# -*- coding: utf-8 -*-
# ELEKTRONN - Neural Network Toolkit
#
# Copyright (c) 2014 - now
# Max-Planck-Institute for Medical Research, Heidelberg, Germany
# Authors: Marius Killinger, Gregor Urban
import numpy as np
import theano
from theano import tensor as T
[docs]def maxabs(t1, t2):
pos = T.where(t1 > t2, t1, t2)
neg = T.where(-t1 > -t2, t1, t2)
ret = T.where(pos >= -neg, pos, neg)
return ret
[docs]def my_max_pool_3d(sym_input, pool_shape=(2, 2, 2)):
""" this one is pure theano. Hence all gradient-related stuff is working! No dimshuffling"""
s = None
if pool_shape[2] > 1:
for i in xrange(pool_shape[2]):
t = sym_input[:, :, :, :, i::pool_shape[2]]
if s is None:
s = t
else:
s = T.maximum(s, t)
else:
s = sym_input
if pool_shape[0] > 1:
temp = s
s = None
for i in xrange(pool_shape[0]):
t = temp[:, i::pool_shape[0], :, :, :]
if s is None:
s = t
else:
s = T.maximum(s, t)
if pool_shape[1] > 1:
temp = s
s = None
for i in xrange(pool_shape[1]):
t = temp[:, :, :, i::pool_shape[1], :]
if s is None:
s = t
else:
s = T.maximum(s, t)
sym_ret = s
return sym_ret
[docs]def maxout(conv_out, factor=2, mode='max', axis=1):
"""
Pools axis 1 (the channels) of ``conv_out`` by ``factor``.
I.e. the number of channels is decreased by this factor.
The pooling can either be done as ``max`` or ``maxabs``.
Spatial dimensions are unchanged
"""
assert factor > 1
if mode == 'max':
compare = T.maximum
elif mode == 'maxabs':
compare = maxabs
if axis == 1:
ret = conv_out[:, 0::factor]
for i in xrange(1, factor):
t = conv_out[:, i::factor]
ret = compare(ret, t)
elif axis == 2:
ret = conv_out[:, :, 0::factor]
for i in xrange(1, factor):
t = conv_out[:, :, i::factor]
ret = compare(ret, t)
return ret
[docs]def pooling2d(conv_out, pool_shape=(2, 2), mode='max'):
"""
Pools axis 2,3 (x,y) of ``conv_out`` by respective ``pool_shape``.
I.e. the spatial extent is decreased by this factor.
The pooling can either be done as ``max`` or ``maxabs``.
"""
if mode == 'max':
compare = T.maximum
elif mode == 'maxabs':
compare = maxabs
ret = conv_out
if pool_shape[1] > 1:
ret = conv_out[:, :, :, 0::pool_shape[1]]
for i in xrange(1, pool_shape[1]):
t = conv_out[:, :, :, i::pool_shape[1]]
ret = compare(ret, t)
if pool_shape[0] > 1:
accum = ret[:, :, 0::pool_shape[0], :]
for i in xrange(1, pool_shape[0]):
t = ret[:, :, i::pool_shape[0], :, ]
accum = compare(accum, t)
ret = accum
return ret
[docs]def pooling3d(conv_out, pool_shape=(2, 2, 2), mode='max'):
"""
Pools axis 2,3 (x,y) of ``conv_out`` by respective ``pool_shape``.
I.e. the spatial extent is decreased by this factor.
The pooling can either be done as ``max`` or ``maxabs``.
"""
if mode == 'max':
compare = T.maximum
elif mode == 'maxabs':
compare = maxabs
ret = conv_out #(1,4,1,4,4)
if pool_shape[2] > 1:
ret = conv_out[:, :, :, :, 0::pool_shape[2]]
for i in xrange(1, pool_shape[2]):
t = conv_out[:, :, :, :, i::pool_shape[2]]
ret = compare(ret, t)
if pool_shape[0] > 1:
accum = ret[:, 0::pool_shape[0], :, :, :]
for i in xrange(1, pool_shape[0]):
t = ret[:, i::pool_shape[0], :, :, :]
accum = compare(accum, t)
ret = accum
if pool_shape[1] > 1:
accum = ret[:, :, :, 0::pool_shape[1], :]
for i in xrange(1, pool_shape[1]):
t = ret[:, :, :, i::pool_shape[1], :]
accum = compare(accum, t)
ret = accum
return ret
if __name__ == "__main__":
import sys, os
sys.path.append(os.path.expanduser("~/devel/ELEKTRONN/"))
from elektronn.training.trainutils import timeit
sym_input = T.TensorType(dtype='float32', broadcastable=[False] * 5)()
pool = (2, 2, 2)
sym_ret = my_max_pool_3d(sym_input, pool) #my_max_pool_3d_stupid(sym_input)
f_maxp_3d_ref = timeit(theano.function([sym_input], sym_ret), 4)
sym_ret = pooling3d(sym_input, pool, 'max') #my_max_pool_3d_stupid(sym_input)
f_maxp_3d = timeit(theano.function([sym_input], sym_ret), 4)
sym_ret = pooling3d(sym_input, pool, 'maxabs') #my_max_pool_3d_stupid(sym_input)
f_maxp_3d_new = timeit(theano.function([sym_input], sym_ret), 4)
# sym_ret = maxabsPool3d(sym_input, pool)#my_max_pool_3d_stupid(sym_input)
# f_maxabsp_3d= timeit(theano.function([sym_input],sym_ret), 4)
for i in range(0, 10, 2):
inp = np.random.rand(1, 16 + 32 * i, 1, 16 + 32 * i, 16 + 32 *
i).astype('float32') - 0.5
print inp.shape
a0 = f_maxp_3d_ref(inp)
a1 = f_maxp_3d(inp)
a2 = f_maxp_3d_new(inp)
#a3 = f_maxabsp_3d(inp)
assert np.all(np.equal(a1, a0))
#print a1.shape
r1 = a1[0, :, 0]
r2 = a2[0, :, 0]
#r3 = a3[0,:,0]
r0 = inp[0, :, 0]
#d = r3 - r4
