Merge pull request #209 from azavea/lf/rv1-parallel

Add parallel prediction for old version of object detection

Author: lewfish

src/config-samples/detection/train/mobilenet.config

@@ -141,6 +141,7 @@ model {
 
 train_config: {
   batch_size: 8
+  num_steps: 40000 # Set to 0 for an indefinite number of steps
   optimizer {
     rms_prop_optimizer: {
       learning_rate: {

src/rv/detection/commands/merge_predictions.py

@@ -11,46 +11,42 @@
     MyTemporaryDirectory)
 
 
-def get_annotations_paths(projects_path, temp_dir):
-    annotations_paths = []
-    with open(projects_path, 'r') as projects_file:
-        projects = json.load(projects_file)
-        for project_ind, project in enumerate(projects):
-            annotations_uri = project['annotations']
-            annotations_path = download_if_needed(
-                annotations_uri, temp_dir)
-            annotations_paths.append(annotations_path)
-    return annotations_paths
-
-
-def merge_annotations(annotations_list):
-    all_annotations = copy.deepcopy(annotations_list[0])
-    for annotations in annotations_list[1:]:
-        all_annotations['features'].extend(annotations['features'])
-    return all_annotations
+def _merge_predictions(predictions_list):
+    merged_predictions = copy.deepcopy(predictions_list[0])
+    for predictions in predictions_list[1:]:
+        merged_predictions['features'].extend(predictions['features'])
+    return merged_predictions
 
 
 @click.command()
 @click.argument('projects_uri')
-@click.argument('output_uri')
+@click.argument('output_dir_uri')
 @click.option('--save-temp', is_flag=True)
-def merge_predictions(projects_uri, output_uri, save_temp):
+def merge_predictions(projects_uri, output_dir_uri, save_temp):
     prefix = temp_root_dir
     temp_dir = os.path.join(prefix, 'merge-predictions') if save_temp else None
     with MyTemporaryDirectory(temp_dir, prefix) as temp_dir:
         projects_path = download_if_needed(projects_uri, temp_dir)
-        output_path = get_local_path(output_uri, temp_dir)
-
-        annotation_paths = get_annotations_paths(projects_path, temp_dir)
-        annotations_list = []
-        for annotation_path in annotation_paths:
-            with open(annotation_path, 'r') as annotation_file:
-                annotations_list.append(json.load(annotation_file))
-
-        annotations = merge_annotations(annotations_list)
-        with open(output_path, 'w') as output_file:
-            json.dump(annotations, output_file, indent=4)
-        upload_if_needed(output_path, output_uri)
+
+        # For each project:
+        # download the predictions files, merge them, and upload the merged
+        # predictions.
+        projects = json.load(open(projects_path))
+        for project in projects:
+            predictions_list = []
+            for image_ind, image in enumerate(project['images']):
+                predictions_uri = os.path.join(
+                    output_dir_uri, project['id'],
+                    '{}.json'.format(image_ind))
+                predictions_path = download_if_needed(
+                    predictions_uri, temp_dir)
+                predictions_list.append(json.load(open(predictions_path)))
+
+            output_uri = project['annotations']
+            output_path = get_local_path(output_uri, temp_dir)
+            predictions = _merge_predictions(predictions_list)
+            json.dump(predictions, open(output_path, 'w'))
+            upload_if_needed(output_path, output_uri)
 
 
 if __name__ == '__main__':

src/rv/detection/commands/predict_array.py

@@ -14,34 +14,46 @@
 @click.argument('inference_graph_uri')
 @click.argument('label_map_uri')
 @click.argument('projects_uri')
+@click.argument('output_dir_uri')
 @click.option('--mask-uri', default=None,
               help='URI for mask GeoJSON file to use as filter for detections')
 @click.option('--channel-order', nargs=3, type=int,
               default=default_channel_order, help='Index of RGB channels')
 @click.option('--chip-size', default=300)
 @click.option('--score-thresh', default=0.5,
               help='Score threshold of predictions to keep')
-@click.option('--merge-thresh', default=0.05,
+@click.option('--merge-thresh', default=0.5,
               help='IOU threshold for merging predictions')
 @click.option('--save-temp', is_flag=True)
 def predict_array(inference_graph_uri, label_map_uri, projects_uri,
-                  mask_uri, channel_order, chip_size, score_thresh,
-                  merge_thresh, save_temp):
-    job_index = int(os.environ['AWS_BATCH_JOB_ARRAY_INDEX'])
+                  output_dir_uri, mask_uri, channel_order, chip_size,
+                  score_thresh, merge_thresh, save_temp):
+    job_ind = int(os.environ['AWS_BATCH_JOB_ARRAY_INDEX'])
 
     prefix = temp_root_dir
     temp_dir = os.path.join(prefix, 'predict-array') if save_temp else None
     with MyTemporaryDirectory(temp_dir, prefix) as temp_dir:
         projects_path = download_if_needed(projects_uri, temp_dir)
         with open(projects_path, 'r') as projects_file:
             projects = json.load(projects_file)
-            if job_index >= len(projects):
-                raise ValueError(
-                    'There are {} projects and job_index is {}!'.format(
-                        len(projects), job_index))
-            project = projects[job_index]
-            image_uris = project['images']
-            output_uri = project['annotations']
+
+            def get_image_coords():
+                image_ind = 0
+                for project_ind, project in enumerate(projects):
+                    for project_image_ind, image_uri in enumerate(project['images']):
+                        if job_ind == image_ind:
+                            return project_ind, project_image_ind
+                        image_ind += 1
+
+            # Run predict for single image and generate ouput_uri based on
+            # the index of the image within the project.
+            project_ind, project_image_ind = get_image_coords()
+            project = projects[project_ind]
+            image_uri = project['images'][project_image_ind]
+            image_uris = [image_uri]
+            output_uri = os.path.join(
+                output_dir_uri, project['id'],
+                '{}.json'.format(project_image_ind))
             output_debug_uri = None
 
