class Model(object):
def __init__(self, config,
debug_information=False,
is_train=True):
self.debug = debug_information
self.XO={}
self.pen=None
self.config = config
self.batch_size = self.config.batch_size
self.img_size = self.config.data_info[0]
self.c_dim = self.config.data_info[2]
self.q_dim = self.config.data_info[3]
self.a_dim = self.config.data_info[4]
self.conv_info = self.config.conv_info
self.acc=0
self.feat_count = 64
self.ilp_params = None
# create placeholders for the input
self.img = tf.placeholder(
name='img', dtype=tf.float32,
shape=[self.batch_size, self.img_size, self.img_size, self.c_dim],
)
self.q = tf.placeholder(
name='q', dtype=tf.float32, shape=[self.batch_size, self.q_dim],
)
self.a = tf.placeholder(
name='a', dtype=tf.float32, shape=[self.batch_size, self.a_dim],
)
self.is_training = tf.placeholder_with_default(bool(is_train), [], name='is_training')
self.build(is_train=is_train)
def load_ilp_config(self):
parser = argparse.ArgumentParser()
batch_size = self.batch_size
# extra_feat_count=16
# parser.add_argument('--CHECK_CONVERGENCE',default=1,help='Check for convergence',type=int)
# parser.add_argument('--SHOW_PRED_DETAILS',default=0,help='Print predicates definition details',type=int)
parser.add_argument('--BS',default=16,help='Batch Size',type=int)
parser.add_argument('--T',default=1 ,help='Number of forward chain',type=int)
# parser.add_argument('--LR_SC', default={ (-1000,2):.005 , (2,1e5):.01} , help='Learning rate schedule',type=dict)
# parser.add_argument('--BINARAIZE', default=1 , help='Enable binrizing at fast convergence',type=int)
# parser.add_argument('--MAX_DISP_ITEMS', default=10 , help='Max number of facts to display',type=int)
# parser.add_argument('--DISP_BATCH_VALUES',default=[],help='Batch Size',type=list)
# parser.add_argument('--W_DISP_TH', default=.2 , help='Display Threshold for weights',type=int)
# parser.add_argument('--ITER', default=400000, help='Maximum number of iteration',type=int)
# parser.add_argument('--ITER2', default=200, help='Epoch',type=int)
# parser.add_argument('--PRINTPRED',default=1,help='Print predicates',type=int)
# parser.add_argument('--PRINT_WEIGHTS',default=0,help='Print raw weights',type=int)
parser.add_argument('--MAXTERMS',default=6 ,help='Maximum number of terms in each clause',type=int)
parser.add_argument('--L1',default=0 ,help='Penalty for maxterm',type=float)
parser.add_argument('--L2',default=0 ,help='Penalty for distance from binary for weights',type=float)
parser.add_argument('--L3',default=0 ,help='Penalty for distance from binary for each term',type=float)
parser.add_argument('--L2LOSS',default=0,help='Use L2 instead of cross entropy',type=int)
parser.add_argument('--SYNC',default=0,help='Synchronized Update',type=int)
# parser.add_argument('--ALLTIMESTAMP',default=0 ,help='Add loss for each timestamp',type=int)
# parser.add_argument('--FILT_TH_MEAN', default=.5 , help='Fast convergence total loss threshold MEAN',type=float)
# parser.add_argument('--FILT_TH_MAX', default=.5 , help='Fast convergence total loss threshold MAX',type=float)
# parser.add_argument('--OPT_TH', default=.05 , help='Per value accuracy threshold',type=float)
# parser.add_argument('--PLOGENT', default=.50 , help='Crossentropy coefficient',type=float)
# parser.add_argument('--BETA1', default=.90 , help='ADAM Beta1',type=float)
# parser.add_argument('--BETA2', default=.999 , help='ADAM Beta2',type=float)
# parser.add_argument('--EPS', default=1e-6, help='ADAM Epsillon',type=float)
parser.add_argument('--GPU', default=1, help='Use GPU',type=int)
# parser.add_argument('--LOGDIR', default='./logs/Logic', help='Log Dir',type=str)
parser.add_argument('--TB', default=0, help='Use Tensorboard',type=int)
parser.add_argument('--SEED',default=0,help='Random seed',type=int)
# parser.add_argument('--ADDGRAPH', default=1, help='Add graph to Tensorboard',type=int)
# parser.add_argument('--CLIP_NORM', default=0, help='Clip gradient',type=float)
self.args_ilp = parser.parse_args()
def define_preds(self):
nC=6
nD=16
C = ['%d'%i for i in range(nC)]
D = ['%d'%i for i in range(nD)]
self.Constants = dict( { 'C':C, 'D':D}) #, 'N':['%d'%i for i in range(6)] })
self.predColl = PredCollection (self.Constants)
self.predColl.add_pred(dname='pos' ,arguments=['C','D' ])
self.predColl.add_pred(dname='question' ,arguments=['C'])
self.predColl.add_pred(dname='eq' ,arguments=['D','D' ])
self.predColl.add_pred(dname='ltD' ,arguments=['D','D','D'])
self.predColl.add_pred(dname='gtD' ,arguments=['D','D','D'])
self.predColl.add_pred(dname='left' ,arguments=['D'])
self.predColl.add_pred(dname='button' ,arguments=['D'])
for i in range(self.q_dim):
self.predColl.add_pred(dname='is_q_%d'%i ,arguments=[])
for i in range(nC):
self.predColl.add_pred(dname='is_color_%d'%i ,arguments=['C'])
# ,pFunc =
# DNF('obj',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['is_l_0(A)'],predColl=self.predColl,fast=True) , use_neg=False, Fam='or')
self.predColl.add_pred(dname='rectangle' ,arguments=['C'])
self.predColl.add_pred(dname='exist' ,arguments=['C'])
self.predColl.add_pred(dname='eqC' ,arguments=['C','C'],variables=['D','D'] ,pFunc =
DNF('eqC',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['pos(A,C), pos(B,D), eq(C,D)'],predColl=self.predColl,fast=True) , use_neg=True, Fam='or')
self.predColl.add_pred(dname='closer',arguments=['C','C','C'], variables=['D','D','D'] ,pFunc =
DNF('closer',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['pos(A,D), pos(B,E), pos(C,F), exist(A), exist(B), exist(C), ltD(D,E,F)'],predColl=self.predColl,fast=True) , use_neg=False, Fam='or')
self.predColl.add_pred(dname='farther',arguments=['C','C','C'], variables=['D','D','D'] ,pFunc =
DNF('farther',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['pos(A,D), pos(B,E), pos(C,F), exist(A), exist(B), exist(C), gtD(D,E,F)'],predColl=self.predColl,fast=True) , use_neg=False, Fam='or')
self.predColl.add_pred(dname='closest',arguments=['C','C'], variables=['C'] ,pFunc =
DNF('closest',terms=4,init=[1,.1,-1,.1],sig=2,init_terms=['closer(A,C,B)','not exist(A)','not exist(B)','eqC(A,B)'],predColl=self.predColl,fast=True,neg=True) , use_neg=True, Fam='eq')
self.predColl.add_pred(dname='farthest',arguments=['C','C'], variables=['C'] ,pFunc =
DNF('farthest',terms=4,init=[1,.1,-1,.1],sig=2,init_terms=['farther(A,C,B)','not exist(A)','not exist(B)','eqC(A,B)'],predColl=self.predColl,fast=True,neg=True) , use_neg=True, Fam='eq')
exc = ['CL_%d'%i for i in range(self.a_dim)]
exc=[]
for k in range(0,self.a_dim):
# # # # exc = ['CL_%d'%i for i in range(k,self.a_dim)]
# # # # self.predColl.add_pred(dname='CL0_%d'%k,oname='CL_%d'%k,arguments=[] , variables=[] , pFunc = DNF('CL0_%d'%k,predColl=self.predColl,terms=10 ,init=[-1,.1,-1,.1],sig=2) ,use_neg=True, Fam='eq', exc_preds=exc )
# # self.predColl.add_pred(dname='CL1_%d'%k,oname='CL_%d'%k,arguments=[] , variables=['D'] , pFunc = DNF('CL1_%d'%k,predColl=self.predColl,terms=6 ,init=[-1,.1,-1,.1],sig=2) ,use_neg=True, Fam='eq',exc_conds=[('*','rep1') ] )
if k==0:
post_terms=[]
else:
post_terms=[ ('and', 'not CL_%d()'%j ) for j in range(k)]
post_terms=[]
self.predColl.add_pred(dname='CL_%d'%k,oname='CL_%d'%k,arguments=[] , variables=['C','C','D'] , pFunc = DNF('CL_%d'%k,predColl=self.predColl,terms=14,init=[-1,-1,-1,.1],sig=2, post_terms=post_terms) ,use_neg=True, Fam='eq',exc_conds=[('*','rep1') ] ,exc_preds=exc ) #
# # # # self.predColl.add_pred(dname='CL_%d'%k,oname='CL_%d'%k,arguments=[] , variables=['D','D'] , pFunc = MLP('CL_%d'%k,dims=[200,1], acts=[tf.nn.relu,tf.sigmoid] ) ,use_neg=False, Fam='eq', exc_preds=exc )
self.predColl.initialize_predicates()
self.bg = Background( self.predColl )
# self.bg.add_backgroud('notExist', ('%d'%(nD-1),))
for i in range(nC):
self.bg.add_backgroud('is_color_%d'%i, ('%d'%i,))
for i in range(nD):
ri,ci=int(i//4),int(i%4)
if ri>=2:
self.bg.add_backgroud('button', ('%d'%i,))
if ci<2:
self.bg.add_backgroud('left', ('%d'%i,))
# self.bg.add_backgroud('-%d'%i , ('%d'%i,) )
self.bg.add_backgroud('eq', ('%d'%i,'%d'%i))
for j in range(nD):
rj,cj=int(j//4),int(j%4)
for k in range(nD):
rk,ck=int(k//4),int(k%4)
d1=(ri-rj)**2+(ci-cj)**2
d2=(ri-rk)**2+(ci-ck)**2
if(d1<d2 and i!=j and i!=k and j!=k):
self.bg.add_backgroud('ltD', ('%d'%i,'%d'%j,'%d'%k))
if(d1>d2 and i!=j and i!=k and j!=k):
self.bg.add_backgroud('gtD', ('%d'%i,'%d'%j,'%d'%k))
a = '%d'%i
# self.bg.add_backgroud('is_c_%d'%i , (a,) )
# self.bg.add_backgroud('is_r_%d'%i , (a,) )
# self.bg.add_backgroud('eqC' , (a,a) )
# self.bg.add_backgroud('eqR' , (a,a) )
# for j in range(nC):
# if i<j:
# self.bg.add_backgroud('ltC', ('%d'%i, '%d'%j))
# self.bg.add_backgroud('ltR', ('%d'%i, '%d'%j))
bg_set=[]
self.X0=OrderedDict()
for p in self.predColl.outpreds:
if p.oname not in bg_set:
tmp = tf.expand_dims( tf.constant( self.bg.get_X0(p.oname) ,tf.float32) , 0)
self.X0[p.oname] = tf.tile( tmp , [self.batch_size,1] )
print('displaying config setting...')
