evalute for CNN model

[yhgon]

Dear @xujinfan

I want to reproduce The search space is 5 in convolutional layers, 25 in fully-connected layers for RCL,DEN and PGN.

plz comment for define target variable for evaluate function.

TODO, I made basic 5 CNN model with 2 CNN(5x5,32), (5x5,64) + 2 Maxpool + 1 FCN(1024) as below :
https://github.com/yhgon/Reinforced-Continual-Learning/blob/master/RCL_CNN.py#L53

with tf.Graph().as_default() as g:
    with tf.name_scope("before"):
	inputs = tf.placeholder(shape=(None, 784), dtype=tf.float32)
	y = tf.placeholder(shape=(None, 10), dtype=tf.float32)
	w1 = tf.Variable(tf.truncated_normal([5,5,1,32], stddev=0.1))
	b1 = tf.Variable(tf.constant(0.1, shape=(32,)))
	w2 = tf.Variable(tf.truncated_normal([5,5,32,64], stddev=0.1))
	b2 = tf.Variable(tf.constant(0.1, shape=(64,)))
	w3 = tf.Variable(tf.truncated_normal([2*2*64,1024], stddev=0.1))
	b3 = tf.Variable(tf.constant(0.1, shape=(1024, )))
	w4 = tf.Variable(tf.truncated_normal([1024,10], stddev=0.1))
	b4 = tf.Variable(tf.constant(0.1, shape=(10,)))
	## model
	inputs_shape= inputs.get_shape().as_list()
	print("DEBUG input_shape before:",inputs_shape)
	inputx=tf.reshape(inputs, shape=[-1,28,28,1]) # 28x28
	inputs_shape= inputx.get_shape().as_list()
	print("DEBUG input_shape after :",inputs_shape)
	conv1 = tf.nn.relu(tf.nn.conv2d(inputx, w1,    strides=[1,2,2,1], padding='SAME') + b1)
	conv1 = tf.nn.max_pool(conv1, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')
	conv2 = tf.nn.relu(tf.nn.conv2d(conv1, w2,     strides=[1,2,2,1], padding='SAME') + b2)
	conv2 = tf.nn.max_pool(conv2, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')
	conv2_shape = conv2.get_shape().as_list()
	print("DEBUG conv2_shape before :",conv2_shape)
	conv2 = tf.reshape(conv2, [-1, conv2_shape[1] * conv2_shape[2] * conv2_shape[3]])
	conv2_shape = conv2.get_shape().as_list()
	print("DEBUG conv2_shape after :",conv2_shape)
	fcn= tf.nn.relu(tf.matmul(conv2, w3) + b3)
	output3=tf.matmul(fcn, w4) + b4
	loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=output3)) + \
	       0.0001*(tf.nn.l2_loss(w1) + tf.nn.l2_loss(w2) + tf.nn.l2_loss(w3) + tf.nn.l2_loss(w4))

task 0 for initial training works well as below :

--batch_size 1024 --n_epochs 30 --n_tasks 10 --n_layers 2 --hidden_size 10 --num_layers 2 --max_trials 40 --actions_num 2
DEBUG : task0/10 IF
DEBUG input_shape before: [None, 784]
DEBUG input_shape after : [None, 28, 28, 1]
DEBUG conv2_shape before : [None, 2, 2, 64]
DEBUG conv2_shape after : [None, 256]

task 0/10  epoch 29  train_step 53248/55000
task 0/10  epoch 29  train_step 54272/55000
task 0/10  epoch 29  train_step 55296/55000
task 0/10 test accuracy: 0.9629999995231628 IF
task 1/10 trial 0/40 *********actions  for [13, 14] ELSE 


DEBUG :  0 (5, 1, 32)
DEBUG :  1 ()
DEBUG :  2 (5, 32, 64)
DEBUG :  3 ()
DEBUG :  4 (1024,)
DEBUG :  5 ()
DEBUG : TODO 

I wonder how to set action for 2d conv and range for var_list.

• conv1 : 5x5 conv filter , 32
• conv2 : 5x5 conv , 64
• FCN : 1024

current actions for [13, 14] use (5,1,32) instead of (5x5,1,3) of (5x5,1,32)
IMHO, right behavior for taget would be control 32.

task 1/10 trial 0/40 *********actions  for  [13, 14] ELSE 
DEBUG :  0 (5, 1, 32)
DEBUG :  1 ()
DEBUG :  2 (5, 32, 64)
DEBUG :  3 ()
DEBUG :  4 (1024,)
DEBUG :  5 ()
DEBUG : TODO 

in FCN task1, var_list is below :

#DEBUG :  0 (312,)
#DEBUG :  1 ()
#DEBUG :  2 (128,)
#DEBUG :  3 ()
#DEBUG :  4 (10,)
#DEBUG :  5 ()  

how to control the target?
in your paper, do you also control conv filter size such as 3x3, 5x5, 1x1 ?