embed_dim = 1000
lstm_dim = 1000
mlp_hidden_dims = 500
################################################################################
# segmentation network
################################################################################
# Inputs
text_seq_batch = tf.placeholder(tf.int32, [T, N]) # one batch per sentence
imcrop_batch = tf.placeholder(tf.float32, [N, 512, 512, 3])
_ = segmodel.text_objseg_upsample32s(text_seq_batch,
imcrop_batch,
num_vocab,
embed_dim,
lstm_dim,
mlp_hidden_dims,
vgg_dropout=False,
mlp_dropout=False)
load_var = {
var.op.name: var
for var in tf.all_variables()
if not var.op.name.startswith('classifier/upsample32s')
}
snapshot_loader = tf.train.Saver(load_var)
with tf.variable_scope('classifier', reuse=True):
upsample32s_w = tf.get_variable('upsample32s/weights')
init_upsample32s_w = tf.assign(upsample32s_w,
segmodel.generate_bilinear_filter(32))
num_vocab = 8803
embed_dim = 1000
lstm_dim = 1000
mlp_hidden_dims = 500
################################################################################
# segmentation network
################################################################################
# Inputs
text_seq_batch = tf.placeholder(tf.int32, [T, N]) # one batch per sentence
imcrop_batch = tf.placeholder(tf.float32, [N, 512, 512, 3])
_ = segmodel.text_objseg_upsample32s(text_seq_batch, imcrop_batch,
num_vocab, embed_dim, lstm_dim, mlp_hidden_dims,
vgg_dropout=False, mlp_dropout=False)
load_var = {var.op.name: var for var in tf.all_variables()
if not var.op.name.startswith('classifier/upsample32s')}
snapshot_loader = tf.train.Saver(load_var)
with tf.variable_scope('classifier', reuse=True):
upsample32s_w = tf.get_variable('upsample32s/weights')
init_upsample32s_w = tf.assign(upsample32s_w, segmodel.generate_bilinear_filter(32))
snapshot_saver = tf.train.Saver()
with tf.Session() as sess:
snapshot_loader.restore(sess, lowres_model)
sess.run(init_upsample32s_w)
snapshot_saver.save(sess, highres_model)