def create_resnet50_model_ops(model, loss_scale):
initializer = (pFP16Initializer if args.dtype == 'float16'
else Initializer)
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=initializer,
BiasInitializer=initializer,
enable_tensor_core=args.enable_tensor_core):
pred = resnet.create_resnet50(
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
no_bias=True,
no_loss=True,
)
if args.dtype == 'float16':
pred = model.net.HalfToFloat(pred, pred + '_fp32')
softmax, loss = model.SoftmaxWithLoss([pred, 'label'],
['softmax', 'loss'])
loss = model.Scale(loss, scale=loss_scale)
brew.accuracy(model, [softmax, "label"], "accuracy")
return [loss]
def create_resnext_model_ops(model, loss_scale):
initializer = (PseudoFP16Initializer
if args.dtype == 'float16' else Initializer)
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=initializer,
BiasInitializer=initializer,
enable_tensor_core=args.enable_tensor_core,
float16_compute=args.float16_compute):
pred = resnet.create_resnext(
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
num_layers=args.num_layers,
num_groups=args.resnext_num_groups,
num_width_per_group=args.resnext_width_per_group,
no_bias=True,
no_loss=True,
)
if args.dtype == 'float16':
pred = model.net.HalfToFloat(pred, pred + '_fp32')
softmax, loss = model.SoftmaxWithLoss([pred, 'label'],
['softmax', 'loss'])
loss = model.Scale(loss, scale=loss_scale)
brew.accuracy(model, [softmax, "label"], "accuracy", top_k=1)
brew.accuracy(model, [softmax, "label"], "accuracy_top5", top_k=5)
return [loss]
def create_model(model, loss_scale):
initializer = (PseudoFP16Initializer
if args.data_type == 'float16' else Initializer)
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=initializer,
BiasInitializer=initializer,
enable_tensor_core=False,
float16_compute=False):
pred = resnet.create_resnet50(
model,
"data",
num_input_channels=args.channels,
num_labels=args.num_labels,
# num_groups=args.resnext_num_groups,
# num_width_per_group=args.resnext_width_per_group,
no_bias=True,
no_loss=True)
# If we're using float on 2B, then inflate to the 4B representation
if args.data_type == 'float16':
pred = model.net.HalfToFloat(pred, pred + '_fp32')
# Compute the softmax probabilities and the loss
softmax, loss = model.SoftmaxWithLoss([pred, 'label'],
['softmax', 'loss'])
# Noralize the loss, and compute the top_k accuracies for k \in {1, 5}
loss = model.Scale(loss, scale=loss_scale)
brew.accuracy(model, [softmax, "label"], "accuracy", top_k=1)
brew.accuracy(model, [softmax, "label"], "accuracy_top5", top_k=5)
return [loss]
def create_resnet50_model_ops(model, loss_scale):
initializer = (pFP16Initializer if args.dtype == 'float16'
else Initializer)
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=initializer,
BiasInitializer=initializer,
enable_tensor_core=args.enable_tensor_core):
pred = resnet.create_resnet50(
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
no_bias=True,
no_loss=True,
)
if args.dtype == 'float16':
pred = model.net.HalfToFloat(pred, pred + '_fp32')
softmax, loss = model.SoftmaxWithLoss([pred, 'label'],
['softmax', 'loss'])
loss = model.Scale(loss, scale=loss_scale)
brew.accuracy(model, [softmax, "label"], "accuracy")
return [loss]
def Resnext101(model, loss_scale, dtype='float'):
initializer = (PseudoFP16Initializer
if dtype == 'float16' else Initializer)
with brew.arg_scope(
[brew.conv, brew.fc],
WeightInitializer=initializer,
BiasInitializer=initializer,
):
# residual network
pred = resnet.create_resnext(
model,
"data",
num_input_channels=3,
num_labels=1000,
label="label",
num_layers=101,
num_groups=32,
num_width_per_group=4,
no_bias=True,
no_loss=True,
)
if dtype == 'float16':
pred = model.net.HalfToFloat(pred, pred + '_fp32')
softmax, loss = model.SoftmaxWithLoss([pred, 'label'],
['softmax', 'loss'])
prefix = model.net.Proto().name
loss = model.net.Scale(loss, prefix + "_loss", scale=loss_scale)
brew.accuracy(model, [softmax, "label"], prefix + "_accuracy")
return [loss]
def create_inception_model_ops(model, loss_scale=1.0):
[softmax,
loss] = inceptionv4.create_Inceptionv4(model, "data", num_labels,
"label")
prefix = model.net.Proto().name
loss = model.net.Scale(loss, prefix + "_loss", scale=loss_scale)
