Transfer learning

conf/base/parameters.yml

@@ -43,13 +43,13 @@ data_renaming: # TODO: remove later when renaming finished
 
 modelling:
   # obj detection
-  detection_model: "yolov3-tiny"
+  detection_model: "yolov3_tf"
   detection_implementation: "cvlib"
   detection_iou_threshold: 0.05
   detection_confidence_threshold: 0.2
-  # TODO: change nms threshold to iou threshold	
+  # TODO: change nms threshold to iou threshold
   detection_nms_threshold: 0.2
-  
+
   # tracking
   selected_labels: ["car", "truck", "bus", "motorbike"]
   opencv_tracker_type: "csrt"

conf/base/paths.yml

@@ -7,15 +7,21 @@ s3_paths:
   s3_camera_details: "ref/camera_details/camera_details.json"
 
   s3_frame_level: "frame_level/" # TODO DELETE THIS
-  s3_profile: "dssg" # TODO: change this for user? 
+  s3_profile: "dssg" # TODO: change this for user?
   s3_creds: "dev_s3" # TODO: CHANGE TO JUST S3
   s3_detection_model: "ref/model_conf/"
+  s3_cvat_annotations: "ref/annotations/cvat/"
+  s3_detrac_annotations: "ref/annotations/detrac/"
+  s3_detrac_images: "raw/images/detrac/"
+  s3_cvat_training_annotations: "ref/annotations/cvat_train/"
 
 local_paths:
   temp_video: "data/temp/videos/"
+  temp_raw_images: "data/temp/raw_images/"
   temp_raw_video: "data/temp/raw_videos/"
   temp_frame_level: "data/temp/frame_level/"
   temp_video_level: "data/temp/video_level/"
+  temp_annotation: "data/temp/annotation/"
   temp_setup: "data/temp/setup/"
 
   video_names: "data/ref/video_names/"

conf/base/training_parameters.yml

@@ -0,0 +1,34 @@
+training:
+  num_batches : 10
+  letterbox_resize : True  # Whether to use letterbox resize, i.e., keep the original aspect ratio in the resized img.
+  total_epochs : 1000
+  train_evaluation_step : 100  # Evaluate on the training batch after some steps.
+  val_evaluation_epoch : 2  # Evaluate on the validation dataset after some epochs. Set to None to evaluate all epoch.
+  save_epoch : 10  # Save the model after some epochs.
+  batch_norm_decay : 0.99  # decay in bn ops
+  weight_decay : 0.0005  # l2 weight decay
+  global_step : 0  # used when resuming training
+  warm_up_epoch : 3 # set to larger value if gradient explodes
+  num_threads : 10  # Number of threads for image processing used in tf.data pipeline.
+  prefetech_buffer : 5  # Prefetech_buffer used in tf.data pipeline.
+  trained_model_name : 'yolov3_traffic'
+
+learning:
+  optimizer_name : 'momentum'  # Chosen from [sgd, momentum, adam, rmsprop]
+  save_optimizer : True  # Whether to save the optimizer parameters into the checkpoint file.
+  learning_rate_init : 0.0001
+  lr_type : 'piecewise'  # Chosen from [exponential, piecewise]
+  lr_decay_epoch : 5  # Epochs after which learning rate decays. Int or float. Used when chosen `exponential` lr_type.
+  lr_decay_factor : 0.96  # The learning rate decay factor. Used when chosen `exponential` lr_type.
+  lr_lower_bound : 0.000001  # The minimum learning rate.
+  pw_boundaries : [30, 50]  # epoch based boundaries
+  pw_values : [0.0001, 0.00003, 0.00001] # FIRST VALUE MUST BE LEARNING_RATE_INIT
+
+validation:
+  # nms
+  nms_threshold : 0.45  # iou threshold in nms operation
+  score_threshold : 0.01  # threshold of the prob of the classes in nms operation, i.e. score = pred_confs * pred_probs.
+  nms_topk : 150  # keep at most nms_topk outputs after nms
+
+  # mAP eval
+  eval_threshold : 0.5  # the iou threshold applied in mAP evaluation

requirements.txt

@@ -20,3 +20,4 @@ sqlalchemy
 pandas==0.24.2
 python-crontab>=2.3.8, <3.0
 seaborn>=0.9
+tqdm==4.33.0

src/run_transfer_learning.py

@@ -0,0 +1,33 @@
+from traffic_analysis.d00_utils.load_confs import load_paths, load_credentials, \
+    load_parameters, load_training_parameters
+from traffic_analysis.d04_modelling.transfer_learning.training_data_loader import TrainingDataLoader, TransferDataset
+from traffic_analysis.d04_modelling.transfer_learning.train_tensorflow_model import transfer_learn
+
+paths = load_paths()
+creds = load_credentials()
+params = load_parameters()
+train_params = load_training_parameters()
+
+training_data_loader = TrainingDataLoader(datasets=[TransferDataset.cvat, TransferDataset.detrac],
+                                          creds=creds,
+                                          paths=paths)
+
+fraction_for_training = 0.8
+x_train, y_train, x_test, y_test = training_data_loader.get_train_and_test(fraction_for_training)
+
+
+saved_text_files_dir = paths['temp_annotation']
+with open(saved_text_files_dir + 'train.txt', 'w') as f:
+    for item in y_train:
+        f.write("%s\n" % item)
+
+with open(saved_text_files_dir + 'test.txt', 'w') as f:
+    for item in y_test:
+        f.write("%s\n" % item)
+
+transfer_learn(paths=paths,
+               params=params,
+               train_params=train_params,
+               train_file='train.txt',
+               test_file='test.txt',
+               selected_labels=params['selected_labels'])

src/traffic_analysis/d00_utils/load_confs.py

@@ -30,6 +30,12 @@ def load_app_parameters():
     return {**params['visualization']}
 
