im_batch = tf.placeholder(tf.float32, [1, None, None, 3])
bbox_batch = tf.placeholder(tf.float32, [None, 5])
spatial_batch = tf.placeholder(tf.float32, [None, 5])
text_seq_batch = tf.placeholder(tf.int32, [T, None])
scores = visgeno_attention_model.visgeno_attbilstm_net(im_batch, bbox_batch,
spatial_batch,
text_seq_batch,
num_vocab, embed_dim,
lstm_dim, False, False)
np.random.seed(3)
reader = DataReader(imdb_file,
vocab_file,
im_mean,
shuffle=False,
max_bbox_num=10000,
max_rel_num=10000)
################################################################################
# Snapshot and log
################################################################################
# Snapshot saver
snapshot_saver = tf.train.Saver()
# Start Session
sess = tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(
allow_growth=True)))
# Run Initialization operations
processed_W[l_name])
assign_B = tf.assign(tf.get_variable(l_name + '/biases'),
processed_B[l_name])
init_ops += [assign_W, assign_B]
processed_params.close()
# Initialize word embedding matrix
embedding_mat_val = np.load(wordembed_params)
with tf.variable_scope('lstm', reuse=True):
embedding_mat = tf.get_variable("embedding_mat", [num_vocab, embed_dim])
init_we = tf.assign(embedding_mat, embedding_mat_val)
init_ops += [init_we]
# Load data
reader = DataReader(imdb_file, vocab_file, im_mean)
################################################################################
# Snapshot and log
################################################################################
# Snapshot saver
snapshot_saver = tf.train.Saver()
# Log writer
if not os.path.isdir(log_dir):
os.mkdir(log_dir)
log_writer = tf.summary.FileWriter(log_dir, tf.get_default_graph())
loss_cls_ph = tf.placeholder(tf.float32, [])
lr_ph = tf.placeholder(tf.float32, [])
scores_ph = tf.placeholder(tf.float32, None)