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 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 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 model_build_fun_deploy(model, loss_scale):
last_out = create_resnet(model=model,
data='data',
num_input_channels=config.IMG_CHANNELS,
num_groups=args.groups,
num_labels=config.CLASS_NUM,
is_test=True)
brew.softmax(model, last_out, "softmax")
return []
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 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_head_nodes(self, model, v, dim_in, fc_name, loss_scale=1.0):
"""Adds dense and softmax head nodes.
:param model_helper.ModelHelper model: Current model to use.
:param obj v: Input blobs.
:param int dim_in: Number of input features.
:param str fc_name: Name of a fully connected operator.
:param float loss_scale: For multi-GPU case.
:return: List with one head node. A softmax node if `phase` is `inference`
else `loss`.
"""
v = brew.fc(model, v, fc_name, dim_in=dim_in, dim_out=self.num_classes)
if self.dtype == 'float16':
print(
"[INFO] Converting logits from float16 to float32 for softmax layer"
)
v = model.net.HalfToFloat(v, v + '_fp32')
if self.phase == 'inference':
softmax = brew.softmax(model, v, 'softmax')
head_nodes = [softmax]
else:
softmax, loss = model.SoftmaxWithLoss([v, 'softmax_label'],
['softmax', 'loss'])
prefix = model.net.Proto().name
loss = model.Scale(loss, prefix + "_loss", scale=loss_scale)
head_nodes = [loss]
return head_nodes
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.
Para encadenar layers, la dimension de salida tiene que coincidir con a dimension de entrada
de la siguiente layer (sin contar las pool, solo las conv)
'''
#Each layer is created using brew, which in turn is using its operator hooks to instantiate each Op.
# 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 = 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)
# 100 * 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)
#TODO: RELU despues de cada conv layer -> no mejora mucho la accuracy
relu = brew.relu(model, fc3, fc3)
pred = brew.fc(model, relu, 'pred', 500, 10)
softmax = brew.softmax(model, pred, 'softmax')
return softmax
def test_caffe2_simple_model(self):
model = 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 = c2_graph.model_to_graph_def(
model,
blob_name_tracker=blob_name_tracker,
shapes={},
show_simplified=False,
)
compare_proto(graph, self)
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 / 350 x 350 -> 346 x 346
conv1 = brew.conv(model, data, 'conv1', dim_in=1, dim_out=20, kernel=3)
# Image size: 346 x 346 -> 173 x 173
pool1 = brew.max_pool(model, conv1, 'pool1', kernel=2, stride=2)
# Image size: 174 x 174
conv2 = brew.conv(model, pool1, 'conv2', dim_in=20, dim_out=50, kernel=5)
# Image size: 174 x 174 -> 87 x 87
pool2 = brew.max_pool(model, conv2, 'pool2', kernel=2, stride=2)
# 50 * 4 * 4 stands for dim_out from previous layer multiplied by the image size
# Here, the data is flattened from a tensor of dimension 50x4x4 to a vector of length 50*4*4
fc3 = brew.fc(model, pool2, 'fc3', dim_in=50 * 87 * 87, dim_out=500)
relu3 = brew.relu(model, fc3, 'relu3')
# Last FC Layer
pred = brew.fc(model, relu3, 'pred', dim_in=500, dim_out=10)
