def MLP(order, cudnn_ws, mkl):
model = ModelHelper(name="benchmark")
d = 256
depth = 20
width = 3
for i in range(depth):
for j in range(width):
current = "fc_{}_{}".format(i, j) if i > 0 else "data"
next_ = "fc_{}_{}".format(i + 1, j)
brew.fc(model,
current,
next_,
dim_in=d,
dim_out=d,
weight_init=('XavierFill', {}),
bias_init=('XavierFill', {}))
brew.sum(model, ["fc_{}_{}".format(depth, j) for j in range(width)],
["sum"])
brew.fc(model,
"sum",
"last",
dim_in=d,
dim_out=1000,
weight_init=('XavierFill', {}),
bias_init=('XavierFill', {}))
xent = model.LabelCrossEntropy(["last", "label"], "xent")
if not mkl:
model.AveragedLoss(xent, "loss")
return model, d
def test_release_blobs_when_used(self):
m = model_helper.ModelHelper()
fc1 = brew.fc(m, "data", "x", dim_in=2, dim_out=2)
fc2 = brew.fc(m, fc1, "y", dim_in=2, dim_out=2)
fc3 = brew.fc(m, fc1, "z", dim_in=2, dim_out=2)
fc4 = brew.fc(m, fc2, "u", dim_in=2, dim_out=2)
m.net.Alias(["u"], ["u_alias"])
brew.sum(m, [fc3, fc4], "out")
with_frees = memonger.release_blobs_when_used(m.net.Proto(),
set("data"))
expect_frees = {"x", "y", "z"} # out is external output
# and u is aliased so cannot be freed
found_frees = set()
for op in with_frees.op:
if op.type == "Free":
self.assertFalse(op.input[0] in found_frees) # no double frees
found_frees.add(op.input[0])
else:
# Check a freed blob is not used anymore
for inp in op.input:
self.assertFalse(inp in found_frees)
for outp in op.output:
self.assertFalse(outp in found_frees)
self.assertEqual(expect_frees, found_frees)
def test_verify_graph_inequality(self, input_dim, output_dim, batch_size):
m = model_helper.ModelHelper()
m.Proto().type = "dag"
m.Proto().num_workers = 4
with core.NameScope("name_x"):
fc1 = brew.fc(m, "data", "x", dim_in=input_dim, dim_out=output_dim)
fc2 = brew.fc(m, fc1, "y", dim_in=output_dim, dim_out=output_dim)
fc3 = brew.fc(m, fc1, "z", dim_in=output_dim, dim_out=output_dim)
brew.sum(m, [fc2, fc3], "out")
m2 = model_helper.ModelHelper()
m2.Proto().type = "dag"
m2.Proto().num_workers = 4
with core.NameScope("name_x"):
fc1 = brew.fc(m2,
"data",
"x",
dim_in=input_dim,
dim_out=output_dim)
fc2 = brew.fc(m2, fc1, "y", dim_in=output_dim, dim_out=output_dim)
fc3 = brew.fc(m2, fc1, "y", dim_in=output_dim, dim_out=output_dim)
brew.sum(m2, [fc2, fc3], "out")
self.assertFalse(
memonger.verify_graph_equality(m.net.Proto(), m2.net.Proto()))
def test_release_blobs_when_used(self):
m = model_helper.ModelHelper()
fc1 = brew.fc(m, "data", "x", dim_in=2, dim_out=2)
fc2 = brew.fc(m, fc1, "y", dim_in=2, dim_out=2)
fc3 = brew.fc(m, fc1, "z", dim_in=2, dim_out=2)
fc4 = brew.fc(m, fc2, "u", dim_in=2, dim_out=2)
m.net.Alias(["u"], ["u_alias"])
brew.sum(m, [fc3, fc4], "out")
with_frees = memonger.release_blobs_when_used(m.net.Proto(), set("data"))
expect_frees = {"x", "y", "z"} # out is external output
# and u is aliased so cannot be freed
found_frees = set()
for op in with_frees.op:
if op.type == "Free":
self.assertFalse(op.input[0] in found_frees) # no double frees
found_frees.add(op.input[0])
else:
# Check a freed blob is not used anymore
for inp in op.input:
self.assertFalse(inp in found_frees)
for outp in op.output:
self.assertFalse(outp in found_frees)
self.assertEqual(expect_frees, found_frees)
def MLP(order, cudnn_ws, device):
model = ModelHelper(name="benchmark")
d = 256
depth = 20
width = 3
for i in range(depth):
for j in range(width):
current = "fc_{}_{}".format(i, j) if i > 0 else "data"
next_ = "fc_{}_{}".format(i + 1, j)
brew.fc(
model,
current, next_,
dim_in=d, dim_out=d,
weight_init=('XavierFill', {}),
bias_init=('XavierFill', {}))
brew.sum(model, ["fc_{}_{}".format(depth, j) for j in range(width)], ["sum"])
brew.fc(model, "sum", "last",
dim_in=d, dim_out=1000,
weight_init=('XavierFill', {}),
bias_init=('XavierFill', {}))
xent = model.LabelCrossEntropy(["last", "label"], "xent")
if device != 'MKL':
model.AveragedLoss(xent, "loss")
return model, d
def MLP(order, gpu_engine_ws):
model = model_helper.ModelHelper(name="MLP")
d = 256
depth = 20
width = 3
for i in range(depth):
for j in range(width):
current = "fc_{}_{}".format(i, j) if i > 0 else "data"
next_ = "fc_{}_{}".format(i + 1, j)
brew.fc(
model,
current,
next_,
dim_in=d,
dim_out=d,
weight_init=('XavierFill', {}),
bias_init=('XavierFill', {}),
)
brew.sum(model, ["fc_{}_{}".format(depth, j) for j in range(width)],
["sum"])
brew.fc(
model,
"sum",
"last",
dim_in=d,
dim_out=1000,
weight_init=('XavierFill', {}),
bias_init=('XavierFill', {}),
)
xent = model.net.LabelCrossEntropy(["last", "label"], "xent")
model.net.AveragedLoss(xent, "loss")
return model, d
def test_fast_memonger_unique_outputs(self):
m = model_helper.ModelHelper()
fc = []
for i in range(2):
z = brew.fc(m,
"data{}".format(i),
"fc".format(i),
dim_in=2,
dim_out=2)
fc.append(z)
r = []
# Trick is here to have same input appear twice in a same Sum
for x in fc:
for y in fc:
r.append(brew.sum(m, [x, y], 1))
concated = brew.concat(m, r, "concated")
brew.relu(m, concated, "merged")
static_blobs = \
[o for op in m.param_init_net.Proto().op for o in op.output] + \
["merged"] + ["data{}".format(i) for i in range(len(fc))]
