Merge pull request #27 from mlgill/memory_mapped_data_loading

Memory mapped data loading

prosit/__init__.py

@@ -1,4 +1,4 @@
-from . import io
+from . import io_local
 from . import constants
 from . import model
 from . import alignment

prosit/io_local.py

@@ -0,0 +1,29 @@
+from . import utils
+
+
+def get_array(tensor, keys):
+    utils.check_mandatory_keys(tensor, keys)
+    return [tensor[key] for key in keys]
+
+
+def to_hdf5(dictionary, path):
+    import h5py
+
+    with h5py.File(path, "w") as f:
+        for key, data in dictionary.items():
+            f.create_dataset(key, data=data, dtype=data.dtype, compression="gzip")
+
+
+def from_hdf5(path, n_samples=None):
+    from keras.utils import HDF5Matrix
+    import h5py
+
+    # Get a list of the keys for the datasets
+    with h5py.File(path, 'r') as f:
+        dataset_list = list(f.keys())
+
+    # Assemble into a dictionary
+    data = dict()
+    for dataset in dataset_list:
+        data[dataset] = HDF5Matrix(path, dataset, start=0, end=n_samples, normalizer=None)
+    return data

prosit/prediction.py

@@ -3,14 +3,14 @@
 import numpy as np
 
 from . import model as model_lib
-from . import io
+from . import io_local
 from . import constants
 from . import sanitize
 
 
 def predict(data, d_model):
     # check for mandatory keys
-    x = io.get_array(data, d_model["config"]["x"])
+    x = io_local.get_array(data, d_model["config"]["x"])
 
     keras.backend.set_session(d_model["session"])
     with d_model["graph"].as_default():

prosit/server.py

@@ -7,7 +7,7 @@
 import tensorflow as tf
 
 from . import model
-from . import io
+from . import io_local
 from . import constants
 from . import tensorize
 from . import prediction

prosit/training.py

@@ -1,6 +1,6 @@
 import os
 
-from . import io
+from . import io_local
 from . import losses
 from . import model as model_lib
 from . import constants
@@ -28,8 +28,8 @@ def train(tensor, model, model_config, callbacks):
     else:
         loss = losses.get(model_config["loss"])
     optimizer = model_config["optimizer"]
-    x = io.get_array(tensor, model_config["x"])
-    y = io.get_array(tensor, model_config["y"])
+    x = io_local.get_array(tensor, model_config["x"])
+    y = io_local.get_array(tensor, model_config["y"])
     model.compile(optimizer=optimizer, loss=loss)
     model.fit(
         x=x,
@@ -48,6 +48,6 @@ def train(tensor, model, model_config, callbacks):
     model_dir = constants.MODEL_DIR
 
     model, model_config = model_lib.load(model_dir, trained=True)
-    tensor = io.from_hdf5(data_path)
+    tensor = io_local.from_hdf5(data_path)
     callbacks = get_callbacks(model_dir)
     train(tensor, model, model_config, callbacks)