# Softmax Layer
softmax = brew.softmax(model, pred, 'softmax')
return softmax
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 = 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=50, kernel=5)
# Image size: 8 x 8 -> 4 x 4
pool2 = brew.max_pool(model, 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 = brew.relu(model, fc3, fc3)
pred = brew.fc(model, fc3, 'pred', 500, 10)
softmax = brew.softmax(model, pred, 'softmax')
return softmax
def AddLeNetModel(model, data, num_classes, image_height, image_width,
image_channels):
# Image size: 28 x 28 -> 24 x 24
############# Change dim_in value if images are more than 1 color channel
conv1 = brew.conv(model, data, 'conv1', dim_in=1, dim_out=20, kernel=5)
h, w = update_dims(height=image_height,
width=image_width,
kernel=5,
stride=1,
pad=0)
# Image size: 24 x 24 -> 12 x 12
pool1 = brew.max_pool(model, conv1, 'pool1', kernel=2, stride=2)
h, w = update_dims(height=h, width=w, kernel=2, stride=2, pad=0)
# Image size: 12 x 12 -> 8 x 8
conv2 = brew.conv(model, pool1, 'conv2', dim_in=20, dim_out=50, kernel=5)
h, w = update_dims(height=h, width=w, kernel=5, stride=1, pad=0)
# Image size: 8 x 8 -> 4 x 4
pool2 = brew.max_pool(model, conv2, 'pool2', kernel=2, stride=2)
h, w = update_dims(height=h, width=w, kernel=2, stride=2, pad=0)
# 50 * 4 * 4 stands for dim_out from previous layer multiplied by the image size
############# Change dim_in value if images are not 28x28
fc3 = brew.fc(model, pool2, 'fc3', dim_in=50 * h * w, dim_out=500)
fc3 = brew.relu(model, fc3, fc3)
pred = brew.fc(model, fc3, 'pred', 500, num_classes)
softmax = brew.softmax(model, pred, 'softmax')
return softmax
def AddLeNetModel(model, data):
conv1 = brew.conv(model,
data,
'conv1',
dim_in=1,
dim_out=32,
kernel=3,
pad=1)
conv1 = brew.relu(model, conv1, conv1)
conv2 = brew.conv(model,
conv1,
'conv2',
dim_in=32,
dim_out=128,
kernel=3,
pad=1)
conv2 = brew.relu(model, conv2, conv2)
conv3 = brew.conv(model,
conv2,
'conv3',
dim_in=128,
dim_out=256,
kernel=3,
pad=1)
conv3 = brew.relu(model, conv3, conv3)
fc3 = brew.fc(model, conv3, 'fc3', dim_in=256 * 28 * 28, dim_out=512)
fc3 = brew.relu(model, fc3, fc3)
pred = brew.fc(model, fc3, 'pred', 512, 10)
softmax = brew.softmax(model, pred, 'softmax')
return softmax
def add_lenet_fc_topk(model, data, **kwargs):
'''
Based on original LeNet structure, according to meProp, we add topk
gradient selecting layer for fc layers
'''
# resolving arguments
k = kwargs.get('k', 10)
# Image size: 28 x 28 -> 24 x 24
conv1 = brew.conv(model, data, 'conv1', dim_in=3, 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=50, kernel=5)
# Image size: 8 x 8 -> 4 x 4
pool2 = brew.max_pool(model, conv2, 'pool2', kernel=2, stride=2)
# DEBUG PRINT
model.net.Print(pool2, [])
# set topkgradhook for the hidden fc layers
fc3 = brew.fc(model, pool2, 'fc3', dim_in=50 * 4 * 4, dim_out=500)
fc3_hook = model.net.TopKGradHook(fc3, 'fc3_hook', k=k)
relu = brew.relu(model, fc3_hook, fc3_hook)
# don't set topkgradhook for the output fc layer
pred = brew.fc(model, relu, 'pred', 500, 10)
softmax = brew.softmax(model, pred, 'softmax')
return softmax
def add_lenet_conv_topk(model, data, **kwargs):
'''
Based on original LeNet structure, according to meProp, we add topk
