edited cleaner

csv_img_to_tfrecord.py

@@ -71,7 +71,7 @@ def create_tf_example(group, path, label_map_dict):
 def main():
     for directory in ['train','eval']:
 
-        image_path = os.path.join(os.getcwd(), 'data/{}/images'.format(directory))
+        image_path = os.path.join(os.getcwd(), 'data/{}'.format(directory))
         csv_path = os.path.join(os.getcwd(), 'data/{}_labels.csv'.format(directory,directory))
         output_path = os.path.join(os.getcwd(), 'data/{}.record'.format(directory))
         label_map_dict = label_map_util.get_label_map_dict(os.path.join(os.getcwd(), 'data/label_map.pbtxt'))

egohands_dataset_cleaner.py

@@ -236,6 +236,10 @@ def download_egohands_dataset(dataset_url, dataset_path):
             rename_files("egohands/_LABELLED_SAMPLES/")
 
 def final_finish():
+    f = open("data/label_map.pbtxt","w") 
+    f.write("item {\n  id: 1\n  name: 'hand'\n}")
+    f.close()
+
     sh.rmtree('egohands')
     os.remove(EGO_HANDS_FILE)
     print('> creating the dataset complete\n\

howto_wiki

@@ -10,8 +10,8 @@ export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim
 # From the tensorflow/models/research directory
 python object_detection/train.py \
     --logtostderr \
-    --pipeline_config_path=/home/gustav/workspace/training_hands/model/ssd_mobilehandsnet.config \
-    --train_dir=/home/gustav/workspace/training_hands/model/train
+    --pipeline_config_path=/home/nvidia/workspace/deeptraining_hands/model/ssd_mobilehandsnet.config \
+    --train_dir=/home/nvidia/workspace/deeptraining_hands/model/train
 
 ## EVALUATION
 # From the tensorflow/models/research directory
@@ -58,8 +58,8 @@ Train on a single GPU and eval on CPU, In object_detection/trainer.py
 add after "session_config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=False)" :
 
 session_config.gpu_options.allow_growth = True
-session_config.gpu_options.per_process_gpu_memory_fraction = 0.8 #optional
-session_config.gpu_options.allocator_type ='BFC' # optional
+session_config.gpu_options.per_process_gpu_memory_fraction = 0.9
+session_config.gpu_options.allocator_type ='BFC'
 
 For Evaluation Using only CPU: In object_detection/eval.py add behind "FLAGS = flags.FLAG" :
 

model/ssd_mobilehandsnet.config

@@ -155,7 +155,7 @@ train_config: {
       epsilon: 1.0
     }
   }
-  fine_tune_checkpoint: "/home/gustav/workspace/training_hands/model/ssd_mobilenet_checkpoints/model.ckpt"
+  fine_tune_checkpoint: "/home/nvidia/workspace/deeptraining_hands/model/ssd_mobilenet_checkpoints/model.ckpt"
   from_detection_checkpoint: true
   # Note: The below line limits the training process to 200K steps, which we
   # empirically found to be sufficient enough to train the pets dataset. This
@@ -174,24 +174,24 @@ train_config: {
 
 train_input_reader: {
   tf_record_input_reader {
-    input_path: "/home/gustav/workspace/training_hands/data/train.record"
+    input_path: "/home/nvidia/workspace/deeptraining_hands/data/train.record"
   }
-  label_map_path: "/home/gustav/workspace/training_hands/data/label_map.pbtxt"
+  label_map_path: "/home/nvidia/workspace/deeptraining_hands/data/label_map.pbtxt"
 }
 
 eval_config: {
-  num_examples: 1
-  num_visualizations: 1
+  num_examples: 400
+  num_visualizations: 3
   # Note: The below line limits the evaluation process to 10 evaluations.
   # Remove the below line to evaluate indefinitely.
   max_evals: 10
 }
 
 eval_input_reader: {
   tf_record_input_reader {
-    input_path: "/home/gustav/workspace/training_hands/data/eval.record"
+    input_path: "/home/nvidia/workspace/deeptraining_hands/data/eval.record"
   }
-  label_map_path: "/home/gustav/workspace/training_hands/data/label_map.pbtxt"
+  label_map_path: "/home/nvidia/workspace/deeptraining_hands/data/label_map.pbtxt"
   shuffle: false
   num_readers: 1
   num_epochs: 1