Merge pull request #54 from BirdVox/export-context

Export context

birdvoxdetect/core.py

@@ -45,6 +45,7 @@ def process_file(
         output_dir=None,
         export_clips=False,
         export_confidence=False,
+        export_context=False,
         export_faults=False,
         export_logger=False,
         threshold=50.0,
@@ -223,10 +224,14 @@ def process_file(
         if not os.path.exists(clips_dir):
             os.makedirs(clips_dir)
 
-    # Append confidence to list of per-chunk confidences.
+    # Initialize list of per-chunk confidences.
     if export_confidence:
         chunk_confidences = []
 
+    # Initialize list of context arrays.
+    if export_context:
+        contexts = []
+
     # Print chunk duration.
     logger.info("Chunk duration: {} seconds".format(chunk_duration))
     logger.info("")
@@ -254,6 +259,10 @@ def process_file(
         deque_context = np.percentile(
             concat_deque, percentiles, axis=1, overwrite_input=True)
 
+        # Append context.
+        if export_context:
+            contexts.append(deque_context)
+
         # Compute sensor fault features.
         # Median is 4th order statistic. Restrict to lowest 120 mel-freq bins
         context_median = deque_context[4, :120]
@@ -415,6 +424,10 @@ def process_file(
             concat_deque, percentiles,
             axis=1, out=deque_context, overwrite_input=True)
 
+        # Append context.
+        if export_context:
+            contexts.append(deque_context)
+
         # Compute sensor fault features.
         # Median is 4th order statistic. Restrict to lowest 120 mel-freq bins
         context_median = deque_context[4, :120]
@@ -580,7 +593,7 @@ def process_file(
         chunk_confidence_length = int(frame_rate*full_length/sr)
         chunk_confidence = np.full(chunk_confidence_length, np.nan)
 
-        if has_context and (n_chunks==1):
+        if (export_context or has_context) and (n_chunks==1):
             deque_context = np.percentile(
                 chunk_pcen, percentiles, axis=1, overwrite_input=True)
             logging.warning(
@@ -601,6 +614,9 @@ def process_file(
                 concat_deque, percentiles,
                 axis=1, out=deque_context, overwrite_input=True)
 
+        # Append context.
+        if export_context:
+            contexts.append(deque_context)
 
     if not has_sensor_fault:
         # Define trimming length for last chunk.
@@ -734,6 +750,21 @@ def process_file(
             # Increment pointer.
             chunk_pointer = next_chunk_pointer
 
+    # Export context.
+    if export_context:
+        # Define output path for context.
+        context_path = get_output_path(
+            filepath, suffix + "context.hdf5", output_dir=output_dir)
+
+        # Stack context over time.
+        context_array = np.stack(contexts, axis=0)
+
+        # Export context.
+        with h5py.File(context_path, "w") as f:
+            f["context"] = context_array
+            f["chunk_duration"] = chunk_duration
+            f["frame_rate"] = frame_rate
+
     # Print final messages.
     if threshold is not None:
         df = pd.read_csv(checklist_path)
@@ -747,6 +778,9 @@ def process_file(
     if export_confidence:
         event_str = "The event detection curve is available at: {}"
         logger.info(event_str.format(confidence_path))
+    if export_context:
+        event_str = "The context array is available at: {}"
+        logger.info(event_str.format(context_path))
     if export_faults:
         event_str = "The list of sensor faults is available at: {}"
         logger.info(event_str.format(faultlist_path))

birdvoxdetect/version.py

@@ -1,2 +1,2 @@
 short_version = '0.2'
-version = '0.3.0a3'
+version = '0.3.0a4'

tests/test_core.py

@@ -126,6 +126,19 @@ def test_process_file():
     assert confidence.shape == (199,)
     shutil.rmtree(tmpdir)
 
+    # export context
+    tmpdir = tempfile.mkdtemp()
+    process_file(
+        os.path.join(TEST_AUDIO_DIR, POSITIVE_MD5 + '.wav'),
+        output_dir=tmpdir,
+        export_context=True)
+    context_path = os.path.join(
+        tmpdir, POSITIVE_MD5 + '_context.hdf5')
+    assert os.path.exists(context_path)
+    with h5py.File(context_path, "r") as f:
+        confidence = f["context"][()]
+    shutil.rmtree(tmpdir)
+
     # export list of sensor faults
     tmpdir = tempfile.mkdtemp()
     process_file(