2022.09.09 21:06

A/config.yaml

@@ -0,0 +1,40 @@
+DATA:
+  BATCH_SIZE: 8
+  FILTER:
+    MAX_DEPTH: 65
+    MAX_OCCLUSION: 2
+    MAX_TRUNCATION: 0.5
+    MIN_DEPTH: 2
+    MIN_HEIGHT: 25
+  NUM_WORKERS: 4
+  ROOT: /home/user/SSD/KITTI
+  TEST_SPLIT: val
+  TRAIN_SPLIT: train
+DESCRIPTION: MonoCon Default Configuration
+GPU_ID: 0
+MODEL:
+  BACKBONE:
+    IMAGENET_PRETRAINED: true
+    NUM_LAYERS: 34
+  HEAD:
+    MAX_OBJS: 30
+    NUM_CLASSES: 3
+OUTPUT_DIR: A
+PERIOD:
+  EVAL_PERIOD: 10
+  LOG_PERIOD: 50
+SEED: 0
+SOLVER:
+  CLIP_GRAD:
+    ENABLE: true
+    MAX_NORM: 35
+    NORM_TYPE: 2.0
+  OPTIM:
+    LR: 0.000225
+    NUM_EPOCHS: 200
+    WEIGHT_DECAY: 1.0e-05
+  SCHEDULER:
+    ENABLE: true
+USE_BENCHMARK: false
+VERSION: v1.0.3
+

README.MD

@@ -10,8 +10,10 @@ Unchecked items are currently being prepared.
 - [x] Single-GPU Training
 - [x] KITTI Evaluation
 - [ ] Multi-GPU Training
+- [ ] Fancy Logger
 - [x] Visualization (2D Bounding Box + Projected 3D Bounding Box)
 - [x] Visualization (Bird Eye's View)
+- [ ] Video Inference using KITTI Raw Data Sequences
 
 
 ## Preparations
@@ -79,6 +81,7 @@ The structure of the data files should be as below.
 ## Usage
 ### Training
 Just edit the items in ```config/monocon_configs.py``` before execution.  
+If you want less GPU memory consumption, set ```_C.USE_BENCHMARK``` to ```False```.
 ```bash
 python train.py
 ```
@@ -159,7 +162,7 @@ If possible, use a value greater than 4 as the batch size.
 
 
 ## Change Log
-This repository was last updated to **v1.0.2** on **2022.08.30**.  
+This repository was last updated to **v1.0.3** on **2022.09.09**.  
 Check [changelog.md](changelog.MD) for detailed update history.
 
 

changelog.MD

@@ -1,5 +1,8 @@
 ## Changelog
 
+### v1.0.3 (2022.09.09)
+- Made ```torch.backend.cudnn.benchmark``` option configurable in ```monocon_config.py```.
+
 ### v1.0.2 (2022.08.30)
 - A Transform class ```Resize3D``` for resize augmentation has been added. (**Issue #1**)
 

config/monocon_configs.py

@@ -3,12 +3,14 @@
 
 _C = CN()
 
-_C.VERSION = 'v1.0.2'
+_C.VERSION = 'v1.0.3'
 _C.DESCRIPTION = "MonoCon Default Configuration"
 
-_C.OUTPUT_DIR = ""                              # Output Directory
-_C.SEED = -1                                    # -1: Random Seed Selection
-_C.GPU_ID = 0                                   # Index of GPU to use
+_C.OUTPUT_DIR = ""                               # Output Directory
+_C.SEED = -1                                     # -1: Random Seed Selection
+_C.GPU_ID = 0                                    # Index of GPU to use
+
+_C.USE_BENCHMARK = True                          # Value of 'torch.backends.cudnn.benchmark' and 'torch.backends.cudnn.enabled'
 
 
 # Data
@@ -19,10 +21,6 @@
 _C.DATA.TRAIN_SPLIT = 'train'
 _C.DATA.TEST_SPLIT = 'val'
 
-_C.DATA.STATS = CN()
-_C.DATA.STATS.MEAN = [123.675, 116.28, 103.53]
-_C.DATA.STATS.STD = [58.395, 57.12, 57.375]
-
 _C.DATA.FILTER = CN()
 _C.DATA.FILTER.MIN_HEIGHT = 25
 _C.DATA.FILTER.MIN_DEPTH = 2

losses/l1_loss.py

@@ -9,6 +9,7 @@
 from losses.utils import weighted_loss
 
 
+# TODO Handling an exception so that it can be handled even when the target data is empty
 @weighted_loss
 def l1_loss(pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
     assert pred.size() == target.size() and target.numel() > 0

test.py

@@ -32,15 +32,20 @@
     torch.backends.cuda.matmul.allow_tf32 = False
     torch.backends.cudnn.allow_tf32 = False
 
-torch.backends.cudnn.enabled = True
-torch.backends.cudnn.benchmark = True
-
 
 # Load Config
 cfg = load_cfg(args.config_file)
 cfg.GPU_ID = args.gpu_id
 
 
+# Set Benchmark
+# If this is set to True, it may consume more memory. (Default: True)
+if cfg.get('USE_BENCHMARK', True):
+    torch.backends.cudnn.enabled = True
+    torch.backends.cudnn.benchmark = True
+    tprint(f"CuDNN Benchmark is enabled.")
+
+
 # Set Random Seed
 seed = cfg.get('SEED', -1)
 seed = generate_random_seed(seed)

train.py

@@ -14,14 +14,19 @@
     torch.backends.cuda.matmul.allow_tf32 = False
     torch.backends.cudnn.allow_tf32 = False
 
-torch.backends.cudnn.enabled = True
-torch.backends.cudnn.benchmark = True
-
 
 # Get Config from 'config/monocon_configs.py'
 cfg = get_default_cfg()
 
 
+# Set Benchmark
+# If this is set to True, it may consume more memory. (Default: True)
+if cfg.get('USE_BENCHMARK', True):
+    torch.backends.cudnn.enabled = True
+    torch.backends.cudnn.benchmark = True
+    tprint(f"CuDNN Benchmark is enabled.")
+
+
 # Set Random Seed
 seed = cfg.get('SEED', -1)
 seed = generate_random_seed(seed)