def max_pool_forward_im2col(x, pool_param):
"""
An implementation of the forward pass for max pooling based on im2col.
This isn't much faster than the naive version, so it should be avoided if
possible.
"""
N, C, H, W = x.shape
pool_height, pool_width = pool_param['pool_height'], pool_param[
'pool_width']
stride = pool_param['stride']
assert (H - pool_height) % stride == 0, 'Invalid height'
assert (W - pool_width) % stride == 0, 'Invalid width'
out_height = (H - pool_height) / stride + 1
out_width = (W - pool_width) / stride + 1
x_split = x.reshape(N * C, 1, H, W)
x_cols = im2col(x_split, pool_height, pool_width, padding=0, stride=stride)
x_cols_argmax = np.argmax(x_cols, axis=0)
x_cols_max = x_cols[x_cols_argmax, np.arange(x_cols.shape[1])]
out = x_cols_max.reshape(out_height, out_width, N, C).transpose(2, 3, 0, 1)
cache = (x, x_cols, x_cols_argmax, pool_param)
return out, cache
def get_part_info(self, image):
# Gathers information for partial patches; return a dict
# whose keys are masks for observable pixels wrt a patch,
# and values are indices of those pixels wrt the image
patchSize = int(np.sqrt(self.D))
H, W = image.shape
HFull = H + (patchSize - 1) * 2
WFull = W + (patchSize - 1) * 2
imgFull = np.reshape(np.arange(HFull * WFull), (HFull, WFull))
PgnFull = im2col(imgFull, patchSize).T
NFull = PgnFull.shape[0]
PgnPart = dict()
for n in xrange(NFull):
h, w = np.unravel_index(PgnFull[n], (HFull, WFull))
hMask = np.logical_and(h >= patchSize - 1, h <= HFull - patchSize)
wMask = np.logical_and(w >= patchSize - 1, w <= WFull - patchSize)
mask = np.logical_and(hMask, wMask)
if not np.all(mask):
h = h[mask] - (patchSize - 1)
w = w[mask] - (patchSize - 1)
idx = np.ravel_multi_index(np.array([h, w]), (H, W))
if tuple(mask) in PgnPart:
PgnPart[tuple(mask)] = np.vstack(
(PgnPart[tuple(mask)], idx))
else:
PgnPart[tuple(mask)] = np.array([idx])
return PgnPart
def _initU(self, y):
patchSize = int(np.sqrt(self.D))
patches = im2col(y, patchSize)
u = np.mean(patches, axis=0)
uPart = dict()
for mask, idx in self.PgnPart.items():
uPart[mask] = np.mean(y.ravel()[idx], axis=1)
return u, uPart
def forward(self, x, train=True):
n, c, h, w = x.shape
self.batch_size = n
out_h, out_w = filter_out_size(h, w, self.fh, self.fw, self.stride_h, self.stride_w, self.padding)
col = im2col(x, self.fh, self.fw, self.stride_h, self.stride_w, self.padding)
col = col.reshape(-1, self.fh*self.fw)
out = np.mean(col, axis=1)
out = out.reshape(n, out_h, out_w, c).transpose(0, 3, 1, 2)
return out
def forward(self, x):
N, C, H, W = x.shape
out_h = int(1 + (H - self.pool_h) / self.stride)
out_w = int(1 + (W - self.pool_w) / self.stride)
col = im2col(x, self.pool_h, self.pool_w, self.stride, self.pad)
col = col.reshape(-1, self.pool_h * self.pool_w)
out = np.max(col, axis=1)
out = out.reshape(N, out_h, out_w, C).transpose(0, 3, 1, 2)
return out
def forward(self, x):
FN, C, FH = self.W.shape
N, C, H, W = x.shape
out_h = int(1 + (H + 2 * self.pad - FH) / self.stride)
out_w = int(1 + (W + 2 * self.pad - FW) / self.stride)