brew.accuracy(model, [softmax, "label"], prefix + "_accuracy")
return [loss]
def add_training_operators(softmax, model, device_opts) :
with core.DeviceScope(device_opts):
xent = model.LabelCrossEntropy([softmax, "label"], 'xent')
loss = model.AveragedLoss(xent, "loss")
brew.accuracy(model, [softmax, "label"], "accuracy")
model.AddGradientOperators([loss])
opt = optimizer.build_sgd(model, base_learning_rate=0.01, policy="step", stepsize=1, gamma=0.999) # , momentum=0.9
def add_accuracy(self, model, output, label, device_opts, eval_metric):
with core.DeviceScope(device_opts):
if eval_metric == 'accuracy':
accuracy = brew.accuracy(model, [output, label], "accuracy")
elif eval_metric == 'top_k_accuracy':
accuracy = brew.accuracy(model, [output, label],
"accuracy",
top_k=3)
return accuracy
def create_mobilenet_model_ops(model, loss_scale):
[softmax, loss] = mobilenet.create_mobilenet(
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
label="label")
# loss = model.Scale(loss, scale=loss_scale)
brew.accuracy(model, [softmax, "label"], "accuracy")
def create_resnet50_model_ops(model, loss_scale=1.0):
# residual network
[softmax, loss] = resnet.create_resnet50(model,
"data",
num_input_channels=3,
num_labels=num_labels,
label="label",
no_bias=True, )
prefix = model.net.Proto().name
loss = model.net.Scale(loss, prefix + "_loss", scale=loss_scale)
brew.accuracy(model, [softmax, "label"], prefix + "_accuracy")
return [loss]
def AddForwardPassOps(model):
"""Add forward pass ops and return a list of losses."""
conv1 = brew.conv(model, 'data', 'conv1', 1, 20, 5)
pool1 = brew.max_pool(model, conv1, 'pool1', kernel=2, stride=2)
conv2 = brew.conv(model, pool1, 'conv2', 20, 50, 5)
pool2 = brew.max_pool(model, conv2, 'pool2', kernel=2, stride=2)
fc3 = brew.fc(model, pool2, 'fc3', 50 * 4 * 4, 500)
fc3 = brew.relu(model, fc3, fc3)
pred = brew.fc(model, fc3, 'pred', 500, 10)
softmax, loss = model.SoftmaxWithLoss([pred, 'label'], ['softmax', 'loss'])
brew.accuracy(model, [softmax, 'label'], 'accuracy')
return [loss]
def create_resnet50_model_ops(model, loss_scale):
[softmax, loss] = resnet.create_resnet50(
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
label="label",
no_bias=True,
)
loss = model.Scale(loss, scale=loss_scale)
brew.accuracy(model, [softmax, "label"], "accuracy")
return [loss]
def create_resnet50_model_ops(model, loss_scale=1.0):
# Creates a residual network
[softmax, loss] = resnet.create_resnet50(
model,
"data",
num_input_channels=3,
num_labels=num_labels,
label="label",
)
prefix = model.net.Proto().name
loss = model.net.Scale(loss, prefix + "_loss", scale=loss_scale)
brew.accuracy(model, [softmax, "label"], prefix + "_accuracy")
return [loss]
def create_target_model_ops(model, loss_scale):
initializer = (PseudoFP16Initializer if args.dtype == 'float16'
else Initializer)
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=initializer,
BiasInitializer=initializer,
enable_tensor_core=args.enable_tensor_core,
float16_compute=args.float16_compute):
pred = add_se_model(model, model_config, "data", is_test=False)
if args.dtype == 'float16':
pred = model.net.HalfToFloat(pred, pred + '_fp32')
loss = add_softmax_loss(model, pred, 'label')
brew.accuracy(model, ['softmax', 'label'], 'accuracy')
return [loss]
def add_accuracy(model, softmax, label):
""" compute model classification accuracy """
accuracy = brew.accuracy(model, [softmax, label], "accuracy")
accuracy_5 = model.net.Accuracy(
[softmax, label],
"accuracy_5",
top_k=5,
)
return (accuracy, accuracy_5)
def AddTrainingOperators(model, softmax, label, device_opts):
with core.DeviceScope(device_opts):
xent = model.LabelCrossEntropy([softmax, label], 'xent')
# Compute the expected loss
loss = model.AveragedLoss(xent, "loss")
brew.accuracy(model, [softmax, label], "accuracy")
# Use the average loss we just computed to add gradient operators to the model
model.AddGradientOperators([loss])
# Use SGD optimizer
optimizer.build_sgd(
model,
base_learning_rate=0.1,
weight_decay=1e-5,
gamma=0.999,
policy='step',
stepsize=50,
nesterov=1,
)
def AddForwardPassOps(model):
"""Add forward pass ops and return a list of losses."""
