def predict(self,address,dp):
img = scipy.ndimage.imread(address)
print img.shape
size_1 = int(img.shape[1]*1.0/img.shape[0]*224.0)
offset = int((size_1-224.0)/2)
# print size_1,offset
img = scipy.misc.imresize(img, (224,size_1), interp='bilinear', mode=None)[:,offset:offset+224,:]
print img.shape
img = img.reshape(224**2,3).T.reshape(1,3,224,224)
img = img.astype("double")
nn.show_images(img,(1,1),unit = True)
assert nn.backend == nn.GnumpyBackend
img = nn.garray(img - dp.data_mean.reshape(1,3,256,256)[:,:,:224,:224])
if dp.mode == "vgg": img = img[:,::-1,:,:]
self.feedforward(img)
H_cpu = self.H[-1].as_numpy_array()
mask = nn.NumpyBackend.k_sparsity_mask(H_cpu,5,1)
# H_cpu *= mask
index = np.nonzero(mask)
# print index[1]
# print H_cpu[:100]
for i in xrange(len(index[1])):
print index[1][i],dp.words(index[1][i]),H_cpu[index][i]
def predict(self, address, dp):
img = scipy.ndimage.imread(address)
print img.shape
size_1 = int(img.shape[1] * 1.0 / img.shape[0] * 224.0)
offset = int((size_1 - 224.0) / 2)
# print size_1,offset
img = scipy.misc.imresize(img, (224, size_1),
interp='bilinear',
mode=None)[:, offset:offset + 224, :]
print img.shape
img = img.reshape(224**2, 3).T.reshape(1, 3, 224, 224)
img = img.astype("double")
nn.show_images(img, (1, 1), unit=True)
assert nn.backend == nn.GnumpyBackend
img = nn.garray(img -
dp.data_mean.reshape(1, 3, 256, 256)[:, :, :224, :224])
if dp.mode == "vgg": img = img[:, ::-1, :, :]
self.feedforward(img)
H_cpu = self.H[-1].as_numpy_array()
mask = nn.NumpyBackend.k_sparsity_mask(H_cpu, 5, 1)
# H_cpu *= mask
index = np.nonzero(mask)
# print index[1]
# print H_cpu[:100]
for i in xrange(len(index[1])):
print index[1][i], dp.words(index[1][i]), H_cpu[index][i]
def show_filters(self,*size):
w1,b1=self.weights[1]
w2,b2=self.weights[2]
if self.size - 1 in self.cfg.index_convolution:
# plt.figure(num=None, figsize=(30,90), dpi=80, facecolor='w', edgecolor='k')
nn.show_images(np.swapaxes(w2.as_numpy_array(),0,3)[:,:,:,:size[0]*size[1]],(size[0],size[1]),unit=True)
return
if 1 in self.cfg.index_dense:
plt.figure(num=None, figsize=(30,90), dpi=80, facecolor='w', edgecolor='k')
if type(self.cfg[0].shape)==int: nn.show_images(w1[:,:size[0]*size[1]].T,size) #MNIST dense
elif self.cfg[0].shape[0]==1: nn.show_images(w1[:,:size[0]*size[1]].T,size) #MNIST dense
elif self.cfg[0].shape[0]==3: nn.show_images(w1[:,:size[0]*size[1]].reshape(3,32,32,size[0]*size[1]),size[::-1]) #CIFAR10 dense
def visualize_conv(self,X):
w1,b1 = self.weights[1]
w2,b2 = self.weights[2]
x=X[:,:,:,:1]
self.feedforward(x)
plt.figure(num=None, figsize=(15,5), dpi=80, facecolor='w', edgecolor='k')
plt.subplot(131)
nn.show_images(self.H[2][:,:,:,:1],(1,1))
plt.subplot(132)
nn.show_images(w1[:,:,:,:16],(4,4))
plt.subplot(133)
nn.show_images(np.swapaxes(w2.as_numpy_array(),0,3)[:,:,:,:16],(4,4),unit=True)
plt.show()
plt.figure(num=None, figsize=(15,5), dpi=80, facecolor='w', edgecolor='k')
plt.subplot(131)
self.plot_train(a=2)
plt.subplot(132)
nn.show_images(np.swapaxes(self.H[1][:16,:,:,:1].as_numpy_array(),0,3),(4,4),bg="white")
plt.subplot(133)
# if self.dataset == "cifar": nn.show_images(self.H1[:,:,:,:1].sum(0).reshape(1,32,32,1),(1,1))
# elif self.dataset == "mnist": nn.show_images(self.H1[:,:,:,:1].sum(0).reshape(1,28,28,1),(1,1))
plt.show()
def visualize(self,X,u=None,s=None):
if self.H[-1].ndim ==4:
self.visualize_conv(X)
return
x = self.data_provider(X,0,1)
w1,b1=self.weights[1]
plt.figure(num=None, figsize=(15,5), dpi=80, facecolor='w', edgecolor='k')
plt.subplot(131); self.plot_train(a=2)
plt.subplot(132); self.plot_test(a=2)
plt.subplot(133)
if 1 in self.cfg.index_dense: #dense
if self.cfg[0].shape==[1, 28, 28] or self.cfg[0].shape==784:
if w1.shape[1]>25: nn.show_images(w1[:,:25].T,(5,5)) #MNIST dense
else: nn.show_images(w1[:,:].T,(5,w1.shape[1]/5)) #MNIST softmax
elif self.cfg[0].shape==[3, 32, 32]:
if w1.shape[1]>25: nn.show_images(w1[:,:25].reshape(3,32,32,25),(5,5)) #CIFAR10 dense
else: nn.show_images(w1[:,:].reshape(3,32,32,10),(5,2)) #CIFAR10 softmax
elif self.cfg[0].shape==[3, 8, 8]: #CIFAR10 dense patches
