Merge pull request #75 from cimat-ris/use_config_files

Use config files

.gitignore

@@ -131,3 +131,6 @@ dmypy.json
 # Dataset related
 *sdd_raw*
 *.pickle
+*.pth
+*.pdf
+*.csv

README.md

@@ -2,33 +2,40 @@
 [![Tests status](https://github.com/cimat-ris/trajpred-bdl/actions/workflows/python-app.yml/badge.svg)](https://github.com/cimat-ris/trajpred-bdl/actions/workflows/python-app.yml)
 # trajpred-bdl
 
+## To install the libraries
+
+```bash
+pip -e install .
+```
+
 ## Training
 
 To train a simple deterministic model:
 
-```
+```bash
 python scripts/train_deterministic.py
 ```
 
 To train a simple deterministic model with variances as output (DG):
 
-```
+```bash
  python scripts/train_deterministic_gaussian.py
 ```
 
 To train a model made of an ensemble of DGs (DGE):
 
-```
+```bash
 python scripts/train_ensembles.py
 ```
 
 To train a deterministic model with dropout at inference (DD):
-```
+
+```bash
 python scripts/train_dropout.py
 ```
 
 To train a deterministic-variational model (DV):
-```
+```bash
 python scripts/train_variational.py
 ```
 
@@ -37,15 +44,15 @@ python scripts/train_variational.py
 
 With any of the training scripts above, you can use the '--no-retrain' option to produce testing results
 
-```
+```bash
 python scripts/train_ensembles.py --no-retrain --pickle  --examples 10
 ```
 
 ## Calibration: a postprocess step
 
 After a model is trained, it saves it's results in a `pickle` file, then the calibration for it uses the output from a trained model and can be executed as follows:
 
-```
+```bash
 # training the desired model
 $ python scripts/train_torch_deterministic_gaussian.py --pickle --no-retrain
 
@@ -72,14 +79,15 @@ The `test_calibration.py` script uses Isotonic regression to compute the calibra
 
 * Clone the [bitrap repository](https://github.com/umautobots/bidireaction-trajectory-prediction).
 * The train/test partition from the [Trajectron++](https://github.com/StanfordASL/Trajectron-plus-plus) repository are now present in the datasets/trj++ directory as .pkl files.
-* Modify *bitrap_np_ETH.yml* lines 30 and set the path to where the .json file is located. You may also change BATCH_SIZE or NUM_WORKERS   
+* Modify *bitrap_np_ETH.yml* lines 30 and set the path to where the .json file is located. You may also change BATCH_SIZE or NUM_WORKERS  
+
 * To train bitrap, run
-```
+```bash
 python scripts/train_bitrap.py --config_file bitrap_np_ETH.yml --seed n
 ```
 By changing the seed, you will be building different models for an ensemble.
 
 * To generate data calibration from bitrap, run
-```
+```bash
 python tests/test_bitrap.py
 ```

cfg/bitrap_np_eth.yml

@@ -27,7 +27,7 @@ MODEL:
   DEC_HIDDEN_SIZE: 256
 DATASET:
   NAME: 'eth'    
-  ETH_CONFIG: '/home/jbhayet/opt/repositories/devel/bidireaction-trajectory-prediction/configs/ETH_UCY.json'
+  ETH_CONFIG: '../bitrap/configs/ETH_UCY.json'
   ROOT: 'datasets/trj++'
   TRAJECTORY_PATH: 'datasets/trj++'
 DATALOADER:

cfg/bitrap_np_hotel.yml

@@ -27,7 +27,7 @@ MODEL:
   DEC_HIDDEN_SIZE: 256
 DATASET:
   NAME: 'hotel'    
-  ETH_CONFIG: '/home/jbhayet/opt/repositories/devel/bidireaction-trajectory-prediction/configs/ETH_UCY.json'
+  ETH_CONFIG: '../bitrap/configs/ETH_UCY.json'
   ROOT: 'datasets/trj++'
   TRAJECTORY_PATH: 'datasets/trj++'
 DATALOADER:

cfg/bitrap_np_univ.yml

@@ -27,7 +27,7 @@ MODEL:
   DEC_HIDDEN_SIZE: 256
 DATASET:
   NAME: 'univ'    
-  ETH_CONFIG: '/home/jbhayet/opt/repositories/devel/bidireaction-trajectory-prediction/configs/ETH_UCY.json'
+  ETH_CONFIG: '../bitrap/configs/ETH_UCY.json'
   ROOT: 'datasets/trj++'
   TRAJECTORY_PATH: 'datasets/trj++'
 DATALOADER:

cfg/bitrap_np_zara1.yml

@@ -27,7 +27,7 @@ MODEL:
   DEC_HIDDEN_SIZE: 256
 DATASET:
   NAME: 'zara1'    
-  ETH_CONFIG: '/home/jbhayet/opt/repositories/devel/bidireaction-trajectory-prediction/configs/ETH_UCY.json'
+  ETH_CONFIG: '../bitrap/configs/ETH_UCY.json'
   ROOT: 'datasets/trj++'
   TRAJECTORY_PATH: 'datasets/trj++'
 DATALOADER:

cfg/bitrap_np_zara2.yml

@@ -27,7 +27,7 @@ MODEL:
   DEC_HIDDEN_SIZE: 256
 DATASET:
   NAME: 'zara2'    
-  ETH_CONFIG: '/home/jbhayet/opt/repositories/devel/bidireaction-trajectory-prediction/configs/ETH_UCY.json'
+  ETH_CONFIG: '../bitrap/configs/ETH_UCY.json'
   ROOT: 'datasets/trj++'
   TRAJECTORY_PATH: 'datasets/trj++'
 DATALOADER:

cfg/deterministic_dropout_ethucy.yaml

@@ -0,0 +1,31 @@
+dataset:
+  pickle_dir: 'pickle/'  
+  validation_proportion: 0.1
+  use_neighbors: False
+  batch_size: 256        
+  person_max: 70 # Maximum number of persons in a frame
+  obs_len: 8 # Observation length (trajlet size)
+  pred_len: 12 # Prediction length
+  delim: "," # Delimiter		
+  dt: 0.4 # Delta time (time between two discrete time samples)
+  max_overlap: 1 # Maximal overlap between trajets
+model:
+  num_layers: 2
+  dropout_rate: 0.2
+  hidden_dim: 128
+  embedding_dim: 128
+  input_dim: 2
+  output_dim: 2  
+train:
+  initial_lr: 0.001
+  epochs: 220 #220
+  no_retrain: False
+  teacher_forcing: False
+  save_dir : "training_checkpoints/"
+  model_name: "deterministic"
+misc:
+  plot_losses: True
+  plot_dir: "images/"
+  show_test: False
+  samples_test: 10
+  model_samples: 10

cfg/deterministic_ethucy.yaml

@@ -0,0 +1,35 @@
+dataset:
+  id_dataset: 0 # 0: ETHUCY
+  id_test: 2
+  pickle: False
+  pickle_dir: 'pickle/'  
+  validation_proportion: 0.1
+  use_neighbors: False
+  batch_size: 512        
+  person_max: 70 # Maximum number of persons in a frame
+  obs_len: 8 # Observation length (trajlet size)
+  pred_len: 12 # Prediction length
+  delim: "," # Delimiter		
+  dt: 0.4 # Delta time (time between two discrete time samples)
+  max_overlap: 1 # Maximal overlap between trajets
+model:
+  num_layers: 2
+  dropout_rate: 0.2
+  hidden_dim: 128
+  embedding_dim: 8
+  input_dim: 2
+  output_dim: 2  
+train:
+  initial_lr: 0.001
+  epochs: 800
+  no_retrain: False
+  teacher_forcing: False
+  save_dir : "training_checkpoints/"
+  model_name: "deterministic"
+misc:
+  plot_losses: True
+  plot_dir: "images/"
+  show_test: True
+  samples_test: 10
+  log_level: 20
+  seed: 1234

cfg/deterministic_gaussian_ethucy.yaml

@@ -0,0 +1,31 @@
+dataset:
+  pickle_dir: 'pickle/'  
+  validation_proportion: 0.1
+  use_neighbors: True
+  batch_size: 256        
+  person_max: 70 # Maximum number of persons in a frame
+  obs_len: 8 # Observation length (trajlet size)
+  pred_len: 12 # Prediction length
+  delim: "," # Delimiter		
+  dt: 0.4 # Delta time (time between two discrete time samples)
+  max_overlap: 1 # Maximal overlap between trajets
+model:
+  num_layers: 2
+  dropout_rate: 0.2
+  hidden_dim: 128
+  embedding_dim: 128
+  input_dim: 2
+  output_dim: 2  
+train:
+  initial_lr: 0.001
+  epochs: 220
+  no_retrain: False
+  teacher_forcing: False
+  save_dir : "training_checkpoints/"
+  model_name: "deterministic_gaussian"
+misc:
+  plot_losses: False
+  plot_dir: "images/"
+  show_test: False
+  samples_test: 10
+  model_samples: 3

cfg/deterministic_gaussian_sdd.yaml

@@ -0,0 +1,36 @@
+dataset:
+  id_dataset: 1 # 0: ETHUCY
+  id_test: 2
+  pickle: True
+  pickle_dir: 'pickle/'  
+  validation_proportion: 0.1
+  use_neighbors: False
+  batch_size: 256        
+  person_max: 70 # Maximum number of persons in a frame
+  obs_len: 8 # Observation length (trajlet size)
+  pred_len: 12 # Prediction length
+  delim: "," # Delimiter		
+  dt: 0.4 # Delta time (time between two discrete time samples)
+  max_overlap: 1 # Maximal overlap between trajets
+model:
+  num_layers: 2
+  dropout_rate: 0.2
+  hidden_dim: 128
+  embedding_dim: 128
+  input_dim: 2
+  output_dim: 2  
+train:
+  initial_lr: 0.001
+  epochs: 1 #220
+  no_retrain: True
+  teacher_forcing: False
+  save_dir : "training_checkpoints/"
+  model_name: "deterministic_gaussian"
+misc:
+  plot_losses: False
+  plot_dir: "images/"
+  show_test: True
+  samples_test: 10
+  log_level: 20
+  seed: 1234
+  model_samples: 3

