def train(cfg_file: str, ckpt=None) -> None:
# Load the config file
cfg = load_config(cfg_file)
# Set the random seed
set_seed(seed=cfg["training"].get("random_seed", 42))
# Load the data - Trg as (batch, # of frames, joints + 1 )
train_data, dev_data, test_data, src_vocab, trg_vocab = load_data(cfg=cfg)
# Build the Progressive Transformer model
model = build_model(cfg, src_vocab=src_vocab, trg_vocab=trg_vocab)
if ckpt is not None:
use_cuda = cfg["training"].get("use_cuda", False)
model_checkpoint = load_checkpoint(ckpt, use_cuda=use_cuda)
# Build model and load parameters from the checkpoint
model.load_state_dict(model_checkpoint["model_state"])
# for training management, e.g. early stopping and model selection
trainer = TrainManager(model=model, config=cfg)
# Store copy of original training config in model dir
shutil.copy2(cfg_file, trainer.model_dir + "/config.yaml")
# Log all entries of config
log_cfg(cfg, trainer.logger)
# Train the model
trainer.train_and_validate(train_data=train_data, valid_data=dev_data)
# Test the model with the best checkpoint
test(cfg_file)
def perform_single_test(network, train_config, seed, data_collect_freq):
observer = TestObserver(freq=data_collect_freq)
helpers.set_seed(seed=seed)
best_network = trainer.train_network(network=network,
config=train_config,
observer=observer)
return observer.get_results(), best_network
def __init__(self, train_set, validation_set, batch_size, epochs, lr, momentum, criterion, seed=None):
helpers.set_seed(seed)
self.train_set_loader = torch.utils.data.DataLoader(
train_set,
batch_size=batch_size,
shuffle=True
)
self.validation_set_loader = torch.utils.data.DataLoader(
validation_set,
batch_size=batch_size,
shuffle=True
)
self.epochs = epochs
self.lr = lr
self.momentum = momentum
self.criterion = criterion
outputs[t] = output
#decide if we are going to use teacher forcing or not
teacher_force = random.random() < teacher_forcing_ratio
#get the highest predicted token from our predictions
top1 = output.argmax(1)
#if teacher forcing, use actual next token as next input
#if not, use predicted token
input = trg[t] if teacher_force else top1
return outputs
set_seed(42)
BATCH_SIZE = 128
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
train_iterator, valid_iterator, test_iterator, SRC, TRG = get_data(
BATCH_SIZE, device, reverse=True)
INPUT_DIM = len(SRC.vocab)
OUTPUT_DIM = len(TRG.vocab)
ENC_EMB_DIM = 256
DEC_EMB_DIM = 256
HID_DIM = 512
N_LAYERS = 4
ENC_DROPOUT = 0.5
DEC_DROPOUT = 0.5
EXPERIMENT_SEEDS = range(1)
print(f'\n======================='
+ '\n Debugging mode:'
+ '\n=======================\n')
else:
EXPERIMENT_SEEDS = range(N_SEEDS)
print(f'Each experiment will be run with {len(EXPERIMENT_SEEDS)} different seeds.\n')
print('These are the embedding strategies that will be tested:')
for e in ALL_EMBEDDINGS:
print(e)
print('')
# Experiment loop:
for strategy in ALL_EMBEDDINGS:
for seed in EXPERIMENT_SEEDS:
# Set a seed for reproducibility
set_seed(seed_value=seed)
print(f'\nSeed: {seed + 1}/{len(EXPERIMENT_SEEDS)}.')
experiment = Experiment(
seed_value=seed,
embedding_strategy=strategy,
name=EXPERIMENT_NAME,
config=deepcopy(CONFIG))
experiment.run()