if u==None: nn.show_images(w1[:,:25].reshape(3,8,8,25),(5,5))
else: nn.show_images(whiten_undo(w1[:,:25].T.as_numpy_array(),u,s).T.reshape(3,8,8,25),(5,5),unit=True)
elif self.cfg[0].shape==[1, 8, 8]: #MNIST dense patches
nn.show_images(w1[:,:25].T.as_numpy_array().T.reshape(1,8,8,16),(4,4),unit=True)
else:
if self.cfg[0].shape[0]<4: nn.show_images(w1[:,:,:,:16],(4,4)) #convnet
plt.show()
if self.cfg[-1].activation==nn.linear: #autoencoder
if self.cfg[0].shape==784: #MNIST dense
plt.subplot(121); nn.show_images(self.H[0][:1,:],(1,1))
plt.subplot(122); nn.show_images(self.H[-1][:1,:],(1,1));
if self.cfg[0].shape==[1, 28, 28]: #MNIST dense
plt.subplot(121); nn.show_images(self.H[0][:,:,:,0].reshape(1, 28, 28,1))
plt.subplot(122); nn.show_images(self.H[-1][0].T.reshape(1,28,28,1),(1,1));
elif self.cfg[0].shape==[3, 32, 32]: #CIFAR10
plt.subplot(121); nn.show_images(self.H[0][:,:,:,0].reshape(3, 32, 32,1),(1,1))
plt.subplot(122); nn.show_images(self.H[-1][0].T.reshape(3,32,32,1),(1,1))
elif self.cfg[0].shape==[1, 8, 8]: #MNIST patches
plt.subplot(121); nn.show_images(self.H[0][:,:,:,8:9],(1,1))
plt.subplot(122); nn.show_images(self.H[-1][8].T.reshape(1,8,8,1),(1,1))
elif self.cfg[0].shape==[3, 8, 8]: #CIFAR10 patches
if u==None:
plt.subplot(121); nn.show_images(self.H[0][:,:,:,0].reshape(3, 8, 8,1),(1,1))
plt.subplot(122); nn.show_images(self.H[-1][0].T.reshape(3,8,8,1),(1,1))
else:
plt.subplot(121); nn.show_images(whiten_undo(X.reshape(192,1000000).T[0].as_numpy_array(),u,s).T.reshape(3,8,8,1),(1,1),unit=True)
self.feedforward(x.reshape(3,8,8,1))
plt.subplot(122); nn.show_images(whiten_undo(self.H[-1][0].as_numpy_array(),u,s).T.reshape(3,8,8,1),(1,1),unit=True)
plt.show()
if not(1 in self.cfg.index_dense):
self.feedforward(x)
plt.figure(num=None, figsize=(15,5), dpi=80, facecolor='w', edgecolor='k')
plt.subplot(121); nn.show_images(np.swapaxes(self.H[1][:16,:,:,:1].as_numpy_array(),0,3),(4,4),bg="white")
plt.subplot(122); nn.show_images(np.swapaxes(self.H[2][:16,:,:,:1].as_numpy_array(),0,3),(4,4),bg="white")
plt.show()
def show_filters(self,size=(10,20),scale = 1,unit = 1):
w1,b1=self.weights[1]
w2,b2=self.weights[-1]
if self.cfg.arch == "dense" and self.cfg.learning == "auto":
# plt.figure(num=None, figsize=(30,90), dpi=80, facecolor='w', edgecolor='k')
if self.cfg.dataset == "mnist":
# print w2.shape
nn.show_images(w2[:size[0]*size[1]],(size[0],size[1]),scale=scale,unit=unit)
elif self.cfg.dataset == "cifar10":
nn.show_images(w2[:size[0]*size[1]].reshape(size[0]*size[1],3,32,32),(size[0],size[1]),scale=scale,unit=unit)
elif self.cfg.dataset == "frey":
nn.show_images(w2[:size[0]*size[1]].reshape(size[0]*size[1],3,20,20),(size[0],size[1]),scale=scale,unit=unit)
elif self.cfg.dataset == "cifar10-patch": #CIFAR10 dense patches
size_ = int(np.sqrt(self.cfg[0].shape/3))
nn.show_images(w2[:size[0]*size[1]].reshape(size[0]*size[1],3,size_,size_),(size[0],size[1]),scale=scale,unit=unit)
elif self.cfg.dataset == "natural": #CIFAR10 dense patches
size = int(np.sqrt(self.cfg[0].shape))
nn.show_images(w2[:256].reshape(256,1,size,size),(16,16),unit=unit,scale=scale)
if self.cfg.learning == "auto" and self.cfg.arch == "conv":
# plt.figure(num=None, figsize=(30,90), dpi=80, facecolor='w', edgecolor='k')
# if self.cfg.dataset == "mnist":
# nn.show_images(np.swapaxes(w2.as_numpy_array(),0,1)[:size[0]*size[1],:,:,:],(size[0],size[1]),unit=unit)
# print w2.shape
if self.cfg[-1].type=="convolution":
num_filters = w2.shape[1]
if size==None:
nn.show_images(np.swapaxes(w2.as_numpy_array(),0,1)[:num_filters,:,:,:],(4,num_filters/4),scale=scale)
else:
print 'ali',np.swapaxes(w2.as_numpy_array(),0,1).shape
nn.show_images(np.swapaxes(w2.as_numpy_array(),0,1)[:size[0]*size[1],:,:,:],(size[0],size[1]),scale=scale)
if self.cfg[-1].type=="deconvolution":
# print 'ali',size
num_filters = w2.shape[0]
nn.show_images(w2.as_numpy_array()[:num_filters,:,:,:],(4,num_filters/4),unit=1,scale=scale) # elif:
# nn.show_images(np.swapaxes(w2.as_numpy_array(),0,1)[:num_filters,:,:,:],(4,num_filters/4))
if self.cfg.learning == "disc" and self.cfg.arch == "dense":
