Merge branch 'master' into batchwise_dropout

PyTorch/sparseconvnet/legacy/__init__.py

@@ -10,12 +10,13 @@ from .inputBatch import InputBatch
 from .sparseConvNetTensor import SparseConvNetTensor
 from .sparseModule import SparseModule
 from .averagePooling import AveragePooling
-from .batchNormalization import BatchNormalization
+from .batchNormalization import BatchNormalization, BatchNormReLU, BatchNormLeakyReLU, BatchNormalizationInTensor
 from .batchwiseDropout import BatchwiseDropout
 from .concatTable import ConcatTable
 from .convolution import Convolution
 from .cAddTable import CAddTable
 from .deconvolution import Deconvolution
+from .denseToSparse import DenseToSparse
 from .identity import Identity
 from .joinTable import JoinTable
 from .leakyReLU import LeakyReLU
@@ -27,3 +28,4 @@ from .sparseToDense import SparseToDense
 from .validConvolution import ValidConvolution
 from .networkArchitectures import *
 from .classificationTrainValidate import ClassificationTrainValidate
+from .misc import *

PyTorch/sparseconvnet/legacy/affineReLUTrivialConvolution.py

@@ -28,13 +28,11 @@ class AffineReLUTrivialConvolution(SparseModule):
         self.nOut = nOut
         self.affineWeight = torch.Tensor(nIn).fill_(1)
         self.affineBias = torch.Tensor(nIn).zero_()
-        self.convWeight = torch.Tensor(
-            nIn, nOut).normal_(
-            0, math.sqrt(
-                2.0 / nIn))
+        std = math.sqrt(2.0 / nIn)
+        self.convWeight = torch.Tensor(nIn, nOut).normal_(0, std)
         self.gradAffineWeight = torch.Tensor(nIn).fill_(0)
-        self.gradAffineBias = torch.Tensor(nIn).zero_()
-        self.gradConvWeight = torch.Tensor(nIn, nOut).zero_()
+        self.gradAffineBias = torch.Tensor(nIn).zero_(0.333)
+        self.gradConvWeight = torch.Tensor(nIn, nOut).fill_(std)
         self.additiveGrad = additiveGrad
         self.output = SparseConvNetTensor(torch.Tensor())
         self.gradInput = torch.Tensor()
@@ -46,7 +44,7 @@ class AffineReLUTrivialConvolution(SparseModule):
     def updateOutput(self, input):
         self.output.metadata = input.metadata
         self.output.spatial_size = input.spatial_size
-        typed_fn(input, 'AffineReluTrivialConvolution_updateOutput')(
+        typed_fn(input.features, 'AffineReluTrivialConvolution_updateOutput')(
             input.features,
             self.output.features,
             self.affineWeight,
@@ -59,7 +57,7 @@ class AffineReLUTrivialConvolution(SparseModule):
 
     def backward(self, input, gradOutput, scale=1):
         assert scale == 1
-        typed_fn(input, 'AffineReluTrivialConvolution_backward')(
+        typed_fn(input.features, 'AffineReluTrivialConvolution_backward')(
             input.features,
             self.gradInput,
             gradOutput,

PyTorch/sparseconvnet/legacy/averagePooling.py

@@ -17,7 +17,6 @@ class AveragePooling(SparseModule):
         self.dimension = dimension
         self.pool_size = toLongTensor(dimension, pool_size)
         self.pool_stride = toLongTensor(dimension, pool_stride)
-        self.pool_volume = self.pool_size.prod()
         self.nFeaturesToDrop = nFeaturesToDrop
         self.output = SparseConvNetTensor(torch.Tensor())
         self.gradInput = torch.Tensor()
@@ -26,7 +25,7 @@ class AveragePooling(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, 'AveragePooling_updateOutput')(
+        dim_typed_fn(self.dimension, input.features, 'AveragePooling_updateOutput')(
             input.spatial_size,
             self.output.spatial_size,
             self.pool_size,
@@ -40,7 +39,7 @@ class AveragePooling(SparseModule):
 
     def updateGradInput(self, input, gradOutput):
         dim_typed_fn(
-            self.dimension, input, 'AveragePooling_updateGradInput')(
+            self.dimension, input.features, 'AveragePooling_updateGradInput')(
             input.spatial_size,
             self.output.spatial_size,
             self.pool_size,