col = im2col(x, FH, FW, self.stride, self.pad)
col_W = self.W.reshape(FN, -1).T
out = np.dot(col_w) + self.b
out = out.reshape(N, out_h, out_w, -1).transpose(0, 3, 1, 2)
return out
def forward(self, x, train=True):
oc, _, fh, fw = self.w.shape
n, _, h, w = x.shape
out_h, out_w = filter_out_size(h, w, fh, fw, self.stride_h, self.stride_w, self.padding)
col = im2col(x, fh, fw, self.stride_h, self.stride_w, self.padding)
col_w = self.w.reshape(oc, -1).T
out = np.dot(col, col_w) + self.b
out = out.reshape(n, out_h, out_w, -1).transpose(0, 3, 1, 2)
self.x = x
self.col = col
self.col_w = col_w
return out
def forward(self, din):
#save info to use backward procedure
self.din_shape = din.shape
#compute forward
N, C, H, W = din.shape
oh = (H + 2 * self.pad - self.fh) // self.stride + 1
ow = (W + 2 * self.pad - self.fw) // self.stride + 1
self.din = im2col(din, self.fh, self.fw, self.stride, self.pad)
col_w = self.w.reshape(self.nf, -1).transpose()
conv_result = np.dot(self.din, col_w)
conv_result += self.b
return conv_result.reshape(N, oh, ow, self.nf).transpose(0, 3, 1, 2)
def forward(self, din):
self.din_shape = din.shape
if len(self.din_shape) == 4:
N, C, H, W = self.din_shape
else:
C, H, W = self.din_shape
oh = int(1 + (H - self.fh) / self.stride)
ow = int(1 + (W - self.fw) / self.stride)
imcol = im2col(din, self.fh, self.fw, self.stride, self.pad)
#N*oh*ow*C, fh*fw로 모양 변화
imcol = imcol.reshape(-1, self.fh * self.fw)
self.mask = np.argmax(imcol, axis=1)
dout = imcol.max(axis=1)
return dout.reshape(N, oh, ow, C).transpose(0, 3, 1, 2)
def forward(self, x):
FN, C, FH, FW = self.W.shape
N, C, H, W = x.shape
_, out_h, out_w = self.output_size((C, H, W))
col = im2col(x, FH, FW, self.stride, self.pad)
col_W = self.W.reshape(FN, -1).T
out = np.dot(col, col_W) + self.b
out = out.reshape(N, out_h, out_w, -1).transpose(0, 3, 1, 2)
self.x = x
self.col = col
self.col_W = col_W
return out
def update_v(self,
beta,
x,
u,
uPart,
resp,
respPart,
patchLst=None,
IP=None):
# fully observable patches
if IP is None:
IP = self.calcIterationParams(beta)
D, K, GP, patchSize = self.D, self.K, self.GP, int(np.sqrt(self.D))
Px_minus_u = im2col(x, patchSize) - u
if patchLst is not None:
Px_minus_u = Px_minus_u[:, patchLst]
NFull = Px_minus_u.shape[1]
v = np.zeros((NFull, D))
for k in xrange(K):
idx_k = np.flatnonzero(resp == k)
if len(idx_k) == 0:
continue
cho = (IP.Rc[k] * beta, bool(IP.Rlower[k]))
v[idx_k] = cho_solve(cho,
Px_minus_u[:, idx_k],
overwrite_b=True,
check_finite=False).T
# partially observable patches
vPart = dict()
for mask, idx in self.PgnPart.items():
maskLst = np.array(list(mask), dtype=bool)
IPPart = self.calcIterationParams(beta, mask=maskLst)
NPart = len(uPart[mask])
CT_Px_minus_u = np.zeros((D, NPart))
CT_Px_minus_u[maskLst, :] = x.ravel()[idx].T - uPart[mask]
this_v = np.zeros((NPart, D))
for k in xrange(K):
idx_k = np.flatnonzero(respPart[mask] == k)
if len(idx_k) == 0:
continue
cho = (IPPart.Rc[k] * beta, bool(IPPart.Rlower[k]))
this_v[idx_k] = cho_solve(cho,
CT_Px_minus_u[:, idx_k],
overwrite_b=True,