if self.cfg[1].type=="dense":
plt.figure(num=None, figsize=(30,90), dpi=80, facecolor='w', edgecolor='k')
if type(self.cfg[0].shape)==int: nn.show_images(w1[:,:size[0]*size[1]].T,size) #MNIST dense
elif self.cfg[0].shape[0]==1: nn.show_images(w1[:,:size[0]*size[1]].T,size) #MNIST dense
elif self.cfg[0].shape[0]==3: nn.show_images(w1[:,:size[0]*size[1]].reshape(3,32,32,size[0]*size[1]),size[::-1]) #CIFAR10 dense
if self.cfg.learning == "disc" and self.cfg.arch == "conv":
nn.show_images(w1[:size[0]*size[1],:,:,:],size,scale=scale,unit=unit)
plt.show()
def visualize(self,dp,visual_params):
w1,b1 = self.weights[1]
w2,b2 = self.weights[-1]
if visual_params['save']:
if self.cfg.learning == "disc":
num_filters = w1.shape[0]
nn.show_images(w1,(4,num_filters/4))
if self.cfg.arch == "dense" and self.cfg.learning == "auto":
# plt.figure(num=None, figsize=(30,90), dpi=80, facecolor='w', edgecolor='k')
size=(10,20)
if self.cfg.dataset == "mnist":
# print size, w2.shape
nn.show_images(w2[:size[0]*size[1]],(size[0],size[1]),unit=1,scale=2)
elif self.cfg.dataset in ("cifar10","svhn-ram"):
nn.show_images(w2[:size[0]*size[1]].reshape(size[0]*size[1],3,32,32),(size[0],size[1]),unit=1)
elif self.cfg.dataset == "faces":
nn.show_images(w2[:size[0]*size[1]].reshape(size[0]*size[1],1,32,32),(size[0],size[1]),unit=1)
elif self.cfg.dataset == "faces48":
nn.show_images(w2[:size[0]*size[1]].reshape(size[0]*size[1],1,48,48),(size[0],size[1]),unit=1)
elif self.cfg.dataset == "frey":
nn.show_images(w2[:size[0]*size[1]].reshape(size[0]*size[1],1,20,20),(size[0],size[1]),unit=1)
elif self.cfg.dataset == "cifar10-patch": #CIFAR10 dense patches
size = int(np.sqrt(self.cfg[0].shape/3))
nn.show_images(w2[:256].reshape(256,3,size,size),(16,16),unit=1,scale=2)
else: #CIFAR10 dense patches
size = int(np.sqrt(self.cfg[0].shape))
# print size
nn.show_images(w2[:200].reshape(200,1,size,size),(10,20),unit=1,scale=2)
# nn.show_images(w2[:64].reshape(64,1,size,size),(8,8),unit=1,scale=2)
if self.cfg.learning == "auto" and self.cfg.arch == "conv":
# print w2.as_numpy_array()[:num_filters,:,:,:].shape
num_filters = w2.shape[0]
# print w2.shape
nn.show_images(w2.as_numpy_array()[:num_filters,:,:,:],(4,num_filters/4),unit=1,scale=2)
# plt.subplot(212)
# num_filters = w1.shape[0]
# nn.show_images(w1.as_numpy_array()[:num_filters,:,:,:],(4,num_filters/4),unit=1)
# print w1.shape
# nn.show_images(np.swapaxes(w2.as_numpy_array(),0,1)[:num_filters,:,:,:],(4,num_filters/4),unit=1)
# plt.show()
plt.savefig(self.cfg.directory+self.cfg.name+".png",format="png")
plt.close()
else:
if not nn.is_interactive():
if self.cfg.learning == "auto" and not(self.cfg.dataset in ("cifar10-second","svhn-second","mnist-second")):
plt.figure(num=1, figsize=(15,10), dpi=80, facecolor='w', edgecolor='k')
else:
plt.figure(num=1, figsize=(15,5), dpi=80, facecolor='w', edgecolor='k')
# X = dp.X_id(0)
# x = nn.data_convertor(X,0,1)
w1,b1=self.weights[1]
w2,b2=self.weights[-1]
if self.cfg.arch == "dense" and self.cfg.learning == "disc": #dense
if nn.is_interactive(): plt.figure(num=1, figsize=(15,5), dpi=80, facecolor='w', edgecolor='k')
plt.figure(1)
plt.subplot(131); self.plot_train()
plt.subplot(132); self.plot_test()
plt.subplot(133)
if self.cfg.dataset == "mnist":
if w1.shape[1]>25: nn.show_images(w1[:,:25].T,(5,5)) #MNIST dense
else: nn.show_images(w1[:,:].T,(5,w1.shape[1]/5)) #MNIST softmax
elif self.cfg.dataset in ("cifar10","svhn-ram"):
# print w1.shape
if w1.shape[1]>25: nn.show_images(w1.T[:25,:].reshape(25,3,32,32),(5,5)) #CIFAR10 dense
else: nn.show_images(w1[:,:].reshape(3,32,32,10),(5,2)) #CIFAR10 softmax
elif self.cfg.dataset in ("svhn-torch"):
# print w1.shape
if w1.shape[1]>25: nn.show_images(w1.T[:25,:].reshape(25,3,32,32),(5,5),yuv=True) #CIFAR10 dense
else: nn.show_images(w1[:,:].reshape(3,32,32,10),(5,2),yuv=True) #CIFAR10 softmax
elif self.cfg.dataset == "cifar10-patches": #CIFAR10 dense patches
if u==None: nn.show_images(w1[:,:25].reshape(3,8,8,25),(5,5))
else: nn.show_images(whiten_undo(w1[:,:25].T.as_numpy_array(),u,s).T.reshape(3,8,8,25),(5,5),unit=True)
