def createResidualLayer(self, customLayer=None):
core1 = nn.Sequential()
if customLayer:
core2 = customLayer
else:
core2 = nn.Identity()
core = nn.ConcatTable()
core.add(core1)
core.add(core2)
layer = nn.Sequential()
layer.add(core)
layer.add(nn.CAddTable())
layer.add(nn.ReLU())
return layer, core1, core2
def init_modules(self):
n = nn.Sequential()
n = nn.SpatialConvolution(1, 1, 1,1, 1,1, 1,1)
n = nn.SpatialBatchNormalization(1)
n = nn.ReLU()
n = nn.Sigmoid()
n = nn.SpatialMaxPooling(1,1, 1,1, 1,1)
n = nn.SpatialAveragePooling(1,1, 1,1, 1,1)
n = nn.Linear(1, 1)
n = nn.Dropout(0.1)
n = nn.SoftMax()
n = nn.Identity()
n = nn.Reshape(1)
n = nn.BCECriterion()
n = nn.MSECriterion()
tnet = {}
net = caffe.Net(
'test_c.prototxt',
'/work/cv3/sankaran/caffe-exp/faster-rcnn-exp/VGG16_e2e_pascal/models/ms_tv_vgg16_faster_rcnn_iter_100000.caffemodel',
caffe.TEST)
print net.params['fc6'][0].data.shape
print net.params['fc6'][1].data.shape
tnet["conv1_1"] = nn.SpatialConvolutionMM(3, 64, 3, 3, 1, 1, 1, 1)
transferW(tnet["conv1_1"].weight,
net.params['conv1_1'][0].data.reshape(64, 27))
transferB(tnet["conv1_1"].bias, net.params['conv1_1'][1].data)
tnet["relu1_1"] = nn.ReLU()
tnet["conv1_2"] = nn.SpatialConvolutionMM(64, 64, 3, 3, 1, 1, 1, 1)
transferW(tnet["conv1_2"].weight,
net.params['conv1_2'][0].data.reshape(64, 576))
transferB(tnet["conv1_2"].bias, net.params['conv1_2'][1].data)
tnet["relu1_2"] = nn.ReLU()
tnet["pool1"] = nn.SpatialMaxPooling(2, 2, 2, 2, 0, 0)
tnet["conv2_1"] = nn.SpatialConvolutionMM(64, 128, 3, 3, 1, 1, 1, 1)
transferW(tnet["conv2_1"].weight,
net.params['conv2_1'][0].data.reshape(128, 576))
transferB(tnet["conv2_1"].bias, net.params['conv2_1'][1].data)
tnet["relu2_1"] = nn.ReLU()
tnet["conv2_2"] = nn.SpatialConvolutionMM(128, 128, 3, 3, 1, 1, 1, 1)
myeval('linear.output')
# print('linearCl.output', linear.output)
output = linear.forward(a)
print('output.dims()', output.dims())
print('output.size()', output.size())
outputFloat = output.float()
print('outputFloat', outputFloat)
print('output', output)
mlp = nn.Sequential()
mlp.add(nn.SpatialConvolutionMM(1, 16, 5, 5, 1, 1, 2, 2))
mlp.add(nn.ReLU())
mlp.add(nn.SpatialMaxPooling(3, 3, 3, 3))
mlp.add(nn.SpatialConvolutionMM(16, 32, 5, 5, 1, 1, 2, 2))
mlp.add(nn.ReLU())
mlp.add(nn.SpatialMaxPooling(2, 2, 2, 2))
mlp.add(nn.Reshape(32 * 4 * 4))
mlp.add(nn.Linear(32 * 4 * 4, 150))
mlp.add(nn.Tanh())
mlp.add(nn.Linear(150, 10))
mlp.add(nn.LogSoftMax())
mlp.cl()
print('mlp', mlp)
myeval('mlp.output')
input = PyTorch.FloatTensor(128, 1, 28, 28).uniform().cl()
def test_cltorch():
if 'ALLOW_NON_GPUS' in os.environ:
PyClTorch.setAllowNonGpus(True)
# a = PyTorch.foo(3,2)
# print('a', a)
# print(PyTorch.FloatTensor(3,2))
a = PyClTorch.ClTensor([3, 4, 9])
assert a[0] == 3
assert a[1] == 4
assert a[2] == 9
print('a', a)
a = PyClTorch.ClTensor([[3, 5, 7], [9, 2, 4]])
print('a', a)
print('a[0]', a[0])
print('a[0][0]', a[0][0])
assert a[0][0] == 3
assert a[1][0] == 9
assert a[1][2] == 4
PyTorch.manualSeed(123)
a = PyTorch.FloatTensor(4, 3).uniform()
print('a', a)
a_cl = a.cl()
print(type(a_cl))
assert str(type(a_cl)) == '<class \'PyClTorch.ClTensor\'>'
print('a_cl[0]', a_cl[0])
print('a_cl[0][0]', a_cl[0][0])
assert a[0][0] == a_cl[0][0]
assert a[0][1] == a_cl[0][1]
assert a[1][1] == a_cl[1][1]
print('a.dims()', a.dims())
print('a.size()', a.size())
print('a', a)
assert a.dims() == 2
assert a.size()[0] == 4
assert a.size()[1] == 3
a_sum = a.sum()
a_cl_sum = a_cl.sum()
assert abs(a_sum - a_cl_sum) < 1e-4
a_cl2 = a_cl + 3.2
assert abs(a_cl2[1][0] - a[1][0] - 3.2) < 1e-4
b = PyClTorch.ClTensor()
print('got b')
myeval('b')
assert b.dims() == -1
b.resizeAs(a)
myeval('b')
assert b.dims() == 2
assert b.size()[0] == 4
assert b.size()[1] == 3
print('run uniform')
b.uniform()
myeval('b')
print('create new b')
b = PyClTorch.ClTensor()
print('b.dims()', b.dims())
print('b.size()', b.size())
print('b', b)
c = PyTorch.FloatTensor().cl()
print('c.dims()', c.dims())
print('c.size()', c.size())
print('c', c)
assert b.dims() == -1
assert b.size() is None
print('creating Linear...')
