def log_like(var_par, draw, value, k):
l = int(len(var_par) / 2)
mu, cov = var_par[:l], np.exp(var_par[l:])
samples = draw * cov + mu
pi = softmax(samples[:k])
d = len(value)
thetas = softmax(samples[k:].reshape([k, d]), axis=1)
n = np.sum(value)
logps = np.log(pi) + np.dot(
np.log(thetas), value) + gammaln(n + 1) - np.sum(gammaln(value + 1))
return logsumexp(logps)
def logprior(var_par, draw, k):
# Log of prior probabilities
d = int((len(var_par) / 2 - k) / k)
l = int(len(var_par) / 2)
alpha = np.ones(k) # Prior for mixture probabilities
beta = np.ones(d) # Prior for multinomials
mu, cov = var_par[:l], np.exp(var_par[l:])
samples = draw * cov + mu
pi = softmax(samples[:k])
thetas = softmax(samples[k:].reshape([k, d]), axis=1)
return dirichlet.logpdf(pi, alpha) + np.sum(
[dirichlet.logpdf(theta, beta) + ldt(theta)
for theta in thetas]) + ldt(pi)
def log_like(var_par, draw, value, k): d = len(value) value = value[:-1] l = int(len(var_par)/2) mu, cov = var_par[:l], np.exp(var_par[l:]) samples = draw*cov + mu pi = softmax(samples[:k]) lam = np.exp(samples[k:2*k]) thetas = softmax(samples[2*k:].reshape([k,d]), axis=1) thetas1 = thetas[:,:-1] thetas2 = thetas[:,-1] logps = np.log(pi) + np.dot(np.log(thetas1.T*lam).T, value) + lam*(thetas2-1) - np.sum(gammaln(value+1)) return logsumexp(logps)
def np_log_like(var_par, draw, data, k): d = data.shape[1] gln_test_values = np.sum(gammaln(data[:,:-1]+1), axis=1) data = data[:,:-1] l = int(len(var_par)/2) mu, cov = var_par[:l], np.exp(var_par[l:]) samples = draw*cov + mu pi = softmax(samples[:k]) lam = np.exp(samples[k:2*k]) thetas = softmax(samples[2*k:].reshape([k,d]), axis=1) thetas1 = thetas[:,:-1] thetas2 = thetas[:,-1] logps = (np.log(pi) + np.dot(data, np.log(thetas1.T*lam)) + lam*(thetas2-1)).T - gln_test_values return np.sum(logsumexp(logps, axis=0))
def pred_like(test_data, var_par, n_samples, k):
# Method for prediction likelihood
N_test, d = test_data.shape
l = int(len(var_par) / 2)
gln_test_n = gammaln(test_data.sum(axis=1) + 1)
gln_test_values = np.sum(gammaln(test_data + 1), axis=1)
mu, cov = var_par[:l], np.exp(var_par[l:])
like_matrix = np.empty([N_test, n_samples])
samples = draw_samples(var_par, n_samples)
for s, sample in enumerate(samples):
pi = softmax(sample[:k])
thetas = softmax(sample[k:].reshape([k, d]), axis=1)
logps = (np.log(pi) + np.dot(test_data, np.log(
thetas.T))).T + gln_test_n - gln_test_values
like_matrix[:, s] = logsumexp2(np.stack(logps)[:, :N_test], axis=0)
return np.mean(logsumexp2(like_matrix, axis=1) - np.log(n_samples))
def get_frame_info(self, frame):
actor = frame['policy']
distribution = [np.around(softmax(head), decimals=3) for head in actor]
logits = [np.around(head, decimals=3) for head in actor]
value = np.around(frame['value'], decimals=3)
value_info = "reward={}, manipulated_reward={}, value={}\n".format(frame['reward'], frame['manipulated_reward'], value)
actor_info = "logits={}, distribution={}\n".format(logits, distribution)
action_info = "action={}\n".format(frame['action'])
extra_info = "extra={}\n".format(frame['extra'])
frame_info = { "log": value_info + actor_info + action_info + extra_info }
if flags.save_episode_screen and frame['screen'] is not None:
frame_info["screen"] = frame['screen']
return frame_info