elif self.cfg.dataset == "mnist-patches": #MNIST dense patches
nn.show_images(w1[:,:25].T.as_numpy_array().T.reshape(1,8,8,16),(4,4),unit=True)
else:
channel = self.H[0].shape[1]
size = self.H[0].shape[2]
if w1.shape[1]>25: nn.show_images(w1.T[:25,:].reshape(25,channel,size,size),(5,5))
else: nn.show_images(w1[:,:].reshape(10,channel,size,size),(5,2))
if self.cfg.arch == "dense" and self.cfg.learning == "auto" and not self.cfg.dataset_extra: #dense
if nn.is_interactive(): plt.figure(num=1, figsize=(15,10), dpi=80, facecolor='w', edgecolor='k')
plt.figure(1)
plt.subplot2grid((2,3), (0,0), colspan=1); self.plot_train()
plt.subplot2grid((2,3), (0,1), colspan=2)
if self.cfg.dataset == "mnist":
nn.show_images(w2[:50],(5,10))
# nn.show_images(w2[:25],(5,5))
elif self.cfg.dataset in ("cifar10","svhn-ram"):
# print w2.shape
nn.show_images(w2[:50].reshape(50,3,32,32),(5,10))
elif self.cfg.dataset == "svhn-torch": #CIFAR10 dense patches
nn.show_images(w2[:50].reshape(50,3,32,32),(5,10),yuv=True)
elif self.cfg.dataset == "cifar10-patch": #CIFAR10 dense patches
size = int(np.sqrt(self.H[0].shape[1]/3))
# print size
nn.show_images(w2[:50].reshape(50,3,size,size),(5,10))
# if u==None: nn.show_images(w1[:,:25].reshape(3,8,8,25),(5,5))
# else: nn.show_images(whiten_undo(w1[:,:25].T.as_numpy_array(),u,s).T.reshape(3,8,8,25),(5,5),unit=True)
elif self.cfg.dataset == "mnist-patches": #MNIST dense patches
nn.show_images(w1[:,:25].T.as_numpy_array().T.reshape(1,8,8,16),(4,4),unit=True)
else: #CIFAR10 dense patches
size = int(np.sqrt(self.H[0].shape[1]))
# print w2[:50].shape,size
nn.show_images(w2[:50].reshape(50,1,size,size),(5,10))
plt.subplot2grid((2,3), (1,0), colspan=1);
if self.cfg.dataset in ("natural","mnist"):
nn.show_images(w1[:,:25].T,(5,5))
# w1,b1 = self.weights[1]
# w2,b2 = self.weights[2]
# print w1[:5,:5]
# print w2[:5,:5].T
# print "------"
plt.subplot2grid((2,3), (1,1), colspan=1);
if self.cfg.dataset == "mnist":
nn.show_images(self.H[0][0].reshape(1,1,28,28),(1,1))
elif self.cfg.dataset in ("cifar10","svhn-ram"):
nn.show_images(self.H[0][0].reshape(1,3,32,32),(1,1))
elif self.cfg.dataset == "svhn-torch":
nn.show_images(self.H[0][0].reshape(1,3,32,32),(1,1),yuv=True)
elif self.cfg.dataset == "cifar10-patch": #CIFAR10 dense patches
size = int(np.sqrt(self.H[0].shape[1]/3))
nn.show_images(self.H[0][0].reshape(1,3,size,size),(1,1))
else: #CIFAR10 dense patches
size = int(np.sqrt(self.H[0].shape[1]))
nn.show_images(self.H[0][0].reshape(1,1,size,size),(1,1))
plt.subplot2grid((2,3), (1,2), colspan=1);
if self.cfg.dataset == "mnist":
nn.show_images(self.H[-1][0].reshape(1,1,28,28),(1,1))
elif self.cfg.dataset in ("cifar10","svhn-ram"):
nn.show_images(self.H[-1][0].reshape(1,3,32,32),(1,1))
elif self.cfg.dataset == "svhn-torch":
nn.show_images(self.H[-1][0].reshape(1,3,32,32),(1,1),yuv=True)
elif self.cfg.dataset == "cifar10-patch": #CIFAR10 dense patches
size = int(np.sqrt(self.H[0].shape[1]/3))
nn.show_images(self.H[-1][0].reshape(1,3,size,size),(1,1))
else: #CIFAR10 dense patches
size = int(np.sqrt(self.H[0].shape[1]))
nn.show_images(self.H[-1][0].reshape(1,1,size,size),(1,1))
if self.cfg.arch == "conv" and self.cfg.learning == "disc":
if nn.is_interactive(): plt.figure(num=1, figsize=(15,5), dpi=80, facecolor='w', edgecolor='k')
plt.figure(1)
plt.subplot(131); self.plot_train()
plt.subplot(132); self.plot_test()
plt.subplot(133)
nn.show_images(w1[:16,:,:,:],(4,4))
if self.cfg.arch == "conv" and self.cfg.learning == "auto":
if nn.is_interactive(): plt.figure(num=1, figsize=(15,10), dpi=80, facecolor='w', edgecolor='k')
plt.figure(1)
# w2,b2 = self.weights[-1]
# x=X[:,:,:,:1]
# self.feedforward(x)
if self.cfg.dataset in ("cifar10-second","svhn-second","mnist-second"): #CIFAR10
plt.subplot(131);
# print self.H[0].shape,self.H[-1].shape,self.H[-1].max()
nn.show_images(np.swapaxes(self.H[0][:1,:16,:,:].as_numpy_array(),0,1),(4,4),bg="white")
plt.subplot(132);
nn.show_images(np.swapaxes(self.H[-1][:1,:16,:,:].as_numpy_array(),0,1),(4,4),bg="white")
plt.subplot(133);
nn.show_images(np.swapaxes(self.H[-2][:1,:16,:,:].as_numpy_array(),0,1),(4,4),bg="white")
