Skip to content

GitLab

Commit d77687a6 authored by Benjamin Graham's avatar Benjamin Graham Committed by Benjamin Thomas Graham
Browse files

Rename ValidConvolutions to SubmanifoldConvolutions, update for PyTorch 0.4 Tensor/Variable merge

parent 297e04c0
Show whitespace changes
Inline Side-by-side
Showing with 401 additions and 278 deletions
PyTorch/sparseconvnet/legacy/maxPooling.py
View file @ d77687a6
......@@ -26,7 +26,10 @@ class MaxPooling(SparseModule):
self.output.metadata = input.metadata
self.output.spatial_size =\
(input.spatial_size - self.pool_size) / self.pool_stride + 1
dim_typed_fn(self.dimension, input.features, 'MaxPooling_updateOutput')(
dim_typed_fn(
self.dimension,
input.features,
'MaxPooling_updateOutput')(
input.spatial_size,
self.output.spatial_size,
self.pool_size,
......@@ -39,7 +42,10 @@ class MaxPooling(SparseModule):
return self.output
def updateGradInput(self, input, gradOutput):
dim_typed_fn(self.dimension, input.features, 'MaxPooling_updateGradInput')(
dim_typed_fn(
self.dimension,
input.features,
'MaxPooling_updateGradInput')(
input.spatial_size,
self.output.spatial_size,
self.pool_size,
......
PyTorch/sparseconvnet/legacy/misc.py
View file @ d77687a6
......@@ -10,12 +10,13 @@ from .sparseModule import SparseModule
from ..sparseConvNetTensor import SparseConvNetTensor
from .batchNormalization import BatchNormalization
class Tanh(SparseModule):
def __init__(self):
SparseModule.__init__(self)
self.module=nn.Tanh()
self.module = nn.Tanh()
self.output = SparseConvNetTensor()
self.output.features=self.module.output
self.output.features = self.module.output
self.gradInput = self.module.gradInput
def updateOutput(self, input):
......@@ -25,19 +26,20 @@ class Tanh(SparseModule):
return self.output
def updateGradInput(self, input, gradOutput):
self.module.updateGradInput(input.features,gradOutput)
self.module.updateGradInput(input.features, gradOutput)
return self.gradInput
def type(self, t, tensorCache=None):
if t:
self.module.type(t,tensorCache)
self.output.features=self.module.output
self.module.type(t, tensorCache)
self.output.features = self.module.output
self.gradInput = self.module.gradInput
class ELU(SparseModule):
def __init__(self):
SparseModule.__init__(self)
self.module=nn.ELU()
self.module = nn.ELU()
self.output = SparseConvNetTensor()
self.gradInput = self.module.gradInput
......@@ -48,14 +50,20 @@ class ELU(SparseModule):
return self.output
def updateGradInput(self, input, gradOutput):
self.module.updateGradInput(input.features,gradOutput)
self.module.updateGradInput(input.features, gradOutput)
return self.gradInput
def type(self, t, tensorCache=None):
if t:
self.module.type(t,tensorCache)
self.output.features=self.module.output
self.module.type(t, tensorCache)
self.output.features = self.module.output
self.gradInput = self.module.gradInput
def BatchNormELU(nPlanes, eps=1e-4, momentum=0.9):
return Sequential().add(BatchNormalization(nPlanes,eps,momentum)).add(ELU())
return Sequential().add(
BatchNormalization(
nPlanes,
eps,
momentum)).add(
ELU())
PyTorch/sparseconvnet/legacy/networkArchitectures.py
View file @ d77687a6
......@@ -20,7 +20,7 @@ from .cAddTable import CAddTable
from .convolution import Convolution
from .deconvolution import Deconvolution
from .denseNetBlock import DenseNetBlock
from .validConvolution import ValidConvolution
from .submanifoldConvolution import SubmanifoldConvolution
from .networkInNetwork import NetworkInNetwork
from .batchNormalization import BatchNormReLU, BatchNormalizationInTensor
from .maxPooling import MaxPooling
......@@ -60,7 +60,7 @@ def DeepCNet(dimension, nInputPlanes, nPlanes, bn=True):
def SparseVggNet(dimension, nInputPlanes, layers):
"""
VGG style nets
Use valid convolutions
Use submanifold convolutions
Also implements 'Plus'-augmented nets
"""
nPlanes = nInputPlanes
......@@ -71,19 +71,19 @@ def SparseVggNet(dimension, nInputPlanes, layers):
elif x[0] == 'MP':
m.add(MaxPooling(dimension, x[1], x[2]))
elif x[0] == 'C' and len(x) == 2:
m.add(ValidConvolution(dimension, nPlanes, x[1], 3, False))
m.add(SubmanifoldConvolution(dimension, nPlanes, x[1], 3, False))
nPlanes = x[1]
m.add(BatchNormReLU(nPlanes))
elif x[0] == 'C' and len(x) == 3:
m.add(ConcatTable()
.add(
ValidConvolution(dimension, nPlanes, x[1], 3, False)
SubmanifoldConvolution(dimension, nPlanes, x[1], 3, False)
)
.add(
Sequential()
.add(Convolution(dimension, nPlanes, x[2], 3, 2, False))
.add(BatchNormReLU(x[2]))
.add(ValidConvolution(dimension, x[2], x[2], 3, False))
.add(SubmanifoldConvolution(dimension, x[2], x[2], 3, False))
.add(BatchNormReLU(x[2]))
.add(Deconvolution(dimension, x[2], x[2], 3, 2, False))
)).add(JoinTable([x[1], x[2]]))
......@@ -92,28 +92,28 @@ def SparseVggNet(dimension, nInputPlanes, layers):
elif x[0] == 'C' and len(x) == 4:
m.add(ConcatTable()
.add(
ValidConvolution(dimension, nPlanes, x[1], 3, False)
SubmanifoldConvolution(dimension, nPlanes, x[1], 3, False)
)
.add(
Sequential()
.add(Convolution(dimension, nPlanes, x[2], 3, 2, False))
.add(BatchNormReLU(x[2]))