PyTorch/sparseconvnet/legacy/batchNormalization.py

@@ -21,7 +21,6 @@ from . import SparseModule
 from ..utils import toLongTensor, typed_fn, optionalTensor, nullptr
 from .sparseConvNetTensor import SparseConvNetTensor
 
-
 class BatchNormalization(SparseModule):
     def __init__(
             self,
@@ -31,7 +30,6 @@ class BatchNormalization(SparseModule):
             affine=True,
             leakiness=1):
         SparseModule.__init__(self)
-        assert nPlanes % 4 == 0
         self.nPlanes = nPlanes
         self.eps = eps
         self.momentum = momentum
@@ -44,16 +42,16 @@ class BatchNormalization(SparseModule):
         if affine:
             self.weight = torch.Tensor(nPlanes).fill_(1)
             self.bias = torch.Tensor(nPlanes).fill_(0)
-            self.gradWeight = torch.Tensor(nPlanes)
-            self.gradBias = torch.Tensor(nPlanes)
+            self.gradWeight = torch.Tensor(nPlanes).fill_(0)
+            self.gradBias = torch.Tensor(nPlanes).fill_(0.333)
         self.output = SparseConvNetTensor(torch.Tensor())
         self.gradInput = torch.Tensor()
 
     def updateOutput(self, input):
-        assert input.features.size(1) == self.nPlanes
+        assert input.features.ndimension()==0 or input.features.size(1) == self.nPlanes
         self.output.metadata = input.metadata
         self.output.spatial_size = input.spatial_size
-        typed_fn(input, 'BatchNormalization_updateOutput')(
+        typed_fn(input.features, 'BatchNormalization_updateOutput')(
             input.features,
             self.output.features,
             self.saveMean,
@@ -71,7 +69,7 @@ class BatchNormalization(SparseModule):
     def backward(self, input, gradOutput, scale=1):
         assert scale == 1
         assert self.train
-        typed_fn(input, 'BatchNormalization_backward')(
+        typed_fn(input.features, 'BatchNormalization_backward')(
             input.features,
             self.gradInput,
             self.output.features,
@@ -114,7 +112,14 @@ class BatchNormReLU(BatchNormalization):
         s = 'BatchNormReLU(' + str(self.nPlanes) + ',eps=' + str(self.eps) + \
             ',momentum=' + str(self.momentum) + ',affine=' + str(self.affine) + ')'
         return s
+class BatchNormLeakyReLU(BatchNormalization):
+    def __init__(self, nPlanes, eps=1e-4, momentum=0.9):
+        BatchNormalization.__init__(self, nPlanes, eps, momentum, True, 0.333)
 
+    def __repr__(self):
+        s = 'BatchNormReLU(' + str(self.nPlanes) + ',eps=' + str(self.eps) + \
+            ',momentum=' + str(self.momentum) + ',affine=' + str(self.affine) + ')'
+        return s
 
 class BatchNormalizationInTensor(BatchNormalization):
     def __init__(
@@ -131,7 +136,7 @@ class BatchNormalizationInTensor(BatchNormalization):
             1, self.output_column_offset, self.nPlanes)
         self.output.metadata = input.metadata
         self.output.spatial_size = input.spatial_size
-        typed_fn(input, 'BatchNormalizationInTensor_updateOutput')(
+        typed_fn(input.features, 'BatchNormalizationInTensor_updateOutput')(
             input.features,
             o,
             self.saveMean,
@@ -152,7 +157,7 @@ class BatchNormalizationInTensor(BatchNormalization):
         o = self.output.features.narrow(
             1, self.output_column_offset, self.nPlanes)
         d_o = gradOutput.narrow(1, self.output_column_offset, self.nPlanes)
-        typed_fn(input, 'BatchNormalization_backward')(
+        typed_fn(input.features, 'BatchNormalization_backward')(
             input.features,
             self.gradInput,
             o,

