def from_checkpoint_directory(path):
# type: (Path) -> MachineWithSingleNetwork
configuration = np.load(path / 'experiment-configuration.npy',
allow_pickle=True).item() # type: Any
data_directory = configuration['args'].data_directory # type: str
dataset = load_data(data_directory,
shuffle=False,
load_only_dicts=True)
with tf.variable_scope(configuration['which_network']):
raw_model = seq2seq_model(dataset.vocabulary_size,
50,
28,
cell_type='LSTM',
memory_dim=300,
num_layers=4,
dropout=0)
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.9)
session = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
best = MachineWithSingleNetwork.get_best_checkpoint_identifier(path)
best = str(path / 'best' / 'saved-model-attn-{}'.format(best))
raw_model.load_parameters(session, best)
return MachineWithSingleNetwork(configuration=configuration,
dataset=dataset,
raw_model=raw_model,
tf_session=session)
best_checkpoint = None
for checkpoint_name in os.listdir(
os.path.join(args.checkpoint_directory, 'best')):
if 'meta' in checkpoint_name:
this_checkpoint = int(checkpoint_name[17:].split('.')[0])
if best_checkpoint is None or this_checkpoint > best_checkpoint:
best_checkpoint = this_checkpoint
print "Resuming at", best_checkpoint, "..."
else:
best_checkpoint = args.resume_at
# Load data
dataset = load_data(args.data_directory, shuffle=False, load_only_dicts=True)
dictionary = dataset.get_tl_dictionary()
# Build the network
# scope = 'typo' if 'typo' in args.data_directory else 'ids'
scope = 'typo'
with tf.variable_scope(scope):
seq2seq = model(
dataset.vocabulary_size,
args.embedding_dim,
args.max_output_seq_len,
cell_type=args.cell_type,
memory_dim=args.memory_dim,
num_layers=args.num_layers,
dropout=0,