linear = nn.Linear(3, 5).float()
print('created linear')
print('linear:', linear)
myeval('linear.output')
myeval('linear.output.dims()')
myeval('linear.output.size()')
myeval('linear.output.nElement()')
linear_cl = linear.clone().cl()
print('type(linear.output)', type(linear.output))
print('type(linear_cl.output)', type(linear_cl.output))
assert str(type(linear.output)) == '<class \'PyTorch._FloatTensor\'>'
assert str(type(linear_cl.output)) == '<class \'PyClTorch.ClTensor\'>'
# myeval('type(linear)')
# myeval('type(linear.output)')
myeval('linear_cl.output.dims()')
myeval('linear_cl.output.size()')
# myeval('linear.output')
assert str(type(linear)) == '<class \'PyTorchAug.Linear\'>'
assert str(type(linear_cl)) == '<class \'PyTorchAug.Linear\'>'
# assert str(type(linear.output)) == '<class \'PyClTorch.ClTensor\'>'
# assert linear.output.dims() == -1 # why is this 0? should be -1???
# assert linear.output.size() is None # again, should be None?
a_cl = PyClTorch.ClTensor(4, 3).uniform()
# print('a_cl', a_cl)
output_cl = linear_cl.forward(a_cl)
# print('output', output)
assert str(type(output_cl)) == '<class \'PyClTorch.ClTensor\'>'
assert output_cl.dims() == 2
assert output_cl.size()[0] == 4
assert output_cl.size()[1] == 5
a = a_cl.float()
output = linear.forward(a)
assert str(type(output)) == '<class \'PyTorch._FloatTensor\'>'
assert output.dims() == 2
assert output.size()[0] == 4
assert output.size()[1] == 5
print('a.size()', a.size())
print('a_cl.size()', a_cl.size())
assert a[1][0] == a_cl[1][0]
assert a[2][1] == a_cl[2][1]
mlp = nn.Sequential()
mlp.add(nn.SpatialConvolutionMM(1, 16, 5, 5, 1, 1, 2, 2))
mlp.add(nn.ReLU())
mlp.add(nn.SpatialMaxPooling(3, 3, 3, 3))
mlp.add(nn.SpatialConvolutionMM(16, 32, 5, 5, 1, 1, 2, 2))
mlp.add(nn.ReLU())
mlp.add(nn.SpatialMaxPooling(2, 2, 2, 2))
mlp.add(nn.Reshape(32 * 4 * 4))
mlp.add(nn.Linear(32 * 4 * 4, 150))
mlp.add(nn.Tanh())
mlp.add(nn.Linear(150, 10))
mlp.add(nn.LogSoftMax())
mlp.float()
mlp_cl = mlp.clone().cl()
print('mlp_cl', mlp_cl)
# myeval('mlp.output')
input = PyTorch.FloatTensor(128, 1, 28, 28).uniform()
input_cl = PyClTorch.FloatTensorToClTensor(
input.clone()) # This is a bit hacky...