check_finite=False).T
vPart[mask] = this_v
return v, vPart
def forward(self, x):
# Convolution
filternum, channel, filter_h, filter_w = self.W.shape
batchsize, channel, height, width = x.shape
conv_out_h = 1 + int((height + 2*self.conv_pad - filter_h) / self.conv_stride)
conv_out_w = 1 + int((width + 2*self.conv_pad - filter_w) / self.conv_stride)
self.x = x
self.x_col = im2col(x, filter_h, filter_w, self.conv_stride, self.conv_pad)
self.W_col = self.W.reshape(filternum, -1).T
self.u_col = np.dot(self.x_col, self.W_col)
self.u = self.u_col.reshape(batchsize, conv_out_h, conv_out_w, -1).transpose(0, 3, 1, 2)
out = self.u
return out
def forward(self, x):
FN, C, FH, FW = self.W.shape
N, C, H, W = x.shape
out_h = 1 + int((H + 2 * self.pad - FH) / self.stride)
out_w = 1 + int((W + 2 * self.pad - FW) / self.stride)
col = im2col(x, FH, FW, self.stride, self.pad)
col_W = self.W.reshape(FN, -1).T
out = np.dot(col, col_W) + self.b
out = out.reshape(N, out_h, out_w, -1).transpose(0, 3, 1, 2)
self.x = x
self.col = col
self.col_W = col_W
return out
def forward(self, x): N, C, H, W = x.shape ## <- (N, FN, Conv_OH, Conv_OW) out_h = 1 + int( (H - self.pool_h) / self.stride ) ## pooling OH -> OH out_w = 1 + int( (W - self.pool_w) / self.stride ) ## pooling OW -> OW col = im2col(x, self.pool_h, self.pool_w, self.stride, self.pad) # (N * OH * OW, C * PH * PW) col = col.reshape(-1, self.pool_h * self.pool_w) # (N * OH * OW * C, PH * PW) arg_max = np.argmax(col, axis=1) # Array of indices into the array, one-dimension out = np.max(col, axis=1) # (N * OH * OW * C, 1) out = out.reshape(N, out_h, out_w, C).transpose(0, 3, 1, 2) # (N, C, OH, OW) self.x = x self.arg_max = arg_max return out
def forward(self, x):
W_, b = self.params
FN, C, FH, FW = self.params[0].shape
N, C, H, W = x.shape
out_h = int(1 + (H + 2 * self.pad - FH) / self.stride)
out_w = int(1 + (W + 2 * self.pad - FW) / self.stride)
col = im2col(x, FH, FW, self.stride, self.pad)
col_W = W_.reshape(FN, -1).T
out = np.dot(col, col_W) + b
out = out.reshape(N, out_h, out_w, -1).transpose(0, 3, 1, 2)
self.x = x
self.col = col
self.col_W = col_W
return out
def forward(self, x):
N, C, H, W = x.shape
out_h = int(1 + (H - self.pool_h) / self.stride)
out_w = int(1 + (W - self.pool_w) / self.stride)
# 展開
col = im2col(x, self.pool_h, self.pool_w, self.stride, self.pad)
col = col.reshape(-1, self.pool_h * self.pool_w)
arg_max = np.argmax(col, axis=1)
# Maxプーリング
out = np.max(col, axis=1)
out = out.reshape(N, out_h, out_w, C).transpose(0, 3, 1, 2)
self.x = x
self.arg_max = arg_max
return out
def forward(self, x):
FN, C, FH, FW = self.W.shape # フィルターのパラメータ。次元順に分割して格納
N, C, H, W = x.shape # 入力データのパラメータ。次元順に分割して格納
out_h = int((H + 2 * self.pad - FH) / self.stride + 1)
out_W = int((W + 2 * self.pad - FW) / self.stride + 1)
col = im2col(x, FH, FW, self.stride, self.pad)
col_w = self.W.reshape(FN, -1).T # フィルターの展開
out = np.dot(col, col_w) + self.b
out = out.reshape(N, out_h, out_W, -1).transpose(0, 3, 1, 2)
# 逆伝播に使用
self.x = x
self.col = col
self.col_W = col_W
return out
