def test_model(saved_files_path=None):
with open(os.path.join(saved_files_path, 'configs.yaml'), 'r') as yamlfile:
opts = yaml.safe_load(yamlfile)
print(opts)
model_opts = opts['model_opts']
data_opts = opts['data_opts']
net_opts = opts['net_opts']
tte = model_opts['time_to_event'] if isinstance(model_opts['time_to_event'], int) else \
model_opts['time_to_event'][1]
data_opts['min_track_size'] = model_opts['obs_length'] + tte
if model_opts['dataset'] == 'pie':
imdb = PIE(data_path=os.environ.copy()['PIE_PATH'])
imdb.get_data_stats()
elif model_opts['dataset'] == 'jaad':
imdb = JAAD(data_path=os.environ.copy()['JAAD_PATH'])
else:
raise ValueError("{} dataset is incorrect".format(
model_opts['dataset']))
method_class = action_prediction(model_opts['model'])(**net_opts)
#beh_seq_train = imdb.generate_data_trajectory_sequence('train', **data_opts)
#saved_files_path = method_class.train(beh_seq_train, **train_opts, model_opts=model_opts)
beh_seq_test = imdb.generate_data_trajectory_sequence('test', **data_opts)
acc, auc, f1, precision, recall = method_class.test(
beh_seq_test, saved_files_path)
regularizer_val=0.001,
lstm_dropout=0.4,
lstm_recurrent_dropout=0.2,
convlstm_num_filters=64,
convlstm_kernel_size=2)
pretrained_model_path = 'data/pie/intention/context_loc_pretrained'
data_path = './PIE_dataset'
img_path = './PIE_dataset/images'
set_id = 'set02'
vid_id = 'video_0003'
ped_id = '2_3_217'
#['2_3_194', '2_3_195', '2_3_196', '2_3_198', '2_3_199', '2_3_200', '2_3_201', '2_3_202', '2_3_203', '2_3_204', '2_3_205', '2_3_206', '2_3_207', '2_3_208', '2_3_209', '2_3_211', '2_3_212', '2_3_213', '2_3_214', '2_3_215', '2_3_216', '2_3_217', '2_3_210', '2_3_197']
imdb = PIE(data_path=data_path)
annt = imdb._get_annotations(set_id, vid_id)
ped = annt['ped_annotations']
ped_frames = ped[ped_id]['frames'][:30]
ped_frames = [i for i in range(63)]
#bbox = ped['ped_id']['bbox']
save_path = './bag1_results'
# 15 consecutive frames
# load bbox
bbox_sequences = ped[ped_id]['bbox'][:30]
VERSION = 'v0'
#Version v0, Total time: 31.7845561504364s, Per frame: 0.5045167642926413s
#Version v1, Total time: 24.467329025268555s, Per frame: 0.38837030198838973s
def train_predict(
dataset='pie',
train_test=2,
intent_model_path='data/pie/intention/context_loc_pretrained'):
data_opts = {
'fstride': 1,
'sample_type': 'all',
'height_rng': [0, float('inf')],
'squarify_ratio': 0,
'data_split_type': 'default', # kfold, random, default
'seq_type': 'trajectory',
'min_track_size': 61,
'random_params': {
'ratios': None,
'val_data': True,
'regen_data': True
},
'kfold_params': {
'num_folds': 5,
'fold': 1
}
}
t = PIEPredict()
pie_path = os.environ.copy()['PIE_PATH']
if dataset == 'pie':
imdb = PIE(data_path=pie_path)
traj_model_opts = {
'normalize_bbox': True,
'track_overlap': 0.5,
'observe_length': 15,
'predict_length': 45,
'enc_input_type': ['bbox'],
'dec_input_type': ['intention_prob', 'obd_speed'],
'prediction_type': ['bbox']
}
speed_model_opts = {
'normalize_bbox': True,
'track_overlap': 0.5,
'observe_length': 15,
'predict_length': 45,
'enc_input_type': ['obd_speed'],
'dec_input_type': [],
'prediction_type': ['obd_speed']
}
traj_model_path = 'data/pie/trajectory/loc_intent_speed_pretrained'
speed_model_path = 'data/pie/speed/speed_pretrained'
if train_test < 2:
beh_seq_val = imdb.generate_data_trajectory_sequence(
'val', **data_opts)
beh_seq_train = imdb.generate_data_trajectory_sequence(
'train', **data_opts)
traj_model_path = t.train(beh_seq_train, beh_seq_val,
**traj_model_opts)
speed_model_path = t.train(beh_seq_train, beh_seq_val,
**speed_model_opts)
if train_test > 0:
