def test_topological_sort_longest_path_multi_target(self):
# two outputs: conv2 and data4
m = model_helper.ModelHelper()
# 0
m.Copy("conv0_w_comp", "conv0_w")
# 1
conv0 = brew.conv(m, "data", "conv0", 32, 32, 4)
# 2
m.Copy("conv2_w", "conv2_w")
# 3
brew.conv(m, conv0, "conv2", 16, 32, 4)
# 4
m.Copy("data1", "data2")
# 5
m.Copy("data2", "data3")
g = memonger.compute_interference_graph(m.net.Proto().op)
orders_org = memonger.topological_sort_traversal(g)
orders_gt_org = [4, 5, 2, 0, 1, 3]
self.assertEqual(orders_gt_org, list(orders_org))
orders = memonger.topological_sort_traversal_longest_path(g)
# longer path is in front of the shorter one
orders_gt = [0, 1, 2, 3, 4, 5]
self.assertEqual(orders_gt, list(orders))
def test_model_helper(self):
X = np.random.rand(64, 32, 32, 3).astype(np.float32) - 0.5
workspace.FeedBlob("x", X)
my_arg_scope = {'order': 'NHWC'}
model = ModelHelper(name="test_model", arg_scope=my_arg_scope)
with brew.arg_scope(
brew.conv,
stride=2,
pad=2,
weight_init=('XavierFill', {}),
bias_init=('ConstantFill', {})
):
brew.conv(
model=model,
blob_in="x",
blob_out="out",
dim_in=3,
dim_out=64,
kernel=3,
)
model.Validate()
workspace.RunNetOnce(model.param_init_net)
workspace.RunNetOnce(model.net)
out = workspace.FetchBlob("out")
self.assertEqual(out.shape, (64, 17, 17, 64))
def AddLeNetModel(model, data):
'''
This part is the standard LeNet model: from data to the softmax prediction.
For each convolutional layer we specify dim_in - number of input channels
and dim_out - number or output channels. Also each Conv and MaxPool layer changes the
image size. For example, kernel of size 5 reduces each side of an image by 4.
While when we have kernel and stride sizes equal 2 in a MaxPool layer, it divides
each side in half.
'''
# Image size: 28 x 28 -> 24 x 24
conv1 = brew.conv(model, data, 'conv1', dim_in=1, dim_out=20, kernel=5)
# Image size: 24 x 24 -> 12 x 12
pool1 = model.net.MaxPool(conv1, 'pool1', kernel=2, stride=2)
# Image size: 12 x 12 -> 8 x 8
conv2 = brew.conv(model, pool1, 'conv2', dim_in=20, dim_out=50, kernel=5)
# Image size: 8 x 8 -> 4 x 4
pool2 = model.net.MaxPool(conv2, 'pool2', kernel=2, stride=2)
# 50 * 4 * 4 stands for dim_out from previous layer multiplied by the image size
fc3 = brew.fc(model, pool2, 'fc3', dim_in=50 * 4 * 4, dim_out=500)
fc3 = model.net.Relu(fc3, 'relu3')
pred = brew.fc(model, fc3, 'pred', 500, 10)
softmax = model.net.Softmax(pred, 'softmax')
return softmax
def test_mobile_exporter(self):
model = ModelHelper(name="mobile_exporter_test_model")
# Test LeNet
brew.conv(model, 'data', 'conv1', dim_in=1, dim_out=20, kernel=5)
brew.max_pool(model, 'conv1', 'pool1', kernel=2, stride=2)
brew.conv(model, 'pool1', 'conv2', dim_in=20, dim_out=50, kernel=5)
brew.max_pool(model, 'conv2', 'pool2', kernel=2, stride=2)
brew.fc(model, 'pool2', 'fc3', dim_in=50 * 4 * 4, dim_out=500)
brew.relu(model, 'fc3', 'fc3')
brew.fc(model, 'fc3', 'pred', 500, 10)
brew.softmax(model, 'pred', 'out')
# Create our mobile exportable networks
workspace.RunNetOnce(model.param_init_net)
init_net, predict_net = mobile_exporter.Export(
workspace, model.net, model.params
)
# Populate the workspace with data
np_data = np.random.rand(1, 1, 28, 28).astype(np.float32)
workspace.FeedBlob("data", np_data)
workspace.CreateNet(model.net)
workspace.RunNet(model.net)
ref_out = workspace.FetchBlob("out")
# Clear the workspace
workspace.ResetWorkspace()
# Populate the workspace with data
workspace.RunNetOnce(init_net)
# Fake "data" is populated by init_net, we have to replace it
workspace.FeedBlob("data", np_data)
# Overwrite the old net
workspace.CreateNet(predict_net, True)
workspace.RunNet(predict_net.name)
manual_run_out = workspace.FetchBlob("out")
np.testing.assert_allclose(
ref_out, manual_run_out, atol=1e-10, rtol=1e-10
)
# Clear the workspace
workspace.ResetWorkspace()
# Predictor interface test (simulates writing to disk)
predictor = workspace.Predictor(
init_net.SerializeToString(), predict_net.SerializeToString()
)
# Output is a vector of outputs but we only care about the first and only result
predictor_out = predictor.run([np_data])
assert len(predictor_out) == 1
predictor_out = predictor_out[0]
np.testing.assert_allclose(
ref_out, predictor_out, atol=1e-10, rtol=1e-10
)
def simple_cnn():
model = ModelHelper(name="r", arg_scope={"order": "NCHW", "is_test": True})
brew.conv(
model, "data", 'conv1', 3, 16, kernel=3, stride=1
)
brew.spatial_bn(
model, 'conv1', 'conv1_spatbn', 16, epsilon=1e-3
)
brew.relu(model, 'conv1_spatbn', 'relu1')
return model, (1, 3, 32, 32)
def testShapeInferenceConvNet(self):
model = model_helper.ModelHelper(name="convtest")
model.NHWC2NCHW("data", "data_nchw")
brew.conv(model, "data_nchw", 'conv1', 3, 64,
weight_init=("MSRAFill", {}), kernel=7,
stride=2, pad=3, no_bias=0)
brew.spatial_bn(model, 'conv1', 'conv1_spatbn_relu', 64, epsilon=1e-3, is_test=False)
brew.relu(model, 'conv1_spatbn_relu', 'conv1_spatbn_relu')
brew.max_pool(model, 'conv1_spatbn_relu', 'pool1', kernel=3, stride=2)
brew.fc(model, 'pool1', 'fc', dim_in=(64 * 56 * 56), dim_out=100)
brew.dropout(model, 'fc', 'fc_drop', is_test=False)
model.Sigmoid('fc_drop', 'fc_sigm')
brew.softmax(model, 'fc_sigm', 'softmax')
model.LabelCrossEntropy(['softmax', 'label'], 'xent')
loss = model.AveragedLoss('xent', 'loss')
model.AddGradientOperators([loss])
LR = model.param_init_net.ConstantFill(
[], 'LR', shape=[1], value=0.1
)
for param in model.GetParams():
param_grad = model.param_to_grad[param]
param_momentum = model.param_init_net.ConstantFill(
[param], param + '_momentum', value=0.0
)
model.net.MomentumSGDUpdate(
[param_grad, param_momentum, LR, param],
[param_grad, param_momentum, param],
)
workspace.FeedBlob(
"data",
np.random.rand(16, 227, 227, 3).astype(np.float32),
)
workspace.FeedBlob(
"label",
(100 * np.random.rand(16)).astype(np.int32),
)
workspace.FeedBlob(
"label",
(100 * np.random.rand(16)).astype(np.int32),
)
# Then do automatic comparison test: run the next once to
# initialize everything
self.InferTensorRunAndCompare(model)
def test_optimizer_context(self):
from caffe2.python import brew, optimizer
from caffe2.python.model_helper import ModelHelper
model = ModelHelper(name="test", arg_scope={'order': 'NCHW'})
count = optimizer._optimizer_instance_count['SgdOptimizer']
cnv_optim = SgdOptimizer(0.15)
weight_optim = SgdOptimizer(0.2)
bias_optim = SgdOptimizer(0.1)
with UseOptimizer(cnv_optim):
cnv = brew.conv(model, 'data', 'cnv', 32, 32, 4)
with UseOptimizer({'WEIGHT': weight_optim, 'BIAS': bias_optim}):
a = brew.fc(model, cnv, 'a', 100, 200)
pred = brew.fc(model, a, 'b', 200, 5)
(softmax, loss) = model.SoftmaxWithLoss(
[pred, 'label'],
['softmax', 'loss'],
)
model.AddGradientOperators([loss])
add_weight_decay(model, weight_decay=1e-4)
# use the following optimizer if none specified in param_info
build_sgd(model, 0.11)
expected_weight_grad = {'b_w_grad', 'a_w_grad', 'cnv_w_grad'}
expected_learning_rate = {
"SgdOptimizer_{}_lr_cpu".format(count): -0.15,
"SgdOptimizer_{}_lr_cpu".format(count + 1): -0.2,
"SgdOptimizer_{}_lr_cpu".format(count + 2): -0.1,
"SgdOptimizer_{}_lr_cpu".format(count + 3): -0.11
}
for op in model.net.Proto().op:
# Check the proto that all weights are decayed and not non-weights
# are decayed.
if op.type == 'WeightedSum' and 'wd_0_0' in op.input:
if op.output[0] not in expected_weight_grad:
print(
"Unexpected param for weight_decay: {}".
format(op.output[0])
)
self.assertTrue(op.output[0] in expected_weight_grad)
expected_weight_grad.remove(op.output[0])
# Check the learning rate for each parameter
if op.type == 'LearningRate':
val = 0
for arg in op.arg:
if arg.name == 'base_lr':
val = arg.f
self.assertAlmostEqual(
val,
expected_learning_rate[op.output[0]]
)
self.assertEqual(
expected_weight_grad,
set(),
"Not all weights were decayed: {}".format(expected_weight_grad)
)
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_param_consistence(self):
model = ModelHelper(name='test_mode')
cnv = brew.conv(model, 'data', 'cnv', 32, 32, 4)
step_model = ModelHelper(name='step_model', param_model=model)
a = brew.fc(step_model, cnv, 'a', 100, 200)
brew.fc(model, a, 'b', 200, 5)
# test the _parameters_info is shared between model and step_model
self.assertEqual(model._parameters_info, step_model._parameters_info)
def Conv(self, *args, **kwargs):
return brew.conv(
self,
*args,
use_cudnn=self.use_cudnn,
order=self.order,
cudnn_exhaustive_search=self.cudnn_exhaustive_search,
ws_nbytes_limit=self.ws_nbytes_limit,
**kwargs
)
def test_topological_sort_longest_path(self):
m = model_helper.ModelHelper()
# 0
m.Copy("conv0_w_comp", "conv0_w")
# 1
conv0 = brew.conv(m, "data", "conv0", 32, 32, 4)
# 2
m.Copy("conv2_w", "conv2_w")
# 3
brew.conv(m, conv0, "conv2", 16, 32, 4)
g = memonger.compute_interference_graph(m.net.Proto().op)
orders_org = memonger.topological_sort_traversal(g)
orders_gt_org = [2, 0, 1, 3]
self.assertEqual(orders_gt_org, orders_org)
orders = memonger.topological_sort_traversal_longest_path(g)
# longer path is in front of the shorter one
orders_gt = [0, 1, 2, 3]
self.assertEqual(orders_gt, orders)
def create_resnet_32x32(
model, data, num_input_channels, num_groups, num_labels, is_test=False
):
'''
Create residual net for smaller images (sec 4.2 of He et. al (2015))
num_groups = 'n' in the paper
'''
# conv1 + maxpool
brew.conv(
model, data, 'conv1', num_input_channels, 16, kernel=3, stride=1
)
brew.spatial_bn(
model, 'conv1', 'conv1_spatbn', 16, epsilon=1e-3, is_test=is_test
)
brew.relu(model, 'conv1_spatbn', 'relu1')
# Number of blocks as described in sec 4.2
filters = [16, 32, 64]
builder = ResNetBuilder(model, 'relu1', is_test=is_test)
prev_filters = 16
for groupidx in range(0, 3):
for blockidx in range(0, 2 * num_groups):
builder.add_simple_block(
prev_filters if blockidx == 0 else filters[groupidx],
filters[groupidx],
down_sampling=(True if blockidx == 0 and
groupidx > 0 else False))
prev_filters = filters[groupidx]
# Final layers
brew.average_pool(
model, builder.prev_blob, 'final_avg', kernel=8, stride=1
)
brew.fc(model, 'final_avg', 'last_out', 64, num_labels)
softmax = brew.softmax(model, 'last_out', 'softmax')
return softmax
def test_extract_simple(self):
from caffe2.python import brew
from caffe2.python.model_helper import ModelHelper, ExtractPredictorNet
model = ModelHelper(name="test", arg_scope={'order': 'NCHW'})
[data, label] = brew.image_input(
model,
"reader", ["xx/data", "label"],
is_test=1,
)
cnv = brew.conv(model, data, 'cnv', 32, 32, 4)
a = brew.fc(model, cnv, 'a', 100, 200)
pred = brew.fc(model, a, 'pred', 200, 5)
brew.softmax(model, [pred, label], "softmax")
(predict_net, export_blobs) = ExtractPredictorNet(
net_proto=model.net.Proto(),
input_blobs=["xx/data"],
output_blobs=["pred"],
renames={"xx/data": "image"},
)
export_blobs = set(export_blobs)
ops = list(predict_net.Proto().op)
for op in ops:
self.assertFalse(op.type == "Softmax")
self.assertFalse("xx/data" in op.input)
# Note: image input should not be included
self.assertEquals(ops[0].type, "Conv")
self.assertEquals(ops[1].type, "FC")
self.assertEquals(ops[2].type, "FC")
self.assertEquals(len(ops), 3)
# test rename happened
self.assertEquals(ops[0].input[0], "image")
# Check export blobs
self.assertTrue("image" not in export_blobs)
self.assertTrue("xx/data" not in export_blobs)
self.assertEqual(set([str(p) for p in model.params]), export_blobs)
# Check external inputs/outputs
self.assertTrue("image" in predict_net.Proto().external_input)
self.assertEquals(set(["pred"]), set(predict_net.Proto().external_output))
self.assertEqual(
set(predict_net.Proto().external_input) -
set([str(p) for p in model.params]), set(["image"])
)
def add_conv(self, in_filters, out_filters, kernel, stride=1, pad=0):