optimized_net = memonger.optimize_inference_fast(
m.Proto(), static_blobs)
for op in optimized_net.op:
self.assertEqual(len(op.output), len(set(op.output)), str(op))
def cross_gated_global_pool(
self,
blob_in,
dim_in,
prefix='',
ratio=8,
reduced_dim=2,
res_gate=False,
):
gp_blob = self.model.ReduceBackMean(blob_in,
prefix + '_g_pool',
num_reduce_dim=reduced_dim)
fc1 = self.model.FC(gp_blob, prefix + '_fc1', dim_in, dim_in // ratio)
fc1_relu = self.model.Relu(fc1, prefix + '_fc1_relu')
fc2 = self.model.FC(fc1_relu, prefix + '_fc2', dim_in // ratio, dim_in)
sig = self.model.Sigmoid(fc2, fc2 + '_sig')
shortcut_blob = self.prev_blob
self.prev_blob = self.model.Mul([self.prev_blob, sig],
[prefix + 'g_pool_out'],
broadcast=1,
axis=0)
if res_gate:
self.prev_blob = brew.sum(self.model,
[shortcut_blob, self.prev_blob],
prefix + 'res_pool_out')
return self.prev_blob
def add_simple_block(
self,
input_filters,
num_filters,
down_sampling=False,
spatial_batch_norm=True
):
self.comp_idx = 0
shortcut_blob = self.prev_blob
# 3x3
self.add_conv(
input_filters,
num_filters,
kernel=3,
stride=(1 if down_sampling is False else 2),
pad=1
)
if spatial_batch_norm:
self.add_spatial_bn(num_filters)
self.add_relu()
last_conv = self.add_conv(num_filters, num_filters, kernel=3, pad=1)
if spatial_batch_norm:
last_conv = self.add_spatial_bn(num_filters)
# Increase of dimensions, need a projection for the shortcut
if (num_filters != input_filters):
shortcut_blob = brew.conv(
self.model,
shortcut_blob,
'shortcut_projection_%d' % self.comp_count,
input_filters,
num_filters,
weight_init=("MSRAFill", {}),
kernel=1,
stride=(1 if down_sampling is False else 2),
no_bias=self.no_bias,
)
if spatial_batch_norm:
shortcut_blob = brew.spatial_bn(
self.model,
shortcut_blob,
'shortcut_projection_%d_spatbn' % self.comp_count,
num_filters,
epsilon=1e-3,
is_test=self.is_test,
)
self.prev_blob = brew.sum(
self.model, [shortcut_blob, last_conv],
'comp_%d_sum_%d' % (self.comp_count, self.comp_idx)
)
self.comp_idx += 1
self.add_relu()
# Keep track of number of high level components if this ResNetBuilder
self.comp_count += 1
def add_simple_block(
self,
input_filters,
num_filters,
down_sampling=False,
spatial_batch_norm=True
):
self.comp_idx = 0
shortcut_blob = self.prev_blob
# 3x3
self.add_conv(
input_filters,
num_filters,
kernel=3,
stride=(1 if down_sampling is False else 2),
pad=1
)
if spatial_batch_norm:
self.add_spatial_bn(num_filters)
self.add_relu()
last_conv = self.add_conv(num_filters, num_filters, kernel=3, pad=1)
if spatial_batch_norm:
last_conv = self.add_spatial_bn(num_filters)
# Increase of dimensions, need a projection for the shortcut
if (num_filters != input_filters):
shortcut_blob = brew.conv(
self.model,
shortcut_blob,
'shortcut_projection_%d' % self.comp_count,
input_filters,
num_filters,
weight_init=("MSRAFill", {}),
kernel=1,
stride=(1 if down_sampling is False else 2),
no_bias=self.no_bias,
)
if spatial_batch_norm:
shortcut_blob = brew.spatial_bn(
self.model,
shortcut_blob,
'shortcut_projection_%d_spatbn' % self.comp_count,
num_filters,
epsilon=1e-3,
is_test=self.is_test,
)
self.prev_blob = brew.sum(
self.model, [shortcut_blob, last_conv],
'comp_%d_sum_%d' % (self.comp_count, self.comp_idx)
)
self.comp_idx += 1
self.add_relu()
# Keep track of number of high level components if this ResNetBuilder
self.comp_count += 1
def test_forward_optim_tree_dag_traversal(self):
input_dim = 4
output_dim = 4
batch_size = 4
m = model_helper.ModelHelper()
m.Proto().type = "dag"
m.Proto().num_workers = 4
with core.NameScope("name_x"):
fc1 = brew.fc(m,
"data",
"fc1",
dim_in=input_dim,
dim_out=output_dim)
fc2 = brew.fc(m, fc1, "fc2", dim_in=output_dim, dim_out=output_dim)
fc3 = brew.fc(m, fc2, "fc3", dim_in=output_dim, dim_out=output_dim)
fc4 = brew.fc(m, fc3, "fc4", dim_in=output_dim, dim_out=output_dim)
fc5 = brew.fc(m, fc4, "fc5", dim_in=output_dim, dim_out=output_dim)
# Branch
fc3b = brew.fc(m,
fc2,
"fc3b",
dim_in=output_dim,
dim_out=output_dim)
fc4b = brew.fc(m,
fc3b,
"fc4b",
dim_in=output_dim,
dim_out=output_dim)
fc5b = brew.fc(m,
fc4b,
"fc5b",
dim_in=output_dim,
dim_out=output_dim)
fc5sum = brew.sum(m, [fc5, fc5b], "fc5sum")
fc5.Relu([], fc5sum) \
.Softmax([], "pred1") \
.LabelCrossEntropy(["label"], ["xent1"]) \
.AveragedLoss([], "loss1")
fc6 = brew.fc(m, fc5, "fc6", dim_in=output_dim, dim_out=output_dim)
fc6.Relu([], fc6) \
.Softmax([], "pred2") \
.LabelCrossEntropy(["label"], ["xent2"]) \
.AveragedLoss([], "loss2")
blobs_before = count_blobs(m.net.Proto())
# adding name_x/fc5_w as heads (which belongs to non-root op)
# to make sure that dag traversal always starts from root ops
optim_proto = memonger.optimize_inference_for_dag(
m.net, ["name_x/fc5_w", "name_x/data"], "name_x")
blobs_after = count_blobs(optim_proto)
self.assertLess(blobs_after, blobs_before)
def add_residual_block(self, in_blob, out_dim):
self.conv_prelu(in_blob, out_dim)
self.prev_dim = out_dim
self.conv_prelu(self.prev_blob, out_dim)
self.prev_dim = out_dim
self.prev_blob = brew.sum(
self.model, [self.prev_blob, in_blob],
"res{}_{}".format(self.comp_idx, self.comp_count - 1))