def update_u(self, beta, x, v, vPart, patchLst=None):
# fully observable patches
D, GP, patchSize = self.D, self.GP, int(np.sqrt(self.D))
beta2inv = 1.0 / beta**2
gamma2 = 1.0 / (1.0 / GP.s2 + D * beta2inv)
patches = im2col(x, patchSize)
if patchLst is not None:
patches = patches[:, patchLst]
Px_minus_v = patches.T - v
u = gamma2 * (GP.r / GP.s2 + beta2inv * np.sum(Px_minus_v, axis=1))
# partially observable patches
uPart = dict()
for mask, idx in self.PgnPart.items():
maskLst = np.array(list(mask), dtype=bool)
NPart, DPart = idx.shape
gamma2 = 1.0 / (1.0 / GP.s2 + DPart * beta2inv)
Px_minus_v = x.ravel()[idx] - vPart[mask][:, maskLst]
uPart[mask] = gamma2 * (GP.r / GP.s2 +
beta2inv * np.sum(Px_minus_v, axis=1))
return u, uPart
def forward(self, x, train_flg=False):
# Convolution
FN, C, FH, FW = self.W.shape
N, C, H, W = x.shape
if self.first_flg:
self.W /= np.sqrt(self.T / self.win_size)
self.first_flg = False
out_h = 1 + int((H - FH)/self.T)
out_w = 1 + int((W - FW)/self.T)
col = im2col(x, FH, FW, self.T, 0)
col_W = self.W.reshape(FN, -1).T
self.u_col = np.dot(col, col_W) + self.b
self.u = self.u_col.reshape(N, out_h, out_w, -1).transpose(0, 3, 1, 2)
self.x = x
self.x_col = col
self.W_col = col_W
self.u_sum = np.sum(self.u, axis=3).reshape((N,self.N))
if self.activation == 'tanh':
self.conv_y = np.tanh(self.u_sum)
elif self.activation == 'Relu':
self.mask = (self.u_sum <= 0)
self.conv_y = self.u_sum
self.conv_y[self.mask] = 0
elif self.activation == 'None':
self.conv_y = self.u_sum
else:
print('Error : Activation function ' + self.activation + ' is undefined.')
self.arg_max = np.argmax(self.conv_y, axis=1)
y = np.max(self.conv_y, axis=1)
self.pool_y = y
self.y = self.u
return y
def forward(self, x): # input shape: (batch_num, channel, height, width) # filter shape: (filter_num, channel, filter_height, filter_width) # output shape: (batch_num, filter_num, output_height, output_width) N, C, H, W = x.shape FN, C, FH, FW = self.W.shape out_h = 1 + int( (H + 2*self.pad - FH) / self.stride ) # OH out_w = 1 + int( (W + 2*self.pad - FW) / self.stride ) # OW col = im2col(x, FH, FW, self.stride, self.pad) # (N * OH * OW, C * FH * FW) col_W = self.W.reshape(FN, -1).T # (FN, C * FH * FW) -> (C * FH * FW, FN) out = np.dot(col, col_W) + self.b # (N * OH * OW, FN) out = out.reshape(N, out_h, out_w, -1).transpose(0, 3, 1, 2) # (N, FN, OH, OW) self.x = x self.col = col self.col_W = col_W return out
def forward(self, x):
N, C, H, W = x.shape
out_h = int(1 + (H + 2 * self.pad - self.pool_h) / self.stride)
out_w = int(1 + (W + 2 * self.pad - self.pool_w) / self.stride)
col = im2col(x, self.pool_h, self.pool_w, self.stride, self.pad)
#(N * out_h * out_w, C * pool_h * pool_w) -> (N * out_h * out_w * C, pool_h * pool_w)
col = col.reshape(-1, self.pool_h * self.pool_w)
arg_max = np.argmax(col, axis=1)
out = np.max(col, axis=1)
out = out.reshape(N, out_h, out_w, C).transpose(0, 3, 1, 2)
if self.flatten:
out = out.reshape(N, -1)
self.x = x
self.arg_max = arg_max
self.out_h = out_h
self.out_w = out_w
self.C = C
