def backward(self, dout):
# dout : [BS, in_D, out_H, out_W]
# self.in_col : [f_H*f_W*in_D, out_H*out_W*BS]
dout_reshape = dout.transpose((2,3,0,1)).reshape((-1, self.in_D)).T.reshape((1,-1),order='F') # [ out_H *out_W * BS*in_D]
din_col = self.W * dout_reshape # [f_H*f_W, out_H *out_W * BS * in_D]
din_col = din_col.reshape((self.f_H * self.f_W * self.in_D, -1), order='F') # [f_H*f_W * in_D, out_H *out_W * BS]
din = col2im(din_col, self.input_shape, [self.f_H, self.f_W], [self.s_H, self.s_W], [self.p_H, self.p_W])
return din # [BS, in_D, in_H, in_W]
def back_propagate(self, dout):
dout_reshaped = dout.transpose(1, 2, 3, 0).reshape(
self.in_D, 1, self.out_H * self.out_W *
self.BS) #shape=(BS,in_D,out_H,out_W)->(in_D,out_H*out_W*BS)
din_col = (self.iomat * dout_reshaped).reshape(
self.in_D * self.f_H * self.f_W, self.out_H * self.out_W * self.BS)
din = col2im(din_col, self.input_shape, [self.f_H, self.f_W],
[self.stride_H, self.stride_W], [self.pad_H, self.pad_W])
return din
def backward(self, dout):
# dout : [BS, out_D, out_H, out_W]
# self.in_col : [f_H*f_W*in_D, out_H*out_W*BS]
self.db = np.sum(dout, axis=(0,2,3)).reshape((self.out_D,1)) # [out_D, 1]
dout_reshape = np.transpose(dout, (1,2,3,0)).reshape((self.out_D,-1)) #[out_D, out_H *out_W * BS]
self.dw = np.matmul(dout_reshape, self.in_col.T) #[out_D, in_D*f_H*f_W]
din_col = np.matmul(self.w_col.T, dout_reshape) # [f_H*f_W*in_D, out_H *out_W * BS]
din = col2im(din_col, self.input_shape, [self.f_H, self.f_W], [self.s_H, self.s_W], [self.p_H, self.p_W])
return din #[BS, in_D, in_H, in_W]
def max_pooling_backward(x, dout, pool_params):
H, W, D, N = x.shape
x_reshaped = x.reshape(H, W, 1, -1)
x_col = im2col(x_reshaped, pool_params['HF'],
pool_params['WF'], pool_params['pad'], pool_params['stride'])
x_col_argmax = np.argmax(x_col, axis=0)
dx_col = np.zeros_like(x_col)
dx_col[x_col_argmax, np.arange(x_col.shape[1])] = dout.ravel()
dx_shaped = col2im(dx_col, x_reshaped.shape, pool_params['HF'], pool_params['WF'],
pool_params['pad'], stride=pool_params['stride'])
dx = dx_shaped.reshape(x.shape)
return [dx]
def conv_backward(x, w, b, conv_param, dout):
HF, WF, DF, NF = w.shape
x_col = im2col(x, HF, WF, conv_param['pad'], conv_param['stride'])
w_col = w.transpose(3, 0, 1, 2).reshape((NF, -1))
db = np.sum(dout, axis=(0, 1, 3))
dout = dout.transpose(2, 0, 1, 3)
dout = dout.reshape((w_col.shape[0], x_col.shape[-1]))
dx_col = w_col.T.dot(dout)
dw_col = dout.dot(x_col.T)
dx = col2im(dx_col, x.shape, HF, WF, conv_param['pad'], conv_param['stride'])
dw = dw_col.reshape((dw_col.shape[0], HF, WF, DF))
dw = dw.transpose(1, 2, 3, 0)
return [dx, dw, db]
def max_pooling_backward(x, dout, pool_params):
H, W, D, N = x.shape
x_reshaped = x.reshape(H, W, 1, -1)
x_col = im2col(x_reshaped, pool_params['HF'], pool_params['WF'],
pool_params['pad'], pool_params['stride'])
x_col_argmax = np.argmax(x_col, axis=0)
dx_col = np.zeros_like(x_col)
dx_col[x_col_argmax, np.arange(x_col.shape[1])] = dout.ravel()
dx_shaped = col2im(dx_col,
x_reshaped.shape,
pool_params['HF'],
pool_params['WF'],
pool_params['pad'],
stride=pool_params['stride'])
dx = dx_shaped.reshape(x.shape)
return [dx]