self.comp_idx += 1
self.prev_blob = brew.conv(
self.model,
self.prev_blob,
'comp_%d_conv_%d' % (self.comp_count, self.comp_idx),
in_filters,
out_filters,
weight_init=("MSRAFill", {}),
kernel=kernel,
stride=stride,
pad=pad,
no_bias=self.no_bias,
)
return self.prev_blob
def test_tensorboard_graphs(self):
model = model_helper.ModelHelper(name="overfeat")
data, label = brew.image_input(
model, ["db"], ["data", "label"], is_test=0
)
with core.NameScope("conv1"):
conv1 = brew.conv(model, data, "conv1", 3, 96, 11, stride=4)
relu1 = brew.relu(model, conv1, conv1)
pool1 = brew.max_pool(model, relu1, "pool1", kernel=2, stride=2)
with core.NameScope("classifier"):
fc = brew.fc(model, pool1, "fc", 4096, 1000)
pred = brew.softmax(model, fc, "pred")
xent = model.LabelCrossEntropy([pred, label], "xent")
loss = model.AveragedLoss(xent, "loss")
model.AddGradientOperators([loss], skip=1)
c2_dir = tempfile.mkdtemp()
tf_dir = tempfile.mkdtemp()
with open(os.path.join(c2_dir, "init"), "w") as f:
f.write(str(model.param_init_net.Proto()))
with open(os.path.join(c2_dir, "net"), "w") as f:
f.write(str(model.net.Proto()))
runner = click.testing.CliRunner()
result = runner.invoke(
tb.cli,
["tensorboard-graphs",
"--c2-netdef", os.path.join(c2_dir, "init"),
"--c2-netdef", os.path.join(c2_dir, "net"),
"--tf-dir", tf_dir])
self.assertEqual(result.exit_code, 0)
entries = list(os.walk(tf_dir))
self.assertEqual(len(entries), 1)
((d, _, (fname,)),) = entries
self.assertEqual(tf_dir, d)
events = list(tf.train.summary_iterator(os.path.join(tf_dir, fname)))
self.assertEqual(len(events), 3)
events = events[1:]
nets = [model.param_init_net, model.net]
for i, (event, net) in enumerate(zip(events, nets), start=1):
self.assertEqual(event.step, i)
self.assertEqual(event.wall_time, i)
g = tf.GraphDef()
g.ParseFromString(event.graph_def)
self.assertMultiLineEqual(
str(g),
str(tb_exporter.nets_to_graph_def([net])))
def test_simple_transform(self, input_dim, output_dim, batch_size):
m = model_helper.ModelHelper()
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)
conv = brew.conv(m, fc2, "conv",
dim_in=output_dim,
dim_out=output_dim,
use_cudnn=True,
engine="CUDNN",
kernel=3)
conv.Relu([], conv)\
.Softmax([], "pred") \
.LabelCrossEntropy(["label"], ["xent"]) \
.AveragedLoss([], "loss")
transformed_net_proto = workspace.ApplyTransform(
"ConvToNNPack",
m.net.Proto())
self.assertEqual(transformed_net_proto.op[2].engine, "NNPACK")
def test_apply_transform_if_faster(self, value):
init_net = core.Net("init_net")
init_net.ConstantFill([], ["data"], shape=[5, 5, 5, 5], value=value)
init_net.ConstantFill([], ["conv_w"], shape=[5, 5, 3, 3], value=value)
init_net.ConstantFill([], ["conv_b"], shape=[5], value=value)
self.assertEqual(
workspace.RunNetOnce(init_net.Proto().SerializeToString()), True)
m = model_helper.ModelHelper()
conv = brew.conv(m, "data", "conv",
dim_in=5,
dim_out=5,
kernel=3,
use_cudnn=True,
engine="CUDNN")
conv.Relu([], conv)\
.Softmax([], "pred") \
.AveragedLoss([], "loss")
self.assertEqual(
workspace.RunNetOnce(m.net.Proto().SerializeToString()), True)
proto = workspace.ApplyTransformIfFaster(
"ConvToNNPack",
m.net.Proto(),
init_net.Proto())
self.assertEqual(
workspace.RunNetOnce(proto.SerializeToString()), True)
proto = workspace.ApplyTransformIfFaster(
"ConvToNNPack",
m.net.Proto(),
init_net.Proto(),
warmup_runs=10,
main_runs=100,
improvement_threshold=2.0)
self.assertEqual(
workspace.RunNetOnce(proto.SerializeToString()), True)
def test_weight_decay(self):
from caffe2.python import brew
from caffe2.python.model_helper import ModelHelper
model = ModelHelper(name="test", arg_scope={'order': 'NCHW'})
cnv = brew.conv(model, 'data', 'cnv', 32, 32, 4)
a = brew.fc(model, cnv, 'a', 100, 200)
pred = brew.fc(model, a, 'b', 200, 5)
(softmax, loss) = model.SoftmaxWithLoss(
[pred, 'label'],
['softmax', 'loss'],
)
model.AddGradientOperators([loss])
add_weight_decay(model, weight_decay=1e-4)
build_sgd(model, 0.11)
expected_weight_grad = {'b_w_grad', 'a_w_grad', 'cnv_w_grad'}
# Check the proto that all weights are decayed and not non-weights
# are decayed.
for op in model.net.Proto().op:
if op.type == 'WeightedSum' and 'wd_0_0' in op.input:
if op.output[0] not in expected_weight_grad:
print(
"Unexpected param for weight_decay: {}".
format(op.output[0])
)
self.assertTrue(op.output[0] in expected_weight_grad)
expected_weight_grad.remove(op.output[0])
self.assertEqual(
expected_weight_grad,
set(),
"Not all weights were decayed: {}".format(expected_weight_grad)
)
def _MiniAlexNetNoDropout(self, order):
# First, AlexNet using the cnn wrapper.
model = model_helper.ModelHelper(name="alexnet")
conv1 = brew.conv(
model,
"data",
"conv1",
3,
16,
11,
("XavierFill", {}),
("ConstantFill", {}),
stride=4,
pad=0
)
relu1 = brew.relu(model, conv1, "relu1")
norm1 = brew.lrn(model, relu1, "norm1", size=5, alpha=0.0001, beta=0.75)
pool1 = brew.max_pool(model, norm1, "pool1", kernel=3, stride=2)
conv2 = brew.group_conv(
model,
pool1,
"conv2",
16,
32,
5,
("XavierFill", {}),
("ConstantFill", {"value": 0.1}),
group=2,
stride=1,
pad=2
)
relu2 = brew.relu(model, conv2, "relu2")
norm2 = brew.lrn(model, relu2, "norm2", size=5, alpha=0.0001, beta=0.75)
pool2 = brew.max_pool(model, norm2, "pool2", kernel=3, stride=2)
conv3 = brew.conv(
model,
pool2,
"conv3",
32,
64,
3,
("XavierFill", {'std': 0.01}),
("ConstantFill", {}),
pad=1
)
relu3 = brew.relu(model, conv3, "relu3")
conv4 = brew.group_conv(
model,
relu3,
"conv4",
64,
64,
3,
("XavierFill", {}),
("ConstantFill", {"value": 0.1}),
group=2,
pad=1
)
relu4 = brew.relu(model, conv4, "relu4")
conv5 = brew.group_conv(
model,
relu4,
"conv5",
64,
32,
3,
("XavierFill", {}),
("ConstantFill", {"value": 0.1}),
group=2,
pad=1
)
relu5 = brew.relu(model, conv5, "relu5")
pool5 = brew.max_pool(model, relu5, "pool5", kernel=3, stride=2)
fc6 = brew.fc(
model, pool5, "fc6", 1152, 1024, ("XavierFill", {}),
("ConstantFill", {"value": 0.1})
)
relu6 = brew.relu(model, fc6, "relu6")
fc7 = brew.fc(
model, relu6, "fc7", 1024, 1024, ("XavierFill", {}),
("ConstantFill", {"value": 0.1})
)
relu7 = brew.relu(model, fc7, "relu7")
fc8 = brew.fc(
model, relu7, "fc8", 1024, 5, ("XavierFill", {}),
("ConstantFill", {"value": 0.0})
)
pred = brew.softmax(model, fc8, "pred")
xent = model.LabelCrossEntropy([pred, "label"], "xent")
loss = model.AveragedLoss([xent], ["loss"])
model.AddGradientOperators([loss])
return model
def _InceptionModule(
model, input_blob, input_depth, output_name, conv1_depth, conv3_depths,
conv5_depths, pool_depth
):
# path 1: 1x1 conv
conv1 = brew.conv(
model, input_blob, output_name + ":conv1", input_depth, conv1_depth, 1,
('XavierFill', {}), ('ConstantFill', {})
)
conv1 = brew.relu(model, conv1, conv1)
# path 2: 1x1 conv + 3x3 conv
conv3_reduce = brew.conv(
model, input_blob, output_name + ":conv3_reduce", input_depth,
conv3_depths[0], 1, ('XavierFill', {}), ('ConstantFill', {})
)
conv3_reduce = brew.relu(model, conv3_reduce, conv3_reduce)
conv3 = brew.conv(
model,
conv3_reduce,
output_name + ":conv3",
conv3_depths[0],
conv3_depths[1],
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
conv3 = brew.relu(model, conv3, conv3)
# path 3: 1x1 conv + 5x5 conv
conv5_reduce = brew.conv(
model, input_blob, output_name + ":conv5_reduce", input_depth,
conv5_depths[0], 1, ('XavierFill', {}), ('ConstantFill', {})
)
conv5_reduce = brew.relu(model, conv5_reduce, conv5_reduce)
conv5 = brew.conv(
model,
conv5_reduce,
output_name + ":conv5",
conv5_depths[0],
conv5_depths[1],
5,
('XavierFill', {}),
('ConstantFill', {}),
pad=2
)
conv5 = brew.relu(model, conv5, conv5)
# path 4: pool + 1x1 conv
pool = brew.max_pool(
model,
input_blob,
output_name + ":pool",
kernel=3,
stride=1,
pad=1
)
pool_proj = brew.conv(
model, pool, output_name + ":pool_proj", input_depth, pool_depth, 1,
('XavierFill', {}), ('ConstantFill', {})
)
pool_proj = brew.relu(model, pool_proj, pool_proj)
output = brew.concat(model, [conv1, conv3, conv5, pool_proj], output_name)
return output
def _MiniAlexNetNoDropout(self, order):
# First, AlexNet using the cnn wrapper.
model = model_helper.ModelHelper(name="alexnet")
conv1 = brew.conv(model,
"data",
"conv1",
3,
16,
11, ("XavierFill", {}), ("ConstantFill", {}),
stride=4,
pad=0)
relu1 = brew.relu(model, conv1, "relu1")
norm1 = brew.lrn(model,
relu1,
"norm1",
size=5,
alpha=0.0001,
beta=0.75)
pool1 = brew.max_pool(model, norm1, "pool1", kernel=3, stride=2)
conv2 = brew.group_conv(model,
pool1,
"conv2",
16,
32,
5, ("XavierFill", {}), ("ConstantFill", {
"value": 0.1
}),
group=2,
stride=1,
pad=2)
relu2 = brew.relu(model, conv2, "relu2")
norm2 = brew.lrn(model,
relu2,
"norm2",
size=5,
alpha=0.0001,
beta=0.75)
pool2 = brew.max_pool(model, norm2, "pool2", kernel=3, stride=2)
conv3 = brew.conv(model,
pool2,
"conv3",
32,
64,
3, ("XavierFill", {
'std': 0.01
}), ("ConstantFill", {}),
pad=1)
relu3 = brew.relu(model, conv3, "relu3")
conv4 = brew.group_conv(model,
relu3,
"conv4",
64,
64,
3, ("XavierFill", {}), ("ConstantFill", {
"value": 0.1
}),
group=2,
pad=1)
relu4 = brew.relu(model, conv4, "relu4")
conv5 = brew.group_conv(model,
relu4,
"conv5",
64,
32,
3, ("XavierFill", {}), ("ConstantFill", {
"value": 0.1
}),
group=2,
pad=1)
relu5 = brew.relu(model, conv5, "relu5")
pool5 = brew.max_pool(model, relu5, "pool5", kernel=3, stride=2)
fc6 = brew.fc(model, pool5, "fc6", 1152, 1024, ("XavierFill", {}),
("ConstantFill", {
"value": 0.1
}))
relu6 = brew.relu(model, fc6, "relu6")
fc7 = brew.fc(model, relu6, "fc7", 1024, 1024, ("XavierFill", {}),
("ConstantFill", {
"value": 0.1
}))
relu7 = brew.relu(model, fc7, "relu7")
fc8 = brew.fc(model, relu7, "fc8", 1024, 5, ("XavierFill", {}),
("ConstantFill", {
"value": 0.0
}))
pred = brew.softmax(model, fc8, "pred")
xent = model.LabelCrossEntropy([pred, "label"], "xent")
loss = model.AveragedLoss([xent], ["loss"])
model.AddGradientOperators([loss])
return model
def AddNetModel_6(model, data):
# 0 params, dims [1, 146, 146]
input1 = data
# 72 params, dims [8, 144, 144]
conv2 = brew.conv(model,
input1,
'conv2',
dim_in=1,
dim_out=8,
kernel=3,
stride=1)
# 0 params, dims [8, 144, 144]
relu3 = brew.relu(model, conv2, 'relu3')
# 128 params, dims [4, 72, 72]
conv4 = brew.conv(model,
relu3,
'conv4',
dim_in=8,
dim_out=4,
kernel=2,
stride=2)
# 1600 params, dims [16, 68, 68]
conv5 = brew.conv(model,
conv4,
'conv5',
dim_in=4,
dim_out=16,
kernel=5,
stride=1)
# 0 params, dims [16, 68, 68]
relu6 = brew.relu(model, conv5, 'relu6')
# 512 params, dims [8, 34, 34]
conv7 = brew.conv(model,
relu6,
'conv7',
dim_in=16,
dim_out=8,
kernel=2,
stride=2)
# 2304 params, dims [32, 32, 32]
conv8 = brew.conv(model,
conv7,
'conv8',
dim_in=8,
dim_out=32,
kernel=3,
stride=1)
# 0 params, dims [32, 32, 32]
relu9 = brew.relu(model, conv8, 'relu9')
# 2048 params, dims [16, 16, 16]
conv10 = brew.conv(model,
relu9,
'conv10',
dim_in=32,
dim_out=16,
kernel=2,
stride=2)
# 9216 params, dims [64, 14, 14]
conv11 = brew.conv(model,
conv10,
'conv11',
dim_in=16,
dim_out=64,
kernel=3,
stride=1)
# 0 params, dims [64, 14, 14]
relu12 = brew.relu(model, conv11, 'relu12')
# 8192 params, dims [32, 7, 7]
conv13 = brew.conv(model,
relu12,
'conv13',
dim_in=64,
dim_out=32,
kernel=2,
stride=2)
# 0 params, dims [32, 7, 7]
relu14 = brew.relu(model, conv13, 'relu14')
# 301056 params, dims [192]
fc15 = brew.fc(model, relu14, 'fc15', dim_in=1568, dim_out=192)
# 0 params, dims [192]
relu16 = brew.relu(model, fc15, 'relu16')
# 384 params, dims [2]
output = brew.fc(model, relu16, 'output', dim_in=192, dim_out=2)
# 325512 parameters total
return output
def test_convolution_sync(self, net_type, num_workers, engine, gc, dc):
m = ModelHelper(name="test_model")
n = 1
d = 2
depth = 3
iters = 5
h = 5
w = 5
workspace.ResetWorkspace()
use_cudnn = engine == "CUDNN"
np.random.seed(1701)
