config = load_config(args.config)
lr_ph = tf.placeholder(tf.float32)
y_ph = tf.placeholder(tf.float32, [None, nway])
# 64: train_n_classes
trainnet = TransferNet(args.model_name, 64, isTr=True)
testnet = TransferNet(args.model_name, 64, isTr=False, reuse=True)
feat_emb = trainnet.outputs['embedding']
ff = trainnet.dense(feat_emb, nway, name='specific_dense')
pred = tf.nn.softmax(ff)
loss = cross_entropy(pred, y_ph)
ff = testnet.outputs['embedding']
ff = testnet.dense(ff, nway, name='specific_dense', reuse=True)
pred = tf.nn.softmax(ff)
acc = tf_acc(pred, y_ph)
rvlist = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
args.model_name + '/*')
opt = tf.train.AdamOptimizer(lr_ph)
update_op = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_op):
train_op = opt.minimize(loss)
# placeholder for initializing
var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
var_holder = [tf.placeholder(tf.float32, v.shape) for v in var_list]
init_ops = [var.assign(val) for var, val in zip(var_list, var_holder)]
saver = tf.train.Saver(rvlist)
embeds = [models[tn].outputs['embedding'] \
for tn in range(len(TRAIN_DATASET))]
embeds = tf.convert_to_tensor(embeds)
preds = [models[tn].outputs['pred'] \
for tn in range(len(TRAIN_DATASET))]
if args.transductive:
pemb = Lrn2evl(args.ens_name, embeds, nway)
pemb = tf.reshape(pemb, [-1,1,1])
else:
pemb = Lrn2evl(args.ens_name, embeds, nway, transduction=False)
pemb = tf.reshape(pemb, [len(TRAIN_DATASET), nway*qsize, 1])
ens_pred = tf.reduce_sum(pemb * preds, axis=0)
acc = tf_acc(ens_pred, qy_ph)
saver = tf.train.Saver()
sess = tf.Session()
sess.run(tf.global_variables_initializer())
load_loc = os.path.join(args.model_dir,
args.project_name,
args.ens_name + '.ckpt')
saver.restore(sess, load_loc)
ep_gen = EpisodeGenerator(args.dataset_dir, 'test', config)
target_dataset = ep_gen.dataset_list
#target_dataset = ['miniImagenet']
means = np.zeros([len(TRAIN_DATASET)+2, len(target_dataset)])
stds = np.zeros([len(TRAIN_DATASET)+2, len(target_dataset)])
for tdi, dset in enumerate(target_dataset):
fe = Lrn2evl(args.ens_name, embeds, nway) # (M,1)
fe = tf.reshape(fe, [-1, 1, 1]) # (M,1,1)
else:
fe = Lrn2evl(args.ens_name, embeds, nway,
transduction=False) # (MNQ,1)
fe = tf.reshape(fe,
[len(TRAIN_DATASET), nway * qsize, 1]) # (M,NQ,1)
pe = tf.reduce_sum(fe * preds, axis=0)
pe = tf.nn.softmax(pe)
train_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope=args.project_name + '/' +
args.ens_name + '/*')
acc = tf_acc(pe, qy_ph)
loss = cross_entropy(pe, qy_ph)
opt = tf.train.AdamOptimizer(lr_ph)
train_op = opt.minimize(loss, var_list=train_vars)
saver = tf.train.Saver()
sess = tf.Session()
sess.run(tf.global_variables_initializer())
if not args.from_ckpt:
for tn, tset in enumerate(TRAIN_DATASET):
saved_loc = os.path.join(args.model_dir, args.project_name,
tset + '.ckpt')
loaders[tn].restore(sess, saved_loc)
print('model_{} restored from {}'.format(tn, saved_loc))
else:
out_loc = os.path.join(args.model_dir, args.project_name,