Commenting

LPDI-EPFL · Aug 28, 2019 · 05d15e6 · 05d15e6
1 parent 52c7994
commit 05d15e6
Show file tree

Hide file tree

Showing 3 changed files with 12 additions and 5 deletions.
diff --git a/source/data_preparation/04b-make_ligand_tfrecords.py b/source/data_preparation/04b-make_ligand_tfrecords.py
@@ -52,6 +52,7 @@
     for i, pdb in enumerate(train_pdbs):
         print("Working on", pdb)
         try:
+            # Load precomputed data
             input_feat = np.load(
                 os.path.join(precom_dir, pdb + "_", "p1_input_feat.npy")
             )

diff --git a/source/masif_ligand/masif_ligand_evaluate_test.py b/source/masif_ligand/masif_ligand_evaluate_test.py
@@ -16,6 +16,7 @@
 """
 
 params = masif_opts["ligand"]
+# Load testing data
 testing_data = tf.contrib.data.TFRecordDataset(
     os.path.join(params["tfrecords_dir"], "testing_data_sequenceSplit_30.tfrecord")
 )
@@ -30,7 +31,7 @@
 
 
 with tf.Session() as sess:
-    # Build trained network
+    # Build network
     learning_obj = MaSIF_ligand(
         sess,
         params["max_distance"],
@@ -39,6 +40,7 @@
         feat_mask=params["feat_mask"],
         costfun=params["costfun"],
     )
+    # Load pretrained network
     learning_obj.saver.restore(learning_obj.session, output_model)
 
     num_test_samples = 290
@@ -62,6 +64,7 @@
         pdb_logits_softmax = []
         pdb_labels = []
         for ligand in range(n_ligands):
+            # Rows indicate point number and columns ligand type
             pocket_points = np.where(labels[:, ligand] != 0.0)[0]
             label = np.max(labels[:, ligand]) - 1
             pocket_labels = np.zeros(7, dtype=np.float32)
@@ -77,7 +80,7 @@
             samples_data_loss = []
             # Make 100 predictions
             for i in range(100):
-                # sample = samples[int(i*32):int((i+1)*32)]
+                # Sample pocket randomly
                 sample = np.random.choice(pocket_points, 32, replace=False)
                 feed_dict = {
                     learning_obj.input_feat: data_element[0][sample, :, :],

diff --git a/source/masif_ligand/masif_ligand_train.py b/source/masif_ligand/masif_ligand_train.py
@@ -84,7 +84,6 @@
             npoints = pocket_points.shape[0]
             if npoints < 32:
                 continue
-            sample = np.random.choice(pocket_points, 32, replace=False)
             # For evaluating take the first 32 points of the pocket
             feed_dict = {
                 learning_obj.input_feat: data_element[0][pocket_points[:32], :, :],
@@ -112,7 +111,9 @@
                 num_epoch, np.mean(training_losses), np.median(training_losses)
             )
         )
+        # Generate confusion matrix
         training_conf_mat = confusion_matrix(training_ytrue, training_ypred)
+        # Compute accuracy
         training_accuracy = float(np.sum(np.diag(training_conf_mat))) / np.sum(
             training_conf_mat
         )
@@ -123,6 +124,7 @@
         validation_ytrue = []
         validation_ypred = []
         print("Calulating validation loss")
+        # Compute accuracy on the validation set
         for num_val_sample in range(num_validation_samples):
             try:
                 data_element = sess.run(validation_next_element)
@@ -138,7 +140,6 @@
             npoints = pocket_points.shape[0]
             if npoints < 32:
                 continue
-            sample = np.random.choice(pocket_points, 32, replace=False)
             feed_dict = {
                 learning_obj.input_feat: data_element[0][pocket_points[:32], :, :],
                 learning_obj.rho_coords: np.expand_dims(data_element[1], -1)[
@@ -195,7 +196,6 @@
             npoints = pocket_points.shape[0]
             if npoints < 32:
                 continue
-            sample = np.random.choice(pocket_points, 32, replace=False)
             feed_dict = {
                 learning_obj.input_feat: data_element[0][pocket_points[:32], :, :],
                 learning_obj.rho_coords: np.expand_dims(data_element[1], -1)[
@@ -228,9 +228,11 @@
         )
         print(testing_conf_mat)
         print("Testing accuracy:", testing_accuracy)
+        # Stop training if number of iterations has reached 40000
         if total_iterations == 40000:
             break
 
+        # Train the network
         training_losses = []
         training_ytrue = []
         training_ypred = []
@@ -251,6 +253,7 @@
             npoints = pocket_points.shape[0]
             if npoints < 32:
                 continue
+            # Sample 32 points randomly
             sample = np.random.choice(pocket_points, 32, replace=False)
             feed_dict = {
                 learning_obj.input_feat: data_element[0][sample, :, :],