return out
def update_z(self, beta, logPi, x, u, uPart, patchLst=None, IP=None):
# fully observable patches
if IP is None:
IP = self.calcIterationParams(beta)
D, K, GP, patchSize = self.D, self.K, self.GP, int(np.sqrt(self.D))
Px_minus_u = im2col(x, patchSize) - u
if patchLst is not None:
Px_minus_u = Px_minus_u[:, patchLst]
NFull = Px_minus_u.shape[1]
resp = np.tile(logPi + 0.5 * (IP.logdetSigma + GP.logdetLam),
(NFull, 1))
for k in xrange(K):
tmp = solve_triangular(beta**2 * IP.Rc[k],
Px_minus_u,
lower=IP.Rlower[k],
check_finite=False)
resp[:, k] += .5 * np.einsum('dn,dn->n', tmp, tmp)
resp = np.argmax(resp, axis=1)
# partially observable patches
respPart = dict()
for mask, idx in self.PgnPart.items():
maskLst = np.array(list(mask), dtype=bool)
IPPart = self.calcIterationParams(beta, mask=maskLst)
NPart = idx.shape[0]
CT_Px_minus_u = np.zeros((D, NPart))
CT_Px_minus_u[maskLst, :] = x.ravel()[idx].T - uPart[mask]
this_resp = np.tile(
logPi + 0.5 * (IPPart.logdetSigma + GP.logdetLam), (NPart, 1))
for k in xrange(K):
tmp = solve_triangular(beta**2 * IPPart.Rc[k],
CT_Px_minus_u,
lower=IPPart.Rlower[k],
check_finite=False)
this_resp[:, k] += .5 * np.einsum('dn,dn->n', tmp, tmp)
respPart[mask] = np.argmax(this_resp, axis=1)
return resp, respPart
def train_image_specific_topics(self,
y,
sigma,
Niter=50,
Kfresh=100,
pixelMask=None):
print('Training %d image-specific clusters...' % Kfresh)
D, patchSize, GP = self.D, int(np.sqrt(self.D)), self.GP
# gather fully observable patches
if pixelMask is None: # gray-scale image denoising
v = im2col(y, patchSize)
else: # color image inpainting
C = 3
patchMask = np.logical_not(
np.any(im2col(pixelMask, patchSize), axis=0))
v = np.hstack(
tuple([
im2col(y[:, :, c], patchSize)[:, patchMask]
for c in xrange(C)
]))
v -= np.mean(v, axis=0)
v = v.T
testData = GroupXData(X=v, doc_range=[0, len(v)], nDocTotal=1)
testData.name = 'test_image_patches'
# set up hyper-parameters and run Bregman k-means
cached_B_name = 'models/HDP/B.mat'
xBar = loadmat(cached_B_name)['Cov']
xBar2 = loadmat(cached_B_name)['Cov2']
tmp0 = (np.diag(xBar) + sigma**2)**2
tmp1 = np.diag(xBar2) + 6 * np.diag(xBar) * sigma**2 + 3 * sigma**4
nu = D + 3 + 2 * np.sum(tmp0) / np.sum(tmp1 - tmp0)
B = (nu - D - 1) * (xBar + sigma**2 * np.eye(D))
obsModel = ZeroMeanGaussObsModel(D=D,
min_covar=1e-8,
inferType='memoVB',
B=B,
nu=nu)
Z, Mu, Lscores = runKMeans_BregmanDiv(testData.X,
Kfresh,
obsModel,
Niter=Niter,
assert_monotonic=False)
Korig = self.K
Kall = np.max(Z) + Korig + 1
Kfresh = Kall - Korig
Z += Korig
# load SuffStats of training images
trainSS = loadSuffStatBag('models/HDP/SS.dump')
trainSS.insertEmptyComps(Kfresh)
# construct SuffStats of the test image
DocTopicCount = np.bincount(Z, minlength=int(Kall)).reshape((1, Kall))
DocTopicCount = np.array(DocTopicCount, dtype=np.float64)
resp = np.zeros((len(Z), Kall))
resp[np.arange(len(Z)), Z] = 1.0
testLP = dict(resp=resp, DocTopicCount=DocTopicCount)
alphaPi0 = np.hstack(
(GP.alphaPi0, GP.alphaPi0Rem / (Kfresh + 1) * np.ones(Kfresh)))
