def run_test(model_arch_module, dataset_dir, model_path, result_dir, top_k=1):
import data_config as cfg
os.makedirs(result_dir, exist_ok=True)
model = model_arch_module.build_model_arch()
_, _, test_data = data_reader.read_data_set_dir(dataset_dir, cfg.one_hot,
64)
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, model_path)
overall_truth_indices = []
overall_pred_probs = []
overall_correct_classification = 0.
for step in range(test_data.batch_count):
images, one_hot = test_data.next_batch()
truth_indexes = np.argmax(one_hot, 1)
pred = model.predict(sess, images)[0]
pred_indexes = np.argmax(pred, 1)
if top_k == 1:
correct_classification = (
truth_indexes == pred_indexes).astype(np.float32)
else:
correct_classification = tf.nn.in_top_k(pred,
truth_indexes,
k=top_k)
correct_classification = sess.run(
correct_classification).astype(np.float32)
overall_truth_indices.extend(truth_indexes)
overall_pred_probs.extend(pred)
overall_correct_classification = overall_correct_classification + correct_classification.sum(
)
print("Step %d, accuracy %f" %
(step, correct_classification.mean()))
overall_acc = overall_correct_classification / float(
test_data.data_set_size)
print("Overall accuracy: %f" % overall_acc)
with open(os.path.join(result_dir, 'truth_indices.pkl'), 'wb') as f:
pickle.dump(overall_truth_indices, f)
with open(os.path.join(result_dir, 'pred_probs.pkl'), 'wb') as f:
pickle.dump(overall_pred_probs, f)
def run_test_visual(model_arch_module, dataset_dir, model_path):
import data_visualizer as dv
import data_config as cfg
model = model_arch_module.build_model_arch()
_, _, test_data = data_reader.read_data_set_dir(dataset_dir, cfg.one_hot,
24)
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, model_path)
labels = cfg.one_hot_labels
for step in range(test_data.batch_count):
images, one_hot = test_data.next_batch()
truth_indexes = np.argmax(one_hot, 1)
pred = model.predict(sess, images)[0]
pred_indexes = np.argmax(pred, 1)
pred_labels = [labels[i] for i in pred_indexes]
dv.show_images_with_truth(images, pred_labels, truth_indexes,
pred_indexes)
def run_trainer(model_arch_module, num_epochs, batch_size, dataset_path,
model_name, run_name):
import tensorflow as tf
import data_reader
import data_config as cfg
model = model_arch_module.build_model_arch()
train_data, val_data, _ = data_reader.read_data_set_dir(
dataset_path, cfg.one_hot, batch_size)
accuracies_input, losses_input, train_mean, val_mean = get_mean_op()
global_step = 0
file_writer = tf.summary.FileWriter('logs/%s/' % run_name)
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
file_writer.add_graph(sess.graph)
try:
for epoch in range(num_epochs):
train_accuracies = []
train_losses = []
for step in range(train_data.batch_count):
batch_images, batch_labels = train_data.next_batch()
if step % 10 == 0:
train_accuracy, loss, summary = model.train_step(
sess, batch_images, batch_labels, run_summary=True)
file_writer.add_summary(summary, global_step)
print(
'Epoch %d, step %d, global step %d, training accuracy: %f, training loss %f'
% (epoch, step, global_step, train_accuracy, loss))
else:
train_accuracy, loss = model.train_step(
sess,
batch_images,
batch_labels,
run_summary=False)
train_accuracies.append(train_accuracy)
train_losses.append(loss)
global_step += 1
mean_acc, mean_loss, summ = sess.run(train_mean,
feed_dict={
accuracies_input:
train_accuracies,
losses_input:
train_losses
})
print('Epoch %d: Training accuracy: %f, loss %f' %
(epoch, mean_acc, mean_loss))
file_writer.add_summary(summ, epoch)
val_accuracies = []
val_losses = []
for step in range(val_data.batch_count):
batch_images, batch_labels = val_data.next_batch()
val_accuracy, loss, summary = model.evaluate(
sess, batch_images, batch_labels)
val_accuracies.append(val_accuracy)
val_losses.append(loss)
mean_acc, mean_loss, summ = sess.run(val_mean,
feed_dict={
accuracies_input:
val_accuracies,
losses_input:
val_losses
})
print('Epoch %d: Validation accuracy: %f, loss %f' %
(epoch, mean_acc, mean_loss))
file_writer.add_summary(summ, epoch)
saver.save(sess,
save_path='./model/%s' % model_name,
global_step=epoch)
except KeyboardInterrupt:
print('Training cancelled intentionally.')
print('Stop training at %d steps' % global_step)
file_writer.close()