PyTorch/sparseconvnet/legacy/convolution.py

@@ -20,24 +20,29 @@ class Convolution(SparseModule):
         self.filter_size = toLongTensor(dimension, filter_size)
         self.filter_volume = self.filter_size.prod()
         self.filter_stride = toLongTensor(dimension, filter_stride)
+        std = (2.0 / nIn / self.filter_volume)**0.5
         self.weight = torch.Tensor(
-            nIn * self.filter_volume, nOut
-        ).normal_(0, (2.0 / nIn / self.filter_volume)**0.5)
-        self.gradWeight = torch.Tensor(nIn * self.filter_volume, nOut)
+            nIn *
+            self.filter_volume,
+            nOut).normal_(
+            0,
+            std)
+        self.gradWeight = torch.Tensor(
+            nIn * self.filter_volume, nOut).fill_(std)
         if bias:
             self.bias = torch.Tensor(nOut).zero_()
-            self.gradBias = torch.Tensor(nOut)
+            self.gradBias = torch.Tensor(nOut).zero_()
         self.output = SparseConvNetTensor(torch.Tensor())
         self.gradInput = torch.Tensor()
 
     def updateOutput(self, input):
-        assert 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 - self.filter_size) / self.filter_stride + 1
         s.forward_pass_multiplyAdd_count +=\
             dim_typed_fn(
-                self.dimension, input, 'Convolution_updateOutput')(
+                self.dimension, input.features, 'Convolution_updateOutput')(
                 input.spatial_size,
                 self.output.spatial_size,
                 self.filter_size,
@@ -55,7 +60,7 @@ class Convolution(SparseModule):
     def backward(self, input, gradOutput, scale=1):
         assert scale == 1
         dim_typed_fn(
-            self.dimension, input, 'Convolution_backward')(
+            self.dimension, input.features, 'Convolution_backward')(
             input.spatial_size,
             self.output.spatial_size,
             self.filter_size,

PyTorch/sparseconvnet/legacy/deconvolution.py

@@ -22,24 +22,26 @@ class Deconvolution(SparseModule):
         self.filter_size = toLongTensor(dimension, filter_size)
         self.filter_stride = toLongTensor(dimension, filter_stride)
         self.filter_volume = self.filter_size.prod()
+        std = (2.0 / nIn / self.filter_volume)**0.5
         self.weight = torch.Tensor(
             nIn * self.filter_volume, nOut
-        ).normal_(0, (2.0 / nIn / self.filter_volume)**0.5)
-        self.gradWeight = torch.Tensor(nIn * self.filter_volume, nOut)
+        ).normal_(0, std)
+        self.gradWeight = torch.Tensor(
+            nIn * self.filter_volume, nOut).fill_(std)
         if bias:
             self.bias = torch.Tensor(nOut).zero_()
-            self.gradBias = torch.Tensor(nOut)
+            self.gradBias = torch.Tensor(nOut).zero_()
         self.output = SparseConvNetTensor(torch.Tensor())
         self.gradInput = torch.Tensor()
 
     def updateOutput(self, input):
-        assert 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 - 1) * self.filter_stride + self.filter_size
         s.forward_pass_multiplyAdd_count +=\
             dim_typed_fn(
-                self.dimension, input, 'Deconvolution_updateOutput')(
+                self.dimension, input.features, 'Deconvolution_updateOutput')(
                 input.spatial_size,
                 self.output.spatial_size,
                 self.filter_size,
@@ -57,7 +59,7 @@ class Deconvolution(SparseModule):
     def backward(self, input, gradOutput, scale=1):
         assert scale == 1
         dim_typed_fn(
-            self.dimension, input, 'Deconvolution_backward')(
+            self.dimension, input.features, 'Deconvolution_backward')(
             input.spatial_size,
             self.output.spatial_size,
             self.filter_size,

