megadetector 5.0.7__py3-none-any.whl → 5.0.9__py3-none-any.whl

classification/train_classifier_tf.py

@@ -1,27 +1,16 @@
-########
-# 
-# train_classifier_tf.py
-#
-# Train an EfficientNet classifier.
-#
-# Currently the implementation of multi-label multi-class classification is
-# non-functional.
-#
-# During training, start tensorboard from within the classification/ directory:
-#    tensorboard --logdir run --bind_all --samples_per_plugin scalars=0,images=0
-#
-########
+"""
 
-#%% Example usage
+train_classifier_tf.py
 
-"""    
-    python train_classifier_tf.py run_idfg /ssd/crops_sq \
-        -m "efficientnet-b0" --pretrained --finetune --label-weighted \
-        --epochs 50 --batch-size 512 --lr 1e-4 \
-        --seed 123 \
-        --logdir run_idfg
-"""
+Train an EfficientNet classifier.
 
+Currently the implementation of multi-label multi-class classification is
+non-functional.
+
+During training, start tensorboard from within the classification/ directory:
+   tensorboard --logdir run --bind_all --samples_per_plugin scalars=0,images=0
+
+"""
 
 #%% Imports and constants
 
@@ -47,7 +36,6 @@ from classification.train_utils import (
     imgs_with_confidences, load_dataset_csv, prefix_all_keys)
 from md_visualization import plot_utils
 
-
 AUTOTUNE = tf.data.experimental.AUTOTUNE
 
 # match pytorch EfficientNet model names
@@ -63,6 +51,17 @@ EFFICIENTNET_MODELS: Mapping[str, Mapping[str, Any]] = {
 }
 
 
+#%% Example usage
+
+"""    
+    python train_classifier_tf.py run_idfg /ssd/crops_sq \
+        -m "efficientnet-b0" --pretrained --finetune --label-weighted \
+        --epochs 50 --batch-size 512 --lr 1e-4 \
+        --seed 123 \
+        --logdir run_idfg
+"""
+
+
 #%% Support functions
 
 def create_dataset(

classification/train_utils.py

@@ -1,13 +1,13 @@
-########
-# 
-# train_utils.py
-#
-# Utility functions useful for training a classifier.
-#
-# This script should NOT depend on any other file within this repo. It should
-# especially be agnostic to PyTorch vs. TensorFlow.
-#
-########
+"""
+
+train_utils.py
+
+Utility functions useful for training a classifier.
+
+This script should NOT depend on any other file within this repo. It should
+especially be agnostic to PyTorch vs. TensorFlow.
+
+"""
 
 #%% Imports
 

data_management/annotations/annotation_constants.py

@@ -1,37 +1,20 @@
-########
-#
-# annotation_constants.py
-#
-# Shared constants used to interpret annotation output
-#
-# Categories assigned to bounding boxes.  Used throughout our repo; do not change unless
-# you are Dan or Siyu.  In fact, do not change unless you are both Dan *and* Siyu.
-# 
-# We use integer indices here; this is different than the API output .json file,
-# where indices are string integers.
-#
-########
-
-NUM_DETECTOR_CATEGORIES = 3  # this is for choosing colors, so ignoring the "empty" class
-
-# This is the label mapping used for our incoming iMerit annotations
-# Only used to parse the incoming annotations. In our database, the string name is used to avoid confusion
-annotation_bbox_categories = [
-    {'id': 0, 'name': 'empty'},
-    {'id': 1, 'name': 'animal'},
-    {'id': 2, 'name': 'person'},
-    {'id': 3, 'name': 'group'},  # group of animals
-    {'id': 4, 'name': 'vehicle'}
-]
+"""
+
+annotation_constants.py
 
-annotation_bbox_category_id_to_name = {}
-annotation_bbox_category_name_to_id = {}
+Defines default categories for MegaDetector output boxes.
 
-for cat in annotation_bbox_categories:
-    annotation_bbox_category_id_to_name[cat['id']] = cat['name']
-    annotation_bbox_category_name_to_id[cat['name']] = cat['id']
+Used throughout the repo; do not change unless you are Dan or Siyu.  In fact, do not change unless 
+you are both Dan *and* Siyu.
 