# Build a binary tree of conv layers, summing at each node.
for i in reversed(range(depth)):
for j in range(2**i):
bottom_1 = "{}_{}".format(i + 1, 2 * j)
bottom_2 = "{}_{}".format(i + 1, 2 * j + 1)
mid_1 = "{}_{}_m".format(i + 1, 2 * j)
mid_2 = "{}_{}_m".format(i + 1, 2 * j + 1)
top = "{}_{}".format(i, j)
w1, b1, w2, b2 = np.random.randn(4).tolist()
brew.conv(
m,
bottom_1,
mid_1,
dim_in=d,
dim_out=d,
kernel=3,
weight_init=("ConstantFill", {
"value": w1
}),
bias_init=("ConstantFill", {
"value": b1
}),
cudnn_state=np.random.randint(0, 3),
stride=1,
pad=1,
deterministic=1,
use_cudnn=use_cudnn,
engine=engine,
)
brew.conv(
m,
bottom_2,
mid_2,
dim_in=d,
dim_out=d,
kernel=3,
stride=1,
pad=1,
weight_init=("ConstantFill", {
"value": w2
}),
bias_init=("ConstantFill", {
"value": b2
}),
deterministic=1,
cudnn_state=np.random.randint(0, 3),
use_cudnn=use_cudnn,
engine=engine,
)
m.net.Sum([mid_1, mid_2], top)
m.net.Flatten(["0_0"], ["0_0_flat"])
m.net.SquaredL2Distance(["0_0_flat", "label"], "xent")
m.net.AveragedLoss("xent", "loss")
input_to_grad = m.AddGradientOperators(["loss"])
m.Proto().device_option.CopyFrom(gc)
m.param_init_net.Proto().device_option.CopyFrom(gc)
m.Proto().type = net_type
m.Proto().num_workers = num_workers
self.ws.run(m.param_init_net)
def run():
import numpy as np
np.random.seed(1701)
input_blobs = ["{}_{}".format(depth, j) for j in range(2**depth)]
for input_blob in input_blobs:
self.ws.create_blob(input_blob).feed(np.random.randn(
n, d, h, w).astype(np.float32),
device_option=gc)
self.ws.create_blob("label").feed(np.random.randn(
n, d * h * w).astype(np.float32),
device_option=gc)
self.ws.run(m.net)
gradients = [
self.ws.blobs[str(input_to_grad[input_blob])].fetch()
for input_blob in input_blobs
]
return gradients
outputs = [run() for _ in range(iters)]
for output in outputs[1:]:
np.testing.assert_array_equal(outputs[0], output)
np.testing.assert_allclose(np.sum(np.square(output)),
1763719461732352.0,
rtol=1e-5)
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 VGGA(order, cudnn_ws, device):
my_arg_scope = {'order': order, 'use_cudnn': True,
'cudnn_exhaustive_search': True,
'ws_nbytes_limit': str(cudnn_ws)}
model = ModelHelper(name='vgg-a', arg_scope=my_arg_scope)
conv1 = brew.conv(
model,
"data",
"conv1",
3,
64,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu1 = brew.relu(model, conv1, "conv1")
pool1 = brew.max_pool(model, relu1, "pool1", kernel=2, stride=2)
conv2 = brew.conv(
model,
pool1,
"conv2",
64,
128,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu2 = brew.relu(model, conv2, "conv2")
pool2 = brew.max_pool(model, relu2, "pool2", kernel=2, stride=2)
conv3 = brew.conv(
model,
pool2,
"conv3",
128,
256,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu3 = brew.relu(model, conv3, "conv3")
conv4 = brew.conv(
model,
relu3,
"conv4",
256,
256,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu4 = brew.relu(model, conv4, "conv4")
pool4 = brew.max_pool(model, relu4, "pool4", kernel=2, stride=2)
conv5 = brew.conv(
model,
pool4,
"conv5",
256,
512,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu5 = brew.relu(model, conv5, "conv5")
conv6 = brew.conv(
model,
relu5,
"conv6",
512,
512,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu6 = brew.relu(model, conv6, "conv6")
pool6 = brew.max_pool(model, relu6, "pool6", kernel=2, stride=2)
conv7 = brew.conv(
model,
pool6,
"conv7",
512,
512,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu7 = brew.relu(model, conv7, "conv7")
conv8 = brew.conv(
model,
relu7,
"conv8",
512,
512,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu8 = brew.relu(model, conv8, "conv8")
pool8 = brew.max_pool(model, relu8, "pool8", kernel=2, stride=2)
fcix = brew.fc(
model, pool8, "fcix", 512 * 7 * 7, 4096, ('XavierFill', {}),
('ConstantFill', {})
)
reluix = brew.relu(model, fcix, "fcix")
fcx = brew.fc(
model, reluix, "fcx", 4096, 4096, ('XavierFill', {}), ('ConstantFill', {})
)
relux = brew.relu(model, fcx, "fcx")
fcxi = brew.fc(
model, relux, "fcxi", 4096, 1000, ('XavierFill', {}), ('ConstantFill', {})
)
pred = brew.softmax(model, fcxi, "pred")
xent = model.LabelCrossEntropy([pred, "label"], "xent")
if device != 'MKL':
loss = model.AveragedLoss(xent, "loss")
return model, 231
def nate_VGG_d(model, data):
# (?) - what is default weight fillers of these layers
# (?) - should i add dropout to the fc layers?
# (*) - make sure the output dimension is kosher
# Shape here = 3x224x224
# Conv group
conv1_1 = brew.conv(model,
data,
'conv1_1',
dim_in=3,
dim_out=64,
kernel=3,
pad=1)
relu1_1 = brew.relu(model, conv1_1, 'relu1_1')
conv1_2 = brew.conv(model,
relu1_1,
'conv1_2',
dim_in=64,
dim_out=64,
kernel=3,
pad=1)
relu1_2 = brew.relu(model, conv1_2, 'relu1_2')
# Shape here = 64x224x224
# Max Pool
pool1 = brew.max_pool(model, relu1_2, 'pool1', kernel=2, stride=2)
# Shape here = 64x112x112
# Conv group
conv2_1 = brew.conv(model,
pool1,
'conv2_1',
dim_in=64,
dim_out=128,
kernel=3,
pad=1)
relu2_1 = brew.relu(model, conv2_1, 'relu2_1')
conv2_2 = brew.conv(model,
relu2_1,
'conv2_2',
dim_in=128,
dim_out=128,
kernel=3,
pad=1)
relu2_2 = brew.relu(model, conv2_2, 'relu2_2')
# Shape here = 128x112x112
# Max Pool
pool2 = brew.max_pool(model, relu2_2, 'pool2', kernel=2, stride=2)
# Shape here = 128x56x56
# Conv group
conv3_1 = brew.conv(model,
pool2,
'conv3_1',
dim_in=128,
dim_out=256,
kernel=3,
pad=1)
relu3_1 = brew.relu(model, conv3_1, 'relu3_1')
conv3_2 = brew.conv(model,
relu3_1,
'conv3_2',
dim_in=256,
dim_out=256,
kernel=3,
pad=1)
relu3_2 = brew.relu(model, conv3_2, 'relu3_2')
conv3_3 = brew.conv(model,
relu3_2,
'conv3_3',
dim_in=256,
dim_out=256,
kernel=3,
pad=1)
relu3_3 = brew.relu(model, conv3_3, 'relu3_3')
# Shape here = 256x56x56
# Max Pool
pool3 = brew.max_pool(model, relu3_3, 'pool3', kernel=2, stride=2)
# Shape here = 256x28x28
# Conv group
conv4_1 = brew.conv(model,
pool3,
'conv4_1',
dim_in=256,
dim_out=512,
kernel=3,
pad=1)
relu4_1 = brew.relu(model, conv4_1, 'relu4_1')
conv4_2 = brew.conv(model,
relu4_1,
'conv4_2',
dim_in=512,
dim_out=512,
kernel=3,
pad=1)
relu4_2 = brew.relu(model, conv4_2, 'relu4_2')
conv4_3 = brew.conv(model,
relu4_2,
'conv4_3',
dim_in=512,
dim_out=512,
kernel=3,
pad=1)
relu4_3 = brew.relu(model, conv4_3, 'relu4_3')
# Shape here = 512x28x28
# Max Pool
pool4 = brew.max_pool(model, relu4_3, 'pool4', kernel=2, stride=2)
# Shape here = 512x14x14
# Conv group
conv5_1 = brew.conv(model,
pool4,
'conv5_1',
dim_in=512,
dim_out=512,
kernel=3,
pad=1)
relu5_1 = brew.relu(model, conv5_1, 'relu5_1')
conv5_2 = brew.conv(model,
relu5_1,
'conv5_2',
dim_in=512,
dim_out=512,
kernel=3,
pad=1)
relu5_2 = brew.relu(model, conv5_2, 'relu5_2')
conv5_3 = brew.conv(model,
relu5_2,
'conv5_3',
dim_in=512,
dim_out=512,
kernel=3,
pad=1)
relu5_3 = brew.relu(model, conv5_3, 'relu5_3')
# Shape here = 512x14x14
# Max Pool
pool5 = brew.max_pool(model, relu5_3, 'pool5', kernel=2, stride=2)
# Shape here = 512x7x7
fc1 = brew.fc(model, pool5, 'fc1', dim_in=512 * 7 * 7, dim_out=4096)
relu1 = brew.relu(model, fc1, 'relu1')
# Shape here = 1x4096
fc2 = brew.fc(model, relu1, 'fc2', dim_in=4096, dim_out=4096)
relu2 = brew.relu(model, fc2, 'relu2')
# Shape here = 1x4096
fc3 = brew.fc(model, relu2, 'fc3', dim_in=4096, dim_out=11)
# Shape here = 1x11
softmax = brew.softmax(model, fc3, 'softmax')
return softmax
def create_mobilenet(
model, data, num_input_channels, num_labels, label, is_test=False
):
'''
Create residual net for smaller images (sec 4.2 of He et. al (2015))
num_groups = 'n' in the paper
'''
# conv1
brew.conv(
model,
data,
'conv1',
3,
32,
weight_init=("MSRAFill", {}),
kernel=3,
stride=2,
pad=2,
no_bias=True,
)
brew.spatial_bn(
model, 'conv1', 'conv1_spatbn', 32, epsilon=1e-3, is_test=is_test
)
brew.relu(model, 'conv1_spatbn', 'relu1')
builder = MobileNetBuilder(model, 'relu1', no_bias=True, is_test=is_test)
# block1
builder.add_simple_block(input_filters=32, output_filters=64, down_sampling=False, spatial_batch_norm=True)
# block2
builder.add_simple_block(input_filters=64, output_filters=128, down_sampling=True, spatial_batch_norm=True)
# block3
builder.add_simple_block(input_filters=128, output_filters=128, down_sampling=False, spatial_batch_norm=True)
# block4
builder.add_simple_block(input_filters=128, output_filters=256, down_sampling=True, spatial_batch_norm=True)
# block5
builder.add_simple_block(input_filters=256, output_filters=256, down_sampling=False, spatial_batch_norm=True)
# block6
builder.add_simple_block(input_filters=256, output_filters=512, down_sampling=True, spatial_batch_norm=True)
# block7-11
for i in xrange(7, 12):
builder.add_simple_block(input_filters=512, output_filters=512, down_sampling=False, spatial_batch_norm=True)
# block12
builder.add_simple_block(input_filters=512, output_filters=1024, down_sampling=True, spatial_batch_norm=True)
# block13
builder.add_simple_block(input_filters=1024, output_filters=1024, down_sampling=False, spatial_batch_norm=True)
# Final layers
brew.average_pool(
model, builder.prev_blob, 'final_avg', kernel=7, stride=7)
last_out = brew.fc(model, 'final_avg', 'last_out', 1024, num_labels)
if (label is not None):
(softmax, loss) = model.SoftmaxWithLoss(
[last_out, label],
['softmax', 'loss'],
)
return (softmax, loss)
else:
return brew.softmax(model, 'last_out', 'softmax')
def Add_Action_Tufts_Model(model,
num_classes,
image_height,
image_width,
image_channels,
is_test=0):
################################## Block 1 ############################
# Convolutional layer 1
# conv1_1 = brew.conv(model, 'data', 'conv1_1', dim_in=image_channels, dim_out=64, kernel=3, stride=2, pad=0)
# h,w = update_dims(height=image_height, width=image_width, kernel=3, stride=2, pad=0)
# ReLU layer 1
# relu1_1 = brew.relu(model, conv1_1, 'relu1_1')
# Batch normalization layer 1
# bn1_1 = brew.spatial_bn(model, relu1_1, 'bn1_1', dim_in=64, epsilon=1e-3, momentum=0.1, is_test=is_test)
# Drop out with p=0.25
# dropout1_1 = brew.dropout(model, bn1_1, 'dropout1_1', ratio=0.25, is_test=is_test)
# Convolutional layer 2
# conv1_2 = brew.conv(model, dropout1_1, 'conv1_2', dim_in=64, dim_out=64, kernel=3, stride=1, pad=0)
# h,w = update_dims(height=h, width=w, kernel=3, stride=1, pad=0)
# ReLU layer 1
# relu1_2 = brew.relu(model, conv1_2, 'relu1_2')
# Batch normalization layer 1
# bn1_2 = brew.spatial_bn(model, relu1_2, 'bn1_2', dim_in=64, epsilon=1e-3, momentum=0.1, is_test=is_test)
# Drop out with p=0.25
# dropout1_2 = brew.dropout(model, bn1_2, 'dropout1_2', ratio=0.25, is_test=is_test)
##################################### Block 2 ##########################
# Convolutional layer 3
conv2_1 = brew.conv(model,
'data',
'conv2_1',
dim_in=image_channels,
dim_out=128,
kernel=3,
stride=2,
pad=0)
h, w = update_dims(height=image_height,
width=image_width,
kernel=3,
stride=2,
pad=0)
# ReLU layer 1
relu2_1 = brew.relu(model, conv2_1, 'relu2_1')
# Batch normalization layer 1
bn2_1 = brew.spatial_bn(model,
relu2_1,
'bn2_1',
dim_in=128,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
# Drop out with p=0.25
dropout2_1 = brew.dropout(model,
bn2_1,
'dropout2_1',
ratio=0.25,
is_test=is_test)
# Convolutional layer 4
conv2_2 = brew.conv(model,
dropout2_1,
'conv2_2',
dim_in=128,
dim_out=128,
kernel=3,
stride=1,
pad=0)
h, w = update_dims(height=h, width=w, kernel=3, stride=1, pad=0)
# ReLU layer 1
relu2_2 = brew.relu(model, conv2_2, 'relu2_2')
# Batch normalization layer 1
bn2_2 = brew.spatial_bn(model,
relu2_2,
'bn2_2',
dim_in=128,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
# Drop out with p=0.25
dropout2_2 = brew.dropout(model,
bn2_2,
'dropout2_2',
ratio=0.25,
is_test=is_test)