             _predict(inference_graph_uri, label_map_uri, image_uris,
@@ -50,6 +62,5 @@ def predict_array(inference_graph_uri, label_map_uri, projects_uri,
                      save_temp)
 
 
-
 if __name__ == '__main__':
     predict_array()

src/rv/detection/commands/train.py

@@ -1,9 +1,11 @@
-from os.path import join, dirname, splitext
+from os.path import join, basename
 import os
 from subprocess import Popen
 import zipfile
 from threading import Timer
 from urllib.parse import urlparse
+import glob
+import re
 
 import click
 
@@ -14,6 +16,30 @@
 from rv.detection.commands.settings import temp_root_dir
 
 
+def get_last_checkpoint_path(train_root_dir):
+    index_paths = glob.glob(join(train_root_dir, 'train', '*.index'))
+    checkpoint_ids = []
+    for index_path in index_paths:
+        match = re.match(r'model.ckpt-(\d+).index', basename(index_path))
+        checkpoint_ids.append(int(match.group(1)))
+    checkpoint_id = max(checkpoint_ids)
+    checkpoint_path = join(
+        train_root_dir, 'train', 'model.ckpt-{}'.format(checkpoint_id))
+    return checkpoint_path
+
+
+def export_inference_graph(train_root_dir, config_path, inference_graph_path):
+    checkpoint_path = get_last_checkpoint_path(train_root_dir)
+    print('Exporting checkpoint {}...'.format(checkpoint_path))
+    train_process = Popen([
+        'python', '/opt/src/tf/object_detection/export_inference_graph.py',
+        '--input_type', 'image_tensor',
+        '--pipeline_config_path', config_path,
+        '--checkpoint_path', checkpoint_path,
+        '--inference_graph_path', inference_graph_path])
+    train_process.wait()
+
+
 @click.command()
 @click.argument('config_uri')
 @click.argument('train_dataset_uri')
@@ -41,6 +67,7 @@ def train(config_uri, train_dataset_uri, val_dataset_uri, model_checkpoint_uri,
         make_dir(train_root_dir)
         train_dir = join(train_root_dir, 'train')
         eval_dir = join(train_root_dir, 'eval')
+        inference_graph_path = join(train_root_dir, 'inference_graph.pb')
 
         def process_zip_file(uri, temp_dir, link_dir):
             if uri.endswith('.zip'):
@@ -84,9 +111,15 @@ def sync_train_dir(delete=True):
             'tensorboard', '--logdir={}'.format(train_root_dir)],
             preexec_fn=on_parent_exit('SIGTERM'))
 
+        # After training finishes due to num_steps exceeded,
+        # kill monitor processes, export inference graph, and upload.
         train_process.wait()
-        eval_process.wait()
-        tensorboard_process.wait()
+        eval_process.kill()
+        tensorboard_process.kill()
+        export_inference_graph(
+            train_root_dir, config_path, inference_graph_path)
+        if urlparse(train_uri).scheme == 's3':
+            sync_dir(train_root_dir, train_uri, delete=True)
 
 
 if __name__ == '__main__':

src/workflows/detection/parallel_predict.py

@@ -0,0 +1,65 @@
+import json
+import os
+import click
+
+from rv.utils.files import (
+    download_if_needed, MyTemporaryDirectory)
+from rv.utils.batch import _batch_submit
+from rv.detection.commands.settings import temp_root_dir
+
+
+def make_predict_array_cmd(inference_graph_uri, label_map_uri, projects_uri,
+                           output_dir_uri):
+    return 'python -m rv.detection.run predict_array {} {} {} {}'.format(
+        inference_graph_uri, label_map_uri, projects_uri, output_dir_uri)
+
+
+def make_merge_predictions_cmd(projects_uri, output_dir_uri):
+    return 'python -m rv.detection.run merge_predictions {} {}'.format(
+        projects_uri, output_dir_uri)
+
+
+def get_nb_images(projects):
+    nb_images = 0
+    for project in projects:
+        nb_images += len(project['images'])
+    return nb_images
+
+
+@click.command()
+@click.argument('projects_uri')
+@click.argument('label_map_uri')
+@click.argument('inference_graph_uri')
+@click.argument('output_dir_uri')
+@click.option('--branch-name', default='develop')
+@click.option('--attempts', default=1)
+@click.option('--cpu', is_flag=True)
+def parallel_predict(projects_uri, label_map_uri, inference_graph_uri,
+                     output_dir_uri,
+                     branch_name, attempts, cpu):
+    prefix = temp_root_dir
+    temp_dir = os.path.join(prefix, 'parallel-predict')
+    with MyTemporaryDirectory(temp_dir, prefix) as temp_dir:
+        # Load projects and count number of images
+        projects_path = download_if_needed(projects_uri, temp_dir)
+        projects = json.load(open(projects_path))
+        nb_images = get_nb_images(projects)
+
+        # Submit an array job with nb_images elements.
+        command = make_predict_array_cmd(
+            inference_graph_uri, label_map_uri, projects_uri, output_dir_uri)
+        '''
+        predict_job_id = _batch_submit(
+            branch_name, command, attempts=attempts, cpu=cpu,
+            array_size=nb_images)
+        '''
+        # Submit a dependent merge_predictions job.
+        command = make_merge_predictions_cmd(
+            projects_uri, output_dir_uri)
+        _batch_submit(
+            branch_name, command, attempts=attempts, cpu=cpu)
+            #parent_job_ids=[predict_job_id])
+
+
+if __name__ == '__main__':
+    parallel_predict()