-# MegaDetector outputs
+We use integer IDs here; this is different from the MD .json file format,
+where indices are string integers.
+
+"""
+
+#%% Constants
+
+# MegaDetector output categories (the "empty" category is implicit)
 detector_bbox_categories = [
     {'id': 0, 'name': 'empty'},
     {'id': 1, 'name': 'animal'},
@@ -39,9 +22,13 @@ detector_bbox_categories = [
     {'id': 3, 'name': 'vehicle'}
 ]
 
+# This is used for choosing colors, so it ignores the "empty" class.
+NUM_DETECTOR_CATEGORIES = len(detector_bbox_categories) - 1
+
 detector_bbox_category_id_to_name = {}
 detector_bbox_category_name_to_id = {}
 
 for cat in detector_bbox_categories:
     detector_bbox_category_id_to_name[cat['id']] = cat['name']
     detector_bbox_category_name_to_id[cat['name']] = cat['id']
+

data_management/camtrap_dp_to_coco.py

@@ -0,0 +1,238 @@
+"""
+
+camtrap_dp_to_coco.py
+
+Parse a very limited subset of the Camtrap DP data package format:
+
+https://camtrap-dp.tdwg.org/
+
+...and convert to COCO format.  Assumes that all required metadata files have been
+put in the same directory (which is standard).
+
+Does not currently parse bounding boxes, just attaches species labels to images.
+
+Currently supports only sequence-level labeling.
+
+"""
+
+#%% Imports and constants
+
+import os
+import json
+import pandas as pd
+
+from dateutil import parser as dateparser
+
+from collections import defaultdict
+
+
+#%% Functions
+
+def camtrap_dp_to_coco(camtrap_dp_folder,output_file=None):
+    """
+    Convert the Camtrap DP package in [camtrap_dp_folder] to COCO.
+    
+    Does not validate images, just converts.  Use integrity_check_json_db to validate
+    the resulting COCO file.  
+    
+    Optionally writes the results to [output_file]
+    """
+    
+    required_files = ('datapackage.json','deployments.csv','events.csv','media.csv','observations.csv')
+    
+    for fn in required_files:
+        fn_abs = os.path.join(camtrap_dp_folder,fn)
+        assert os.path.isfile(fn_abs), 'Could not find required file {}'.format(fn_abs)
+        
+    with open(os.path.join(camtrap_dp_folder,'datapackage.json'),'r') as f:
+        datapackage = json.load(f)
+        
+    assert datapackage['profile'] == 'https://raw.githubusercontent.com/tdwg/camtrap-dp/1.0/camtrap-dp-profile.json', \
+        'I only know how to parse Camtrap DP 1.0 packages'
+
+    deployments_file = None
+    events_file = None
+    media_file = None
+    observations_file = None
+    
+    resources = datapackage['resources']
+    for r in resources:
+        if r['name'] == 'deployments':
+            deployments_file = r['path']
+        elif r['name'] == 'media':
+            media_file = r['path']
+        elif r['name'] == 'events':
+            events_file = r['path']
+        elif r['name'] == 'observations':
+            observations_file = r['path']
+
+    assert deployments_file is not None, 'No deployment file specified'
+    assert events_file is not None, 'No events file specified'
+    assert media_file is not None, 'No media file specified'
+    assert observations_file is not None, 'No observation file specified'
+    
+    deployments_df = pd.read_csv(os.path.join(camtrap_dp_folder,deployments_file))
+    events_df = pd.read_csv(os.path.join(camtrap_dp_folder,events_file))
+    media_df = pd.read_csv(os.path.join(camtrap_dp_folder,media_file))
+    observations_df = pd.read_csv(os.path.join(camtrap_dp_folder,observations_file))
+    
+    print('Read {} deployment lines'.format(len(deployments_df)))
+    print('Read {} events lines'.format(len(events_df)))
+    print('Read {} media lines'.format(len(media_df)))
+    print('Read {} observation lines'.format(len(observations_df)))
+    
+    media_id_to_media_info = {}
+    
+    # i_row = 0; row = media_df.iloc[i_row]
+    for i_row,row in media_df.iterrows():
+        media_info = {}
+        media_info['file_name'] = os.path.join(row['filePath'],row['fileName']).replace('\\','/')
+        media_info['location'] = row['deploymentID']
+        media_info['id'] = row['mediaID']
+        media_info['datetime'] = row['timestamp']
+        media_info['datetime'] = dateparser.parse(media_info['datetime'])
+        media_info['frame_num'] = -1
+        media_info['seq_num_frames'] = -1
+        media_id_to_media_info[row['mediaID']] = media_info
+        
+    event_id_to_media_ids = defaultdict(list)
+    
+    # i_row = 0; row = events_df.iloc[i_row]
+    for i_row,row in events_df.iterrows():
+        media_id = row['mediaID']
+        assert media_id in media_id_to_media_info
+        event_id_to_media_ids[row['eventID']].append(media_id)
+    
+    event_id_to_category_names = defaultdict(set)
+    
+    # i_row = 0; row = observations_df.iloc[i_row]
+    for i_row,row in observations_df.iterrows():
+        