output = mlp.forward(input)
# myeval('input[0]')
output_cl = mlp_cl.forward(input_cl)
# myeval('output[0]')
assert (output_cl.float() - output).abs().max() < 1e-4
def test_pynn():
PyTorch.manualSeed(123)
linear = nn.Linear(3, 5)
linear
print('linear', linear)
print('linear.weight', linear.weight)
print('linear.output', linear.output)
print('linear.gradInput', linear.gradInput)
input = PyTorch.DoubleTensor(4, 3).uniform()
print('input', input)
output = linear.updateOutput(input)
print('output', output)
gradInput = linear.updateGradInput(input, output)
print('gradInput', gradInput)
criterion = nn.ClassNLLCriterion()
print('criterion', criterion)
print('dir(linear)', dir(linear))
mlp = nn.Sequential()
mlp.add(linear)
output = mlp.forward(input)
print('output', output)
# import sys
# sys.path.append('thirdparty/python-mnist')
from mnist import MNIST
import numpy
import array
numpy.random.seed(123)
mlp = nn.Sequential()
mlp.add(nn.SpatialConvolutionMM(1, 16, 5, 5, 1, 1, 2, 2))
res = mlp.add(nn.ReLU())
print('res', res)
mlp.add(nn.SpatialMaxPooling(3, 3, 3, 3))
mlp.add(nn.SpatialConvolutionMM(16, 32, 3, 3, 1, 1, 1, 1))
mlp.add(nn.ReLU())
mlp.add(nn.SpatialMaxPooling(2, 2, 2, 2))
mlp.add(nn.Reshape(32 * 4 * 4))
mlp.add(nn.Linear(32 * 4 * 4, 150))
mlp.add(nn.Tanh())
mlp.add(nn.Linear(150, 10))
mlp.add(nn.LogSoftMax())
criterion = nn.ClassNLLCriterion()
print('got criterion')
learningRate = 0.02
mndata = MNIST('data/mnist')
imagesList, labelsB = mndata.load_training()
images = numpy.array(imagesList).astype(numpy.float64)
# print('imagesArray', images.shape)
# print(images[0].shape)
labelsf = array.array('d', labelsB.tolist())
imagesTensor = PyTorch.asDoubleTensor(images)
# imagesTensor = PyTorch.FloatTensor(100,784)
# labels = numpy.array(20,).astype(numpy.int32)
# labelsTensor = PyTorch.FloatTensor(100).fill(1)
# print('labels', labels)
# print(imagesTensor.size())
def printStorageAddr(name, tensor):
print('printStorageAddr START')
storage = tensor.storage()
if storage is None:
print(name, 'storage is None')
else:
print(name, 'storage is ', hex(storage.dataAddr()))
print('printStorageAddr END')
labelsTensor = PyTorch.asDoubleTensor(labelsf)
labelsTensor += 1
# print('calling size on imagestensor...')
# print(' (called size)')
desiredN = 128
maxN = int(imagesTensor.size()[0])
desiredN = min(maxN, desiredN)
imagesTensor = imagesTensor.narrow(0, 0, desiredN)
labelsTensor = labelsTensor.narrow(0, 0, desiredN)
print('imagesTensor.size()', imagesTensor.size())
print('labelsTensor.size()', labelsTensor.size())
N = int(imagesTensor.size()[0])
print('type(imagesTensor)', type(imagesTensor))
size = PyTorch.LongStorage(4)
size[0] = N
size[1] = 1
size[2] = 28
size[3] = 28
imagesTensor.resize(size)
imagesTensor /= 255.0
imagesTensor -= 0.2
print('imagesTensor.size()', imagesTensor.size())
print('start training...')
for epoch in range(4):
numRight = 0
for n in range(N):
# print('n', n)
input = imagesTensor[n]
label = labelsTensor[n]
labelTensor = PyTorch.DoubleTensor(1)
labelTensor[0] = label
# print('label', label)
output = mlp.forward(input)
prediction = PyTorch.getDoublePrediction(output)
# print('prediction', prediction)
if prediction == label:
numRight += 1
criterion.forward(output, labelTensor)
mlp.zeroGradParameters()
gradOutput = criterion.backward(output, labelTensor)
mlp.backward(input, gradOutput)
mlp.updateParameters(learningRate)
# PyTorch.collectgarbage()
# if n % 100 == 0:
# print('n=', n)
print('epoch ' + str(epoch) + ' accuracy: ' + str(numRight * 100.0 / N) + '%')
print "Transfering RPN layers"
proto = 'frcnn_vgg1024.prototxt'
tnet = {}
#net = caffe.Net('test.prototxt','/work/cv3/sankaran/faster-rcnn/data/imagenet_models/VGG_CNN_M_1024.v2.caffemodel',caffe.TEST)
#net = caffe.Net('test_c.prototxt','vgg_cnn_m_1024_faster_rcnn.caffemodel',caffe.TEST)
net = caffe.Net(proto,args.model,caffe.TEST)
