best_epoch_losses = [1000, 1000, 1000, 1000, 1000]
with tf.Session(config=tf.ConfigProto(log_device_placement=False)) as sess:
# initialize all variables/parameters:
init = tf.global_variables_initializer()
sess.run(init)
for epoch in range(no_of_epochs):
print("epoch: %d/%d" % (epoch + 1, no_of_epochs))
# run an epoch and get all batch losses:
batch_losses = []
for step, (imgs, onehot_labels) in enumerate(tqdm(train_data_iterator())):
# create a feed dict containing the batch data:
batch_feed_dict = model.create_feed_dict(imgs_batch=imgs,
early_drop_prob=0.01, late_drop_prob=0.1,
onehot_labels_batch=onehot_labels)
# compute the batch loss and compute & apply all gradients w.r.t to
# the batch loss (without model.train_op in the call, the network
# would NOT train, we would only compute the batch loss):
batch_loss, _ = sess.run([model.loss, model.train_op],
feed_dict=batch_feed_dict)
batch_losses.append(batch_loss)
print("step: %d/%d, training batch loss: %g" % (step + 1, no_of_batches, batch_loss))
# compute the train epoch loss:
train_epoch_loss = np.mean(batch_losses)
# save the train epoch loss:
train_loss_per_epoch.append(train_epoch_loss)
for i in range(batch_size):
img_path = seq_frame_paths[batch_pointer + i]
img_paths.append(img_path)
print(img_path)
# read the image:
img = cv2.imread(img_path, -1)
print(type(img))
img = cv2.resize(img, (img_width, img_height))
img = img - train_mean_channels
batch_imgs[i] = img
batch_pointer += batch_size
batch_feed_dict = model.create_feed_dict(imgs_batch=batch_imgs,
early_drop_prob=0.0, late_drop_prob=0.0)
# run a forward pass and get the logits:
logits = sess.run(model.logits, feed_dict=batch_feed_dict)
print("step: %d/%d" % (step+1, no_of_batches))
# save all predicted label images overlayed on the input frames to results_dir:
predictions = np.argmax(logits, axis=3)
for i in range(batch_size):
pred_img = predictions[i]
pred_img_color = label_img_to_color(pred_img)
img = batch_imgs[i] + train_mean_channels
img_file_name = img_paths[i].split("/")[-1]