+ str(len(str(len(test_dataloader)))) \
+ '} out of ' \
+ str(len(test_dataloader)) \
+ '...'
for k in range(num_chains):
test_pred_probs = np.empty([len(test_dataloader), num_classes])
nums_dropped_samples = np.empty([len(test_dataloader), num_classes],
dtype=np.int64)
for i, (x, _) in enumerate(test_dataloader):
print(verbose_msg.format(k + 1, i + 1))
for j in range(num_classes):
y = torch.zeros([1, num_classes], dtype=dtype)
y[0, j] = 1.
integral, num_dropped_samples = model.predictive_posterior(
chain_lists.vals['sample'][k], x, y)
test_pred_probs[i, j] = integral.item()
nums_dropped_samples[i, j] = num_dropped_samples
np.savetxt(
sampler_output_run_paths[k].joinpath('pred_posterior_on_test.csv'),
test_pred_probs,
delimiter=',')
np.savetxt(sampler_output_run_paths[k].joinpath(
'pred_posterior_on_test_num_dropped_samples.csv'),
nums_dropped_samples,
fmt='%d',
delimiter=',')
keys=['sample'],
dtype=dtype)
# %% Drop burn-in samples
for i in range(num_chains):
chain_lists.vals['sample'][i] = chain_lists.vals['sample'][i][
pred_iter_thres:]
# %% Compute chain means
means = chain_lists.mean()
# %% Make and save predictions
for k in range(num_chains):
test_pred_probs = np.empty([len(test_dataloader), num_classes])
for i, (x, _) in enumerate(test_dataloader):
for j in range(num_classes):
y = torch.zeros([1, num_classes], dtype=dtype)
y[0, j] = 1.
integral, _ = model.predictive_posterior([means[k, :]], x, y)
test_pred_probs[i, j] = integral.item()
test_preds = np.argmax(test_pred_probs, axis=1)
np.savetxt(sampler_output_run_paths[k].joinpath('preds_via_mean.txt'),
test_preds,
fmt='%d')