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=pFP16Initializer,
BiasInitializer=pFP16Initializer):
conv1 = brew.conv(model, 'data', 'conv1', 1, 20, 5)
pool1 = brew.max_pool(model, conv1, 'pool1', kernel=2, stride=2)
conv2 = brew.conv(model, pool1, 'conv2', 20, 50, 5)
pool2 = brew.max_pool(model, conv2, 'pool2', kernel=2, stride=2)
fc3 = brew.fc(model, pool2, 'fc3', 50 * 4 * 4, 500)
fc3 = brew.relu(model, fc3, fc3)
pred = brew.fc(model, fc3, 'pred', 500, 10)
# Cast back to fp32 for remaining ops
pred = model.net.HalfToFloat(pred, pred + '_fp32')
softmax, loss = model.SoftmaxWithLoss([pred, 'label'], ['softmax', 'loss'])
brew.accuracy(model, [softmax, 'label'], 'accuracy')
return [loss]
def create_resnet50_model_ops(model, loss_scale):
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=Initializer,
BiasInitializer=Initializer,
enable_tensor_core=0):
pred = resnet.create_resnet50(
model,
"data",
num_input_channels=num_channels,
num_labels=num_labels,
no_bias=True,
no_loss=True,
)
softmax, loss = model.SoftmaxWithLoss([pred, 'label'],
['softmax', 'loss'])
loss = model.Scale(loss, scale=loss_scale)
brew.accuracy(model, [softmax, "label"], "accuracy")
return [loss]
def AddAccuracy(model, softmax, label):
'''模型准确率.
参数:
model: 模型结构
softmax: 分类数据
lable:图像标签
返回:
accuracy: 识别准确率
'''
accuracy = brew.accuracy(model, [softmax, label], 'accuracy')
return accuracy
def AddLeNetModel(model, data, label=None, no_loss=False, embedding_size=2, class_num=10, margin=0):
'''
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=32, kernel=5)
prelu1 = brew.relu(model, conv1, "prelu1")
# Image size: 24 x 24 -> 12 x 12
pool1 = brew.max_pool(model, prelu1, 'pool1', kernel=2, stride=2)
# Image size: 12 x 12 -> 8 x 8
conv2 = brew.conv(model, pool1, 'conv2', dim_in=32, dim_out=64, kernel=5)
prelu2 = brew.relu(model, conv2, "prelu2")
# Image size: 8 x 8 -> 4 x 4
pool2 = brew.max_pool(model, prelu2, '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=64 * 4 * 4, dim_out=256)
prelu3 = brew.relu(model, fc3, "prelu3")
embedding = brew.fc(model, prelu3, 'embedding', 256, embedding_size)
if no_loss:
return embedding
if label is not None:
output = brew.lsoftmax(model, [embedding, label], "fc4", embedding_size, class_num,
margin=margin,
base=float(10),#200
lambda_min=float(0)) #5
fc4 = output[0]
softmax, loss = model.SoftmaxWithLoss([fc4, label], ['softmax', 'loss'])
accuracy = brew.accuracy(model, [softmax, label], "accuracy")
return [loss, accuracy]
else:
fc4 = brew.fc(model, embedding, 'fc4', embedding_size, class_num)
return brew.softmax(model, fc4, "softmax")
def add_accuracy(model, softmax, device_opts):
with core.DeviceScope(device_opts):
accuracy = brew.accuracy(model, [softmax, "label"], "accuracy")
return accuracy
def AddAccuracy(model, softmax, label):
accuracy = brew.accuracy(model, [softmax, label], "accuracy")
return accuracy
# Load init and predict nets
#
test_model = model_helper.ModelHelper(name="test_model", arg_scope=arg_scope, init_params=False)
data, _ = AddInputLayer(test_model, 1, TEST_LMDB, "lmdb")