+        if row['observationLevel'] != 'event':
+            raise ValueError("I don't know how to parse image-level events yet")
+            
+        if row['observationType'] == 'blank':
+            event_id_to_category_names[row['eventID']].add('empty')
+        elif row['observationType'] == 'unknown':
+            event_id_to_category_names[row['eventID']].add('unknown')
+        elif row['observationType'] == 'human':
+            assert row['scientificName'] == 'Homo sapiens'
+            event_id_to_category_names[row['eventID']].add(row['scientificName'])
+        else:
+            assert row['observationType'] == 'animal'
+            assert isinstance(row['scientificName'],str)
+            event_id_to_category_names[row['eventID']].add(row['scientificName'])
+    
+    # Sort images within an event into frame numbers
+    #
+    # event_id = next(iter(event_id_to_media_ids))
+    for event_id in event_id_to_media_ids.keys():
+        media_ids_this_event = event_id_to_media_ids[event_id]
+        media_info_this_event = [media_id_to_media_info[media_id] for media_id in media_ids_this_event]
+        media_info_this_event = sorted(media_info_this_event, key=lambda x: x['datetime'])
+        for i_media,media_info in enumerate(media_info_this_event):
+            media_info['frame_num'] = i_media
+            media_info['seq_num_frames'] = len(media_info_this_event)
+            media_info['seq_id'] = event_id
+            
+    # Create category names
+    category_name_to_category_id = {'empty':0}
+    for event_id in event_id_to_category_names:
+        category_names_this_event = event_id_to_category_names[event_id]
+        for name in category_names_this_event:
+            if name not in category_name_to_category_id:
+                category_name_to_category_id[name] = len(category_name_to_category_id)
+    
+    # Move everything into COCO format
+    images = list(media_id_to_media_info.values())
+    
+    categories = []
+    for name in category_name_to_category_id:
+        categories.append({'name':name,'id':category_name_to_category_id[name]})
+    info = {'version':1.0,'description':datapackage['name']}
+    
+    # Create annotations
+    annotations = []
+    
+    for event_id in event_id_to_media_ids.keys():
+        i_ann = 0
+        media_ids_this_event = event_id_to_media_ids[event_id]
+        media_info_this_event = [media_id_to_media_info[media_id] for media_id in media_ids_this_event]
+        categories_this_event = event_id_to_category_names[event_id]
+        for im in media_info_this_event:
+            for category_name in categories_this_event:
+                ann = {}
+                ann['id'] = event_id + '_' + str(i_ann)
+                i_ann += 1
+                ann['image_id'] = im['id']
+                ann['category_id'] = category_name_to_category_id[category_name]
+                ann['sequence_level_annotation'] = True
+                annotations.append(ann)
+    
+    coco_data = {}
+    coco_data['images'] = images
+    coco_data['annotations'] = annotations
+    coco_data['categories'] = categories
+    coco_data['info'] = info
+    
+    for im in coco_data['images']:
+        im['datetime'] = str(im['datetime'] )
+        
+    if output_file is not None:
+        with open(output_file,'w') as f:
+            json.dump(coco_data,f,indent=1)
+    
+    return coco_data
+            
+    
+#%% Interactive driver
+
+if False:
+
+    pass
+
+    #%%
+    
+    camtrap_dp_folder = r'C:\temp\pilot2\pilot2'
+    coco_file = os.path.join(camtrap_dp_folder,'test-coco.json')
+    coco_data = camtrap_dp_to_coco(camtrap_dp_folder,
+                                   output_file=coco_file)
+    
+    #%% Validate
+    
+    from data_management.databases.integrity_check_json_db import integrity_check_json_db, IntegrityCheckOptions
+    
+    options = IntegrityCheckOptions()
+    
+    options.baseDir = camtrap_dp_folder
+    options.bCheckImageSizes = False
+    options.bCheckImageExistence = True
+    options.bFindUnusedImages = True
+    options.bRequireLocation = True
+    options.iMaxNumImages = -1
+    options.nThreads = 1
+    options.verbose = True
+    
+    sortedCategories, data, errorInfo = integrity_check_json_db(coco_file,options)
+
+    #%% Preview
+    
+    from md_visualization.visualize_db import DbVizOptions, visualize_db
+    
+    options = DbVizOptions()
+    options.parallelize_rendering = True
+    options.parallelize_rendering_with_threads = True
+    options.parallelize_rendering_n_cores = 10
+    
+    preview_dir = r'c:\temp\camtrapdp-preview'
+    htmlOutputFile,image_db = visualize_db(coco_file, preview_dir, camtrap_dp_folder, options=options)
+    
+    from md_utils.path_utils import open_file
+    open_file(htmlOutputFile)
+    
+    
+#%% Command-line driver
+
+# TODO
+
+    