return self.prev_blob
def test_verify_graph_inequality(self, input_dim, output_dim, batch_size):
m = model_helper.ModelHelper()
m.Proto().type = "dag"
m.Proto().num_workers = 4
with core.NameScope("name_x"):
fc1 = brew.fc(m, "data", "x", dim_in=input_dim, dim_out=output_dim)
fc2 = brew.fc(m, fc1, "y", dim_in=output_dim, dim_out=output_dim)
fc3 = brew.fc(m, fc1, "z", dim_in=output_dim, dim_out=output_dim)
brew.sum(m, [fc2, fc3], "out")
m2 = model_helper.ModelHelper()
m2.Proto().type = "dag"
m2.Proto().num_workers = 4
with core.NameScope("name_x"):
fc1 = brew.fc(m2, "data", "x", dim_in=input_dim, dim_out=output_dim)
fc2 = brew.fc(m2, fc1, "y", dim_in=output_dim, dim_out=output_dim)
fc3 = brew.fc(m2, fc1, "y", dim_in=output_dim, dim_out=output_dim)
brew.sum(m2, [fc2, fc3], "out")
self.assertFalse(memonger.verify_graph_equality(m.net.Proto(), m2.net.Proto()))
def test_forward_optim_tree_daggy(self, input_dim, output_dim, batch_size):
m = model_helper.ModelHelper()
m.Proto().type = "dag"
m.Proto().num_workers = 4
with core.NameScope("name_x"):
fc1 = brew.fc(m, "data", "fc1", dim_in=input_dim, dim_out=output_dim)
fc2 = brew.fc(m, fc1, "fc2", dim_in=output_dim, dim_out=output_dim)
fc3 = brew.fc(m, fc2, "fc3", dim_in=output_dim, dim_out=output_dim)
fc4 = brew.fc(m, fc3, "fc4", dim_in=output_dim, dim_out=output_dim)
fc5 = brew.fc(m, fc4, "fc5", dim_in=output_dim, dim_out=output_dim)
# Branch
fc3b = brew.fc(m, fc2, "fc3b", dim_in=output_dim, dim_out=output_dim)
fc4b = brew.fc(m, fc3b, "fc4b", dim_in=output_dim, dim_out=output_dim)
fc5b = brew.fc(m, fc4b, "fc5b", dim_in=output_dim, dim_out=output_dim)
fc5sum = brew.sum(m, [fc5, fc5b], "fc5sum")
fc5.Relu([], fc5sum) \
.Softmax([], "pred1") \
.LabelCrossEntropy(["label"], ["xent1"]) \
.AveragedLoss([], "loss1")
fc6 = brew.fc(m, fc5, "fc6", dim_in=output_dim, dim_out=output_dim)
fc6.Relu([], fc6) \
.Softmax([], "pred2") \
.LabelCrossEntropy(["label"], ["xent2"]) \
.AveragedLoss([], "loss2")
blobs_before = count_blobs(m.net.Proto())
optim_proto = memonger.optimize_inference_for_dag(
m.net, ["name_x/data"], "name_x"
)
blobs_after = count_blobs(optim_proto)
self.assertLess(blobs_after, blobs_before)
# Test networks produce exactly same results
data = np.random.randn(batch_size, input_dim).astype(np.float32)
label = np.random.randint(
low=0, high=output_dim, size=(batch_size,)).astype(np.int32)
workspace.RunNetOnce(m.param_init_net)
workspace.FeedBlob("name_x/data", data)
workspace.FeedBlob("name_x/label", label)
workspace.RunNetOnce(m.net)
loss1 = workspace.FetchBlob("name_x/loss1")
loss2 = workspace.FetchBlob("name_x/loss2")
workspace.RunNetOnce(optim_proto)
optimized_loss1 = workspace.FetchBlob("name_x/loss1")
optimized_loss2 = workspace.FetchBlob("name_x/loss2")
np.testing.assert_almost_equal(loss1, optimized_loss1)
np.testing.assert_almost_equal(loss2, optimized_loss2)
def test_forward_optim_tree_daggy(self, input_dim, output_dim, batch_size):
m = model_helper.ModelHelper()
m.Proto().type = "dag"
m.Proto().num_workers = 4
with core.NameScope("name_x"):
fc1 = brew.fc(m, "data", "fc1", dim_in=input_dim, dim_out=output_dim)
fc2 = brew.fc(m, fc1, "fc2", dim_in=output_dim, dim_out=output_dim)
fc3 = brew.fc(m, fc2, "fc3", dim_in=output_dim, dim_out=output_dim)
fc4 = brew.fc(m, fc3, "fc4", dim_in=output_dim, dim_out=output_dim)
fc5 = brew.fc(m, fc4, "fc5", dim_in=output_dim, dim_out=output_dim)
# Branch
fc3b = brew.fc(m, fc2, "fc3b", dim_in=output_dim, dim_out=output_dim)
fc4b = brew.fc(m, fc3b, "fc4b", dim_in=output_dim, dim_out=output_dim)
fc5b = brew.fc(m, fc4b, "fc5b", dim_in=output_dim, dim_out=output_dim)
fc5sum = brew.sum(m, [fc5, fc5b], "fc5sum")
fc5.Relu([], fc5sum) \
.Softmax([], "pred1") \
.LabelCrossEntropy(["label"], ["xent1"]) \
.AveragedLoss([], "loss1")
fc6 = brew.fc(m, fc5, "fc6", dim_in=output_dim, dim_out=output_dim)
fc6.Relu([], fc6) \
.Softmax([], "pred2") \
.LabelCrossEntropy(["label"], ["xent2"]) \
.AveragedLoss([], "loss2")
blobs_before = count_blobs(m.net.Proto())
optim_proto = memonger.optimize_inference_for_dag(
m.net, ["name_x/data"], "name_x"
)
blobs_after = count_blobs(optim_proto)
self.assertLess(blobs_after, blobs_before)
# Test networks produce exactly same results
data = np.random.randn(batch_size, input_dim).astype(np.float32)
label = np.random.randint(
low=0, high=output_dim, size=(batch_size,)).astype(np.int32)
workspace.RunNetOnce(m.param_init_net)
workspace.FeedBlob("name_x/data", data)
workspace.FeedBlob("name_x/label", label)
workspace.RunNetOnce(m.net)
loss1 = workspace.FetchBlob("name_x/loss1")
loss2 = workspace.FetchBlob("name_x/loss2")
workspace.RunNetOnce(optim_proto)
optimized_loss1 = workspace.FetchBlob("name_x/loss1")
optimized_loss2 = workspace.FetchBlob("name_x/loss2")
np.testing.assert_almost_equal(loss1, optimized_loss1)
np.testing.assert_almost_equal(loss2, optimized_loss2)
def block_head(self, model, v, towers, num_in_channels, num_out_channels,
scale=1.0, relu=True, name='block_head_node'):
tower_mixed = brew.concat(model, towers, blob_out=name+'_tower_mixed')
tower_out = self.conv_factory(model, tower_mixed, num_in_channels,