##################################### Block 3 ############################
# Convolutional layer 5
conv3_1 = brew.conv(model,
dropout2_2,
'conv3_1',
dim_in=128,
dim_out=256,
kernel=3,
stride=2,
pad=0)
h, w = update_dims(height=h, width=w, kernel=3, stride=2, pad=0)
# ReLU layer 1
relu3_1 = brew.relu(model, conv3_1, 'relu3_1')
# Batch normalization layer 1
bn3_1 = brew.spatial_bn(model,
relu3_1,
'bn3_1',
dim_in=256,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
# Drop out with p=0.25
dropout3_1 = brew.dropout(model,
bn3_1,
'dropout3_1',
ratio=0.25,
is_test=is_test)
# Convolutional layer 4
conv3_2 = brew.conv(model,
dropout3_1,
'conv3_2',
dim_in=256,
dim_out=256,
kernel=3,
stride=1,
pad=0)
h, w = update_dims(height=h, width=w, kernel=3, stride=1, pad=0)
# ReLU layer 1
relu3_2 = brew.relu(model, conv3_2, 'relu3_2')
# Batch normalization layer 1
bn3_2 = brew.spatial_bn(model,
relu3_2,
'bn3_2',
dim_in=256,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
# Drop out with p=0.25
dropout3_2 = brew.dropout(model,
bn3_2,
'dropout3_2',
ratio=0.25,
is_test=is_test)
# Global average pooling
pool1 = brew.average_pool(model, dropout3_2, 'pool1', global_pooling=True)
# Fully connected layers
pred = brew.fc(model, pool1, 'fc1', dim_in=256, dim_out=num_classes)
# Softmax layer
softmax, loss = model.SoftmaxWithLoss([pred, 'label'], ['softmax', 'loss'])
brew.accuracy(model, [softmax, 'label'], 'accuracy')
model.net.MultiClassAccuracy([softmax, 'label'],
['accuracy_per_class', 'amount_per_class'])
return [loss]
def meta_conv_conv_topk(
model,
inputs,
dim_in,
dim_out,
kernel,
pad,
stride,
topk=5,
no_bias=False,
is_test=False,
has_relu=False,
module_seq=None, # group seq
sub_seq=None, # block seq
branch_seq=None,
conv_seq=None
):
'''
add basic conv module of resnet 50
'''
# set projection branch
if branch_seq == "2": # branch of 2 normal part
conv_name = "conv{}".format(conv_seq)
elif branch_seq == "1" : # branch of 1 projection part
conv_name = "proj"
# convolution layer
conv = brew.conv(
model,
inputs,
'group{}_block{}_{}'.format(module_seq, sub_seq, conv_name),
dim_in=dim_in,
dim_out=dim_out,
kernel=kernel,
stride=stride,
pad=pad,
no_bias=no_bias,
)
conv_hook = model.net.TopKGradHook(
conv,
'group{}_block{}_{}_hook'.format(module_seq, sub_seq, conv_name),
k=topk,
)
# spaial batchnormalization layer
bn = brew.spatial_bn(
model,
conv_hook,
'group{}_block{}_{}_bn'.format(module_seq, sub_seq, conv_name),
# conv, # in-place
dim_in = dim_out,
epsilon = 1e-05,
is_test = is_test,
)
# ReLU layer
if has_relu:
relu = brew.relu(model, bn, bn)
return relu
else:
return bn
model = model_helper.ModelHelper('xx')
def recalculateDim(size, kernel, stride, pad):
nSize = ((size - kernel + (2 * pad)) / stride) + 1
return nSize
s = 224
dims = 3, 10, 10, 10, 10, 20, 20, 20, 64, 128, 3
layer = 'data'
for i in range(10):
layer = brew.conv(model, layer, 'conv_%d' % i, dim_in=dims[i], dim_out=dims[i + 1], kernel=5, stride=2)
s = recalculateDim(s, 5, 2, 0)
layer = brew.relu(model, layer, 'relu_%d' % i)
if i % 3 == 0:
layer = brew.max_pool(model, layer, 'pool_%d' % i, kernel=2, stride=2)
s = recalculateDim(s, 2, 2, 0)
layer = brew.fc(model, layer, 'last_fc', dim_in=s*s*dims[-2], dim_out=dims[-1])
softmax = brew.softmax(model, layer, 'softmax')
def OverFeat(order, cudnn_ws):
my_arg_scope = {
'order': order,
'use_cudnn': True,
'cudnn_exhaustive_search': True,
}
if cudnn_ws:
my_arg_scope['ws_nbytes_limit'] = cudnn_ws
model = model_helper.ModelHelper(
name="overfeat",
arg_scope=my_arg_scope,
)
conv1 = brew.conv(
model,
"data",
"conv1",
3,
96,
11,
('XavierFill', {}),
('ConstantFill', {}),
stride=4,
)
relu1 = brew.relu(model, conv1, "conv1")
pool1 = brew.max_pool(model, relu1, "pool1", kernel=2, stride=2)
conv2 = brew.conv(model, pool1, "conv2", 96, 256, 5, ('XavierFill', {}),
('ConstantFill', {}))
relu2 = brew.relu(model, conv2, "conv2")
pool2 = brew.max_pool(model, relu2, "pool2", kernel=2, stride=2)
conv3 = brew.conv(
model,
pool2,
"conv3",
256,
512,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1,
)
relu3 = brew.relu(model, conv3, "conv3")
conv4 = brew.conv(
model,
relu3,
"conv4",
512,
1024,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1,
)
relu4 = brew.relu(model, conv4, "conv4")
conv5 = brew.conv(
model,
relu4,
"conv5",
1024,
1024,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1,
)
relu5 = brew.relu(model, conv5, "conv5")
pool5 = brew.max_pool(model, relu5, "pool5", kernel=2, stride=2)
fc6 = brew.fc(model, pool5, "fc6", 1024 * 6 * 6, 3072, ('XavierFill', {}),
('ConstantFill', {}))
relu6 = brew.relu(model, fc6, "fc6")
fc7 = brew.fc(model, relu6, "fc7", 3072, 4096, ('XavierFill', {}),
('ConstantFill', {}))
relu7 = brew.relu(model, fc7, "fc7")
fc8 = brew.fc(model, relu7, "fc8", 4096, 1000, ('XavierFill', {}),
('ConstantFill', {}))
pred = brew.softmax(model, fc8, "pred")
xent = model.net.LabelCrossEntropy([pred, "label"], "xent")
model.net.AveragedLoss(xent, "loss")
return model, 231
def test_convolution_sync(self, net_type, num_workers, do, engine):
m = ModelHelper(name="test_model")
n = 1
d = 2
depth = 3
iters = 5
h = 5
w = 5
workspace.ResetWorkspace()
use_cudnn = (engine == 'CUDNN')
np.random.seed(1701)
# Build a binary tree of conv layers, summing at each node.
for i in reversed(range(depth)):
for j in range(2 ** i):
bottom_1 = "{}_{}".format(i + 1, 2 * j)
bottom_2 = "{}_{}".format(i + 1, 2 * j + 1)
mid_1 = "{}_{}_m".format(i + 1, 2 * j)
mid_2 = "{}_{}_m".format(i + 1, 2 * j + 1)
top = "{}_{}".format(i, j)
w1, b1, w2, b2 = np.random.randn(4).tolist()
brew.conv(
m, bottom_1, mid_1,
dim_in=d, dim_out=d,
kernel=3,
weight_init=('ConstantFill', dict(value=w1)),
bias_init=('ConstantFill', dict(value=b1)),
cudnn_state=np.random.randint(0, 3),
stride=1,
pad=1,
deterministic=1,
use_cudnn=use_cudnn,
engine=engine)
brew.conv(
m, bottom_2, mid_2,
dim_in=d, dim_out=d,
kernel=3,
stride=1,
pad=1,
weight_init=('ConstantFill', dict(value=w2)),
bias_init=('ConstantFill', dict(value=b2)),
deterministic=1,
cudnn_state=np.random.randint(0, 3),
use_cudnn=use_cudnn,
engine=engine)
m.net.Sum([mid_1, mid_2], top)
m.net.Flatten(["0_0"], ["0_0_flat"])
m.net.SquaredL2Distance(["0_0_flat", "label"], "xent")
m.net.AveragedLoss("xent", "loss")
input_to_grad = m.AddGradientOperators(["loss"])
m.Proto().device_option.CopyFrom(do)
m.param_init_net.Proto().device_option.CopyFrom(do)
m.Proto().type = net_type
m.Proto().num_workers = num_workers
self.ws.run(m.param_init_net)
def run():
import numpy as np
np.random.seed(1701)
input_blobs = ["{}_{}".format(depth, j) for j in range(2 ** depth)]
for input_blob in input_blobs:
self.ws.create_blob(input_blob).feed(
np.random.randn(n, d, h, w).astype(np.float32),
device_option=do)
self.ws.create_blob("label").feed(
np.random.randn(n, d * h * w).astype(np.float32),
device_option=do)
self.ws.run(m.net)
gradients = [
self.ws.blobs[str(input_to_grad[input_blob])].fetch()
for input_blob in input_blobs]
return gradients
outputs = [run() for _ in range(iters)]
for output in outputs[1:]:
np.testing.assert_array_equal(outputs[0], output)
np.testing.assert_allclose(
np.sum(np.square(output)),
1763719461732352.0,
rtol=1e-5)
def create_resnet50(
model,
data,
num_input_channels,
num_labels,
label=None,
is_test=False,
no_loss=False,
no_bias=0,
conv1_kernel=7,
conv1_stride=2,
final_avg_kernel=7,
):
# conv1 + maxpool
brew.conv(
model,
data,
'conv1',
num_input_channels,
64,
weight_init=("MSRAFill", {}),
kernel=conv1_kernel,
stride=conv1_stride,
pad=3,
no_bias=no_bias
)
brew.spatial_bn(
model,
'conv1',
'conv1_spatbn_relu',
64,
epsilon=1e-3,
momentum=0.1,
is_test=is_test
)
brew.relu(model, 'conv1_spatbn_relu', 'conv1_spatbn_relu')
brew.max_pool(model, 'conv1_spatbn_relu', 'pool1', kernel=3, stride=2)
# Residual blocks...
builder = ResNetBuilder(model, 'pool1', no_bias=no_bias,
is_test=is_test, spatial_bn_mom=0.1)
# conv2_x (ref Table 1 in He et al. (2015))
builder.add_bottleneck(64, 64, 256)
builder.add_bottleneck(256, 64, 256)
builder.add_bottleneck(256, 64, 256)
# conv3_x
builder.add_bottleneck(256, 128, 512, down_sampling=True)
for _ in range(1, 4):
builder.add_bottleneck(512, 128, 512)
# conv4_x
builder.add_bottleneck(512, 256, 1024, down_sampling=True)
for _ in range(1, 6):
builder.add_bottleneck(1024, 256, 1024)
# conv5_x
builder.add_bottleneck(1024, 512, 2048, down_sampling=True)
builder.add_bottleneck(2048, 512, 2048)
builder.add_bottleneck(2048, 512, 2048)
# Final layers
final_avg = brew.average_pool(
model,
builder.prev_blob,
'final_avg',
kernel=final_avg_kernel,
stride=1,
)
# Final dimension of the "image" is reduced to 7x7
last_out = brew.fc(
model, final_avg, 'last_out_L{}'.format(num_labels), 2048, num_labels
)
if no_loss:
return last_out
# If we create model for training, use softmax-with-loss
if (label is not None):
(softmax, loss) = model.SoftmaxWithLoss(
[last_out, label],
["softmax", "loss"],
)
return (softmax, loss)
else:
# For inference, we just return softmax
return brew.softmax(model, last_out, "softmax")
def test_simple_model(self):
model = model_helper.ModelHelper(name="mnist")
# how come those inputs don't break the forward pass =.=a
workspace.FeedBlob("data",
np.random.randn(1, 3, 64, 64).astype(np.float32))
workspace.FeedBlob("label", np.random.randn(1, 1000).astype(np.int))
with core.NameScope("conv1"):
conv1 = brew.conv(model,
"data",
'conv1',
dim_in=1,
dim_out=20,
kernel=5)
# Image size: 24 x 24 -> 12 x 12
pool1 = brew.max_pool(model, conv1, 'pool1', kernel=2, stride=2)
# Image size: 12 x 12 -> 8 x 8
conv2 = brew.conv(model,
pool1,
'conv2',
dim_in=20,
dim_out=100,
kernel=5)
# Image size: 8 x 8 -> 4 x 4
pool2 = brew.max_pool(model, conv2, 'pool2', kernel=2, stride=2)
with core.NameScope("classifier"):
# 50 * 4 * 4 stands for dim_out from previous layer multiplied by the image size
fc3 = brew.fc(model, pool2, 'fc3', dim_in=100 * 4 * 4, dim_out=500)
relu = brew.relu(model, fc3, fc3)
pred = brew.fc(model, relu, 'pred', 500, 10)
softmax = brew.softmax(model, pred, 'softmax')
xent = model.LabelCrossEntropy([softmax, "label"], 'xent')
# compute the expected loss
loss = model.AveragedLoss(xent, "loss")
model.net.RunAllOnMKL()
model.param_init_net.RunAllOnMKL()
model.AddGradientOperators([loss], skip=1)
blob_name_tracker = {}
graph = tb.model_to_graph_def(
model,
blob_name_tracker=blob_name_tracker,
shapes={},
show_simplified=False,
)
#self.assertEqual(
# blob_name_tracker['GRADIENTS/conv1/conv1_b_grad'],
# 'conv1/conv1_b_grad',
#)
self.maxDiff = None
# We can't guarantee the order in which they appear, so we sort
# both before we compare them
with open('tests/expect/caffe_mnist.expect') as f:
EXPECTED_MNIST = f.read()
sep = "node {"
expected = "\n".join(
sorted(sep + "\n " + part.strip()
for part in EXPECTED_MNIST.strip().split(sep)
if part.strip()))
actual = "\n".join(
sorted(sep + "\n " + part.strip()
for part in str(graph).strip().split(sep) if part.strip()))
def OverFeat(order, cudnn_ws, device):
my_arg_scope = {'order': order, 'use_cudnn': True,
'cudnn_exhaustive_search': True,
'ws_nbytes_limit': str(cudnn_ws)}
model = ModelHelper(name='overfeat', arg_scope=my_arg_scope)
conv1 = brew.conv(
model,
"data",
"conv1",
3,
96,
11,
('XavierFill', {}),
('ConstantFill', {}),
stride=4
)
relu1 = brew.relu(model, conv1, "conv1")
pool1 = brew.max_pool(model, relu1, "pool1", kernel=2, stride=2)
conv2 = brew.conv(
model, pool1, "conv2", 96, 256, 5, ('XavierFill', {}), ('ConstantFill', {})
)
relu2 = brew.relu(model, conv2, "conv2")
pool2 = brew.max_pool(model, relu2, "pool2", kernel=2, stride=2)
conv3 = brew.conv(
model,
pool2,
"conv3",
256,
