def random_seed(seed):
"""Execute code inside this with-block using the specified random seed.
Sets the seed for random, numpy.random and torch (CPU).
WARNING: torch GPU seeds are NOT set!
Does not affect the state of random number generators outside this block.
Not thread-safe.
Args:
seed (int)
"""
state = RandomState()
random.seed(seed) # alter state
np.random.seed(seed)
torch.manual_seed(seed)
yield
state.set_global()
def __init__(self, model, optim, src_dict, trg_dict, model_params=None, gpu_ids=None, random_seed=None):
self._logger = logging.getLogger('nmmt.NMTEngineTrainer')
self._log_level = logging.INFO
self._model = model
self._optim = optim
self._src_dict = src_dict
self._trg_dict = trg_dict
self._model_params = model_params
self._gpu_ids = gpu_ids
if random_seed is not None:
torch.manual_seed(random_seed)
random.manual_seed_all(random_seed)
# Public-editable options
self.log_interval = 50 # Log status every 'log_interval' updates
self.batch_size = 64 # Maximum batch size
self.max_generator_batches = 32 # Maximum batches of words in a seq to run the generator on in parallel.
self.max_epochs = 40 # Maximum number of training epochs
self.min_epochs = 10 # Minimum number of training epochs
self.start_epoch = 1 # The epoch from which to start
self.min_perplexity_decrement = .02 # If perplexity decrement is lower than this percentage, stop training
# load mosi
mm_dset = MultiModalDataset(dataset, task, approach)
# get dataset length
l = len(mm_dset)
# split dataset to train-valid-test
test_size = int(0.2 * l)
train_size = l - 2 * test_size
# reproducability
if DEVICE == "cuda:1":
torch.backends.cudnn.deterministic = True
torch.manual_seed(64)
mm_train, mm_valid, mm_test = random_split(mm_dset,
[train_size, test_size, test_size])
# use dataloaders wrappers
train_loader = DataLoader(mm_train, batch_size=BATCH_SIZE, shuffle=True)
valid_loader = DataLoader(mm_valid)
test_loader = DataLoader(mm_test)
#######################################
#### text rnn hyperparams ######
######################################
text_hyperparameters = []
input_size = 300 # glove size
hidden_size = 64 * 2 # hidden state size
num_layers = 1 # how many stacked rnn's
bidirectional = True