data_management/cct_json_utils.py

@@ -1,12 +1,12 @@
-########
-#
-# cct_json_utils.py
-#
-# Utilities for working with COCO Camera Traps .json databases
-#
-# https://github.com/agentmorris/MegaDetector/blob/master/data_management/README.md#coco-cameratraps-format
-#
-########
+"""
+
+cct_json_utils.py
+
+Utilities for working with COCO Camera Traps .json databases
+
+https://github.com/agentmorris/MegaDetector/blob/master/data_management/README.md#coco-cameratraps-format
+
+"""
 
 #%% Constants and imports
 
@@ -15,9 +15,6 @@ import os
 
 from tqdm import tqdm
 from collections import defaultdict, OrderedDict
-from typing import Any, Dict, Iterable, List, Mapping, Optional, Union
-
-JSONObject = Mapping[str, Any]
 
 
 #%% Classes
@@ -31,18 +28,32 @@ class CameraTrapJsonUtils:
     def annotations_to_string(annotations, cat_id_to_name):
         """
         Given a list of annotations and a mapping from class IDs to names, produces
-        a concatenated class list, always sorting alphabetically.
+        a comma-delimited string containing a list of class names, sorted alphabetically.
+        
+        Args:
+            annotations (list): a list of annotation dicts
+            cat_id_to_name (dict): a dict mapping category IDs to category names
+            
+        Returns:
+            str: a comma-delimited list of class names
         """
         
-        class_names = CameraTrapJsonUtils.annotationsToClassnames(annotations, cat_id_to_name)
+        class_names = CameraTrapJsonUtils.annotations_to_class_names(annotations, cat_id_to_name)
         return ','.join(class_names)
 
 
     @staticmethod
-    def annotations_to_classnames(annotations, cat_id_to_name):
+    def annotations_to_class_names(annotations, cat_id_to_name):
         """
         Given a list of annotations and a mapping from class IDs to names, produces
-        a list of class names, always sorting alphabetically.
+        a list of class names, sorted alphabetically.
+        
+        Args:
+            annotations (list): a list of annotation dicts
+            cat_id_to_name (dict): a dict mapping category IDs to category names
+            
+        Returns:
+            list: a list of class names present in [annotations]
         """
         
         # Collect all names
@@ -53,18 +64,19 @@ class CameraTrapJsonUtils:
 
 
     @staticmethod
-    def order_db_keys(db: JSONObject) -> OrderedDict:
+    def order_db_keys(db):
         """
         Given a dict representing a JSON database in the COCO Camera Trap
-        format, return an OrderedDict with keys in the order of 'info',
+        format, returns an OrderedDict with keys in the order of 'info',
         'categories', 'annotations' and 'images'. When this OrderedDict is
         serialized with json.dump(), the order of the keys are preserved.
 
