def map_segments(self):
# merge overlapping segments
segments = [align(addr, size, grow=True) for addr, size, _ in self.loader.segments()]
merged = []
for addr, size in segments:
left = addr
right = addr + size
for a2, s2 in merged:
if (addr >= a2 and addr <= a2 + s2) or (addr < a2 and addr + size > a2):
left = min(a2, addr)
right = max(right, a2 + s2)
merged.remove((a2, s2))
merged.append((left, right - left))
for addr, size in merged:
self.uc.mem_map(addr, size)
for addr, size, data in self.loader.segments():
# FIXME: weird, if I don't touch the data before write it segfaults on ARM
# Issue #15
binascii.hexlify(data)
self.uc.mem_write(addr, data)
# TODO: ask loader for stack size/location
self.stack = self.uc.mmap(STACK_SIZE, addr_hint=STACK_BASE)
self.uc.reg_write(self.arch.sp, self.stack + STACK_SIZE - self.bsz)
# test_path = 'data/mini_mnist'
test_path = 'data/mnist.test'
X_train, Y_train = read_dense_data(open(train_path))
print >> sys.stderr, 'read training data done.'
X_train = np.matrix(X_train)
Y_train = [int(y) for y in Y_train]
Y_train = np.matrix(Y_train).T
print >> sys.stderr, 'create training matrix done.'
X_test, Y_test = read_dense_data(open(test_path))
print >> sys.stderr, 'read test data done'
X_test = np.matrix(X_test)
Y_test = [int(y) for y in Y_test]
Y_test = np.matrix(Y_test).T
print >> sys.stderr, 'create test matrix done.'
X_train, X_test = align(X_train, X_test)
X_train, X_test = normalize(X_train, X_test)
clf = NeuralNetwork()
clf.train(X_train, Y_train)
# clf.train(X_test, Y_test)
acc_train = clf.test(X_train, Y_train)
acc_test = clf.test(X_test, Y_test)
print >> sys.stderr, 'Training accuracy for Neural Network : %lf%%' % (
100.0 * acc_train)
print >> sys.stderr, 'Test accuracy for Neural Network : %lf%%' % (
100.0 * acc_test)
def _dump(self):
"""
Returns all data in this U8 archive as bytes
"""
header = self.U8Header()
rootnode = self.U8Node()
# constants
header.tag = b'U\xAA8-'
header.rootnode_offset = 0x20
header.zeroes = b'\x00' * 16
rootnode.type = 0x0100
nodes = []
strings = b'\x00'
data = b''
for item, value in self.files:
node = self.U8Node()
recursion = item.count('/')
if recursion < 0:
recursion = 0
name = item.split('/')[-1]
node.name_offset = len(strings)
strings += name.encode('latin-1') + b'\x00'
if value is None: # directory
node.type = 0x0100
node.data_offset = recursion
node.size = len(nodes) + 1
for one, two in self.files:
if one[:len(item)] == item: # find nodes in the folder
node.size += 1
else: # file
node.type = 0x0000
node.data_offset = len(data)
data += value + (
b'\x00' * (align(len(value), 32) - len(value))
) # 32 seems to work best for fuzzyness? I'm still really not sure
node.size = len(value)
nodes.append(node)
header.header_size = ((len(nodes) + 1) * len(rootnode)) + len(strings)
header.data_offset = align(header.header_size + header.rootnode_offset,
64)
rootnode.size = len(nodes) + 1
for i in range(len(nodes)):
if nodes[i].type == 0x0000:
nodes[i].data_offset += header.data_offset
fd = b''
fd += header.pack()
fd += rootnode.pack()
for node in nodes:
fd += node.pack()
fd += strings
fd += b'\x00' * (header.data_offset - header.rootnode_offset -
header.header_size)
fd += data
return fd
X_train, Y_train = read_dense_data(open(train_path))
print >> sys.stderr, 'read training data done.'
X_train = np.matrix(X_train)
Y_train = [int(y) for y in Y_train]
Y_train = np.matrix(Y_train).T
print >> sys.stderr, 'create training matrix done.'
X_test, Y_test = read_dense_data(open(test_path))
print >> sys.stderr, 'read test data done'
X_test = np.matrix(X_test)
Y_test = [int(y) for y in Y_test]
Y_test = np.matrix(Y_test).T
print >> sys.stderr, 'create test matrix done.'
X_train, X_test = align(X_train, X_test)
'''
X_all = np.row_stack([X_train, X_test])
print X_all.shape
mean = X_all.mean(0)
std = X_all.std(0)
del X_all
X_train = 1.0 * (X_train - mean) / (std + 0.0001)
X_test = 1.0 * (X_test - mean) / (std + 0.0001)
'''
clf = SoftmaxRegression()
clf.train(X_train, Y_train)
# clf.train(X_test, Y_test)