# for arg in vars(args):
# print( '{}-{}'.format ( arg, getattr(args, arg) ) )
self.mdl = ILPRLEngine( args=self.args_ilp ,predColl=self.predColl ,bgs=None )
def get_feed_dict(self, batch_chunk, step=None, is_training=None):
fd = {
self.img: batch_chunk['img'], # [B, h, w, c]
self.q: batch_chunk['q'], # [B, n]
self.a: batch_chunk['a'], # [B, m]
}
if is_training is not None:
fd[self.is_training] = is_training
return fd
def build(self, is_train=True):
n = self.a_dim
conv_info = self.conv_info
# build loss and accuracy {{{
def build_loss(logits, labels):
# Cross-entropy loss
loss = tf.nn.softmax_cross_entropy_with_logits(logits=logits*10, labels=labels)
# loss = tf.reduce_sum( neg_ent_loss (labels,logits) , -1 )
# Classification accuracy
correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(labels, 1))
self.acc = tf.cast(correct_prediction, tf.float32)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
return tf.reduce_mean(loss), accuracy
# }}}
# def concat_coor(o, i, d):
# coor = tf.tile(tf.expand_dims(
# [float(int(i / d)) / d, (i % d) / d], axis=0), [self.batch_size, 1])
# o = tf.concat([o, tf.to_float(coor)], axis=1)
# return o
# def g_theta(o_i, o_j, q, scope='g_theta', reuse=True):
# with tf.variable_scope(scope, reuse=reuse) as scope:
# if not reuse: log.warn(scope.name)
# g_1 = fc(tf.concat([o_i, o_j, q], axis=1), 256, name='g_1')
# g_2 = fc(g_1, 256, name='g_2')
# g_3 = fc(g_2, 256, name='g_3')
# g_4 = fc(g_3, 256, name='g_4')
# return g_4
# Classifier: takes images as input and outputs class label [B, m]
def CONV(img, q, scope='CONV'):
nD = 16
nC=6
with tf.variable_scope(scope) as scope:
log.warn(scope.name)
self.X0['question'] = q[:,:nC]
x1 = tf.layers.conv2d( img, 32*2, 3, strides=(2,2) , activation='tanh', padding='valid')
x2 = tf.layers.conv2d( x1, 32*2, 3, strides=(2,2), activation='tanh' , padding='valid' )
x2 = tf.layers.conv2d( x2, 32*2, 3, strides=(2,2), activation='tanh' , padding='valid' )
sz=x2.shape[1].value
feat = tf.reshape(x2, [-1,sz*sz,32*2])
def feat2scores2(feat,nX,nF):
# pos = tf.constant( np.eye(nF) , tf.float32 )
w = weight_variable( [ feat.shape[-1].value,nF] )
b = bias_variable([nF,])
m = tf.matmul( feat,w+b)
# mm = tf.matmul( m , pos)
mm = tf.nn.softmax( m,1)
mmm = tf.transpose( mm, [0,2,1] )
return tf.layers.dense( mmm, nX, tf.nn.softmax )
with tf.variable_scope('conv_pos') as scope:
# Fs = feat2scores2(feat,nC)
# rect = tf.layers.dense( Fs, 3, tf.nn.softmax )
rect = feat2scores2(feat,nC,3)
rect = tf.transpose( rect, [0,2,1] )
# with tf.variable_scope('obj'):
# # obj = tf.layers.dense( obj, 2, tf.nn.softmax )
with tf.variable_scope('cl'):
pos = feat2scores2(feat,nC,nD)
# Fs = feat2scores2(feat,nC)
# pos = tf.layers.dense( feat, nD, tf.nn.softmax )
pos = tf.transpose( pos, [0,2,1] )
# cl1,cl2 = get_conv(1,[nL1,nL2],24)
# self.X0['obj'] = obj[:,0,:]
self.X0['rectangle'] = rect[:,:,1]
self.X0['exist'] = 1.0-rect[:,:,0]
self.X0['pos'] = tf.reshape( pos, [-1, nC*nD])
for i in range(self.q_dim):
self.X0['is_q_%d'%i] = q[:,i:(i+1)]
with tf.variable_scope('myscope'):
self.XO,L3 = self.mdl.getTSteps(self.X0)
os = tf.concat( [self.XO['CL_%d'%i] for i in range(self.a_dim)],-1)
return os
# return fc(os, self.a_dim, activation_fn=None, name='fc_3')
# # os=os+ fc(os, 1, activation_fn=tf.sigmoid, name='fc_2')*0
# # return os
# all_g = tf.concat( [self.XO[i.oname] for i in self.predColl.outpreds],-1)
# all_g = fc( 2*all_g-1, 256, activation_fn=tf.nn.relu, name='fc_1')
# # all_g = slim.dropout(all_g, keep_prob=0.5, is_training=is_train, scope='fc_3/')
# all_g = fc(all_g, 256, activation_fn=tf.nn.relu, name='fc_2')
# # all_g = slim.dropout(all_g, keep_prob=0.8, is_training=is_train, scope='fc_2')
# return fc(all_g, self.a_dim, activation_fn=None, name='fc_3')
# # all_g = tf.concat( [XO[i.oname] for i in self.predColl.outpreds if 'aux' in i.dname],-1)
# os = tf.concat( [self.XO['CL_%d'%i] for i in range(self.a_dim)],-1)
# # os=os+ fc(os, 1, activation_fn=tf.sigmoid, name='fc_2')*0
# return os*5
# return all_g
# return all_g
def f_phi(g, scope='f_phi'):
with tf.variable_scope(scope) as scope:
log.warn(scope.name)
fc_1 = fc(g, 256, name='fc_1')
fc_2 = fc(fc_1, 256, name='fc_2')
# fc_2 = slim.dropout(fc_2, keep_prob=0.5, is_training=is_train, scope='fc_3/')
fc_3 = fc(fc_2, n, activation_fn=None, name='fc_3')
return fc_3
self.load_ilp_config()
self.define_preds()
logits = CONV(self.img, self.q, scope='CONV')
# logits = f_phi(g, scope='f_phi')
self.all_preds = tf.nn.softmax(logits)
self.loss, self.accuracy = build_loss(logits, self.a)
# Add summaries
def draw_iqa(img, q, target_a, pred_a):
fig, ax = tfplot.subplots(figsize=(6, 6))
ax.imshow(img)
ax.set_title(question2str(q))
ax.set_xlabel(answer2str(target_a)+answer2str(pred_a, 'Predicted'))
return fig
try:
tfplot.summary.plot_many('IQA/',
draw_iqa, [self.img, self.q, self.a, self.all_preds],
max_outputs=4,
collections=["plot_summaries"])
except:
pass
tf.summary.scalar("loss/accuracy", self.accuracy)
tf.summary.scalar("loss/cross_entropy", self.loss)
log.warn('Successfully loaded the model.')