gradient selecting layer for conv layers
'''
# resolving arguments
config = kwargs.get('config', {})
k = kwargs.get('k', 10)
# Image size: 28 x 28 -> 24 x 24
conv1 = brew.conv(model, data, 'conv1', dim_in=3, dim_out=20, kernel=5)
conv1_hook = model.net.TopKGradHook(conv1, 'conv1_hook', k=k)
# Image size: 24 x 24 -> 12 x 12
pool1 = brew.max_pool(model, conv1_hook, '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)
conv2_hook = model.net.TopKGradHook(conv2, 'conv2_hook', k=k)
# Image size: 8 x 8 -> 4 x 4
pool2 = brew.max_pool(model, conv2_hook, '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)
relu = brew.relu(model, fc3, fc3)
pred = brew.fc(model, relu, 'pred', 500, config['model_arch']['num_classes'])
softmax = brew.softmax(model, pred, 'softmax')
return softmax
def AddModel(model, data):
conv1 = brew.conv(model,
data,
'conv1',
dim_in=input_channels,
dim_out=10,
kernel=3,
stride=1,
pad=1,
training_mode=1)
deconv1 = brew.conv_transpose(model,
conv1,
'deconv1',
dim_in=10,
dim_out=10,
kernel=2,
stride=2,
pad=0,
training_mode=1)
fc1 = brew.fc(model,
deconv1,
'fc1',
dim_in=10 * 56 * 56,
dim_out=3)
softmax = brew.softmax(model, fc1, 'softmax')
return softmax
def ScaffoldModelCNN_AlexNet(model, classCount, data, imageDimension, amountOfChannels): brew.conv(model, data, 'conv1', amountOfChannels, 96, 11, stride=4, pad=0) brew.relu(model, 'conv1', 'conv1') brew.max_pool(model, 'norm1', 'pool1', kernel=3, stride=2, pad=0) brew.conv(model, 'pool1', 'conv2', 48, 256, 5, stride=1, pad=2) brew.relu(model, 'conv2', 'conv2') brew.max_pool(model, 'norm2', 'pool2', kernel=3, stride=2, pad=0) brew.conv(model, 'pool2', 'conv3', 256, 384, 3, stride=1, pad=1) brew.relu(model, 'conv3', 'conv3') brew.conv(model, 'conv3', 'conv4', 192, 384, 3, stride=1, pad=1) brew.relu(model, 'conv4', 'conv4') brew.conv(model, 'conv4', 'conv5', 192, 256, 3, stride=1, pad=1) brew.relu(model, 'conv5', 'conv5') brew.max_pool(model, 'conv5', 'pool5', kernel=3, stride=2, pad=1) brew.fc(model, 'pool5', 'fc6', 9216, 4096) brew.relu(model, 'fc6', 'fc6') brew.fc(model, 'fc6', 'fc7', 4096, 4096) brew.relu(model, 'fc7', 'fc7') brew.fc(model, 'fc7', 'fc8', 4096, 1000) return brew.softmax(model, 'fc8', 'softmax')
def add_layers(model, data):
# Image size: 32 x 32 x 3 -> 30 x 30 x 60
conv1 = brew.conv(model, data, 'conv1', dim_in=3, dim_out=64, kernel=3)
#add relu
conv1 = brew.relu(model, conv1, conv1)
# Image size: 30 x 30 -> 15 x 15
pool1 = brew.max_pool(model, conv1, 'pool1', kernel=2, stride=2)
# Image size: 15 x 15 -> 12 x 12
conv2 = brew.conv(model, pool1, 'conv2', dim_in=64, dim_out=128, kernel=3)
#add relu
conv2 = brew.relu(model, conv2, conv2)
# Image size: 12 x 12 -> 6 x 6
pool2 = brew.max_pool(model, conv2, 'pool2', kernel=2, stride=2)
# 32 * 6 * 6 stands for dim_out from previous layer multiplied by the image size
# Hidden layer
fc3 = brew.fc(model, pool2, 'fc3', dim_in=128 * 6 * 6, dim_out=512)
#add relu
fc3 = brew.relu(model, fc3, fc3)
#CO = model.param_init_net.ConstantFill([], "CO", shape=[1], value=0.18)
#fc3 = brew.prelu(model, fc3, fc3)
#output layer
pred = brew.fc(model, fc3, 'pred', 512, 10)
#add softmax
probs = brew.softmax(model, pred, 'probs')
return probs
def AddNet(model, data):
# image size: 28x28 -> 24x24
conv1 = brew.conv(model,