.add(ValidConvolution(dimension, x[2], x[2], 3, False))
.add(SubmanifoldConvolution(dimension, x[2], x[2], 3, False))
.add(BatchNormReLU(x[2]))
.add(Deconvolution(dimension, x[2], x[2], 3, 2, False))
)
.add(Sequential()
.add(Convolution(dimension, nPlanes, x[3], 3, 2, False))
.add(BatchNormReLU(x[3]))
.add(ValidConvolution(dimension, x[3], x[3], 3, False))
.add(SubmanifoldConvolution(dimension, x[3], x[3], 3, False))
.add(BatchNormReLU(x[3]))
.add(Convolution(dimension, x[3], x[3], 3, 2, False))
.add(BatchNormReLU(x[3]))
.add(ValidConvolution(dimension, x[3], x[3], 3, False))
.add(SubmanifoldConvolution(dimension, x[3], x[3], 3, False))
.add(BatchNormReLU(x[3]))
.add(Deconvolution(dimension, x[3], x[3], 3, 2, False))
.add(BatchNormReLU(x[3]))
.add(ValidConvolution(dimension, x[3], x[3], 3, False))
.add(SubmanifoldConvolution(dimension, x[3], x[3], 3, False))
.add(BatchNormReLU(x[3]))
.add(Deconvolution(dimension, x[3], x[3], 3, 2, False))
)).add(JoinTable([x[1], x[2], x[3]]))
......@@ -145,7 +145,7 @@ def SparseResNet(dimension, nInputPlanes, layers):
m.add(
ConcatTable().add(
Sequential().add(
ValidConvolution(
SubmanifoldConvolution(
dimension,
nPlanes,
n,
......@@ -158,7 +158,7 @@ def SparseResNet(dimension, nInputPlanes, layers):
stride,
False)) .add(
BatchNormReLU(n)) .add(
ValidConvolution(
SubmanifoldConvolution(
dimension,
n,
n,
......@@ -173,14 +173,14 @@ def SparseResNet(dimension, nInputPlanes, layers):
ConcatTable().add(
Sequential().add(
BatchNormReLU(nPlanes)) .add(
ValidConvolution(
SubmanifoldConvolution(
dimension,
nPlanes,
n,
3,
False)) .add(
BatchNormReLU(n)) .add(
ValidConvolution(
SubmanifoldConvolution(
dimension,
n,
n,
......@@ -195,7 +195,7 @@ def SparseResNet(dimension, nInputPlanes, layers):
def SparseDenseNet(dimension, nInputPlanes, layers):
"""
SparseConvNet meets DenseNets using valid convolutions
SparseConvNet meets DenseNets using submanifold convolutions
Could do with a less confusing name
"""
nPlanes = nInputPlanes
......
PyTorch/sparseconvnet/legacy/sparseToDense.py
View file @ d77687a6
......@@ -27,12 +27,15 @@ class SparseToDense(SparseModule):
self.dimension = dimension
self.output = torch.Tensor()
self.gradInput = torch.FloatTensor()
self.nPlanes=nPlanes
self.nPlanes = nPlanes
def updateOutput(self, input):
if not self.nPlanes:
self.nPlanes=input.features.size(1)
dim_typed_fn(self.dimension, input.features, 'SparseToDense_updateOutput')(
self.nPlanes = input.features.size(1)
dim_typed_fn(
self.dimension,
input.features,
'SparseToDense_updateOutput')(
input.spatial_size,
input.metadata.ffi,
input.features,
......@@ -42,7 +45,10 @@ class SparseToDense(SparseModule):
return self.output
def updateGradInput(self, input, gradOutput):
dim_typed_fn(self.dimension, input.features, 'SparseToDense_updateGradInput')(
dim_typed_fn(
self.dimension,
input.features,
'SparseToDense_updateGradInput')(
input.spatial_size,
input.metadata.ffi,
input.features,
......
PyTorch/sparseconvnet/legacy/validConvolution.py → PyTorch/sparseconvnet/legacy/submanifoldConvolution.py
View file @ d77687a6
......@@ -10,7 +10,8 @@ from . import SparseModule
from ..utils import toLongTensor, dim_typed_fn, optionalTensor, nullptr
from ..sparseConvNetTensor import SparseConvNetTensor
class ValidConvolution(SparseModule):
class SubmanifoldConvolution(SparseModule):
def __init__(self, dimension, nIn, nOut, filter_size, bias):
SparseModule.__init__(self)
self.dimension = dimension
......@@ -31,11 +32,11 @@ class ValidConvolution(SparseModule):
self.gradInput = torch.Tensor()
def updateOutput(self, input):
assert input.features.ndimension()==0 or input.features.size(1) == self.nIn
assert input.features.ndimension() == 0 or input.features.size(1) == self.nIn
self.output.metadata = input.metadata
self.output.spatial_size = input.spatial_size
s.forward_pass_multiplyAdd_count +=\
dim_typed_fn(self.dimension, input.features, 'ValidConvolution_updateOutput')(
dim_typed_fn(self.dimension, input.features, 'SubmanifoldConvolution_updateOutput')(
input.spatial_size,
self.filter_size,
input.metadata.ffi,
......@@ -51,7 +52,10 @@ class ValidConvolution(SparseModule):
def backward(self, input, gradOutput, scale=1):
assert scale == 1
dim_typed_fn(self.dimension, input.features, 'ValidConvolution_backward')(
dim_typed_fn(
self.dimension,
input.features,
'SubmanifoldConvolution_backward')(
input.spatial_size,
self.filter_size,
input.metadata.ffi,
......@@ -60,7 +64,9 @@ class ValidConvolution(SparseModule):
gradOutput,
self.weight,
self.gradWeight,
optionalTensor(self, 'gradBias'),
optionalTensor(
self,
'gradBias'),
self.filter_volume,
torch.cuda.IntTensor() if input.features.is_cuda else nullptr)
return self.gradInput
......@@ -78,7 +84,8 @@ class ValidConvolution(SparseModule):
self.gradBias = self.gradBias.type(t)
def __repr__(self):
s = 'ValidConvolution ' + str(self.nIn) + '->' + str(self.nOut) + ' C'
s = 'SubmanifoldConvolution ' + \
str(self.nIn) + '->' + str(self.nOut) + ' C'
if self.filter_size.max() == self.filter_size.min():
s = s + str(self.filter_size[0])
else:
......