class ILP_MODEL(object):
def __init__(self, DIM1, DIM2, F_COUNT=3):
self.DIM1 = DIM1
self.DIM2 = DIM2
self.F_COUNT = F_COUNT
self.args = self.load_ilp_config()
self.define_preds()
self.Xo = None
self.X0 = None
def load_ilp_config(self):
param = dotdict({})
param.BS = 1
param.T = 1
param.W_DISP_TH = .2
param.GPU = 1
return param
def define_preds(self):
X = ['%d' % i for i in range(self.DIM1)]
Y = ['%d' % i for i in range(self.DIM2)]
self.Constants = dict({'N': X, 'Y': Y})
self.predColl = PredCollection(self.Constants)
self.predColl.add_pred(dname='sameX', arguments=['N', 'N'])
self.predColl.add_pred(dname='sameY', arguments=['Y', 'Y'])
self.predColl.add_pred(dname='X_U', arguments=['N'])
self.predColl.add_pred(dname='X_D', arguments=['N'])
self.predColl.add_pred(dname='Y_L', arguments=['Y'])
self.predColl.add_pred(dname='Y_R', arguments=['Y'])
self.predColl.add_pred(dname='ltY', arguments=['Y', 'Y'])
self.predColl.add_pred(dname='close', arguments=['Y', 'Y'])
self.predColl.add_pred(dname='agent', arguments=['N', 'Y'])
self.predColl.add_pred(dname='predLR', arguments=['N', 'Y'])
self.predColl.add_pred(dname='predRL', arguments=['N', 'Y'])
self.predColl.add_pred(dname='food', arguments=['N', 'Y'])
count_type = 'max'
self.predColl.add_pred(dname='pred',
arguments=['N', 'Y'],
variables=[],
pFunc=DNF(
'pred',
init_terms=['predLR(A,B)', 'predRL(A,B)'],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq',
count_type='or',
arg_funcs=[])
self.predColl.add_pred(dname='agentX',
arguments=['N'],
variables=['Y'],
pFunc=DNF('agentX',
init_terms=['agent(A,B)'],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq',
count_type='or',
arg_funcs=[])
self.predColl.add_pred(dname='agentY',
arguments=['Y'],
variables=['N'],
pFunc=DNF('agentY',
init_terms=['agent(B,A)'],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq',
count_type='or',
arg_funcs=[])
self.predColl.add_pred(
dname='C1',
arguments=[],
variables=['N', 'N'],
pFunc=DNF('C1',
terms=1,
init=[-1, .1, -1, .1],
sig=2,
init_terms=['agentX(A), agentX(B), not sameX(A,B)'],
predColl=self.predColl,
fast=True),
use_neg=True,
Fam='eq',
count_type='max',
arg_funcs=[])
self.predColl.add_pred(
dname='C2',
arguments=[],
variables=['Y', 'Y'],
pFunc=DNF('C2',
terms=1,
init=[-1, .1, -1, .1],
sig=2,
init_terms=['agentY(A), agentY(B), not close(A,B)'],
predColl=self.predColl,
fast=True),
use_neg=True,
Fam='eq',
count_type='max',
arg_funcs=[])
self.predColl.add_pred(dname='C3',
arguments=[],
variables=['N', 'Y'],
pFunc=DNF('C3',
terms=1,
init=[-1, .1, -1, .1],
sig=2,
init_terms=[
'predLR(A,B), predRL(A,B)',
'agent(A,B), predRL(A,B)',
'agent(A,B), predLR(A,B)',
],
predColl=self.predColl,
fast=True),
use_neg=True,
Fam='eq',
count_type='max',
arg_funcs=[])
self.predColl.add_pred(dname='C4',
arguments=[],
variables=['N', 'Y'],
pFunc=DNF('C4',
terms=1,
init=[-1, .1, -1, .1],
sig=2,
init_terms=['agent(A,B)'],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq',
count_type='max',
arg_funcs=[])
excs = [
'action_noop', 'action_up', 'action_down', 'action_left',
'action_right', 'Q'
]
w = [-1, .1, -3, .1]
self.predColl.add_pred(dname='en_up',
arguments=[],
variables=['N', 'Y', 'Y'],
pFunc=DNF(
'en_up',
terms=1,
init=[-1, .1, -1, .1],
sig=2,
init_terms=[
'agent(A,B), X_U(A)',
'agent(A,B), pred(M_A,C), close(B,C)'
],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq',
count_type=count_type,
count_th=100,
arg_funcs=['M'])
self.predColl.add_pred(dname='en_down',
arguments=[],
variables=['N', 'Y', 'Y'],
pFunc=DNF(
'en_down',
terms=1,
init=[-1, .1, -1, .1],
sig=2,
init_terms=[
'agent(A,B), X_D(A)',
'agent(A,B), pred(P_A,C), close(B,C)'
],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq',
count_type=count_type,
count_th=100,
arg_funcs=['P'])
self.predColl.add_pred(
dname='en_right',
arguments=[],
variables=['N', 'Y', 'Y'],
pFunc=DNF('en_right',
terms=1,
init=[-1, .1, -1, .1],
sig=2,
init_terms=[
'agent(A,B), Y_R(B)',
'agent(A,B), pred(A,C), close(B,C), ltY(B,C)'
],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq',
count_type=count_type,
count_th=100,
arg_funcs=[])
self.predColl.add_pred(
dname='en_left',
arguments=[],
variables=['N', 'Y', 'Y'],
pFunc=DNF('en_left',
terms=1,
init=[-1, .1, -1, .1],
sig=2,
init_terms=[
'agent(A,B), Y_L(B)',
'agent(A,B), pred(A,C), close(B,C), ltY(C,B)'
],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq',
count_type=count_type,
count_th=100,
arg_funcs=[])
self.predColl.add_pred(
dname='en_noop',
arguments=[],
variables=[],
pFunc=DNF(
'en_noop',
terms=1,
init=[-1, .1, -1, .1],
sig=2,
#init_terms=['agent(A,B), predRL(A,C), ltY(C,B)','agent(A,B), predLR(A,C), ltY(B,C)' ],
init_terms=['en_right()', 'en_left()'],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq',
count_type=count_type,
count_th=100)
pt = [('and', 'not en_noop()')]
self.predColl.add_pred(dname='action_noop',
arguments=[],
variables=['N', 'Y', 'Y'],
pFunc=DNF('action_noop',
terms=8,
init=w,
sig=2,
predColl=self.predColl,
fast=False,
post_terms=pt),
use_neg=True,
Fam='eq',
exc_preds=excs,
count_type=count_type,
arg_funcs=['M'])
pt = [('and', 'not en_up()')]
self.predColl.add_pred(dname='action_up',
arguments=[],
variables=['N', 'Y', 'Y'],
pFunc=DNF('action_up',
terms=8,
init=w,
sig=2,
predColl=self.predColl,
fast=False,
post_terms=pt),
use_neg=True,
Fam='eq',
exc_preds=excs,
count_type=count_type,
count_th=100,
arg_funcs=['M'])
pt = [('and', 'not en_right()')]
self.predColl.add_pred(dname='action_right',
arguments=[],
variables=['N', 'Y', 'Y'],
pFunc=DNF('action_right',
terms=8,
init=w,
sig=2,
predColl=self.predColl,
fast=False,
post_terms=pt),
use_neg=True,
Fam='eq',
exc_preds=excs,
count_type=count_type,
arg_funcs=['M'])
pt = [('and', 'not en_left()')]
self.predColl.add_pred(dname='action_left',
arguments=[],
variables=['N', 'Y', 'Y'],
pFunc=DNF('action_left',
terms=8,
init=w,
sig=2,
predColl=self.predColl,
fast=False,
post_terms=pt),
use_neg=True,
Fam='eq',
exc_preds=excs,
count_type=count_type,
arg_funcs=['M'])
pt = [('and', 'not en_down()')]
self.predColl.add_pred(dname='action_down',
arguments=[],
variables=['N', 'Y', 'Y'],
pFunc=DNF('action_down',
terms=8,
init=w,
sig=2,
predColl=self.predColl,
fast=False,
post_terms=pt),
use_neg=True,
Fam='eq',
exc_preds=excs,
count_type=count_type,
arg_funcs=['M'])
self.predColl.initialize_predicates()
self.bg = Background(self.predColl)
self.bg.add_backgroud('X_U', ('%d' % 0, ))
self.bg.add_backgroud('X_D', ('%d' % (self.DIM1 - 1), ))
self.bg.add_backgroud('Y_L', ('%d' % 0, ))
self.bg.add_backgroud('Y_R', ('%d' % (self.DIM2 - 1), ))
for i in range(self.DIM1):
self.bg.add_backgroud('sameX', (
'%d' % i,
'%d' % i,
))
for i in range(self.DIM2):
self.bg.add_backgroud('sameY', (
'%d' % i,
'%d' % i,
))
for i in range(self.DIM2):
for j in range(self.DIM2):
if i < j:
self.bg.add_backgroud('ltY', (
'%d' % i,
'%d' % j,
))
if abs(i - j) < 2:
self.bg.add_backgroud('close', (
'%d' % i,
'%d' % j,
))
print('displaying config setting...')