         Args:
-            db: dict representing a JSON database in the COCO Camera Trap format
+            db (dict): a JSON database in the COCO Camera Trap format
 
         Returns:
-            the same db but as an OrderedDict with keys ordered for readability
+            dict: the same content as [db] but as an OrderedDict with keys ordered for 
+                readability
         """
         
         ordered = OrderedDict([
@@ -76,10 +88,20 @@ class CameraTrapJsonUtils:
 
 
     @staticmethod
-    def annotations_groupby_image_field(db_indexed, image_field='seq_id'):
+    def group_annotations_by_image_field(db_indexed, image_field='seq_id'):
         """
         Given an instance of IndexedJsonDb, group annotation entries by a field in the
-        image entry.
+        image entry.  Typically used to find all the annotations associated with a sequence.
+        
+        Args:
+            db_indexed (IndexedJsonDb): an initialized IndexedJsonDb, typically loaded from a 
+                COCO Camera Traps .json file
+            image_field (str, optional): a field by which to group annotations (defaults
+                to 'seq_id')
+        
+        Returns:
+            dict: a dict mapping objects (typically strings, in fact typically sequence IDs) to
+                lists of annotations
         """
         
         image_id_to_image_field = {}
@@ -95,27 +117,24 @@ class CameraTrapJsonUtils:
 
 
     @staticmethod
-    def get_entries_from_locations(db: JSONObject, locations: Iterable[str]
-                                   ) -> Dict[str, Any]:
+    def get_entries_for_locations(db, locations):
         """
-        Given a dict representing a JSON database in the COCO Camera Trap format, return a dict
+        Given a dict representing a JSON database in the COCO Camera Trap format, returns a dict
         with the 'images' and 'annotations' fields in the CCT format, each is an array that only
-        includes entries in the original `db` that are in the `locations` set.
+        includes entries in the original [db] that are in the [locations] set.
+        
         Args:
-            db: a dict representing a JSON database in the COCO Camera Trap format
-            locations: a set or list of locations to include; each item is a string
+            db (dict): a dict representing a JSON database in the COCO Camera Trap format
+            locations (set): a set or list of locations to include; each item is a string
 
         Returns:
-            a dict with the 'images' and 'annotations' fields in the CCT format
+            dict: a dict with the 'images' and 'annotations' fields in the CCT format
         """
         
         locations = set(locations)
         print('Original DB has {} image and {} annotation entries.'.format(
             len(db['images']), len(db['annotations'])))
-        new_db: Dict[str, Any] = {
-            'images': [],
-            'annotations': []
-        }
+        new_db = { 'images': [], 'annotations': [] }
         new_images = set()
         for i in db['images']:
             # cast location to string as the entries in locations are strings
@@ -139,12 +158,26 @@ class IndexedJsonDb:
     a .json database.
     """
 
-    def __init__(self, json_filename: Union[str, JSONObject],
-                 b_normalize_paths: bool = False,
-                 filename_replacements: Optional[Mapping[str, str]] = None,
-                 b_convert_classes_to_lower: bool = True):
+    def __init__(self, 
+                 json_filename,
+                 b_normalize_paths=False,
+                 filename_replacements=None,
+                 b_convert_classes_to_lower=True,
+                 b_force_forward_slashes=True):
         """
-        json_filename can also be an existing json db
+        Constructor for IndexedJsonDb that loads from a .json file or CCT-formatted dict.
+        
+        Args:
+            json_filename (str): filename to load, or an already-loaded dict
+            b_normalize_paths (bool, optional): whether to invoke os.path.normpath on 
+                all filenames.  Not relevant if b_force_forward_slashes is True.
+            filename_replacements (dict, optional): a set of string --> string mappings
+                that will trigger replacements in all filenames, typically used to remove
+                leading folders
+            b_convert_classes_to_lower (bool, optional): whether to convert all class
+                names to lowercase
+            b_force_forward_slashes (bool, optional): whether to convert backslashes to
+                forward slashes in all path names            
         """
         
         if isinstance(json_filename, str):
@@ -162,11 +195,15 @@ class IndexedJsonDb:
             for c in self.db['categories']:
                 c['name'] = c['name'].lower()
 
