M, N,
K)
sgemm(False, False, 1.0, fc7_bias_multiplier, 0, 1, fc7_bias, 0, N, 1.0,
fc7, 0, N, M, N, 1)
fc7 = ReluForward(fc7)
N = fc8.shape[1]
K = np.prod(fc7.shape[1:])
M = fc7.shape[0]
sgemm(False, True, 1.0, fc7, 0, K, fc8_filters, 0, K, 0.0, fc8, 0, N, M, N,
K)
sgemm(False, False, 1.0, fc8_bias_multiplier, 0, 1, fc8_bias, 0, N, 1.0,
fc8, 0, N, M, N, 1)
prob = SoftmaxForward(fc8)
return prob
def launch(self, symbol_table, wait_for):
bottom = symbol_table[sources[0].name]
bot_offset = np.prod(bottom.shape[1:])
weights = symbol_table[sources[1].name]
bias = symbol_table[sources[2].name]
top = symbol_table[sinks[0].name]
top_offset = np.prod(top.shape[1:])
m = weights.shape[0]
n = np.prod(top.shape[2:])
k = np.prod(weights.shape[1:])
# cl.clFinish(queues[0])
evts = []
if is_1x1:
for i in range(bottom.shape[0]):
evt = sgemm(False,
False,
1.0,
weights,
0,
k,
bottom,
i * bot_offset,
n,
0.0,
top,
i * top_offset,
n,
m,
n,
k,
queues[i % len(queues)],
wait_for=wait_for)
evt = sgemm(False,
False,
1.0,
bias,
0,
1,
bias_multiplier,
0,
n,
1.0,
top,
i * top_offset,
n,
m,
n,
1,
queues[i % len(queues)],
wait_for=evt)
evts.append(evt)
else:
for i in range(bottom.shape[0]):
evt = im2col(bottom.ocl_buf,
col_datas[i % len(queues)].ocl_buf,
i * bot_offset).on(
queues[i % len(queues)],
(padded, ),
wait_for=wait_for)
evt = sgemm(False,
False,
1.0,
weights,
0,
k,
col_datas[i % len(queues)],
0,
n,
0.0,
top,
i * top_offset,
n,
m,
n,
k,
queues[i % len(queues)],
wait_for=evt)
evt = sgemm(False,
False,
1.0,
bias,
0,
1,
bias_multiplier,
0,
n,
1.0,
top,
i * top_offset,
n,
m,
n,
1,
queues[i % len(queues)],
wait_for=evt)
evts.append(evt)
return evts