def conv_backward(x, w, b, conv_param, dout):
HF, WF, DF, NF = w.shape
x_col = im2col(x, HF, WF, conv_param['pad'], conv_param['stride'])
w_col = w.transpose(3, 0, 1, 2).reshape((NF, -1))
db = np.sum(dout, axis=(0, 1, 3))
dout = dout.transpose(2, 0, 1, 3)
dout = dout.reshape((w_col.shape[0], x_col.shape[-1]))
dx_col = w_col.T.dot(dout)
dw_col = dout.dot(x_col.T)
dx = col2im(dx_col, x.shape, HF, WF, conv_param['pad'],
conv_param['stride'])
dw = dw_col.reshape((dw_col.shape[0], HF, WF, DF))
dw = dw.transpose(1, 2, 3, 0)
return [dx, dw, db]
def back_propagate(self, dout):
db = np.sum(dout, axis=(0, 2, 3))
self.db = db.reshape(self.out_D, 1) #shape=(out_D,1)
dout_reshaped = dout.transpose(1, 2, 3, 0).reshape(
self.out_D,
-1) #shape=(BS,out_D,out_H,out_W)->(out_D,out_H*out_W*BS)
self.dW_col = np.matmul(dout_reshaped,
self.X_col.T) #shape=(out_D,f_H*f_W*in_D)
din_col = np.matmul(
self.W_col.T, dout_reshaped) #shape=(f_H*f_W*in_D,out_H*out_W*BS)
din = col2im(din_col, self.input_shape, [self.f_H, self.f_W],
[self.stride_H, self.stride_W], [self.pad_H, self.pad_W])
return din
def max_pooling_backward(x, dout, pool_params):
print "in max_pooling_backward"
print "dout.shape", dout.shape
print "x.shape", x.shape
H, W, D, N = x.shape
x_reshaped = x.reshape(H, W, 1, -1)
x_col = im2col(x_reshaped, pool_params['HF'], pool_params['WF'],
pool_params['pad'], pool_params['stride'])
x_col_argmax = np.argmax(x_col, axis=0)
dx_col = np.zeros_like(x_col) #和x_col同样纬度的0矩阵
print " 1 dx_col.shape", dx_col.shape
dx_col[x_col_argmax, np.arange(x_col.shape[1])] = dout.ravel() #把dout平铺
print " 2 dx_col.shape", dx_col.shape
dx_shaped = col2im(dx_col,
x_reshaped.shape,
pool_params['HF'],
pool_params['WF'],
pool_params['pad'],
stride=pool_params['stride'])
dx = dx_shaped.reshape(x.shape)
return [dx]
def conv_backward(x, w, b, conv_param, dout):
print " in conv_backward"
HF, WF, DF, NF = w.shape
print "dout.shape", dout.shape
x_col = im2col(x, HF, WF, conv_param['pad'], conv_param['stride'])
#转换后变成(HF*WF*DF,N*Hout*Wout)
print "x_col.shape", x_col.shape
w_col = w.transpose(3, 0, 1, 2).reshape((NF, -1)) #每一行是一个卷积核
#w_col的维度是(NF,HF,WF,DF),reshape成(NF,HF*WF*DF)
db = np.sum(dout, axis=(0, 1, 3))
dout = dout.transpose(2, 0, 1, 3) #(NF,Hout,Wout, N)
dout = dout.reshape((w_col.shape[0], x_col.shape[-1])) #(NF,N*Hout*Wout)
dx_col = w_col.T.dot(dout) #当前层的残差 , （HF*WF*DF, N*Hout*Wout),和x_col的维度相同
dw_col = dout.dot(x_col.T) #当前层关于卷积核的梯度
dx = col2im(dx_col, x.shape, HF, WF, conv_param['pad'],
conv_param['stride'])
dw = dw_col.reshape((dw_col.shape[0], HF, WF, DF))
dw = dw.transpose(1, 2, 3, 0)
return [dx, dw, db]
def decompose_filter(parent_filter_wt, filters=16):
lamda = 0.0001
error = 1e-7
c1 = parent_filter_wt.shape[1]
c2 = parent_filter_wt.shape[0]
k = parent_filter_wt.shape[2]
k1 = k
k2 = k
k_expanded = k1 + k2 - 1
pad_zero = nn.ZeroPad2d((k_expanded - k) / 2)
new_weight = pad_zero(parent_filter_wt)
# new_weight = add_noise(new_weight, new_weight)
new_weight = new_weight.cpu().numpy()
# output is the original parent filter which is generated by
# convolving an image (=img_col) by filter (=kernel)
# output_col is the 2D representation of an output generated after matrix
# multiplication
# output = new_weight
# wt = parent_filter_wt.cpu().numpy()
# mean = wt.mean()
# std = wt.std()
# var = wt.var()
output = np.concatenate(