data,
"conv1",
dim_in=1,
dim_out=20,
kernel=5,
stride=1)
# image size: 24x24 -> 12x12
pool1 = brew.max_pool(model, conv1, "pool1", kernel=2, stride=2)
# image size: 12x12 -> 8x8
conv2 = brew.conv(model,
pool1,
"conv2",
dim_in=20,
dim_out=100,
kernel=5,
stride=1)
# image size: 8x8 -> 4x4
pool2 = brew.max_pool(model, conv2, "pool2", kernel=2, stride=2)
# image size: 100x4x4 -> 500
fc3 = brew.fc(model, pool2, "fc3", dim_in=100 * 4 * 4, dim_out=500)
relu = brew.relu(model, fc3, fc3)
# image size: 500 -> 10
fc4 = brew.fc(model, fc3, "fc4", dim_in=500, dim_out=10)
# softmax
softmax = brew.softmax(model, fc4, "softmax")
return softmax
def AddMLP_BN(model, data):
'''
Implement MLP model on MNIST
'''
# number of nuerons in fc layer
num_units = 4096
# NCHW: 64 x 1 x 28 x 28 -> 64 x 1 x 28 x 28
drop1 = brew.dropout(model, data, 'drop1', ratio=0.5, is_test=0)
# NCHW: 64 x 1 x 28 x 28 -> 64 x 4096
fc2 = brew.fc(model, drop1, 'fc2', dim_in=1 * 28 * 28, dim_out=num_units)
fc2_reshaped = model.Reshape([fc2], ['fc2_reshaped'], shape=(4, 2))
# bn2 = brew.spatial_bn(model, fc2, 'bn2', num_units, epsilon=1e-3, momentum=0.9, is_test=is_test)
#relu2 = brew.relu(model, fc2, 'relu2')
#fc3 = brew.fc(model, relu2, 'fc3', dim_in=num_units, dim_out=num_units)
# bn3 = brew.spatial_bn(model, fc3, 'bn3', bn_units, epsilon=1e-3, momentum=0.9, is_test=is_test)
#relu3 = brew.relu(model, fc3, 'relu3')
#fc4 = brew.fc(model, relu3, 'fc4', dim_in=num_units, dim_out=num_units)
# bn4 = brew.spatial_bn(model, fc4, 'bn4', bn_units, epsilon=1e-3, momentum=0.9, is_test=is_test)
#relu4 = brew.relu(model, fc4, 'relu4')
#fc5 = brew.fc(model, relu4, 'fc5', dim_in=num_units, dim_out=10) # 10 for 10-classes
# bn5 = brew.spatial_bn(model, fc5, 'bn5', 10, epsilon=1e-3, momentum=0.9, is_test=is_test)
softmax = brew.softmax(model, fc2, 'softmax')
return softmax
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 = 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=50, kernel=5)
# Image size: 8 x 8 -> 4 x 4
pool2 = brew.max_pool(model, 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 = brew.relu(model, fc3, fc3)
pred = brew.fc(model, fc3, 'pred', 500, 10)
softmax = brew.softmax(model, pred, 'softmax')
return softmax
def AddLeNetModel(model, data, save_png=True):
"""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.
"""
conv1 = brew.conv(model, data, 'conv1', dim_in=1, dim_out=20, kernel=5)
pool1 = brew.max_pool(model, conv1, 'pool1', kernel=2, stride=2)
conv2 = brew.conv(model, pool1, 'conv2', dim_in=20, dim_out=50, kernel=5)
pool2 = brew.max_pool(model, conv2, 'pool2', kernel=2, stride=2)
fc3 = brew.fc(model, pool2, 'fc3', dim_in=50 * 4 * 4, dim_out=500)
relu3 = brew.relu(model, fc3, 'relu3')
pred = brew.fc(model, relu3, 'pred', dim_in=500, dim_out=10)
softmax = brew.softmax(model, pred, 'softmax')
if save_png:
graph = net_drawer.GetPydotGraph(model.net)
graph.write_png("LeNet.png")
return softmax
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 OverFeat(order, cudnn_ws, mkl):
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 not mkl:
loss = model.AveragedLoss(xent, "loss")
return model, 231
def test_simple_model(self):
model = model_helper.ModelHelper(name="mnist")
