def fine_tune_faces_unites():
# constants used in fine-tuning
training_steps = 200
num_batch = 100
lr = 0.01
# read the pretrained weights
weight_vh1 = np.loadtxt('./pretrain/yale_faces/Weight_matrix_1.txt')
weight_vh2 = np.loadtxt('./pretrain/yale_faces/Weight_matrix_2.txt')
weight_vh3 = np.loadtxt('./pretrain/yale_faces/Weight_matrix_3.txt')
weight_vh4 = np.loadtxt('./pretrain/yale_faces/Weight_matrix_4.txt')
# construct the fine-tuning input weights
w_list_4_layer = [weight_vh1, weight_vh2, weight_vh3, weight_vh4]
# read the pretrained bias
# hidden bias
bias1 = np.array([np.loadtxt('./pretrain/yale_faces/Hidden_bias_1.txt')
]).transpose()
bias2 = np.array([np.loadtxt('./pretrain/yale_faces/Hidden_bias_2.txt')
]).transpose()
bias3 = np.array([np.loadtxt('./pretrain/yale_faces/Hidden_bias_3.txt')
]).transpose()
bias4 = np.array([np.loadtxt('./pretrain/yale_faces/Hidden_bias_4.txt')
]).transpose()
# visual bias
bias5 = np.array([np.loadtxt('./pretrain/yale_faces/Visible_bias_4.txt')
]).transpose()
bias6 = np.array([np.loadtxt('./pretrain/yale_faces/Visible_bias_3.txt')
]).transpose()
bias7 = np.array([np.loadtxt('./pretrain/yale_faces/Visible_bias_2.txt')
]).transpose()
bias8 = np.array([np.loadtxt('./pretrain/yale_faces/Visible_bias_1.txt')
]).transpose()
# construct the fine-tuning input bias
b_list_4_layer = [bias1, bias2, bias3, bias4, bias5, bias6, bias7, bias8]
#N = np.array([2016,1400,896,504,224])
data_train, data_test = face_read(2000)
data_train_finetune = np.transpose(data_train)
# Do the fine-tuning
total_cost_record_4_lay, w_list_output_4_lay, b_list_output_4_lay = fine_tuning(
w_list_4_layer, b_list_4_layer, data_train_finetune, num_batch,
training_steps, lr)
# write the weight matrix
for i in np.arange(4):
filename = 'W_matrix_' + str(i + 1) + '_Final.txt'
np.savetxt(filename, w_list_output_4_lay[i])
# write the bias
for j in np.arange(8):
filename = 'B_bias_' + str(j + 1) + '_Final.txt'
np.savetxt(filename, b_list_output_4_lay[j])
# write the total cost
np.savetxt('Total_Cost_Final.txt', total_cost_record_4_lay)
np.transpose(ini_value_w2), np.transpose(ini_value_w1)] w_list_4_layer = [ini_value_w1, ini_value_w2, ini_value_w3, ini_value_w4, \ np.transpose(ini_value_w4),np.transpose(ini_value_w3),np.transpose(ini_value_w2),np.transpose(ini_value_w1)] ini_value_b1 = rng.randn(2000,1) ini_value_b2 = rng.randn(1000,1) ini_value_b3 = rng.randn(500,1) ini_value_b4 = rng.randn(30,1) ini_value_b5 = rng.randn(500,1) ini_value_b6 = rng.randn(1000,1) ini_value_b7 = rng.randn(2000,1) ini_value_b8 = rng.randn(784,1) b_list_3_layer = [ini_value_b1, ini_value_b2, ini_value_b3, ini_value_b6, ini_value_b7, ini_value_b8] b_list_4_layer = [ini_value_b1, ini_value_b2, ini_value_b3, ini_value_b4, ini_value_b5, ini_value_b6, ini_value_b7, ini_value_b8] simulate_data = rng.randn(784,simulate_data_num) # 50 data points #total_cost_record_3_lay, w_list_output_3_lay, b_list_output_3_lay = fine_tuning(w_list_3_layer, b_list_3_layer, simulate_data, num_batch, training_steps, lr) total_cost_record_4_lay, w_list_output_4_lay, b_list_output_4_lay = fine_tuning(w_list_4_layer, b_list_4_layer, simulate_data, num_batch, training_steps, lr) end_time = time.time() print(end_time-start_time) fig, ax = plt.subplots() #ax.scatter(range(len(total_cost_record_3_lay)),total_cost_record_3_lay) ax.scatter(range(len(total_cost_record_4_lay)),total_cost_record_4_lay) plt.show()
def PNN_test_HSMS_ratio_3(
I_MS_LR, I_PAN, inputImg, param, net, path, mode,
epochs): #iniz è uguale a 0 QUI ERA IL PROBLEMA FACEVA UN OVERRIDE
test_dir_out = path['test_dir_out']
FTnetwork_dir_out = path['ftnetwork_dir_out']
# import pdb; pdb.set_trace()
# san paolo hs/ms
param['L'] = 15 # modifica del 23/04
# param['ratio']=3
param['ratio'] = 3 # modificato il 04/04
param[
'lr'] = 0.0001 # prec 10e-4 #modificato il 25/04
param['patchSize'] = 33 #preced = 33
# param['padSize'] = 8
# per il debug
# import pdb; pdb.set_trace()
if 'inputType' not in param.keys():
param['inputType'] = 'MS_PAN'
#fine tuning
if epochs != 0:
# pdb.set_trace()
fine_tuning(I_MS_LR, I_PAN, param, epochs, FTnetwork_dir_out)
ft_model_path = FTnetwork_dir_out + '/PNN_model.mat'
FT_model = sio.loadmat(ft_model_path, squeeze_me=True) #provo25/04
from PNN_testing_model import Network, ConvLayer
layer = []
for j in range(0, len(FT_model['layers']), 2):
layer.append(
ConvLayer(FT_model['layers'][j], FT_model['layers'][j + 1]))
net = Network(layer)
# layer=[]
# for j in range(0,len(param['layers']),2):
# layer.append(ConvLayer(param['layers'][j], param['layers'][j+1]))
# net=Network(layer)
# pdb.set_trace()
if mode != 'full':
I_MS_LR, I_PAN = downgrade_images(I_MS_LR, I_PAN, param['ratio'],
param['sensor'])
I_PAN = np.expand_dims(I_PAN, axis=0)
NDxI_LR = []
mav_value = 2**(np.float32(param['L']))
# compute radiometric indexes
if param['inputType'] == 'MS_PAN_NDxI':
if I_MS_LR.shape[0] == 8:
NDxI_LR = np.stack(((I_MS_LR[4, :, :] - I_MS_LR[7, :, :]) /
(I_MS_LR[4, :, :] + I_MS_LR[7, :, :]),
(I_MS_LR[0, :, :] - I_MS_LR[7, :, :]) /
(I_MS_LR[0, :, :] + I_MS_LR[7, :, :]),
(I_MS_LR[2, :, :] - I_MS_LR[3, :, :]) /
(I_MS_LR[2, :, :] + I_MS_LR[3, :, :]),
(I_MS_LR[5, :, :] - I_MS_LR[0, :, :]) /
(I_MS_LR[5, :, :] + I_MS_LR[0, :, :])),
axis=0)
else:
NDxI_LR = np.stack(((I_MS_LR[3, :, :] - I_MS_LR[2, :, :]) /
(I_MS_LR[3, :, :] + I_MS_LR[2, :, :]),
(I_MS_LR[1, :, :] - I_MS_LR[3, :, :]) /
(I_MS_LR[1, :, :] + I_MS_LR[3, :, :])),
axis=0)
# %input preparation
if param['typeInterp'] == 'interp23tap':
# import pdb; pdb.set_trace()
I_MS = interp23(I_MS_LR, param['ratio'])
# if len(NDxI_LR)!=0:
# NDxI = interp23(NDxI_LR, param['ratio'])
# else:
# sys.exit('interpolation not supported')
if param['inputType'] == 'MS':
I_in = I_MS.astype('single') / mav_value
