Refactor out shared LSTM/GGNN training loop.

github.com//issues/69

Author: ChrisCummins

programl/task/dataflow/dataflow.py

@@ -21,7 +21,7 @@
 import warnings
 from typing import Tuple
 
-from labm8.py import app, pbutil
+from labm8.py import app, humanize, pbutil
 from sklearn.exceptions import UndefinedMetricWarning
 
 from programl.proto import checkpoint_pb2, epoch_pb2
@@ -208,3 +208,50 @@ def CreateLoggingDirectories(
     (log_dir / "checkpoints").mkdir()
     (log_dir / "graph_loader").mkdir()
     return log_dir
+
+
+def run_training_loop(log_dir, epochs, start_epoch_step, model):
+    for (
+        epoch_step,
+        (train_graph_count, train_graph_cumsum, train_batches),
+    ) in enumerate(epochs, start=start_epoch_step):
+        start_time = time.time()
+        hr_graph_cumsum = f"{humanize.Commas(train_graph_cumsum)} graphs"
+
+        train_results = model.RunBatches(
+            epoch_pb2.TRAIN,
+            train_batches,
+            log_prefix=f"Train to {hr_graph_cumsum}",
+            total_graph_count=train_graph_count,
+        )
+        val_results = model.RunBatches(
+            epoch_pb2.VAL,
+            val_batches.batches,
+            log_prefix=f"Val at {hr_graph_cumsum}",
+            total_graph_count=FLAGS.val_graph_count,
+        )
+
+        # Write the epoch to file as an epoch list. This may seem redundant since
+        # epoch list contains a single item, but it means that we can easily
+        # concatenate a sequence of these epoch protos to produce a valid epoch
+        # list using: `cat *.EpochList.pbtxt > epochs.pbtxt`
+        epoch = epoch_pb2.EpochList(
+            epoch=[
+                epoch_pb2.Epoch(
+                    walltime_seconds=time.time() - start_time,
+                    epoch_num=epoch_step,
+                    train_results=train_results,
+                    val_results=val_results,
+                )
+            ]
+        )
+        print(epoch, end="")
+
+        epoch_path = log_dir / "epochs" / f"{epoch_step:03d}.EpochList.pbtxt"
+        pbutil.ToFile(epoch, epoch_path)
+        app.Log(1, "Wrote %s", epoch_path)
+
+        checkpoint_path = log_dir / "checkpoints" / f"{epoch_step:03d}.Checkpoint.pb"
+        pbutil.ToFile(model.SaveCheckpoint(), checkpoint_path)
+
+    return log_dir

programl/task/dataflow/ggnn.py

@@ -173,50 +173,7 @@ def TrainDataflowGGNN(
         )
     )
 
-    for (
-        epoch_step,
-        (train_graph_count, train_graph_cumsum, train_batches),
-    ) in enumerate(epochs, start=start_epoch_step):
-        start_time = time.time()
-        hr_graph_cumsum = f"{humanize.Commas(train_graph_cumsum)} graphs"
-
-        train_results = model.RunBatches(
-            epoch_pb2.TRAIN,
-            train_batches,
-            log_prefix=f"Train to {hr_graph_cumsum}",
-            total_graph_count=train_graph_count,
-        )
-        val_results = model.RunBatches(
-            epoch_pb2.VAL,
-            val_batches.batches,
-            log_prefix=f"Val at {hr_graph_cumsum}",
-            total_graph_count=val_graph_count,
-        )
-
-        # Write the epoch to file as an epoch list. This may seem redundant since
-        # epoch list contains a single item, but it means that we can easily
-        # concatenate a sequence of these epoch protos to produce a valid epoch
-        # list using: `cat *.EpochList.pbtxt > epochs.pbtxt`
-        epoch = epoch_pb2.EpochList(
-            epoch=[
-                epoch_pb2.Epoch(
-                    walltime_seconds=time.time() - start_time,
-                    epoch_num=epoch_step,
-                    train_results=train_results,
-                    val_results=val_results,
-                )
-            ]
-        )
-        print(epoch, end="")
-
-        epoch_path = log_dir / "epochs" / f"{epoch_step:03d}.EpochList.pbtxt"
-        pbutil.ToFile(epoch, epoch_path)
-        app.Log(1, "Wrote %s", epoch_path)
-
-        checkpoint_path = log_dir / "checkpoints" / f"{epoch_step:03d}.Checkpoint.pb"
-        pbutil.ToFile(model.SaveCheckpoint(), checkpoint_path)
-
-    return log_dir
+    return dataflow.run_training_loop(log_dir, epochs, start_epoch_step, model)
 
 
 def TestDataflowGGNN(

programl/task/dataflow/train_lstm.py

@@ -160,50 +160,7 @@ def TrainDataflowLSTM(
         )
     )
 
-    for (
-        epoch_step,
-        (train_graph_count, train_graph_cumsum, train_batches),
-    ) in enumerate(epochs, start=start_epoch_step):
-        start_time = time.time()
-        hr_graph_cumsum = f"{humanize.Commas(train_graph_cumsum)} graphs"
-
-        train_results = model.RunBatches(
-            epoch_pb2.TRAIN,
-            train_batches,
-            log_prefix=f"Train to {hr_graph_cumsum}",
-            total_graph_count=train_graph_count,
-        )
-        val_results = model.RunBatches(
-            epoch_pb2.VAL,
-            val_batches.batches,
-            log_prefix=f"Val at {hr_graph_cumsum}",
-            total_graph_count=FLAGS.val_graph_count,
-        )
-
-        # Write the epoch to file as an epoch list. This may seem redundant since
-        # epoch list contains a single item, but it means that we can easily
-        # concatenate a sequence of these epoch protos to produce a valid epoch
-        # list using: `cat *.EpochList.pbtxt > epochs.pbtxt`
-        epoch = epoch_pb2.EpochList(
-            epoch=[
-                epoch_pb2.Epoch(
-                    walltime_seconds=time.time() - start_time,
-                    epoch_num=epoch_step,
-                    train_results=train_results,
-                    val_results=val_results,
-                )
-            ]
-        )
-        print(epoch, end="")
-
-        epoch_path = log_dir / "epochs" / f"{epoch_step:03d}.EpochList.pbtxt"
-        pbutil.ToFile(epoch, epoch_path)
-        app.Log(1, "Wrote %s", epoch_path)
-
-        checkpoint_path = log_dir / "checkpoints" / f"{epoch_step:03d}.Checkpoint.pb"
-        pbutil.ToFile(model.SaveCheckpoint(), checkpoint_path)
-
-    return log_dir
+    return dataflow.run_training_loop(log_dir, epochs, start_epoch_step, model)
 
 
 def TestDataflowLSTM(