data, label = brew.image_input(model, ["db"], ["data", "label"],
order="NCHW",
use_gpu_transform=False,
is_test=0)
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.RunAllOnGPU()
model.param_init_net.RunAllOnGPU()
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
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()))
self.assertMultiLineEqual(actual, expected)
def AddLeNetModel(model, data):
fc1 = brew.fc(model, data, 'fc1', dim_in=784, dim_out=1000)
fc1 = brew.relu(model, fc1, fc1)
fc2 = brew.fc(model, fc1, 'fc2', dim_in=1000, dim_out=1000)
fc2 = brew.relu(model, fc2, fc2)
pred = brew.fc(model, fc2, 'fc3', dim_in=1000, dim_out=10)
softmax = brew.softmax(model, pred, 'softmax')
return softmax
def Add_Original_CIFAR10_Model(model,
data,
num_classes,
image_height,
image_width,
image_channels,
save_png=True):
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)
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)
relu1 = brew.relu(model, pool1, 'relu1')
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)
relu2 = brew.relu(model, conv2, 'relu2')
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)
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)
relu3 = brew.relu(model, conv3, 'relu3')
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)
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 = brew.softmax(model, fc2, 'softmax')
if save_png:
graph = net_drawer.GetPydotGraph(model.net, rankdir="LR")
graph.write_png("CIFAR10_Model.png")
return softmax
def AddMLPModel(model, data):
size = INPUT_WIDTH * INPUT_HEIGHT * INPUT_LAYERS
sizes = [INPUT_WIDTH, INPUT_HEIGHT * 2, size * 2, CLASSES]
layer = data
for i in range(len(sizes) - 1):
layer = brew.fc(model, layer, 'dense_{}'.format(i), dim_in=sizes[i], dim_out=sizes[i + 1])
layer = brew.relu(model, layer, 'relu_{}'.format(i))
softmax = brew.softmax(model, layer, 'softmax')
return softmax
def AddMLPModel(model, data):
size = 28 * 28 * 1
sizes = [size, size * 2, size * 2, 10]
layer = data
for i in range(len(sizes) - 1):
layer = brew.fc(model, layer, 'dense_{}'.format(i), dim_in=sizes[i], dim_out=sizes[i + 1])
layer = brew.relu(model, layer, 'relu_{}'.format(i))
softmax = brew.softmax(model, layer, 'softmax')
return softmax
def AddLeNetModel(model, data):
conv1 = brew.conv(model, data, 'conv1', dim_in=1, dim_out=20, kernel=5)
pool1 = brew.max_pool(model, conv1, 'pool1', kernel=2, stride=2)
conv2 = brew.conv(model, pool1, 'conv2', dim_in=20, dim_out=50, kernel=5)
pool2 = brew.max_pool(model, conv2, 'pool2', kernel=2, stride=2)
fc3 = brew.fc(model, pool2, 'fc3', dim_in=50 * 4 * 4, dim_out=500)
fc3 = brew.relu(model, fc3, fc3)
pred = brew.fc(model, fc3, 'pred', 500, 10)
softmax = brew.softmax(model, pred, 'softmax')
def add_softmax(self, prev_blob, label=None):
if label is not None:
(softmax, loss) = self.model.SoftmaxWithLoss(
[prev_blob, label],
['softmax', 'loss'],
)
return (softmax, loss)
else:
return brew.softmax(self.model, prev_blob, 'softmax')
def AddAlexNet(model, data):
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, 10, ('XavierFill', {}),
('ConstantFill', {}))
pred = brew.softmax(model, fc8, "pred")
return pred
def AddLeNetModel(model, data, device_opts):
'''
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.