self.mdl = ILPRLEngine(args=self.args,
predColl=self.predColl,
bgs=None)
def run(self, states):
bs = tf.shape(states[0])[0]
self.X0 = OrderedDict()
for p in self.predColl.outpreds:
tmp = tf.expand_dims(
tf.constant(self.bg.get_X0(p.oname), tf.float32), 0)
self.X0[p.oname] = tf.tile(tmp, [bs, 1])
self.X0['agent'] = states[0]
self.X0['predLR'] = states[1]
self.X0['predRL'] = states[2]
self.X0['food'] = states[3]
self.Xo, L3 = self.mdl.getTSteps(self.X0)
xo_r = {}
for p in self.Xo:
xo_r[p] = tf.reshape(self.Xo[p], [-1] + [
len(self.predColl.constants[i])
for i in self.predColl[p].arguments
])
move = [
self.Xo[i] for i in [
'action_noop', 'action_up', 'action_right', 'action_left',
'action_down'
]
]
# move = [ 1.0-self.Xo[i] for i in ['en_noop','en_up','en_right','en_left', 'en_down']]
# move[0] = tf.zeros_like( move[0] )-5
return tf.concat(move, -1), self.X0, self.Xo, xo_r
class ILP_MODEL(object):
def __init__(self, num_box, is_train=True):
self.num_box = num_box
self.args = self.load_ilp_config()
self.define_preds()
self.Xo = None
self.X0 = None
def load_ilp_config(self):
param = dotdict({})
param.BS = 1
param.T = 1
param.W_DISP_TH = .1
param.GPU = 1
return param
def define_preds(self):
Box = ['%d' % i for i in range(self.num_box + 1)]
Ds = ['%d' % i for i in range(self.num_box + 1)]
self.Constants = dict({'C': Box, 'D': Ds})
self.predColl = PredCollection(self.Constants)
self.predColl.add_pred(dname='posH', arguments=['C', 'D'])
self.predColl.add_pred(dname='posV', arguments=['C', 'D'])
self.predColl.add_pred(dname='is_one', arguments=['D'])
self.predColl.add_pred(dname='lt', arguments=['D', 'D'])
self.predColl.add_pred(dname='inc', arguments=['D', 'D'])
self.predColl.add_pred(dname='same', arguments=['C', 'C'])
self.predColl.add_pred(dname='is_floor', arguments=['C'])
self.predColl.add_pred(dname='is_blue', arguments=['C'])
self.predColl.add_pred(dname='same_col',
arguments=['C', 'C'],
variables=['D'],
pFunc=DNF('same_col',
terms=1,
init=[1, .1, -1, .1],
sig=2,
init_terms=['posH(A,C), posH(B,C)'],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq')
self.predColl.add_pred(
dname='above',
arguments=['C', 'C'],
variables=['D', 'D'],
pFunc=DNF(
'above',
terms=1,
init=[1, .1, -1, .1],
sig=2,
init_terms=['same_col(A,B), posV(A,C), posV(B,D), lt(D,C)'],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq')
self.predColl.add_pred(
dname='below',
arguments=['C', 'C'],
variables=['D', 'D'],
pFunc=DNF(
'below',
terms=1,
init=[1, .1, -1, .1],
sig=2,
init_terms=['same_col(A,B), posV(A,C), posV(B,D), lt(C,D)'],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq')
self.predColl.add_pred(
dname='on',
arguments=['C', 'C'],
variables=['D', 'D'],
pFunc=DNF('on',
terms=2,
init=[1, .1, -1, .1],
sig=2,
init_terms=[
'is_floor(B), posV(A,C), is_one(C)',
'same_col(A,B), posV(A,C), posV(B,D), inc(D,C)'
],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq')
self.predColl.add_pred(dname='isCovered',
arguments=['C'],
variables=['C'],
pFunc=DNF(
'isCovered',
terms=1,
init=[1, .1, -1, .1],
sig=2,
init_terms=['on(B,A), not is_floor(A)'],
predColl=self.predColl,
fast=True),
use_neg=True,
Fam='eq')
self.predColl.add_pred(
dname='lower',
oname='lower',
arguments=['C', 'C'],
variables=['D', 'D'],
pFunc=DNF('lower',
terms=1,
init=[1, .1, -1, .1],
sig=2,
init_terms=['posV(A,C), posV(B,D), lt(C,D)'],
predColl=self.predColl,
fast=True),
use_neg=True,
Fam='eq')
self.predColl.add_pred(
dname='moveable',
oname='moveable',
arguments=['C', 'C'],
variables=[],
pFunc=DNF(
'moveable',
terms=1,
init=[1, .1, -1, .1],
sig=2,
init_terms=[
'not isCovered(A), not isCovered(B), not same(A,B), not is_floor(A), not on(A,B), not is_blue(A), not is_floor(B), not lower(B,A)'
],
predColl=self.predColl,
fast=True),
use_neg=True,
Fam='eq')
pt = [
('and', 'moveable(A,B)'),
]
self.predColl.add_pred(dname='move',
oname='move',
arguments=['C', 'C'],
variables=[],
pFunc=DNF('move',
terms=4,
init=[1, -1, -1, .1],
sig=2,
predColl=self.predColl,
fast=False,
post_terms=pt),
use_neg=True,
Fam='eq',
exc_preds=[],
exc_conds=[('*', 'rep1')])
self.predColl.initialize_predicates()
self.bg = Background(self.predColl)
#define backgrounds
self.bg.add_backgroud('is_floor', ('0', ))
self.bg.add_backgroud('is_one', ('1', ))
self.bg.add_backgroud('is_blue', ('1', ))
for i in range(self.num_box + 1):
self.bg.add_backgroud('same', ('%d' % i, '%d' % i))
if '%d' % (i + 1) in Ds:
self.bg.add_backgroud('inc', ('%d' % i, '%d' % (i + 1)))
for i in range(self.num_box + 1):
for j in range(self.num_box + 1):
if i < j:
self.bg.add_backgroud('lt', ('%d' % i, '%d' % (j)))
print('displaying config setting...')
self.mdl = ILPRLEngine(args=self.args,
predColl=self.predColl,
bgs=None)
def run(self, state_in_x, state_in_y):
bs = tf.shape(state_in_x)[0]
self.X0 = OrderedDict()
for p in self.predColl.outpreds:
tmp = tf.expand_dims(
tf.constant(self.bg.get_X0(p.oname), tf.float32), 0)
self.X0[p.oname] = tf.tile(tmp, [bs, 1])
flx = np.zeros((self.num_box + 1, self.num_box + 1), dtype=np.float32)
flx[0, :] = 1
xx = tf.pad(state_in_x,
[[0, 0], [1, 0], [1, 0]]) + flx[np.newaxis, :, :]
self.X0['posH'] = tf.reshape(xx, [-1, (self.num_box + 1)**2])
fly = np.zeros((self.num_box + 1, self.num_box + 1), dtype=np.float32)
fly[0, 0] = 1
yy = tf.pad(state_in_y,
[[0, 0], [1, 0], [1, 0]]) + fly[np.newaxis, :, :]
self.X0['posV'] = tf.reshape(yy, [-1, (self.num_box + 1)**2])
self.Xo, L3 = self.mdl.getTSteps(self.X0)
return self.Xo
class ILP_MODEL(object):
def __init__(self, num_box, is_train=True):
self.num_box = num_box
self.args = self.load_ilp_config()
self.define_preds()
self.Xo = None
self.X0 = None
self.has_key = None
def reset(self):
self.has_key = None
def load_ilp_config(self):
param = dotdict({})
param.BS = 1
param.T = 1
param.W_DISP_TH = .1
param.GPU = 1
return param
def define_preds(self):
nCOLOR = 10
Colors = [i for i in range(nCOLOR)]
Pos = [i for i in range(12)]
# Ds = ['%d'%i for i in range(self.num_box)]
self.Constants = dict({
'C': Colors,
'P': Pos,
'Q': Pos
}) #, 'N':['%d'%i for i in range(6)] })
self.predColl = PredCollection(self.Constants)
self.predColl.add_pred(dname='color', arguments=['P', 'Q', 'C'])
for i in range(nCOLOR):
self.predColl.add_pred(dname='is%d' % i, arguments=['C'])
# self.predColl.add_pred(dname='Pos0' ,arguments=['P'])
# self.predColl.add_pred(dname='Pos11' ,arguments=['P'])
self.predColl.add_pred(dname='has_key', arguments=['C'])
self.predColl.add_pred(dname='neq', arguments=['C', 'C'])
# self.predColl.add_pred(dname='lt' ,arguments=['P','P'])
# self.predColl.add_pred(dname='incp' ,arguments=['P','P'])
self.predColl.add_pred(dname='incq', arguments=['Q', 'Q'])
self.predColl.add_pred(dname='isBK',
arguments=['P', 'Q'],
variables=['C'],
pFunc=DNF('isBK',
terms=1,
init=[1, .1, -1, .1],
sig=2,
init_terms=['color(A,B,C), is0(C)'],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq')
self.predColl.add_pred(dname='isAgent',
arguments=['P', 'Q'],
variables=['C'],
pFunc=DNF('isAgent',
terms=1,
init=[1, .1, -1, .1],
sig=2,
init_terms=['color(A,B,C), is1(C)'],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq')
self.predColl.add_pred(dname='isGem',
arguments=['P', 'Q'],
variables=['C'],
pFunc=DNF('isGem',
terms=1,
init=[1, .1, -1, .1],
sig=2,
init_terms=['color(A,B,C), is2(C)'],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq')
self.predColl.add_pred(
dname='isItem',
arguments=['P', 'Q'],
variables=[],
pFunc=DNF('isItem',
terms=1,
init=[1, .1, -1, .1],
sig=2,
init_terms=['not isBK(A,B), not isAgent(A,B)'],
predColl=self.predColl,
fast=True),
use_neg=True,
Fam='eq')
# self.predColl.add_pred(dname='sameColor' ,arguments=['P','P','P','P'], variables=['C'] ,pFunc =
# DNF('sameColor',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['color(A,B,E), color(C,D,E)'],predColl=self.predColl,fast=True) , use_neg=False, Fam='or')
self.predColl.add_pred(
dname='locked',
arguments=['P', 'Q'],
variables=['Q'],
pFunc=DNF('locked',
terms=1,
init=[1, .1, -1, .1],
sig=2,
init_terms=['isItem(A,B), isItem(A,C), incq(B,C)'],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq')
self.predColl.add_pred(
dname='isLock',
arguments=['P', 'Q'],
variables=['Q'],
pFunc=DNF('isLock',
terms=1,
init=[1, .1, -1, .1],
sig=2,
init_terms=['isItem(A,B), isItem(A,C), incq(C,B)'],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='eq')