num_filters=num_out_channels,
kernel=1, relu=relu, name=name+'tower_out')
#v = v + scale * tower_out
scaled = model.Scale(tower_out, name + '_scale', scale=scale)
v = brew.sum(model, [v, scaled], name+'_sum')
#
if relu is True:
v = brew.relu(model, v, name + '_relu')
return v
def add_simple_block(self,
input_filters,
num_filters,
down_sampling=False,
spatial_batch_norm=True):
self.comp_idx = 0
shortcut_blob = self.prev_blob
# 3x3
self.add_conv(input_filters,
num_filters,
kernel=3,
stride=(1 if down_sampling is False else 2),
pad=1)
if spatial_batch_norm:
self.add_spatial_bn(num_filters)
self.add_relu()
last_conv = self.add_conv(num_filters, num_filters, kernel=3, pad=1)
if spatial_batch_norm:
last_conv = self.add_spatial_bn(num_filters)
if (num_filters != input_filters):
shortcut_blob = brew.conv(
self.model,
shortcut_blob,
'shortcut_projection_%d' % self.comp_count,
input_filters,
num_filters,
weight_init=("MSRAFill", {}),
kernel=1,
stride=(1 if down_sampling is False else 2),
no_bias=self.no_bias,
)
if spatial_batch_norm:
shortcut_blob = brew.spatial_bn(
self.model,
shortcut_blob,
'shortcut_projection_%d_spatbn' % self.comp_count,
num_filters,
epsilon=1e-3,
is_test=self.is_test,
)
self.prev_blob = brew.sum(
self.model, [shortcut_blob, last_conv],
'comp_%d_sum_%d' % (self.comp_count, self.comp_idx))
self.comp_idx += 1
self.add_relu()
self.comp_count += 1
def add_simple_block(
self,
input_filters,
num_filters,
down_sampling=False,
spatial_batch_norm=True,
):
self.comp_idx = 0
shortcut_blob = self.prev_blob
if spatial_batch_norm:
self.add_spatial_bn(input_filters)
pre_relu = self.add_relu()
self.add_conv(
input_filters,
num_filters,
kernel=3,
stride=(1 if down_sampling is False else 2),
pad=1,
)
if spatial_batch_norm:
self.add_spatial_bn(num_filters)
self.add_relu()
last_conv = self.add_conv(num_filters, num_filters, kernel=3, pad=1)
# Increase of dimensions, need a projection for the shortcut
if (num_filters != input_filters):
shortcut_blob = brew.conv(
self.model,
pre_relu,
'shortcut_projection_%d' % self.comp_count,
input_filters,
num_filters,
weight_init=("MSRAFill", {}),
kernel=1,
stride=(1 if down_sampling is False else 2),
no_bias=self.no_bias,
)
self.prev_blob = brew.sum(
self.model,
[shortcut_blob, last_conv],
'comp_%d_sum_%d' % (self.comp_count, self.comp_idx),
)
self.comp_idx += 1
self.comp_count += 1
def test_fast_memonger_unique_outputs(self):
m = model_helper.ModelHelper()
fc = []
for i in range(2):
z = brew.fc(
m, "data{}".format(i), "fc".format(i), dim_in=2, dim_out=2)
fc.append(z)
r = []
# Trick is here to have same input appear twice in a same Sum
for x in fc:
for y in fc:
r.append(brew.sum(m, [x, y], 1))
concated = brew.concat(m, r, "concated")
brew.relu(m, concated, "merged")
static_blobs = \
[o for op in m.param_init_net.Proto().op for o in op.output] + \
["merged"] + ["data{}".format(i) for i in range(len(fc))]
optimized_net = memonger.optimize_inference_fast(
m.Proto(), static_blobs)
for op in optimized_net.op:
self.assertEqual(len(op.output), len(set(op.output)), str(op))
def fused_gated_global_pool(
self,
blob_in,
dim_in,
prefix='',
ratio=8,
down_rate=2,
res_gate=False,
):
fc1 = self.model.FC(blob_in, prefix + '_fc1', dim_in, dim_in // ratio)
fc1_relu = self.model.Relu(fc1, prefix + '_fc1_relu')
fc2 = self.model.FC(fc1_relu, prefix + '_fc2', dim_in // ratio,
dim_in // down_rate)
sig = self.model.Sigmoid(fc2, fc2 + '_sig')
shortcut_blob = self.prev_blob
self.prev_blob = self.model.Mul([self.prev_blob, sig],
[prefix + 'g_pool_out'],
broadcast=1,
axis=0)
if res_gate:
self.prev_blob = brew.sum(self.model,
[shortcut_blob, self.prev_blob],
prefix + 'res_pool_out')
return self.prev_blob
def add_bottleneck(
self,
input_filters, # num of feature maps from preceding layer
base_filters, # num of filters internally in the component
output_filters, # num of feature maps to output
stride=1,
group=1,
spatial_batch_norm=True,
):
self.comp_idx = 0
shortcut_blob = self.prev_blob
# 1x1
self.add_conv(
input_filters,
base_filters,
kernel=1,
stride=1,
)
if spatial_batch_norm:
self.add_spatial_bn(base_filters)
self.add_relu()
# 3x3 (note the pad, required for keeping dimensions)
self.add_conv(
base_filters,
base_filters,
kernel=3,
stride=stride,
group=group,
pad=1,
)
if spatial_batch_norm:
self.add_spatial_bn(base_filters)
self.add_relu()
# 1x1
last_conv = self.add_conv(base_filters, output_filters, kernel=1)
if spatial_batch_norm:
last_conv = self.add_spatial_bn(output_filters)
# Summation with input signal (shortcut)
# When the number of feature maps mismatch between the input
# and output (this usually happens when the residual stage
# changes), we need to do a projection for the short cut
if output_filters != input_filters:
shortcut_blob = brew.conv(
self.model,
shortcut_blob,
'shortcut_projection_%d' % self.comp_count,
input_filters,
output_filters,
weight_init=("MSRAFill", {}),
kernel=1,
stride=stride,
no_bias=self.no_bias,
)
if spatial_batch_norm:
shortcut_blob = brew.spatial_bn(
self.model,
shortcut_blob,
'shortcut_projection_%d_spatbn' % self.comp_count,
output_filters,
epsilon=self.bn_epsilon,