# print self.H[-1]
else:
plt.subplot(231);
nn.show_images(self.H[0][0,:,:,:].reshape(1,self.H[0].shape[1],self.H[0].shape[2],self.H[0].shape[3]),(1,1));
plt.subplot(232);
nn.show_images(self.H[-1][0,:,:,:].reshape(1,self.H[-1].shape[1],self.H[-1].shape[2],self.H[-1].shape[3]),(1,1));
plt.subplot(233);
self.plot_train()
plt.subplot(234)
# if self.H[1].shape[1]>=16:
# H1 = self.H[1].as_numpy_array()
# H1 = H1.reshape(16*100,28*28)
# print np.nonzero(H1)[0].shape
nn.show_images(np.swapaxes(self.H[-2][:1,:16,:,:].as_numpy_array(),0,1),(4,4),bg="white")
# else:
# nn.show_images(np.swapaxes(self.H[1][:1,:8,:,:].as_numpy_array(),0,1),(2,4),bg="white")
plt.subplot(235)
# if w1.shape[0]>16:
nn.show_images(w1[:16,:,:,:],(4,4))
# else:
# nn.show_images(w1[:8,:,:,:],(2,4))
plt.subplot(236)
# if w2.shape[0]>=16:
# print w2.shape
# if self.cfg.dataset == "svhn-torch":
# nn.show_images(np.swapaxes(w2.as_numpy_array(),0,1)[:16,:,:,:],(4,4),unit=1,yuv=1)
if self.cfg[-1].type=="convolution":
nn.show_images(np.swapaxes(w2.as_numpy_array(),0,1)[:16,:,:,:],(4,4),unit=1)
if self.cfg[-1].type=="deconvolution":
nn.show_images(w2[:16,:,:,:],(4,4),unit=1)
# else:
# nn.show_images(np.swapaxes(w2.as_numpy_array(),0,1)[:,:8,:,:],(2,4),unit=True)
if nn.is_interactive(): plt.show()
else: plt.draw(); plt.pause(.01)
if self.cfg.dataset_extra=="generate": #dense
for k in self.cfg.index_dense:
if self.cfg[k].l2_activity!=None: index = k
if nn.is_interactive(): plt.figure(num=1, figsize=(15,10), dpi=80, facecolor='w', edgecolor='k')
plt.figure(1)
plt.subplot2grid((2,3), (0,0), colspan=1); self.plot_train()
plt.subplot2grid((2,3), (0,1), colspan=2)
if self.cfg.dataset == "mnist":
nn.show_images(w2[:50],(5,10))
plt.subplot2grid((2,3), (1,1), colspan=1);
# if self.cfg.dataset == "mnist":
# print self.H[index].shape
x = self.H[index][:,0].as_numpy_array()
y = self.H[index][:,1].as_numpy_array()
plt.plot(x,y,'bo')
# plt.grid()
x = self.T_sort[:,0].as_numpy_array()
y = self.T_sort[:,1].as_numpy_array()
plt.plot(x,y,'ro')
# plt.grid()
# x = self.test_rand[:,0].as_numpy_array()
# y = self.test_rand[:,1].as_numpy_array()
# plt.plot(x,y,'go')
plt.grid()
# nn.show_images(self.H[0][0].reshape(1,1,28,28),(1,1))
plt.subplot2grid((2,3), (1,2), colspan=1);
if self.cfg.dataset == "mnist":
nn.show_images(self.H[-1][0].reshape(1,1,28,28),(1,1))
if self.cfg.dataset == "mnist" and self.cfg.dataset_extra in ("vae","generate"):
plt.figure(num=5, figsize=(15,10), dpi=80, facecolor='w', edgecolor='k')
temp = nn.randn((64,784))
self.test_mode = True
self.feedforward(temp)
self.test_mode = False
nn.show_images(self.H[-1].reshape(64,1,28,28),(8,8))
if visual_params['save']:
plt.savefig(self.cfg.directory+self.cfg.name+"_samples.png",format="png")
else:
plt.draw(); plt.pause(.01)
def show_filters(self, size=(10, 20), scale=1, unit=1):
w1, b1 = self.weights[1]
w2, b2 = self.weights[-1]
if self.cfg.arch == "dense" and self.cfg.learning == "auto":
# plt.figure(num=None, figsize=(30,90), dpi=80, facecolor='w', edgecolor='k')
if self.cfg.dataset == "mnist":
# print w2.shape
nn.show_images(w2[:size[0] * size[1]], (size[0], size[1]),
scale=scale,
unit=unit)
elif self.cfg.dataset == "cifar10":
nn.show_images(w2[:size[0] * size[1]].reshape(
size[0] * size[1], 3, 32, 32), (size[0], size[1]),
scale=scale,
unit=unit)
elif self.cfg.dataset == "frey":
nn.show_images(w2[:size[0] * size[1]].reshape(
size[0] * size[1], 3, 20, 20), (size[0], size[1]),
scale=scale,
unit=unit)
elif self.cfg.dataset == "cifar10-patch": #CIFAR10 dense patches
size_ = int(np.sqrt(self.cfg[0].shape / 3))
nn.show_images(w2[:size[0] * size[1]].reshape(
size[0] * size[1], 3, size_, size_), (size[0], size[1]),
scale=scale,
unit=unit)
elif self.cfg.dataset == "natural": #CIFAR10 dense patches
size = int(np.sqrt(self.cfg[0].shape))
nn.show_images(w2[:256].reshape(256, 1, size, size), (16, 16),
unit=unit,
scale=scale)
if self.cfg.learning == "auto" and self.cfg.arch == "conv":
