def test_frcnn_roi_pooling(self):
'''
need to unzip test_roi_pooling.pkl.zip first
'''
tf.reset_default_graph()
with open(self.datadir + 'test_roi_pooling.pkl', 'rb') as fid:
df = pickle.load(fid)
reshapein, rois, pool_5_ref = df
frcnn = FRCNN(DEBUG=True)
input_ph = tf.placeholder(dtype=tf.float32,
shape=[None, None, None, None])
input_ph1 = tf.placeholder(dtype=tf.float32, shape=[None, None])
pool5 = frcnn.frcnn_roi_pooling(input_ph,
input_ph1,
7,
7,
1. / 16,
name='pool_5')
with tf.Session() as sess:
pool_5_r = sess.run(pool5,
feed_dict={
input_ph: reshapein,
input_ph1: rois[0]
})
assert (np.equal(pool_5_ref, pool_5_r).all())
def test_frcnn_proposal_target(self):
tf.reset_default_graph()
with open(self.datadir + 'test_proposal_target.pkl', 'rb') as fid:
df = pickle.load(fid)
with open(self.datadir + 'train_cfg.pkl') as fid:
cfg1 = pickle.load(fid)
roiin,gt, \
rois_ref,labels_ref,bbox_targets_ref,bbox_iw_ref,bbox_ow_ref = df
frcnn = FRCNN(DEBUG=True)
cfg.TRAIN = cfg1.TRAIN
n_classes = 21
rois, labels, bbox_targets, bbox_iw, bbox_ow = \
frcnn.frcnn_proposal_target( roiin, gt, n_classes )
with tf.Session() as sess:
rois_r, labels_r, bbox_targets_r, bbox_iw_r, bbox_ow_r = \
sess.run([ rois, labels, bbox_targets, bbox_iw, bbox_ow ])
assert (np.equal(rois_r, rois_ref).all())
assert (np.equal(labels_r, labels_ref).all())
assert (np.equal(bbox_targets_r, bbox_targets_ref).all())
assert (np.equal(bbox_iw_r, bbox_iw_ref).all())
assert (np.equal(bbox_ow_r, bbox_ow_ref).all())
def test_frcnn_head(self):
frcnn = FRCNN(DEBUG=True)
input_ph = tf.placeholder(dtype=tf.float32, shape=[256, 7, 7, 512])
cls, clsp, bbox = frcnn.frcnn_head(input_ph,
num_classes=21,
isTrain=True)
assert (cls.shape == [256, 21])
assert (clsp.shape == [256, 21])
assert (bbox.shape == [256, 4 * 21])
def setup(self):
'''
connections:
data
|
[ VGG16BackBone ]
|
conv5_3, gt_boxes, im_info
| | |
| | [ RPN ]
| | |
| | rois, rpn_cross_entropy, rpn_loss_box
| | |
[ FRCNN ]
|
cls_prob,bbox_pred, cross_entropy,loss_box
'''
self.layers = {}
##### backbone
self.vggbk = VGG16BackBone(self.data)
conv5_3 = self.vggbk.get_output()
##### RPN
rpn_rois, rpn_cross_entro, rpn_loss_box = \
RPN().create(
conv5_3,
self.im_info,
self.gt_boxes,
isTrain = self.trainable
).get_outputs()
self.layers['rpn_cross_entropy'] = rpn_cross_entro
self.layers['rpn_loss_box'] = rpn_loss_box
##### FRCNN
self.frcnn = FRCNN().create(
conv5_3,
rpn_rois,
self.gt_boxes,
num_classes = n_classes,
isTrain = self.trainable
)
cls_prob, bbox_pred, cross_entropy, loss_box = self.frcnn.get_outputs()
self.layers['cls_score'] = tf.get_default_graph().get_tensor_by_name( 'cls_score/BiasAdd:0' )
self.layers['bbox_pred'] = bbox_pred
self.layers['cls_prob'] = cls_prob
self.layers['cross_entropy'] = cross_entropy
self.layers['loss_box'] = loss_box
self.layers['rois'] = rpn_rois
def test_frcnn_testmode(self):
tf.reset_default_graph()
reshapein = tf.placeholder(dtype=tf.float32,
shape=(None, None, None, 512))
rpn_rois = tf.placeholder(dtype=tf.float32, shape=(None, 5))
frcnn = FRCNN(DEBUG=True)
model = frcnn.create(reshapein,
rpn_rois,
num_classes=21,
isTrain=False)
cls_prob, bbox_pred, cross_entropy, loss_box = model.get_outputs()
def test_restore_params(self):
tf.reset_default_graph()
frcnn = FRCNN(DEBUG=True)
input_ph = tf.placeholder(dtype=tf.float32, shape=[256, 7, 7, 512])
frcnn.frcnn_head(input_ph, num_classes=21, isTrain=True)
weights = os.path.join('../data/pretrain_model', 'vgg16_weights.npz')
with tf.Session() as sess:
frcnn.restore_params(sess, weights)
def test_frcnn_all(self):
'''
it is too larget to store test_all.pkl
to regenerate:
get input output from frcnn and all weights/bias tensors
make sure dropout keep prob is 1
before npr.choice in sampling_roi function, reset rand seed to 0