# self.predColl.add_pred(dname='locked1' ,arguments=['P','P'], variables=['P'] ,pFunc =
# DNF('locked1',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['isItem(A,B), isItem(C,B), inc(C,B)'],predColl=self.predColl,fast=True) , use_neg=False, Fam='or')
#self.predColl.add_pred(dname='LockColor' ,arguments=['C','C'], variables=['P','Q','Q'] ,pFunc =
# DNF('LockColor',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['isItem(C,D), isItem(C,E), color(C,D,A), color(C,E,B), incq(D,E)'],predColl=self.predColl,fast=True) , use_neg=False, Fam='eq')
#self.predColl.add_pred(dname='loosekey' ,arguments=['P','Q'], variables=[] ,pFunc =
# DNF('loosekey',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['isItem(A,B), not isLock(A,B), not locked(A,B)'],predColl=self.predColl,fast=True) , use_neg=True, Fam='eq')
#self.predColl.add_pred(dname='key_color' ,arguments=['P','Q','C'], variables=['Q'] ,pFunc =
## DNF('key_color',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['isItem(A,B), isLock(A,B), color(A,B,C)',
# 'isItem(A,B), locked(A,D), incq(D,B), color(A,D,C)' ],predColl=self.predColl,fast=True) , use_neg=True, Fam='eq')
#self.predColl.add_pred(dname='inGoal' ,arguments=['C'], variables=['C'] ,pFunc =
# DNF('inGoal',terms=2,init=[1,.1,-1,.1],sig=2,init_terms=['is2(A)' , 'inGoal(B), LockColor(B,A)'],predColl=self.predColl,fast=True) , use_neg=False, Fam='or')
# self.predColl.add_pred(dname='valid',oname='valid',arguments=['P','P'] , variables=['C'] ,pFunc =
# DNF('valid',terms=2,init_terms=['loosekey(A,B)', 'isItem(A,B), isLock(A,B), color(A,B,C), has_key(C)'],sig=2,predColl=self.predColl,fast=True ) , use_neg=True, Fam='eq')
#,post_terms=[('and','isItem(A,B)')] )
# self.predColl.add_pred(dname='move',oname='move',arguments=['P','P'] , variables=['C'] ,pFunc =
# DNF('move',terms=4,init=[-1,.1,-1,.1],sig=2,predColl=self.predColl,fast=False,post_terms=[('and','not isBK(A,B)')]) , use_neg=True, Fam='eq')
# self.predColl.add_pred(dname='move',oname='move',arguments=['P','P'] , variables=['C','C'] ,pFunc =
# DNF('move',terms=8,init=[1,-1,-2,.3],sig=1,predColl=self.predColl,fast=False,post_terms=[('and','isItem(A,B)')]) , use_neg=True, Fam='eq', exc_preds=['move']+[ 'is%d'%i for i in range(nCOLOR)],exc_conds=[('*','rep1') ])
# DNF('move',terms=5,init=[1,-3,-1,.1],sig=2,predColl=self.predColl,fast=False,post_terms=[('and','isItem(A,B)')]) , use_neg=False, Fam='eq', exc_preds=['move']+[ 'is%d'%i for i in range(nCOLOR)],exc_conds=[('*','rep1') ])
# DNF('move',terms=10,init=[1,-1,-1,.1],sig=2,predColl=self.predColl,fast=False,post_terms=[('and','isItem(A,B)'),('or','loosekey(A,B)')]) , use_neg=True, Fam='eq', exc_preds=['move', 'color'])
# DNF('move',terms=5,init_terms=['color(A,B,C), inGoal(C), loosekey(A,B)','key_color(A,B,D), color(A,B,C), inGoal(D), isLock(A,B), inGoal(D), LockColor(D,C), has_key(C)'],sig=2,predColl=self.predColl,fast=True ) , use_neg=True, Fam='eq')
# DNF('move',terms=2,init_terms=['loosekey(A,B)', 'isLock(A,B), color(A,B,C), has_key(C)'],sig=2,predColl=self.predColl,fast=True ) , use_neg=True, Fam='eq')
# for i in range(10):
# self.predColl.add_pred(dname='aux%d'%i,oname='aux%d'%i,arguments=['P','Q'] , variables=['C','C'] ,pFunc =
# CONJ('aux%d'%i,init=[-1,1],init_terms=[],sig=1,predColl=self.predColl,post_terms=[]) , use_neg=False, Fam='eq', exc_preds=['move']+[ '%d'%i for i in range(nCOLOR)]+['aux%d'%i for i in range(10)],exc_conds=[('*','rep1') ])
# it=[]
#self.predColl.add_pred(dname='Q',oname='Q',arguments=['P','Q'] , variables=['P', 'C' ] ,pFunc =
# DNF('Q',terms=6,init=[-1,.1,-1,1],init_terms=[],sig=1,predColl=self.predColl ) , use_neg=True, Fam='eq', exc_preds=['move','Q']+[ 'is%d'%i for i in range(nCOLOR)],exc_conds=[('*','rep1') ])
#self.predColl.add_pred(dname='move',oname='move',arguments=['P','Q'] , variables=['C' ] ,pFunc =
# DNF('move',terms=6,init=[1,-1,-2,1],init_terms=[],sig=2,predColl=self.predColl,post_terms=[('and','isItem(A,B)'),('and','inGoal(C)') ]) , use_neg=True, Fam='eq', exc_preds=['move','Q']+[ 'is%d'%i for i in range(nCOLOR)],exc_conds=[('*','rep1') ])
pt = [('and', 'isItem(A,B)')]
pt = []
self.predColl.add_pred(dname='move',
oname='move',
arguments=['P', 'Q'],
variables=['C'],
pFunc=DNF('move',
terms=6,
init=[-1, .1, -1, .1],
init_terms=[],
sig=1,
predColl=self.predColl,
post_terms=pt),
use_neg=True,
Fam='eq',
exc_preds=['move', 'Q'] +
['is%d' % i for i in range(nCOLOR)],
exc_conds=[('*', 'rep1')])
# for i in range(5):
# if i==0:
# tp="eq"
# # it=['color(A,B,C), inGoal(C), loosekey(A,B)']
# else:
# tp="or"
# it=[]
# vvv=1
# vv=-1
# Vars = ['C','C' ]
# if i==5:
# Vars=['C','C']
# vvv=2
# vv=-1
# # if i==1:
# # it=['key_color(A,B,D), color(A,B,C), isLock(A,B), inGoal(D), LockColor(D,C), has_key(C)']
# self.predColl.add_pred(dname='move%d'%i,oname='move',arguments=['P','Q'] , variables=Vars ,pFunc =
# CONJ('move%d'%i,init=[vv,vvv],init_terms=it,sig=1,predColl=self.predColl,post_terms=[('and','isItem(A,B)')]) , use_neg=False, Fam=tp, exc_preds=['move']+[ '%d'%i for i in range(nCOLOR)],exc_conds=[('*','rep1') ])
# self.predColl.add_pred(dname='move2',oname='move',arguments=['P','P'] , variables=['C'] ,pFunc =
# CONJ('move2',init=[-1,.1],sig=2,predColl=self.predColl,fast=False) , use_neg=True, Fam='eq', exc_preds=['move'],exc_conds=[('*','rep1') ])
# MLP('move',dims=[64,64,1], acts=[relu1,relu1,tf.nn.sigmoid] ) , use_neg=False, Fam='eq')
# DNF('move',terms=10,init=[1,-1,-1,.1],sig=2,predColl=self.predColl,fast=False,post_terms=[('and','isItem(A,B)'),('or','loosekey(A,B)')]) , use_neg=True, Fam='eq', exc_preds=['move', 'color'])
# self.predColl.add_pred(dname='move',oname='move',arguments=['P','P'] , variables=['C'] ,pFunc =
# DNF('move',terms=2,init_terms=['loosekey(A,B)', 'isLock(A,B), color(A,B,C), has_key(C)'],sig=2,predColl=self.predColl,fast=True ) , use_neg=True, Fam='eq')
# self.predColl.add_pred(dname='has_key' ,arguments=['C'], variables=['P','P'] ,pFunc =
# DNF('has_key',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['loosekey(B,C), move(B,C), color(B,C,A)'],predColl=self.predColl,fast=True) , use_neg=True, Fam='or')
# self.predColl.add_pred(dname='enableLeft' ,arguments=['P','P'], variables=['P'] ,pFunc =
# DNF('enableLeft',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['agent(A,B), inc(C,A), not pos0(C)'],predColl=self.predColl,fast=True) , use_neg=True, Fam='or')
# self.predColl.add_pred(dname='enableRight' ,arguments=['P','P'], variables=['P'] ,pFunc =
# DNF('enableRight',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['agent(A,B), not locked(A,B), not Pos0(A)'],predColl=self.predColl,fast=True) , use_neg=True, Fam='or')
# self.predColl.add_pred(dname='enableLeft' ,arguments=['P','P'], variables=['P'] ,pFunc =
# DNF('enableLeft',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['agent(A,B), not locked(A,B), not Pos0(A)'],predColl=self.predColl,fast=True) , use_neg=True, Fam='or')
# self.predColl.add_pred(dname='enableLeft' ,arguments=['P','P'], variables=['P'] ,pFunc =
# DNF('enableLeft',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['agent(A,B), not locked(A,B), not Pos0(A)'],predColl=self.predColl,fast=True) , use_neg=True, Fam='or')
# self.predColl.add_pred(dname='moveLeft',oname='moveLeft',arguments=[] , variables=['P','P','C'] ,pFunc =
# DNF('moveLeft',terms=6,init=[1,.1,-1,.1],sig=2,predColl=self.predColl,fast=False) , use_neg=True, Fam='eq')
# self.predColl.add_pred(dname='moveRight',oname='moveRight',arguments=[] , variables=['P','P','C'] ,pFunc =
# DNF('moveRight',terms=6,init=[1,.1,-1,.1],sig=2,predColl=self.predColl,fast=False) , use_neg=True, Fam='eq')
# self.predColl.add_pred(dname='moveLeft',oname='moveLeft',arguments=[] , variables=['P','P','C'] ,pFunc =
# DNF('moveLeft',terms=6,init=[1,.1,-1,.1],sig=2,predColl=self.predColl,fast=False) , use_neg=True, Fam='eq')
# self.predColl.add_pred(dname='moveLeft',oname='moveLeft',arguments=[] , variables=['P','P','C'] ,pFunc =
# DNF('moveLeft',terms=6,init=[1,.1,-1,.1],sig=2,predColl=self.predColl,fast=False) , use_neg=True, Fam='eq')
self.predColl.initialize_predicates()
self.bg = Background(self.predColl)
#define backgrounds
self.bg.add_backgroud('Pos0', (0, ))
self.bg.add_backgroud('Pos11', (11, ))
for i in range(nCOLOR):
self.bg.add_backgroud('is%d' % i, (i, ))
for j in range(nCOLOR):
if i != j:
self.bg.add_backgroud('neq', (i, i))
for i in range(12):
self.bg.add_backgroud('eq', (i, i))
for j in range(12):
if i < j:
self.bg.add_backgroud('lt', (i, j))
if j == i + 1:
# self.bg.add_backgroud('incp',(i,j) )
self.bg.add_backgroud('incq', (i, j))
print('displaying config setting...')