init_net_proto = caffe2_pb2.NetDef()
with open(INIT_NET, "rb") as f:
init_net_proto.ParseFromString(f.read())
test_model.param_init_net = test_model.param_init_net.AppendNet(core.Net(init_net_proto))
predict_net_proto = caffe2_pb2.NetDef()
with open(PREDICT_NET, "rb") as f:
predict_net_proto.ParseFromString(f.read())
test_model.net = test_model.net.AppendNet(core.Net(predict_net_proto))
accuracy = brew.accuracy(test_model, ["softmax", "label"], "accuracy")
#
# Test loop
#
workspace.RunNetOnce(test_model.param_init_net)
workspace.CreateNet(test_model.net, overwrite=True)
avg_accuracy = 0.0
test_iters = 10000
for i in range(test_iters):
workspace.RunNet(test_model.net)
acc = workspace.FetchBlob('accuracy')
avg_accuracy += acc
if (i+1) % 500 == 0:
def AddAccuracy(model, softmax, label):
"""Adds an accuracy op to the model"""
accuracy = brew.accuracy(model, [softmax, label], "accuracy")
return accuracy
def Accuracy(self, *args, **kwargs):
return brew.accuracy(self, *args, **kwargs)
def CivilNet(name, train_test_deplopy=0):
arg_scope = {
'order': 'NCHW',
'use_cudnn': True,
'cudnn_exhaustive_search': True,
'ws_nbytes_limit': (64 * 1024 * 1024)
}
model = model_helper.ModelHelper(name=name, arg_scope=arg_scope)
model._device_type = caffe2_pb2.CUDA
model._device_prefix = "gpu"
model._shared_model = False
model._devices = [0]
device_opt = core.DeviceOption(caffe2_pb2.CUDA, 0)
#for deploy
if train_test_deplopy == 2:
with core.DeviceScope(device_opt):
with core.NameScope("{}_{}".format(model._device_prefix, 0)):
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=Initializer,
BiasInitializer=Initializer,
enable_tensor_core=False,
float16_compute=False):
resnet.create_resnet50(model,
"data",
num_input_channels=3,
num_labels=args.num_labels,
no_bias=True,
no_loss=False)
workspace.RunNetOnce(model.param_init_net)
workspace.CreateNet(model.net)
return model
reader_name = "reader" if train_test_deplopy == 0 else "test_reader"
reader_data = args.train_data if train_test_deplopy == 0 else args.test_data
reader = model.CreateDB(reader_name,
db=reader_data,
db_type='lmdb',
num_shards=1,
shard_id=0)
is_test = True if train_test_deplopy == 1 else False
loss = None
with core.DeviceScope(device_opt):
with core.NameScope("{}_{}".format(model._device_prefix, 0)):
AddImageInput(model, reader, batch_size=32, is_test=is_test)
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=Initializer,
BiasInitializer=Initializer,
enable_tensor_core=False,
float16_compute=False):
pred = resnet.create_resnet50(model,
"data",
num_input_channels=3,
num_labels=args.num_labels,
no_bias=True,
no_loss=True)
softmax, loss = model.SoftmaxWithLoss([pred, 'label'],
['softmax', 'loss'])
brew.accuracy(model, [softmax, "label"], "accuracy")
#for test
if train_test_deplopy == 1:
workspace.RunNetOnce(model.param_init_net)
workspace.CreateNet(model.net)
return model
#for train
loss_grad = {}
losses_by_gpu = {}
losses_by_gpu[0] = [loss]
#add grad
def create_grad(lossp):
return model.ConstantFill(lossp, str(lossp) + "_grad", value=1.0)
