Update datafolder to work with bob9's videoLikeContainer file format

Datafolder used to work with saved FrameContainers, which were hdf5 files containing a dataset per frame, plus a FrameIndexes array:

FrameContainer.hdf5:
    - frame1: CxHxW   
    - frame2: CxHxW 
    ...  
    - FrameIndexes: N

FrameContainers have been deprecated in favor of bob.bio.video.VideoLikeContainer which are stored as hdf5 with a single dataset containing all frames:

VideoLikeContainer.h5:
    - data: NxCxHxW 
    - indices: N 

This MR adapts the datafolder class to the new format with the least changes. It also includes the custom_function of the data_folder_generic.py to allow for custom modifications of the loaded data, taken from DataFolderGeneric (which is now effectively merged into the DataFolder class)

In order to be more robust to future changes to the data format, the loading part of the DataFolder could be using the VideoLikeContainer loading function instead of directly accessing the file.

bob/learn/pytorch/datasets/data_folder.py

@@ -23,7 +23,13 @@ import h5py
 
 
 # ==============================================================================
-def get_file_names_and_labels(files, data_folder, extension=".hdf5", hldi_type="pad"):
+def get_file_names_and_labels(
+    files,
+    data_folder,
+    extension=".h5",
+    hldi_type="pad",
+    allow_missing_files=True,
+):
     """
     Get absolute names of the corresponding file objects and their class labels,
     as well as keys defining name of the frame to load the data from.
@@ -39,33 +45,34 @@ def get_file_names_and_labels(files, data_folder, extension=".hdf5", hldi_type="
         A directory containing the training data.
 
     extension : str
-        Extension of the data files. Default: ".hdf5" .
+        Extension of the data files. Default: ".h5" .
 
     hldi_type : str
         Type of the high level database interface. Default: "pad".
         Note: this is the only type supported at the moment.
 
+    allow_missing_files : bool
+        If False, will raise an error if a file of the database is missing. 
+        If True only a warning will be printed.
+
     Returns
     -------
 
-    file_names_labels_keys : [(str, int, str)]
+    file_names_labels_indices : [(str, int, int)]
         A list of tuples, where each tuple contain an absolute filename,
-        a corresponding label of the class, and a key defining the name of the
-        frame to extract the data from.
+        a corresponding label of the class, and the index of frame to extract
+        the data from.
     """
 
-    file_names_labels_keys = []
+    file_names_labels_indices = []
 
     if hldi_type == "pad":
 
         for f in files:
 
             if f.attack_type is None:
-
                 label = 1
-
             else:
-
                 label = 0
 
             file_name = os.path.join(data_folder, f.path + extension)
@@ -74,25 +81,29 @@ def get_file_names_and_labels(files, data_folder, extension=".hdf5", hldi_type="
 
                 with h5py.File(file_name, "r") as f_h5py:
 
-                    file_keys = list(f_h5py.keys())
-
-                # removes the 'FrameIndexes' key
-                file_keys = [f for f in file_keys if f != "FrameIndexes"]
+                    n_frames = len(f_h5py["data"])  # shape[0]
 
                 # elements of tuples in the below list are as follows:
                 # a filename a key is extracted from,
                 # a label corresponding to the file,
                 # a key defining a frame from the file.
-                file_names_labels_keys = file_names_labels_keys + [
-                    (file_name, label, key)
-                    for file_name, label, key in zip(
-                        [file_name] * len(file_keys),
-                        [label] * len(file_keys),
-                        file_keys,
-                    )
-                ]
+                file_names_labels_indices.extend(
+                    [
+                        (file_name, label, index)
+                        for file_name, label, index in zip(
+                            [file_name] * n_frames,
+                            [label] * n_frames,
+                            range(n_frames),
+                        )
+                    ]
+                )
 
-    return file_names_labels_keys
+            else:
+                if not allow_missing_files:
+                    raise ValueError(file_name + " is not a file")
+                print("Missing file: " + file_name)
+
+    return file_names_labels_indices
 