momentum=self.bn_momentum,
is_test=self.is_test,
)
self.prev_blob = brew.sum(
self.model, [shortcut_blob, last_conv],
'comp_%d_sum_%d' % (self.comp_count, self.comp_idx))
self.comp_idx += 1
self.add_relu()
# Keep track of number of high level components if this ResNetBuilder
self.comp_count += 1
return output_filters
def add_bottleneck(
self,
input_filters, # num of feature maps from preceding layer
base_filters, # num of filters internally in the component
output_filters, # num of feature maps to output
down_sampling=False,
spatial_batch_norm=True,
):
self.comp_idx = 0
shortcut_blob = self.prev_blob
# 1x1
self.add_conv(input_filters, base_filters, kernel=1, stride=1)
if spatial_batch_norm:
self.add_spatial_bn(base_filters)
self.add_relu()
# 3x3 (note the pad, required for keeping dimensions)
self.add_conv(base_filters,
base_filters,
kernel=3,
stride=(1 if down_sampling is False else 2),
pad=1)
if spatial_batch_norm:
self.add_spatial_bn(base_filters)
self.add_relu()
# 1x1
last_conv = self.add_conv(base_filters, output_filters, kernel=1)
if spatial_batch_norm:
last_conv = self.add_spatial_bn(output_filters)
# Summation with input signal (shortcut)
# If we need to increase dimensions (feature maps), need to
# do a projection for the short cut
if (output_filters > input_filters):
shortcut_blob = brew.conv(
self.model,
shortcut_blob,
'shortcut_projection_%d' % self.comp_count,
input_filters,
output_filters,
weight_init=("MSRAFill", {}),
kernel=1,
stride=(1 if down_sampling is False else 2),
no_bias=self.no_bias,
)
if spatial_batch_norm:
shortcut_blob = brew.spatial_bn(
self.model,
shortcut_blob,
'shortcut_projection_%d_spatbn' % self.comp_count,
output_filters,
epsilon=1e-3,
momentum=self.spatial_bn_mom,
is_test=self.is_test,
)
self.prev_blob = brew.sum(
self.model, [shortcut_blob, last_conv],
'comp_%d_sum_%d' % (self.comp_count, self.comp_idx))
self.comp_idx += 1
self.add_relu()
# Keep track of number of high level components if this ResNetBuilder
self.comp_count += 1
def create_caffe2_model(model,
input_shape,
use_cudnn=True,
init_params=False,
keras_channel_last=True):
arg_scope = {'order': 'NCHW', 'use_cudnn': use_cudnn}
caffe2_model = model_helper.ModelHelper(name='model',
init_params=init_params,
arg_scope=arg_scope)
num_conv_layers = 0
layer_num = 0
layer_sizes = {}
prev_layer_name = ''
for layer in model.layers:
inb_node = layer._inbound_nodes[0]
num_input_layers = len(inb_node.inbound_layers)
input_name_list = []
for ii in range(0, num_input_layers):
inp_layer = inb_node.inbound_layers[ii]
input_name_list.append(inp_layer.name)
prev_layer_name = inp_layer.name
if isinstance(inp_layer, keras.layers.Flatten):
pass
#pinb_node = inp_layer._inbound_nodes[0]
#prev_layer_name = pinb_node.inbound_layers[0].name
name = layer.name
config = layer.get_config()
inputShape = layer.input_shape
outputShape = layer.output_shape
if isinstance(layer, keras.engine.input_layer.InputLayer):
input_sizes = (input_shape[2], input_shape[3])
layer_sizes[name] = input_sizes
else:
if (input_name_list[0] not in layer_sizes):
raise ValueError("Can't find layer size for ",
input_name_list[0])
else:
input_sizes = layer_sizes[input_name_list[0]]
layer_dim = len(outputShape)
if (layer_dim == 4):
if (keras_channel_last):
out_sizes = (outputShape[1], outputShape[2])
else:
out_sizes = (outputShape[2], outputShape[3])
elif (layer_dim == 2):
out_sizes = (0, 0) #flattened
else:
raise ValueError(
'Unsupported layer dimension : {0}'.format(layer_dim))
if isinstance(layer, keras.layers.Flatten):
tmp_prev = prev_layer_name
if (keras_channel_last):
tmp_prev = prev_layer_name + '_transpose' #nb, img_h, img_w, chan <-- nb, chan, img_h, img_w
c2_layer = brew.transpose(caffe2_model,
prev_layer_name,
tmp_prev,
axes=(0, 2, 3, 1))
c2_layer = caffe2_model.net.Flatten(tmp_prev, name)
#print('FLatten previous layer ', prev_layer_name, ' current layer ', name , 'inputshape ', inputShape)
layer_sizes[name] = out_sizes
elif isinstance(layer, keras.layers.Dropout):
#print('name is ', name, ' prev_layer_name ', prev_layer_name)
c2_layer = caffe2_model.net.Dropout(prev_layer_name,
name,
is_test=True
#ratio=config['rate']
)
#same size
layer_sizes[name] = input_sizes
elif (isinstance(layer, keras.layers.convolutional.Conv2D)):
dim_in = inputShape[-1]
dim_out = outputShape[-1]
kernel = config['kernel_size'][0]
stride = config['strides'][0]
if (config['padding'] == 'same'):
pad_sizes = get_padding_sizes(input_sizes,
config['kernel_size'],
config['strides'])
elif (config['padding'] == 'valid'):
pad_sizes = ((0, 0), (0, 0))
else:
raise ValueError('unsupported padding')
#print('pad sizes ', pad_sizes)
layer_sizes[name] = out_sizes
c2_layer = brew.conv(caffe2_model,
prev_layer_name,
name,
dim_in=dim_in,
dim_out=dim_out,
kernel=kernel,
stride=stride,
pad_l=pad_sizes[0][0],
pad_r=pad_sizes[0][1],
pad_t=pad_sizes[1][0],
pad_b=pad_sizes[1][1])
if config['activation'] == 'linear':
pass
elif config['activation'] == 'relu':
c2_layer = brew.relu(caffe2_model, name, name)
elif config['activation'] == 'softmax':
#c2_layer = brew.softmax(caffe2_model, name, name)
c2_layer = brew.softmax(caffe2_model, name, 'softmax')
else:
raise ValueError(
'The only supported activation for conv layer is relu')
elif isinstance(layer, keras.layers.MaxPooling2D):
kernel = config['pool_size'][0]
stride = config['strides'][0]
pad_size = ((0, 0), (0, 0))
layer_sizes[name] = out_sizes
c2_layer = brew.max_pool(caffe2_model,
prev_layer_name,
name,
kernel=kernel,
stride=stride)
elif isinstance(layer, keras.layers.AveragePooling2D):
kernel = config['pool_size'][0]
stride = config['strides'][0]
pad_size = ((0, 0), (0, 0))
layer_sizes[name] = out_sizes
c2_layer = brew.average_pool(caffe2_model,
prev_layer_name,
name,
kernel=kernel,
stride=stride)
elif isinstance(layer, keras.layers.BatchNormalization):
dim_in = inputShape[-1]
epsilon = config['epsilon']
momentum = config['momentum']
c2_layer = brew.spatial_bn(caffe2_model,
prev_layer_name,
name,
dim_in=dim_in,
epsilon=epsilon,
momentum=momentum,
is_test=True)
#same size
layer_sizes[name] = input_sizes
elif (isinstance(layer, keras.layers.core.Dense)):
dim_in = inputShape[-1]
dim_out = outputShape[-1]
#print('input shape for dense is ', inputShape)
if (len(inputShape) == 2): #flattened input
c2_layer = brew.fc(caffe2_model,
prev_layer_name,
name,
dim_in=dim_in,
dim_out=dim_out)
else: #fully convolutional input
c2_layer = brew.conv(caffe2_model,
prev_layer_name,
name,
dim_in=dim_in,
dim_out=dim_out,
kernel=1,
stride=1)
activation = config['activation']
if activation == 'relu':
c2_layer = brew.relu(caffe2_model, name, name)
elif activation == 'softmax':
c2_layer = brew.softmax(caffe2_model, name, 'softmax')
elif activation == 'linear':
pass #
else:
raise ValueError(
'The only supported activations for fc layer are relu and softmax'
)
#same size
layer_sizes[name] = input_sizes
elif (isinstance(layer, keras.layers.advanced_activations.LeakyReLU)):
dim_in = inputShape[-1]
c2_layer = caffe2_model.net.LeakyRelu(prev_layer_name,
name,
alpha=config['alpha'])
#same size
layer_sizes[name] = input_sizes
elif (isinstance(layer, keras.layers.merge.Add)):
c2_layer = brew.sum(caffe2_model,
[input_name_list[0], input_name_list[1]], name)
#same size
layer_sizes[name] = input_sizes
layer_num = layer_num + 1
if (layer_num == len(model.layers)):
caffe2_model.net.AddExternalOutput(c2_layer)
return caffe2_model
def Sum(self, *args, **kwargs):
return brew.sum(self, *args, **kwargs)
def res_block_1_conv_topk(
model,
inputs,
dim_in,
dim_out,
topk=5,
no_bias=False,
is_test=False,
module_seq=None,
sub_seq=None,
):
# branch of 1 (projection)
branch1_conv = meta_conv_conv_topk(
model,
inputs,
dim_in,
dim_out,
kernel=1,
pad=0,
stride=2,
topk=topk,
no_bias=no_bias,
is_test=is_test,
has_relu=False,
module_seq=module_seq,
sub_seq=sub_seq,
branch_seq='1',
conv_seq='',
)
# branch of 2 (normal)
branch2_conv1 = meta_conv_conv_topk(
model,
inputs,
dim_in,
dim_out,
kernel=3,
pad=1,
stride=2,
topk=topk,
no_bias=no_bias,
is_test=is_test,
has_relu=True,
module_seq=module_seq,
sub_seq=sub_seq,
branch_seq='2',
conv_seq='0',
)
branch2_conv2 = meta_conv_conv_topk(
model,
branch2_conv1,
dim_out,
dim_out,
kernel=3,
pad=1,
stride=1,
topk=topk,
no_bias=no_bias,
is_test=is_test,
has_relu=False,
module_seq=module_seq,
sub_seq=sub_seq,
branch_seq='2',
conv_seq='1',
)
# sum
branch_sum = brew.sum(
model,
[branch2_conv2, branch1_conv],
# branch2_conv2 # in-place
["group{}_conv{}_sum".format(module_seq, sub_seq)]
)
branch_relu = brew.relu(
model,
branch_sum,
branch_sum
)
return branch_relu
def create_gru_unit(self, emb_ls, user_emb_ids, model, tag, seq_q, hid_q):
(tag_layer, tag_in, tag_out) = tag
emb_ls_str = []
for user_emb_id in user_emb_ids:
emb_ls_str.append(emb_ls[user_emb_id])
tag_cat = tag_layer + ":::_rnn_inputs"
tag_cat_info = tag_cat + "_info"
rnn_inputs, info = model.net.Concat(emb_ls_str,
[tag_cat, tag_cat_info])
rnn_shape = model.net.Reshape(
rnn_inputs, [tag_layer + ":::rnn_shape", "old_shape"],
shape=(len(user_emb_ids), -1, self.input_size))
gates_t_w_data = np.random.randn(
self.args.hidden_size, self.args.hidden_size).astype(np.float32)
gates_t_b_data = np.random.randn(self.args.hidden_size).astype(
np.float32)
i2h_w_data = np.random.randn(self.args.hidden_size,
self.input_size).astype(np.float32)
i2h_b_data = np.random.randn(self.args.hidden_size).astype(np.float32)
workspace.FeedBlob('rnn_0/gates_t_w', gates_t_w_data)
workspace.FeedBlob('rnn_0/gates_t_b', gates_t_b_data)
workspace.FeedBlob('rnn_0/i2h_w', i2h_w_data)
workspace.FeedBlob('rnn_0/i2h_b', i2h_b_data)
if seq_q:
model.net.DequeueBlobs(seq_q, "seq_lengths")
if hid_q:
model.net.DequeueBlobs(hid_q, "initial_h")
rnn_0_out, _ = rnn_cell.BasicRNN(model,
tag_layer + ":::rnn_shape",
'seq_lengths', ['initial_h'],
self.input_size,
self.args.hidden_size,
"rnn_0",
activation="tanh",
forward_only=True)
output = brew.fc(self.model,
rnn_0_out,
None,
dim_in=self.args.hidden_size,
dim_out=self.args.hidden_size,
axis=2,
engine=self.args.engine,
max_num_tasks=self.args.fc_workers)
output = brew.softmax(self.model, output, axis=2)
output = brew.sum(self.model, rnn_0_out, output, axis=2)
# TODO: Need to make input_h_data an input to the overall model due to
# batch-size
gates_t_w_data = np.random.randn(