# Explicitly need to create gradients on GPU 0
device = core.DeviceOption(model._device_type, 0)
with core.DeviceScope(device):
for l in losses_by_gpu[0]:
lg = create_grad(l)
loss_grad[str(l)] = str(lg)
model.AddGradientOperators(loss_grad)
#end add grad
optimizer.add_weight_decay(model, args.weight_decay)
stepsz = int(30 * args.epoch_size / 32)
opt = optimizer.build_multi_precision_sgd(model,
args.base_learning_rate,
momentum=0.9,
nesterov=1,
policy="step",
stepsize=stepsz,
gamma=0.1)
model._optimizer = opt
workspace.RunNetOnce(model.param_init_net)
workspace.CreateNet(model.net)
return model
def AddAccuracy(model, softmax, label, device_opts):
with core.DeviceScope(device_opts):
accuracy = brew.accuracy(model, [softmax, label], "accuracy")
return accuracy
def AddAccuracy(model, softmax, label):
"""Adds an accuracy op to the model"""
accuracy = brew.accuracy(model, [softmax, label], "accuracy")
return accuracy
def AddForwardPassOps(model, loss_scale, dtype):
"""Add forward pass ops and return a list of losses."""
initializer = (pFP16Initializer
if dtype == DataType.FLOAT16 else Initializer)
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=initializer,
BiasInitializer=initializer):
conv1 = brew.conv(model,
'data',
'conv1',
3,
32,
5,
pad=2,
weight_init=('GaussianFill', {
'std': 0.0001,
'mean': 0.0
}))
pool1 = brew.max_pool(model, conv1, 'pool1', kernel=3, stride=2)
relu1 = brew.relu(model, pool1, 'relu1')
conv2 = brew.conv(model,
relu1,
'conv2',
32,
32,
5,
pad=2,
weight_init=('GaussianFill', {
'std': 0.01
}))
conv2 = brew.relu(model, conv2, conv2)
pool2 = brew.average_pool(model, conv2, 'pool2', kernel=3, stride=2)
conv3 = brew.conv(model,
pool2,
'conv3',
32,
64,
5,
pad=2,
weight_init=('GaussianFill', {
'std': 0.01
}))
conv3 = brew.relu(model, conv3, conv3)
pool3 = brew.average_pool(model, conv3, 'pool3', kernel=3, stride=2)
fc1 = brew.fc(model,
pool3,
'fc1',
64 * 3 * 3,
64,
weight_init=('GaussianFill', {
'std': 0.1
}))
fc2 = brew.fc(model,
fc1,
'fc2',
64,
10,
weight_init=('GaussianFill', {
'std': 0.1
}))
if dtype == DataType.FLOAT16:
fc2 = model.net.HalfToFloat(fc2, fc2 + '_fp32')
softmax, loss = model.SoftmaxWithLoss([fc2, 'label'], ['softmax', 'loss'])
loss = model.Scale(loss, loss, scale=loss_scale)
brew.accuracy(model, [softmax, 'label'], 'accuracy')
return [loss]
def create_model_ops_testable(model, loss_scale, is_test=False):
initializer = (PseudoFP16Initializer
if args.dtype == 'float16' else Initializer)
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=initializer,
BiasInitializer=initializer,
enable_tensor_core=args.enable_tensor_core,
float16_compute=args.float16_compute):
if args.model == "cifar10":
if args.image_size != 32:
log.warn("Cifar10 expects a 32x32 image.")
pred = models.cifar10.create_cifar10(
model,
"data",
image_channels=args.num_channels,
num_classes=args.num_labels,
image_height=args.image_size,
image_width=args.image_size,
)
elif args.model == "resnet32x32":
if args.image_size != 32:
log.warn("ResNet32x32 expects a 32x32 image.")
