load_from = save_to
logposterior = lambda aa: model.logposterior_func2(x=x, z=aa)
print('optimiznig local', samp_i)
z = optimize_local_expressive_only_sample(logposterior,
model,
x,
save_to=save_to,
load_from=load_from)
z = z.view(-1, z_size)
z = z.data.cpu().numpy()
# print (z)
plot_kde(ax, samps=z, xlimits=xlimits, ylimits=ylimits, cmap='Reds')
# #Plot prob
# col +=1
# ax = plt.subplot2grid((rows,cols), (samp_i,col), frameon=False)
# Ws, logpW, logqW = model.sample_W() #_ , [1], [1]
# func = lambda zs: log_bernoulli(model.decode(Ws, Variable(torch.unsqueeze(zs,1))), Variable(torch.unsqueeze(samp,0)))+ Variable(torch.unsqueeze(lognormal4(torch.Tensor(zs), torch.zeros(2), torch.zeros(2)), 1))
# plot_isocontours2_exp(ax, func, cmap='Greens', legend=legend)
# if samp_i==0: ax.annotate('p(z,x|W2)', xytext=(.1, 1.1), xy=(0, 1), textcoords='axes fraction')
# func = lambda zs: lognormal4(torch.Tensor(zs), torch.zeros(2), torch.zeros(2))
# plot_isocontours(ax, func, cmap='Blues', alpha=.3)
# func = lambda zs: lognormal4(torch.Tensor(zs), torch.squeeze(mean.data), torch.squeeze(logvar.data))
# plot_isocontours(ax, func, cmap='Reds')
# #Plot prob
# col +=1
x = x.view(1,784)
save_to = this_dir+'/local_params'+str(samp_i)+'.pt'
load_from = save_to
logposterior = lambda aa: model.logposterior_func2(x=x,z=aa)
print ('optimiznig local', samp_i)
z = optimize_local_expressive_only_sample(logposterior, model, x, save_to=save_to, load_from=load_from)
z = z.view(-1,z_size)
z = z.data.cpu().numpy()
# print (z)
plot_kde(ax,samps=z,xlimits=xlimits,ylimits=ylimits,cmap='Reds')
# #Plot prob
# col +=1
# ax = plt.subplot2grid((rows,cols), (samp_i,col), frameon=False)
# Ws, logpW, logqW = model.sample_W() #_ , [1], [1]
# func = lambda zs: log_bernoulli(model.decode(Ws, Variable(torch.unsqueeze(zs,1))), Variable(torch.unsqueeze(samp,0)))+ Variable(torch.unsqueeze(lognormal4(torch.Tensor(zs), torch.zeros(2), torch.zeros(2)), 1))
# plot_isocontours2_exp(ax, func, cmap='Greens', legend=legend)
print '\nPosterior', p_i, posterior_names[p_i]
posterior = ttp.posterior_class(posteriors[p_i])
ax = plt.subplot2grid((rows,columns), (p_i,0), frameon=False)#, colspan=3)
plot_isocontours(ax, posterior.run_log_post, cmap='Blues')
if p_i == 0: ax.annotate('Posterior', xytext=(.3, 1.1), xy=(0, 1), textcoords='axes fraction')
for q_i in range(len(models)):
print model_names[q_i]
ax = plt.subplot2grid((rows,columns), (p_i,q_i+1), frameon=False)#, colspan=3)
model = models[q_i](posteriors[p_i])
# model.train(10000, save_to=home+'/Documents/tmp/vars.ckpt')
model.train(9999000, save_to='')
samps = model.sample(1000)
plot_kde(ax, samps, cmap='Reds')
plot_isocontours(ax, posterior.run_log_post, cmap='Blues', alpha=alpha)
if p_i == 0: ax.annotate(model_names[q_i], xytext=(.38, 1.1), xy=(0, 1), textcoords='axes fraction')
# plt.show()
plt.savefig(home+'/Documents/tmp/plots.png')
print 'saved'
load_from = save_to
logposterior = lambda aa: model.logposterior_func2(x=x, z=aa)
print('optimiznig local flow', samp_i)
# z = optimize_local_expressive_only_sample(logposterior, model, x, save_to=save_to, load_from=load_from)
# z = optimize_local_expressive_only_sample_2(logposterior, model, x)
z = optimize_local_expressive_only_sample(logposterior, model, x)
z = z.view(-1, z_size)
z = z.data.cpu().numpy()
# plot_kde(ax,samps=z,xlimits=xlimits,ylimits=ylimits,cmap='Reds')
plot_kde(ax,
samps=z,
xlimits=xlimits,
ylimits=ylimits,
cmap='Blues')
# flow = optimize_local_flow1(logposterior, model, x)
# func = lambda zs: flow.logprob(torch.Tensor(zs).cuda())
# _ = plot_isocontours_new(ax, func, cmap='Blues',xlimits=xlimits,ylimits=ylimits, legend=False, levels=[])
# #Plot prob
# col +=1
# ax = plt.subplot2grid((rows,cols), (samp_i,col), frameon=False)
# Ws, logpW, logqW = model.sample_W() #_ , [1], [1]
# func = lambda zs: log_bernoulli(model.decode(Ws, Variable(torch.unsqueeze(zs,1))), Variable(torch.unsqueeze(samp,0)))+ Variable(torch.unsqueeze(lognormal4(torch.Tensor(zs), torch.zeros(2), torch.zeros(2)), 1))
# plot_isocontours2_exp(ax, func, cmap='Greens', legend=legend)
# if samp_i==0: ax.annotate('p(z,x|W2)', xytext=(.1, 1.1), xy=(0, 1), textcoords='axes fraction')
# func = lambda zs: lognormal4(torch.Tensor(zs), torch.zeros(2), torch.zeros(2))
load_from = home + '/Documents/tmp/' + posterior_names[
p_i] + '_' + model_names[q_i] + '.ckpt'
else:
load_from = ''
print model_names[q_i]
ax = plt.subplot2grid((rows, columns), (q_i + 1, p_i + 1),
frameon=False) #, colspan=3)
model = models[q_i](posteriors[p_i])
if train_:
model.train(100000, save_to=save_to, load_from=load_from)
model.load_params(load_from)
samps = model.sample(100)
plot_kde(ax, samps, cmap='Blues')
plot_isocontours(ax,
posterior.run_log_post,
cmap='Greys',
alpha=alpha)
# if p_i == 0: ax.annotate(model_names[q_i], xytext=(.38, 1.1), xy=(0, 1), textcoords='axes fraction')
# plt.show()
plt.savefig(home + '/Documents/tmp/plots.png')
print 'saved'
plt.savefig(home + '/Documents/tmp/plots.eps')
print 'saved'
plt.savefig(home + '/Documents/tmp/plots.pdf')
print 'saved'
# if __name__ == '__main__':