self.args.hidden_size, self.args.hidden_size).astype(np.float32)
gates_t_b_data = np.random.randn(self.args.hidden_size).astype(
np.float32)
i2h_w_data = np.random.randn(self.args.hidden_size,
self.args.hidden_size).astype(np.float32)
i2h_b_data = np.random.randn(self.args.hidden_size).astype(np.float32)
workspace.FeedBlob('rnn_1/gates_t_w', gates_t_w_data)
workspace.FeedBlob('rnn_1/gates_t_b', gates_t_b_data)
workspace.FeedBlob('rnn_1/i2h_w', i2h_w_data)
workspace.FeedBlob('rnn_1/i2h_b', i2h_b_data)
rnn_1_all_out, rnn_1_out = rnn_cell.BasicRNN(model,
output,
'seq_lengths',
['initial_h'],
self.args.hidden_size,
self.args.hidden_size,
"rnn_1",
activation="tanh",
forward_only=True)
return rnn_1_out
def Sum(self, *args, **kwargs):
return brew.sum(self, *args, **kwargs)
def res_block_2_conv_topk(
model,
inputs,
dim_in,
dim_out,
topk=5,
no_bias=False,
is_test=False,
module_seq=None,
sub_seq=None,
):
# input & output channel check
assert(dim_in == dim_out)
# branch of 2 (normal)
branch2_conv1 = meta_conv_conv_topk(
model,
inputs,
dim_in,
dim_out,
kernel=3,
pad=1,
stride=1,
topk=topk,
no_bias=no_bias,
is_test=is_test,
has_relu=True,
module_seq=module_seq,
sub_seq=sub_seq,
branch_seq='2',
conv_seq='0',
)
branch2_conv2 = meta_conv_conv_topk(
model,
branch2_conv1,
dim_in,
dim_out,
kernel=3,
pad=1,
stride=1,
topk=topk,
no_bias=no_bias,
is_test=is_test,
has_relu=False,
module_seq=module_seq,
sub_seq=sub_seq,
branch_seq='2',
conv_seq='1',
)
# sum
branch_sum = brew.sum(
model,
[branch2_conv2, inputs],
# branch2_conv2 # in-place
["group{}_conv{}_sum".format(module_seq, sub_seq)]
)
branch_relu = brew.relu(
model,
branch_sum,
branch_sum
)
return branch_relu
def add_bottleneck(
self,
input_filters,
base_filters,
output_filters,
down_sampling=False,
spatial_batch_norm=True,
):
self.comp_idx = 0
shortcut_blob = self.prev_blob
# 1x1
self.add_conv(input_filters, base_filters, kernel=1, stride=1)
if spatial_batch_norm:
self.add_spatial_bn(base_filters)
self.add_relu()
# 3x3 (note the pad, required for keeping dimensions)
self.add_conv(base_filters,
base_filters,
kernel=3,
stride=(1 if down_sampling is False else 2),
pad=1)
if spatial_batch_norm:
self.add_spatial_bn(base_filters)
self.add_relu()
# 1x1
last_conv = self.add_conv(base_filters, output_filters, kernel=1)
if spatial_batch_norm:
last_conv = self.add_spatial_bn(output_filters)
if (output_filters > input_filters):
shortcut_blob = brew.conv(
self.model,
shortcut_blob,
'shortcut_projection_%d' % self.comp_count,
input_filters,
output_filters,
weight_init=("MSRAFill", {}),
kernel=1,
stride=(1 if down_sampling is False else 2),
no_bias=self.no_bias,
)
if spatial_batch_norm:
shortcut_blob = brew.spatial_bn(
self.model,
shortcut_blob,
'shortcut_projection_%d_spatbn' % self.comp_count,
output_filters,
epsilon=1e-3,
momentum=self.spatial_bn_mom,
is_test=self.is_test,
)
self.prev_blob = brew.sum(
self.model, [shortcut_blob, last_conv],
'comp_%d_sum_%d' % (self.comp_count, self.comp_idx))
self.comp_idx += 1
self.add_relu()
self.comp_count += 1
def add_bottleneck(
self,
input_filters, # num of feature maps from preceding layer
base_filters, # num of filters internally in the component
output_filters, # num of feature maps to output
down_sampling=False,
spatial_batch_norm=True,
):
self.comp_idx = 0
shortcut_blob = self.prev_blob
# 1x1
self.add_conv(
input_filters,
base_filters,
kernel=1,
stride=1
)
if spatial_batch_norm:
self.add_spatial_bn(base_filters)
self.add_relu()
# 3x3 (note the pad, required for keeping dimensions)
self.add_conv(
base_filters,
base_filters,
kernel=3,
stride=(1 if down_sampling is False else 2),
pad=1
)
if spatial_batch_norm:
self.add_spatial_bn(base_filters)
self.add_relu()
# 1x1
last_conv = self.add_conv(base_filters, output_filters, kernel=1)
if spatial_batch_norm:
last_conv = self.add_spatial_bn(output_filters)
# Summation with input signal (shortcut)
# If we need to increase dimensions (feature maps), need to
# do a projection for the short cut
if (output_filters > input_filters):
shortcut_blob = brew.conv(
self.model,
shortcut_blob,
'shortcut_projection_%d' % self.comp_count,
input_filters,
output_filters,
weight_init=("MSRAFill", {}),
kernel=1,
stride=(1 if down_sampling is False else 2),
no_bias=self.no_bias,
)
if spatial_batch_norm:
shortcut_blob = brew.spatial_bn(
self.model,
shortcut_blob,
'shortcut_projection_%d_spatbn' % self.comp_count,
output_filters,
epsilon=1e-3,
momentum=self.spatial_bn_mom,
is_test=self.is_test,
)
self.prev_blob = brew.sum(
self.model, [shortcut_blob, last_conv],
'comp_%d_sum_%d' % (self.comp_count, self.comp_idx)
)
self.comp_idx += 1
self.add_relu()
# Keep track of number of high level components if this ResNetBuilder
self.comp_count += 1
def residual_unit(self,
model,
v,
num_in_channels,
num_filter,
stride,
dim_match,
name,
bottle_neck=True,
bn_mom=0.9,
is_inference=False):
"""Return ResNet Unit symbol for building ResNet
Parameters
----------
data : str
Input data
num_filter : int
Number of output channels
bnf : int
Bottle neck channels factor with regard to num_filter