PyTorch/sparseconvnet/maxPooling.py
View file @ d77687a6
......@@ -4,11 +4,12 @@
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
from torch.autograd import Function, Variable
from torch.autograd import Function
from torch.nn import Module
from .utils import *
from .sparseConvNetTensor import SparseConvNetTensor
class MaxPoolingFunction(Function):
@staticmethod
def forward(
......@@ -21,15 +22,10 @@ class MaxPoolingFunction(Function):
pool_size,
pool_stride,
nFeaturesToDrop):
ctx.input_features=input_features
ctx.input_metadata=input_metadata
ctx.input_spatial_size = input_spatial_size
ctx.output_spatial_size = output_spatial_size
ctx.input_metadata = input_metadata
ctx.dimension = dimension
ctx.pool_size = pool_size
ctx.pool_stride = pool_stride
ctx.nFeaturesToDrop = nFeaturesToDrop
ctx.output_features = input_features.new()
output_features = input_features.new()
dim_typed_fn(dimension, input_features, 'MaxPooling_updateOutput')(
input_spatial_size,
output_spatial_size,
......@@ -37,27 +33,40 @@ class MaxPoolingFunction(Function):
pool_stride,
input_metadata.ffi,
input_features,
ctx.output_features,
output_features,
nFeaturesToDrop,
torch.cuda.IntTensor() if input_features.is_cuda else nullptr)
return ctx.output_features
ctx.save_for_backward(
input_features,
output_features,
input_spatial_size,
output_spatial_size,
pool_size,
pool_stride)
return output_features
@staticmethod
def backward(ctx, grad_output):
grad_input=Variable(grad_output.data.new())
input_features,\
output_features,\
input_spatial_size,\
output_spatial_size,\
pool_size,\
pool_stride = ctx.saved_tensors
grad_input = grad_output.new()
dim_typed_fn(
ctx.dimension, ctx.input_features, 'MaxPooling_updateGradInput')(
ctx.input_spatial_size,
ctx.output_spatial_size,
ctx.pool_size,
ctx.pool_stride,
ctx.dimension, input_features, 'MaxPooling_updateGradInput')(
input_spatial_size,
output_spatial_size,
pool_size,
pool_stride,
ctx.input_metadata.ffi,
ctx.input_features,
grad_input.data,
ctx.output_features,
grad_output.data,
input_features,
grad_input,
output_features,
grad_output,
ctx.nFeaturesToDrop,
torch.cuda.IntTensor() if ctx.input_features.is_cuda else nullptr)
torch.cuda.IntTensor() if input_features.is_cuda else nullptr)
return grad_input, None, None, None, None, None, None, None
......@@ -68,29 +77,39 @@ class MaxPooling(Module):
self.pool_size = toLongTensor(dimension, pool_size)
self.pool_stride = toLongTensor(dimension, pool_stride)
self.nFeaturesToDrop = nFeaturesToDrop
def forward(self, input):
output = SparseConvNetTensor()
output.metadata = input.metadata
output.spatial_size = (
input.spatial_size - self.pool_size) / self.pool_stride + 1
assert ((output.spatial_size-1)*self.pool_stride+self.pool_size==input.spatial_size).all()
output.features = MaxPoolingFunction().apply(
input.features, input.metadata, input.spatial_size,
output.spatial_size, self.dimension,self.pool_size,self.pool_stride,
assert ((output.spatial_size - 1) * self.pool_stride +
self.pool_size == input.spatial_size).all()
output.features = MaxPoolingFunction.apply(
input.features,
input.metadata,
input.spatial_size,
output.spatial_size,
self.dimension,
self.pool_size,
self.pool_stride,
self.nFeaturesToDrop)
return output
def input_spatial_size(self, out_size):
return (out_size - 1) * self.pool_stride + self.pool_size
def __repr__(self):
s = 'MaxPooling'
if self.pool_size.max() == self.pool_size.min() and\
self.pool_stride.max() == self.pool_stride.min():
s = s + str(self.pool_size[0]) + '/' + str(self.pool_stride[0])
s = s + str(self.pool_size[0].item()) + \
'/' + str(self.pool_stride[0].item())
else:
s = s + '(' + str(self.pool_size[0])
s = s + '(' + str(self.pool_size[0].item())
for i in self.pool_size[1:]:
s = s + ',' + str(i)
s = s + ')/(' + str(self.pool_stride[0])
s = s + ')/(' + str(self.pool_stride[0].item())
for i in self.pool_stride[1:]:
s = s + ',' + str(i)
s = s + ')'
......