alphaPi0Rem = GP.alphaPi0Rem / (Kfresh + 1)
testLP = updateLPGivenDocTopicCount(testLP, DocTopicCount, alphaPi0,
alphaPi0Rem)
testSS = self.patchModel.get_global_suff_stats(
testData, testLP, doPrecompEntropy=1, doTrackTruncationGrowth=1)
xxT = np.zeros((Kall, D, D))
for k in xrange(Korig, Kall):
idx = Z == k
tmp = np.einsum('nd,ne->de', v[idx], v[idx])
tmp -= testSS.N[k] * sigma**2 * np.eye(D)
val, vec = np.linalg.eig(tmp)
val[val < EPS] = EPS
xxT[k] = np.dot(vec, np.dot(np.diag(val), vec.T))
testSS.setField('xxT', xxT, dims=('K', 'D', 'D'))
testSS.setUIDs(trainSS.uids)
# combine training and test SS; update model parameters
combinedSS = trainSS + testSS
self.patchModel.update_global_params(combinedSS)
self.calcGlobalParams()
def inpaint(self, y, pixelMask, T=20, **kwargs):
self.print_inpainting_info(y)
betas = 1.0 / np.sqrt(10 * np.array([1, 2, 16, 128, 512]))
D, patchSize = self.D, int(np.sqrt(self.D))
if np.any(pixelMask[:patchSize]) or np.any(pixelMask[:, :patchSize]) or \
np.any(pixelMask[-patchSize+1:]) or np.any(pixelMask[:, -patchSize+1:]):
raise ValueError(
'The current implementation does not support inpainting boundary pixels!'
)
self.PgnPart = dict()
self.train_image_specific_topics(y,
betas[-1],
pixelMask=pixelMask,
**kwargs)
mask_unseen = np.any(im2col(pixelMask, patchSize), axis=0)
mask_seen = np.logical_not(mask_unseen)
result = y.copy()
result[pixelMask] = self.GP.r
if y.ndim == 3:
C = 3
else:
raise TypeError(
'The current implementation only supports color-image inpainting!'
)
for c in xrange(C):
print('Inpainting channel %d/%d...' % (c + 1, C))
x, u, uPart, logPi = self.init_x_u_logPi(result[:, :, c])
resp_seen, respPart = self.update_z(betas[-1],
logPi,
x,
u,
uPart,
patchLst=mask_seen)
v_seen, vPart = self.update_v(betas[-1],
x,
u,
uPart,
resp_seen,
respPart,
patchLst=mask_seen)
u_seen, uPart = self.update_u(betas[-1],
x,
v_seen,
vPart,
patchLst=mask_seen)
for i, beta in enumerate(betas):
print(' beta value %d/%d' % (i + 1, len(betas)))
IP = self.calcIterationParams(beta)
for t in xrange(T):
resp_unseen, respPart = self.update_z(beta,
logPi,
x,
u,
uPart,
patchLst=mask_unseen,
IP=IP)
v_unseen, vPart = self.update_v(beta,
x,
u,
uPart,
resp_unseen,
respPart,
patchLst=mask_unseen,
IP=IP)
u_unseen, uPart = self.update_u(beta,
x,
v_unseen,
vPart,
patchLst=mask_unseen)
logPi = self.update_pi(
self.get_N(np.concatenate((resp_seen, resp_unseen)),
respPart))
NFull = len(mask_seen)
v, u = np.zeros((NFull, D)), np.zeros(NFull)
v[mask_seen] = v_seen
v[mask_unseen] = v_unseen
u[mask_seen] = u_seen
u[mask_unseen] = u_unseen
x = self.update_x_by_inpainting(result[:, :, c], v, u,
pixelMask)
print ' inner iteration %d/%d' % (t + 1, T)
result[:, :, c] = x
result = ycbcr2rgb(self.clip_pixel_intensity(result))
return result
def forward(self, x):
#print('a')
# Convolution
filternum, channel, filter_h, filter_w = self.W.shape
batchsize, channel, height, width = x.shape
conv_out_h = 1 + int(
(height + 2 * self.conv_pad - filter_h) / self.conv_stride)
conv_out_w = 1 + int(
