def SaabTransform(self, X, saab, train, layer):
SaabArg = self.SaabArgs[layer]
X = self.Neighbor(X, layer)
S = list(X.shape)
X = X.reshape(-1, S[-1])
if SaabArg['num_AC_kernels'] != -1:
S[-1] = SaabArg['num_AC_kernels']
if train == True:
saab = Saab(num_kernels=SaabArg['num_AC_kernels'],
useDC=SaabArg['useDC'],
needBias=SaabArg['needBias'])
saab.fit(X)
# use batch to avoid memory error
batch_size = int(X.shape[0] / 5)
x1 = saab.transform(X[:1 * batch_size])
x2 = saab.transform(X[1 * batch_size:2 * batch_size])
x3 = saab.transform(X[2 * batch_size:3 * batch_size])
x4 = saab.transform(X[3 * batch_size:4 * batch_size])
x5 = saab.transform(X[4 * batch_size:])
del X
X = np.concatenate((x1, x2, x3, x4, x5), axis=0)
del x1, x2, x3, x4, x5
X = X.reshape(S)
X = self.Pooling(X, layer)
return saab, X
def SaabTransform(self, X, saab, train, layer):
shrinkArg, SaabArg = self.shrinkArgs[layer], self.SaabArgs[layer]
assert ('func'
in shrinkArg.keys()), "shrinkArg must contain key 'func'!"
X = shrinkArg['func'](X, shrinkArg)
S = list(X.shape)
X = X.reshape(-1, S[-1])
if SaabArg['num_AC_kernels'] != -1:
S[-1] = SaabArg['num_AC_kernels']
if train == True:
saab = Saab(num_kernels=SaabArg['num_AC_kernels'],
useDC=SaabArg['useDC'],
needBias=SaabArg['needBias'])
saab.fit(X)
print("fit finish")
# batch_size = 10000
num_batches = 15
batch_size = int(X.shape[0] / num_batches)
flag = False
transformed = None
dc = []
for i in range(num_batches - 1):
print("batch :", i)
X[i * batch_size:(i + 1) * batch_size], dc_tmp = saab.transform(
X[i * batch_size:(i + 1) * batch_size])
#transformed_tmp = transformed_tmp.reshape(S)
dc.append(dc_tmp)
X[(num_batches - 1) * batch_size:], dc_tmp = saab.transform(
X[(num_batches - 1) * batch_size:])
dc.append(dc_tmp)
dc = np.concatenate(dc, axis=0)
X = X.reshape(S)
# print(transformed.shape)
print("current layer", layer)
#batch_size = int(X.shape[0] / num_batches)
#poolwin = shrinkArg['poolwin']
#X = view_as_windows(X,(1,poolwin,poolwin,1),(1,poolwin,poolwin,1))
#print(X.shape)
#X = X.reshape(X.shape[0],X.shape[1],X.shape[2],X.shape[3],-1)
#X = shrinkArg['method'](X,axis = -1)
print("shape before pooling", X.shape)
if layer != 2:
pool_win = shrinkArg['poolwin']
pool_method = shrinkArg['method']
num_batches = 6
X = poolImage_batch(X, num_batches, pool_win, pool_method)
print("shape of pooled X: ", X.shape)
return saab, X, dc
def SaabTransform(self, X, saab, train, layer):
shrinkArg, SaabArg = self.shrinkArgs[layer], self.SaabArgs[layer]
assert ('func'
in shrinkArg.keys()), "shrinkArg must contain key 'func'!"
X = shrinkArg['func'](X, shrinkArg)
S = list(X.shape)
X = X.reshape(-1, S[-1])
if SaabArg['num_AC_kernels'] != -1:
S[-1] = SaabArg['num_AC_kernels']
if train == True:
isInteger, bits, opType, whichPCA = False, 8, 'int32', 'numpy'
if 'isInteger' in SaabArg.keys():
isInteger = SaabArg['isInteger']
if 'bits' in SaabArg.keys():
bits = SaabArg['bits']
if 'opType' in SaabArg.keys():
opType = SaabArg['opType']
if 'whichPCA' in SaabArg.keys():
whichPCA = SaabArg['opType']
saab = Saab(num_kernels=SaabArg['num_AC_kernels'],
useDC=SaabArg['useDC'],
needBias=SaabArg['needBias'],
isInteger=isInteger,
bits=bits,
opType=opType)
saab.fit(X, whichPCA=whichPCA)
transformed = saab.transform(X).reshape(S)
return saab, transformed
class Pixelhop():
def __init__(self, dilate, pad, SaabArg, batch=None):
self.saab = Saab(num_kernels=SaabArg['num_AC_kernels'],
useDC=SaabArg['useDC'],
needBias=SaabArg['needBias'])
self.dilate = np.array([dilate]).tolist()
self.pad = pad
self.batch = batch
self.trained = False
def fit(self, X):
if self.batch == None:
X = PixelHop_Neighbour(X, self.dilate, self.pad)
else:
X = Batch_PixelHop_Neighbour(X, self.dilate, self.pad, self.batch)
X = X.reshape(-1, X.shape[-1])
self.saab.fit(X)
self.trained = True
def transform(self, X):
assert (self.trained == True), "Call fit first!"