PyTorch/sparseconvnet/metadata.py
View file @ d77687a6
......@@ -6,7 +6,7 @@
"""
Store Metadata relating to which spatial locations are active at each scale.
Convolutions, valid convolutions and 'convolution reversing' deconvolutions
Convolutions, submanifold convolutions and 'convolution reversing' deconvolutions
all coexist within the same MetaData object as long as each spatial size
only occurs once.
......
PyTorch/sparseconvnet/networkArchitectures.py
View file @ d77687a6
......@@ -21,7 +21,7 @@ from .tables import *
def SparseVggNet(dimension, nInputPlanes, layers):
"""
VGG style nets
Use valid convolutions
Use submanifold convolutions
Also implements 'Plus'-augmented nets
"""
nPlanes = nInputPlanes
......@@ -216,7 +216,7 @@ def ResNetUNet(dimension, nPlanes, reps, depth=4):
def __init__(self):
nn.Module.__init__(self)
self.sparseModel = scn.Sequential().add(
scn.ValidConvolution(3, nInputFeatures, 64, 3, False)).add(
scn.SubmanifoldConvolution(3, nInputFeatures, 64, 3, False)).add(
scn.ResNetUNet(3, 64, 2, 4))
self.linear = nn.Linear(64, nClasses)
def forward(self,x):
......
PyTorch/sparseconvnet/networkInNetwork.py
View file @ d77687a6
......@@ -5,11 +5,12 @@
# LICENSE file in the root directory of this source tree.
import sparseconvnet
from torch.autograd import Function, Variable
from torch.autograd import Function
from torch.nn import Module, Parameter
from .utils import *
from .sparseConvNetTensor import SparseConvNetTensor
class NetworkInNetworkFunction(Function):
@staticmethod
def forward(
......@@ -17,37 +18,44 @@ class NetworkInNetworkFunction(Function):
input_features,
weight,
bias):
ctx.input_features=input_features
ctx.weight=weight
ctx.bias=bias
ctx.output_features=input_features.new()
output_features = input_features.new()
ctx.save_forbackwards(input_features,
output_features,
weight,
bias)
sparseconvnet.forward_pass_multiplyAdd_count +=\
typed_fn(input_features, 'NetworkInNetwork_updateOutput')(
input_features,
ctx.output_features,
output_features,
weight,
bias if bias is not None else nullptr)
sparseconvnet.forward_pass_hidden_states += ctx.output_features.nelement()
return ctx.output_features
sparseconvnet.forward_pass_hidden_states += output_features.nelement()
return output_features
@staticmethod
def backward(ctx, grad_output):
grad_input=Variable(grad_output.data.new())
grad_weight=Variable(grad_output.data.new().resize_as_(ctx.weight).zero_())
if ctx.bias is None:
grad_bias=None
input_features,\
output_features,\
weight,\
bias = ctx.saved_tensors
grad_input = grad_output.new()
grad_weight = grad_output.new().resize_as_(weight).zero_()
if bias is None:
grad_bias = None
else:
grad_bias = Variable(grad_output.data.new().resize_as_(ctx.bias))
typed_fn(ctx.input_features, 'NetworkInNetwork_updateGradInput')(
grad_input.data,
grad_output.data,
ctx.weight)
typed_fn(ctx.input_features, 'NetworkInNetwork_accGradParameters')(
ctx.input_features,
grad_output.data,
grad_weight.data,
grad_bias = grad_output.new().resize_as_(bias)
typed_fn(input_features, 'NetworkInNetwork_updateGradInput')(
grad_input,
grad_output,
weight)
typed_fn(input_features, 'NetworkInNetwork_accGradParameters')(
input_features,
grad_output,
grad_weight,
grad_bias.data if grad_bias is not None else nullptr)
return grad_input, grad_weight, grad_bias
class NetworkInNetwork(Module):
def __init__(self, nIn, nOut, bias=False):
Module.__init__(self)
......@@ -62,13 +70,14 @@ class NetworkInNetwork(Module):
self.bias = Parameter(torch.Tensor(nOut).
_())
else:
self.bias=None
self.bias = None
def forward(self, input):
assert input.features.ndimension()==0 or input.features.size(1) == self.nIn
assert input.features.ndimension() == 0 or input.features.size(1) == self.nIn
output = SparseConvNetTensor()
output.metadata = input.metadata
output.spatial_size = input.spatial_size
output.features=NetworkInNetworkFunction().apply(
output.features = NetworkInNetworkFunction.apply(
input.features,
self.weight,
self.bias)
......
PyTorch/sparseconvnet/sequential.py
View file @ d77687a6
......@@ -7,11 +7,13 @@
from torch.nn import Sequential as S
from .utils import set
class Sequential(S):
def input_spatial_size(self, out_size):
for m in reversed(self._modules):
out_size = self._modules[m].input_spatial_size(out_size)
return out_size
def add(self, module):
self._modules[str(len(self._modules))]=module
self._modules[str(len(self._modules))] = module
return self
PyTorch/sparseconvnet/sparseConvNetTensor.py
View file @ d77687a6
......@@ -22,7 +22,12 @@ class SparseConvNetTensor(object):
spatial_size = self.spatial_size
t = torch.LongTensor()
dim_fn(self.metadata.dimension, 'getSpatialLocations')(self.metadata.ffi, spatial_size, t)
dim_fn(
self.metadata.dimension,
'getSpatialLocations')(
self.metadata.ffi,
spatial_size,
t)
return t
def type(self, t=None):
......@@ -50,5 +55,8 @@ class SparseConvNetTensor(object):
def to_variable(self, requires_grad=False, volatile=False):
"Convert self.features to a variable for use with modern PyTorch interface."
