x_dist=x_dist,
batchnorm=batchnorm,
mc_samples=mc_samps,
l2_reg=l2_reg,
learning_paradigm=learning_paradigm,
name=model_token,
ckpt=model_token,
loss_ratio=loss_ratio)
#
if model_name == 'gm_dgm':
model = gm_dgm(n_x,
n_y,
n_z,
n_hidden,
x_dist=x_dist,
batchnorm=batchnorm,
mc_samples=mc_samps,
l2_reg=l2_reg,
learning_paradigm=learning_paradigm,
name=model_token,
ckpt=model_token,
loss_ratio=loss_ratio)
#
if model_name == 'gmm_vae':
model = gmm_vae(n_x,
n_y,
n_w,
n_z,
n_hidden,
x_dist=x_dist,
batchnorm=batchnorm,
mc_samples=mc_samps,
verbose = 3
Data.reset_counters()
results=[]
for i in range(num_runs):
print("Starting work on run: {}".format(i))
Data.reset_counters()
np.random.seed(2)
tf.set_random_seed(2)
tf.reset_default_graph()
model_token = token+'-'+str(i)+'---'
if model_name == 'm2':
model = m2(n_x, n_y, n_z, n_hidden, x_dist=x_dist, batchnorm=batchnorm, mc_samples=mc_samps, l2_reg=l2_reg, learning_paradigm=learning_paradigm, name=model_token, ckpt = model_token)
if model_name == 'gm_dgm':
model = gm_dgm(n_x, n_y, n_z, n_hidden, x_dist=x_dist, batchnorm=batchnorm, alpha=alpha, mc_samples=mc_samps, l2_reg=l2_reg, learning_paradigm=learning_paradigm, name=model_token, ckpt = model_token, prior=prior[0:n_y]/float(sum(prior[0:n_y])), loss_ratio=loss_ratio, output_dir=output_dir)
if learning_paradigm == 'semisupervised' or 'semi-unsupervised':
model.loss = model.compute_loss()
elif learning_paradigm == 'unsupervised':
model.loss = model.compute_unsupervised_loss()
elif model.learning_paradigm == 'supervised':
model.loss = model.compute_supervised_loss()
model.train(Data, n_epochs, l_bs, u_bs, lr, eval_samps=eval_samps, binarize=binarize, verbose=1)
results.append(model.curve_array)
np.save(os.path.join(output_dir,'curve_'+token+'_'+str(i)+'.npy'), model.curve_array)
y_pred_test = predict_new(Data.data['x_test'])[0]
conf_mat = confusion_matrix(Data.data['y_test'].argmax(1), y_pred_test.argmax(1))
np.save(os.path.join(output_dir,'conf_mat_'+token+'_'+str(i)+'.npy'), conf_mat)
np.savez(os.path.join(output_dir,'y_preds_labels_'+token+'_'+str(i)+'.npz'), y_true=Data.data['y_test'].argmax(1), y_pred=y_pred_test.argmax(1), y_labels = y_test[1])
n_z,
x_dist=x_dist,
mc_samples=mc_samps,
alpha=alpha,
l2_reg=l2_reg,
learning_paradigm=learning_paradigm,
name=model_token,
ckpt=model_token,
output_dir=output_dir)
if model_name == 'gm_dgm':
model = gm_dgm(n_x,
n_y,
n_z,
x_dist=x_dist,
mc_samples=mc_samps,
alpha=alpha,
l2_reg=l2_reg,
learning_paradigm=learning_paradigm,
name=model_token,
ckpt=model_token,
prior=prior[0:n_y] / float(sum(prior[0:n_y])),
output_dir=output_dir)
if learning_paradigm == 'semisupervised' or 'semi-unsupervised':
model.loss = model.compute_loss()
elif learning_paradigm == 'unsupervised':
model.loss = model.compute_unsupervised_loss()
elif model.learning_paradigm == 'supervised':
model.loss = model.compute_supervised_loss()
model.train(Data,
n_epochs,