def record(seed, use_seed, device, epochs, batch_size, representations,
metrics):
"""
Start and record the learning phase of the neural network.
"""
from user_data import datasets, models
if not os.path.exists(RECORDS_PATH):
os.mkdir(RECORDS_PATH)
if use_seed:
if seed is None:
seed = DEFAULT_SEED
else:
click.echo(f'Seed set to {seed} for Python, Numpy and PyTorch.\n')
set_seeds(seed)
else:
click.echo('Not using a seed. Results will not be reproducible...\n')
device = set_device(device)
click.echo('Loaded models:')
for model in models.keys():
click.echo(f'- {model}')
click.echo('')
click.echo('Loaded datasets:')
for ds in datasets:
click.echo(f'- {ds.name}')
click.echo('')
if metrics == 'all':
metrics = metrics_list.keys()
else:
metrics = re.sub(r"[^\w]", ' ', metrics).split()
#click.echo('Metrics:')
#for metric in metrics:
# click.echo(f'- {metric}')
#click.echo('')
if representations == 'all':
representations = list(projections_list.keys())
else:
representations = re.sub(r"[^\w]", ' ', representations).split()
click.echo('Representations:')
for representation in representations:
click.echo(f'- {representation}')
click.echo('')
start_record(models, datasets, device, batch_size, representations, epochs,
metrics)
#V0 is two level loss straight, no balance
#V1 is the switching loss on 2 loss (top and both level)
#V1.5 is a training on the second level, then finetune with information on first level. Model used will need to be hand picked
#V2 is the iterating version of 3 losses (top low both levels)
#V3 is module learning it's loss weights
sentence_hidden_state_size = 20
paragraph_hidden_state_size = 10
nb_layers = 1
patience = 20
cooldown = 5
batch_size=4
epoch = 70
seed = 420
drop_out = 0.5
set_seeds(seed)
optimizer = "adam"
test_name = "paper_loss_phs{}_shs{}_patience{}_cd{}_{}l_bs{}_op{}_do{}_noupscale_seed{}_test".format(
paragraph_hidden_state_size, sentence_hidden_state_size, patience, cooldown, nb_layers, batch_size, optimizer,drop_out, seed
)
train_pickled = pickle.load(open("./data/{}/{}/preprocessed_train.pkl".format(aggregation, dataset),'rb'))
valid_pickled = pickle.load(open("./data/{}/{}/preprocessed_valid.pkl".format(aggregation, dataset), 'rb'))
test_pickled = pickle.load(open("./data/{}/{}/preprocessed_test.pkl".format(aggregation, dataset), 'rb'))
# preprocessed_train = [datapoint for _, _, _, _,datapoint in train_pickled]
# preprocessed_valid = [datapoint for _, _, _, _, datapoint in valid_pickled]