512,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu3 = brew.relu(model, conv3, "conv3")
conv4 = brew.conv(
model,
relu3,
"conv4",
512,
1024,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu4 = brew.relu(model, conv4, "conv4")
conv5 = brew.conv(
model,
relu4,
"conv5",
1024,
1024,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu5 = brew.relu(model, conv5, "conv5")
pool5 = brew.max_pool(model, relu5, "pool5", kernel=2, stride=2)
fc6 = brew.fc(
model, pool5, "fc6", 1024 * 6 * 6, 3072, ('XavierFill', {}),
('ConstantFill', {})
)
relu6 = brew.relu(model, fc6, "fc6")
fc7 = brew.fc(
model, relu6, "fc7", 3072, 4096, ('XavierFill', {}), ('ConstantFill', {})
)
relu7 = brew.relu(model, fc7, "fc7")
fc8 = brew.fc(
model, relu7, "fc8", 4096, 1000, ('XavierFill', {}), ('ConstantFill', {})
)
pred = brew.softmax(model, fc8, "pred")
xent = model.LabelCrossEntropy([pred, "label"], "xent")
if device != 'MKL':
loss = model.AveragedLoss(xent, "loss")
return model, 231
def TestModel_ops(model, data, label):
# Image size: 227 x 227 -> 224 x 224
conv1_s2 = brew.conv(model,
data,
'conv1_s2',
dim_in=3,
dim_out=32,
kernel=4,
stride=2)
conv1_s2_s = brew.spatial_bn(model,
conv1_s2,
'conv1_s2_s',
32,
epsilon=1e-3,
momentum=0.9)
conv1_dw = brew.conv(model,
conv1_s2_s,
'conv1_dw',
dim_in=32,
dim_out=32,
kernel=3,
pad=2)
# conv1_dw_s = brew.spatial_bn(model, conv1_dw, 'conv1_dw_s', 32, epsilon=1e-3, momentum=0.9)
conv1_s1 = brew.conv(model,
conv1_dw,
'conv1_s1',
dim_in=32,
dim_out=64,
kernel=1)
conv1_s1_s = brew.spatial_bn(model,
conv1_s1,
'conv1_s1_s',
64,
epsilon=1e-3,
momentum=0.9)
conv2_s2 = brew.conv(model,
conv1_s1_s,
'conv2_s2',
dim_in=64,
dim_out=64,
kernel=3,
pad=2,
stride=2)
conv2_dw = brew.conv(model,
conv2_s2,
'conv2_dw',
dim_in=64,
dim_out=128,
kernel=1)
conv2_k3 = brew.conv(model,
conv2_dw,
'conv2_k3',
dim_in=128,
dim_out=128,
kernel=3,
pad=2)
conv2_s1 = brew.conv(model,
conv2_k3,
'conv2_s1',
dim_in=128,
dim_out=128,
kernel=1)
conv3_s2 = brew.conv(model,
conv2_s1,
'conv3_s2',
dim_in=128,
dim_out=128,
kernel=3,
pad=2,
stride=2)
conv3_dw = brew.conv(model,
conv3_s2,
'conv3_dw',
dim_in=128,
dim_out=256,
kernel=1)
conv3_k3 = brew.conv(model,
conv3_dw,
'conv3_k3',
dim_in=256,
dim_out=256,
kernel=3,
pad=2)
conv3_s1 = brew.conv(model,
conv3_k3,
'conv3_s1',
dim_in=256,
dim_out=256,
kernel=1)
conv4_s2 = brew.conv(model,
conv3_s1,
'conv4_s2',
dim_in=256,
dim_out=256,
kernel=3,
pad=2,
stride=2)
conv4_dw = brew.conv(model,
conv4_s2,
'conv4_dw',
dim_in=256,
dim_out=512,
kernel=1)
#5
conv5_s2 = brew.conv(model,
conv4_dw,
'conv5_s2',
dim_in=512,
dim_out=512,
kernel=3,
pad=2)
conv5_dw = brew.conv(model,
conv5_s2,
'conv5_dw',
dim_in=512,
dim_out=512,
kernel=1)
#4
conv6_s2 = brew.conv(model,
conv5_dw,
'conv6_s2',
dim_in=512,
dim_out=512,
kernel=3,
pad=2)
conv6_dw = brew.conv(model,
conv6_s2,
'conv6_dw',
dim_in=512,
dim_out=512,
kernel=1)
#3
conv7_s2 = brew.conv(model,
conv6_dw,
'conv7_s2',
dim_in=512,
dim_out=512,
kernel=3,
pad=2)
conv7_dw = brew.conv(model,
conv7_s2,
'conv7_dw',
dim_in=512,
dim_out=512,
kernel=1)
#2
conv8_s2 = brew.conv(model,
conv7_dw,
'conv8_s2',
dim_in=512,
dim_out=512,
kernel=3,
pad=2)
conv8_dw = brew.conv(model,
conv8_s2,
'conv8_dw',
dim_in=512,
dim_out=512,
kernel=1)
#1
conv9_s2 = brew.conv(model,
conv8_dw,
'conv9_s2',
dim_in=512,
dim_out=512,
kernel=3,
pad=2)
conv9_dw = brew.conv(model,
conv9_s2,
'conv9_dw',
dim_in=512,
dim_out=512,
kernel=1)
conv10_s2 = brew.conv(model,
conv9_dw,
'conv10_s2',
dim_in=512,
dim_out=512,
kernel=3,
pad=2,
stride=2)
conv10_dw = brew.conv(model,
conv10_s2,
'conv10_dw',
dim_in=512,
dim_out=1024,
kernel=1) # out 1024
conv10_k3 = brew.conv(model,
conv10_dw,
'conv10_k3',
dim_in=1024,
dim_out=1024,
kernel=3)
conv10_s1 = brew.conv(model,
conv10_k3,
'conv10_s1',
dim_in=1024,
dim_out=1024,
kernel=1)
pool1 = brew.average_pool(model, conv10_s1, 'pool1', kernel=7, stride=7)
# pool1 = brew.max_pool(model, conv10_s1, 'pool1', kernel=7, stride=7)
fc1 = brew.fc(model, pool1, 'fc1', dim_in=1 * 1 * 1024, dim_out=num_labels)
# softmax = brew.softmax(model, fc1, 'softmax')
[softmax, loss] = model.SoftmaxWithLoss(
[fc1, label],
["softmax", "loss"],
)
return [softmax, loss]
def create_vgg(
model,
data,
num_input_channels,
num_labels,
num_layers=11,
is_test=False,
):
if num_layers == 11: # VGG configuration A
first_layers_count = 1
last_layers_count = 2
elif num_layers == 13: # VGG configuration D
first_layers_count = 2
last_layers_count = 2
elif num_layers == 16: # VGG configuration D
first_layers_count = 2
last_layers_count = 3
elif num_layers == 19: # VGG configuration E
first_layers_count = 2
last_layers_count = 4
else:
raise NotImplementedError(
"not currently supported: try one of {11, 13, 16, 19}, corresponding to VGG A, B, D, and E."
)
conv1 = brew.conv(
model,
data,
"conv1",
num_input_channels,
64,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1,
)
relu1 = brew.relu(model, conv1, "conv1")
for i in range(0, first_layers_count - 1):
conv1 = brew.conv(
model,
relu1,
"conv1{}".format(i),
64,
64,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1,
)
relu1 = brew.relu(model, conv1, "conv1{}".format(i))
pool1 = brew.max_pool(model, relu1, "pool1", kernel=2, stride=2)
conv2 = brew.conv(
model,
pool1,
"conv2",
64,
128,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1,
)
relu2 = brew.relu(model, conv2, "conv2")
for i in range(0, first_layers_count - 1):
conv2 = brew.conv(
model,
relu2,
"conv2{}".format(i),
128,
128,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1,
)
relu2 = brew.relu(model, conv2, "conv2{}".format(i))
pool2 = brew.max_pool(model, relu2, "pool2", kernel=2, stride=2)
conv3 = brew.conv(
model,
pool2,
"conv3",
128,
256,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1,
)
relu3 = brew.relu(model, conv3, "conv3")
for i in range(0, last_layers_count - 1):
conv4 = brew.conv(
model,
relu3,
"conv4{}".format(i),
256,
256,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1,
)
relu4 = brew.relu(model, conv4, "conv4{}".format(i))
pool4 = brew.max_pool(model, relu4, "pool4", kernel=2, stride=2)
conv5 = brew.conv(
model,
pool4,
"conv5",
256,
512,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1,
)
relu5 = brew.relu(model, conv5, "conv5")
for i in range(0, last_layers_count - 1):
conv6 = brew.conv(
model,
relu5,
"conv6{}".format(i),
512,
512,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1,
)
relu6 = brew.relu(model, conv6, "conv6{}".format(i))
pool6 = brew.max_pool(model, relu6, "pool6", kernel=2, stride=2)
conv7 = brew.conv(
model,
pool6,
"conv7",
512,
512,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1,
)
relu7 = brew.relu(model, conv7, "conv7")
for i in range(0, last_layers_count - 1):
conv8 = brew.conv(
model,
relu7,
"conv8{}".format(i),
512,
512,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1,
)
relu8 = brew.relu(model, conv8, "conv8{}".format(i))
pool8 = brew.max_pool(model, relu8, "pool8", kernel=2, stride=2)
fcix = brew.fc(model, pool8, "fcix", 512 * 7 * 7, 4096, ('XavierFill', {}),
('ConstantFill', {}))
reluix = brew.relu(model, fcix, "fcix")
fcx = brew.fc(model, reluix, "fcx", 4096, 4096, ('XavierFill', {}),
('ConstantFill', {}))
relux = brew.relu(model, fcx, "fcx")
fcxi = brew.fc(model, relux, "fcxi", 4096, num_labels, ('XavierFill', {}),
('ConstantFill', {}))
return fcxi
def create_unet_model(m, device_opts, is_test):
base_n_filters = 16
kernel_size = 3
pad = (kernel_size - 1) / 2
do_dropout = True
num_classes = 3
weight_init = ("MSRAFill", {})
with core.DeviceScope(device_opts):
contr_1_1 = brew.spatial_bn(m,
brew.relu(
m,
brew.conv(m,
'data',
'conv_1_1',
dim_in=num_classes,
dim_out=base_n_filters,
kernel=kernel_size,
pad=pad,
weight_init=weight_init),
'nonl_1_1'),
'contr_1_1',
dim_in=base_n_filters,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
contr_1_2 = brew.spatial_bn(m,
brew.relu(
m,
brew.conv(m,
contr_1_1,
'conv_1_2',
dim_in=base_n_filters,
dim_out=base_n_filters,
kernel=kernel_size,
pad=pad,
weight_init=weight_init),
'nonl_1_2'),
'contr_1_2',
dim_in=base_n_filters,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
pool1 = brew.max_pool(m, contr_1_2, 'pool1', kernel=2, stride=2)
contr_2_1 = brew.spatial_bn(m,
brew.relu(
m,
brew.conv(m,
pool1,
'conv_2_1',
dim_in=base_n_filters,
dim_out=base_n_filters * 2,
kernel=kernel_size,
pad=pad,
weight_init=weight_init),
'nonl_2_1'),
'contr_2_1',
dim_in=base_n_filters * 2,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
contr_2_2 = brew.spatial_bn(m,
brew.relu(
m,
brew.conv(m,
contr_2_1,
'conv_2_2',
dim_in=base_n_filters * 2,
dim_out=base_n_filters * 2,
kernel=kernel_size,
pad=pad,
weight_init=weight_init),
'nonl_2_2'),
'contr_2_2',
dim_in=base_n_filters * 2,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
pool2 = brew.max_pool(m, contr_2_2, 'pool2', kernel=2, stride=2)
contr_3_1 = brew.spatial_bn(m,
brew.relu(
m,
brew.conv(m,
pool2,
'conv_3_1',
dim_in=base_n_filters * 2,
dim_out=base_n_filters * 4,
kernel=kernel_size,
pad=pad,
weight_init=weight_init),
'nonl_3_1'),
'contr_3_1',
dim_in=base_n_filters * 4,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
contr_3_2 = brew.spatial_bn(m,
brew.relu(
m,
brew.conv(m,
contr_3_1,
'conv_3_2',
dim_in=base_n_filters * 4,
dim_out=base_n_filters * 4,
kernel=kernel_size,
pad=pad,
weight_init=weight_init),
'nonl_3_2'),
'contr_3_2',
dim_in=base_n_filters * 4,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
pool3 = brew.max_pool(m, contr_3_2, 'pool3', kernel=2, stride=2)
contr_4_1 = brew.spatial_bn(m,
brew.relu(
m,
brew.conv(m,
pool3,
'conv_4_1',
dim_in=base_n_filters * 4,
dim_out=base_n_filters * 8,
kernel=kernel_size,
pad=pad,
weight_init=weight_init),
'nonl_4_1'),
'contr_4_1',
dim_in=base_n_filters * 8,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
contr_4_2 = brew.spatial_bn(m,
brew.relu(
m,
brew.conv(m,
contr_4_1,
'conv_4_2',
dim_in=base_n_filters * 8,
dim_out=base_n_filters * 8,
kernel=kernel_size,
pad=pad,
weight_init=weight_init),
'nonl_4_2'),
'contr_4_2',
dim_in=base_n_filters * 8,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
pool4 = brew.max_pool(m, contr_4_2, 'pool4', kernel=2, stride=2)
if do_dropout:
pool4 = brew.dropout(m, pool4, 'drop', ratio=0.4)
encode_5_1 = brew.spatial_bn(m,
brew.relu(
m,
brew.conv(m,
pool4,
'conv_5_1',
dim_in=base_n_filters * 8,
dim_out=base_n_filters * 16,
kernel=kernel_size,
pad=pad,
weight_init=weight_init),
'nonl_5_1'),
'encode_5_1',
dim_in=base_n_filters * 16,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
encode_5_2 = brew.spatial_bn(m,
brew.relu(
m,
brew.conv(m,
encode_5_1,
'conv_5_2',
dim_in=base_n_filters * 16,
dim_out=base_n_filters * 16,
kernel=kernel_size,
pad=pad,
weight_init=weight_init),
'nonl_5_2'),
'encode_5_2',
dim_in=base_n_filters * 16,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
upscale5 = brew.conv_transpose(m,
encode_5_2,
'upscale5',
dim_in=base_n_filters * 16,
dim_out=base_n_filters * 16,
kernel=2,
stride=2,
weight_init=weight_init)
concat6 = brew.concat(m, [upscale5, contr_4_2], 'concat6') #, axis=1)
expand_6_1 = brew.spatial_bn(
m,
brew.relu(
m,
brew.conv(m,
concat6,
'conv_6_1',
dim_in=base_n_filters * 8 * 3,
dim_out=base_n_filters * 8,
kernel=kernel_size,
pad=pad,
weight_init=weight_init), 'nonl_6_1'),
'expand_6_1',
dim_in=base_n_filters * 8,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
expand_6_2 = brew.spatial_bn(m,
brew.relu(
m,
brew.conv(m,
expand_6_1,
'conv_6_2',
dim_in=base_n_filters * 8,
dim_out=base_n_filters * 8,
kernel=kernel_size,
pad=pad,
weight_init=weight_init),
'nonl_6_2'),
'expand_6_2',
dim_in=base_n_filters * 8,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
upscale6 = brew.conv_transpose(m,
