# PixlHop ++
for i in range(
22
): # pixelHop 0, 1: {TH1: 0.05, TH2: 0.005} finished # features: 0 finished
print(i)
# p2_1 = Pixelhop2(depth=2, TH1=0.0000001, TH2=0.00000001, SaabArgs=SaabArgs, shrinkArgs=shrinkArgs,
# concatArg=concatArg).fit(train_selected[:, :, :, i][:, :, :, None])
# f = open("/mnt/zhengwen/image_steganalysis/dataset/codes/PixelHopStegoPts_5_5_2nd_3rd_layer_" + str(i) + "_PIXEL_sample.pkl", 'wb')
# pickle.dump(p2_1, f)
# f.close()
p2_1 = Pixelhop2(depth=2,
TH1=0.0000001,
TH2=0.00000001,
SaabArgs=SaabArgs,
shrinkArgs=shrinkArgs,
concatArg=concatArg).fit(train_selected[:, :, :,
i][:, :, :, None])
f = open(
"/mnt/zhengwen/image_steganalysis/dataset/codes/PixelHopStegoPts_5_5_2nd_3rd_layer_"
+ str(i) + "_PIXEL_sample_ALL.pkl", 'wb')
pickle.dump(p2_1, f)
f.close()
feature_p2_1 = p2_1.transform(train_selected[:, :, :, i][:, :, :, None])
print(i, feature_p2_1[1].shape)
# np.save("/mnt/zhengwen/image_steganalysis/dataset/codes/end_features_from_" + str(i) + "_PIXEL_0_sample.npy", feature_p2_1[0])
# np.save("/mnt/zhengwen/image_steganalysis/dataset/codes/end_features_from_" + str(i) + "_PIXEL_1_sample.npy", feature_p2_1[1])
np.save(
"/mnt/zhengwen/image_steganalysis/dataset/codes/end_features_from_" +
'func': Shrink,
'win': 3
}]
concatArg = {'func': Concat}
train_patches = np.concatenate((original_vectors, stego_vectors), axis=0)
train_labels = np.concatenate(
(0 * np.ones(len(original_vectors)), 1 * np.ones(len(stego_vectors))),
axis=0)
idx = np.random.permutation(len(train_patches))
train_patches = train_patches[idx]
train_labels = train_labels[idx]
p2_original = Pixelhop2(depth=2,
TH1=0.005,
TH2=0.001,
SaabArgs=SaabArgs,
shrinkArgs=shrinkArgs,
concatArg=concatArg)
p2_original.fit(train_patches[:, :, :, None])
f = open('/mnt/zhengwen/new_trial/p2_original_week8_HONGSHUO', 'wb')
pickle.dump(p2_original, f)
f.close()
train_features = p2_original.transform(train_patches[:, :, :, None])
pixel_hop_2 = train_features[1]
n_channels = pixel_hop_2.shape[-1]
N_train = pixel_hop_2.shape[0]
params = {
return X
# set args for PixelHop++
SaabArgs = [{
'num_AC_kernels': -1,
'needBias': True,
'useDC': True,
'batch': None
}]
shrinkArgs = [{'func': Shrink, 'win': 5}]
concatArg = {'func': Concat}
p2 = Pixelhop2(depth=1,
TH1=1e-20,
TH2=1e-30,
SaabArgs=SaabArgs,
shrinkArgs=shrinkArgs,
concatArg=concatArg).fit(steg_patches)
# f = open("/mnt/zhengwen/image_steganalysis/dataset/codes/PixelHopStegoPts_5_5_1st_layer.pkl", 'wb') # overlapping
f = open(
"/mnt/zhengwen/image_steganalysis/dataset/codes/PixelHopStegoPts_5_5_1st_layer_sampled.pkl",
'wb') # sampled
# f = open("/mnt/zhengwen/image_steganalysis/dataset/codes/PixelHopStegoPts_7_7_1st_layer.pkl", 'wb')
pickle.dump(p2, f)
f.close()
features_cover = p2.transform(cover_patches)
features_steg = p2.transform(steg_patches)
features_cover = np.squeeze(features_cover)
}, {
'func': Shrink,
'win': 5
}, {
'func': Shrink,
'win': 5
}, {
'func': Shrink,
'win': 5
}]
concatArg = {'func': Concat}
# PixlHop ++
p2 = Pixelhop2(depth=4,
TH1=0.005,
TH2=0.0005,
SaabArgs=SaabArgs,
shrinkArgs=shrinkArgs,
concatArg=concatArg).fit(train_img)
p2.construct_count()
# f = open("PixelHopUniform.pkl", 'wb') # 3 PixelHop, win: 5, TH1:0.005, TH2:0.005, CH1: 15, CH2: 20, CH3: 25, NUM_TRAIN=500
f = open("PixelHopUniform_4PH.pkl", 'wb')
pickle.dump(p2, f)
f.close()
# test_feature = p2.transform(test_images)
#
# print(train_feature[0].shape, train_feature[1].shape)
# print("------- DONE -------\n")
#
'needBias': True,
'useDC': False,
'batch': None
}, {
'num_AC_kernels': -1,
'needBias': True,
'useDC': False,
'batch': None
}]
shrinkArgsTrain = [{'func': Shrink, 'win': 5}, {'func': Shrink, 'win': 5}]
concatArg = {'func': Concat}
p1_max = Pixelhop2(depth=1,
TH1=1e-20,
TH2=1e-30,
SaabArgs=SaabArgs,
shrinkArgs=shrinkArgsTrain,
concatArg=concatArg).fit(
blocks_cover_train[kmeans.labels_ == steg_top_class])
f = open(
"/mnt/zhengwen/image_steganalysis/dataset/codes/PXH_patch_energy_MAX.pkl",
'wb') # 1st hop
pickle.dump(p1_max, f)
f.close()
features_cover_max_p1 = p1_max.transform(
blocks_cover_train[kmeans.labels_ == steg_top_class])
features_stego_max_p1 = p1_max.transform(
blocks_stego_train[kmeans.labels_ == steg_top_class])
feature_max_p1 = np.concatenate(
'needBias': True,
'useDC': False,
'batch': None
}, {
'num_AC_kernels': -1,
'needBias': True,
'useDC': False,
'batch': None
}]
shrinkArgsTrain = [{'func': Shrink, 'win': 5}, {'func': Shrink, 'win': 5}]
concatArg = {'func': Concat}
p_single = Pixelhop2(depth=1,
TH1=1e-20,
TH2=1e-30,
SaabArgs=SaabArgs,
shrinkArgs=shrinkArgsTrain,
concatArg=concatArg).fit(
single_circle[None, :, :, None].astype('double'))
p_multip = Pixelhop2(depth=1,
TH1=1e-20,
TH2=1e-30,
SaabArgs=SaabArgs,
shrinkArgs=shrinkArgsTrain,
concatArg=concatArg).fit(
multip_circle[None, :, :, None].astype('double'))
channel_idx = 20
single_BY_single_feature = np.squeeze(
p_single.transform(single_circle[None, :, :, None])[0])[:, :, channel_idx]