neg+=negtm
pickle.dump([neg, pos], open('negpos.pkl', 'wb'))
"""
neg, pos = pickle.load(open('negpos.pkl', 'rb'))
#neg = create_neg(input_shape)[None,:,:,:]
#pos= create_pos(input_shape)[None,:,:,:]
inputs = tf.placeholder(tf.float32,
shape=(batch_size, input_shape[0], input_shape[1], 1))
modify = []
for i in range(1, 4):
modify.append('conv%d' % i)
logits, net, activations, modifys = sc(inputs, modify=modify)
print modifys
modifyv = {}
for i in range(1, 4):
name = 'conv%d' % i
print name
modifyv[name] = np.ones(activations[name].shape)
with tf.Session() as sess:
saver = tf.train.Saver()
#saver.restore(sess,'ckpts5/39900.ckpt')
#saver.restore(sess,'ckpts_manual_noise_gauss/1.ckpt')
#saver.restore(sess,'ckpts_manual/1.ckpt')
saver.restore(sess, 'ckpts_manual_noise_gauss2/1.ckpt')
fd = {modifys['conv%d' % i]: modifyv['conv%d' % i] for i in range(1, 4)}
fdpos = {inputs: pos}
fdpos.update(fd)
from sc_manual import sc
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
from sample_maker import create_pos, create_neg
# Create mini-batch for demo
input_shape = (8, 8)
batch_size = 1
ch = 3
gradients = OrderedDict()
activations = OrderedDict()
inputs = tf.placeholder(tf.float32,
shape=(batch_size, input_shape[0], input_shape[1], 1))
logits, net, activations = sc(inputs)
with tf.Session() as sess:
saver = tf.train.Saver()
#saver.restore(sess,'ckpts_manual_noise/1.ckpt')
#saver.restore(sess,'ckpts_manual_noise_0mean/1.ckpt')
saver.restore(sess, 'ckpts_manual_noise_gauss2/1.ckpt')
for tv in tf.trainable_variables():
tvv = sess.run(tv)
if tv.name.find('weights') > -1:
_, _, bchs, tchs = tvv.shape
for bch, tch in product(range(bchs), range(tchs)):
print tv.name, bch, tch
print tvv[:, :, bch, tch]
if tv.name.find('biases') > -1:
tchs = tvv.shape[0]
for tch in range(tchs):