(new_weight, np.zeros((c2, c2 - c1, k_expanded, k_expanded))), axis=1)
output = np.concatenate(
(output, np.zeros((filters - c2, c2, k_expanded, k_expanded))), axis=0)
# NOISE_RATIO = 1e-5
# noise_range = NOISE_RATIO * np.ptp(parent_filter_wt.flatten())
# noise = np.random.uniform(-noise_range, noise_range, size=output.shape)
# output = output + noise
output_col = output.reshape(filters, -1).T
# output_col2 = np.random.normal(0, 1e-2, size=output_col.shape)
# print np.linalg.norm(output_col2 - output_col)
# exit()
# Below 2 lines can be removed
# kernel is equivalent to filter f1 which will convolve image (=img_col)
kernel = np.random.normal(0, 1e-3, size=(filters, c1, k1, k1))
kernel = np.concatenate(
(kernel, np.zeros((filters, filters - c1, k1, k1))), axis=1)
# kernel = np.random.choice(output.flatten(), size=(c2, filters, k1, k1))
kernel_col = kernel.reshape(filters, -1).T
# img is the f2 filter treated as image to be convolved by f1(=kernel)
# img_col is the 2D representation of a filter for matrix multiplication
img = np.random.normal(0, 1e-3, size=(c2, filters, k2, k2))
# # img = np.random.choice(output.flatten(), size=(c2, filters, k2, k2))
img_col = im2col.im2col(img, k1, k1, stride=1, padding=k_expanded - k)
# img_col = np.random.normal(
# 0, 1e-2, size=(k_expanded * k_expanded * c2, k1 * k1 * filters))
# img_col_original = img_col.copy()
# kernel_col = np.linalg.lstsq(img_col, output_col, rcond=None)[0]
print kernel_col.shape
print img_col.shape
print output_col.shape
print 'before calculating prod: ',
print np.linalg.norm(np.dot(img_col, kernel_col) - output_col)
# exit()
for i in range(10):
img_col = np.linalg.solve(
np.dot(kernel_col, kernel_col.T) + lamda * np.eye(
kernel_col.shape[0]),
np.dot(kernel_col, output_col.T)).T
kernel_col = np.linalg.solve(
img_col.T.dot(img_col) + lamda * np.eye(img_col.shape[1]),
np.dot(img_col.T, output_col))
print np.linalg.norm(np.dot(img_col, kernel_col) - output_col)
if np.linalg.norm(np.dot(img_col, kernel_col) - output_col) < error:
break
x1 = img_col
# c = 0.25
# kernel_col = kernel_col * c
# img_col = img_col / c
# Using Weighted ALS
# print output_col
# z = output_col > 0
# z = z.astype(np.float32)
# for n in range(20):
# for i, zi in enumerate(z):
# img_col[i] = np.linalg.solve(
# np.dot(kernel_col, np.dot(np.diag(zi), kernel_col.T)) + lamda * np.eye(kernel_col.shape[0]),
# np.dot(kernel_col, np.dot(np.diag(zi), output_col[i].T))).T
#
# for j, zj in enumerate(z.T):
# kernel_col[:, j] = np.linalg.solve(
# np.dot(img_col.T, np.dot(np.diag(zj), img_col)) + lamda * np.eye(img_col.shape[1]),
# np.dot(img_col.T, np.dot(np.diag(zj), output_col[:, j])))
#
# print np.linalg.norm(np.dot(img_col, kernel_col) - output_col)
# if np.linalg.norm(np.dot(img_col, kernel_col) - output_col) < error:
# break
print 'after calculating prod: ',
new_prod = np.dot(img_col, kernel_col)
print np.linalg.norm(new_prod - output_col)
kernel = kernel_col.T.reshape(filters, filters, k1, k1)
kernel = kernel[:, :c1, ...]