expand_6_2,
'upscale6',
dim_in=base_n_filters * 8,
dim_out=base_n_filters * 8,
kernel=2,
stride=2,
weight_init=weight_init)
concat7 = brew.concat(m, [upscale6, contr_3_2], 'concat7')
expand_7_1 = brew.spatial_bn(
m,
brew.relu(
m,
brew.conv(m,
concat7,
'conv_7_1',
dim_in=base_n_filters * 4 * 3,
dim_out=base_n_filters * 4,
kernel=kernel_size,
pad=pad,
weight_init=weight_init), 'nonl_7_1'),
'expand_7_1',
dim_in=base_n_filters * 4,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
expand_7_2 = brew.spatial_bn(m,
brew.relu(
m,
brew.conv(m,
expand_7_1,
'conv_7_2',
dim_in=base_n_filters * 4,
dim_out=base_n_filters * 4,
kernel=kernel_size,
pad=pad,
weight_init=weight_init),
'nonl_7_2'),
'expand_7_2',
dim_in=base_n_filters * 4,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
upscale7 = brew.conv_transpose(m,
expand_7_2,
'upscale7',
dim_in=base_n_filters * 4,
dim_out=base_n_filters * 4,
kernel=2,
stride=2,
weight_init=weight_init)
concat8 = brew.concat(m, [upscale7, contr_2_2], 'concat8')
expand_8_1 = brew.spatial_bn(
m,
brew.relu(
m,
brew.conv(m,
concat8,
'conv_8_1',
dim_in=base_n_filters * 2 * 3,
dim_out=base_n_filters * 2,
kernel=kernel_size,
pad=pad,
weight_init=weight_init), 'nonl_8_1'),
'expand_8_1',
dim_in=base_n_filters * 2,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
expand_8_2 = brew.spatial_bn(m,
brew.relu(
m,
brew.conv(m,
expand_8_1,
'conv_8_2',
dim_in=base_n_filters * 2,
dim_out=base_n_filters * 2,
kernel=kernel_size,
pad=pad,
weight_init=weight_init),
'nonl_8_2'),
'expand_8_2',
dim_in=base_n_filters * 2,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
upscale8 = brew.conv_transpose(m,
expand_8_2,
'upscale8',
dim_in=base_n_filters * 2,
dim_out=base_n_filters * 2,
kernel=2,
stride=2,
weight_init=weight_init)
concat9 = brew.concat(m, [upscale8, contr_1_2], 'concat9')
expand_9_1 = brew.spatial_bn(m,
brew.relu(
m,
brew.conv(m,
concat9,
'conv_9_1',
dim_in=base_n_filters * 3,
dim_out=base_n_filters,
kernel=kernel_size,
pad=pad,
weight_init=weight_init),
'nonl_9_1'),
'expand_9_1',
dim_in=base_n_filters,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
expand_9_2 = brew.spatial_bn(m,
brew.relu(
m,
brew.conv(m,
expand_9_1,
'conv_9_2',
dim_in=base_n_filters,
dim_out=base_n_filters,
kernel=kernel_size,
pad=pad,
weight_init=weight_init),
'nonl_9_2'),
'expand_9_2',
dim_in=base_n_filters,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
output_segmentation = brew.conv(m,
expand_9_2,
'output_segmentation',
dim_in=base_n_filters,
dim_out=num_classes,
kernel=1,
pad=0,
stride=1,
weight_init=weight_init)
m.net.AddExternalOutput(output_segmentation)
output_sigmoid = m.Sigmoid(output_segmentation, 'output_sigmoid')
m.net.AddExternalOutput(output_sigmoid)
return output_segmentation
def alexnet():
model = ModelHelper(name="r", arg_scope={"order": "NCHW", "is_test": True})
conv1 = brew.conv(
model,
"data",
"conv1",
3,
64,
11, ('XavierFill', {}), ('ConstantFill', {}),
stride=4,
pad=2
)
relu1 = brew.relu(model, conv1, "conv1")
pool1 = brew.max_pool(model, relu1, "pool1", kernel=3, stride=2, pad=0,
legacy_pad=3)
lrn1 = brew.lrn(
model, pool1, "pool1_lrn", size=5, alpha=1.0e-4, beta=0.75, bias=1.0)
conv2 = brew.conv(
model,
lrn1,
"conv2",
64,
192,
5,
('XavierFill', {}),
('ConstantFill', {}),
pad=2
)
relu2 = brew.relu(model, conv2, "conv2")
pool2 = brew.max_pool(model, relu2, "pool2", kernel=3, stride=2)
lrn2 = brew.lrn(
model, pool2, "pool2_lrn", size=5, alpha=1.0e-4, beta=0.75, bias=1.0)
conv3 = brew.conv(
model,
lrn2,
"conv3",
192,
384,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu3 = brew.relu(model, conv3, "conv3")
conv4 = brew.conv(
model,
relu3,
"conv4",
384,
256,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu4 = brew.relu(model, conv4, "conv4")
conv5 = brew.conv(
model,
relu4,
"conv5",
256,
256,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu5 = brew.relu(model, conv5, "conv5")
pool5 = brew.max_pool(model, relu5, "pool5", kernel=3, stride=2)
fc6 = brew.fc(
model,
pool5, "fc6", 256 * 6 * 6, 4096, ('XavierFill', {}),
('ConstantFill', {})
)
relu6 = brew.relu(model, fc6, "fc6")
fc7 = brew.fc(
model, relu6, "fc7", 4096, 4096, ('XavierFill', {}), ('ConstantFill', {})
)
relu7 = brew.relu(model, fc7, "fc7")
drop7 = brew.dropout(model, relu7, "fc7_dropout", is_test=1, ratio=0.5)
fc8 = brew.fc(
model, drop7, "fc8", 4096, 1000, ('XavierFill', {}), ('ConstantFill', {})
)
relu8 = brew.relu(model, fc8, "fc8")
_ = brew.dropout(model, relu8, "fc8_dropout", is_test=1, ratio=0.5)
return model, [(1, 3, 224, 224)]
def create_alexnetv0(
model,
data,
num_input_channels,
num_labels,
is_test=False,
):
conv1 = brew.conv(
model,
data,
"conv1",
num_input_channels, # dim_in
96, # dim_out
11, # kernel
('XavierFill', {}),
('ConstantFill', {}),
stride=4,
pad=2)
relu1 = brew.relu(model, conv1, "conv1")
norm1 = brew.lrn(model, relu1, "norm1", size=5, alpha=0.0001, beta=0.75)
pool1 = brew.max_pool(model, norm1, "pool1", kernel=3, stride=2)
conv2 = brew.conv(model,
pool1,
"conv2",
96,
256,
5, ('XavierFill', {}), ('ConstantFill', {}),
pad=2)
relu2 = brew.relu(model, conv2, "conv2")
norm2 = brew.lrn(model, relu2, "norm2", size=5, alpha=0.0001, beta=0.75)
pool2 = brew.max_pool(model, norm2, "pool2", kernel=3, stride=2)
conv3 = brew.conv(model,
pool2,
"conv3",
256,
384,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu3 = brew.relu(model, conv3, "conv3")
conv4 = brew.conv(model,
relu3,
"conv4",
384,
384,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu4 = brew.relu(model, conv4, "conv4")
conv5 = brew.conv(model,
relu4,
"conv5",
384,
256,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu5 = brew.relu(model, conv5, "conv5")
pool5 = brew.max_pool(model, relu5, "pool5", kernel=3, stride=2)
fc6 = brew.fc(model, pool5, "fc6", 256 * 6 * 6, 4096, ('XavierFill', {}),
('ConstantFill', {}))
relu6 = brew.relu(model, fc6, "fc6")
dropout1 = brew.dropout(model,
relu6,
'dropout1',
ratio=0.5,
is_test=is_test)
fc7 = brew.fc(model, dropout1, "fc7", 4096, 4096, ('XavierFill', {}),
('ConstantFill', {}))
relu7 = brew.relu(model, fc7, "fc7")
dropout2 = brew.dropout(model,
relu7,
'dropout2',
ratio=0.5,
is_test=is_test)
fc8 = brew.fc(model, dropout2, "fc8", 4096, num_labels, ('XavierFill', {}),
('ConstantFill', {}))
return fc8
def vgg16_model(self, model, data, precision):
initializer = Initializer if precision == 'fp32' else pFP16Initializer
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=initializer,
BiasInitializer=initializer,
enable_tensor_core=True):
conv1_1 = brew.conv(model,
data,
'conv1_1',
dim_in=3,
dim_out=64,
kernel=3,
pad=1)
conv1_1 = brew.relu(model, conv1_1, conv1_1)
conv1_2 = brew.conv(model,
conv1_1,
'conv1_2',
dim_in=64,
dim_out=64,
kernel=3,
pad=1)
conv1_2 = brew.relu(model, conv1_2, conv1_2)
pool1 = brew.max_pool(model, conv1_2, 'pool1', kernel=2, stride=2)
conv2_1 = brew.conv(model,
pool1,
'conv2_1',
dim_in=64,
dim_out=128,
kernel=3,
pad=1)
conv2_1 = brew.relu(model, conv2_1, conv2_1)
conv2_2 = brew.conv(model,
conv2_1,
'conv2_2',
dim_in=128,
dim_out=128,
kernel=3,
pad=1)
conv2_2 = brew.relu(model, conv2_2, conv2_2)
pool2 = brew.max_pool(model, conv2_2, 'pool2', kernel=2, stride=2)
conv3_1 = brew.conv(model,
pool2,
'conv3_1',
dim_in=128,
dim_out=256,
kernel=3,
pad=1)
conv3_1 = brew.relu(model, conv3_1, conv3_1)
conv3_2 = brew.conv(model,
conv3_1,
'conv3_2',
dim_in=256,
dim_out=256,
kernel=3,
pad=1)
conv3_2 = brew.relu(model, conv3_2, conv3_2)
conv3_3 = brew.conv(model,
conv3_2,
'conv3_3',
dim_in=256,
dim_out=256,
kernel=3,
pad=1)
conv3_3 = brew.relu(model, conv3_3, conv3_3)
pool3 = brew.max_pool(model, conv3_3, 'pool3', kernel=2, stride=2)
conv4_1 = brew.conv(model,
pool3,
'conv4_1',
dim_in=256,
dim_out=512,
kernel=3,
pad=1)
conv4_1 = brew.relu(model, conv4_1, conv4_1)
conv4_2 = brew.conv(model,
conv4_1,
'conv4_2',
dim_in=512,
dim_out=512,
kernel=3,
pad=1)
conv4_2 = brew.relu(model, conv4_2, conv4_2)
conv4_3 = brew.conv(model,
conv4_2,
'conv4_3',
dim_in=512,
dim_out=512,
kernel=3,
pad=1)
conv4_3 = brew.relu(model, conv4_3, conv4_3)
pool4 = brew.max_pool(model, conv4_3, 'pool4', kernel=2, stride=2)
conv5_1 = brew.conv(model,
pool4,
'conv5_1',
dim_in=512,
dim_out=512,
kernel=3,
pad=1)
conv5_1 = brew.relu(model, conv5_1, conv5_1)
conv5_2 = brew.conv(model,
conv5_1,
'conv5_2',
dim_in=512,
dim_out=512,
kernel=3,
pad=1)
conv5_2 = brew.relu(model, conv5_2, conv5_2)
conv5_3 = brew.conv(model,
conv5_2,
'conv5_3',
dim_in=512,
dim_out=512,
kernel=3,
pad=1)
conv5_3 = brew.relu(model, conv5_3, conv5_3)
pool5 = brew.max_pool(model, conv5_3, 'pool5', kernel=2, stride=2)
fc6 = brew.fc(model, pool5, 'fc6', dim_in=25088, dim_out=4096)
fc6 = brew.relu(model, fc6, fc6)
fc7 = brew.fc(model, fc6, 'fc7', dim_in=4096, dim_out=4096)
fc7 = brew.relu(model, fc7, fc7)
pred = brew.fc(model, fc7, 'pred', 4096, 1000)
softmax = brew.softmax(model, pred, 'softmax')
return softmax
def alexnet():
model = ModelHelper(name="r", arg_scope={"order": "NCHW", "is_test": True})
conv1 = brew.conv(
model,
"data",
"conv1",
3,
64,
11, ('XavierFill', {}), ('ConstantFill', {}),
stride=4,
pad=2
)
relu1 = brew.relu(model, conv1, "conv1")
pool1 = brew.max_pool(model, relu1, "pool1", kernel=3, stride=2, pad=0,
legacy_pad=3)
lrn1 = brew.lrn(
model, pool1, "pool1_lrn", size=5, alpha=1.0e-4, beta=0.75, bias=1.0)
conv2 = brew.conv(
model,
lrn1,
"conv2",
64,
192,
5,
('XavierFill', {}),
('ConstantFill', {}),
pad=2
)
relu2 = brew.relu(model, conv2, "conv2")
pool2 = brew.max_pool(model, relu2, "pool2", kernel=3, stride=2)
lrn2 = brew.lrn(
model, pool2, "pool2_lrn", size=5, alpha=1.0e-4, beta=0.75, bias=1.0)
conv3 = brew.conv(
model,
lrn2,
"conv3",
192,
384,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu3 = brew.relu(model, conv3, "conv3")
conv4 = brew.conv(
model,
relu3,
"conv4",
384,
256,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu4 = brew.relu(model, conv4, "conv4")
conv5 = brew.conv(
model,
relu4,
"conv5",
256,
256,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu5 = brew.relu(model, conv5, "conv5")
pool5 = brew.max_pool(model, relu5, "pool5", kernel=3, stride=2)
fc6 = brew.fc(
model,
pool5, "fc6", 256 * 6 * 6, 4096, ('XavierFill', {}),
('ConstantFill', {})
)
relu6 = brew.relu(model, fc6, "fc6")
fc7 = brew.fc(
model, relu6, "fc7", 4096, 4096, ('XavierFill', {}), ('ConstantFill', {})
)
relu7 = brew.relu(model, fc7, "fc7")
drop7 = brew.dropout(model, relu7, "fc7_dropout", is_test=1, ratio=0.5)
fc8 = brew.fc(
model, drop7, "fc8", 4096, 1000, ('XavierFill', {}), ('ConstantFill', {})
)
relu8 = brew.relu(model, fc8, "fc8")
brew.dropout(model, relu8, "fc8_dropout", is_test=1, ratio=0.5)
return model, [(1, 3, 224, 224)]
def __init__(self):
self.Init_Net = None
self.Temple_Net = None
self.Track_Net = None
self.Net_Enum = [
'init', 'temple', 'track', 'adjust', 'CorrelationFilter'
]
curinput = 'data'
for net in self.Net_Enum:
model = model_helper.ModelHelper(
name=net,
arg_scope=cnn_arg_scope,
)
if net == 'init' or net == 'temple' or net == 'track':
p = brew.conv(model,
'data',
'conv_1',
3,
192,
kernel=11,
stride=2)
p = brew.spatial_bn(model,
'conv_1',
'bn_1',
192,
is_test=True,
epsilon=1e-5,
momentum=0.1)
p = brew.relu(model, 'bn_1', 'bn_1')
p = brew.max_pool(model, 'bn_1', 'pool_1', kernel=3, stride=2)
p = brew.conv(model, 'pool_1', 'conv_2', 192, 512, kernel=5)
p = brew.spatial_bn(model,
'conv_2',
'bn_2',
512,
is_test=True,
epsilon=1e-5,
momentum=0.1)
p = brew.relu(model, 'bn_2', 'bn_2')
p = brew.max_pool(model, 'bn_2', 'pool_2', kernel=3, stride=2)
p = brew.conv(model, 'pool_2', 'conv_3', 512, 768, kernel=3)
p = brew.spatial_bn(model,
'conv_3',
'bn_3',
768,
is_test=True,