cfg/deterministic_variational_ethucy.yaml

@@ -0,0 +1,37 @@
+dataset:
+  id_dataset: 0 # 0: ETHUCY
+  id_test: 2
+  pickle: False
+  pickle_dir: 'pickle/'  
+  validation_proportion: 0.1
+  use_neighbors: False
+  batch_size: 256        
+  person_max: 70 # Maximum number of persons in a frame
+  obs_len: 8 # Observation length (trajlet size)
+  pred_len: 12 # Prediction length
+  delim: "," # Delimiter		
+  dt: 0.4 # Delta time (time between two discrete time samples)
+  max_overlap: 1 # Maximal overlap between trajets
+model:
+  num_layers: 2
+  dropout_rate: 0.2
+  hidden_dim: 128
+  embedding_dim: 128
+  input_dim: 2
+  output_dim: 2  
+train:
+  initial_lr: 0.001
+  epochs: 2 #220
+  num_mctrain: 10
+  no_retrain: False
+  teacher_forcing: False
+  save_dir : "training_checkpoints/"
+  model_name: "deterministic_variational"
+misc:
+  plot_losses: True
+  plot_dir: "images/"
+  show_test: True
+  samples_test: 10
+  log_level: 20
+  seed: 1234
+  model_samples: 10

datasets/edinburgh/preprocess_edinburgh.py

@@ -0,0 +1,37 @@
+import os, sys
+import pandas as pd
+import matplotlib.pyplot as plt
+
+from opentraj.loaders.loader_edinburgh import load_edinburgh
+sys.path.append('.')
+
+months = {
+    #'Jan' : {'01','02','04','05','06','07','08','10','11','12','13','14','15','16','17','18','19'},
+    #'May' : {'29','30','31'},
+    #'Jun' : {'02','03','04','05','06','08','09','11','12','14','16','17','18','20','22','24','25','26','29','30'},
+    #'Jul' : {'01','02','04','11','12','13','14','17','18','19','20','21','22','23','25','26','27','28','29','30'},
+    #'Aug' : {'01','24','25','26','27','28','29','30'},
+    'Sep' : {'01','02','04','05','06','10','11','12','13','14','16','18','19','20','21','22','23','24','25','27','28','29','30'},
+    'Oct': {'02','03','04','05','06','07','08','09','10','11','12','13','14','15'},
+    'Dec' : {'06','11','14','15','16','18','19','20','21','22','23','24','29','30','31'}
+}
+# Fixme: set proper OpenTraj directory
+edi_root = os.path.join('../../../OpenTraj', 'datasets', 'Edinburgh','annotations')
+edi_data = {key: pd.DataFrame() for key in months.keys() }
+
+for month, videos_per_month in months.items():
+    scene_video_ids         = [day+month for day in videos_per_month]
+    traj_datasets_per_scene = []
+
+    for scene_video_id in scene_video_ids:
+        annot_file = os.path.join(edi_root,'tracks.'+scene_video_id+'.txt')
+        print(annot_file)
+        itraj_dataset = load_edinburgh(annot_file, title="Edinburgh",
+                                  use_kalman=False, scene_id=scene_video_id, sampling_rate=4)  # original framerate=9
+        trajs = list(itraj_dataset.get_trajectories())
+        traj_datasets_per_scene.append(pd.concat([itraj_dataset.data.iloc[:, : 4], itraj_dataset.data.iloc[:, 8:9]], axis=1))
+
+    if len(traj_datasets_per_scene)>0:
+        df = pd.concat(traj_datasets_per_scene)
+        df.to_pickle(month+'.pickle')
+

datasets/sdd/preprocess_sdd.py

@@ -2,6 +2,8 @@
 import pickle
 import pandas as pd
 
+sys.path.append('../../../OpenTraj')
+sys.path.append('../../../OpenTraj/opentraj')
 from opentraj.toolkit.loaders.loader_sdd import load_sdd
 sys.path.append('.')
 
@@ -16,8 +18,7 @@
     'quad' : 4
 }
 # Fixme: set proper OpenTraj directory
-sdd_root = os.path.join('/<dirTo>/OpenTraj', 'datasets', 'SDD')
-
+sdd_root = os.path.join('../../../OpenTraj', 'datasets', 'SDD')
 sdd_data = {key: pd.DataFrame() for key in scenes.keys() }
 
 for scene_name, total_videos_per_scene in scenes.items():
@@ -36,6 +37,7 @@
                         drop_lost_frames=False, use_kalman=False, label='Pedestrian')
         traj_datasets_per_scene.append(pd.concat([itraj_dataset.data.iloc[:, : 4], itraj_dataset.data.iloc[:, 8:9]], axis=1))
 
-    pickle_out = open(scene_name+'.pickle',"wb")
-    pickle.dump(pd.concat(traj_datasets_per_scene), pickle_out, protocol=2)
-    pickle_out.close()
+
+    df = pd.concat(traj_datasets_per_scene)
+    df.to_pickle(scene_name+'.pickle')
+