PyTorch/sparseconvnet/legacy/denseToSparse.py

+# Copyright 2016-present, Facebook, Inc.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+Function to convert a Dense Input into a sparse input.
+If possible, avoid using this module; build the hidden layer using InputBatch.
+
+Parameters:
+dimension : of the input field
+"""
+
+import torch
+from . import SparseModule
+from ..utils import dim_fn, nullptr
+from .sparseConvNetTensor import SparseConvNetTensor
+from .metadata import Metadata
+
+class DenseToSparse(SparseModule):
+    def __init__(self, dimension):
+        SparseModule.__init__(self)
+        self.dimension = dimension
+        self.output = SparseConvNetTensor(torch.Tensor(),Metadata(dimension))
+        self.gradInput = torch.Tensor()
+
+    def updateOutput(self, input):
+        a=input
+        aa=a.permute(*([0,]+list(range(2,2+self.dimension))+[1,])).clone()
+        self.aas=aa.size()
+        nz=aa.abs().sum(self.dimension+1).view(aa.size()[0:-1])
+        s=torch.LongTensor(nz.stride()).view(1,self.dimension+1)
+        nz=nz.nonzero()
+        s=s.type_as(nz)
+        aa=aa.view(-1,a.size(1))
+        self.aas2=aa.size()
+        self.r=(nz*s.expand_as(nz)).sum(1).view(-1)
+        self.output.features=aa.index_select(0,self.r)
+        self.output.spatial_size=torch.LongTensor(list(input.size()[2:]))
+        dim_fn(self.dimension, 'createMetadataForDenseToSparse')(
+            self.output.metadata.ffi,
+            self.output.spatial_size,
+            nz.cpu(),
+            input.size(0))
+        return self.output
+
+    def updateGradInput(self, input, gradOutput):
+        self.gradInput.resize_(self.aas2).zero_()
+        self.gradInput.index_copy_(0,self.r,gradOutput)
+        self.gradInput=self.gradInput.view(self.aas).permute(*([0,self.dimension+1]+list(range(1,self.dimension+1))))
+        return self.gradInput
+
+    def clearState(self):
+        SparseModule.clearState(self)
+        self.aas=None
+        self.r=None
+    def __repr__(self):
+        return 'DenseToSparse(' + str(self.dimension) + ')'

PyTorch/sparseconvnet/legacy/leakyReLU.py

@@ -22,7 +22,7 @@ class LeakyReLU(SparseModule):
     def updateOutput(self, input):
         self.output.metadata = input.metadata
         self.output.spatial_size = input.spatial_size
-        typed_fn(input, 'LeakyReLU_updateOutput')(
+        typed_fn(input.features, 'LeakyReLU_updateOutput')(
             input.features,
             self.output.features,
             self.leakage)

PyTorch/sparseconvnet/legacy/maxPooling.py

@@ -26,7 +26,7 @@ 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, 'MaxPooling_updateOutput')(
+        dim_typed_fn(self.dimension, input.features, 'MaxPooling_updateOutput')(
             input.spatial_size,
             self.output.spatial_size,
             self.pool_size,
@@ -39,7 +39,7 @@ class MaxPooling(SparseModule):
         return self.output
 
     def updateGradInput(self, input, gradOutput):
-        dim_typed_fn(self.dimension, input, 'MaxPooling_updateGradInput')(
+        dim_typed_fn(self.dimension, input.features, 'MaxPooling_updateGradInput')(
             input.spatial_size,
             self.output.spatial_size,
             self.pool_size,

PyTorch/sparseconvnet/legacy/metadata.py

@@ -24,16 +24,16 @@ ffi = cffi.FFI()
 
 class Metadata(object):
     def __init__(self, dimension, ptr=0):
-        #print('make meta',dimension, ptr)
         self.dimension = dimension
         self.ffi = ffi.new('void *[1]')
         scn_writePtr(ptr, self.ffi)
         self.ffigc = ffi.gc(self.ffi, dim_fn(self.dimension, 'freeMetadata'))
 
     def set_(self):
-        if hasattr(self, 'ffi'):
-            del self.ffigc
-            del self.ffi
+        dim_fn(self.dimension, 'freeMetadata')(self.ffi)
+        # if hasattr(self, 'ffi'):
+        #     del self.ffigc
+        #     del self.ffi
 
     def __reduce__(self):
         if hasattr(self, 'ffi'):

PyTorch/sparseconvnet/legacy/misc.py

+# Copyright 2016-present, Facebook, Inc.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch.legacy.nn as nn
+from .sequential import Sequential
+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.output = SparseConvNetTensor()
+        self.output.features=self.module.output
+        self.gradInput = self.module.gradInput
+
+    def updateOutput(self, input):
+        self.output.metadata = input.metadata
+        self.output.spatial_size = input.spatial_size
+        self.module.forward(input.features)
+        return self.output
+
+    def updateGradInput(self, input, 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.gradInput = self.module.gradInput
+
+class ELU(SparseModule):
+    def __init__(self):
+        SparseModule.__init__(self)
+        self.module=nn.ELU()
+        self.output = SparseConvNetTensor()
+        self.gradInput = self.module.gradInput
+
+    def updateOutput(self, input):
+        self.output.metadata = input.metadata
+        self.output.spatial_size = input.spatial_size
+        self.module.forward(input.features)
+        return self.output
+
+    def updateGradInput(self, input, 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.gradInput = self.module.gradInput
+
+def BatchNormELU(nPlanes, eps=1e-4, momentum=0.9):
+    return Sequential().add(BatchNormalization(nPlanes,eps,momentum)).add(ELU())

PyTorch/sparseconvnet/legacy/networkArchitectures.py

@@ -116,7 +116,7 @@ def SparseVggNet(dimension, nInputPlanes, layers):
                      .add(ValidConvolution(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]}))
+                     )).add(JoinTable([x[1], x[2], x[3]]))
             nPlanes = x[1] + x[2] + x[3]
             m.add(BatchNormReLU(nPlanes))
     return m

PyTorch/sparseconvnet/legacy/networkInNetwork.py

@@ -16,11 +16,12 @@ class NetworkInNetwork(SparseModule):
         SparseModule.__init__(self)
         self.nIn = nIn
         self.nOut = nOut
-        self.weight = torch.Tensor(nIn, nOut).normal_(0, (2.0 / self.nIn)**0.5)
-        self.gradWeight = torch.Tensor(nIn, nOut)
+        std = (2.0 / self.nIn)**0.5
+        self.weight = torch.Tensor(nIn, nOut).normal_(0, std)
+        self.gradWeight = torch.Tensor(nIn, nOut).fill_(std)
         if bias:
-            self.bias = torch.Tensor(nOut).fill_(0)
-            self.gradBias = torch.Tensor(nOut)
+            self.bias = torch.Tensor(nOut).zero_()
+            self.gradBias = torch.Tensor(nOut).zero_()
         self.output = SparseConvNetTensor(torch.Tensor())
         self.gradInput = torch.Tensor()
 
@@ -28,7 +29,7 @@ class NetworkInNetwork(SparseModule):
         self.output.metadata = input.metadata
         self.output.spatial_size = input.spatial_size
         s.forward_pass_multiplyAdd_count +=\
-            typed_fn(input, 'NetworkInNetwork_updateOutput')(
+            typed_fn(input.features, 'NetworkInNetwork_updateOutput')(
                 input.features,
                 self.output.features,
                 self.weight,
@@ -37,7 +38,7 @@ class NetworkInNetwork(SparseModule):
         return self.output
 
     def updateGradInput(self, input, gradOutput):
-        typed_fn(input, 'NetworkInNetwork_updateGradInput')(
+        typed_fn(input.features, 'NetworkInNetwork_updateGradInput')(
             self.gradInput,
             gradOutput,
             self.weight)
@@ -45,7 +46,7 @@ class NetworkInNetwork(SparseModule):
 
     def accGradParameters(self, input, gradOutput, scale=1):
         assert scale == 1
-        typed_fn(input, 'NetworkInNetwork_accGradParameters')(
+        typed_fn(input.features, 'NetworkInNetwork_accGradParameters')(
             input.features,
             gradOutput,
             self.gradWeight,

PyTorch/sparseconvnet/legacy/sparseConvNetTensor.py

@@ -29,7 +29,7 @@ class SparseConvNetTensor(object):
 
     def set_(self):
         self.features.set_(self.features.storage_type()())
-        self.metadata = None
+        self.metadata.set_()
         self.spatialSize = None
 
     def __repr__(self):

PyTorch/sparseconvnet/legacy/sparseToDense.py

@@ -22,23 +22,27 @@ from .sparseConvNetTensor import SparseConvNetTensor
 
 
 class SparseToDense(SparseModule):
-    def __init__(self, dimension):
+    def __init__(self, dimension, nPlanes=None):
         SparseModule.__init__(self)
         self.dimension = dimension
         self.output = torch.Tensor()
         self.gradInput = torch.FloatTensor()
+        self.nPlanes=nPlanes
 
     def updateOutput(self, input):
-        dim_typed_fn(self.dimension, input, 'SparseToDense_updateOutput')(
+        if not self.nPlanes:
+            self.nPlanes=input.features.size(1)
+        dim_typed_fn(self.dimension, input.features, 'SparseToDense_updateOutput')(
             input.spatial_size,
             input.metadata.ffi,
             input.features,
             self.output,
-            torch.cuda.IntTensor() if input.features.is_cuda else nullptr)
+            torch.cuda.IntTensor() if input.features.is_cuda else nullptr,
+            self.nPlanes)
         return self.output
 
     def updateGradInput(self, input, gradOutput):
-        dim_typed_fn(self.dimension, input, 'SparseToDense_updateGradInput')(
+        dim_typed_fn(self.dimension, input.features, 'SparseToDense_updateGradInput')(
             input.spatial_size,
             input.metadata.ffi,
             input.features,