epsilon=1e-5,
momentum=0.1)
p = brew.relu(model, 'bn_3', 'bn_3')
p = brew.conv(model, 'bn_3', 'conv_4', 768, 768, kernel=3)
p = brew.spatial_bn(model,
'conv_4',
'bn_4',
768,
is_test=True,
epsilon=1e-5,
momentum=0.1)
p = brew.relu(model, 'bn_4', 'bn_4')
p = brew.conv(model, 'bn_4', 'conv_5', 768, 512, kernel=3)
p = brew.spatial_bn(model,
'conv_5',
'feature_out',
512,
is_test=True,
epsilon=1e-5,
momentum=0.1)
#rpn sub networks
if net == 'init':
p = brew.conv(model,
'feature_out',
'conv_r1',
512,
512 * 4 * 5,
kernel=3)
p = brew.conv(model,
'feature_out',
'conv_r2',
512,
512,
kernel=3)
p = brew.conv(model,
'feature_out',
'conv_cls1',
512,
512 * 2 * 5,
kernel=3)
p = brew.conv(model,
'feature_out',
'conv_cls2',
512,
512,
kernel=3)
self.add_inference_inputs(model)
self.Init_Net = model.param_init_net
self.InitModel = model
elif net == 'temple':
p = brew.conv(model,
'feature_out',
'conv_r1',
512,
512 * 4 * 5,
kernel=3)
p = brew.conv(model,
'feature_out',
'conv_cls1',
512,
512 * 2 * 5,
kernel=3)
self.add_inference_inputs(model)
self.Temple_Net = model.net
self.TempleModel = model
elif net == 'track':
p = brew.conv(model,
'feature_out',
'conv_r2',
512,
512,
kernel=3)
p = brew.conv(model,
'feature_out',
'conv_cls2',
512,
512,
kernel=3)
self.add_inference_inputs(model)
self.Track_Net = model.net
self.TrackModel = model
elif net == 'adjust':
p = brew.conv(model,
'conv_conv_r2',
'r2_out',
4 * 5,
5 * 4,
kernel=1)
self.add_inference_inputs(model)
self.addJustModel = model
elif net == 'CorrelationFilter':
p = brew.conv(model,
'conv_r2',
'conv_conv_r2',
4 * 5,
5 * 4,
kernel=1)
self.add_inference_inputs(model)
self.CorrelationFilterModel = model
def AlexNet(order, cudnn_ws):
my_arg_scope = {
'order': order,
'use_cudnn': True,
'cudnn_exhaustive_search': True,
}
if cudnn_ws:
my_arg_scope['ws_nbytes_limit'] = cudnn_ws
model = model_helper.ModelHelper(
name="alexnet",
arg_scope=my_arg_scope,
)
conv1 = brew.conv(
model,
"data",
"conv1",
3,
64,
11, ('XavierFill', {}), ('ConstantFill', {}),
stride=4,
pad=2
)
relu1 = brew.relu(model, conv1, "conv1")
pool1 = brew.max_pool(model, relu1, "pool1", kernel=3, stride=2)
conv2 = brew.conv(
model,
pool1,
"conv2",
64,
192,
5,
('XavierFill', {}),
('ConstantFill', {}),
pad=2
)
relu2 = brew.relu(model, conv2, "conv2")
pool2 = brew.max_pool(model, relu2, "pool2", kernel=3, stride=2)
conv3 = brew.conv(
model,
pool2,
"conv3",
192,
384,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu3 = brew.relu(model, conv3, "conv3")
conv4 = brew.conv(
model,
relu3,
"conv4",
384,
256,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu4 = brew.relu(model, conv4, "conv4")
conv5 = brew.conv(
model,
relu4,
"conv5",
256,
256,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu5 = brew.relu(model, conv5, "conv5")
pool5 = brew.max_pool(model, relu5, "pool5", kernel=3, stride=2)
fc6 = brew.fc(
model,
pool5, "fc6", 256 * 6 * 6, 4096, ('XavierFill', {}),
('ConstantFill', {})
)
relu6 = brew.relu(model, fc6, "fc6")
fc7 = brew.fc(
model, relu6, "fc7", 4096, 4096, ('XavierFill', {}), ('ConstantFill', {})
)
relu7 = brew.relu(model, fc7, "fc7")
fc8 = brew.fc(
model, relu7, "fc8", 4096, 1000, ('XavierFill', {}), ('ConstantFill', {})
)
pred = brew.softmax(model, fc8, "pred")
xent = model.net.LabelCrossEntropy([pred, "label"], "xent")
model.net.AveragedLoss(xent, "loss")
return model, 224
def create_alexnet(
model,
data,
num_input_channels,
num_labels,
is_test=False,
):
conv1 = brew.conv(
model,
data,
"conv1",
num_input_channels, # dim_in
64, # dim_out
11, # kernel
('XavierFill', {}),
('ConstantFill', {}),
stride=4,
pad=2)
relu1 = brew.relu(model, conv1, "conv1")
norm1 = brew.lrn(model, relu1, "norm1", size=5, alpha=0.0001, beta=0.75)
pool1 = brew.max_pool(model, norm1, "pool1", kernel=3, stride=2)
conv2 = brew.conv(model,
pool1,
"conv2",
64,
192,
5, ('XavierFill', {}), ('ConstantFill', {}),
pad=2)
relu2 = brew.relu(model, conv2, "conv2")
norm2 = brew.lrn(model, relu2, "norm2", size=5, alpha=0.0001, beta=0.75)
pool2 = brew.max_pool(model, norm2, "pool2", kernel=3, stride=2)
conv3 = brew.conv(model,
pool2,
"conv3",
192,
384,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu3 = brew.relu(model, conv3, "conv3")
conv4 = brew.conv(model,
relu3,
"conv4",
384,
256,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu4 = brew.relu(model, conv4, "conv4")
conv5 = brew.conv(model,
relu4,
"conv5",
256,
256,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu5 = brew.relu(model, conv5, "conv5")
pool5 = brew.max_pool(model, relu5, "pool5", kernel=3, stride=2)
fc6 = brew.fc(model, pool5, "fc6", 256 * 6 * 6, 4096, ('XavierFill', {}),
('ConstantFill', {}))
relu6 = brew.relu(model, fc6, "fc6")
dropout1 = brew.dropout(model,
relu6,
'dropout1',
ratio=0.5,
is_test=is_test)
fc7 = brew.fc(model, dropout1, "fc7", 4096, 4096, ('XavierFill', {}),
('ConstantFill', {}))
relu7 = brew.relu(model, fc7, "fc7")
dropout2 = brew.dropout(model,
relu7,
'dropout2',
ratio=0.5,
is_test=is_test)
fc8 = brew.fc(model, dropout2, "fc8", 4096, num_labels, ('XavierFill', {}),
('ConstantFill', {}))
# pred = brew.softmax(model, fc8, "pred")
# xent = model.net.LabelCrossEntropy([pred, "label"], "xent")
# model.net.AveragedLoss(xent, "loss")
return fc8
def VGGA(order, cudnn_ws, device):
my_arg_scope = {
'order': order,
'use_cudnn': True,
'cudnn_exhaustive_search': True,
'ws_nbytes_limit': str(cudnn_ws)
}
model = ModelHelper(name='vgg-a', arg_scope=my_arg_scope)
conv1 = brew.conv(model,
"data",
"conv1",
3,
64,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu1 = brew.relu(model, conv1, "conv1")
pool1 = brew.max_pool(model, relu1, "pool1", kernel=2, stride=2)
conv2 = brew.conv(model,
pool1,
"conv2",
64,
128,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu2 = brew.relu(model, conv2, "conv2")
pool2 = brew.max_pool(model, relu2, "pool2", kernel=2, stride=2)
conv3 = brew.conv(model,
pool2,
"conv3",
128,
256,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu3 = brew.relu(model, conv3, "conv3")
conv4 = brew.conv(model,
relu3,
"conv4",
256,
256,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu4 = brew.relu(model, conv4, "conv4")
pool4 = brew.max_pool(model, relu4, "pool4", kernel=2, stride=2)
conv5 = brew.conv(model,
pool4,
"conv5",
256,
512,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu5 = brew.relu(model, conv5, "conv5")
conv6 = brew.conv(model,
relu5,
"conv6",
512,
512,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu6 = brew.relu(model, conv6, "conv6")
pool6 = brew.max_pool(model, relu6, "pool6", kernel=2, stride=2)
conv7 = brew.conv(model,
pool6,
"conv7",
512,
512,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu7 = brew.relu(model, conv7, "conv7")
conv8 = brew.conv(model,
relu7,
"conv8",
512,
512,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu8 = brew.relu(model, conv8, "conv8")
pool8 = brew.max_pool(model, relu8, "pool8", kernel=2, stride=2)
fcix = brew.fc(model, pool8, "fcix", 512 * 7 * 7, 4096, ('XavierFill', {}),
('ConstantFill', {}))
reluix = brew.relu(model, fcix, "fcix")
fcx = brew.fc(model, reluix, "fcx", 4096, 4096, ('XavierFill', {}),
('ConstantFill', {}))
relux = brew.relu(model, fcx, "fcx")
fcxi = brew.fc(model, relux, "fcxi", 4096, 1000, ('XavierFill', {}),
('ConstantFill', {}))
pred = brew.softmax(model, fcxi, "pred")
xent = model.LabelCrossEntropy([pred, "label"], "xent")
if device != 'MKL':
loss = model.AveragedLoss(xent, "loss")
return model, 231
def Inception(order, cudnn_ws, device):
my_arg_scope = {
'order': order,
'use_cudnn': True,
'cudnn_exhaustive_search': True,
'ws_nbytes_limit': str(cudnn_ws)
}
model = ModelHelper(name="inception", arg_scope=my_arg_scope)
conv1 = brew.conv(model,
"data",
"conv1",
3,
64,
7, ('XavierFill', {}), ('ConstantFill', {}),
stride=2,
pad=3)
relu1 = brew.relu(model, conv1, "conv1")
pool1 = brew.max_pool(model, relu1, "pool1", kernel=3, stride=2, pad=1)
conv2a = brew.conv(model, pool1, "conv2a", 64, 64, 1, ('XavierFill', {}),
('ConstantFill', {}))
conv2a = brew.relu(model, conv2a, conv2a)
conv2 = brew.conv(model,
conv2a,
"conv2",
64,
192,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu2 = brew.relu(model, conv2, "conv2")
pool2 = brew.max_pool(model, relu2, "pool2", kernel=3, stride=2, pad=1)
# Inception modules
inc3 = _InceptionModule(model, pool2, 192, "inc3", 64, [96, 128], [16, 32],
32)
inc4 = _InceptionModule(model, inc3, 256, "inc4", 128, [128, 192],
[32, 96], 64)
pool5 = brew.max_pool(model, inc4, "pool5", kernel=3, stride=2, pad=1)
inc5 = _InceptionModule(model, pool5, 480, "inc5", 192, [96, 208],
[16, 48], 64)
inc6 = _InceptionModule(model, inc5, 512, "inc6", 160, [112, 224],
[24, 64], 64)
inc7 = _InceptionModule(model, inc6, 512, "inc7", 128, [128, 256],
[24, 64], 64)
inc8 = _InceptionModule(model, inc7, 512, "inc8", 112, [144, 288],
[32, 64], 64)
inc9 = _InceptionModule(model, inc8, 528, "inc9", 256, [160, 320],
[32, 128], 128)
pool9 = brew.max_pool(model, inc9, "pool9", kernel=3, stride=2, pad=1)
inc10 = _InceptionModule(model, pool9, 832, "inc10", 256, [160, 320],
[32, 128], 128)
inc11 = _InceptionModule(model, inc10, 832, "inc11", 384, [192, 384],
[48, 128], 128)
pool11 = brew.average_pool(model, inc11, "pool11", kernel=7, stride=1)
fc = brew.fc(model, pool11, "fc", 1024, 1000, ('XavierFill', {}),
('ConstantFill', {}))
# It seems that Soumith's benchmark does not have softmax on top
# for Inception. We will add it anyway so we can have a proper
# backward pass.
pred = brew.softmax(model, fc, "pred")
xent = model.LabelCrossEntropy([pred, "label"], "xent")
if device != 'MKL':
loss = model.AveragedLoss(xent, "loss")
return model, 224
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 Inception(model, loss_scale, dtype='float'):
initializer = (PseudoFP16Initializer
if dtype == 'float16' else Initializer)
with brew.arg_scope(
[brew.conv, brew.fc],
WeightInitializer=initializer,
BiasInitializer=initializer,
):
conv1 = brew.conv(
model,
"data",
"conv1",
3,
64,
7,
('XavierFill', {}),
('ConstantFill', {}),
stride=2,
pad=3,
)
relu1 = brew.relu(model, conv1, "conv1")
pool1 = brew.max_pool(model, relu1, "pool1", kernel=3, stride=2, pad=1)
conv2a = brew.conv(model, pool1, "conv2a", 64, 64, 1,
('XavierFill', {}), ('ConstantFill', {}))
conv2a = brew.relu(model, conv2a, conv2a)
conv2 = brew.conv(
model,
conv2a,
"conv2",
64,
192,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1,
)
relu2 = brew.relu(model, conv2, "conv2")
pool2 = brew.max_pool(model, relu2, "pool2", kernel=3, stride=2, pad=1)
# Inception modules
inc3 = _InceptionModule(model, pool2, 192, "inc3", 64, [96, 128],
[16, 32], 32)
inc4 = _InceptionModule(model, inc3, 256, "inc4", 128, [128, 192],
[32, 96], 64)
pool5 = brew.max_pool(model, inc4, "pool5", kernel=3, stride=2, pad=1)
inc5 = _InceptionModule(model, pool5, 480, "inc5", 192, [96, 208],
[16, 48], 64)
inc6 = _InceptionModule(model, inc5, 512, "inc6", 160, [112, 224],
[24, 64], 64)
inc7 = _InceptionModule(model, inc6, 512, "inc7", 128, [128, 256],
[24, 64], 64)
inc8 = _InceptionModule(model, inc7, 512, "inc8", 112, [144, 288],
[32, 64], 64)
inc9 = _InceptionModule(model, inc8, 528, "inc9", 256, [160, 320],
[32, 128], 128)
pool9 = brew.max_pool(model, inc9, "pool9", kernel=3, stride=2, pad=1)
inc10 = _InceptionModule(model, pool9, 832, "inc10", 256, [160, 320],
[32, 128], 128)
inc11 = _InceptionModule(model, inc10, 832, "inc11", 384, [192, 384],
[48, 128], 128)
pool11 = brew.average_pool(model, inc11, "pool11", kernel=7, stride=1)
fc = brew.fc(model, pool11, "fc", 1024, 1000, ('XavierFill', {}),
('ConstantFill', {}))
if dtype == 'float16':
fc = model.net.HalfToFloat(fc, fc + '_fp32')