stride : tuple
Stride used in convolution
dim_match : Boolean
True means channel number between input and output is the same, otherwise means differ
name : str
Base name of the operators
workspace : int
Workspace used in convolution operator
"""
if bottle_neck:
# Branch 1
if dim_match:
shortcut = v
else:
shortcut = brew.conv(model,
v,
name + '_sc',
num_in_channels,
num_filter,
kernel=1,
stride=stride,
no_bias=True)
shortcut = brew.spatial_bn(model,
shortcut,
name + '_sc_bn',
num_filter,
eps=2e-5,
momentum=bn_mom,
is_test=is_inference)
# Branch 2
interim_filters = int(num_filter * 0.25)
# Block 1
conv1 = brew.conv(model,
v,
name + '_conv1',
num_in_channels,
interim_filters,
kernel=1,
pad=0,
stride=1,
no_bias=True)
bn1 = brew.spatial_bn(model,
conv1,
name + '_bn1',
interim_filters,
eps=2e-5,
momentum=bn_mom,
is_test=is_inference)
act1 = brew.relu(model, bn1, name + '_relu1')
# Block 2
conv2 = brew.conv(model,
act1,
name + '_conv2',
interim_filters,
int(num_filter * 0.25),
kernel=3,
pad=1,
stride=stride,
no_bias=True)
bn2 = brew.spatial_bn(model,
conv2,
name + '_bn2',
interim_filters,
eps=2e-5,
momentum=bn_mom,
is_test=is_inference)
act2 = brew.relu(model, bn2, name + '_relu2')
# Block 3
conv3 = brew.conv(model,
act2,
name + '_conv3',
interim_filters,
num_filter,
kernel=1,
pad=0,
stride=1,
no_bias=True)
bn3 = brew.spatial_bn(model,
conv3,
name + '_bn3',
num_filter,
eps=2e-5,
momentum=bn_mom,
is_test=is_inference)
# Element-wise summation and ReLU
output = brew.sum(model, [shortcut, bn3], name + '_sum')
output = brew.relu(model, output, name + '_relu')
return output
else:
# Branch 1
if dim_match:
shortcut = v
else:
shortcut = brew.conv(model,
v,
name + '_sc_conv',
num_in_channels,
num_filter,
kernel=1,
stride=stride,
no_bias=True)
shortcut = brew.spatial_bn(model,
shortcut,
name + '_sc_bn',
num_filter,
eps=2e-5,
momentum=bn_mom,
is_test=is_inference)
# Branch 2
# Block 1
conv1 = brew.conv(model,
v,
name + '_conv1',
num_in_channels,
num_filter,
kernel=3,
pad=1,
stride=stride,
no_bias=True)
bn1 = brew.spatial_bn(model,
conv1,
name + '_bn1',
num_filter,
eps=2e-5,
momentum=bn_mom,
is_test=is_inference)
act1 = brew.relu(model, bn1, name + '_relu1')
# Block 2
conv2 = brew.conv(model,
act1,
name + '_conv2',
num_filter,
num_filter,
kernel=3,
pad=1,
stride=1,
no_bias=True)
bn2 = brew.spatial_bn(model,
conv2,
name + '_bn2',
num_filter,
eps=2e-5,
momentum=bn_mom,
is_test=is_inference)
# Element-wise summation and ReLU
output = brew.sum(model, [shortcut, bn2], name + '_sum')
output = brew.relu(model, output, name + '_relu')
return output
def add_bottleneck(
self,
input_filters, # num of feature maps from preceding layer
output_filters, # num of feature maps to output
base_filters, # num of filters internally in the component
down_sampling=False,
spatial_batch_norm=True,
):
self.comp_idx = 0
shortcut_blob = self.prev_blob
# 1x1
self.add_conv(
input_filters,
base_filters,
kernel=1,
)
if spatial_batch_norm:
self.add_spatial_bn(base_filters)
self.add_relu()
# 3x3
self.add_conv(
base_filters,
base_filters,
kernel=3,
stride=(2 if down_sampling else 1),
pad=1,
)
if spatial_batch_norm:
self.add_spatial_bn(base_filters)
self.add_relu()
# 1x1
last_conv = self.add_conv(
base_filters,
output_filters,
kernel=1,
)
if spatial_batch_norm:
last_conv = self.add_spatial_bn(output_filters)
# Summation with input signal (shortcut)
# If we need to increase dimensions (feature maps), need to
# do do a projection for the short cut
if (output_filters > input_filters or down_sampling):
shortcut_blob = brew.conv(
self.model,
shortcut_blob,
'%sshortcut_projection_%d' % (self.prefix, self.comp_count),
input_filters,
output_filters,
kernel=1,
weight_init=("MSRAFill", {}),
stride=(2 if down_sampling else 1),
no_bias=self.no_bias,
)
if spatial_batch_norm:
shortcut_blob = brew.spatial_bn(
self.model,
shortcut_blob,
'%sshortcut_projection_%d_spatbn' %
(self.prefix, self.comp_count),
output_filters,
epsilon=1e-3,
momentum=self.spatial_bn_mom,
is_test=self.is_test,
)
self.prev_blob = brew.sum(
self.model, [shortcut_blob, last_conv],
'%scomp_%d_sum_%d' % (self.prefix, self.comp_count, self.comp_idx))
self.comp_idx += 1
self.add_relu()
self.comp_count += 1
def res_block_2(
model,
inputs,
dim_in,
dim_mid,
dim_out,
no_bias=False,
is_test=False,
module_seq=None,
sub_seq=None,
):
# branch2 (right)
branch2_conv1 = meta_conv(
model,
inputs,
dim_in,
dim_mid,
kernel=1,
pad=0,
stride=1,
no_bias=no_bias,
is_test=is_test,
has_relu=True,
module_seq=module_seq,
sub_seq=sub_seq,
branch_seq='2',
conv_seq='a',
)
branch2_conv2 = meta_conv(
model,
branch2_conv1,
dim_mid,
dim_mid,
kernel=3,
pad=1,
stride=1,
no_bias=no_bias,
is_test=is_test,
has_relu=True,
module_seq=module_seq,
sub_seq=sub_seq,
branch_seq='2',
conv_seq='b',
)
branch2_conv3 = meta_conv(
model,
branch2_conv2,
dim_mid,
dim_out,
kernel=1,
pad=0,
stride=1,
no_bias=no_bias,
is_test=is_test,
has_relu=False,
module_seq=module_seq,
sub_seq=sub_seq,
branch_seq='2',
conv_seq='c',
)
# sum
branch_sum = brew.sum(model, [branch2_conv3, inputs], branch2_conv3)
branch_relu = brew.relu(model, branch_sum, branch_sum)
return branch_relu