'''
with core.DeviceScope(device_opts):
############# Change dim_in value if images are more than 1 color channel
# Image size: 64x64
conv1 = brew.conv(model, data, 'conv1', dim_in=1, dim_out=32, kernel=5)
h, w = update_dims(height=image_height,
width=image_width,
kernel=5,
stride=1,
pad=0)
# Image size: 60x60
pool1 = brew.max_pool(model, conv1, 'pool1', kernel=2, stride=2)
h, w = update_dims(height=h, width=w, kernel=2, stride=2, pad=0)
relu1 = brew.relu(model, pool1, 'relu1')
# Image size: 30x30
conv2 = brew.conv(model,
relu1,
'conv2',
dim_in=32,
dim_out=64,
kernel=5)
h, w = update_dims(height=h, width=w, kernel=5, stride=1, pad=0)
# Image size: 26x26
pool2 = brew.max_pool(model, conv2, 'pool2', kernel=2, stride=2)
h, w = update_dims(height=h, width=w, kernel=2, stride=2, pad=0)
relu2 = brew.relu(model, pool2, 'relu2')
# Image size: 13x13
conv3 = brew.conv(model,
relu2,
'conv3',
dim_in=64,
dim_out=64,
kernel=5)
h, w = update_dims(height=h, width=w, kernel=5, stride=1, pad=0)
# Image size: 9x9
pool3 = brew.max_pool(model, conv3, 'pool3', kernel=2, stride=2)
h, w = update_dims(height=h, width=w, kernel=2, stride=2, pad=0)
relu3 = brew.relu(model, pool3, 'relu3')
# Image size: 4x4
# 50 * 4 * 4 stands for dim_out from previous layer multiplied by the image size
############# Change dim_in value if images are not 28x28
fc3 = brew.fc(model, relu3, 'fc3', dim_in=64 * h * w, dim_out=500)
fc3 = brew.relu(model, fc3, fc3)
pred = brew.fc(model, fc3, 'pred', 500, num_classes)
softmax = brew.softmax(model, pred, 'softmax')
return softmax
def test_get_complete_net(self):
model = model_helper.ModelHelper("test_orig")
conv = brew.conv(
model,
"input",
"conv",
dim_in=3,
dim_out=16,
weight_init=("MSRAFill", {}),
kernel=3,
stride=1,
pad=0,
)
conv = brew.spatial_bn(model,
conv,
"conv_bn",
16,
epsilon=1e-3,
is_test=False)
conv = brew.relu(model, conv, "conv_relu")
pred = brew.fc(model, conv, "pred", dim_in=16 * 3 * 3, dim_out=10)
brew.softmax(model, pred, "softmax")
net = model.GetCompleteNet()
model2 = model_helper.ModelHelper("test_new")
model2.ConstructInitTrainNetfromNet(net)
net = model.param_init_net
net2 = model2.param_init_net
for op1, op2 in zip(net.Proto().op, net2.Proto().op):
op1.debug_info = op1.debug_info + "/param_init_net"
self.assertEqual(
op1, op2, "op mismatch between {}\n and {}\n".format(op1, op2))
net = model.net
net2 = model2.net
for op1, op2 in zip(net.Proto().op, net2.Proto().op):
self.assertEqual(
op1, op2, "op mismatch between {}\n and {}\n".format(op1, op2))
# this is not guaranteed in other situations where user define own net
self.assertEqual(
sorted(map(str, net.external_inputs)),
sorted(map(str, net2.external_inputs)),
)
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 _calc_attention_weights(
model,
attention_logits_transposed,
scope,
encoder_lengths=None,
):
if encoder_lengths is not None:
attention_logits_transposed = model.net.SequenceMask(
[attention_logits_transposed, encoder_lengths],
['masked_attention_logits'],
mode='sequence',
)
# [batch_size, encoder_length, 1]
attention_weights_3d = brew.softmax(
model,
attention_logits_transposed,
s(scope, 'attention_weights_3d'),
engine='CUDNN',
axis=1,
)
return attention_weights_3d
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 _calc_attention_weights(
model,
attention_logits_transposed,
scope,
encoder_lengths=None,
):
if encoder_lengths is not None:
attention_logits = model.net.Squeeze(
[attention_logits_transposed],
[attention_logits_transposed],
dims=[2],
)
flat_attention_logits = model.net.UnpackSegments(
[encoder_lengths, attention_logits],
'flat_attention_logits',
)
masked_attention_logits = model.net.PackSegments(
[encoder_lengths, flat_attention_logits],
'masked_attention_logits',
pad_minf=True,
)
attention_logits_transposed = model.net.ExpandDims(
[masked_attention_logits],
[masked_attention_logits],
dims=[2],
)
# [batch_size, encoder_length, 1]
attention_weights_3d = brew.softmax(
model,
attention_logits_transposed,
s(scope, 'attention_weights_3d'),
engine='CUDNN',
axis=1,
)
return attention_weights_3d
def Softmax(self, *args, **kwargs):
return brew.softmax(self, *args, use_cudnn=self.use_cudnn, **kwargs)
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 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 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 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 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 _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
