--- a
+++ b/medseg_dl/model/training.py
@@ -0,0 +1,98 @@
+import tensorflow as tf
+import os
+import logging
+from medseg_dl.utils import utils_misc
+
+
+# from tqdm import tqdm
+
+
+def sess_train(spec_pipeline, spec_model, params):
+
+    # Add an op to initialize the variables
+    init_op_vars = tf.global_variables_initializer()
+
+    # Fetch global step of default graph
+    global_step = tf.train.get_global_step()
+
+    # Add ops to save and restore all variables
+    saver_recent = tf.train.Saver(max_to_keep=10)  # keeps the last 10 ckpts
+
+    # generate summary writer
+    writer = tf.summary.FileWriter(params.dict['dir_logs_train'])
+    logging.info(f'saving log to {params.dict["dir_logs_train"]}')
+
+    # Define fetched variables
+    fetched_train = {'loss_value': spec_model['loss'],
+                     'train_op': spec_model['train_op'],
+                     'update_metrics_op_train': spec_model['update_op_metrics'],
+                     'summary_train': spec_model['summary_op_train'],
+                     'gstep': global_step}
+
+    fetched_metrics_train = {'metrics': spec_model['metrics_values'],
+                             'summary_metrics': spec_model['summary_op_metrics']}
+
+    if params.dict['b_viewer_train']:
+        fetched_train.update({'images': spec_pipeline['images'],
+                              'labels': spec_pipeline['labels'],
+                              'positions': spec_pipeline['positions'],
+                              'probs': spec_model['probs']})
+
+    # set growth option
+    config = tf.ConfigProto()
+    config.gpu_options.allow_growth = True
+
+    with tf.Session(config=config) as sess:
+        _ = tf.summary.FileWriter(params.dict['dir_graphs_train'], sess.graph)
+        logging.info(f'Graph saved in {params.dict["dir_graphs_train"]}')
+
+        sess.run(init_op_vars)  # init global variables
+        first_epoch = 0
+        total_steps = 0
+
+        # Reload weights from directory if previous training is restored
+        if params.dict['b_restore']:
+            if os.path.isdir(params.dict['dir_ckpts']):
+                file_ckpt = tf.train.latest_checkpoint(params.dict['dir_ckpts'])
+                first_epoch = int(os.path.basename(file_ckpt).split('-')[1]) + 1
+                logging.info(f'Restoring parameters from {file_ckpt}')
+                saver_recent.restore(sess, file_ckpt)
+
+        # Epochs
+        for epoch in range(first_epoch, params.dict['num_epochs']):
+
+            # training
+            logging.info(f'Epoch {epoch + 1}/{params.dict["num_epochs"]}: training')
+            sess.run(spec_pipeline['init_op_iter'])  # initialize dataset
+            sess.run(spec_model['init_op_metrics'])  # reset metrics
+
+            # training step
+            #pbar = tqdm(total=total_steps)
+            while True:
+                try:
+                    results = sess.run(fetched_train)  # perform mini-batch update
+                    gstep = results['gstep']
+                    loss_value = results['loss_value']
+                    logging.info(f'Step {gstep}, loss: {loss_value}')
+                    writer.add_summary(results['summary_train'], global_step=gstep)  # write loss per batch
+
+                    # allow viewing of data
+                    if params.dict['b_viewer_train']:
+                        shown_index = 2
+                        print(f'Visualized patch has a position of {results["positions"][shown_index,...]}')
+                        utils_misc.show_results(results['images'][shown_index, ...], results['labels'][shown_index, ...], results['probs'][shown_index, ...])
+                    #pbar.update(1)
+                except tf.errors.OutOfRangeError:
+                    #pbar.close()
+                    break
+
+            # fetch aggregated metrics values
+            results = sess.run(fetched_metrics_train)
+            writer.add_summary(results['summary_metrics'], global_step=epoch)  # write metrics per epoch
+
+            # save model every epoch:
+            # if epoch % 5 == 0:
+            save_path = saver_recent.save(sess,
+                                          os.path.join(params.dict['dir_ckpts'], 'model.ckpt'),
+                                          global_step=epoch)
+            logging.info(f'Model saved in {save_path}')