# It seems that Soumith's benchmark does not have softmax on top
# for Inception. We will add it anyway so we can have a proper
# backward pass.
pred = brew.softmax(model, fc, "pred")
xent = model.net.LabelCrossEntropy([pred, "label"], "xent")
loss = model.net.AveragedLoss(xent, "loss")
return [loss]
def Add_Original_CIFAR10_Model(model, data, num_classes, image_height,
image_width, image_channels):
# Convolutional layer 1
conv1 = brew.conv(model,
data,
'conv1',
dim_in=image_channels,
dim_out=32,
kernel=5,
stride=1,
pad=2)
h, w = update_dims(height=image_height,
width=image_width,
kernel=5,
stride=1,
pad=2)
# Pooling layer 1
pool1 = brew.max_pool(model, conv1, 'pool1', kernel=3, stride=2)
h, w = update_dims(height=h, width=w, kernel=3, stride=2, pad=0)
# ReLU layer 1
relu1 = brew.relu(model, pool1, 'relu1')
# Convolutional layer 2
conv2 = brew.conv(model,
relu1,
'conv2',
dim_in=32,
dim_out=32,
kernel=5,
stride=1,
pad=2)
h, w = update_dims(height=h, width=w, kernel=5, stride=1, pad=2)
# ReLU layer 2
relu2 = brew.relu(model, conv2, 'relu2')
# Pooling layer 1
pool2 = brew.average_pool(model, relu2, 'pool2', kernel=3, stride=2)
h, w = update_dims(height=h, width=w, kernel=3, stride=2, pad=0)
# Convolutional layer 3
conv3 = brew.conv(model,
pool2,
'conv3',
dim_in=32,
dim_out=64,
kernel=5,
stride=1,
pad=2)
h, w = update_dims(height=h, width=w, kernel=5, stride=1, pad=2)
# ReLU layer 3
relu3 = brew.relu(model, conv3, 'relu3')
# Pooling layer 3
pool3 = brew.average_pool(model, relu3, 'pool3', kernel=3, stride=2)
h, w = update_dims(height=h, width=w, kernel=3, stride=2, pad=0)
# Fully connected layers
fc1 = brew.fc(model, pool3, 'fc1', dim_in=64 * h * w, dim_out=64)
fc2 = brew.fc(model, fc1, 'fc2', dim_in=64, dim_out=num_classes)
# Softmax layer
softmax = brew.softmax(model, fc2, 'softmax')
return softmax
def Inception(order, cudnn_ws, device):
my_arg_scope = {'order': order, 'use_cudnn': True,
'cudnn_exhaustive_search': True,
'ws_nbytes_limit': str(cudnn_ws)}
model = ModelHelper(name="inception", arg_scope=my_arg_scope)
conv1 = brew.conv(
model,
"data",
"conv1",
3,
64,
7,
('XavierFill', {}),
('ConstantFill', {}),
stride=2,
pad=3
)
relu1 = brew.relu(model, conv1, "conv1")
pool1 = brew.max_pool(model, relu1, "pool1", kernel=3, stride=2, pad=1)
conv2a = brew.conv(
model, pool1, "conv2a", 64, 64, 1,
('XavierFill', {}), ('ConstantFill', {})
)
conv2a = brew.relu(model, conv2a, conv2a)
conv2 = brew.conv(
model,
conv2a,
"conv2",
64,
192,
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
relu2 = brew.relu(model, conv2, "conv2")
pool2 = brew.max_pool(model, relu2, "pool2", kernel=3, stride=2, pad=1)
# Inception modules
inc3 = _InceptionModule(
model, pool2, 192, "inc3", 64, [96, 128], [16, 32], 32
)
inc4 = _InceptionModule(
model, inc3, 256, "inc4", 128, [128, 192], [32, 96], 64
)
pool5 = brew.max_pool(model, inc4, "pool5", kernel=3, stride=2, pad=1)
inc5 = _InceptionModule(
model, pool5, 480, "inc5", 192, [96, 208], [16, 48], 64
)
inc6 = _InceptionModule(
model, inc5, 512, "inc6", 160, [112, 224], [24, 64], 64
)
inc7 = _InceptionModule(
model, inc6, 512, "inc7", 128, [128, 256], [24, 64], 64
)
inc8 = _InceptionModule(
model, inc7, 512, "inc8", 112, [144, 288], [32, 64], 64
)
inc9 = _InceptionModule(
model, inc8, 528, "inc9", 256, [160, 320], [32, 128], 128
)
pool9 = brew.max_pool(model, inc9, "pool9", kernel=3, stride=2, pad=1)
inc10 = _InceptionModule(
model, pool9, 832, "inc10", 256, [160, 320], [32, 128], 128
)
inc11 = _InceptionModule(
model, inc10, 832, "inc11", 384, [192, 384], [48, 128], 128
)
pool11 = brew.average_pool(model, inc11, "pool11", kernel=7, stride=1)
fc = brew.fc(
model, pool11, "fc", 1024, 1000,
('XavierFill', {}), ('ConstantFill', {})
)
# It seems that Soumith's benchmark does not have softmax on top
# for Inception. We will add it anyway so we can have a proper
# backward pass.
pred = brew.softmax(model, fc, "pred")
xent = model.LabelCrossEntropy([pred, "label"], "xent")
if device != 'MKL':
loss = model.AveragedLoss(xent, "loss")
return model, 224
import caffe2.contrib.tensorboard.tensorboard as tb
import caffe2.contrib.tensorboard.tensorboard_exporter as tb_exporter
from caffe2.python import brew, core, model_helper
model = model_helper.ModelHelper(name="overfeat")
data, label = brew.image_input(model, ["db"], ["data", "label"], is_test=0)
with core.NameScope("conv1"):
conv1 = brew.conv(model, data, "conv1", 3, 96, 11, stride=4)
relu1 = brew.relu(model, conv1, conv1)
pool1 = brew.max_pool(model, relu1, "pool1", kernel=2, stride=2)
with core.NameScope("conv2"):
conv2 = brew.conv(model, pool1, "conv2", 96, 256, 5)
relu2 = brew.relu(model, conv2, conv2)
pool2 = brew.max_pool(model, relu2, "pool2", kernel=2, stride=2)
with core.NameScope("conv3"):
conv3 = brew.conv(model, pool2, "conv3", 256, 512, 3, pad=1)
relu3 = brew.relu(model, conv3, conv3)
with core.NameScope("conv4"):
conv4 = brew.conv(model, relu3, "conv4", 512, 1024, 3, pad=1)
relu4 = brew.relu(model, conv4, conv4)
with core.NameScope("conv5"):
conv5 = brew.conv(model, relu4, "conv5", 1024, 1024, 3, pad=1)
relu5 = brew.relu(model, conv5, conv5)
pool5 = brew.max_pool(model, relu5, "pool5", kernel=2, stride=2)
with core.NameScope("fc6"):
fc6 = brew.fc(model, pool5, "fc6", 1024 * 6 * 6, 3072)
relu6 = brew.relu(model, fc6, "fc6")
with core.NameScope("fc7"):
fc7 = brew.fc(model, relu6, "fc7", 3072, 4096)
relu7 = brew.relu(model, fc7, "fc7")
def _InceptionModule(
model, input_blob, input_depth, output_name, conv1_depth, conv3_depths,
conv5_depths, pool_depth
):
# path 1: 1x1 conv
conv1 = brew.conv(
model, input_blob, output_name + ":conv1", input_depth, conv1_depth, 1,
('XavierFill', {}), ('ConstantFill', {})
)
conv1 = brew.relu(model, conv1, conv1)
# path 2: 1x1 conv + 3x3 conv
conv3_reduce = brew.conv(
model, input_blob, output_name + ":conv3_reduce", input_depth,
conv3_depths[0], 1, ('XavierFill', {}), ('ConstantFill', {})
)
conv3_reduce = brew.relu(model, conv3_reduce, conv3_reduce)
conv3 = brew.conv(
model,
conv3_reduce,
output_name + ":conv3",
conv3_depths[0],
conv3_depths[1],
3,
('XavierFill', {}),
('ConstantFill', {}),
pad=1
)
conv3 = brew.relu(model, conv3, conv3)
# path 3: 1x1 conv + 5x5 conv
conv5_reduce = brew.conv(
model, input_blob, output_name + ":conv5_reduce", input_depth,
conv5_depths[0], 1, ('XavierFill', {}), ('ConstantFill', {})
)
conv5_reduce = brew.relu(model, conv5_reduce, conv5_reduce)
conv5 = brew.conv(
model,
conv5_reduce,
output_name + ":conv5",
conv5_depths[0],
conv5_depths[1],
5,
('XavierFill', {}),
('ConstantFill', {}),
pad=2
)
conv5 = brew.relu(model, conv5, conv5)
# path 4: pool + 1x1 conv
pool = brew.max_pool(
model,
input_blob,
output_name + ":pool",
kernel=3,
stride=1,
pad=1
)
pool_proj = brew.conv(
model, pool, output_name + ":pool_proj", input_depth, pool_depth, 1,
('XavierFill', {}), ('ConstantFill', {})
)
pool_proj = brew.relu(model, pool_proj, pool_proj)
output = brew.concat(model, [conv1, conv3, conv5, pool_proj], output_name)
return output
def create_resnet152(
model,
data,
num_input_channels,
num_labels,
label=None,
is_test=False,
no_loss=False,
no_bias=0,
conv1_kernel=7,
conv1_stride=2,
final_avg_kernel=7,
):
# conv1 + maxpool
brew.conv(model,
data,
'conv1',
num_input_channels,
64,
weight_init=("MSRAFill", {}),
kernel=conv1_kernel,
stride=conv1_stride,
pad=3,
no_bias=no_bias)
brew.spatial_bn(model,
'conv1',
'conv1_spatbn_relu',
64,
epsilon=1e-3,
momentum=0.1,
is_test=is_test)
brew.relu(model, 'conv1_spatbn_relu', 'conv1_spatbn_relu')
brew.max_pool(model, 'conv1_spatbn_relu', 'pool1', kernel=3, stride=2)
# Residual blocks...
builder = ResNetBuilder(model,
'pool1',
no_bias=no_bias,
is_test=is_test,
spatial_bn_mom=0.1)
# conv2_x (ref Table 1 in He et al. (2015))
builder.add_bottleneck(64, 64, 256)
builder.add_bottleneck(256, 64, 256)
builder.add_bottleneck(256, 64, 256)
# conv3_x
builder.add_bottleneck(256, 128, 512, down_sampling=True)
for _ in range(1, 8):
builder.add_bottleneck(512, 128, 512)
# conv4_x
builder.add_bottleneck(512, 256, 1024, down_sampling=True)
for _ in range(1, 36):
builder.add_bottleneck(1024, 256, 1024)
# conv5_x
builder.add_bottleneck(1024, 512, 2048, down_sampling=True)
builder.add_bottleneck(2048, 512, 2048)
builder.add_bottleneck(2048, 512, 2048)
# Final layers
final_avg = brew.average_pool(
model,
builder.prev_blob,
'final_avg',
kernel=final_avg_kernel,
stride=1,
)
# Final dimension of the "image" is reduced to 7x7
last_out = brew.fc(model, final_avg, 'last_out_L{}'.format(num_labels),
2048, num_labels)
if no_loss:
return last_out
# If we create model for training, use softmax-with-loss
if (label is not None):
(softmax, loss) = model.SoftmaxWithLoss(
[last_out, label],
["softmax", "loss"],
)
return (softmax, loss)
else:
# For inference, we just return softmax
return brew.softmax(model, last_out, "softmax")
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 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 AlexNet(order, cudnn_ws, device):
my_arg_scope = {
'order': order,
'use_cudnn': True,
'cudnn_exhaustive_search': True,
'ws_nbytes_limit': str(cudnn_ws)
}
model = ModelHelper(name="alexnet", arg_scope=my_arg_scope)
conv1 = brew.conv(model,
"data",
"conv1",
3,
64,
11, ('XavierFill', {}), ('ConstantFill', {}),
stride=4,
pad=2)
relu1 = brew.relu(model, conv1, "conv1")
pool1 = brew.max_pool(model, relu1, "pool1", kernel=3, stride=2)
conv2 = brew.conv(model,
pool1,
"conv2",
64,
192,
5, ('XavierFill', {}), ('ConstantFill', {}),
pad=2)
relu2 = brew.relu(model, conv2, "conv2")
pool2 = brew.max_pool(model, relu2, "pool2", kernel=3, stride=2)
conv3 = brew.conv(model,
pool2,
"conv3",
192,
384,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu3 = brew.relu(model, conv3, "conv3")
conv4 = brew.conv(model,
relu3,
"conv4",
384,
256,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu4 = brew.relu(model, conv4, "conv4")
conv5 = brew.conv(model,
relu4,
"conv5",
256,
256,
3, ('XavierFill', {}), ('ConstantFill', {}),
pad=1)
relu5 = brew.relu(model, conv5, "conv5")
pool5 = brew.max_pool(model, relu5, "pool5", kernel=3, stride=2)
fc6 = brew.fc(model, pool5, "fc6", 256 * 6 * 6, 4096, ('XavierFill', {}),
('ConstantFill', {}))
relu6 = brew.relu(model, fc6, "fc6")
fc7 = brew.fc(model, relu6, "fc7", 4096, 4096, ('XavierFill', {}),
('ConstantFill', {}))
relu7 = brew.relu(model, fc7, "fc7")
fc8 = brew.fc(model, relu7, "fc8", 4096, 1000, ('XavierFill', {}),
('ConstantFill', {}))
pred = brew.softmax(model, fc8, "pred")
xent = model.LabelCrossEntropy([pred, "label"], "xent")
if device != 'MKL':
loss = model.AveragedLoss(xent, "loss")
return model, 224
def simple_cnn():
model = ModelHelper(name="r", arg_scope={"order": "NCHW", "is_test": True})
brew.conv(model, "data", 'conv1', 3, 16, kernel=3, stride=1)
brew.spatial_bn(model, 'conv1', 'conv1_spatbn', 16, epsilon=1e-3)
brew.relu(model, 'conv1_spatbn', 'relu1')
return model, [(1, 3, 32, 32)]