pred = models.resnet.create_resnet32x32(
model,
"data",
num_layers=args.num_layers,
num_input_channels=args.num_channels,
num_labels=args.num_labels,
is_test=is_test)
elif args.model == "resnet":
if args.image_size != 224:
log.warn(
"ResNet expects a 224x224 image. input image = %d" %
args.image_size)
pred = resnet.create_resnet50(
#args.layers,
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
no_bias=True,
no_loss=True,
)
elif args.model == "vgg":
if args.image_size != 224:
log.warn("VGG expects a 224x224 image.")
pred = vgg.create_vgg(model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
num_layers=args.num_layers,
is_test=is_test)
elif args.model == "googlenet":
if args.image_size != 224:
log.warn("GoogLeNet expects a 224x224 image.")
pred = googlenet.create_googlenet(
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
is_test=is_test)
elif args.model == "alexnet":
if args.image_size != 224:
log.warn("Alexnet expects a 224x224 image.")
pred = alexnet.create_alexnet(
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
is_test=is_test)
elif args.model == "alexnetv0":
if args.image_size != 224:
log.warn("Alexnet v0 expects a 224x224 image.")
pred = alexnet.create_alexnetv0(
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
is_test=is_test)
else:
raise NotImplementedError("Network {} not found.".format(
args.model))
if args.dtype == 'float16':
pred = model.net.HalfToFloat(pred, pred + '_fp32')
softmax, loss = model.SoftmaxWithLoss([pred, 'label'],
['softmax', 'loss'])
loss = model.Scale(loss, scale=loss_scale)
brew.accuracy(model, [softmax, "label"], "accuracy")
return [loss]
def add_accuracy(model, probs, label):
accuracy = brew.accuracy(model, [probs, label], "accuracy")
return accuracy
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')
softmax = AddLeNetModel(train_model, data)
##################################################################################
#### Step 3: Add training operators to the model
# TODO: use the optimizer class here instead of doing sgd by hand
xent = train_model.LabelCrossEntropy(['softmax', 'label'], 'xent')
loss = train_model.AveragedLoss(xent, 'loss')
brew.accuracy(train_model, ['softmax', 'label'], 'accuracy')
train_model.AddGradientOperators([loss])
opt = optimizer.build_sgd(train_model, base_learning_rate=0.1)
for param in train_model.GetOptimizationParamInfo():
opt(train_model.net, train_model.param_init_net, param)
#model.Checkpoint([ITER] + model.params, [], db="mnist_lenet_checkpoint_%05d.lmdb", db_type="lmdb", every=20)
ITER = brew.iter(train_model, "iter")
train_model.Checkpoint([ITER] + train_model.params, [],
db="mnist_lenet_checkpoint_%05d.lmdb",
db_type="lmdb",
every=checkpoint_iters)
##################################################################################
#### Run the training procedure
load_all=1))
##### Externally initialize params so we can extract gradients
# for i,op in enumerate(init_net_proto.op):
# param_name = op.output[0]
# if param_name != 'data':
# print "param_name:", param_name
# assert(op.arg[0].name == "shape")
# tags = (ParameterTags.WEIGHT if param_name.endswith("_w") else ParameterTags.BIAS)
# model.create_param(param_name=op.output[0], shape=op.arg[0].ints, initializer=initializers.ExternalInitializer(), tags=tags)
# Add the "training operators" to the model
softmax = model_defs.Add_CNN_M(model, 'data', device_opts)
xent = model.LabelCrossEntropy([softmax, 'label'], 'xent')
loss = model.AveragedLoss(xent, "loss")
accuracy = brew.accuracy(model, [softmax, 'label'], "accuracy")
model.AddGradientOperators([loss])
# Instatiate test_dataset object
test_dataset = jdh.Jester_Dataset(dictionary_file=test_dictionary, seq_size=10)
# Prime the workspace with some data so we can run init net once
for image, label in test_dataset.read(batch_size=1):
workspace.FeedBlob("data", image)
workspace.FeedBlob("label", label)
break
# run the param init network once
workspace.RunNetOnce(model.param_init_net)
# create the network
workspace.CreateNet(model.net, overwrite=True)
def Accuracy(self, *args, **kwargs):
return brew.accuracy(self, *args, **kwargs)