print 'diff pad',
print k_expanded - k
img_calculated = im2col.col2im(col=img_col, input_shape=(c2, filters, k2, k2),
filter_h=k1, filter_w=k1,
padding=k_expanded - k)
# img_calculated = im2col.recover_input(
# input=img_col, kernel_size=k1, stride=1, outshape=(c2, filters, k2, k2))
img_calculated = img_calculated / 9 # because original matrix elements are added 9 times , for double padding
# img_calculated = img_calculated/[[1, 2, 1], [2, 4, 2], [1, 2, 1]] for zero padding
# img = (img / ([[4, 6, 4], [6, 9, 6], [4, 6, 4]]))/ for single padding
# print 'image_col error: ',
# print np.linalg.norm((img_col - img_col_original))
print 'image error: ',
print np.linalg.norm((img_calculated - img))
#
img_col2 = im2col.im2col(img_calculated, k1, k1, stride=1, padding=k_expanded - k)
print 'after converting, product error = ',
print np.linalg.norm(img_col2 - x1)
# exit()
# exit()
# img = im2col.recover_input(input=img_col, kernel_size=k1, stride=1,
# outshape=(c2, c, k2, k2))
# exit()
# *************************************************************
# img = np.random.normal(0, 1e-2, size=(4, 4, 3, 3))
# original_img = img.copy()
# img_col = im2col.im2col(img, 3, 3, 1, 2)
# original_img_col = img_col.copy()
# # print img_col
# # img_col = np.random.randint(0, 4, size=(100, 36))
# # print img_col[0, 0]
# # kernel_col = np.random.randint(0, 2, size=(36, 4))
# output_col = np.random.normal(0, 1e-2, size=(100, 4))
#
# kernel_col = np.linalg.lstsq(img_col, output_col, rcond=None)[0]
#
# for i in range(100):
# img_col = np.linalg.solve(
# np.dot(kernel_col, kernel_col.T) + lamda * np.eye(
# kernel_col.shape[0]),
# np.dot(kernel_col, output_col.T)).T
#
# kernel_col = np.linalg.solve(
# img_col.T.dot(img_col) + lamda * np.eye(img_col.shape[1]),
# np.dot(img_col.T, output_col))
#
# # print np.linalg.norm(np.dot(img_col, kernel_col) - output_col)
# if np.linalg.norm(np.dot(img_col, kernel_col) - output_col) < error:
# break
#
# print 'before converting after calcultating',
# new_prod = np.dot(img_col, kernel_col)
# print np.linalg.norm(new_prod - output_col)
#
# print np.linalg.norm(img_col - original_img_col)
# img = im2col.col2im(img_col, (4, 4, 3, 3), 3, 3, padding=2)
# img = img / 9
#
# print np.linalg.norm(img - original_img)
# img_col2 = im2col.im2col(img, 3, 3, padding=2)
# new_prod = np.dot(img_col2, kernel_col)
# print np.linalg.norm(new_prod - output_col)
# exit()
# *************************************************************
print parent_filter_wt.shape
print kernel.shape
print img_calculated.shape
exit()
return kernel, img_calculated
