def predict(self, Xt, yt, ids, results):
tic = timeit.default_timer()
y_dist = pd.DataFrame(self.model.predict(Xt),index=yt.index.values)
toc = timeit.default_timer()
pred_time = toc-tic
y_samples = np.apply_along_axis(mdn.sample_from_output, 1, y_dist, self.n_out, self.n_mixes,temp=1.0)
y_hat = pd.DataFrame(y_samples[:,0,:],index=yt.index.values)
# split mixture params
mus, sigs, pi_logits = mdn.split_mixture_params(y_dist, self.n_out, self.n_mixes)
pis = mdn.softmax(pi_logits, t=1)
muc = [list(t) for t in zip(*[iter(mus)]*self.n_mixes)]
sic = [list(t) for t in zip(*[iter(sigs)]*self.n_mixes)]
# store results
for k in y_hat.columns:
results[ids[k]]['ytest'].append(yt.iloc[:,k])
results[ids[k]]['yhat'].append(y_hat.iloc[:,k])
results[ids[k]]['alphas'].append(pis)
results[ids[k]]['mus'].append(muc[k])
results[ids[k]]['sigs'].append(sic[k])
results['pred_time'].append(pred_time)
results['fit_time'].append(self.fit_time)
return y_hat, results
def get_mu_sigma_MDN(params, output_dim, temp, sigma_temp, mixes):
mus, sigs, pi_logits = mdn.split_mixture_params(params, output_dim, mixes)
pis = mdn.softmax(pi_logits, t=temp)
m = mdn.sample_from_categorical(pis)
# Alternative way to sample from categorical:
# m = np.random.choice(range(len(pis)), p=pis)
mus_vector = mus[m * output_dim:(m + 1) * output_dim]
sig_vector = sigs[m * output_dim:(m + 1) *
output_dim] * sigma_temp # adjust for temperature
return mus_vector, sig_vector
def sample_from_one_specific_mixture(mdn, mixture_number, params, output_dim,
total_num_mixes, sigma_temp):
#To make dreams only from 1 of the trained mixtures - helps see what's inside there.
mus, sigs, pi_logits = mdn.split_mixture_params(params, output_dim,
total_num_mixes)
mus_vector = mus[mixture_number * output_dim:(mixture_number + 1) *
output_dim]
sig_vector = sigs[mixture_number * output_dim:(mixture_number + 1) *
output_dim] * sigma_temp # adjust for temperature
cov_matrix = np.identity(output_dim) * sig_vector
sample = np.random.multivariate_normal(mus_vector, cov_matrix, 1)
return sample
def _mixture_components(self, predictions, num_plots=1):
"""
Statistics of mixture components
"""
stats = []
for i in range(len(predictions)):
mus, sigs, pis = mdn.split_mixture_params(
np.squeeze(np.array(predictions))[i], self.OUTPUT_DIMS,
self.n_mixes)
stats.append(np.zeros(12))
stats[i][0] = np.mean(sigs)
stats[i][1] = np.std(sigs)
stats[i][2] = np.max(sigs)
stats[i][3] = np.min(sigs)
stats[i][4] = np.mean(mus)
stats[i][5] = np.std(mus)
stats[i][6] = np.max(mus)
stats[i][7] = np.min(mus)
stats[i][8] = np.mean(pis)
stats[i][9] = np.std(pis)
stats[i][10] = np.max(pis)
stats[i][11] = np.min(pis)
stats = np.array(stats)