# Discover the magic number
net = convVAE(dim_z,x_train,x_test)
get_magic = net.get_flattened(net.x_train[:2,:,:,:])
# actual nework
#net1 = convVAE(dim_z,x_train,x_test,diff = dif,magic = get_magic.shape[1])
net = convVAE(dim_z,x_train,x_test,diff=None,magic = get_magic.shape[1])
# load the parameters from the model:
params = pickle_loader(model_dir+"model.pkl")
for p in params:
net.params.set_value(p.get_value())
# Writing paths
path_array,original_array = sound_write.path_write(net,sample_rate,duration = 3.,data_points = 10)
wav.write("sound/"+experiment_name+"_paths_rewrite.wav",sample_rate,path_array)
wav.write("sound/"+experiment_name+"_paths_original_rewrite.wav",sample_rate,original_array)
# Reconstruction sounds and random latent configurations
random,reconstruction,original = sound_write.sample_write(net,sample_rate,duration = 10)
wav.write("sound/"+experiment_name+"_random_rewrite.wav",sample_rate,random)
wav.write("sound/"+experiment_name+"_reconstruction_rewrite.wav",sample_rate,reconstruction)
wav.write("sound/"+experiment_name+"_reconstruction_original_rewrite.wav",sample_rate,original)
# Weights
plot_params(net.params,weight_plot_dir)
#ou1 = net1.output(x_train[:50])
idxx = np.random.randint(0,x_train.shape[0],50)
# Reconstruction images
ou2 = net.output(x_train[idxx])
for i in range(ou2.shape[0]):
if i <100:
print "magic_value", net.magic
iterations = 100
disc = 1.01
for i in range(iterations):
net1.iterate()
net2.iterate()
rows = 1
columns = 3
print "ITERATION", i
print net2.performance['train'][-1]
net1.dropout_prob.set_value(np.float32(0.0))
net2.dropout_prob.set_value(np.float32(0.0))
path_array, original_array = sound_write.path_write(net2,
880,
duration=0.2,
data_points=10)
wav.write("sound/" + experiment_name + "_paths.wav", sample_rate,
path_array)
wav.write("sound/" + experiment_name + "_paths_original.wav", sample_rate,
original_array)
random, reconstruction, original = sound_write.sample_write(net1,
880,
duration=10)
wav.write("sound/" + experiment_name + "_random.wav", sample_rate, random)
wav.write("sound/" + experiment_name + "_reconstruction.wav", sample_rate,
reconstruction)
wav.write("sound/" + experiment_name + "_reconstruction_original.wav",
sample_rate, original)
print net.performance['train'][-1]
###### Shifting experiments ####
num = 100
h_val = np.zeros([num,net.magic]).astype(np.float32)
for i in range(num):
h_val[i,i]=1
out = net.generate_h(h_val)
plt.plot(out.T)
plt.show()
pdb.set_trace()
############## ALL THE BOOKEEPING #########################
# Writing paths
path_array,original_array = sound_write.path_write(net,880,duration = 0.2,data_points = 10)
wav.write("sound/"+experiment_name+"_paths.wav",sample_rate,path_array)
wav.write("sound/"+experiment_name+"_paths_original.wav",sample_rate,original_array)
# Reconstruction sounds and random latent configurations
random,reconstruction,original = sound_write.sample_write(net,880,duration = 10)
wav.write("sound/"+experiment_name+"_random.wav",sample_rate,random)
wav.write("sound/"+experiment_name+"_reconstruction.wav",sample_rate,reconstruction)
wav.write("sound/"+experiment_name+"_reconstruction_original.wav",sample_rate,original)
# Weights
plot_params(net.params,weight_plot_dir)
#ou1 = net1.output(x_train[:50])
# Reconstruction images
ou2 = net.output(x_train[:50])
for i in range(ou2.shape[0]):
if i <100:
