embeddingsdir = cwd + "../glove.6B/"
embedding_dim = 100
embedding_fdim = 64
embeddings, vocab = load_embeddings(
embeddingsdir=embeddingsdir,
embedding_dim=embedding_dim, # try 50
embedding_fdim=embedding_dim,
seed=None,
vocab=textworld_vocab)
index = np.random.permutation(range(embedding_dim))[:embedding_fdim]
embeddings = embeddings[index, :]
# instantiate network
network = nn.AlphaTextWorldNet(embeddings, vocab)
network(inputs={
'memory_input': tf.constant([[0]], tf.int32),
'cmdlist_input': tf.constant([[0]], tf.int32),
'location_input': tf.constant([0], tf.int32),
'cmdprev_input': tf.constant([[0]], tf.int32),
'ents2id': {
".": 0
},
'entvocab_input': tf.constant([[0]], tf.int32)
},
training=True)
# load latest weights if available
modeldir = cwd + "trained_models/"
models = glob.glob(modeldir + "*.h5")
rewards_list.append(d['reward'])
cmdlist_list.append(d['cmdlist'])
counts_list.append(d['counts'])
memory_list.append(d['memory'])
N = len(cmdlist_list)
idx = np.random.permutation(range(N))
cmdlist_list = [cmdlist_list[i] for i in idx]
memory_list = [memory_list[i] for i in idx]
counts_list = [counts_list[i] for i in idx]
rewards_list = [rewards_list[i] for i in idx]
print("number data points")
print(len(cmdlist_list))
model = nn.AlphaTextWorldNet(embeddings, vocab)
optim = tf.optimizers.Nadam(learning_rate=0.00001,
clipnorm=30.0,
beta_1=0.9,
beta_2=0.98)
ckpt = tf.train.Checkpoint(step=tf.Variable(1), optimizer=optim, net=model)
manager = tf.train.CheckpointManager(ckpt, root + 'ckpts', max_to_keep=100)
ckpt.restore(manager.latest_checkpoint)
if manager.latest_checkpoint:
print("Restored from {}".format(manager.latest_checkpoint))
else:
print("Initializing from scratch.")