beh_seq_test = imdb.generate_data_trajectory_sequence(
'test', **data_opts)
perf_final = t.test_final(beh_seq_test,
traj_model_path=traj_model_path,
speed_model_path=speed_model_path,
intent_model_path=intent_model_path)
t = PrettyTable(['MSE', 'C_MSE'])
t.title = 'Trajectory prediction model (loc + PIE_intent + PIE_speed)'
t.add_row([perf_final['mse-45'], perf_final['c-mse-45']])
print(t)
def train_intent(train_test=1):
data_opts = {
'fstride': 1,
'sample_type': 'all',
'height_rng': [0, float('inf')],
'squarify_ratio': 0,
'data_split_type': 'default', # kfold, random, default
'seq_type': 'intention', # crossing , intention
'min_track_size': 0, # discard tracks that are shorter
'max_size_observe': 15, # number of observation frames
'max_size_predict': 5, # number of prediction frames
'seq_overlap_rate': 0.5, # how much consecutive sequences overlap
'balance': True, # balance the training and testing samples
'crop_type': 'context', # crop 2x size of bbox around the pedestrian
'crop_mode': 'pad_resize', # pad with 0s and resize to VGG input
'encoder_input_type': [],
'decoder_input_type': ['bbox'],
'output_type': ['intention_binary']
}
t = PIEIntent(num_hidden_units=128,
regularizer_val=0.001,
lstm_dropout=0.4,
lstm_recurrent_dropout=0.2,
convlstm_num_filters=64,
convlstm_kernel_size=2)
saved_files_path = ''
imdb = PIE(data_path=os.environ.copy()['PIE_PATH'])
pretrained_model_path = 'data/pie/intention/context_loc_pretrained'
if train_test < 2: # Train
beh_seq_val = imdb.generate_data_trajectory_sequence(
'val', **data_opts)
beh_seq_val = imdb.balance_samples_count(beh_seq_val,
label_type='intention_binary')
beh_seq_train = imdb.generate_data_trajectory_sequence(
'train', **data_opts)
beh_seq_train = imdb.balance_samples_count(
beh_seq_train, label_type='intention_binary')
saved_files_path = t.train(data_train=beh_seq_train,
data_val=beh_seq_val,
epochs=400,
loss=['binary_crossentropy'],
metrics=['accuracy'],
batch_size=128,
optimizer_type='rmsprop',
data_opts=data_opts)
print(data_opts['seq_overlap_rate'])
if train_test > 0: # Test
if saved_files_path == '':
saved_files_path = pretrained_model_path
beh_seq_test = imdb.generate_data_trajectory_sequence(
'test', **data_opts)
acc, f1 = t.test_chunk(beh_seq_test, data_opts, saved_files_path,
False)
t = PrettyTable(['Acc', 'F1'])
t.title = 'Intention model (local_context + bbox)'
t.add_row([acc, f1])
print(t)
K.clear_session()
tf.reset_default_graph()
return saved_files_path
import os
from sf_gru import SFGRU
from pie_data import PIE
data_opts = {
'fstride': 1,
'subset': 'default',
'data_split_type': 'random', # kfold, random, default
'seq_type': 'crossing',
'min_track_size': 75
} ## for obs length of 15 frames + 60 frames tte. This should be adjusted for different setup
imdb = PIE(data_path=os.environ.copy()
['PIE_PATH']) # change with the path to the dataset
model_opts = {
'obs_input_type': ['local_box', 'local_context', 'pose', 'box', 'speed'],
'enlarge_ratio': 1.5,
'pred_target_type': ['crossing'],
'obs_length': 15, # Determines min track size
'time_to_event': 60, # Determines min track size
'dataset': 'pie',
'normalize_boxes': True
}
method_class = SFGRU()
beh_seq_train = imdb.generate_data_trajectory_sequence('train', **data_opts)
saved_files_path = method_class.train(beh_seq_train, model_opts=model_opts)
beh_seq_train = imdb.generate_data_trajectory_sequence('test', **data_opts)
acc, auc, f1, precision, recall = method_class.test(beh_seq_train,
saved_files_path)
def run(config_file=None):
"""
Run train and test on the dataset with parameters specified in configuration file.
