def __init__(self):
super(TouchEmbNet2DML, self).__init__()
self.touch_net = Feat2d(hyp.touch_emb_dim)
self.num_samples = hyp.emb_2D_touch_num_samples
assert (self.num_samples >
0), "num samples should be greater than zero"
self.batch_k = 2
self.sampler = utils_misc.DistanceWeightedSampling(
batch_k=self.batch_k, normalize=False)
self.criterion = utils_misc.MarginLoss()
self.beta = 1.2
def __init__(self):
super(EmbNet2D, self).__init__()
print('EmbNet2D...')
self.batch_k = 2
self.num_samples = hyp.emb_2D_num_samples
assert (self.num_samples > 0)
self.sampler = utils_misc.DistanceWeightedSampling(
batch_k=self.batch_k, normalize=False)
self.criterion = utils_misc.MarginLoss(
) #margin=args.margin,nu=args.nu)
self.beta = 1.2
def __init__(self):
super(TriNet2D, self).__init__()
print('TriNet2D...')
self.batch_k = 2
self.num_samples = hyp.tri_2D_num_samples
assert (self.num_samples > 0)
self.sampler = utils_misc.DistanceWeightedSampling(
batch_k=self.batch_k, normalize=False)
self.criterion = utils_misc.MarginLoss(
) #margin=args.margin,nu=args.nu)
self.beta = 1.2
self.dict_len = 20000
self.neg_pool = utils_misc.SimplePool(self.dict_len)
self.ce = torch.nn.CrossEntropyLoss()
def __init__(self):
"""The idea is here the featnet would be frozen, and using some X number
of images it will form the features for the object, which will be used
to get the closeby embeddings to learn metric learning
"""
super(TouchEmbed2D, self).__init__()
if hyp.do_feat:
print('using the visual feat net to generate visual feature tensor')
self.featnet = FeatNet(input_dim=4) # passing occXs and occXs * unpXs
if hyp.do_touch_feat:
print('using 2d backbone network to generate features from sensor depth image')
self.backbone_2D = VGGNet.Feat2d(touch_emb_dim=hyp.feat_dim, do_bn=hyp.do_bn)
# if I need to add the 3d encoder decoder I need to add the next line
# self.touch_featnet = FeatNet(input_dim=1) # just passing occRs here
if hyp.do_touch_forward:
# now here I need to pass this through the bottle3d architecture to predict
# 1-d vectors
print('using context net to turn 3d context grid into 1d feature tensor')
self.context_net = bottle3D.Bottle3D(in_channel=hyp.feat_dim,\
pred_dim=hyp.feat_dim)
if hyp.do_touch_occ:
print('this should not be turned on')
from IPython import embed; embed()
self.touch_occnet = OccNet()
# if hyp.do_touch_embML:
# self.touch_embnet3D = EmbNet3D() # metric learning for making touch feature tensor same as visual feature tensor
if hyp.do_freeze_feat:
print('freezing visual features')
self.featnet = self.featnet.eval()
assert self.featnet.training == False, "since I am not training FeatNet it should be false"
if hyp.do_freeze_touch_feat:
print('freezing backbone_2D')
self.backbone_2D = self.backbone_2D.eval()
if hyp.do_freeze_touch_forward:
print('freezing context net')
self.context_net = self.context_net.eval()
# hyperparams for embedding training not really needed as of now
if hyp.do_touch_embML:
print('Instantiating Contrastive ML loss')
self.num_pos_samples = hyp.emb_3D_num_samples
self.batch_k = 2
self.n_negatives = 1
assert (self.num_pos_samples > 0)
self.sampler = utils_misc.DistanceWeightedSampling(batch_k=self.batch_k, normalize=False,
num_neg_samples=self.n_negatives)
self.criterion = utils_misc.MarginLoss()
self.beta = 1.2
if hyp.do_moc or hyp.do_eval_recall:
print('Instantiating MOC net')
self.key_touch_featnet = VGGNet.Feat2d(touch_emb_dim=hyp.feat_dim,\
do_bn=hyp.do_bn)
key_weights = self.backbone_2D.state_dict()
self.key_touch_featnet.load_state_dict(key_weights)
self.key_context_net = bottle3D.Bottle3D(in_channel=hyp.feat_dim,\
pred_dim=hyp.feat_dim)
key_context_weights = self.context_net.state_dict()
self.key_context_net.load_state_dict(key_context_weights)
# check that the two networks indeed have the same weights
p1 = get_params(self.backbone_2D)
p2 = get_params(self.key_touch_featnet)
assert check_equal(p1, p2),\
"initially both the touch networks should have same weights"
cp1 = get_params(self.context_net)
cp2 = get_params(self.key_context_net)
assert check_equal(cp1, cp2),\
"initially both the context networks should have same weights"
self.moc_ml_net = MOCTrainingTouch(dict_len=hyp.dict_len,\
num_neg_samples=hyp.num_neg_samples)