## convert this back to a list of arrs
X_final = [X_batch[t] for t in range(T)]
## perform downsampling
feed_dict_input = {i: d for i, d in zip(sequence, X_final)} #size has to be fixed...
feed_dict_input[label_placeholder] = y_batch
test_loss, test_prediction = sess.run([loss, outputs], feed_dict=feed_dict_input)
# print(test_loss)
# test_prediction = sess.run(outputs, feed_dict=feed_dict_input)
test_prediction = test_prediction > 0.5
print('saving figures...')
for k in range(mini_batch_size):
# print(iou.IoU_3D(y[k], test_prediction[k]))
test_iou.append(iou.IoU_3D(y_batch[k], test_prediction[k]))
fig = plt.figure(figsize = (20,8))
ax = fig.add_subplot(1, 2, 1, projection='3d')
ax.voxels(test_prediction[k], edgecolor='k')
ax.set(xlabel='x', ylabel='y', zlabel='z', title='prediction')
ax = fig.add_subplot(1, 2, 2, projection='3d')
ax.voxels(y_batch[k], edgecolor='k', facecolor='blue')
ax.set(xlabel='x', ylabel='y', zlabel='z', title='original model')
plt.savefig('./results/figures/'+ 'test_sample_' + str(k) + '_minibatch_' + str(i) + '.jpg')
plt.close()
print('figures saved!')
X_final = [X_batch[t] for t in t_sample]
feed_dict_input = {i: d for i, d in zip(sequence, X_final)}
feed_dict_input[label_placeholder] = y_batch
#print(feed_dict_input.keys())
#manually insert the label_placeholder
loss_epoch, prediction, _ = sess.run([loss, outputs, optimizer],
feed_dict=feed_dict_input)
prediction = prediction > 0.5
accuracy = np.mean(prediction == y_batch)
training_accuracy.append(accuracy)
epoch_training_iou = []
for j in range(mini_batch_size):
epoch_training_iou.append(iou.IoU_3D(y_batch[j], prediction[j]))
training_iou.append(np.mean(epoch_training_iou))
loss_history.append(loss_epoch)
toc = time.time()
if epoch % print_every == 0:
print('time elapsed for this epoch:', toc - tic)
print('mean iou:', np.mean(epoch_training_iou))
print('loss:', loss_epoch)
print('accuracy:', accuracy)
if (epoch % save_every == 0
and epoch > 0): #save it occassionally in case we stop the run
pickle.dump([loss_history, training_iou, training_accuracy],
open('gray_scale_planes_training_data.p', 'wb'))
X_final = [X_batch[t, :, :, :, :] for t in t_sample]
feed_dict_input = {i: d
for i, d in zip(sequence, X_final)}
feed_dict_input[label_placeholder] = y_batch
#print(feed_dict_input.keys())
#manually insert the label_placeholder
loss_epoch, prediction, _ = sess.run([loss, outputs, optimizer],
feed_dict=feed_dict_input)
prediction = prediction > 0.5
training_accuracy.append(np.mean(prediction == y_batch))
epoch_training_iou = []
for j in range(mini_batch_size):
epoch_training_iou.append(
iou.IoU_3D(y_batch[j, :, :, :], prediction[j, :, :, :]))
training_iou.append(np.mean(epoch_training_iou))
training_iou.append(np.mean(epoch_training_iou))
print(np.mean(training_iou))
print('epoch: ' + str(epoch) + ' loss: ' + str(loss_epoch))
prediction = np.squeeze(prediction)
print(np.mean(prediction == y_batch))
loss_history.append(loss_epoch)
if (epoch % 20 == 0
and epoch > 1): #save it occassionally in case we stop the run
pickle.dump([loss_history, training_iou, training_accuracy],
open('gray_scale_planes_training_data.p', 'wb'))
saver.save(sess,
'./grayscale_plane_R2N2_model_weights',
global_step=epoch)
#manually insert the label_placeholder
sess.run(optimizer, feed_dict=feed_dict_input)
loss_epoch = sess.run(loss, feed_dict=feed_dict_input)
#reconfigure feed_dict to accept a place_holder for y
prediction = sess.run(predictions, feed_dict=feed_dict_input)
print('epoch: ' + str(epoch) + ' loss: ' + str(loss_epoch))
print(prediction.shape)
print(np.mean(prediction == y_batch_flat))
loss_history.append(loss_epoch)
training_accuracy.append(np.mean(prediction == y_batch_flat))
interim_iou = 0
for k in range(mini_batch_size):
interim_iou += (iou.IoU_3D(y[k, :, :, :], prediction[k, :, :, :]))
training_iou.append(interim_iou / mini_batch_size)
print(training_iou)
#predictions = tf.argmax(outputs, axis = 4)
## run test batch
f = open(os.path.join(data_dir, test_batch), 'rb')
batch_sample_2 = pickle.load(f)
X, y = pickle_to_data.unravel_batch_pickle(batch_sample_2)
X_nparr = np.array(X)
batch_size, T, H, W, C = X_nparr.shape
num_batches = int(128 / mini_batch_size)
test_iou = list()
print('run test case')
for i in range(num_batches - 1):