self.mdl = ILPRLEngine(args=self.args,
predColl=self.predColl,
bgs=None)
def run(self, state):
has_key, color = state
bs = tf.shape(has_key)[0]
self.X0 = OrderedDict()
for p in self.predColl.outpreds:
tmp = tf.expand_dims(
tf.constant(self.bg.get_X0(p.oname), tf.float32), 0)
self.X0[p.oname] = tf.tile(tmp, [bs, 1])
self.X0['color'] = color
self.X0['has_key'] = has_key
# if self.has_key is not None:
# self.X0['has_key'] = has_key
self.Xo, L3 = self.mdl.getTSteps(self.X0)
# self.has_key = self.X0['has_key']
return self.Xo["move"]
def runtest(self, state):
has_key, color = state
bs = tf.shape(has_key)[0]
self.X0 = OrderedDict()
for p in self.predColl.outpreds:
tmp = tf.expand_dims(
tf.constant(self.bg.get_X0(p.oname), tf.float32), 0)
self.X0[p.oname] = tf.tile(tmp, [bs, 1])
# if self.has_key is not None:
# self.X0['has_key'] = has_key
self.X0['color'] = color
self.X0['has_key'] = has_key
self.Xo, L3 = self.mdl.getTSteps(self.X0)
# self.has_key = self.X0['has_key']
L1 = 0
L2 = 0
def get_pen(x):
return tf.nn.relu(2 * x -
2) - 2 * tf.nn.relu(2 * x - 1) + tf.nn.relu(
2 * x)
# for p in self.predColl.preds:
# vs = tf.get_collection( p.dname)
# # for wi in vs:
# if '_AND' in wi.name:
# wi = p.pFunc.conv_weight(wi)
# # L2 += tf.reduce_mean( wi*(1.0-wi))
# L2 += tf.reduce_mean( get_pen(wi))
# s = tf.reduce_sum( wi,-1)
# # L1 += tf.reduce_mean( tf.nn.relu( s-7) )
# s = tf.reduce_max( wi,-1)
# L1 += tf.reduce_mean( tf.nn.relu( 1-s) )
return self.Xo, L2 * 0
class ILP_MODEL(object):
def __init__(self, num_box,is_train=True):
self.num_box = num_box
self.args = self.load_ilp_config()
self.define_preds()
self.Xo=None
self.X0=None
self.has_key=None
def reset(self):
self.has_key=None
def load_ilp_config(self):
param = dotdict({})
param.BS = 1
param.T = 1
param.W_DISP_TH = .1
param.GPU = 1
return param
def define_preds(self):
nCOLOR = 10
Colors=[i for i in range(nCOLOR)]
Pos = [i for i in range(12)]
# Ds = ['%d'%i for i in range(self.num_box)]
self.Constants = dict( {'C':Colors,'P':Pos,'Q':Pos}) #, 'N':['%d'%i for i in range(6)] })
self.predColl = PredCollection (self.Constants)
self.predColl.add_pred(dname='color' ,arguments=['P','Q','C'])
for i in range(nCOLOR):
self.predColl.add_pred(dname='is%d'%i ,arguments=['C'])
self.predColl.add_pred(dname='has_key' ,arguments=['C'] )
self.predColl.add_pred(dname='neq' ,arguments=['C','C'])
self.predColl.add_pred(dname='incq' ,arguments=['Q','Q'])
self.predColl.add_pred(dname='isBK' ,arguments=['P','Q'], variables=['C'] ,pFunc =
DNF('isBK',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['color(A,B,C), is0(C)'],predColl=self.predColl,fast=True) , use_neg=False, Fam='eq')
self.predColl.add_pred(dname='isAgent' ,arguments=['P','Q'], variables=['C'] ,pFunc =
DNF('isAgent',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['color(A,B,C), is1(C)'],predColl=self.predColl,fast=True) , use_neg=False, Fam='eq')
self.predColl.add_pred(dname='isGem' ,arguments=['P','Q'], variables=['C'] ,pFunc =
DNF('isGem',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['color(A,B,C), is2(C)'],predColl=self.predColl,fast=True) , use_neg=False, Fam='eq')
self.predColl.add_pred(dname='isItem' ,arguments=['P','Q'], variables=[] ,pFunc =
DNF('isItem',terms=1,init=[1,.1,-1,.1],sig=2,init_terms=['not isBK(A,B), not isAgent(A,B)'],predColl=self.predColl,fast=True) , use_neg=True, Fam='eq')
pt=[]
self.predColl.add_pred(dname='move',oname='move',arguments=['P','Q'] , variables=['C' ] ,pFunc =
DNF('move',terms=6,init=[-1,.1,-1,.1],init_terms=[],sig=1,predColl=self.predColl,post_terms=pt) , use_neg=True, Fam='eq', exc_preds=['move','Q']+[ 'is%d'%i for i in range(nCOLOR)],exc_conds=[('*','rep1') ])
self.predColl.initialize_predicates()
self.bg = Background( self.predColl )
#define backgrounds
self.bg.add_backgroud('Pos0',(0,) )
self.bg.add_backgroud('Pos11',(11,) )
for i in range(nCOLOR):
self.bg.add_backgroud('is%d'%i , (i,))
for j in range(nCOLOR):
if i!=j:
self.bg.add_backgroud('neq',(i,i) )
for i in range(12):
self.bg.add_backgroud('eq',(i,i) )
for j in range(12):
if i<j:
self.bg.add_backgroud('lt',(i,j) )
if j==i+1:
# self.bg.add_backgroud('incp',(i,j) )
self.bg.add_backgroud('incq',(i,j) )
print('displaying config setting...')
self.mdl = ILPRLEngine( args=self.args ,predColl=self.predColl ,bgs=None )
def run( self, state):
has_key,color = state
bs= tf.shape(has_key)[0]
self.X0=OrderedDict()
for p in self.predColl.outpreds:
tmp = tf.expand_dims( tf.constant( self.bg.get_X0(p.oname) ,tf.float32) , 0)
self.X0[p.oname] = tf.tile( tmp , [bs,1] )
self.X0['color'] = color
self.X0['has_key'] = has_key
# if self.has_key is not None:
# self.X0['has_key'] = has_key
self.Xo,L3 = self.mdl.getTSteps(self.X0)
# self.has_key = self.X0['has_key']
return self.Xo["move"]
def runtest( self, state):
has_key,color = state
bs= tf.shape(has_key)[0]
self.X0=OrderedDict()
for p in self.predColl.outpreds:
tmp = tf.expand_dims( tf.constant( self.bg.get_X0(p.oname) ,tf.float32) , 0)
self.X0[p.oname] = tf.tile( tmp , [bs,1] )
# if self.has_key is not None:
# self.X0['has_key'] = has_key
self.X0['color'] = color
self.X0['has_key'] = has_key
self.Xo,L3 = self.mdl.getTSteps(self.X0)
# self.has_key = self.X0['has_key']
L1=0
L2=0
def get_pen(x):
return tf.nn.relu(2*x-2)-2*tf.nn.relu(2*x-1)+tf.nn.relu(2*x)
# for p in self.predColl.preds:
# vs = tf.get_collection( p.dname)
# # for wi in vs:
# if '_AND' in wi.name:
# wi = p.pFunc.conv_weight(wi)
# # L2 += tf.reduce_mean( wi*(1.0-wi))
# L2 += tf.reduce_mean( get_pen(wi))
# s = tf.reduce_sum( wi,-1)
# # L1 += tf.reduce_mean( tf.nn.relu( s-7) )
# s = tf.reduce_max( wi,-1)
# L1 += tf.reduce_mean( tf.nn.relu( 1-s) )
return self.Xo,L2*0
class ILP_MODEL(object):
def __init__(self, DIM1,DIM2,F_COUNT=3):
self.DIM1=DIM1
self.DIM2=DIM2
self.F_COUNT=F_COUNT
self.args = self.load_ilp_config()
self.define_preds()
self.Xo=None
self.X0=None
def load_ilp_config(self):
param = dotdict({})
param.BS = 1
param.T = 1
param.W_DISP_TH = .2
param.GPU = 1
return param
def define_preds(self):
X = ['%d'%i for i in range(self.DIM1)]
Y = ['%d'%i for i in range(self.DIM2)]
self.Constants = dict( {'N':X,'Y':Y})
self.predColl = PredCollection (self.Constants)
# self.predColl.add_pred(dname='agentX' ,arguments=['N'])
# self.predColl.add_pred(dname='agentY' ,arguments=['Y'])
for i in range(self.DIM1):
if i==0 or i==self.DIM1-1:
self.predColl.add_pred(dname='X_%d'%i,arguments=['N'])
for i in range(self.DIM2):
if i==0 or i==self.DIM2-1:
self.predColl.add_pred(dname='Y_%d'%i,arguments=['Y'])
# self.predColl.add_pred(dname='f_1',arguments=['N','Y'], variables=[ ],pFunc =
# DNF('f_1',init_terms=['agentX(A), agentY(B)' ],predColl=self.predColl,fast=True) , use_neg=False, Fam='eq')
self.predColl.add_pred(dname='f_1',arguments=['N','Y'])
self.predColl.add_pred(dname='f_2',arguments=['N','Y'])
self.predColl.add_pred(dname='f_3',arguments=['N','Y'])
self.predColl.add_pred(dname='f_4',arguments=['N','Y'])
self.predColl.add_pred(dname='ltY',arguments=['Y','Y'])
self.predColl.add_pred(dname='close',arguments=['Y','Y'])
count_type = 'max'
excs = ['action_noop', 'action_up', 'action_down','action_left','action_right', 'Q']
w = [-1,.1,-2,.1]
Alicount_type = None