store all tensor values into self.datadir+'test_all.pkl'
'''
tf.reset_default_graph()
with open(self.datadir + 'test_all.pkl', 'rb') as fid:
df = pickle.load(fid)
reshapein , rpn_rois , gt, \
cls_prob_res,bbox_pred_res, \
w0, b0, w1, b1, w2, b2, w3, b3 = df
frcnn = FRCNN(DEBUG=True)
model = frcnn.create(reshapein,
rpn_rois,
gt,
num_classes=21,
isTrain=True)
with open(self.datadir + 'train_cfg.pkl') as fid:
cfg1 = pickle.load(fid)
cfg.TRAIN = cfg1.TRAIN
cls_prob, bbox_pred, cross_entropy, loss_box = model.get_outputs()
with tf.Session() as sess:
sess.run(
tf.assign(
tf.get_default_graph().get_tensor_by_name('fc6/kernel:0'),
w0))
sess.run(
tf.assign(
tf.get_default_graph().get_tensor_by_name('fc6/bias:0'),
b0))
sess.run(
tf.assign(
tf.get_default_graph().get_tensor_by_name('fc7/kernel:0'),
w1))
sess.run(
tf.assign(
tf.get_default_graph().get_tensor_by_name('fc7/bias:0'),
b1))
sess.run(
tf.assign(
tf.get_default_graph().get_tensor_by_name(
'cls_score/kernel:0'), w2))
sess.run(
tf.assign(
tf.get_default_graph().get_tensor_by_name(
'cls_score/bias:0'), b2))
sess.run(
tf.assign(
tf.get_default_graph().get_tensor_by_name(
'bbox_pred/kernel:0'), w3))
sess.run(
tf.assign(
tf.get_default_graph().get_tensor_by_name(
'bbox_pred/bias:0'), b3))
cls_prob_r, bbox_pred_r, \
ce_r, lb_r = sess.run([ cls_prob, bbox_pred, cross_entropy, loss_box ])
assert (np.equal(cls_prob_res, cls_prob_r).all())
assert (np.equal(bbox_pred_res, bbox_pred_r).all())
class VGGnet_train_wrapper(object):
def __init__(self, trainable=True):
self.inputs = []
self.data = tf.placeholder(tf.float32, shape=[None, None, None, 3])
self.im_info = tf.placeholder(tf.float32, shape=[None, 3])
self.gt_boxes = tf.placeholder(tf.float32, shape=[None, 5])
self.keep_prob = tf.placeholder(tf.float32)
self.layers = dict({'data':self.data, 'im_info':self.im_info, 'gt_boxes':self.gt_boxes})
self.trainable = trainable
self.setup()
# create ops and placeholders for bbox normalization process
with tf.variable_scope('bbox_pred', reuse=True):
weights = tf.get_variable("kernel")
biases = tf.get_variable("bias")
self.bbox_weights = tf.placeholder(weights.dtype, shape=weights.get_shape())
self.bbox_biases = tf.placeholder(biases.dtype, shape=biases.get_shape())
self.bbox_weights_assign = weights.assign(self.bbox_weights)
self.bbox_bias_assign = biases.assign(self.bbox_biases)
def get_output(self, layer):
try:
layer = self.layers[layer]
except KeyError:
print self.layers.keys()
raise KeyError('Unknown layer name fed: %s'%layer)
return layer
def load(self, data_path, session, saver, ignore_missing=False):
if data_path.endswith('.ckpt'):
saver.restore(session, data_path)
else:
self.vggbk.restore_params(session, data_path)
def setup(self):
'''
connections:
data
|
[ VGG16BackBone ]
|
conv5_3, gt_boxes, im_info
| | |
| | [ RPN ]
| | |
| | rois, rpn_cross_entropy, rpn_loss_box
| | |
[ FRCNN ]
|
cls_prob,bbox_pred, cross_entropy,loss_box
'''
self.layers = {}
##### backbone
self.vggbk = VGG16BackBone(self.data)
conv5_3 = self.vggbk.get_output()
##### RPN
rpn_rois, rpn_cross_entro, rpn_loss_box = \
RPN().create(
conv5_3,
self.im_info,
self.gt_boxes,
isTrain = self.trainable
).get_outputs()
self.layers['rpn_cross_entropy'] = rpn_cross_entro
self.layers['rpn_loss_box'] = rpn_loss_box
##### FRCNN
self.frcnn = FRCNN().create(
conv5_3,
rpn_rois,
self.gt_boxes,
num_classes = n_classes,
isTrain = self.trainable
)
cls_prob, bbox_pred, cross_entropy, loss_box = self.frcnn.get_outputs()
self.layers['cls_score'] = tf.get_default_graph().get_tensor_by_name( 'cls_score/BiasAdd:0' )
self.layers['bbox_pred'] = bbox_pred
self.layers['cls_prob'] = cls_prob
self.layers['cross_entropy'] = cross_entropy
self.layers['loss_box'] = loss_box
self.layers['rois'] = rpn_rois