print net.params['fc6'][0].data.shape
print net.params['fc6'][1].data.shape
tnet["conv1"] = nn.SpatialConvolutionMM(3,96,7,7,2,2,0,0)
print "conv1"
transferW(tnet["conv1"].weight, net.params['conv1'][0].data.reshape(96,147))
transferB(tnet["conv1"].bias, net.params['conv1'][1].data)
tnet["relu1"] = nn.ReLU()
tnet["norm1"] = nn.SpatialCrossMapLRN(5,0.0005,0.75,2)
tnet["pool1"] = nn.SpatialMaxPooling(3,3,2,2,0,0)
tnet["conv2"] = nn.SpatialConvolutionMM(96,256,5,5,2,2,1,1)
print "conv2"
transferW(tnet["conv2"].weight, net.params['conv2'][0].data.reshape(256,2400))
transferB(tnet["conv2"].bias, net.params['conv2'][1].data)
tnet["relu2"] = nn.ReLU()
tnet["norm2"] = nn.SpatialCrossMapLRN(5,0.0005,0.75,2)
tnet["pool2"] = nn.SpatialMaxPooling(3,3,2,2,0,0)
tnet["conv3"]= nn.SpatialConvolutionMM(256,512,3,3,1,1,1,1)
print "conv3"
transferW(tnet["conv3"].weight, net.params['conv3'][0].data.reshape(512,2304))
transferB(tnet["conv3"].bias, net.params['conv3'][1].data)
def exportModel(self, graph):
order, expanded_order = graph.topologicalSort()
net = nn.Sequential()
for id in order:
node = graph.nodes.get(id)
name = node.type
m = node.params
l = node.learned_params
if m.get('in_channels') :
nInputPlane = m.get('in_channels')
else:
nInputPlane = 1
if m.get('out_channels') :
nOutputPlane = m.get('out_channels')
else:
nOutputPlane = 1
if m.get('num_features') :
nFeatures = m.get('num_features')
else:
nFeatures = 1
if m.get('in_features') :
inputDimension = m.get('in_features')
else:
inputDimension = 1
if m.get('out_features') :
outputDimension = m.get('out_features')
else:
outputDimension = 1
if m.get('p') :
p = m.get('p')
else:
p = 0.1
if m.get('kernel_size') :
kernel_size = m.get('kernel_size')
if type(kernel_size) == type((3,3)):
kW = kernel_size[0]
kH = kernel_size[1]
else:
kW = kernel_size
kH = kernel_size
else:
kW = 1
kH = 1
if m.get('stride') :
stride = m.get('stride')
if type(stride) == type((3,3)):
dW = stride[0]
dH = stride[1]
else:
dW = stride
dH = stride
else:
dW = 1
dH = 1
if m.get('padding') :
padding = m.get('padding')
if type(padding) == type((3,3)):
padW = padding[0]
padH = padding[1]
else:
padW = padding
padH = padding
else:
padW = 0
padH = 0
if m.get('subgraph'):
subgraph = m.get('subgraph')
# copy the network architecture
if name == 'Conv2d':
n = nn.SpatialConvolution(nInputPlane, nOutputPlane, kW, kH, dW, dH, padW, padH)
elif name == 'BatchNorm2d':
n = nn.SpatialBatchNormalization(nFeatures)
elif name == 'ReLU':
n = nn.ReLU()
elif name == 'Sigmoid':
n = nn.Sigmoid()
elif name == 'MaxPool2d':
n = nn.SpatialMaxPooling(kW, kH, dW, dH, padW, padH)
elif name == 'AvgPool2d':
n = nn.SpatialAveragePooling(kW, kH, dW, dH, padW, padH)
elif name == 'Linear':
n = nn.Linear(inputDimension, outputDimension)
elif name == 'Dropout':
n = nn.Dropout(p)
elif name == 'Softmax':
n = nn.SoftMax()
elif name == 'Identity':
n = nn.Identity()
elif name == 'Reshape':
# add params here
n = nn.Reshape()
elif name == 'BCELoss':
n = nn.BCECriterion()
elif name == 'MSELoss':
n = nn.MSECriterion()
elif name == 'Sequential':
n = self.exportModel(subgraph)
elif name == 'ResNet':
n, core, _ = self.createResidualLayer()
core.add(nn.SpatialConvolution(128,128, 3,3, 2,2, 1,1))
core.add(nn.SpatialBatchNormalization(128))
core.add(nn.ReLU())
core.add(nn.SpatialConvolution(128,128, 3,3, 1,1, 1,1))
core.add(nn.SpatialBatchNormalization(128))
else:
# Group or Sequential nodes
if node.group == True:
n = self.exportModel(subgraph)
else:
print('Not Implement', name)
# copy the network weights
weight = l.get('weight')
bias = l.get('bias')
running_mean = l.get('running_mean')
running_var = l.get('running_var')
# copy over the learned params if they exist
self.copyWeights(n.weight, weight)
self.copyWeights(n.bias, bias)
self.copyWeights(n.running_mean, running_mean)
self.copyWeights(n.running_var, running_var)
net.add(n)
return net