 
 # ==============================================================================
@@ -121,7 +132,7 @@ class DataFolder(data.Dataset):
 
     data_folder : str
         A directory containing the training data. Note, that the training data
-        must be stored as a FrameContainers written to the hdf5 files. Other
+        must be stored as a VideoLikeContainer written to the hdf5 files. Other
         formats are currently not supported.
 
     transform : object
@@ -164,13 +175,14 @@ class DataFolder(data.Dataset):
         self,
         data_folder,
         transform=None,
-        extension=".hdf5",
+        extension=".h5",
         bob_hldi_instance=None,
         hldi_type="pad",
         groups=["train", "dev", "eval"],
         protocol="grandtest",
         purposes=["real", "attack"],
         allow_missing_files=True,
+        custom_func=None,
         **kwargs
     ):
         """
@@ -225,6 +237,7 @@ class DataFolder(data.Dataset):
         self.protocol = protocol
         self.purposes = purposes
         self.allow_missing_files = allow_missing_files
+        self.custom_func = custom_func
 
         if bob_hldi_instance is not None:
 
@@ -235,25 +248,26 @@ class DataFolder(data.Dataset):
                 **kwargs
             )
 
-            file_names_labels_keys = get_file_names_and_labels(
+            file_names_labels_indices = get_file_names_and_labels(
                 files=files,
                 data_folder=self.data_folder,
                 extension=self.extension,
                 hldi_type=self.hldi_type,
+                allow_missing_files=self.allow_missing_files,
             )
 
             if self.allow_missing_files:  # return only existing files
 
-                file_names_labels_keys = [
-                    f for f in file_names_labels_keys if os.path.isfile(f[0])
+                file_names_labels_indices = [
+                    f for f in file_names_labels_indices if os.path.isfile(f[0])
                 ]
 
         else:
 
             # TODO - add behaviour similar to image folder
-            file_names_labels_keys = []
+            file_names_labels_indices = []
 
-        self.file_names_labels_keys = file_names_labels_keys
+        self.file_names_labels_indices = file_names_labels_indices
 
     # ==========================================================================
     def __getitem__(self, index):
@@ -278,12 +292,11 @@ class DataFolder(data.Dataset):
             Index of the class.
         """
 
-        path, target, key = self.file_names_labels_keys[index]
+        path, target, frame_index = self.file_names_labels_indices[index]
 
         with h5py.File(path, "r") as f_h5py:
-
             img_array = np.array(
-                f_h5py.get(key + "/array")
+                f_h5py["data"][frame_index]
             )  # The size now is (3 x W x H)
 
         if isinstance(
@@ -295,14 +308,15 @@ class DataFolder(data.Dataset):
                 img_array_tr = np.swapaxes(img_array, 1, 2)
                 img_array_tr = np.swapaxes(img_array_tr, 0, 2)
 
-                np_img = img_array_tr.copy()  # np_img is numpy.ndarray of shape HxWxC
+                np_img = (
+                    img_array_tr.copy()
+                )  # np_img is numpy.ndarray of shape HxWxC
 
             else:  # for gray-scale images
 
                 np_img = np.expand_dims(
-                    img_array_tr, 2
+                    img_array, 2
                 )  # np_img is numpy.ndarray of size HxWx1
-
             if self.transform is not None:
 
                 np_img = self.transform(
@@ -316,6 +330,11 @@ class DataFolder(data.Dataset):
             return img_array_transformed, target
             # NOTE: make sure ``img_array_transformed`` converted to Tensor in your custom ``transform`` function.
 
+        if (
+            self.custom_func is not None
+        ):  # custom function to change the return to something else
+
+            return self.custom_func(np_img, target)
         return np_img, target
 
     # ==========================================================================
@@ -327,4 +346,5 @@ class DataFolder(data.Dataset):
         len : int
             The length of the file list.
         """
-        return len(self.file_names_labels_keys)
+        return len(self.file_names_labels_indices)
+