PyTorch/sparseconvnet/legacy/validConvolution.py

@@ -19,22 +19,24 @@ class ValidConvolution(SparseModule):
         self.nOut = nOut
         self.filter_size = toLongTensor(dimension, filter_size)
         self.filter_volume = self.filter_size.prod()
+        std = (2.0 / nIn / self.filter_volume)**0.5
         self.weight = torch.Tensor(
             nIn * self.filter_volume, nOut
-        ).normal_(0, (2.0 / nIn / self.filter_volume)**0.5)
-        self.gradWeight = torch.Tensor(nIn * self.filter_volume, nOut)
+        ).normal_(0, std)
+        self.gradWeight = torch.Tensor(
+            nIn * self.filter_volume, nOut).fill_(std)
         if bias:
             self.bias = torch.Tensor(nOut).zero_()
-            self.gradBias = torch.Tensor(nOut)
+            self.gradBias = torch.Tensor(nOut).zero_()
         self.output = SparseConvNetTensor(torch.Tensor())
         self.gradInput = torch.Tensor()
 
     def updateOutput(self, input):
-        assert 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, 'ValidConvolution_updateOutput')(
+            dim_typed_fn(self.dimension, input.features, 'ValidConvolution_updateOutput')(
                 input.spatial_size,
                 self.filter_size,
                 input.metadata.ffi,
@@ -49,7 +51,7 @@ class ValidConvolution(SparseModule):
 
     def backward(self, input, gradOutput, scale=1):
         assert scale == 1
-        dim_typed_fn(self.dimension, input, 'ValidConvolution_backward')(
+        dim_typed_fn(self.dimension, input.features, 'ValidConvolution_backward')(
             input.spatial_size,
             self.filter_size,
             input.metadata.ffi,

PyTorch/sparseconvnet/utils.py

@@ -30,18 +30,16 @@ def dim_fn(dimension, name):
 
 
 def typed_fn(t, name):
-    # print('typed_fn',dimension,name)
-    return getattr(scn, 'scn_' + typeTable[t.features.type()] + '_' + name)
-
+    # print('typed_fn',t.features.type(),name)
+    return getattr(scn, 'scn_' + typeTable[t.type()] + '_' + name)
 
 def dim_typed_fn(dimension, t, name):
-    # print('dim_typed_fn',dimension,t,name)
+    # print('dim_typed_fn',dimension,t.features.type(),name)
     return getattr(scn, 'scn_' +
-                   typeTable[t.features.type()] +
+                   typeTable[t.type()] +
                    str(dimension) +
                    name)
 
-
 ffi = FFI()
 nullptr = ffi.NULL
 

Torch/C.lua

@@ -52,12 +52,16 @@ void scn_DIMENSION_setInputSpatialLocations(void **m, THFloatTensor *features,
   THLongTensor *locations, THFloatTensor *vecs, bool overwrite);
 void scn_DIMENSION_getSpatialLocations(void **m, THLongTensor *spatialSize,
   THLongTensor *locations);
-  ]]
+
+]]
 
   for DIMENSION = 1,10 do
     local def = string.gsub(cdef, 'DIMENSION', DIMENSION)
-    if fc then fc:write(def) end
     ffi.cdef(def)
+    if fc then
+      def=string.gsub(def,'bool','_Bool')
+      fc:write(def)
+    end
   end
 
   --types CPU float, double;
@@ -128,21 +132,29 @@ void scn_ARCH_REAL_NetworkInNetwork_accGradParameters(
     def = string.gsub(def, 'THITensor', 'void')
     def = string.gsub(def, 'REAL', v[1])
     def = string.gsub(def, 'THTensor', v[2])
-    if fc then fc:write(def) end
     ffi.cdef(def)
+    if fc then
+      def=string.gsub(def,'bool','_Bool')
+      fc:write(def)
+    end
   end
   if sparseconvnet.cutorch then
     for k,v in ipairs({
         {'float', 'THCudaTensor'},
-        {'double', 'THCudaDoubleTensor'}}) do
+        --{'double', 'THCudaDoubleTensor'}
+                     })
+    do
       local def = cdef
       def = string.gsub(def, 'ARCH', 'gpu')
       def = string.gsub(def, 'THITensor', sparseconvnet.ruleBookBits==64 and
                           'THCudaLongTensor' or 'THCudaIntTensor')
       def = string.gsub(def, 'REAL', v[1])
       def = string.gsub(def, 'THTensor', v[2])
-      if fg then fg:write(def) end
       ffi.cdef(def)
+      if fg then
+        def=string.gsub(def,'bool','_Bool')
+        fg:write(def)
+      end
     end
   end
 