+        # Normalize paths to simplify comparisons later
         if b_normalize_paths:
-            # Normalize paths to simplify comparisons later
             for im in self.db['images']:
                 im['file_name'] = os.path.normpath(im['file_name'])
 
+        if b_force_forward_slashes:
+            for im in self.db['images']:
+                im['file_name'] = im['file_name'].replace('\\','/')
+
         if filename_replacements is not None:
             for s in filename_replacements:
                 # Make custom replacements in filenames, typically used to
@@ -192,7 +229,7 @@ class IndexedJsonDb:
 
         # Image ID --> annotations
         # Each image can potentially multiple annotations, hence using lists
-        self.image_id_to_annotations: Dict[str, List[Dict[str, Any]]]
+        self.image_id_to_annotations = {}
         self.image_id_to_annotations = defaultdict(list)
         for ann in self.db['annotations']:
             self.image_id_to_annotations[ann['image_id']].append(ann)
@@ -200,12 +237,17 @@ class IndexedJsonDb:
     # ...__init__
 
 
-    def get_annotations_for_image(self, image: JSONObject
-                                  ) -> Optional[List[Dict[str, Any]]]:
+    def get_annotations_for_image(self, image):
         """
-        Returns: list of annotations associated with an image,
-            None if the db has not been loaded,
-            [] if no annotations are available
+        Finds all the annnotations associated with the image dict [image].
+        
+        Args:
+            image (dict): an image dict loaded from a CCT .json file.  Only the 'id' field
+                is used.
+            
+        Returns:
+            list: list of annotations associated with this image.  Returns None if the db 
+                has not been loaded, or [] if no annotations are available for this image.
         """
         
         if self.db is None:
@@ -218,11 +260,17 @@ class IndexedJsonDb:
         return image_annotations
 
 
-    def get_classes_for_image(self, image: JSONObject) -> Optional[List[str]]:
+    def get_classes_for_image(self, image):
         """
-        Returns a list of class names associated with [image]
+        Returns a list of class names associated with [image].
 
-        Returns None is the db has not been loaded, [] if no annotations are available
+        Args:
+            image (dict): an image dict loaded from a CCT .json file.  Only the 'id' field
+                is used.
+            
+        Returns:
+            list: list of class names associated with this image.  Returns None if the db 
+                has not been loaded, or [] if no annotations are available for this image.
         """
         
         if self.db is None:
@@ -242,27 +290,27 @@ class IndexedJsonDb:
 
 # ...class IndexedJsonDb
 
-
-#%% Functions
-
 class SequenceOptions:
+    """
+    Options parameterizing the grouping of images into sequences by time.
+    """
     
+    #: Images separated by <= this duration will be grouped into the same sequence.
     episode_interval_seconds = 60.0
 
     
+#%% Functions
+
 def create_sequences(image_info,options=None):
     """
-    Synthesize episodes/sequences/bursts for the images in [image_info].  [image_info]
-    should be a list of dicts in CCT format, i.e. with fields 'file_name','datetime','location'.
-    
-    'filename' should be a string.
-    
-    'datetime' should be a Python datetime object
-    
-    'location' should be a string.
+    Synthesizes episodes/sequences/bursts for the images in [image_info].
     
     Modifies [image_info], populating the 'seq_id', 'seq_num_frames', and 'frame_num' fields
     for each image.
+    
+    Args:
+        image_info (dict): a list of dicts in CCT format, i.e. with fields 'file_name' (str),
+            'datetime' (datetime), and 'location' (str).    
     """
     
     if options is None: