def e1002_3D_ucf101():
set_init_1()
project_variable.model_number = 25
project_variable.experiment_number = 1002
project_variable.sheet_number = 23
project_variable.device = 1
project_variable.end_epoch = 100
project_variable.batch_size = 20
project_variable.batch_size_val_test = 20
project_variable.load_model = True # exp, model, epoch, run
project_variable.load_from_fast = True
project_variable.use_dali = True
project_variable.dali_workers = 32
project_variable.dali_iterator_size = ['all', 'all', 0]
project_variable.nas = False
project_variable.stop_at_collapse = True
project_variable.early_stopping = True
project_variable.optimizer = 'adam'
project_variable.learning_rate = 0.000005
project_variable.use_adaptive_lr = True
main_file.run(project_variable)
def e002_dots():
set_init_1_frames()
project_variable.dataset = 'dots_frames'
# project_variable.dataset = 'ucf101'
# project_variable.model_number = 56 # lenet5 3t
project_variable.model_number = 55 # lenet5 2D
project_variable.experiment_number = 2
project_variable.device = 0
project_variable.end_epoch = 10
project_variable.batch_size = 32
project_variable.batch_size_val_test = 32
# project_variable.inference_only_mode = True
project_variable.inference_only_mode = False
project_variable.load_model = False # loading model from scratch
# project_variable.load_from_fast = True # UNUSED?
project_variable.use_dali = True
project_variable.dali_workers = 32
project_variable.dali_iterator_size = ['all', 'all', 0]
project_variable.nas = False
# project_variable.stop_at_collapse = True
project_variable.stop_at_collapse = False
# project_variable.early_stopping = True
project_variable.early_stopping = False
project_variable.optimizer = 'adam'
# project_variable.learning_rate = 0.05
project_variable.learning_rate = 0.0005
project_variable.use_adaptive_lr = True
main_file.run(project_variable)
def e001_3T_kinetics():
set_init_1()
project_variable.model_number = 23 # googlenet
project_variable.experiment_number = 1
project_variable.device = 2
project_variable.end_epoch = 200
project_variable.batch_size = 18 # 5 about 3000
project_variable.batch_size_val_test = 30 # 30 about 3000
# project_variable.inference_only_mode = True
project_variable.inference_only_mode = False
# project_variable.save_model_every_x_epochs = 1
project_variable.load_model = True # loading pre-trained on ImageNet
project_variable.load_from_fast = True
project_variable.use_dali = True
project_variable.dali_workers = 32
project_variable.dali_iterator_size = ['all', 'all', 0]
project_variable.nas = False
project_variable.stop_at_collapse = True
# project_variable.stop_at_collapse = False
project_variable.early_stopping = True
project_variable.optimizer = 'adam'
project_variable.learning_rate = 0.00005
project_variable.use_adaptive_lr = True
main_file.run(project_variable)
def e1011_3T_ucf101():
set_init_1()
project_variable.nin = True
project_variable.train_nin_mode = 'joint'
project_variable.model_number = 52 # TACoNet
project_variable.experiment_number = 1011
project_variable.sheet_number = 22
project_variable.device = 1
project_variable.end_epoch = 1
project_variable.repeat_experiments = 1
project_variable.batch_size = 20
project_variable.batch_size_val_test = 20
project_variable.load_model = None
project_variable.load_from_fast = True
project_variable.use_dali = True
project_variable.dali_workers = 32
project_variable.dali_iterator_size = ['all', 'all', 0]
project_variable.nas = False
project_variable.stop_at_collapse = True
project_variable.early_stopping = True
project_variable.optimizer = 'adam'
project_variable.learning_rate = 0.00005
project_variable.use_adaptive_lr = True
project_variable.num_out_channels = [0]
main_file.run(project_variable)
def e1014_3T_ucf101():
set_init_1()
project_variable.model_number = 23
project_variable.experiment_number = 1014
project_variable.sheet_number = 23
project_variable.device = 1
project_variable.end_epoch = 100
project_variable.batch_size = 15
project_variable.batch_size_val_test = 15
project_variable.load_model = [1, 23, 4, 0] # exp, model, epoch, run
project_variable.load_from_fast = True
project_variable.use_dali = True
project_variable.dali_workers = 32
project_variable.dali_iterator_size = ['all', 'all', 0]
project_variable.nas = False
project_variable.stop_at_collapse = True
project_variable.early_stopping = True
project_variable.optimizer = 'adam'
project_variable.learning_rate = 0.00005
project_variable.use_adaptive_lr = True
utils.wait_for_gpu(wait=True, device_num=project_variable.device, threshold=9700)
main_file.run(project_variable)
def e101_conv3D_jester():
set_init_1()
project_variable.model_number = 21 # 3Dresnet18 from scratch
project_variable.experiment_number = 101
project_variable.sheet_number = 22
project_variable.device = 2
project_variable.end_epoch = 100
project_variable.repeat_experiments = 1
project_variable.batch_size = 32
project_variable.batch_size_val_test = 32
project_variable.load_model = False
project_variable.load_from_fast = True
project_variable.use_dali = True
project_variable.dali_workers = 32
project_variable.dali_iterator_size = ['all', 'all', 0]
project_variable.nas = False
project_variable.stop_at_collapse = True
project_variable.early_stopping = True
project_variable.optimizer = 'adam'
project_variable.learning_rate = 0.000005
project_variable.use_adaptive_lr = True
project_variable.num_out_channels = [0]
main_file.run(project_variable)
def e1010_3T_ucf101():
set_init_1()
project_variable.nin = False
project_variable.model_number = 51 # convnet3T
project_variable.experiment_number = 38
project_variable.sheet_number = 23
project_variable.device = 1
project_variable.end_epoch = 1
project_variable.repeat_experiments = 1
project_variable.batch_size = 20
project_variable.batch_size_val_test = 20
project_variable.load_model = None
project_variable.load_from_fast = True
project_variable.use_dali = True
project_variable.dali_workers = 32
project_variable.dali_iterator_size = ['all', 'all', 0]
project_variable.nas = False
project_variable.stop_at_collapse = True
project_variable.early_stopping = True
project_variable.optimizer = 'adam'
project_variable.learning_rate = 0.00005
project_variable.use_adaptive_lr = True
project_variable.num_out_channels = [0]
# utils.wait_for_gpu(wait=True, device_num=project_variable.device)
main_file.run(project_variable)
def e45_conv3T_jester():
set_init_1()
project_variable.dataset = 'tiny_jester'
project_variable.nin = True
project_variable.train_nin_mode = 'joint'
project_variable.model_number = 54 # alexnet-taco
project_variable.experiment_number = 45
project_variable.sheet_number = 22
project_variable.device = 0
project_variable.end_epoch = 100
project_variable.repeat_experiments = 3
project_variable.batch_size = 20
project_variable.batch_size_val_test = 20
project_variable.load_model = True
project_variable.load_from_fast = True
project_variable.use_dali = True
project_variable.dali_workers = 32
project_variable.dali_iterator_size = ['all', 'all', 0]
project_variable.nas = False
project_variable.stop_at_collapse = True
project_variable.early_stopping = True
project_variable.optimizer = 'adam'
project_variable.learning_rate = 0.00001
project_variable.use_adaptive_lr = True
project_variable.num_out_channels = [0]
# utils.wait_for_gpu(wait=True, device_num=project_variable.device)
main_file.run(project_variable)
def e40_conv3T_jester():
set_init_1()
project_variable.dataset = 'tiny_jester'
project_variable.nin = False
# project_variable.train_nin_mode = 'nin_only'
project_variable.model_number = 51 # convnet3T
project_variable.experiment_number = 40
project_variable.sheet_number = 22
project_variable.device = 2
project_variable.end_epoch = 100
project_variable.repeat_experiments = 10
project_variable.batch_size = 20
project_variable.batch_size_val_test = 20
project_variable.load_model = None
project_variable.load_from_fast = True
project_variable.use_dali = True
project_variable.dali_workers = 32
project_variable.dali_iterator_size = ['all', 'all', 0]
project_variable.nas = False
project_variable.stop_at_collapse = True
project_variable.early_stopping = True
project_variable.optimizer = 'adam'
project_variable.learning_rate = 0.00005
project_variable.use_adaptive_lr = True
project_variable.num_out_channels = [0]
wait_for_gpu(wait=True, device_num=project_variable.device, threshold=2000)
main_file.run(project_variable)
def e37_conv3T_jester():
set_init_1()
project_variable.nin = True
project_variable.train_nin_mode = 'nin_only'
project_variable.model_number = 50 # RN18 3T
project_variable.experiment_number = 37
project_variable.sheet_number = 22
project_variable.device = 1
project_variable.end_epoch = 5
project_variable.repeat_experiments = 1
project_variable.batch_size = 1
project_variable.batch_size_val_test = 1
project_variable.load_model = [31, 20, 8, 0]
project_variable.load_from_fast = True
project_variable.use_dali = True
project_variable.dali_workers = 32
project_variable.dali_iterator_size = ['all', 'all', 0]
project_variable.nas = False
project_variable.stop_at_collapse = True
project_variable.early_stopping = True
project_variable.optimizer = 'adam'
project_variable.learning_rate = 0.00005
project_variable.use_adaptive_lr = True
project_variable.num_out_channels = [0]
# wait_for_gpu(wait=True, device_num=project_variable.device)
main_file.run(project_variable)
def e009_3T_kinetics():
set_init_1()
project_variable.dataset = 'kinetics400'
project_variable.label_size = 400
project_variable.model_number = 23 # googlenet
project_variable.experiment_number = 9
project_variable.device = 1
project_variable.end_epoch = 200
project_variable.batch_size = 17 # 5 about 3000
# project_variable.batch_size = 10 # 5 about 3000
project_variable.batch_size_val_test = 29 # 30 about 3000
# project_variable.inference_only_mode = True
project_variable.inference_only_mode = False
# project_variable.eval_on = 'test'
project_variable.eval_on = 'val'
# project_variable.load_model = True # loading pre-trained on ImageNet
project_variable.load_model = [6, 23, 7, 0]
project_variable.load_from_fast = True
project_variable.use_dali = True
project_variable.dali_workers = 32
project_variable.dali_iterator_size = ['all', 'all', 0]
project_variable.nas = False
project_variable.stop_at_collapse = True
project_variable.early_stopping = True
project_variable.optimizer = 'adam'
project_variable.learning_rate = 0.00007
project_variable.use_adaptive_lr = True
utils.wait_for_gpu(wait=True,
device_num=project_variable.device,
threshold=9700)
main_file.run(project_variable)
def e002_3T_kinetics():
set_init_1()
project_variable.dataset = 'kinetics400_metaclass'
project_variable.label_size = 39
project_variable.model_number = 23 # googlenet
project_variable.experiment_number = 2
project_variable.device = 1
project_variable.end_epoch = 200
project_variable.batch_size = 17 # 5 about 3000
project_variable.batch_size_val_test = 29 # 30 about 3000
# project_variable.inference_only_mode = True
project_variable.inference_only_mode = False
# project_variable.eval_on = 'test'
project_variable.eval_on = 'val'
# TODO: standardize the test data
# project_variable.save_model_every_x_epochs = 1
project_variable.load_model = True # loading pre-trained on ImageNet
# project_variable.load_model = [2, 23, 10, 0]
project_variable.load_from_fast = True
project_variable.use_dali = True
project_variable.dali_workers = 32
project_variable.dali_iterator_size = ['all', 'all', 0]
project_variable.nas = False
project_variable.stop_at_collapse = True
project_variable.early_stopping = True
project_variable.optimizer = 'adam'
project_variable.learning_rate = 0.00005
project_variable.use_adaptive_lr = False
main_file.run(project_variable)
def e010_dots():
set_init_1_frames()
project_variable.dataset = 'dots_avi'