@@ -226,7 +238,7 @@ void scn_ARCH_REAL_DIMENSIONValidConvolution_backward(
   THTensor *d_bias, long filterVolume, THITensor *rulesBuffer);
   ]]
 
-  for _,v in ipairs({{'float', 'THFloatTensor'}, {'double','THDoubleTensor'}}) do
+  for _,v in ipairs({{'float', 'THFloatTensor'}, {'double', 'THDoubleTensor'}}) do
     for DIMENSION = 1,10 do
       local def = cdef
       def = string.gsub(def, 'ARCH', 'cpu')
@@ -234,14 +246,18 @@ void scn_ARCH_REAL_DIMENSIONValidConvolution_backward(
       def = string.gsub(def, 'THITensor', 'void')
       def = string.gsub(def, 'REAL', v[1])
       def = string.gsub(def, 'THTensor', v[2])
-      if fc then fc:write(def) end
       ffi.cdef(def)
+      if fc then
+        def=string.gsub(def,'bool','_Bool')
+        fc:write(def)
+      end
     end
   end
   if sparseconvnet.cutorch then
     for k,v in ipairs({
         {'float', 'THCudaTensor'},
-        {'double', 'THCudaDoubleTensor'}}) do
+        --{'double', 'THCudaDoubleTensor'}
+    }) do
       for DIMENSION = 1,10 do
         local def = cdef
         def = string.gsub(def, 'ARCH', 'gpu')
@@ -250,8 +266,11 @@ void scn_ARCH_REAL_DIMENSIONValidConvolution_backward(
                             'THCudaLongTensor' or 'THCudaIntTensor')
         def = string.gsub(def, 'REAL', v[1])
         def = string.gsub(def, 'THTensor', v[2])
-        if fg then fg:write(def) end
         ffi.cdef(def)
+        if fg then
+          def=string.gsub(def,'bool','_Bool')
+          fg:write(def)
+        end
       end
     end
   end

Torch/SparseToDense.lua

@@ -5,14 +5,14 @@
 -- LICENSE file in the root directory of this source tree.
 
 --[[
-Function to convert a SparseConvNet hidden layer to a dense convolutional
-layer. Put a SparseToDense convolutional layer (or an ActivePooling layer) at
-the top of your sparse network. The output can then pass to a dense
-convolutional layers or (if the spatial dimensions have become trivial) a
-linear classifier.
+  Function to convert a SparseConvNet hidden layer to a dense convolutional
+  layer. Put a SparseToDense convolutional layer (or an ActivePooling layer) at
+  the top of your sparse network. The output can then pass to a dense
+  convolutional layers or (if the spatial dimensions have become trivial) a
+  linear classifier.
 
-Parameters:
-dimension : of the input field
+  Parameters:
+  dimension : of the input field
 ]]
 
 return function(sparseconvnet)
@@ -20,11 +20,12 @@ return function(sparseconvnet)
   local SparseToDense, parent = torch.class(
     'sparseconvnet.SparseToDense', 'nn.Module', sparseconvnet)
 
-  function SparseToDense:__init(dimension)
+  function SparseToDense:__init(dimension, nPlanes)
     parent.__init(self)
     self.dimension = dimension
     self.output=torch.Tensor()
     self.gradInput={features=torch.Tensor()}
+    self.nPlanes=nPlanes
   end
 
   function SparseToDense:updateOutput(input)
@@ -34,7 +35,8 @@ return function(sparseconvnet)
       input.metadata.ffi,
       input.features:cdata(),
       self.output:cdata(),
-      self.shared.rulesBuffer and self.shared.rulesBuffer:cdata())
+      self.shared.rulesBuffer and self.shared.rulesBuffer:cdata(),
+      self.nPlanes or input.features.size(2))
     return self.output
   end