# TRAINING PARAMS
num_images_train = len(filenames_train)
num_images_valid = len(filenames_valid)
batch_size = 512
num_epochs = 25
num_steps_per_epoch = int(np.floor(num_images_train/batch_size)) # Use entire dataset per epoch; round up to ensure entire dataset is covered if batch_size does not divide into num_images
num_steps_per_epoch_valid = int(np.floor(num_images_valid/batch_size)) # As above
seed_train = 587
seed_valid = seed_train+1
# Now create the training & validation datasets
dataset_train = utils.create_dataset(filenames = filenames_train
, labels = labels_train_onehot
, num_channels = image_depth
, batch_size = batch_size
, shuffle_and_repeat = True
, repeat_count = num_epochs
, seed = seed_train)
dataset_valid = utils.create_dataset(filenames = filenames_valid
, labels = labels_valid_onehot
, num_channels = image_depth
, batch_size = batch_size
, shuffle_and_repeat = True
, repeat_count = num_epochs
, seed = seed_valid)
print("DATASETS CREATED")
### BUILD MODEL
# Big: initial_filters=256, size_final_dense=100
is_cuda = not cfg.NO_CUDA and torch.cuda.is_available()
torch.manual_seed(cfg.TRAIN.SEED)
if is_cuda:
torch.cuda.manual_seed(cfg.TRAIN.SEED)
# TODO: Transforms for the images
transform = transforms.Compose([])
print(transform)
# Define trainset
trainset = create_dataset(cfg,
cfg.DATASET.SPLIT,
'train',
length=cfg.DATASET.LENGTH,
transform=transform)
# Load train data
train_loader = DataLoader(trainset,
batch_size=cfg.TRAIN.BATCH_SIZE,
shuffle=True,
pin_memory=False,
num_workers=cfg.DATASET.WORKERS)
# Start epoch for the model
start_epoch = cfg.TRAIN.START_EPOCH
if cfg.DATASET.VAL:
# Define validation dataset
def conduct_experiment(n_trials,
n_samples,
func,
transition_period,
tolerance_delay,
random_seed=0):
# fix seed for reproducibility
np.random.seed(random_seed)
mu_max = 1000
div_min = 1e-8
div_max = 1e8
df_result = pd.DataFrame()
print("Create Dataset")
dataset, changepoints = create_dataset(n_samples, func, transition_period)
print("ChangeFinder")
row = conduct_CF(n_trials=n_trials,
n_samples=n_samples,
dataset=dataset,
changepoints=changepoints,
tolerance_delay=tolerance_delay)
df_result = pd.concat([df_result, row], axis=0)
print("BOCPD")
row = conduct_BOCPD(n_trials=n_trials,
n_samples=n_samples,
dataset=dataset,
changepoints=changepoints,
tolerance_delay=tolerance_delay)
df_result = pd.concat([df_result, row], axis=0)
print("Hierarchical 0th")
row = conduct_Hierarchical(n_trials=n_trials,
n_samples=n_samples,
dataset=dataset,
changepoints=changepoints,
tolerance_delay=tolerance_delay,
order=0)
df_result = pd.concat([df_result, row], axis=0)
print("Hierarchical 1st")
row = conduct_Hierarchical(n_trials=n_trials,
n_samples=n_samples,
dataset=dataset,
changepoints=changepoints,
tolerance_delay=tolerance_delay,
order=1)
df_result = pd.concat([df_result, row], axis=0)
print("Hierarchical 2nd")
row = conduct_Hierarchical(n_trials=n_trials,
n_samples=n_samples,
dataset=dataset,
changepoints=changepoints,
tolerance_delay=tolerance_delay,
order=2)
df_result = pd.concat([df_result, row], axis=0)
print("AW2S_MDL")
row = conduct_AW2S_MDL(n_trials=n_trials,
n_samples=n_samples,
dataset=dataset,
changepoints=changepoints,
tolerance_delay=tolerance_delay)
df_result = pd.concat([df_result, row], axis=0)
return df_result
# TRAINING PARAMS
num_images_valid = len(filenames_valid)
print("Num Images Valid:", num_images_valid)
batch_size = 32
num_epochs = 1
num_steps_per_epoch_valid = int(
np.ceil(num_images_valid / batch_size)
) # use np.ceil to ensure model predictions cover the whole validation set
seed_train = None #587
seed_valid = None #seed_train+1
# Now create the training & validation datasets
dataset_valid_noshuffle = utils.create_dataset(filenames=filenames_valid,
labels=labels_valid_onehot,
num_channels=image_depth,
batch_size=batch_size,
shuffle_and_repeat=False)
print("DATASETS CREATED")
### PREDICTIONS
# Open tensorflow session
init = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
config = tf.ConfigProto()
#config.log_device_placement=True
with tf.Session(config=config) as sess:
sess.run(init)
print("TF SESSION OPEN")
def conduct_experiment(n_trials,
n_samples,
func,
transition_period,
tolerance_delay,
random_seed=0):
# fix seed for reproducibility
np.random.seed(random_seed)
mu_max = 1000
div_min = 1e-8
div_max = 1e8
df_result = pd.DataFrame()
print("Create Dataset")
dataset, changepoints = create_dataset(n_samples, func, transition_period)
print("ChangeFinder")
row = conduct_CF(n_trials=n_trials,
n_samples=n_samples,
dataset=dataset,
changepoints=changepoints,
tolerance_delay=tolerance_delay)
df_result = pd.concat([df_result, row], axis=0)
print("BOCPD")
row = conduct_BOCPD(n_trials=n_trials,
n_samples=n_samples,
dataset=dataset,
changepoints=changepoints,
tolerance_delay=tolerance_delay)
df_result = pd.concat([df_result, row], axis=0)
print("S-MDL 0th")
params = {
"mu_max": mu_max,
"div_min": div_min,
"div_max": div_max,
"order": 0
}
row = conduct_SDMDL(n_trials=n_trials,
n_samples=n_samples,
dataset=dataset,
changepoints=changepoints,
tolerance_delay=tolerance_delay,
params=params)
df_result = pd.concat([df_result, row], axis=0)
print("S-MDL 1st")
params = {
"mu_max": mu_max,
"div_min": div_min,
"div_max": div_max,
"order": 1
}
row = conduct_SDMDL(n_trials=n_trials,
n_samples=n_samples,
dataset=dataset,
changepoints=changepoints,
tolerance_delay=tolerance_delay,
params=params)
df_result = pd.concat([df_result, row], axis=0)
print("S-MDL 2nd")
params = {
"mu_max": mu_max,
"div_min": div_min,
"div_max": div_max,
"order": 2
}
row = conduct_SDMDL(n_trials=n_trials,
n_samples=n_samples,
dataset=dataset,
changepoints=changepoints,
tolerance_delay=tolerance_delay,
params=params)
df_result = pd.concat([df_result, row], axis=0)
print("FW2S-MDL")
params = {"mu_max": mu_max, "div_min": div_min, "div_max": div_max}
row = conduct_FW2S_MDL(n_trials=n_trials,
n_samples=n_samples,
dataset=dataset,
changepoints=changepoints,
tolerance_delay=tolerance_delay,
params=params)
df_result = pd.concat([df_result, row], axis=0)
return df_result