self.features = Variable(self.features, requires_grad=requires_grad, volatile=volatile)
self.features = Variable(
self.features,
requires_grad=requires_grad,
volatile=volatile)
return self
PyTorch/sparseconvnet/sparseToDense.py
View file @ d77687a6
......@@ -15,11 +15,12 @@ Parameters:
dimension : of the input field,
"""
from torch.autograd import Function, Variable
from torch.autograd import Function
from torch.nn import Module
from .utils import *
from .sparseConvNetTensor import SparseConvNetTensor
class SparseToDenseFunction(Function):
@staticmethod
def forward(
......@@ -29,12 +30,14 @@ class SparseToDenseFunction(Function):
spatial_size,
dimension,
nPlanes):
ctx.input_metadata=input_metadata
ctx.spatial_size=spatial_size
ctx.dimension=dimension
ctx.input_features=input_features
ctx.input_metadata = input_metadata
ctx.dimension = dimension
ctx.save_for_backward(input_features, spatial_size)
output = input_features.new()
dim_typed_fn(ctx.dimension, input_features, 'SparseToDense_updateOutput')(
dim_typed_fn(
ctx.dimension,
input_features,
'SparseToDense_updateOutput')(
spatial_size,
input_metadata.ffi,
input_features,
......@@ -42,28 +45,41 @@ class SparseToDenseFunction(Function):
torch.cuda.IntTensor() if input_features.is_cuda else nullptr,
nPlanes)
return output
@staticmethod
def backward(ctx, grad_output):
grad_input=Variable(grad_output.data.new())
dim_typed_fn(ctx.dimension, ctx.input_features, 'SparseToDense_updateGradInput')(
ctx.spatial_size,
grad_input = grad_output.new()
input_features, spatial_size = ctx.saved_tensors
dim_typed_fn(
ctx.dimension,
input_features,
'SparseToDense_updateGradInput')(
spatial_size,
ctx.input_metadata.ffi,
ctx.input_features,
grad_input.data,
grad_output.data,
torch.cuda.IntTensor() if ctx.input_features.is_cuda else nullptr)
input_features,
grad_input,
grad_output,
torch.cuda.IntTensor() if input_features.is_cuda else nullptr)
return grad_input, None, None, None, None
class SparseToDense(Module):
def __init__(self, dimension, nPlanes):
Module.__init__(self)
self.dimension = dimension
self.nPlanes=nPlanes
self.nPlanes = nPlanes
def forward(self, input):
return SparseToDenseFunction().apply(input.features,input.metadata,input.spatial_size,self.dimension,self.nPlanes)
return SparseToDenseFunction.apply(
input.features,
input.metadata,
input.spatial_size,
self.dimension,
self.nPlanes)
def input_spatial_size(self, out_size):
return out_size
def __repr__(self):
return 'SparseToDense(' + str(self.dimension) + ','+ str(self.nPlanes)+ ')'
return 'SparseToDense(' + str(self.dimension) + \
',' + str(self.nPlanes) + ')'
PyTorch/sparseconvnet/submanifoldConvolution.py
View file @ d77687a6
......@@ -4,16 +4,16 @@
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# 'SubmanifoldConvolution == ValidConvolution'
# 'SubmanifoldConvolution == SubmanifoldConvolution'
import sparseconvnet
from torch.autograd import Function, Variable
from torch.autograd import Function
from torch.nn import Module, Parameter
from .utils import *
from .sparseConvNetTensor import SparseConvNetTensor
class ValidConvolutionFunction(Function):
class SubmanifoldConvolutionFunction(Function):
@staticmethod
def forward(
ctx,
......@@ -24,59 +24,68 @@ class ValidConvolutionFunction(Function):
spatial_size,
dimension,
filter_size):
ctx.input_features=input_features
ctx.input_metadata=input_metadata
ctx.spatial_size=spatial_size
ctx.weight=weight
ctx.bias=bias
ctx.output_features=input_features.new()
ctx.dimension=dimension
ctx.filter_size=filter_size
ctx.input_metadata = input_metadata
ctx.dimension = dimension
# ctx.input_features=input_features
# ctx.spatial_size=spatial_size
# ctx.weight=weight
# ctx.bias=bias
# ctx.filter_size=filter_size
output_features = input_features.new()
ctx.save_for_backward(
input_features,
spatial_size,
weight,
bias,
filter_size)
sparseconvnet.forward_pass_multiplyAdd_count +=\
dim_typed_fn(
dimension, input_features, 'ValidConvolution_updateOutput')(
dimension, input_features, 'SubmanifoldConvolution_updateOutput')(
spatial_size,
filter_size,
input_metadata.ffi,
input_features,
ctx.output_features,
weight,
bias if bias is not None else nullptr,
0, #remove this parameter!!
input_features.data,
output_features.data,
weight.data,
bias.data if bias is not None else nullptr,
0, # remove this parameter!!