self.predColl.add_pred(dname='en_up' ,arguments=[], variables=[ 'N','Y', 'Y'],pFunc =
DNF('en_up',terms=1,init=[-1,.1,-1,.1],sig=2,
init_terms=['f_1(A,B), X_0(A)','f_1(A,B), f_3(M_A,C), close(B,C), ltY(B,C)', 'f_1(A,B), f_2(M_A,C), close(B,C), ltY(C,B)' ],
predColl=self.predColl,fast=True) , use_neg=False, Fam='eq',count_type=count_type, count_th=100,arg_funcs=['M'])
self.predColl.add_pred(dname='en_down' ,arguments=[ ], variables=['N','Y', 'Y'],pFunc =
DNF('en_down',terms=1,init=[-1,.1,-1,.1],sig=2,
init_terms=['f_1(A,B), X_%d(A)'%(self.DIM1-1) , 'f_1(A,B), f_3(P_A,C), close(B,C), ltY(B,C)', 'f_1(A,B), f_2(P_A,C), close(B,C), ltY(C,B)' ]
,predColl=self.predColl,fast=True) , use_neg=False, Fam='eq',count_type=count_type, count_th=100,arg_funcs=['P'])
self.predColl.add_pred(dname='en_right' ,arguments=[ ], variables=['N','Y','Y'],pFunc =
DNF('en_right',terms=1,init=[-1,.1,-1,.1],sig=2,
init_terms=['f_1(A,B), Y_%d(B)'%(self.DIM2-1) ,'f_1(A,B), f_3(A,C), close(B,C), ltY(B,C)' ,'f_1(A,B), f_2(A,C), close(B,C), ltY(B,C)' ],
predColl=self.predColl,fast=True) , use_neg=False, Fam='eq',count_type=count_type, count_th=100,arg_funcs=[ ])
self.predColl.add_pred(dname='en_left' ,arguments=[ ], variables=['N','Y','Y'],pFunc =
DNF('en_left',terms=1,init=[-1,.1,-1,.1],sig=2,
init_terms=['f_1(A,B), Y_%d(B)'%0, 'f_1(A,B), f_2(A,C), close(B,C), ltY(C,B)','f_1(A,B), f_3(A,C), close(B,C), ltY(C,B)' ],
predColl=self.predColl,fast=True) , use_neg=False, Fam='eq',count_type=count_type, count_th=100,arg_funcs=[ ])
self.predColl.add_pred(dname='en_noop' ,arguments=[ ], variables=[],pFunc =
DNF('en_noop',terms=1,init=[-1,.1,-1,.1],sig=2,
#init_terms=['f_1(A,B), f_3(A,C), ltY(C,B)','f_1(A,B), f_2(A,C), ltY(B,C)' ],
init_terms=['en_right()','en_left()' ],
predColl=self.predColl,fast=True) , use_neg=False, Fam='eq',count_type=count_type, count_th=100)
pt = [('and', 'not en_noop()') ]
self.predColl.add_pred(dname='action_noop',arguments=[] , variables=['N','Y','Y' ] ,pFunc =
DNF('action_noop',terms=8,init=w,sig=2,predColl=self.predColl,fast=False,post_terms=pt) , use_neg=True, Fam='eq' ,
exc_preds=excs,count_type=count_type,arg_funcs=['M'])
pt = [('and', 'not en_up()') ]
self.predColl.add_pred(dname='action_up',arguments=[] , variables=['N','Y','Y' ] ,pFunc =
DNF('action_up',terms=8,init=w,sig=2,predColl=self.predColl,fast=False,post_terms=pt) , use_neg=True, Fam='eq' ,
exc_preds=excs,count_type=count_type, count_th=100,arg_funcs=[ 'M'])
pt = [('and', 'not en_right()') ]
self.predColl.add_pred(dname='action_right',arguments=[] , variables=['N','Y','Y' ] ,pFunc =
DNF('action_right',terms=8,init=w,sig=2,predColl=self.predColl,fast=False,post_terms=pt) , use_neg=True, Fam='eq' ,
exc_preds=excs,count_type=count_type,arg_funcs=['M'])
pt = [('and', 'not en_left()') ]
self.predColl.add_pred(dname='action_left',arguments=[] , variables=['N','Y', 'Y' ] ,pFunc =
DNF('action_left',terms=8,init=w,sig=2,predColl=self.predColl,fast=False,post_terms=pt) , use_neg=True, Fam='eq' ,
exc_preds=excs,count_type=count_type, arg_funcs=['M'])
pt = [('and', 'not en_down()') ]
self.predColl.add_pred(dname='action_down',arguments=[] , variables=['N','Y','Y' ] ,pFunc =
DNF('action_down',terms=8,init=w,sig=2,predColl=self.predColl,fast=False,post_terms=pt) , use_neg=True, Fam='eq' ,
exc_preds=excs,count_type=count_type, arg_funcs=['M'])
self.predColl.initialize_predicates()
self.bg = Background( self.predColl )
#define backgrounds
self.bg.add_backgroud('X_%d'%0 ,('%d'%0,) )
self.bg.add_backgroud('X_%d'%(self.DIM1-1) ,('%d'%(self.DIM1-1),) )
self.bg.add_backgroud('Y_%d'%0 ,('%d'%0,) )
self.bg.add_backgroud('Y_%d'%(self.DIM2-1) ,('%d'%(self.DIM2-1),) )
self.bg.add_backgroud('Y_%d'%0 ,('%d'%1,) )
self.bg.add_backgroud('Y_%d'%(self.DIM2-1) ,('%d'%(self.DIM2-2),) )
for i in range(self.DIM2):
for j in range(self.DIM2):
if i<=j+1:
self.bg.add_backgroud('ltY' ,('%d'%i,'%d'%j,) )
if abs(i-j)<3:
self.bg.add_backgroud('close' ,('%d'%i,'%d'%j,) )
print('displaying config setting...')
self.mdl = ILPRLEngine( args=self.args ,predColl=self.predColl ,bgs=None )
def run( self, states):
bs= tf.shape(states[0])[0]
self.X0=OrderedDict()
for p in self.predColl.outpreds:
tmp = tf.expand_dims( tf.constant( self.bg.get_X0(p.oname) ,tf.float32) , 0)
self.X0[p.oname] = tf.tile( tmp , [bs,1] )
self.X0['f_1'] = states[0]
self.X0['f_2'] = states[1]
self.X0['f_3'] = states[2]
self.X0['f_4'] = states[3]
self.Xo,L3 = self.mdl.getTSteps(self.X0)
move = [ self.Xo[i] for i in ['action_noop','action_up','action_right','action_left', 'action_down']]
# move[0] = tf.zeros_like( move[0] )-5
return tf.concat(move,-1),self.X0,self.Xo
class Model(object):
def __init__(self, config, debug_information=False, is_train=True):
self.debug = debug_information
self.XO = {}
self.pen = None
self.config = config
self.batch_size = self.config.batch_size
self.img_size = self.config.data_info[0]
self.c_dim = self.config.data_info[2]
self.q_dim = self.config.data_info[3]
self.a_dim = self.config.data_info[4]
self.conv_info = self.config.conv_info
self.acc = 0
self.feat_count = 64
self.ilp_params = None
# create placeholders for the input
self.img = tf.placeholder(
name='img',
dtype=tf.float32,
shape=[self.batch_size, self.img_size, self.img_size, self.c_dim],
)
self.q = tf.placeholder(
name='q',
dtype=tf.float32,
shape=[self.batch_size, self.q_dim],
)
self.a = tf.placeholder(
name='a',
dtype=tf.float32,
shape=[self.batch_size, self.a_dim],
)
self.is_training = tf.placeholder_with_default(bool(is_train), [],
name='is_training')
self.build(is_train=is_train)
def load_ilp_config(self):
parser = argparse.ArgumentParser()
batch_size = self.batch_size
parser.add_argument('--BS',
default=16 * 2,
help='Batch Size',
type=int)
parser.add_argument('--T',
default=1,
help='Number of forward chain',
type=int)
parser.add_argument('--MAXTERMS',
default=6,
help='Maximum number of terms in each clause',
type=int)
parser.add_argument('--L1',
default=0,
help='Penalty for maxterm',
type=float)
parser.add_argument(
'--L2',
default=0,
help='Penalty for distance from binary for weights',
type=float)
parser.add_argument(
'--L3',
default=0,
help='Penalty for distance from binary for each term',
type=float)
parser.add_argument('--L2LOSS',
default=0,
help='Use L2 instead of cross entropy',
type=int)
parser.add_argument('--SYNC',
default=0,
help='Synchronized Update',
type=int)
parser.add_argument('--GPU', default=1, help='Use GPU', type=int)
parser.add_argument('--TB',
default=0,
help='Use Tensorboard',
type=int)
parser.add_argument('--SEED', default=0, help='Random seed', type=int)
self.args_ilp = parser.parse_args()
def define_preds(self):
nL = 6
nD = 16
D = ['%d' % i for i in range(nD)]
self.Constants = dict({'D': D})
self.predColl = PredCollection(self.Constants)
self.predColl.add_pred(dname='eq', arguments=['D', 'D'])
self.predColl.add_pred(dname='ltD', arguments=['D', 'D', 'D'])
self.predColl.add_pred(dname='gtD', arguments=['D', 'D', 'D'])
self.predColl.add_pred(dname='left', arguments=['D'])
self.predColl.add_pred(dname='button', arguments=['D'])
# self.predColl.add_pred(dname='top' ,arguments=['D'])
# self.predColl.add_pred(dname='right' ,arguments=['D'])
self.predColl.add_pred(dname='obj', arguments=['D'])