 
+def load_training_parameters():
+    with open(project_dir + '/conf/base/training_parameters.yml') as f:
+        params = yaml.safe_load(f)
+    return collapse_dict_hierarchy(params)
+
+
 def load_credentials():
 
     filepath = os.sep.join(

src/traffic_analysis/d02_ref/ref_utils.py

@@ -5,6 +5,8 @@
 import subprocess
 from subprocess import Popen, PIPE
 
+from traffic_analysis.d00_utils.data_loader_s3 import DataLoaderS3
+
 
 def upload_json_to_s3(paths: dict,
                       save_name: str,
@@ -66,4 +68,48 @@ def get_names_of_folder_content_from_s3(bucket_name, prefix, s3_profile):
     end = Time.time()
     elapsed_time = end-start
 
+    assert ((len(files) == 0) or (files[0] != '')), 'set your aws credentials'
+
     return elapsed_time, files
+
+
+def get_s3_video_path_from_xml_name(xml_file_name, s3_creds, paths):
+
+    # Supports old and new naming conventions
+    vals = xml_file_name.split('_')
+    data_loader_s3 = DataLoaderS3(s3_credentials=s3_creds,
+                                  bucket_name=paths['bucket_name'])
+
+    if (len(vals) >= 4):
+        date = vals[1]
+        file_names = [xml_file_name.split('_')[1:][0].replace('-', '') + '-' +
+                      xml_file_name.split('_')[1:][1].replace('-', '')[:6] + '_' +
+                      xml_file_name.split('_')[1:][2],
+                      xml_file_name.split('_')[1:][0] + ' ' +
+                      xml_file_name.split('_')[1:][1].replace('-', ':') + '_' +
+                      xml_file_name.split('_')[1:][2]]
+    else:
+        date = vals[0]
+        file_names = [xml_file_name.split('_')[0].replace('-', '') + '-' +
+                      xml_file_name.split('_')[1].replace('-', '')[:6] + '_' +
+                      xml_file_name.split('_')[2],
+                      xml_file_name.split('_')[0] + ' ' +
+                      xml_file_name.split('_')[1].replace('-', ':') + '_' +
+                      xml_file_name.split('_')[2]]
+    file_to_download = paths['s3_video'] + \
+                       date + '/' + \
+                       file_names[0] + '.mp4'
+
+    if(data_loader_s3.file_exists(file_to_download)):
+        return file_to_download
+
+    else:
+        file_to_download = paths['s3_video'] + \
+                           date + '/' + \
+                           file_names[1] + '.mp4'
+
+        if (data_loader_s3.file_exists(file_to_download)):
+            return file_to_download
+        else:
+            print('Could not download file: ' + xml_file_name)
+            return

src/traffic_analysis/d02_ref/upload_annotation_names_to_s3.py

@@ -1,5 +1,6 @@
 from traffic_analysis.d02_ref.ref_utils import get_names_of_folder_content_from_s3
 from traffic_analysis.d00_utils.data_loader_s3 import DataLoaderS3
+from traffic_analysis.d02_ref.ref_utils import get_s3_video_path_from_xml_name
 
 
 def upload_annotation_names_to_s3(paths: dict,

src/traffic_analysis/d04_modelling/perform_detection_tensorflow.py

@@ -9,7 +9,7 @@
     remove_overlapping_boxes, letterbox_resize
 from traffic_analysis.d04_modelling.transfer_learning.convert_darknet_to_tensorflow import parse_anchors, \
     yolov3_darknet_to_tensorflow
-from traffic_analysis.d04_modelling.transfer_learning.generate_tensorflow_model import YoloV3
+from traffic_analysis.d04_modelling.transfer_learning.tensorflow_model_loader import YoloV3
 from traffic_analysis.d04_modelling.perform_detection_opencv import label_detections, \
     choose_objects_of_selected_labels
 

src/traffic_analysis/d04_modelling/transfer_learning/convert_darknet_to_tensorflow.py

@@ -8,7 +8,7 @@
 import tensorflow as tf
 import numpy as np
 
-from traffic_analysis.d04_modelling.transfer_learning.generate_tensorflow_model import YoloV3
+from traffic_analysis.d04_modelling.transfer_learning.tensorflow_model_loader import YoloV3
 from traffic_analysis.d02_ref.download_detection_model_from_s3 import download_detection_model_from_s3