(width + 2 * self.conv_pad - filter_w) / self.conv_stride)
self.x = x
self.x_col = im2col(x, filter_h, filter_w, self.conv_stride,
self.conv_pad)
self.W_col = self.W.reshape(filternum, -1).T
self.u_col = np.dot(self.x_col, self.W_col) + self.b
self.u = self.u_col.reshape(batchsize, conv_out_h, conv_out_w,
-1).transpose(0, 3, 1, 2)
self.conv_y_col = activation(self.mag * self.u_col)
out1 = self.conv_y_col
if self.batchnorm:
if self.batchnorm:
self.batch_size, self.out_size = self.conv_y_col.shape
self.mu = np.abs(self.conv_y_col).mean(axis=0).reshape((1, -1))
self.mu = g_repmat(self.mu, self.batch_size, 1)
self.min = np.abs(self.conv_y_col).min(axis=0).reshape((1, -1))
self.min = g_repmat(self.min, self.batch_size, 1)
self.sigma_sq = np.mean((np.abs(self.conv_y_col) - self.mu)**2,
axis=0).reshape((1, -1))
self.sigma_sq = g_repmat(self.sigma_sq, self.batch_size, 1)
self.v_amp = (np.abs(self.conv_y_col) - self.min + self.epsilon
) / np.sqrt(self.sigma_sq + self.epsilon)
out1 = self.v_amp * np.exp(1.j * np.angle(self.conv_y_col))
self.conv_y = out1.reshape(batchsize, conv_out_h, conv_out_w,
-1).transpose(0, 3, 1, 2)
out = self.conv_y
# Pooling
if self.pool_or_not:
batchsize, channel, height, width = self.conv_y.shape
pool_out_h = 1 + int(
(height + 2 * self.pool_pad - self.pool_h) / self.pool_stride)
pool_out_w = 1 + int(
(width + 2 * self.pool_pad - self.pool_w) / self.pool_stride)
self.pool_x_col = im2col(self.conv_y, self.pool_h, self.pool_w,
self.pool_stride, self.pool_pad)
self.pool_x_col = self.pool_x_col.reshape(
-1, self.pool_h * self.pool_w)
if self.pool == 'max':
self.arg_max = np.argmax(np.abs(self.pool_x_col), axis=1)
self.pool_y = np.zeros(self.arg_max.shape, dtype='complex64')
self.pool_y[np.arange(
self.arg_max.shape[0])] = self.pool_x_col[
np.arange(self.arg_max.shape[0]),
self.arg_max[np.arange(self.arg_max.shape[0])]]
elif self.pool == 'avg':
self.pool_y = np.mean(self.pool_x_col, axis=1)
self.pool_y = self.pool_y.reshape(batchsize, pool_out_h,
pool_out_w,
channel).transpose(0, 3, 1, 2)
out = self.pool_y
#print('b')
return out
def forward(self, x, train_flg=False):
# Convolution
FN, C, FH, FW = self.W.shape
N, C, H, W = x.shape
if self.batchnorm:
if self.BN_firstflg:
gamma = np.ones((C * H * W))
beta = np.zeros((C * H * W))
self.BN = BatchNormalization(gamma,
beta,
optimizer=self.optimizer)
self.BN.forward(x, train_flg)
self.BN_firstflg = False
else:
self.BN.forward(x, train_flg)
out_h = 1 + int((H + 2 * self.conv_pad - FH) / self.conv_stride)
out_w = 1 + int((W + 2 * self.conv_pad - FW) / self.conv_stride)
col = im2col(x, FH, FW, self.conv_stride, self.conv_pad)
col_W = self.W.reshape(FN, -1).T
self.u_col = np.dot(col, col_W) + self.b
self.u = self.u_col.reshape(N, out_h, out_w, -1).transpose(0, 3, 1, 2)
self.x = x
self.x_col = col
#self.W_col = col_W
if self.activation == 'tanh':
self.conv_y = np.tanh(self.u)
elif self.activation == 'Relu':
self.mask = (self.u <= 0)
self.conv_y = self.u
self.conv_y[self.mask] = 0
else:
print('Error : Activation function ' + self.activation +
' is undefined.')