torch.cuda.IntTensor() if input_features.is_cuda else nullptr)
sparseconvnet.forward_pass_hidden_states += ctx.output_features.nelement()
return ctx.output_features
sparseconvnet.forward_pass_hidden_states += output_features.nelement()
return output_features
@staticmethod
def backward(ctx, grad_output):
grad_input=Variable(grad_output.data.new())
grad_weight=Variable(grad_output.data.new().resize_as_(ctx.weight).zero_())
if ctx.bias is None:
grad_bias=None
input_features, spatial_size, weight, bias, filter_size = ctx.saved_tensors
grad_input = grad_output.new()
grad_weight = grad_output.new().resize_as_(weight).zero_()
if bias is None:
grad_bias = None
else:
grad_bias = Variable(grad_output.data.new().resize_as_(ctx.bias).zero_())
grad_bias = grad_output.new().resize_as_(bias).zero_()
dim_typed_fn(
ctx.dimension, ctx.input_features, 'ValidConvolution_backward')(
ctx.spatial_size,
ctx.filter_size,
ctx.dimension, input_features, 'SubmanifoldConvolution_backward')(
spatial_size,
filter_size,
ctx.input_metadata.ffi,
ctx.input_features,
grad_input.data,
grad_output.data.contiguous(),
ctx.weight,
grad_weight.data,
input_features,
grad_input,
grad_output.contiguous(),
weight,
grad_weight,
grad_bias.data if grad_bias is not None else nullptr,
0, #remove this parameter
torch.cuda.IntTensor() if ctx.input_features.is_cuda else nullptr)
0, # remove this parameter
torch.cuda.IntTensor() if input_features.is_cuda else nullptr)
return grad_input, grad_weight, grad_bias, None, None, None, None
class ValidConvolution(Module):
class SubmanifoldConvolution(Module):
def __init__(self, dimension, nIn, nOut, filter_size, bias):
Module.__init__(self)
self.dimension = dimension
self.nIn = nIn
self.nOut = nOut
self.filter_size = toLongTensor(dimension, filter_size)
self.filter_volume = self.filter_size.prod()
self.filter_volume = self.filter_size.prod().item()
std = (2.0 / nIn / self.filter_volume)**0.5
self.weight = Parameter(torch.Tensor(
nIn * self.filter_volume, nOut
......@@ -87,11 +96,11 @@ class ValidConvolution(Module):
self.bias = None
def forward(self, input):
assert input.features.ndimension()==0 or input.features.size(1) == self.nIn
assert input.features.ndimension() == 0 or input.features.size(1) == self.nIn
output = SparseConvNetTensor()
output.metadata = input.metadata
output.spatial_size = input.spatial_size
output.features=ValidConvolutionFunction.apply(
output.features = SubmanifoldConvolutionFunction.apply(
input.features,
self.weight,
self.bias,
......@@ -102,11 +111,12 @@ class ValidConvolution(Module):
return output
def __repr__(self):
s = 'ValidConvolution ' + str(self.nIn) + '->' + str(self.nOut) + ' C'
s = 'SubmanifoldConvolution ' + \
str(self.nIn) + '->' + str(self.nOut) + ' C'
if self.filter_size.max() == self.filter_size.min():
s = s + str(self.filter_size[0])
s = s + str(self.filter_size[0].item())
else:
s = s + '(' + str(self.filter_size[0])
s = s + '(' + str(self.filter_size[0].item())
for i in self.filter_size[1:]:
s = s + ',' + str(i)
s = s + ')'
......@@ -115,5 +125,6 @@ class ValidConvolution(Module):
def input_spatial_size(self, out_size):
return out_size
class SubmanifoldConvolution(ValidConvolution):
class ValidConvolution(SubmanifoldConvolution):
pass
PyTorch/sparseconvnet/tables.py
View file @ d77687a6
......@@ -4,36 +4,43 @@
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
from torch.autograd import Function, Variable
from torch.autograd import Function
from torch.nn import Module
from .utils import *
from .sparseConvNetTensor import SparseConvNetTensor
class JoinTable(Module):
def forward(self, input):
output = SparseConvNetTensor()
output.metadata = input[0].metadata
output.spatial_size = input[0].spatial_size
output.features=torch.cat([i.features for i in input],1)
output.features = torch.cat([i.features for i in input], 1)
return output
def input_spatial_size(self,out_size):
def input_spatial_size(self, out_size):
return out_size
class AddTable(Module):
def forward(self, input):
output = SparseConvNetTensor()
output.metadata = input[0].metadata
output.spatial_size = input[0].spatial_size
output.features=sum([i.features for i in input])
output.features = sum([i.features for i in input])
return output
def input_spatial_size(self,out_size):
def input_spatial_size(self, out_size):
return out_size
class ConcatTable(Module):
def forward(self, input):
return [module(input) for module in self._modules.values()]
def add(self, module):
self._modules[str(len(self._modules))]=module
self._modules[str(len(self._modules))] = module
return self
def input_spatial_size(self,out_size):
def input_spatial_size(self, out_size):
return self._modules['0'].input_spatial_size(out_size)
PyTorch/sparseconvnet/utils.py
View file @ d77687a6
......@@ -10,8 +10,11 @@ from cffi import FFI
def toLongTensor(dimension, x):
if type(x).__name__ == 'LongTensor':
if hasattr(x, 'type') and x.type() == 'torch.LongTensor':
return x
elif isinstance(x, (list, tuple)):
assert len(x) == dimension
return torch.LongTensor(x)
else:
return torch.LongTensor(dimension).fill_(x)
......@@ -28,16 +31,18 @@ def dim_fn(dimension, name):
def typed_fn(t, name):
# print('typed_fn',t.features.type(),name)
# print('typed_fn',t.type(),name)
return getattr(scn, 'scn_' + typeTable[t.type()] + '_' + name)
def dim_typed_fn(dimension, t, name):
# print('dim_typed_fn',dimension,t.features.type(),name)
# print('dim_typed_fn',dimension,t.type(),name)
return getattr(scn, 'scn_' +
typeTable[t.type()] +
str(dimension) +
name)
ffi = FFI()
nullptr = ffi.NULL
......@@ -69,6 +74,14 @@ def threadDatasetIterator(d):
return iterator
# def threadDatasetIterator(d):
# print('not threads!!!')
# def iterator():
# for x in d:
# yield x
# return iterator
def set(obj):
if hasattr(obj, 'storage_type'):
obj.set_(obj.storage_type()())
......