self.predColl.add_pred(dname='rectangle', arguments=['D'])
#instead of color(D,C) we define a set of is_
for i in range(nL):
self.predColl.add_pred(dname='is_color_%d' % i, arguments=['D'])
# for i in range(nD):
# self.predColl.add_pred(dname='is_d_%d'%i ,arguments=['D'])
for i in range(self.q_dim):
self.predColl.add_pred(dname='is_q_%d' % i, arguments=[])
self.predColl.add_pred(
dname='closer',
arguments=['D', 'D', 'D'],
variables=[],
pFunc=DNF('closer',
terms=1,
init=[1, .1, -1, .1],
sig=2,
init_terms=['obj(A), obj(B), obj(C), ltD(A,B,C)'],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='or')
self.predColl.add_pred(
dname='farther',
arguments=['D', 'D', 'D'],
variables=[],
pFunc=DNF('farther',
terms=1,
init=[1, .1, -1, .1],
sig=2,
init_terms=['obj(A), obj(B), obj(C), gtD(A,B,C)'],
predColl=self.predColl,
fast=True),
use_neg=False,
Fam='or')
self.predColl.add_pred(dname='notClosest',
arguments=['D', 'D'],
variables=['D'],
pFunc=DNF('notClosest',
terms=3,
init=[1, .1, -1, .1],
sig=2,
init_terms=[
'closer(A,C,B)', 'not obj(A)',
'not obj(B)', 'eq(A,B)'
],
predColl=self.predColl,
fast=True,
neg=True),
use_neg=True,
Fam='eq')
self.predColl.add_pred(dname='notFarthest',
arguments=['D', 'D'],
variables=['D'],
pFunc=DNF('notFarthest',
terms=3,
init=[1, .1, -1, .1],
sig=2,
init_terms=[
'farther(A,C,B)', 'not obj(A)',
'not obj(B)', 'eq(A,B)'
],
predColl=self.predColl,
fast=True,
neg=True),
use_neg=True,
Fam='eq')
exc = ['CL_%d' % i for i in range(self.a_dim)]
self.predColl.add_pred(
dname='qa',
oname='qa',
arguments=['D'],
variables=[],
pFunc=DNF('qa',
predColl=self.predColl,
init_terms=[
'is_q_0(), is_color_0(A)', 'is_q_1(), is_color_1(A)',
'is_q_2(), is_color_2(A)', 'is_q_3(), is_color_3(A)',
'is_q_4(), is_color_4(A)', 'is_q_5(), is_color_5(A)'
],
fast=True),
use_neg=True,
Fam='eq',
exc_conds=[('*', 'rep1')],
exc_preds=exc)
for k in range(0, 10):
post_term = [('and', 'qa(A)')]
if k == 6:
post_term.append(('and', 'not rectangle(A)'))
if k == 7:
post_term.append(('and', 'rectangle(A)'))
post_terms = []
self.predColl.add_pred(dname='CL_%d' % k,
oname='CL_%d' % k,
arguments=[],
variables=['D', 'D'],
pFunc=DNF('CL_%d' % k,
terms=12,
init=[-1, -1, -1, -1],
sig=1,
predColl=self.predColl,
post_terms=post_term),
use_neg=True,
Fam='eq',
exc_preds=exc + ['eq', 'ltD', 'gtD'])
self.predColl.initialize_predicates()
self.bg = Background(self.predColl)
for i in range(nD):
self.bg.add_backgroud('eq', ('%d' % i, '%d' % i))
ri, ci = int(i // 4), int(i % 4)
Y1 = (ri + 0.5) * 28 + ri**2
X1 = (ci + 0.5) * 28 + ci**2
if Y1 > 64:
self.bg.add_backgroud('button', ('%d' % i, ))
if X1 < 64:
self.bg.add_backgroud('left', ('%d' % i, ))
for j in range(nD):
rj, cj = int(j // 4), int(j % 4)
for k in range(nD):
rk, ck = int(k // 4), int(k % 4)
Y2 = (rj + 0.5) * 28 + rj**2
X2 = (cj + 0.5) * 28 + cj**2
Y3 = (rk + 0.5) * 28 + rk**2
X3 = (ck + 0.5) * 28 + ck**2
d1 = 1.1 * (X1 - X2)**2 + (Y1 - Y2)**2
d2 = 1.1 * (X1 - X3)**2 + (Y1 - Y3)**2
if (d1 < d2 and i != j and i != k and j != k):
self.bg.add_backgroud('ltD',
('%d' % i, '%d' % j, '%d' % k))
if (d1 > d2 and i != j and i != k and j != k):
self.bg.add_backgroud('gtD',
('%d' % i, '%d' % j, '%d' % k))
bg_set = []
self.X0 = OrderedDict()
for p in self.predColl.outpreds:
if p.oname not in bg_set:
tmp = tf.expand_dims(
tf.constant(self.bg.get_X0(p.oname), tf.float32), 0)
self.X0[p.oname] = tf.tile(tmp, [self.batch_size, 1])
print('displaying config setting...')
self.mdl = ILPRLEngine(args=self.args_ilp,
predColl=self.predColl,
bgs=None)
def get_feed_dict(self, batch_chunk, step=None, is_training=None):
fd = {
self.img: batch_chunk['img'], # [B, h, w, c]
self.q: batch_chunk['q'], # [B, n]
self.a: batch_chunk['a'], # [B, m]
}
if is_training is not None:
fd[self.is_training] = is_training
return fd
def build(self, is_train=True):
n = self.a_dim
conv_info = self.conv_info
# build loss and accuracy {{{
def build_loss(logits, labels):
# Cross-entropy loss
loss = tf.nn.softmax_cross_entropy_with_logits(
logits=logits, labels=labels) + .02 * self.pen
correct_prediction = tf.equal(tf.argmax(logits, 1),
tf.argmax(labels, 1))
self.acc = tf.cast(correct_prediction, tf.float32)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
return tf.reduce_mean(loss), accuracy
def concat_coor(o, i, d):
coor = tf.tile(
tf.expand_dims([float(int(i / d)) / d, (i % d) / d], axis=0),
[self.batch_size, 1])
o = tf.concat([o, tf.to_float(coor)], axis=1)
return o
def CONV(img, q, scope='CONV'):
with tf.variable_scope(scope) as scope:
log.warn(scope.name)
conv_1 = conv2d(img,
conv_info[0],
is_train,
s_h=3,
s_w=3,
name='conv_1',
batch_norm=True,
activation_fn=tf.nn.relu)
conv_2 = conv2d(conv_1,
conv_info[1],
is_train,
s_h=3,
s_w=3,
name='conv_2',
batch_norm=True,
activation_fn=tf.nn.relu)
conv_3 = conv2d(conv_2,
conv_info[2],
is_train,
name='conv_3',
batch_norm=True,
activation_fn=tf.nn.relu)
conv_4 = conv2d(conv_3,
conv_info[3] * 2,
is_train,
name='conv_4',
batch_norm=True,
activation_fn=tf.nn.relu)
d = conv_4.get_shape().as_list()[1]
all_g = []
for i in range(d * d):
o_i = conv_4[:, int(i / d), int(i % d), :]
all_g.append(o_i)
feat = tf.stack(all_g, axis=1)
nD = 16
nL = 6
for i in range(self.q_dim):
self.X0['is_q_%d' % i] = q[:, i:(i + 1)]
def makebin(x, t, fn):
#return fn(x)
fw = tf.cast(tf.greater_equal(x, t), tf.float32)
return custom_grad(fw, fn(x))
f = tf.layers.dense(feat, 1)
self.X0['rectangle'] = makebin(f[:, :, 0], 0.0, tf.sigmoid)
#self.pen = tf.reduce_mean( self.X0['rectangle']*(1.0-self.X0['rectangle']))
f = tf.layers.dense(feat, 7, tf.nn.softmax)
#self.pen += tf.reduce_mean( f*(1.0-f))
for i in range(6):
self.X0['is_color_%d' % i] = makebin(
f[:, :, i + 1], .5, tf.identity)
self.X0['obj'] = (1.0 - makebin(f[:, :, 0], .5, tf.identity))
#self.pen = tf.reduce_mean ( tf.square( tf.reduce_sum(self.X0['obj'],-1)-4.0 ))*.01
with tf.variable_scope('myscope'):
self.XO, L3 = self.mdl.getTSteps(self.X0)
loss_w = 0
loss_sum = 0
vs = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope='CONV/myscope')
for v in vs:
if v is None:
continue
print(v)
w = tf.sigmoid(2 * v)
loss_w += tf.reduce_mean(w * (1 - .0 - w))
loss_sum += tf.reduce_mean(
tf.nn.relu(tf.reduce_sum(w, -1) - 6))
self.pen = loss_w + loss_sum
os = tf.concat(
[self.XO['CL_%d' % i] for i in range(self.a_dim)], -1)
#os = tf.concat( [ custom_grad( sharp_sigmoid(self.XO['CL_%d'%i]-.5,7),self.XO['CL_%d'%i] ) for i in range(self.a_dim)],-1)
return 10 * os
self.load_ilp_config()
self.define_preds()
logits = CONV(self.img, self.q, scope='CONV')
# logits = f_phi(g, scope='f_phi')
self.all_preds = tf.nn.softmax(logits)
self.loss, self.accuracy = build_loss(logits, self.a)
# Add summaries
def draw_iqa(img, q, target_a, pred_a):
fig, ax = tfplot.subplots(figsize=(6, 6))
ax.imshow(img)
ax.set_title(question2str(q))
ax.set_xlabel(
answer2str(target_a) + answer2str(pred_a, 'Predicted'))
return fig
try:
tfplot.summary.plot_many(
'IQA/',
draw_iqa, [self.img, self.q, self.a, self.all_preds],
max_outputs=4,
collections=["plot_summaries"])
except:
pass
tf.summary.scalar("loss/accuracy", self.accuracy)
tf.summary.scalar("loss/cross_entropy", self.loss)
log.warn('Successfully loaded the model.')