# plt.figure(num=None, figsize=(30,90), dpi=80, facecolor='w', edgecolor='k')
# if self.cfg.dataset == "mnist":
# nn.show_images(np.swapaxes(w2.as_numpy_array(),0,1)[:size[0]*size[1],:,:,:],(size[0],size[1]),unit=unit)
# print w2.shape
if self.cfg[-1].type == "convolution":
num_filters = w2.shape[1]
if size == None:
nn.show_images(np.swapaxes(w2.as_numpy_array(), 0,
1)[:num_filters, :, :, :],
(4, num_filters / 4),
scale=scale)
else:
print 'ali', np.swapaxes(w2.as_numpy_array(), 0, 1).shape
nn.show_images(np.swapaxes(w2.as_numpy_array(), 0,
1)[:size[0] * size[1], :, :, :],
(size[0], size[1]),
scale=scale)
if self.cfg[-1].type == "deconvolution":
# print 'ali',size
num_filters = w2.shape[0]
nn.show_images(w2.as_numpy_array()[:num_filters, :, :, :],
(4, num_filters / 4),
unit=1,
scale=scale) # elif:
# nn.show_images(np.swapaxes(w2.as_numpy_array(),0,1)[:num_filters,:,:,:],(4,num_filters/4))
if self.cfg.learning == "disc" and self.cfg.arch == "dense":
if self.cfg[1].type == "dense":
plt.figure(num=None,
figsize=(30, 90),
dpi=80,
facecolor='w',
edgecolor='k')
if type(self.cfg[0].shape) == int:
nn.show_images(w1[:, :size[0] * size[1]].T,
size) #MNIST dense
elif self.cfg[0].shape[0] == 1:
nn.show_images(w1[:, :size[0] * size[1]].T,
size) #MNIST dense
elif self.cfg[0].shape[0] == 3:
nn.show_images(w1[:, :size[0] * size[1]].reshape(
3, 32, 32, size[0] * size[1]),
size[::-1]) #CIFAR10 dense
if self.cfg.learning == "disc" and self.cfg.arch == "conv":
nn.show_images(w1[:size[0] * size[1], :, :, :],
size,
scale=scale,
unit=unit)
plt.show()
def visualize(self, dp, visual_params):
w1, b1 = self.weights[1]
w2, b2 = self.weights[-1]
if visual_params['save']:
if self.cfg.learning == "disc":
num_filters = w1.shape[0]
nn.show_images(w1, (4, num_filters / 4))
if self.cfg.arch == "dense" and self.cfg.learning == "auto":
# plt.figure(num=None, figsize=(30,90), dpi=80, facecolor='w', edgecolor='k')
size = (10, 20)
if self.cfg.dataset == "mnist":
# print size, w2.shape
nn.show_images(w2[:size[0] * size[1]], (size[0], size[1]),
unit=1,
scale=2)
elif self.cfg.dataset in ("cifar10", "svhn-ram"):
nn.show_images(w2[:size[0] * size[1]].reshape(
size[0] * size[1], 3, 32, 32), (size[0], size[1]),
unit=1)
elif self.cfg.dataset == "faces":
nn.show_images(w2[:size[0] * size[1]].reshape(
size[0] * size[1], 1, 32, 32), (size[0], size[1]),
unit=1)
elif self.cfg.dataset == "faces48":
nn.show_images(w2[:size[0] * size[1]].reshape(
size[0] * size[1], 1, 48, 48), (size[0], size[1]),
unit=1)
elif self.cfg.dataset == "frey":
nn.show_images(w2[:size[0] * size[1]].reshape(
size[0] * size[1], 1, 20, 20), (size[0], size[1]),
unit=1)
elif self.cfg.dataset == "cifar10-patch": #CIFAR10 dense patches
size = int(np.sqrt(self.cfg[0].shape / 3))
nn.show_images(w2[:256].reshape(256, 3, size, size),
(16, 16),
unit=1,
scale=2)
else: #CIFAR10 dense patches
size = int(np.sqrt(self.cfg[0].shape))
# print size
nn.show_images(w2[:200].reshape(200, 1, size, size),
(10, 20),
unit=1,
scale=2)
# nn.show_images(w2[:64].reshape(64,1,size,size),(8,8),unit=1,scale=2)
if self.cfg.learning == "auto" and self.cfg.arch == "conv":
# print w2.as_numpy_array()[:num_filters,:,:,:].shape
num_filters = w2.shape[0]
# print w2.shape
nn.show_images(w2.as_numpy_array()[:num_filters, :, :, :],
(4, num_filters / 4),
unit=1,
scale=2)
# plt.subplot(212)
# num_filters = w1.shape[0]
# nn.show_images(w1.as_numpy_array()[:num_filters,:,:,:],(4,num_filters/4),unit=1)
# print w1.shape
# nn.show_images(np.swapaxes(w2.as_numpy_array(),0,1)[:num_filters,:,:,:],(4,num_filters/4),unit=1)
# plt.show()
plt.savefig(self.cfg.directory + self.cfg.name + ".png",
format="png")
plt.close()
else:
if not nn.is_interactive():
if self.cfg.learning == "auto" and not (self.cfg.dataset in (
"cifar10-second", "svhn-second", "mnist-second")):
plt.figure(num=1,
figsize=(15, 10),
dpi=80,
facecolor='w',
edgecolor='k')
else:
plt.figure(num=1,
figsize=(15, 5),
dpi=80,
facecolor='w',
edgecolor='k')
# X = dp.X_id(0)
# x = nn.data_convertor(X,0,1)
w1, b1 = self.weights[1]
w2, b2 = self.weights[-1]
if self.cfg.arch == "dense" and self.cfg.learning == "disc": #dense
if nn.is_interactive():
plt.figure(num=1,
figsize=(15, 5),
dpi=80,
facecolor='w',
edgecolor='k')
plt.figure(1)
plt.subplot(131)
self.plot_train()
plt.subplot(132)
self.plot_test()
plt.subplot(133)
if self.cfg.dataset == "mnist":
if w1.shape[1] > 25:
nn.show_images(w1[:, :25].T, (5, 5)) #MNIST dense
else:
nn.show_images(w1[:, :].T,
(5, w1.shape[1] / 5)) #MNIST softmax
elif self.cfg.dataset in ("cifar10", "svhn-ram"):
# print w1.shape
if w1.shape[1] > 25:
nn.show_images(w1.T[:25, :].reshape(25, 3, 32, 32),
(5, 5)) #CIFAR10 dense
else:
nn.show_images(w1[:, :].reshape(3, 32, 32, 10),
(5, 2)) #CIFAR10 softmax
elif self.cfg.dataset in ("svhn-torch"):
# print w1.shape
if w1.shape[1] > 25:
nn.show_images(w1.T[:25, :].reshape(25, 3, 32, 32),
(5, 5),
yuv=True) #CIFAR10 dense
else:
nn.show_images(w1[:, :].reshape(3, 32, 32, 10), (5, 2),
yuv=True) #CIFAR10 softmax
elif self.cfg.dataset == "cifar10-patches": #CIFAR10 dense patches
if u == None:
nn.show_images(w1[:, :25].reshape(3, 8, 8, 25), (5, 5))
else:
nn.show_images(whiten_undo(
w1[:, :25].T.as_numpy_array(), u,
s).T.reshape(3, 8, 8, 25), (5, 5),
unit=True)
elif self.cfg.dataset == "mnist-patches": #MNIST dense patches
nn.show_images(w1[:, :25].T.as_numpy_array().T.reshape(
1, 8, 8, 16), (4, 4),
unit=True)
else:
channel = self.H[0].shape[1]
size = self.H[0].shape[2]
if w1.shape[1] > 25:
nn.show_images(
w1.T[:25, :].reshape(25, channel, size, size),
(5, 5))
else:
nn.show_images(
w1[:, :].reshape(10, channel, size, size), (5, 2))
if self.cfg.arch == "dense" and self.cfg.learning == "auto" and not self.cfg.dataset_extra: #dense
if nn.is_interactive():
plt.figure(num=1,
figsize=(15, 10),
dpi=80,
facecolor='w',
edgecolor='k')
plt.figure(1)
plt.subplot2grid((2, 3), (0, 0), colspan=1)
self.plot_train()
plt.subplot2grid((2, 3), (0, 1), colspan=2)
if self.cfg.dataset == "mnist":
nn.show_images(w2[:50], (5, 10))
# nn.show_images(w2[:25],(5,5))
elif self.cfg.dataset in ("cifar10", "svhn-ram"):
# print w2.shape
nn.show_images(w2[:50].reshape(50, 3, 32, 32), (5, 10))
elif self.cfg.dataset == "svhn-torch": #CIFAR10 dense patches
nn.show_images(w2[:50].reshape(50, 3, 32, 32), (5, 10),
yuv=True)
elif self.cfg.dataset == "cifar10-patch": #CIFAR10 dense patches
size = int(np.sqrt(self.H[0].shape[1] / 3))
# print size
nn.show_images(w2[:50].reshape(50, 3, size, size), (5, 10))
# if u==None: nn.show_images(w1[:,:25].reshape(3,8,8,25),(5,5))
# else: nn.show_images(whiten_undo(w1[:,:25].T.as_numpy_array(),u,s).T.reshape(3,8,8,25),(5,5),unit=True)
elif self.cfg.dataset == "mnist-patches": #MNIST dense patches
nn.show_images(w1[:, :25].T.as_numpy_array().T.reshape(
1, 8, 8, 16), (4, 4),
unit=True)
else: #CIFAR10 dense patches
size = int(np.sqrt(self.H[0].shape[1]))
# print w2[:50].shape,size
nn.show_images(w2[:50].reshape(50, 1, size, size), (5, 10))
plt.subplot2grid((2, 3), (1, 0), colspan=1)
if self.cfg.dataset in ("natural", "mnist"):
nn.show_images(w1[:, :25].T, (5, 5))
# w1,b1 = self.weights[1]
# w2,b2 = self.weights[2]
# print w1[:5,:5]
# print w2[:5,:5].T
# print "------"
plt.subplot2grid((2, 3), (1, 1), colspan=1)
if self.cfg.dataset == "mnist":
nn.show_images(self.H[0][0].reshape(1, 1, 28, 28), (1, 1))
elif self.cfg.dataset in ("cifar10", "svhn-ram"):
nn.show_images(self.H[0][0].reshape(1, 3, 32, 32), (1, 1))
elif self.cfg.dataset == "svhn-torch":
nn.show_images(self.H[0][0].reshape(1, 3, 32, 32), (1, 1),
yuv=True)
elif self.cfg.dataset == "cifar10-patch": #CIFAR10 dense patches
size = int(np.sqrt(self.H[0].shape[1] / 3))
nn.show_images(self.H[0][0].reshape(1, 3, size, size),
(1, 1))
else: #CIFAR10 dense patches
size = int(np.sqrt(self.H[0].shape[1]))
nn.show_images(self.H[0][0].reshape(1, 1, size, size),
(1, 1))
plt.subplot2grid((2, 3), (1, 2), colspan=1)
if self.cfg.dataset == "mnist":
nn.show_images(self.H[-1][0].reshape(1, 1, 28, 28), (1, 1))
elif self.cfg.dataset in ("cifar10", "svhn-ram"):
nn.show_images(self.H[-1][0].reshape(1, 3, 32, 32), (1, 1))
elif self.cfg.dataset == "svhn-torch":
nn.show_images(self.H[-1][0].reshape(1, 3, 32, 32), (1, 1),
yuv=True)
elif self.cfg.dataset == "cifar10-patch": #CIFAR10 dense patches
size = int(np.sqrt(self.H[0].shape[1] / 3))
nn.show_images(self.H[-1][0].reshape(1, 3, size, size),
(1, 1))
else: #CIFAR10 dense patches
size = int(np.sqrt(self.H[0].shape[1]))
nn.show_images(self.H[-1][0].reshape(1, 1, size, size),
(1, 1))
if self.cfg.arch == "conv" and self.cfg.learning == "disc":
if nn.is_interactive():
plt.figure(num=1,
figsize=(15, 5),
dpi=80,
facecolor='w',
edgecolor='k')
plt.figure(1)
plt.subplot(131)
self.plot_train()
plt.subplot(132)
self.plot_test()
plt.subplot(133)
nn.show_images(w1[:16, :, :, :], (4, 4))
if self.cfg.arch == "conv" and self.cfg.learning == "auto":
if nn.is_interactive():
plt.figure(num=1,
figsize=(15, 10),
dpi=80,
facecolor='w',
edgecolor='k')
plt.figure(1)
# w2,b2 = self.weights[-1]
# x=X[:,:,:,:1]
# self.feedforward(x)
if self.cfg.dataset in ("cifar10-second", "svhn-second",
"mnist-second"): #CIFAR10
plt.subplot(131)
# print self.H[0].shape,self.H[-1].shape,self.H[-1].max()
nn.show_images(np.swapaxes(
self.H[0][:1, :16, :, :].as_numpy_array(), 0, 1),
(4, 4),
bg="white")
plt.subplot(132)
nn.show_images(np.swapaxes(
self.H[-1][:1, :16, :, :].as_numpy_array(), 0, 1),
(4, 4),
bg="white")
plt.subplot(133)
nn.show_images(np.swapaxes(
self.H[-2][:1, :16, :, :].as_numpy_array(), 0, 1),
(4, 4),
bg="white")
# print self.H[-1]
else:
plt.subplot(231)
nn.show_images(
self.H[0][0, :, :, :].reshape(1, self.H[0].shape[1],
self.H[0].shape[2],
self.H[0].shape[3]),
(1, 1))
plt.subplot(232)
nn.show_images(
self.H[-1][0, :, :, :].reshape(1, self.H[-1].shape[1],
self.H[-1].shape[2],
self.H[-1].shape[3]),
(1, 1))
plt.subplot(233)
self.plot_train()
plt.subplot(234)
# if self.H[1].shape[1]>=16:
# H1 = self.H[1].as_numpy_array()
# H1 = H1.reshape(16*100,28*28)
# print np.nonzero(H1)[0].shape
nn.show_images(np.swapaxes(
self.H[-2][:1, :16, :, :].as_numpy_array(), 0, 1),
(4, 4),
bg="white")
# else:
# nn.show_images(np.swapaxes(self.H[1][:1,:8,:,:].as_numpy_array(),0,1),(2,4),bg="white")
plt.subplot(235)
# if w1.shape[0]>16:
nn.show_images(w1[:16, :, :, :], (4, 4))
# else:
# nn.show_images(w1[:8,:,:,:],(2,4))
plt.subplot(236)
# if w2.shape[0]>=16:
# print w2.shape
# if self.cfg.dataset == "svhn-torch":
# nn.show_images(np.swapaxes(w2.as_numpy_array(),0,1)[:16,:,:,:],(4,4),unit=1,yuv=1)
if self.cfg[-1].type == "convolution":
nn.show_images(np.swapaxes(w2.as_numpy_array(), 0,
1)[:16, :, :, :], (4, 4),
unit=1)
if self.cfg[-1].type == "deconvolution":
nn.show_images(w2[:16, :, :, :], (4, 4), unit=1)
# else:
# nn.show_images(np.swapaxes(w2.as_numpy_array(),0,1)[:,:8,:,:],(2,4),unit=True)
if nn.is_interactive(): plt.show()
else:
plt.draw()
plt.pause(.01)
if self.cfg.dataset_extra == "generate": #dense
for k in self.cfg.index_dense:
if self.cfg[k].l2_activity != None: index = k
if nn.is_interactive():
plt.figure(num=1,
figsize=(15, 10),
dpi=80,
facecolor='w',
edgecolor='k')
plt.figure(1)
plt.subplot2grid((2, 3), (0, 0), colspan=1)
self.plot_train()
plt.subplot2grid((2, 3), (0, 1), colspan=2)
if self.cfg.dataset == "mnist":
nn.show_images(w2[:50], (5, 10))
plt.subplot2grid((2, 3), (1, 1), colspan=1)
# if self.cfg.dataset == "mnist":
# print self.H[index].shape
x = self.H[index][:, 0].as_numpy_array()
y = self.H[index][:, 1].as_numpy_array()
plt.plot(x, y, 'bo')
# plt.grid()
x = self.T_sort[:, 0].as_numpy_array()
y = self.T_sort[:, 1].as_numpy_array()
plt.plot(x, y, 'ro')
# plt.grid()
# x = self.test_rand[:,0].as_numpy_array()
# y = self.test_rand[:,1].as_numpy_array()
# plt.plot(x,y,'go')
plt.grid()
# nn.show_images(self.H[0][0].reshape(1,1,28,28),(1,1))
plt.subplot2grid((2, 3), (1, 2), colspan=1)
if self.cfg.dataset == "mnist":
nn.show_images(self.H[-1][0].reshape(1, 1, 28, 28), (1, 1))
if self.cfg.dataset == "mnist" and self.cfg.dataset_extra in (
"vae", "generate"):
plt.figure(num=5,
figsize=(15, 10),
dpi=80,
facecolor='w',
edgecolor='k')
temp = nn.randn((64, 784))
self.test_mode = True
self.feedforward(temp)
self.test_mode = False
nn.show_images(self.H[-1].reshape(64, 1, 28, 28), (8, 8))
if visual_params['save']:
plt.savefig(self.cfg.directory + self.cfg.name +
"_samples.png",
format="png")
else:
plt.draw()
plt.pause(.01)