if not self.pool_or_not: y = self.conv_y
if self.pool_or_not:
# Pooling
N, C, H, W = self.conv_y.shape
out_h = int(1 + (H - self.pool_h) / self.pool_stride)
out_w = int(1 + (W - self.pool_w) / self.pool_stride)
self.conv_y_col = im2col(self.conv_y, self.pool_h, self.pool_w,
self.pool_stride, self.pool_pad)
self.conv_y_col = self.conv_y_col.reshape(
-1, self.pool_h * self.pool_w)
if self.pool == 'max':
arg_max = np.argmax(self.conv_y_col, axis=1)
self.pool_y = np.max(self.conv_y_col, axis=1)
self.arg_max = arg_max
elif self.pool == 'avg':
self.pool_y = np.mean(self.conv_y_col, axis=1)
self.pool_y = self.pool_y.reshape(N, out_h, out_w,
C).transpose(0, 3, 1, 2)
y = self.pool_y
self.y = y
if self.dropout:
y = self.DO.forward(y, train_flg)
if self.gap_layer:
y = self.GAP.forward(y)
return y
import numpy as np
from util import im2col
# X1 : Data 1
x1 = np.random.rand(1, 3, 7, 7) # Data, Channel, Height, Width
print("x1:", x1.shape)
col1 = im2col(x1, 5, 5, stride=1, pad=0) # filter(5x5)
print(col1.shape) # (9,75) # OH * OW (* Data), FilterH * FilterW * Channel
#X2 : Data 10
x2 = np.random.rand(10, 3, 7, 7) # Data, Channel, Height, Width
print("x2:", x2.shape)
col2 = im2col(x2, 5, 5, stride=1, pad=0) # filter(5x5)
print(col2.shape) # (9,75) # OH * OW (* Data), FilterH * FilterW * Chann
#X3
x3 = np.random.randint(10, size=(1, 2, 3, 3))
print("x3:", x3.shape)
print(x3)
col3 = im2col(x3, 2, 2, stride=1, pad=0)
print(col3.shape)
print(col3)
def cnn():
x1 = np.random.rand(1, 3, 7, 7)
col1 = im2col(x1, 5, 5, stride=1, pad=0)
print(col1.shape)
self.x = None self.col = None self.col_W = None # 가중치와 편향 매개변수의 기울기 self.dW = None self.db = None def forward(self, x): FN, C, FH, FW = self.W.shape N, C, H, W = x.shape out_h = 1 + int((H + 2*self.pad - FH) / self.stride) out_w = 1 + int((W + 2*self.pad - FW) / self.stride) col = im2col(x, FH, FW, self.stride, self.pad) col_W = self.W.reshape(FN, -1).T # 내적 하기 위해, -1은 다차원 배열의 원소 수가 변환 후에도 똑같이 유지 되도록 묶어줌. out = np.dot(col, col_W) + self.b out = out.reshape(N, out_h, out_w, -1).transpose(0, 3, 1, 2) self.x = x self.col = col self.col_W = col_W return out def backward(self, dout): FN, C, FH, FW = self.W.shape dout = dout.transpose(0,2,3,1).reshape(-1, FN)