elif param['inputType'] == 'MS_PAN':
# import pdb; pdb.set_trace()
I_in = np.vstack((I_MS, I_PAN)).astype('single') / mav_value
elif param['inputType'] == 'MS_PAN_NDxI':
I_in = np.vstack((I_MS, I_PAN)).astype('single') / mav_value
I_in = np.vstack((I_in, NDxI)).astype('single')
else:
sys.exit('Configuration not supported')
print(I_in.shape)
# import pdb; pdb.set_trace()
I_in_residual = np.expand_dims(I_in, axis=0)
I_in_residual = I_in_residual[:, :I_MS.shape[0], :, :]
import pdb
pdb.set_trace()
I_in = np.pad(I_in, ((0, 0), (param['padSize'] / 2, param['padSize'] / 2),
(param['padSize'] / 2, param['padSize'] / 2)),
mode='edge')
I_in = np.expand_dims(I_in, axis=0)
#Pansharpening MODIFICA 17/04
# param['residual'] = 'false'
# I_MS_residual = np.expand_dims(I_MS,axis=0)
if param['residual'] == 'true' or param['residual'] == 1:
I_out = net.build(I_in) + I_in_residual[:, :I_MS.shape[0], :, :]
# I_out = net.build(I_in) + I_MS_residual
else:
I_out = net.build(I_in)
I_out = I_out * mav_value
return np.squeeze(I_out)
np.loadtxt('/home/zhshang/DeepEncoder/mnist/Visible_bias_2.txt')
]).transpose()
bias8 = np.array([
np.loadtxt('/home/zhshang/DeepEncoder/mnist/Visible_bias_1.txt')
]).transpose()
# construct the fine-tuning input bias
b_list_4_layer = [bias1, bias2, bias3, bias4, bias5, bias6, bias7, bias8]
#N = np.array([784,1000,500,250,2])
data_train, label_train, data_test, label_test = mnist_read(2000)
data_train_finetune = np.transpose(data_train)
# Do the fine-tuning
total_cost_record_4_lay, w_list_output_4_lay, b_list_output_4_lay = fine_tuning(
w_list_4_layer, b_list_4_layer, data_train_finetune, num_batch,
training_steps, lr)
# write the weight matrix
for i in np.arange(4):
filename = 'W_matrix_' + str(i + 1) + '_Final.txt'
np.savetxt(filename, w_list_output_4_lay[i])
# write the bias
for j in np.arange(8):
filename = 'B_bias_' + str(j + 1) + '_Final.txt'
np.savetxt(filename, b_list_output_4_lay[j])
# write the total cost
np.savetxt('Total_Cost_Final.txt', total_cost_record_4_lay)
def PNN_test(I_MS_LR, I_PAN, inputImg, param, net, path, mode, epochs=0):
test_dir_out = path['test_dir_out']
FTnetwork_dir_out = path['ftnetwork_dir_out']
param['L'] = inputImg['L']
param['ratio'] = inputImg['ratio']
if 'inputType' not in param.keys():
param['inputType'] = 'MS_PAN'
#fine tuning
if epochs != 0:
fine_tuning(I_MS_LR, I_PAN, param, epochs, FTnetwork_dir_out)
ft_model_path = FTnetwork_dir_out + '/PNN_model.mat'
FT_model = sio.loadmat(ft_model_path, squeeze_me=True)
from PNN_testing_model import Network, ConvLayer
layer = []
for j in range(0, len(FT_model['layers']), 2):
layer.append(
ConvLayer(FT_model['layers'][j], FT_model['layers'][j + 1]))
net = Network(layer)
if mode != 'full':
I_MS_LR, I_PAN = downgrade_images(I_MS_LR, I_PAN, param['ratio'],
param['sensor'])
I_PAN = np.expand_dims(I_PAN, axis=0)
NDxI_LR = []
mav_value = 2**(np.float32(param['L']))