if self.batch == None:
X = PixelHop_Neighbour(X, self.dilate, self.pad)
else:
X = Batch_PixelHop_Neighbour(X, self.dilate, self.pad, self.batch)
S = X.shape
X = X.reshape(-1, X.shape[-1])
X, DC = self.saab.transform(X)
X = X.reshape(S[0], S[1], S[2], -1)
return X, DC
class PixelHop_Unit():
def __init__(self, X, num_kernels, window=5, stride=1):
self.X = self.Shrink(X, window, stride) #N*28*28*(5*5*1)
self.S = list(self.X.shape)
self.X = self.X.reshape(-1, self.S[-1]) #(N*28*28)*(5*5*1)
self.num_kernels = num_kernels
self.saab = None
self.window = window
self.stride = stride
def train(self):
self.saab = Saab(num_kernels=self.num_kernels, useDC=True)
self.saab.fit(self.X)
def transform(self, X):
X = self.Shrink(X, self.window, self.stride) #N*28*28*(5*5*1)
S = list(X.shape)
X = X.reshape(-1, S[-1]) #(N*28*28)*(5*5*1)
assert (self.saab != None
), "the model hasn't been trained, must call train() first!"
transformed = self.saab.transform(X).reshape(S[0], S[1], S[2],
-1) #N*28*28*25
return self.saab, transformed
def Shrink(self, X, win, stride):
X = view_as_windows(X, (1, win, win, 1), (1, stride, stride, 1))
return X.reshape(X.shape[0], X.shape[1], X.shape[2], -1)
def PixelHop_Unit(X, num_kernels, saab=None, window=5, stride=1, train=True):
print('input shape', X.shape)
X = Shrink(X, 5, 1)
print('extracting patches', X.shape)
S = list(X.shape)
X = X.reshape(-1, S[-1])
if (train == True):
saab = Saab(num_kernels=num_kernels, useDC=True, needBias=True)
saab.fit(X)
transformed = saab.transform(X).reshape(S[0], S[1], S[2], -1)
print('transformed shape', transformed.shape)
return saab, transformed
def SaabTransform(self, X, saab, train, layer):
shrinkArg, SaabArg = self.shrinkArgs[layer], self.SaabArgs[layer]
assert ('func' in shrinkArg.keys()), "shrinkArg must contain key 'func'!"
X = shrinkArg['func'](X, shrinkArg)
S = list(X.shape)
X = X.reshape(-1, S[-1])
if SaabArg['num_AC_kernels'] != -1:
S[-1] = SaabArg['num_AC_kernels']
if train == True:
saab = Saab(num_kernels=SaabArg['num_AC_kernels'], useDC=SaabArg['useDC'], needBias=SaabArg['needBias'])
saab.fit(X)
transformed, dc = saab.transform(X)
transformed = transformed.reshape(S)
return saab, transformed, dc
def fit_hop1(self, images, verbose):
# train the first hop
if verbose:
print('Hop1')
print("Input shape:", images.shape)
saab = Saab(kernel_size=self.kernel_sizes[0], bias_flag=False)
saab.fit(images)
self.saabs['Hop1'] = [saab]
self.energies['Hop1'] = [saab.eigenvalues / sum(saab.eigenvalues)]
n_channels = np.sum(self.energies['Hop1'][0] > self.keep_thr)
output = saab.transform(images, n_channels)
self.features['Hop1'] = [self.max_pooling(output)]
self.info['Hop1'] = [(0, 0, n_channels)]
if verbose:
print("Output shape:", self.features['Hop1'][-1].shape)
def fit_hop_n(self, n, verbose):
# train the nth hop (n > 1)
if verbose:
print('Hop' + str(n))
self.saabs['Hop' + str(n)] = []
self.energies['Hop' + str(n)] = []
self.features['Hop' + str(n)] = []
self.info['Hop' + str(n)] = []
for saab_id in range(len(self.saabs['Hop' + str(n - 1)])):
saab_parent = self.saabs['Hop' + str(n - 1)][saab_id]
energies_parent = self.energies['Hop' + str(n - 1)][saab_id]
features_parent = self.features['Hop' + str(n - 1)][saab_id]
for channel_id in range(len(energies_parent)):
energy = energies_parent[channel_id]
if energy > self.split_thr:
features = features_parent[:, :, :, channel_id][...,
np.newaxis]
if verbose:
print("SaabID:", saab_id, "ChannelID:", channel_id,
"Energy:", energy)
print("Input shape:", features.shape)
saab = Saab(kernel_size=self.kernel_sizes[n - 1],
bias_flag=True)
saab.fit(features)
self.saabs['Hop' + str(n)].append(saab)
energies = saab.eigenvalues / sum(
saab.eigenvalues) * energy
self.energies['Hop' + str(n)].append(energies)
n_channels = np.sum(energies > self.keep_thr)
output = saab.transform(features, n_channels)
self.tmp.append(
(saab_id, channel_id, self.max_pooling(output)))
self.features['Hop' + str(n)].append(
self.max_pooling(output))
self.info['Hop' + str(n)].append(
(saab_id, channel_id, n_channels))
if verbose:
print("Output shape:",
self.features['Hop' + str(n)][-1].shape)
else:
break