README.md
View file @ d77687a6
......@@ -17,7 +17,7 @@ This is the Torch/PyTorch library for training Submanifold Sparse Convolutional
## Spatial sparsity
This library brings [Spatially-sparse convolutional networks](https://github.com/btgraham/SparseConvNet) to Torch/PyTorch. Moreover, it introduces **Submanifold Sparse Convolutions**, that can be used to build computationally efficient sparse VGG/ResNet/DenseNet-style networks.
This library brings [Spatially-sparse convolutional networks](https://github.com/btgraham/SparseConvNet) to PyTorch and [Torch classic](README_Torch.md). Moreover, it introduces **Submanifold Sparse Convolutions**, that can be used to build computationally efficient sparse VGG/ResNet/DenseNet-style networks.
With regular 3x3 convolutions, the set of active (non-zero) sites grows rapidly:<br />
![submanifold](img/i.gif) <br />
......@@ -122,83 +122,9 @@ output = model.forward(input)
# Output is 2x32x10x10: our minibatch has 2 samples, the network has 32 output
# feature planes, and 10x10 is the spatial size of the output.
print(output.size(), output.data.type())
print(output.size(), output.type())
```
## Hello World - (Lua)Torch
Convolutional networks are built with SparseConvNet in the same way as with Torch's nn/cunn/cudnn packages.
```
--Train on the GPU if there is one, otherwise CPU
scn=require 'sparseconvnet'
tensorType = scn.cutorch and 'torch.CudaTensor' or 'torch.FloatTensor'
model = scn.Sequential()
:add(scn.SparseVggNet(2,1,{ --dimension 2, 1 input plane
{'C', 8}, -- 3x3 VSC convolution, 8 output planes, batchnorm, ReLU
{'C', 8}, -- and another
{'MP', 3, 2}, --max pooling, size 3, stride 2
{'C', 16}, -- etc
{'C', 16},
{'MP', 3, 2},
{'C', 24},
{'C', 24},
{'MP', 3, 2}}))
:add(scn.Convolution(2,24,32,3,1,false)) --an SC convolution on top
:add(scn.BatchNormReLU(32))
:add(scn.SparseToDense(2))
:type(tensorType)
--[[
To use the network we must create an scn.InputBatch with right dimensionality.
If we want the output to have spatial size 10x10, we can find the appropriate
input size, give that we uses three layers of MP3/2 max-pooling, and finish
with a SC convoluton
]]
inputSpatialSize=model:suggestInputSize(torch.LongTensor{10,10}) --103x103
input=scn.InputBatch(2,inputSpatialSize)
--Now we build the input batch, sample by sample, and active site by active site.
msg={
" O O OOO O O OO O O OO OOO O OOO ",
" O O O O O O O O O O O O O O O O ",
" OOOOO OO O O O O O O O O O OOO O O O ",
" O O O O O O O O O O O O O O O O O O ",
" O O OOO OOO OOO OO O O OO O O OOO OOO ",
}
input:addSample()
for y,line in ipairs(msg) do
for x = 1,string.len(line) do
if string.sub(line,x,x) == 'O' then
local location = torch.LongTensor{x,y}
local featureVector = torch.FloatTensor{1}
input:setLocation(location,featureVector,0)
end
end
end
--[[
Optional: allow metadata preprocessing to be done in batch preparation threads
to improve GPU utilization.
Parameter:
3 if using MP3/2 or size-3 stride-2 convolutions for downsizeing,
2 if using MP2
]]
input:precomputeMetadata(3)
model:evaluate()
input:type(tensorType)
output = model:forward(input)
--[[
Output is 1x32x10x10: our minibatch has 1 sample, the network has 32 output
feature planes, and 10x10 is the spatial size of the output.
]]
print(output:size())
```
## Examples
......@@ -209,45 +135,33 @@ Examples in the examples folder include
Data will be downloaded/preprocessed on the first run, i.e.
```
cd examples/Assamese_handwriting
th VGGplus.lua
or
cd examples/Assamese_handwriting
python VGGplus.py
```
## Setup
## PyTorch Setup
Tested with Ubuntu 16.04, Python 3 in [Miniconda](https://conda.io/miniconda.html) and PyTorch master (v0.4 with merged Tensors/Variables).
Tested with Ubuntu 16.04. Install [Torch](http://torch.ch/docs/getting-started.html) and/or [PyTorch](http://pytorch.org/) ([Miniconda](https://conda.io/miniconda.html)) then: <br />
```
git clone https://github.com/pytorch/pytorch.git
cd pytorch
python setup.py install
cd ..
apt-get install libsparsehash-dev
git clone git@github.com:facebookresearch/SparseConvNet.git
then
cd SparseConvNet/Torch/
luarocks make sparseconvnet-0.1-1.rockspec
and/or
cd SparseConvNet/PyTorch/
python setup.py develop
python setup.py install
```
To run the examples you may also need to install unrar and TorchNet:
```
apt-get install unrar
and
luarocks install torchnet
or
pip install git+https://github.com/pytorch/tnt.git@master
```
### Links
1. [ICDAR 2013 Chinese Handwriting Recognition Competition 2013](http://www.nlpr.ia.ac.cn/events/CHRcompetition2013/competition/Home.html) First place in task 3, with test error of 2.61%. Human performance on the test set was 4.81%. [Report](http://www.nlpr.ia.ac.cn/events/CHRcompetition2013/competition/ICDAR%202013%20CHR%20competition.pdf)
2. [Spatially-sparse convolutional neural networks, 2014](http://arxiv.org/abs/1409.6070) SparseConvNets for Chinese handwriting recognition
......