# project_variable.dataset = 'ucf101'
project_variable.model_number = 57 # simplenet3t, 1 conv layer
project_variable.experiment_number = 10
project_variable.load_num_frames = 15
project_variable.device = 0
project_variable.end_epoch = 20
project_variable.batch_size = 32
project_variable.batch_size_val_test = 32
# project_variable.inference_only_mode = True
project_variable.inference_only_mode = False
project_variable.load_model = False # loading model from scratch
# project_variable.load_from_fast = True # UNUSED?
project_variable.use_dali = True
project_variable.dali_workers = 32
project_variable.dali_iterator_size = ['all', 'all', 0]
project_variable.nas = False
project_variable.save_all_models = True
# project_variable.stop_at_collapse = True
project_variable.stop_at_collapse = False
# project_variable.early_stopping = True
project_variable.early_stopping = False
project_variable.optimizer = 'sgd'
# project_variable.learning_rate = 0.03
# project_variable.learning_rate = 0.05
project_variable.learning_rate = 0.08
project_variable.use_adaptive_lr = True
main_file.run(project_variable)
def e29_conv3T_jester():
set_init_1()
project_variable.model_number = 20
project_variable.experiment_number = 29
project_variable.sheet_number = 22
project_variable.device = 1
project_variable.end_epoch = 100
project_variable.repeat_experiments = 1
project_variable.batch_size = 32
project_variable.batch_size_val_test = 32
project_variable.load_model = True
project_variable.load_from_fast = True
project_variable.use_dali = True
project_variable.dali_workers = 32
project_variable.dali_iterator_size = ['all', 'all', 0]
project_variable.nas = False
project_variable.stop_at_collapse = True
project_variable.early_stopping = True
project_variable.optimizer = 'adam'
project_variable.learning_rate = 0.00005
project_variable.use_adaptive_lr = True
project_variable.num_out_channels = [0]
go = False
while not go:
gpu_available = get_gpu_memory_map()
if gpu_available[project_variable.device] < 100:
go = True
else:
print('waiting for gpu...')
time.sleep(10)
main_file.run(project_variable)
def generation_loop_with_cnn(device_num=0):
'''
rad_dot_m_t, num_dots_low_t, num_dots_high_t = parameters
2, 10, 20
rad_dot_m_r, num_dots_low_r, num_dots_high_r = parameters
2, 10, 20
num_dots_low_s, num_dots_high_s, radius_dot_min_s_pos_dir, radius_dot_max_s_pos_dir, radius_dot_min_s_neg_dir, radius_dot_max_s_neg_dir = parameters
10, 11, 3, 4, 3, 4
'''
# parameters = [2, 10, 20, 2, 10, 20, 10, 11, 3, 4, 3, 4]
random_acc = 1 / 3
val_acc = 1
e = 0.05
it = 1
while val_acc > random_acc + e:
opt_remove_dataset()
parameters = sample_params()
make_dataset('train', 500, 30, 33, parameters)
make_dataset('val', 200, 30, 33, parameters)
pv = config_pv(device_num)
val_acc = main_file.run(pv)
if val_acc > random_acc + e:
print('%d: val_acc: %f. parameters tried: %s' %
(it, val_acc, str(parameters)))
save_results(val_acc, parameters)
else:
print('OPTIMAL PARAMETERS FOUND!')
print('%d: val_acc: %f. parameters: %s' %
(it, val_acc, str(parameters)))
save_results(val_acc, parameters)
it = it + 1