# compute radiometric indexes
if param['inputType'] == 'MS_PAN_NDxI':
if I_MS_LR.shape[0] == 8:
NDxI_LR = np.stack(((I_MS_LR[4, :, :] - I_MS_LR[7, :, :]) /
(I_MS_LR[4, :, :] + I_MS_LR[7, :, :]),
(I_MS_LR[0, :, :] - I_MS_LR[7, :, :]) /
(I_MS_LR[0, :, :] + I_MS_LR[7, :, :]),
(I_MS_LR[2, :, :] - I_MS_LR[3, :, :]) /
(I_MS_LR[2, :, :] + I_MS_LR[3, :, :]),
(I_MS_LR[5, :, :] - I_MS_LR[0, :, :]) /
(I_MS_LR[5, :, :] + I_MS_LR[0, :, :])),
axis=0)
else:
NDxI_LR = np.stack(((I_MS_LR[3, :, :] - I_MS_LR[2, :, :]) /
(I_MS_LR[3, :, :] + I_MS_LR[2, :, :]),
(I_MS_LR[1, :, :] - I_MS_LR[3, :, :]) /
(I_MS_LR[1, :, :] + I_MS_LR[3, :, :])),
axis=0)
#input preparation
if param['typeInterp'] == 'interp23tap':
I_MS = interp23(I_MS_LR, param['ratio'])
if len(NDxI_LR) != 0:
NDxI = interp23(NDxI_LR, param['ratio'])
else:
sys.exit('interpolation not supported')
if param['inputType'] == 'MS':
I_in = I_MS.astype('single') / mav_value
elif param['inputType'] == 'MS_PAN':
I_in = np.vstack((I_MS, I_PAN)).astype('single') / mav_value
elif param['inputType'] == 'MS_PAN_NDxI':
I_in = np.vstack((I_MS, I_PAN)).astype('single') / mav_value
I_in = np.vstack((I_in, NDxI)).astype('single')
else:
sys.exit('Configuration not supported')
print I_in.shape
I_in_residual = np.expand_dims(I_in, axis=0)
I_in_residual = I_in_residual[:, :I_MS.shape[0], :, :]
I_in = np.pad(I_in, ((0, 0), (param['padSize'] / 2, param['padSize'] / 2),
(param['padSize'] / 2, param['padSize'] / 2)),
mode='edge')
I_in = np.expand_dims(I_in, axis=0)
#Pansharpening
if param['residual']:
I_out = net.build(I_in) + I_in_residual[:, :I_MS.shape[0], :, :]
else:
I_out = net.build(I_in)
I_out = I_out * mav_value
return np.squeeze(I_out)
para_name += [
'FREQUENCY_USER', 'FREQUENCY_ITEM', 'FREQUENCY', 'KEEP_PORB',
'SAMPLE_RATE', 'GRAPH_CONV', 'PREDICTION', 'LOSS_FUNCTION',
'GENERALIZATION', 'OPTIMIZATION', 'IF_TRASFORMATION', 'ACTIVATION',
'POOLING'
]
if all_para[2] == 'SGNN': para_name += ['PROP_DIM', 'PROP_EMB', 'IF_NORM']
# if testing the model, we need to read in test set
if tuning_method == 'test': all_para[11] = para[11] = 'Test'
## read data
data = read_all_data(all_para)
para[10] = data[-1]
## tuning the model
os.environ["CUDA_VISIBLE_DEVICES"] = all_para[0]
if tuning_method == 'tuning':
tuning(path_excel_dir, para_name, para, data, lr_coarse, lamda_coarse,
min_num_coarse, max_num_coarse, min_num_fine, max_num_fine)
if tuning_method == 'fine_tuning':
fine_tuning(path_excel_dir, para_name, para, data, lr_fine, lamda_fine,
min_num_fine, max_num_fine)
if tuning_method == 'cross_tuning':
cross_tuning(path_excel_dir, para_name, para, data, lr_fine,
lamda_fine, min_num_fine, max_num_fine)
if tuning_method == 'coarse_tuning':
coarse_tuning(path_excel_dir, para_name, para, data, lr_coarse,
lamda_coarse, min_num_coarse, max_num_coarse)
if tuning_method == 'test':
test(path_excel_dir, para_name, para, data, iter_num_test)