Args:
config_file: path to configuration file in yaml format
dataset: dataset to train and test the model on (pie, jaad_beh or jaad_all)
"""
print(config_file)
# Read default Config file
configs_default = 'config_files/configs_default.yaml'
with open(configs_default, 'r') as f:
configs = yaml.safe_load(f)
with open(config_file, 'r') as f:
model_configs = yaml.safe_load(f)
# Update configs based on the model configs
for k in ['model_opts', 'net_opts']:
if k in model_configs:
configs[k].update(model_configs[k])
# Calculate min track size
tte = configs['model_opts']['time_to_event'] if isinstance(configs['model_opts']['time_to_event'], int) else \
configs['model_opts']['time_to_event'][1]
configs['data_opts'][
'min_track_size'] = configs['model_opts']['obs_length'] + tte
# update model and training options from the config file
for dataset_idx, dataset in enumerate(
model_configs['exp_opts']['datasets']):
configs['data_opts'][
'sample_type'] = 'beh' if 'beh' in dataset else 'all'
configs['model_opts']['overlap'] = 0.6 if 'pie' in dataset else 0.8
configs['model_opts']['dataset'] = dataset.split('_')[0]
configs['train_opts']['batch_size'] = model_configs['exp_opts'][
'batch_size'][dataset_idx]
configs['train_opts']['lr'] = model_configs['exp_opts']['lr'][
dataset_idx]
configs['train_opts']['epochs'] = model_configs['exp_opts']['epochs'][
dataset_idx]
model_name = configs['model_opts']['model']
# Remove speed in case the dataset is jaad
if 'RNN' in model_name and 'jaad' in dataset:
configs['model_opts']['obs_input_type'] = configs['model_opts'][
'obs_input_type']
for k, v in configs.items():
print(k, v)
# set batch size
if model_name in ['ConvLSTM']:
configs['train_opts']['batch_size'] = 2
if model_name in ['C3D', 'I3D']:
configs['train_opts']['batch_size'] = 16
if model_name in ['PCPA']:
configs['train_opts']['batch_size'] = 8
if 'MultiRNN' in model_name:
configs['train_opts']['batch_size'] = 8
if model_name in ['TwoStream']:
configs['train_opts']['batch_size'] = 16
if configs['model_opts']['dataset'] == 'pie':
imdb = PIE(data_path=os.environ.copy()['PIE_PATH'])
elif configs['model_opts']['dataset'] == 'jaad':
imdb = JAAD(data_path=os.environ.copy()['JAAD_PATH'])
# get sequences
beh_seq_train = imdb.generate_data_trajectory_sequence(
'train', **configs['data_opts'])
beh_seq_val = None
# Uncomment the line below to use validation set
#beh_seq_val = imdb.generate_data_trajectory_sequence('val', **configs['data_opts'])
beh_seq_test = imdb.generate_data_trajectory_sequence(
'test', **configs['data_opts'])
# get the model
method_class = action_prediction(
configs['model_opts']['model'])(**configs['net_opts'])
# train and save the model
saved_files_path = method_class.train(beh_seq_train,
beh_seq_val,
**configs['train_opts'],
model_opts=configs['model_opts'])
# test and evaluate the model
acc, auc, f1, precision, recall = method_class.test(
beh_seq_test, saved_files_path)
# save the results
data = {}
data['results'] = {}
data['results']['acc'] = float(acc)
data['results']['auc'] = float(auc)
data['results']['f1'] = float(f1)
data['results']['precision'] = float(precision)
data['results']['recall'] = float(recall)
write_to_yaml(yaml_path=os.path.join(saved_files_path, 'results.yaml'),
data=data)
data = configs
write_to_yaml(yaml_path=os.path.join(saved_files_path, 'configs.yaml'),
data=data)
print('Model saved to {}'.format(saved_files_path))