README_Torch.md 0 → 100644
View file @ d77687a6
## Hello World - (Lua)Torch
Convolutional networks are built with SparseConvNet in the same way as with Torch's nn/cunn/cudnn packages.
```
--Train on the GPU if there is one, otherwise CPU
scn=require 'sparseconvnet'
tensorType = scn.cutorch and 'torch.CudaTensor' or 'torch.FloatTensor'
model = scn.Sequential()
:add(scn.SparseVggNet(2,1,{ --dimension 2, 1 input plane
{'C', 8}, -- 3x3 VSC convolution, 8 output planes, batchnorm, ReLU
{'C', 8}, -- and another
{'MP', 3, 2}, --max pooling, size 3, stride 2
{'C', 16}, -- etc
{'C', 16},
{'MP', 3, 2},
{'C', 24},
{'C', 24},
{'MP', 3, 2}}))
:add(scn.Convolution(2,24,32,3,1,false)) --an SC convolution on top
:add(scn.BatchNormReLU(32))
:add(scn.SparseToDense(2))
:type(tensorType)
--[[
To use the network we must create an scn.InputBatch with right dimensionality.
If we want the output to have spatial size 10x10, we can find the appropriate
input size, give that we uses three layers of MP3/2 max-pooling, and finish
with a SC convoluton
]]
inputSpatialSize=model:suggestInputSize(torch.LongTensor{10,10}) --103x103
input=scn.InputBatch(2,inputSpatialSize)
--Now we build the input batch, sample by sample, and active site by active site.
msg={
" O O OOO O O OO O O OO OOO O OOO ",
" O O O O O O O O O O O O O O O O ",
" OOOOO OO O O O O O O O O O OOO O O O ",
" O O O O O O O O O O O O O O O O O O ",
" O O OOO OOO OOO OO O O OO O O OOO OOO ",
}
input:addSample()
for y,line in ipairs(msg) do
for x = 1,string.len(line) do
if string.sub(line,x,x) == 'O' then
local location = torch.LongTensor{x,y}
local featureVector = torch.FloatTensor{1}
input:setLocation(location,featureVector,0)
end
end
end
--[[
Optional: allow metadata preprocessing to be done in batch preparation threads
to improve GPU utilization.
Parameter:
3 if using MP3/2 or size-3 stride-2 convolutions for downsizing,
2 if using MP2
]]
input:precomputeMetadata(3)
model:evaluate()
input:type(tensorType)
output = model:forward(input)
--[[
Output is 1x32x10x10: our minibatch has 1 sample, the network has 32 output
feature planes, and 10x10 is the spatial size of the output.
]]
print(output:size())
```
## Torch Setup
Tested with Ubuntu 16.04.
Install [Torch](http://torch.ch/docs/getting-started.html) then: <br />
```
apt-get install libsparsehash-dev
git clone git@github.com:facebookresearch/SparseConvNet.git
then
cd SparseConvNet/Torch/
luarocks make sparseconvnet-0.1-1.rockspec
```
To run the examples you may also need to install unrar and TorchNet:
```
apt-get install unrar
and
luarocks install torchnet
```
Torch/C.lua
View file @ d77687a6
......@@ -227,11 +227,11 @@ void scn_ARCH_REAL_DIMENSIONSparseToDense_updateGradInput(
THTensor *d_input_features, THTensor *d_output_features,
THITensor *rulesBuffer);
double scn_ARCH_REAL_DIMENSIONValidConvolution_updateOutput(
double scn_ARCH_REAL_DIMENSIONSubmanifoldConvolution_updateOutput(
THLongTensor *inputSize, THLongTensor *filterSize, void **m,
THTensor *input_features, THTensor *output_features, THTensor *weight,
THTensor *bias, long filterVolume, THITensor *rulesBuffer);
void scn_ARCH_REAL_DIMENSIONValidConvolution_backward(
void scn_ARCH_REAL_DIMENSIONSubmanifoldConvolution_backward(
THLongTensor *inputSize, THLongTensor *filterSize, void **m,
THTensor *input_features, THTensor *d_input_features,
THTensor *d_output_features, THTensor *weight, THTensor *d_weight,
......
Torch/CAddTable.lua
View file @ d77687a6
......@@ -7,7 +7,7 @@
--[[
Assume all the inputs have identical SparseGrids and input[i].nActive
Assume input[1].nPlanes >= input[i].nPlanes for all i=1,#input
output.validRules is taken from input[1].validRules (could do set union?)
output.submanifoldRules is taken from input[1].submanifoldRules (could do set union?)
(for resnets, make sure the residual link is input[2])
]]
......
Torch/DenseNetBlock.lua
View file @ d77687a6
......@@ -37,7 +37,7 @@ return function(sparseconvnet)
--Module 4*i-1
self:add(sparseconvnet.BatchNormalization(nFeaturesB,nil,nil,true,0))
--Module 4*i
self:add(sparseconvnet.ValidConvolution(dimension, nFeaturesB, growthRate,
self:add(sparseconvnet.SubmanifoldConvolution(dimension, nFeaturesB, growthRate,
3, false))
--Module 4*i+1
self:add(sparseconvnet.BatchNormalizationInTensor(growthRate,nil,nil,
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment