class TouchEmbed2D(nn.Module):
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)
def forward(self, feed, moc_init_done=False, debug=False):
summ_writer = utils_improc.Summ_writer(
writer = feed['writer'],
global_step = feed['global_step'],
set_name= feed['set_name'],
fps=8)
writer = feed['writer']
global_step = feed['global_step']
total_loss = torch.tensor(0.0).cuda()
### ... All things sensor ... ###
sensor_rgbs = feed['sensor_imgs']
sensor_depths = feed['sensor_depths']
center_sensor_H, center_sensor_W = sensor_depths[0][0].shape[-1] // 2, sensor_depths[0][0].shape[-2] // 2
### ... All things sensor end ... ###
# 1. Form the memory tensor using the feat net and visual images.
# check what all do you need for this and create only those things
## .... Input images .... ##
rgb_camRs = feed['rgb_camRs']
rgb_camXs = feed['rgb_camXs']
## .... Input images end .... ##
## ... Hyperparams ... ##
B, H, W, V, S = hyp.B, hyp.H, hyp.W, hyp.V, hyp.S
__p = lambda x: pack_seqdim(x, B)
__u = lambda x: unpack_seqdim(x, B)
PH, PW = hyp.PH, hyp.PW
Z, Y, X = hyp.Z, hyp.Y, hyp.X
Z2, Y2, X2 = int(Z/2), int(Y/2), int(X/2)
## ... Hyperparams end ... ##
## .... VISUAL TRANSFORMS BEGIN .... ##
pix_T_cams = feed['pix_T_cams']
pix_T_cams_ = __p(pix_T_cams)
origin_T_camRs = feed['origin_T_camRs']
origin_T_camRs_ = __p(origin_T_camRs)
origin_T_camXs = feed['origin_T_camXs']
origin_T_camXs_ = __p(origin_T_camXs)
camRs_T_camXs_ = torch.matmul(utils_geom.safe_inverse(
origin_T_camRs_), origin_T_camXs_)
camXs_T_camRs_ = utils_geom.safe_inverse(camRs_T_camXs_)
camRs_T_camXs = __u(camRs_T_camXs_)
camXs_T_camRs = __u(camXs_T_camRs_)
pix_T_cams_ = utils_geom.pack_intrinsics(pix_T_cams_[:, 0, 0], pix_T_cams_[:, 1, 1], pix_T_cams_[:, 0, 2],
pix_T_cams_[:, 1, 2])
pix_T_camRs_ = torch.matmul(pix_T_cams_, camXs_T_camRs_)
pix_T_camRs = __u(pix_T_camRs_)
## ... VISUAL TRANSFORMS END ... ##
## ... SENSOR TRANSFORMS BEGIN ... ##
sensor_origin_T_camXs = feed['sensor_extrinsics']
sensor_origin_T_camXs_ = __p(sensor_origin_T_camXs)
sensor_origin_T_camRs = feed['sensor_origin_T_camRs']
sensor_origin_T_camRs_ = __p(sensor_origin_T_camRs)
sensor_camRs_T_origin_ = utils_geom.safe_inverse(sensor_origin_T_camRs_)
sensor_camRs_T_camXs_ = torch.matmul(utils_geom.safe_inverse(
sensor_origin_T_camRs_), sensor_origin_T_camXs_)
sensor_camXs_T_camRs_ = utils_geom.safe_inverse(sensor_camRs_T_camXs_)
sensor_camRs_T_camXs = __u(sensor_camRs_T_camXs_)
sensor_camXs_T_camRs = __u(sensor_camXs_T_camRs_)
sensor_pix_T_cams = feed['sensor_intrinsics']
sensor_pix_T_cams_ = __p(sensor_pix_T_cams)
sensor_pix_T_cams_ = utils_geom.pack_intrinsics(sensor_pix_T_cams_[:, 0, 0], sensor_pix_T_cams_[:, 1, 1],
sensor_pix_T_cams_[:, 0, 2], sensor_pix_T_cams_[:, 1, 2])
sensor_pix_T_camRs_ = torch.matmul(sensor_pix_T_cams_, sensor_camXs_T_camRs_)
sensor_pix_T_camRs = __u(sensor_pix_T_camRs_)
## .... SENSOR TRANSFORMS END .... ##
## .... Visual Input point clouds .... ##
xyz_camXs = feed['xyz_camXs']
xyz_camXs_ = __p(xyz_camXs)
xyz_camRs_ = utils_geom.apply_4x4(camRs_T_camXs_, xyz_camXs_) # (40, 4, 4) (B*S, N, 3)
xyz_camRs = __u(xyz_camRs_)
assert all([torch.allclose(xyz_camR, inp_xyz_camR) for xyz_camR, inp_xyz_camR in zip(
xyz_camRs, feed['xyz_camRs']
)]), "computation of xyz_camR here and those computed in input do not match"
## .... Visual Input point clouds end .... ##
## ... Sensor input point clouds ... ##
sensor_xyz_camXs = feed['sensor_xyz_camXs']
sensor_xyz_camXs_ = __p(sensor_xyz_camXs)
sensor_xyz_camRs_ = utils_geom.apply_4x4(sensor_camRs_T_camXs_, sensor_xyz_camXs_)
sensor_xyz_camRs = __u(sensor_xyz_camRs_)
assert all([torch.allclose(sensor_xyz, inp_sensor_xyz) for sensor_xyz, inp_sensor_xyz in zip(
sensor_xyz_camRs, feed['sensor_xyz_camRs']
)]), "the sensor_xyz_camRs computed in forward do not match those computed in input"
## ... visual occupancy computation voxelize the pointcloud from above ... ##
occRs_ = utils_vox.voxelize_xyz(xyz_camRs_, Z, Y, X)
occXs_ = utils_vox.voxelize_xyz(xyz_camXs_, Z, Y, X)
occRs_half_ = utils_vox.voxelize_xyz(xyz_camRs_, Z2, Y2, X2)
occXs_half_ = utils_vox.voxelize_xyz(xyz_camXs_, Z2, Y2, X2)
## ... visual occupancy computation end ... NOTE: no unpacking ##
## .. visual occupancy computation for sensor inputs .. ##
sensor_occRs_ = utils_vox.voxelize_xyz(sensor_xyz_camRs_, Z, Y, X)
sensor_occXs_ = utils_vox.voxelize_xyz(sensor_xyz_camXs_, Z, Y, X)
sensor_occRs_half_ = utils_vox.voxelize_xyz(sensor_xyz_camRs_, Z2, Y2, X2)
sensor_occXs_half_ = utils_vox.voxelize_xyz(sensor_xyz_camXs_, Z2, Y2, X2)
## ... unproject rgb images ... ##
unpRs_ = utils_vox.unproject_rgb_to_mem(__p(rgb_camXs), Z, Y, X, pix_T_camRs_)
unpXs_ = utils_vox.unproject_rgb_to_mem(__p(rgb_camXs), Z, Y, X, pix_T_cams_)
## ... unproject rgb finish ... NOTE: no unpacking ##
## ... Make depth images ... ##
depth_camXs_, valid_camXs_ = utils_geom.create_depth_image(pix_T_cams_, xyz_camXs_, H, W)
dense_xyz_camXs_ = utils_geom.depth2pointcloud(depth_camXs_, pix_T_cams_)
dense_xyz_camRs_ = utils_geom.apply_4x4(camRs_T_camXs_, dense_xyz_camXs_)
inbound_camXs_ = utils_vox.get_inbounds(dense_xyz_camRs_, Z, Y, X).float()
inbound_camXs_ = torch.reshape(inbound_camXs_, [B*S, 1, H, W])
valid_camXs = __u(valid_camXs_) * __u(inbound_camXs_)
## ... Make depth images ... ##
## ... Make sensor depth images ... ##
sensor_depth_camXs_, sensor_valid_camXs_ = utils_geom.create_depth_image(sensor_pix_T_cams_,
sensor_xyz_camXs_, H, W)
sensor_dense_xyz_camXs_ = utils_geom.depth2pointcloud(sensor_depth_camXs_, sensor_pix_T_cams_)
sensor_dense_xyz_camRs_ = utils_geom.apply_4x4(sensor_camRs_T_camXs_, sensor_dense_xyz_camXs_)
sensor_inbound_camXs_ = utils_vox.get_inbounds(sensor_dense_xyz_camRs_, Z, Y, X).float()
sensor_inbound_camXs_ = torch.reshape(sensor_inbound_camXs_, [B*hyp.sensor_S, 1, H, W])
sensor_valid_camXs = __u(sensor_valid_camXs_) * __u(sensor_inbound_camXs_)
### .. Done making sensor depth images .. ##
### ... Sanity check ... Write to tensorboard ... ###
summ_writer.summ_oneds('2D_inputs/depth_camXs', torch.unbind(__u(depth_camXs_), dim=1))
summ_writer.summ_oneds('2D_inputs/valid_camXs', torch.unbind(valid_camXs, dim=1))
summ_writer.summ_rgbs('2D_inputs/rgb_camXs', torch.unbind(rgb_camXs, dim=1))
summ_writer.summ_rgbs('2D_inputs/rgb_camRs', torch.unbind(rgb_camRs, dim=1))
summ_writer.summ_occs('3d_inputs/occXs', torch.unbind(__u(occXs_), dim=1), reduce_axes=[2])
summ_writer.summ_unps('3d_inputs/unpXs', torch.unbind(__u(unpXs_), dim=1),\
torch.unbind(__u(occXs_), dim=1))
# A different approach for viewing occRs of sensors
sensor_occRs = __u(sensor_occRs_)
vis_sensor_occRs = torch.max(sensor_occRs, dim=1, keepdim=True)[0]
# summ_writer.summ_occs('3d_inputs/sensor_occXs', torch.unbind(__u(sensor_occXs_), dim=1),
# reduce_axes=[2])
summ_writer.summ_occs('3d_inputs/sensor_occRs', torch.unbind(vis_sensor_occRs, dim=1), reduce_axes=[2])
### ... code for visualizing sensor depths and sensor rgbs ... ###
# summ_writer.summ_oneds('2D_inputs/depths_sensor', torch.unbind(sensor_depths, dim=1))
# summ_writer.summ_rgbs('2D_inputs/rgbs_sensor', torch.unbind(sensor_rgbs, dim=1))
# summ_writer.summ_oneds('2D_inputs/validXs_sensor', torch.unbind(sensor_valid_camXs, dim=1))
if summ_writer.save_this:
unpRs_ = utils_vox.unproject_rgb_to_mem(__p(rgb_camXs), Z, Y, X, matmul2(pix_T_cams_, camXs_T_camRs_))
unpRs = __u(unpRs_)
occRs_ = utils_vox.voxelize_xyz(xyz_camRs_, Z, Y, X)
summ_writer.summ_occs('3d_inputs/occRs', torch.unbind(__u(occRs_), dim=1), reduce_axes=[2])
summ_writer.summ_unps('3d_inputs/unpRs', torch.unbind(unpRs, dim=1),\
torch.unbind(__u(occRs_), dim=1))
### ... Sanity check ... Writing to tensoboard complete ... ###
results = list()
mask_ = None
### ... Visual featnet part .... ###
if hyp.do_feat:
featXs_input = torch.cat([__u(occXs_), __u(occXs_)*__u(unpXs_)], dim=2) # B, S, 4, H, W, D
featXs_input_ = __p(featXs_input)
freeXs_ = utils_vox.get_freespace(__p(xyz_camXs), occXs_half_)
freeXs = __u(freeXs_)
visXs = torch.clamp(__u(occXs_half_) + freeXs, 0.0, 1.0)
if type(mask_) != type(None):
assert(list(mask_.shape)[2:5] == list(featXs_input.shape)[2:5])
featXs_, validXs_, _ = self.featnet(featXs_input_, summ_writer, mask=occXs_)
# total_loss += feat_loss # Note no need of loss
validXs, featXs = __u(validXs_), __u(featXs_) # unpacked into B, S, C, D, H, W
# bring everything to ref_frame
validRs = utils_vox.apply_4x4_to_voxs(camRs_T_camXs, validXs)
visRs = utils_vox.apply_4x4_to_voxs(camRs_T_camXs, visXs)
featRs = utils_vox.apply_4x4_to_voxs(camRs_T_camXs, featXs) # This is now in memory coordinates
emb3D_e = torch.mean(featRs[:, 1:], dim=1) # context, or the features of the scene
emb3D_g = featRs[:, 0] # this is to predict, basically I will pass emb3D_e as input and hope to predict emb3D_g
vis3D_e = torch.max(validRs[:, 1:], dim=1)[0] * torch.max(visRs[:, 1:], dim=1)[0]
vis3D_g = validRs[:, 0] * visRs[:, 0]
#### ... I do not think I need this ... ####
results = {}
# # if hyp.do_eval_recall:
# # results['emb3D_e'] = emb3D_e
# # results['emb3D_g'] = emb3D_g
# #### ... Check if you need the above
summ_writer.summ_feats('3D_feats/featXs_input', torch.unbind(featXs_input, dim=1), pca=True)
summ_writer.summ_feats('3D_feats/featXs_output', torch.unbind(featXs, dim=1), pca=True)
summ_writer.summ_feats('3D_feats/featRs_output', torch.unbind(featRs, dim=1), pca=True)
summ_writer.summ_feats('3D_feats/validRs', torch.unbind(validRs, dim=1), pca=False)
summ_writer.summ_feat('3D_feats/vis3D_e', vis3D_e, pca=False)
summ_writer.summ_feat('3D_feats/vis3D_g', vis3D_g, pca=False)
# I need to aggregate the features and detach to prevent the backward pass on featnet
featRs = torch.mean(featRs, dim=1)
featRs = featRs.detach()
# ... HERE I HAVE THE VISUAL FEATURE TENSOR ... WHICH IS MADE USING 5 EVENLY SPACED VIEWS #
# FOR THE TOUCH PART, I HAVE THE OCC and THE AIM IS TO PREDICT FEATURES FROM THEM #
if hyp.do_touch_feat:
# 1. Pass all the sensor depth images through the backbone network
input_sensor_depths = __p(sensor_depths)
sensor_features_ = self.backbone_2D(input_sensor_depths)
# should normalize these feature tensors
sensor_features_ = l2_normalize(sensor_features_, dim=1)
sensor_features = __u(sensor_features_)
assert torch.allclose(torch.norm(sensor_features_, dim=1), torch.Tensor([1.0]).cuda()),\
"normalization has no effect on you huh."
if hyp.do_eval_recall:
results['sensor_features'] = sensor_features_
results['sensor_depths'] = input_sensor_depths
results['object_img'] = rgb_camRs
results['sensor_imgs'] = __p(sensor_rgbs)
# if moco is used do the same procedure as above but with a different network #
if hyp.do_moc or hyp.do_eval_recall:
# 1. Pass all the sensor depth images through the key network
key_input_sensor_depths = copy.deepcopy(__p(sensor_depths)) # bx1024x1x16x16->(2048x1x16x16)
self.key_touch_featnet.eval()
with torch.no_grad():
key_sensor_features_ = self.key_touch_featnet(key_input_sensor_depths)
key_sensor_features_ = l2_normalize(key_sensor_features_, dim=1)
key_sensor_features = __u(key_sensor_features_)
assert torch.allclose(torch.norm(key_sensor_features_, dim=1), torch.Tensor([1.0]).cuda()),\
"normalization has no effect on you huh."
# doing the same procedure for moco but with a different network end #
# do you want to do metric learning voxel point based using visual features and sensor features
if hyp.do_touch_embML and not hyp.do_touch_forward:
# trial 1: I do not pass the above obtained features through some encoder decoder in 3d
# So compute the location is ref_frame which the center of these depth images will occupy
# at all of these locations I will sample the from the visual tensor. It forms the positive pairs
# negatives are simply everything except the positive
sensor_depths_centers_x = center_sensor_W * torch.ones((hyp.B, hyp.sensor_S))
sensor_depths_centers_x = sensor_depths_centers_x.cuda()
sensor_depths_centers_y = center_sensor_H * torch.ones((hyp.B, hyp.sensor_S))
sensor_depths_centers_y = sensor_depths_centers_y.cuda()
sensor_depths_centers_z = sensor_depths[:, :, 0, center_sensor_H, center_sensor_W]
# Next use Pixels2Camera to unproject all of these together.
# merge the batch and the sequence dimension
sensor_depths_centers_x = sensor_depths_centers_x.reshape(-1, 1, 1) # BxHxW as required by Pixels2Camera
sensor_depths_centers_y = sensor_depths_centers_y.reshape(-1, 1, 1)
sensor_depths_centers_z = sensor_depths_centers_z.reshape(-1, 1, 1)
fx, fy, x0, y0 = utils_geom.split_intrinsics(sensor_pix_T_cams_)
sensor_depths_centers_in_camXs_ = utils_geom.Pixels2Camera(sensor_depths_centers_x, sensor_depths_centers_y,
sensor_depths_centers_z, fx, fy, x0, y0)
# finally use apply4x4 to get the locations in ref_cam
sensor_depths_centers_in_ref_cam_ = utils_geom.apply_4x4(sensor_camRs_T_camXs_, sensor_depths_centers_in_camXs_)
# NOTE: convert them to memory coordinates, the name is xyz so I presume it returns xyz but talk to ADAM
sensor_depths_centers_in_mem_ = utils_vox.Ref2Mem(sensor_depths_centers_in_ref_cam_, Z2, Y2, X2)
sensor_depths_centers_in_mem = sensor_depths_centers_in_mem_.reshape(hyp.B, hyp.sensor_S, -1)
if debug:
print('assert that you are not entering here')
from IPython import embed; embed()
# form a (0, 1) volume here at these locations and see if it resembles a cup
dim1 = X2 * Y2 * Z2
dim2 = X2 * Y2
dim3 = X2
binary_voxel_grid = torch.zeros((hyp.B, X2, Y2, Z2))
# NOTE: Z is the leading dimension
rounded_idxs = torch.round(sensor_depths_centers_in_mem)
flat_idxs = dim2 * rounded_idxs[0, :, 0] + dim3 * rounded_idxs[0, :, 1] + rounded_idxs[0, :, 2]
flat_idxs1 = dim2 * rounded_idxs[1, :, 0] + dim3 * rounded_idxs[1, :, 1] + rounded_idxs[1, :, 2]
flat_idxs1 = flat_idxs1 + dim1
flat_idxs1 = flat_idxs1.long()
flat_idxs = flat_idxs.long()
flattened_grid = binary_voxel_grid.flatten()
flattened_grid[flat_idxs] = 1.
flattened_grid[flat_idxs1] = 1.
binary_voxel_grid = flattened_grid.view(B, X2, Y2, Z2)
assert binary_voxel_grid[0].sum() == len(torch.unique(flat_idxs)), "some indexes are missed here"
assert binary_voxel_grid[1].sum() == len(torch.unique(flat_idxs1)), "some indexes are missed here"
# o3d.io.write_voxel_grid("forward_pass_save/grid0.ply", binary_voxel_grid[0])
# o3d.io.write_voxel_grid("forward_pass_save/grid1.ply", binary_voxel_grid[0])
# need to save these voxels
save_voxel(binary_voxel_grid[0].cpu().numpy(), "forward_pass_save/grid0.binvox")
save_voxel(binary_voxel_grid[1].cpu().numpy(), "forward_pass_save/grid1.binvox")
from IPython import embed; embed()
# use grid sample to get the visual touch tensor at these locations, NOTE: visual tensor features shape is (B, C, N)
visual_tensor_features = utils_samp.bilinear_sample3D(featRs, sensor_depths_centers_in_mem[:, :, 0],
sensor_depths_centers_in_mem[:, :, 1], sensor_depths_centers_in_mem[:, :, 2])
visual_feature_tensor = visual_tensor_features.permute(0, 2, 1)
# pack it
visual_feature_tensor_ = __p(visual_feature_tensor)
C = list(visual_feature_tensor.shape)[-1]
print('C=', C)
# do the metric learning this is the same as before.
# the code is basically copied from embnet3d.py but some changes are being made very minor
emb_vec = torch.stack((sensor_features_, visual_feature_tensor_), dim=1).view(B*self.num_samples*self.batch_k, C)
y = torch.stack([torch.range(0,self.num_samples*B-1), torch.range(0,self.num_samples*B-1)], dim=1).view(self.num_samples*B*self.batch_k)
a_indices, anchors, positives, negatives, _ = self.sampler(emb_vec)
# I need to write my own version of margin loss since the negatives and anchors may not be same dim
d_ap = torch.sqrt(torch.sum((positives - anchors)**2, dim=1) + 1e-8)
pos_loss = torch.clamp(d_ap - beta + self._margin, min=0.0)
# TODO: expand the dims of anchors and tile them and compute the negative loss
# do the pair count where you average by contributors only
# this is your total loss
# Further idea is to check what volumetric locations do each of the depth images corresponds to
# unproject the entire depth image and convert to ref. and then sample.
if hyp.do_touch_forward:
## ... Begin code for getting crops from visual memory ... ##
sensor_depths_centers_x = center_sensor_W * torch.ones((hyp.B, hyp.sensor_S))
sensor_depths_centers_x = sensor_depths_centers_x.cuda()
sensor_depths_centers_y = center_sensor_H * torch.ones((hyp.B, hyp.sensor_S))
sensor_depths_centers_y = sensor_depths_centers_y.cuda()
sensor_depths_centers_z = sensor_depths[:, :, 0, center_sensor_H, center_sensor_W]
# Next use Pixels2Camera to unproject all of these together.
# merge the batch and the sequence dimension
sensor_depths_centers_x = sensor_depths_centers_x.reshape(-1, 1, 1)
sensor_depths_centers_y = sensor_depths_centers_y.reshape(-1, 1, 1)
sensor_depths_centers_z = sensor_depths_centers_z.reshape(-1, 1, 1)
fx, fy, x0, y0 = utils_geom.split_intrinsics(sensor_pix_T_cams_)
sensor_depths_centers_in_camXs_ = utils_geom.Pixels2Camera(sensor_depths_centers_x, sensor_depths_centers_y,
sensor_depths_centers_z, fx, fy, x0, y0)
sensor_depths_centers_in_world_ = utils_geom.apply_4x4(sensor_origin_T_camXs_, sensor_depths_centers_in_camXs_) # not used by the algorithm
## this will be later used for visualization hence saving it here for now
sensor_depths_centers_in_ref_cam_ = utils_geom.apply_4x4(sensor_camRs_T_camXs_, sensor_depths_centers_in_camXs_) # not used by the algorithm
sensor_depths_centers_in_camXs = __u(sensor_depths_centers_in_camXs_).squeeze(2)
# There has to be a better way to do this, for each of the cameras in the batch I want a box of size (ch, cw, cd)
# TODO: rotation is the deviation of the box from the axis aligned do I want this
tB, tN, _ = list(sensor_depths_centers_in_camXs.shape) # 2, 512, _
boxlist = torch.zeros(tB, tN, 9) # 2, 512, 9
boxlist[:, :, :3] = sensor_depths_centers_in_camXs # this lies on the object
boxlist[:, :, 3:6] = torch.FloatTensor([hyp.contextW, hyp.contextH, hyp.contextD])
# convert the boxlist to lrtlist and to cuda
# the rt here transforms the from box coordinates to camera coordinates
box_lrtlist = utils_geom.convert_boxlist_to_lrtlist(boxlist)
# Now I will use crop_zoom_from_mem functionality to get the features in each of the boxes
# I will do it for each of the box separately as required by the api
context_grid_list = list()
for m in range(box_lrtlist.shape[1]):
curr_box = box_lrtlist[:, m, :]
context_grid = utils_vox.crop_zoom_from_mem(featRs, curr_box, 8, 8, 8,
sensor_camRs_T_camXs[:, m, :, :])
context_grid_list.append(context_grid)
context_grid_list = torch.stack(context_grid_list, dim=1)
context_grid_list_ = __p(context_grid_list)
## ... till here I believe I have not introduced any randomness, so the points are still in
## ... End code for getting crops around this center of certain height, width and depth ... ##
## ... Begin code for passing the context grid through 3D CNN to obtain a vector ... ##
sensor_cam_locs = feed['sensor_locs'] # these are in origin coordinates
sensor_cam_quats = feed['sensor_quats'] # this too in in world_coordinates
sensor_cam_locs_ = __p(sensor_cam_locs)
sensor_cam_quats_ = __p(sensor_cam_quats)
sensor_cam_locs_in_R_ = utils_geom.apply_4x4(sensor_camRs_T_origin_, sensor_cam_locs_.unsqueeze(1)).squeeze(1)
# TODO TODO TODO confirm that this is right? TODO TODO TODO
get_r_mat = lambda cam_quat: transformations.quaternion_matrix_py(cam_quat)
rot_mat_Xs_ = torch.from_numpy(np.stack(list(map(get_r_mat, sensor_cam_quats_.cpu().numpy())))).to(sensor_cam_locs_.device).float()
rot_mat_Rs_ = torch.bmm(sensor_camRs_T_origin_, rot_mat_Xs_)
get_quat = lambda r_mat: transformations.quaternion_from_matrix_py(r_mat)
sensor_quats_in_R_ = torch.from_numpy(np.stack(list(map(get_quat, rot_mat_Rs_.cpu().numpy())))).to(sensor_cam_locs_.device).float()
pred_features_ = self.context_net(context_grid_list_,\
sensor_cam_locs_in_R_, sensor_quats_in_R_)
# normalize
pred_features_ = l2_normalize(pred_features_, dim=1)
pred_features = __u(pred_features_)
# if doing moco I have to pass the inputs through the key(slow) network as well #
if hyp.do_moc or hyp.do_eval_recall:
key_context_grid_list_ = copy.deepcopy(context_grid_list_)
key_sensor_cam_locs_in_R_ = copy.deepcopy(sensor_cam_locs_in_R_)
key_sensor_quats_in_R_ = copy.deepcopy(sensor_quats_in_R_)
self.key_context_net.eval()
with torch.no_grad():
key_pred_features_ = self.key_context_net(key_context_grid_list_,\
key_sensor_cam_locs_in_R_, key_sensor_quats_in_R_)
# normalize, normalization is very important why though
key_pred_features_ = l2_normalize(key_pred_features_, dim=1)
key_pred_features = __u(key_pred_features_)
# end passing of the input through the slow network this is necessary for moco #
## ... End code for passing the context grid through 3D CNN to obtain a vector ... ##
## ... Begin code for doing metric learning between pred_features and sensor features ... ##
# 1. Subsample both based on the number of positive samples
if hyp.do_touch_embML:
assert(hyp.do_touch_forward)
assert(hyp.do_touch_feat)
perm = torch.randperm(len(pred_features_)) ## 1024
chosen_sensor_feats_ = sensor_features_[perm[:self.num_pos_samples*hyp.B]]
chosen_pred_feats_ = pred_features_[perm[:self.num_pos_samples*B]]
# 2. form the emb_vec and get pos and negative samples for the batch
emb_vec = torch.stack((chosen_sensor_feats_, chosen_pred_feats_), dim=1).view(hyp.B*self.num_pos_samples*self.batch_k, -1)
y = torch.stack([torch.range(0, self.num_pos_samples*B-1), torch.range(0, self.num_pos_samples*B-1)],\
dim=1).view(B*self.num_pos_samples*self.batch_k) # (0, 0, 1, 1, ..., 255, 255)
a_indices, anchors, positives, negatives, _ = self.sampler(emb_vec)
# 3. Compute the loss, ML loss and the l2 distance betwee the embeddings
margin_loss, _ = self.criterion(anchors, positives, negatives, self.beta, y[a_indices])
total_loss = utils_misc.add_loss('embtouch/emb_touch_ml_loss', total_loss, margin_loss,
hyp.emb_3D_ml_coeff, summ_writer)
# the l2 loss between the embeddings
l2_loss = torch.nn.functional.mse_loss(chosen_sensor_feats_, chosen_pred_feats_)
total_loss = utils_misc.add_loss('embtouch/emb_l2_loss', total_loss, l2_loss,
hyp.emb_3D_l2_coeff, summ_writer)
## ... End code for doing metric learning between pred_features and sensor_features ... ##
## ... Begin code for doing moc inspired ML between pred_features and sensor_features ... ##
if hyp.do_moc and moc_init_done:
moc_loss = self.moc_ml_net(sensor_features_, key_sensor_features_,\
pred_features_, key_pred_features_, summ_writer)
total_loss += moc_loss
## ... End code for doing moc inspired ML between pred_features and sensor_feature ... ##
## ... add code for filling up results needed for eval recall ... ##
if hyp.do_eval_recall and moc_init_done:
results['context_features'] = pred_features_
results['sensor_depth_centers_in_world'] = sensor_depths_centers_in_world_
results['sensor_depths_centers_in_ref_cam'] = sensor_depths_centers_in_ref_cam_
results['object_name'] = feed['object_name']
# I will do precision recall here at different recall values and summarize it using tensorboard
recalls = [1, 5, 10, 50, 100, 200]
# also should not include any gradients because of this
# fast_sensor_emb_e = sensor_features_
# fast_context_emb_e = pred_features_
# slow_sensor_emb_g = key_sensor_features_
# slow_context_emb_g = key_context_features_
fast_sensor_emb_e = sensor_features_.clone().detach()
fast_context_emb_e = pred_features_.clone().detach()
# I will do multiple eval recalls here
slow_sensor_emb_g = key_sensor_features_.clone().detach()
slow_context_emb_g = key_pred_features_.clone().detach()
# assuming the above thing goes well
fast_sensor_emb_e = fast_sensor_emb_e.cpu().numpy()
fast_context_emb_e = fast_context_emb_e.cpu().numpy()
slow_sensor_emb_g = slow_sensor_emb_g.cpu().numpy()
slow_context_emb_g = slow_context_emb_g.cpu().numpy()
# now also move the vis to numpy and plot it using matplotlib
vis_e = __p(sensor_rgbs)
vis_g = __p(sensor_rgbs)
np_vis_e = vis_e.cpu().detach().numpy()
np_vis_e = np.transpose(np_vis_e, [0, 2, 3, 1])
np_vis_g = vis_g.cpu().detach().numpy()
np_vis_g = np.transpose(np_vis_g, [0, 2, 3, 1])
# bring it back to original color
np_vis_g = ((np_vis_g+0.5) * 255).astype(np.uint8)
np_vis_e = ((np_vis_e+0.5) * 255).astype(np.uint8)
# now compare fast_sensor_emb_e with slow_context_emb_g
# since I am doing positive against this
fast_sensor_emb_e_list = [fast_sensor_emb_e, np_vis_e]
slow_context_emb_g_list = [slow_context_emb_g, np_vis_g]
prec, vis, chosen_inds_and_neighbors_inds = compute_precision(
fast_sensor_emb_e_list, slow_context_emb_g_list, recalls=recalls
)
# finally plot the nearest neighbour retrieval and move ahead
if feed['global_step'] % 1 == 0:
plot_nearest_neighbours(vis, step=feed['global_step'],
save_dir='/home/gauravp/eval_results',
name='fast_sensor_slow_context')
# plot the precisions at different recalls
for pr, re in enumerate(recalls):
summ_writer.summ_scalar(f'evrefast_sensor_slow_context/recall@{re}',\
prec[pr])
# now compare fast_context_emb_e with slow_sensor_emb_g
fast_context_emb_e_list = [fast_context_emb_e, np_vis_e]
slow_sensor_emb_g_list = [slow_sensor_emb_g, np_vis_g]
prec, vis, chosen_inds_and_neighbors_inds = compute_precision(
fast_context_emb_e_list, slow_sensor_emb_g_list, recalls=recalls
)
if feed['global_step'] % 1 == 0:
plot_nearest_neighbours(vis, step=feed['global_step'],
save_dir='/home/gauravp/eval_results',
name='fast_context_slow_sensor')
# plot the precisions at different recalls
for pr, re in enumerate(recalls):
summ_writer.summ_scalar(f'evrefast_context_slow_sensor/recall@{re}',\
prec[pr])
# now finally compare both the fast, I presume we want them to go closer too
fast_sensor_list = [fast_sensor_emb_e, np_vis_e]
fast_context_list = [fast_context_emb_e, np_vis_g]
prec, vis, chosen_inds_and_neighbors_inds = compute_precision(
fast_sensor_list, fast_context_list, recalls=recalls
)
if feed['global_step'] % 1 == 0:
plot_nearest_neighbours(vis, step=feed['global_step'],
save_dir='/home/gauravp/eval_results',
name='fast_sensor_fast_context')
for pr, re in enumerate(recalls):
summ_writer.summ_scalar(f'evrefast_sensor_fast_context/recall@{re}',\
prec[pr])
## ... done code for filling up results needed for eval recall ... ##
summ_writer.summ_scalar('loss', total_loss.cpu().item())
return total_loss, results, [key_sensor_features_, key_pred_features_]
class MujocoOffline(nn.Module):
def __init__(self, config):
super(MujocoOffline, self).__init__()
self.config = config
if self.config.do_feat:
print('------- adding featnet --------')
self.featnet = FeatNet(self.config)
if self.config.do_occ:
print('------- adding occnet ---------')
self.occnet = OccNet(self.config)
if self.config.do_view:
print('------- adding viewnet --------')
self.viewnet = ViewNet(self.config)
if self.config.do_det:
print('------- adding detnet ---------')
self.detnet = DetNet(self.config)
# coordinate range
self.coord_cam_front = Coord(-0.5, 0.5, -0.5, 0.5, 0.2, 1.2, 0.0, -0.4)
self.coord_mem = Coord(-0.5, 0.5, -0.5, 0.5, -0.5, 0.5, 0.0, -0.4)
MH, MW, MD = self.config.Y, self.config.X, self.config.Z
MH2, MW2, MD2 = int(MH / 2), int(MW / 2), int(MD / 2)
# voxel size
mem_protos = VoxProto([MH, MW, MD])
halfmem_protos = VoxProto([MH2, MW2, MD2])
#combine
self.mem_coord_cams = VoxCoord(self.coord_cam_front, mem_protos)
self.mem_coord_Rs = VoxCoord(self.coord_mem, mem_protos)
self.halfmem_coord_cams = VoxCoord(self.coord_cam_front,
halfmem_protos)
self.halfmem_coord_Rs = VoxCoord(self.coord_mem, halfmem_protos)
self.feat_mem_coord_cams = None #self.halfmem_coord_Rs
self.feat_mem_coord_Rs = None
self.is_learned_cluster_centers = False
def save_local_variables(self):
output = dict()
return output
def unproject(self, cam_rgbd_inputs, cam_info_inputs):
rgb_camXs, xyz_camXs = cam_rgbd_inputs
pix_T_cams, origin_T_camXs, origin_T_camRs = cam_info_inputs
B, H, W, V, S = self.config.B, self.config.H, self.config.W, self.config.V, self.config.S
PH, PW = self.config.PH, self.config.PW # this is the size of the predicted image
# the next are the memory dimensions, do not know why this naming though
# merge sequence and batch dimensions
__p = lambda x: utils.basic.pack_seqdim(x, B)
# unmerge sequence and batch dimensions
__u = lambda x: utils.basic.unpack_seqdim(x, B)
pix_T_cams_ = __p(pix_T_cams) # packing here the (B,S) => (B*S)
# intrinsic matrix packed and unpacked end
origin_T_camRs_ = __p(origin_T_camRs)
origin_T_camXs_ = __p(origin_T_camXs)
# origin_T_camXs unpacked and packed end
# completed getting inputs now combining them
# 1. Converts from camX to camR which is Adam's coordinate system
# get from camX_T_camR and camR_T_camX and pack unpack it
camRs_T_camXs_ = torch.matmul(utils.geom.safe_inverse(origin_T_camRs_),
origin_T_camXs_)
camXs_T_camRs_ = utils.geom.safe_inverse(camRs_T_camXs_)
camRs_T_camXs = __u(camRs_T_camXs_)
camXs_T_camRs = __u(camXs_T_camRs_)
# end of camX_T_camR and camR_T_camX and pack unpack it
# goes directly from camR to image in each camera image frame
pix_T_cams_ = utils.geom.pack_intrinsics(pix_T_cams_[:, 0, 0],
pix_T_cams_[:, 1, 1],
pix_T_cams_[:, 0, 2],
pix_T_cams_[:, 1, 2])
pix_T_camRs_ = torch.matmul(pix_T_cams_, camXs_T_camRs_)
pix_T_camRs = __u(pix_T_camRs_)
# end of computation for matrix which goes from camR to each camera image frame
# pointclouds in each camera frame
xyz_camXs_ = __p(xyz_camXs)
# pointclouds converted to camR coordinate system
xyz_camRs_ = utils.geom.apply_4x4(camRs_T_camXs_, xyz_camXs_)
xyz_camRs = __u(xyz_camRs_)
# TODO: visualize the point cloud here and check that it makes sense
# get occupancy maps from pointclouds
# QUESTION: what is the space you are discretizing, I mean the extent of the space
occRs_ = utils.vox.voxelize_xyz(xyz_camRs_, self.mem_coord_Rs)
occXs_ = utils.vox.voxelize_xyz(xyz_camXs_, self.mem_coord_cams)
occRs_half_ = utils.vox.voxelize_xyz(xyz_camRs_, self.halfmem_coord_Rs)
occXs_half_ = utils.vox.voxelize_xyz(xyz_camXs_,
self.halfmem_coord_cams)
occRs = __u(occRs_)
occXs = __u(occXs_)
occRs_half = __u(occRs_half_)
occXs_half = __u(occXs_half_)
# unproject depth images, This is done for the color images not the depths
## rgb unprojection, bilinearly samples and fills the grid
my_device = rgb_camXs.device
unpRs_ = utils.vox.unproject_rgb_to_mem(__p(rgb_camXs),
pix_T_camRs_,
self.mem_coord_Rs,
device=my_device)
unpXs_ = utils.vox.unproject_rgb_to_mem(__p(rgb_camXs),
pix_T_cams_,
self.mem_coord_cams,
device=my_device)
unpRs = __u(unpRs_)
unpXs = __u(unpXs_)
unpRs_half_ = utils.vox.unproject_rgb_to_mem(__p(rgb_camXs),
pix_T_camRs_,
self.halfmem_coord_Rs,
device=my_device)
unpRs_half = __u(unpRs_half_)
unp_visRs = utils.improc.get_unps_vis(unpRs_half, occRs_half)
unp_visRs = torch.mean(unp_visRs, dim=1)
# NOTE: still do not know why is this required or where is this used for that matter
depth_camXs_, valid_camXs_ = utils.geom.create_depth_image(
pix_T_cams_, xyz_camXs_, H, W)
dense_xyz_camXs_ = utils.geom.depth2pointcloud(depth_camXs_,
pix_T_cams_)
dense_xyz_camRs_ = utils.geom.apply_4x4(camRs_T_camXs_,
dense_xyz_camXs_)
# this is B*S x H*W x 3
inbound_camXs_ = utils.vox.get_inbounds(dense_xyz_camRs_,
self.mem_coord_cams).float()
inbound_camXs_ = torch.reshape(
inbound_camXs_,
[B * S, 1, H, W
]) # NOTE: Here there is a difference in tensorflow code
inbound_camXs = __u(inbound_camXs_)
depth_camXs = __u(depth_camXs_)
valid_camXs = __u(valid_camXs_) * inbound_camXs
return depth_camXs, valid_camXs, camRs_T_camXs, camXs_T_camRs, unpXs, unpRs, occXs, occRs, occXs_half, occRs_half, unp_visRs
def predict_forward(self, feed):
cam_rgbd_inputs = (feed["rgb_camXs"], feed["xyz_camXs"])
cam_info_inputs = (feed["pix_T_cams"], feed["origin_T_camXs"],
feed["origin_T_camRs"])
depth_camXs, valid_camXs, camRs_T_camXs, camXs_T_camRs, unpXs, unpRs, occXs, occRs, occXs_half, occRs_half = self.unproject(
cam_rgbd_inputs, cam_info_inputs)
B = self.config.B
__p = lambda x: utils.basic.pack_seqdim(x, B)
# unmerge sequence and batch dimensions
__u = lambda x: utils.basic.unpack_seqdim(x, B)
if self.config.do_feat:
rgb_camXs, xyz_camXs = cam_rgbd_inputs
featXs_input = torch.cat([occXs, occXs * unpXs],
dim=2) # B, S, 4, H, W, D
featXs_input_ = __p(featXs_input)
freeXs_ = utils.vox.get_freespace(__p(xyz_camXs), __p(occXs_half),
self.halfmem_coord_cams)
freeXs = __u(freeXs_)
visXs = torch.clamp(occXs_half + freeXs, 0.0, 1.0)
#if type(mask_) != type(None): # featXs_input: B x NVIEWS x 4 x 64 x 64 x 64
assert (list(occXs.shape)[3:6] == list(featXs_input.shape)[3:6])
featXs_, validXs_, feat_loss = self.featnet(
featXs_input_, mask=__p(occXs), set_num=feed['set_num'])
assert feat_loss.item(
) == 0.0, "there is nothing to guide featnet by itself"
# for each view features are being predicted, NOTE that nothing is brought into common view yet
validXs, featXs = __u(validXs_), __u(featXs_)
#### .... BEGIN Converting everything to ref frame .... ####
validRs = utils.vox.apply_4x4_to_voxs(
camRs_T_camXs,
validXs,
mem_coord_As=self.halfmem_coord_cams,
mem_coord_Bs=self.halfmem_coord_Rs)
visRs = utils.vox.apply_4x4_to_voxs(
camRs_T_camXs,
visXs,
mem_coord_As=self.halfmem_coord_cams,
mem_coord_Bs=self.halfmem_coord_Rs)
featRs = utils.vox.apply_4x4_to_voxs(
camRs_T_camXs,
featXs,
mem_coord_As=self.halfmem_coord_cams,
mem_coord_Bs=self.halfmem_coord_Rs)
if self.feat_mem_coord_Rs == None:
self.feat_mem_coord_Rs = self.halfmem_coord_Rs
#### .... featRs_without_target_view contains features from all the views
#### .... warped and brought into common frame and aggregated .... Using
#### .... features occupancy and target view should be predicted .... ####
# B x 32 x H x W x D
featsRs_without_target_view = torch.mean(featRs[:, 1:], dim=1)
if self.config.do_view:
assert self.config.do_feat
PH, PW = self.config.PH, self.config.PW
sy = float(PH) / float(self.config.H)
sx = float(PW) / float(self.config.W)
assert (sx == 0.5)
assert (sy == 0.5)
# projpix_T_cams, are the intrinsics for the projection, just scale the true intrinsics
pix_T_cams = feed["pix_T_cams"]
projpix_T_cams = __u(
utils.geom.scale_intrinsics(__p(pix_T_cams), sx, sy))
# now I will project the predicted feats to target view (warp)
feat_projtarget_view = utils.vox.apply_pixX_T_memR_to_voxR(
projpix_T_cams[:, 0], camXs_T_camRs[:,
0], self.halfmem_coord_Rs,
featsRs_without_target_view, self.config.view_depth, PH, PW)
rgb_X0 = utils.basic.downsample(
rgb_camXs[:, 0], 2) ## NOTE: this is the ground truth
# rgb_e: b x 3 x 64 x 64
view_loss, rgb_e, emb2D_e = self.viewnet(feat_projtarget_view,
rgb_X0,
set_num=feed['set_num'])
#crop object features
bbox_in_ref_cam = feed['bbox_in_ref_cam']
# based on the batch size this would be B, N, 8, 3
min_bounds = bbox_in_ref_cam[:, :, 0, :]
max_bounds = bbox_in_ref_cam[:, :, -1, :]
lengths = torch.abs(max_bounds - min_bounds)
center = (max_bounds + min_bounds) * 0.5
# now form the box and then covert to lrt list
B = self.config.B # since i have only one box
N = 1 # number of objects
# 9 is cx, cy, cz, lx, ly, lz, rx, ry, rz
boxlist = torch.zeros(B, N, 9)
# NOTE: Note: I am assuming here that N = 1 !!!!!!
boxlist[:, :, :3] = center #.unsqueeze(1)
boxlist[:, :, 3:6] = lengths #.unsqueeze(1)
# convert it to lrt list, it contains box length and rt to go
# from box coordinates to ref coordinate system.
box_lrtlist = utils.geom.convert_boxlist_to_lrtlist(boxlist)
# now this is already in the ref coordinate system which was not
# the case with my previous use of the crop_zoom_from_mem func.
# Hence I had previously included camR_T_camXs which is not req here
_, _, box_dim = box_lrtlist.shape
presumably_object_tensor = utils.vox.crop_zoom_from_mem(
featsRs_without_target_view, self.feat_mem_coord_Rs,
torch.reshape(box_lrtlist[:, :, :], [B * N, box_dim]), 32, 32, 32)
_, C, H, W, D = presumably_object_tensor.shape
presumably_object_tensor = torch.reshape(
presumably_object_tensor.permute([0, 2, 3, 4, 1]),
[B, N, H, W, D, C])
# NOTE: As of now I am not doing backprop through this Tensor so
# no need to keep it in gpu anymore
results = dict()
results['object_tensor'] = presumably_object_tensor.cpu().detach(
).numpy()
results[
'featsRs_without_target_view'] = featsRs_without_target_view.permute(
[0, 2, 3, 4, 1]).cpu().detach().numpy()
results['rgb_e'] = rgb_e.permute(0, 2, 3, 1).cpu().detach().numpy()
# Add the plot of this to tensorboard, and also think how can you
# visualize if the correct thing is being returned to you.
return results
def predict_forward_bbox_detector(self, feed):
# here I assume that this function will be called only during inference, so all the sequences should be used
results = dict()
cam_rgbd_inputs = (feed['rgb_camXs'], feed['xyz_camXs'])
cam_info_inputs = (feed['pix_T_cams'], feed['origin_T_camXs'],
feed['origin_T_camRs'])
B = self.config.B
__p = lambda x: utils.basic.pack_seqdim(x, B)
__u = lambda x: utils.basic.unpack_seqdim(x, B)
# if self.config.do_det:
# # NOTE: this does not contain a valid box
# gt_boxes_corners = feed['bbox_in_ref_cam'] ## fill in some random values
# gt_boxesRMem_corners = utils.vox.Ref2Mem(gt_boxes_corners, self.halfmem_coord_Rs)
# gt_boxesRMem_corners = gt_boxesRMem_corners.unsqueeze(1) # this indicates the number of boxes
#
# # I have corners in mem now, I will convert it to boxlist
# gt_boxesRMem_theta = utils.geom.convert_corners_to_axis_aligned_boxlist(gt_boxesRMem_corners)
# scores = torch.ones(self.config.B, 1).float().to(gt_boxesRMem_theta.device)
depth_camXs, valid_camXs, camRs_T_camXs, camXs_T_camRs, unpXs, unpRs, occXs, occRs, occXs_half, occRs_half, unp_visRs = self.unproject(
cam_rgbd_inputs, cam_info_inputs)
# put the model in eval mode here
self.featnet.eval()
assert self.featnet.training == False, "should have batch norm switched off"
self.detnet.eval()
assert self.detnet.training == False, "should have batch norm switched off here"
if self.config.do_feat:
rgb_camXs, xyz_camXs = cam_rgbd_inputs
featXs_input = torch.cat([occXs, occXs * unpXs],
dim=2) # B, S, 4, H, W, D
featXs_input_ = __p(featXs_input)
freeXs_ = utils.vox.get_freespace(__p(xyz_camXs), __p(occXs_half),
self.halfmem_coord_cams)
freeXs = __u(freeXs_)
visXs = torch.clamp(occXs_half + freeXs, 0.0, 1.0)
#if type(mask_) != type(None): # featXs_input: B x NVIEWS x 4 x 64 x 64 x 64
assert (list(occXs.shape)[3:6] == list(featXs_input.shape)[3:6])
with torch.no_grad():
assert self.featnet.training == False
featXs_, validXs_, feat_loss = self.featnet(
featXs_input_, mask=__p(occXs), set_num=feed['set_num'])
validXs, featXs = __u(validXs_), __u(featXs_)
validRs = utils.vox.apply_4x4_to_voxs(
camRs_T_camXs,
validXs,
mem_coord_As=self.halfmem_coord_cams,
mem_coord_Bs=self.halfmem_coord_Rs)
visRs = utils.vox.apply_4x4_to_voxs(
camRs_T_camXs,
visXs,
mem_coord_As=self.halfmem_coord_cams,
mem_coord_Bs=self.halfmem_coord_Rs)
featRs = utils.vox.apply_4x4_to_voxs(
camRs_T_camXs,
featXs,
mem_coord_As=self.halfmem_coord_cams,
mem_coord_Bs=self.halfmem_coord_Rs)
if self.feat_mem_coord_Rs == None:
self.feat_mem_coord_Rs = self.halfmem_coord_Rs
# since only using for prediction all the views can be used? think about this
features = torch.mean(featRs,
dim=1) # across the sequence dimension
if self.config.do_det:
self.axboxlist_memR = gt_boxesRMem_theta.clone()
self.scorelist = scores
# TODO: some parts of this network are confusing to me, understand it thoroughly
with torch.no_grad():
assert self.detnet.training == False
detect_loss, boxlist_memR_e, scorelist_e, tidlist_e, sco, ove = self.detnet(
self.axboxlist_memR,
self.scorelist,
features,
summ_writer=None)
# now that I have the box list I need to convert it into corners and return
# convert it to ref image and return, mem to adam to ref cam
# adam_T_ref = feed['adam_T_camRs'].squeeze(1)
# ref_T_adam = torch.inverse(adam_T_ref)
pred_box_corners_mem = utils.geom.transform_boxes_to_corners(
boxlist_memR_e)
pred_box_corners_adam = utils.geom.apply_4x4_to_corners(
utils.coordTcoord.get_ref_T_mem(B, self.halfmem_coord_Rs),
pred_box_corners_mem)
#pred_box_corners_adam = utils.vox.Mem2Ref(pred_box_corners_mem, self.halfmem_coord_Rs)
# pred_box_ref = utils.geom.apply_4x4_to_corners(ref_T_adam, pred_box_corners_adam)
results['predicted_boxes_adam'] = pred_box_corners_adam
results['scorelist'] = scorelist_e
results['tidlist'] = tidlist_e
return results
def convert_objects_to_features(self, feed):
results = self.predict_forward(feed)
return results['object_tensor']
def dump_one_batch(self, feed):
import pickle
import copy
feed_copy = dict()
i = 0
for key in feed:
if key in ['record', 'writer', 'global_step']:
continue
if torch.is_tensor(feed[key]):
tensor_np = feed[key].cpu()
feed_copy[key] = tensor_np
else:
feed_copy[key] = feed[key]
i += 1
#if i > 1:
# break
with open("tmp/feed.pkl", "wb") as f:
pickle.dump(feed_copy, f)
import ipdb
ipdb.set_trace()
def forward(self, feed):
# feed is the input here, let's see what it has
results = dict()
#self.dump_one_batch(feed)
# Whenever forward is called, this is instantiated which creates summ_writer object
# save this is True if global_step % log_freq == 0
summ_writer = utils.improc.Summ_writer(config=self.config,
writer=feed['writer'],
global_step=feed['global_step'],
set_name=feed['set_name'],
fps=8)
writer = feed['writer']
#global_step = feed['global_step']
total_loss = torch.tensor(0.0).cuda()
cam_rgbd_inputs = (feed["rgb_camXs"], feed["xyz_camXs"])
cam_info_inputs = (feed["pix_T_cams"], feed["origin_T_camXs"],
feed["origin_T_camRs"])
depth_camXs, valid_camXs, camRs_T_camXs, camXs_T_camRs, unpXs, unpRs, occXs, occRs, occXs_half, occRs_half, unp_visRs = self.unproject(
cam_rgbd_inputs, cam_info_inputs)
B = self.config.B
__p = lambda x: utils.basic.pack_seqdim(x, B)
# unmerge sequence and batch dimensions
__u = lambda x: utils.basic.unpack_seqdim(x, B)
#### ... VISUALIZE what we got ... ####
# prepare stuff for the detector
if self.config.do_det:
gt_boxesR_corners = feed['bbox_in_ref_cam']
gt_boxesR_corners_ = __p(gt_boxesR_corners)
# convert the corners into memory coordinates
gt_boxesRMem_corners_ = utils.vox.Ref2Mem(gt_boxesR_corners_,
self.halfmem_coord_Rs)
gt_boxesRMem_corners = __u(gt_boxesRMem_corners_) #.unsqueeze(1)
gt_boxesRMem_theta = utils.geom.convert_corners_to_axis_aligned_boxlist(
gt_boxesRMem_corners)
# finally get the scores, I am assuming 1 since all boxes are visible for me and all boxes for me is 1
scores = torch.ones(B, 1).float().to(gt_boxesRMem_theta.device)
if not feed['set_num'] == 1:
rgb_camXs, xyz_camXs = cam_rgbd_inputs
rgb_camRs = feed["rgb_camRs"]
summ_writer.summ_oneds('2D_inputs/depth_camXs',
torch.unbind(depth_camXs, dim=1))
summ_writer.summ_oneds('2D_inputs/valid_camXs',
torch.unbind(valid_camXs, dim=1))
summ_writer.summ_rgbs('2D_inputs/rgb_camXs',
torch.unbind(rgb_camXs, dim=1))
summ_writer.summ_rgbs('2D_inputs/rgb_camRs',
torch.unbind(rgb_camRs, dim=1))
summ_writer.summ_occs('3d_inputs/occXs',
torch.unbind(occXs, dim=1),
reduce_axes=[2])
summ_writer.summ_unps('3d_inputs/unpXs', torch.unbind(unpXs,
dim=1),
torch.unbind(occXs, dim=1))
if summ_writer.save_this:
# why compute again?
#unpRs_ = utils.vox.unproject_rgb_to_mem(__p(rgb_camXs), utils.basic.matmul2(pix_T_cams_, camXs_T_camRs_), self.mem_coord_Rs)
#unpRs = __u(unpRs_)
#occRs_ = utils.vox.voxelize_xyz(xyz_camRs_, self.mem_coord_Rs)
summ_writer.summ_occs('3d_inputs/occRs',
torch.unbind(occRs, dim=1),
reduce_axes=[2])
summ_writer.summ_unps('3d_inputs/unpRs',
torch.unbind(unpRs, dim=1),
torch.unbind(occRs, dim=1))
else:
rgb_camXs, xyz_camXs = cam_rgbd_inputs
rgb_camRs = feed["rgb_camRs"]
summ_writer.summ_oneds('2D_inputs_val/depth_camXs',
torch.unbind(depth_camXs, dim=1))
summ_writer.summ_oneds('2D_inputs_val/valid_camXs',
torch.unbind(valid_camXs, dim=1))
summ_writer.summ_rgbs('2D_inputs_val/rgb_camXs',
torch.unbind(rgb_camXs, dim=1))
summ_writer.summ_rgbs('2D_inputs_val/rgb_camRs',
torch.unbind(rgb_camRs, dim=1))
summ_writer.summ_occs('3d_inputs_val/occXs',
torch.unbind(occXs, dim=1),
reduce_axes=[2])
summ_writer.summ_unps('3d_inputs_val/unpXs',
torch.unbind(unpXs, dim=1),
torch.unbind(occXs, dim=1))
if summ_writer.save_this:
#unpRs_ = utils.vox.unproject_rgb_to_mem(__p(rgb_camXs), Z, Y, X, utils.basic.matmul2(pix_T_cams_, camXs_T_camRs_))
#unpRs = __u(unpRs_)
#occRs_ = utils.vox.voxelize_xyz(xyz_camRs_, Z, Y, X)
summ_writer.summ_occs('3d_inputs_val/occRs',
torch.unbind(occRs, dim=1),
reduce_axes=[2])
summ_writer.summ_unps('3d_inputs_val/unpRs',
torch.unbind(unpRs, dim=1),
torch.unbind(occRs, dim=1))
# the idea behind view-pred is form memory with the remaining views project it to target view and
# then use this memory to predict the target image
# idea behind occ_prediction is use the memory to predict occupancy in ref view and compare it
# with the ground truth occupancy in the ref view
if self.config.do_feat:
rgb_camXs, xyz_camXs = cam_rgbd_inputs
featXs_input = torch.cat([occXs, occXs * unpXs],
dim=2) # B, S, 4, H, W, D
featXs_input_ = __p(featXs_input)
freeXs_ = utils.vox.get_freespace(__p(xyz_camXs), __p(occXs_half),
self.halfmem_coord_cams)
freeXs = __u(freeXs_)
visXs = torch.clamp(occXs_half + freeXs, 0.0, 1.0)
#if type(mask_) != type(None): # featXs_input: B x NVIEWS x 4 x 64 x 64 x 64
assert (list(occXs.shape)[3:6] == list(featXs_input.shape)[3:6])
featXs_, validXs_, feat_loss = self.featnet(
featXs_input_,
summ_writer,
mask=__p(occXs),
set_num=feed['set_num'])
total_loss += feat_loss
assert feat_loss.item(
) == 0.0, "there is nothing to guide featnet by itself"
# for each view features are being predicted, NOTE that nothing is brought into common view yet
validXs, featXs = __u(validXs_), __u(featXs_)
#### .... BEGIN Converting everything to ref frame .... ####
validRs = utils.vox.apply_4x4_to_voxs(
camRs_T_camXs,
validXs,
mem_coord_As=self.halfmem_coord_cams,
mem_coord_Bs=self.halfmem_coord_Rs)
visRs = utils.vox.apply_4x4_to_voxs(
camRs_T_camXs,
visXs,
mem_coord_As=self.halfmem_coord_cams,
mem_coord_Bs=self.halfmem_coord_Rs)
featRs = utils.vox.apply_4x4_to_voxs(
camRs_T_camXs,
featXs,
mem_coord_As=self.halfmem_coord_cams,
mem_coord_Bs=self.halfmem_coord_Rs)
if self.feat_mem_coord_Rs == None:
self.feat_mem_coord_Rs = self.halfmem_coord_Rs
#### .... END converting everything to ref frame .... ####
### ... Remember _e added at the end means it is estimated ... ###
vis3D_e = torch.max(validRs[:, 1:], dim=1)[0] * torch.max(
visRs[:, 1:], dim=1)[0]
### ... only thing which is using _e is below visualization ... ###
if not feed['set_num'] == 1:
summ_writer.summ_feats('3D_feats/featXs_input',
torch.unbind(featXs_input, dim=1),
pca=True)
summ_writer.summ_feats('3D_feats/featXs_output',
torch.unbind(featXs, dim=1),
pca=True)
summ_writer.summ_feats('3D_feats/featRs_output',
torch.unbind(featRs, dim=1),
pca=True)
summ_writer.summ_feats('3D_feats/validRs',
torch.unbind(validRs, dim=1),
pca=False)
summ_writer.summ_feat('3D_feats/vis3D_e', vis3D_e, pca=False)
else:
summ_writer.summ_feats('3D_feats_val/featXs_input',
torch.unbind(featXs_input, dim=1),
pca=True)
summ_writer.summ_feats('3D_feats_val/featXs_output',
torch.unbind(featXs, dim=1),
pca=True)
summ_writer.summ_feats('3D_feats_val/featRs_output',
torch.unbind(featRs, dim=1),
pca=True)
summ_writer.summ_feats('3D_feats_val/validRs',
torch.unbind(validRs, dim=1),
pca=False)
summ_writer.summ_feat('3D_feats_val/vis3D_e',
vis3D_e,
pca=False)
#### .... featRs_without_target_view contains features from all the views
#### .... warped and brought into common frame and aggregated .... Using
#### .... features occupancy and target view should be predicted .... ####
featsRs_without_target_view = torch.mean(featRs[:, 1:], dim=1)
if self.config.do_generate_data or (self.config.do_validation
and feed['set_num'] == 1):
featRs_with_target_view = torch.mean(featRs, dim=1)
if self.config.do_occ and self.config.occ_do_cheap:
occRs_sup, freeRs_sup, freeXs = utils.vox.prep_occs_supervision(
xyz_camXs,
occRs_half,
occXs_half,
camRs_T_camXs,
self.halfmem_coord_Rs,
self.halfmem_coord_cams,
agg=True)
if feed['set_num'] != 1:
summ_writer.summ_occ('occ_sup/occ_sup',
occRs_sup,
reduce_axes=[2])
summ_writer.summ_occ('occ_sup/free_sup',
freeRs_sup,
reduce_axes=[2])
summ_writer.summ_occs('occ_sup/freeXs_sup',
torch.unbind(freeXs, dim=1),
reduce_axes=[2])
summ_writer.summ_occs('occ_sup/occXs_sup',
torch.unbind(occXs_half, dim=1),
reduce_axes=[2])
else:
summ_writer.summ_occ('occ_sup_val/occ_sup',
occRs_sup,
reduce_axes=[2])
summ_writer.summ_occ('occ_sup_val/free_sup',
freeRs_sup,
reduce_axes=[2])
summ_writer.summ_occs('occ_sup_val/freeXs_sup',
torch.unbind(freeXs, dim=1),
reduce_axes=[2])
summ_writer.summ_occs('occ_sup_val/occXs_sup',
torch.unbind(occXs_half, dim=1),
reduce_axes=[2])
occ_loss, occRs_pred_ = self.occnet(featsRs_without_target_view,
occRs_sup,
freeRs_sup,
torch.max(validRs[:, 1:],
dim=1)[0],
summ_writer,
set_num=feed['set_num'])
occRs_pred = __u(occRs_pred_)
total_loss += occ_loss
if self.config.do_view:
assert self.config.do_feat
# we warped the features into canonical view which is featR
# now we resample to target view which is view (0) and decode
# be sure not to pass in the features of the view to decode
# use featRs_without_target_view as the features in the canonical view
PH, PW = self.config.PH, self.config.PW
sy = float(PH) / float(self.config.H)
sx = float(PW) / float(self.config.W)
assert (sx == 0.5)
assert (sy == 0.5)
# projpix_T_cams, are the intrinsics for the projection, just scale the true intrinsics
pix_T_cams = feed["pix_T_cams"]
projpix_T_cams = __u(
utils.geom.scale_intrinsics(__p(pix_T_cams), sx, sy))
# now I will project the predicted feats to target view (warp)
feat_projtarget_view = utils.vox.apply_pixX_T_memR_to_voxR(
projpix_T_cams[:, 0], camXs_T_camRs[:,
0], self.halfmem_coord_Rs,
featsRs_without_target_view, self.config.view_depth, PH, PW)
rgb_X0 = utils.basic.downsample(
rgb_camXs[:, 0], 2) ## NOTE: this is the ground truth
view_loss, rgb_e, emb2D_e = self.viewnet(feat_projtarget_view,
rgb_X0,
summ_writer,
set_num=feed['set_num'])
total_loss += view_loss
if self.config.do_det:
emb3D_e_R = featsRs_without_target_view
emb3D_g_R = featRs[:,
0, :, :, :, :] # 0 serves as the target view, which want to predict from the context
self.axboxlist_memR = gt_boxesRMem_theta.clone()
self.scorelist = scores
# TODO: some parts of this network are confusing to me, understand it thoroughly
detect_loss, boxlist_memR_e, scorelist_e, tidlist_e, sco, ove = self.detnet(
self.axboxlist_memR, self.scorelist, emb3D_e_R, summ_writer)
total_loss += detect_loss
# unprojecting again from memory to ref cam coordinates
boxlist_camR_e = utils.vox.convert_boxlist_memR_to_camR(
boxlist_memR_e, self.halfmem_coord_Rs)
boxlist_camR_g = utils.vox.convert_boxlist_memR_to_camR(
self.axboxlist_memR, self.halfmem_coord_Rs)
# from gt_corners see that boxlist_camR_g here is fine
corners_max_g = torch.max(gt_boxesR_corners_, axis=1)[0]
corners_min_g = torch.min(gt_boxesR_corners_, axis=1)[0]
comp_center = corners_min_g + (corners_max_g - corners_min_g) / 2.0
comp_lengths = corners_max_g - corners_min_g
from_above_camR = boxlist_camR_g.clone()
assert np.allclose(
from_above_camR[:, :, :3].squeeze(1).cpu().numpy(),
comp_center.cpu().numpy(),
atol=1e-5)
assert np.allclose(from_above_camR[:, :,
3:6].squeeze(1).cpu().numpy(),
comp_lengths.cpu().numpy(),
atol=1e-5)
summ_writer.summ_box_mem_on_mem(
'detnet/gt_boxesR_mem', unp_visRs, self.axboxlist_memR,
self.scorelist,
torch.ones([self.config.B, 1], dtype=torch.int32))
try:
summ_writer.summ_box_mem_on_mem(
'detnet/pred_boxesR_mem', unp_visRs, boxlist_memR_e,
scorelist_e, torch.ones_like(scorelist_e,
dtype=torch.int32))
except Exception as e:
print('------ will handle this later -------')
### plotting of the detection boxes
#B, C, P = list(gt_boxesR_corners.shape)
gt_box_camR = gt_boxesR_corners #.reshape(B, 1, C, P)
origin_T_camRef = feed['origin_T_camRefs'][:, 0]
camRef_T_origin = torch.inverse(origin_T_camRef)
gt_box_camXs = utils.geom.apply_4x4_to_corners(
camRef_T_origin, gt_box_camR)
gt_ref_img = summ_writer.summ_box_by_corners(
'detnet/gt_box',
feed['rgb_camRs'].squeeze(1),
gt_box_camXs,
self.scorelist,
torch.ones([self.config.B, 1], dtype=torch.int32),
feed['pix_T_cams'][:, 0, :, :],
only_return=True)
# plot gt on each input view
gt_imgs = list()
for plt_i in range(self.config.S):
view = utils.geom.apply_4x4_to_corners(
camXs_T_camRs[:, plt_i, :, :], gt_box_camR)
gt_imgs.append(
summ_writer.summ_box_by_corners(
f'detnet/gt_view_{plt_i}',
feed['rgb_camXs'][:, plt_i, :, :, :],
view,
self.scorelist,
torch.ones([self.config.B, 1], dtype=torch.int32),
feed['pix_T_cams'][:, plt_i, :, :],
only_return=True))
gt_imgs.append(gt_ref_img)
# pred_box_corners_adam = utils.geom.transform_boxes_to_corners(boxlist_camR_e)
pred_box_corners_mem = utils.geom.transform_boxes_to_corners(
boxlist_memR_e)
pred_box_corners_adam = utils.geom.apply_4x4_to_corners(
utils.coordTcoord.get_ref_T_mem(B, self.halfmem_coord_Rs),
pred_box_corners_mem)
pred_box_ref = utils.geom.apply_4x4_to_corners(
camRef_T_origin, pred_box_corners_adam)
pred_ref_img = summ_writer.summ_box_by_corners(
'detnet/pred_boxes',
feed['rgb_camRs'].squeeze(1),
pred_box_ref,
scorelist_e.detach(),
torch.ones_like(scorelist_e, dtype=torch.int32),
feed['pix_T_cams'][:, 0, :, :],
only_return=True)
# plot pred on each input view
pred_ims_list = list()
for plt_i in range(self.config.S):
view = utils.geom.apply_4x4_to_corners(
camXs_T_camRs[:, plt_i, :, :], pred_box_corners_adam)
pred_ims_list.append(
summ_writer.summ_box_by_corners(
f'detnet/pred_view_{plt_i}',
feed['rgb_camXs'][:, plt_i, :, :, :],
view,
scorelist_e.detach(),
torch.ones_like(scorelist_e, dtype=torch.int32),
feed['pix_T_cams'][:, plt_i, :, :],
only_return=True))
pred_ims_list.append(pred_ref_img)
gt_imgs = torch.cat(gt_imgs, dim=0)
pred_ims_list = torch.cat(pred_ims_list, dim=0)
gt_grid = make_grid(gt_imgs, nrow=1)
pred_img_grid = make_grid(pred_ims_list, nrow=1)
summ_writer.summ_rgb('detnet/gt_grid', gt_grid.unsqueeze(0))
summ_writer.summ_rgb('detnet/pred_img_grid',
pred_img_grid.unsqueeze(0))
### plotting ends
# overlap = [max(0, min(e0[i], e1[i]) - max(s0[i], s1[i])) for i in range(3)]
# intersection = reduce(lambda x,y:x*y, overlap)
# union = pow(box0[3], 3) + pow(box1[3], 3) - intersection
# print(f'iou is : {intersection/union}')
scorelist_g = self.scorelist[0:1].detach().cpu().numpy()
boxlist_e = boxlist_camR_e[0:1].detach().cpu().numpy()
boxlist_g = boxlist_camR_g[0:1].detach().cpu().numpy()
scorelist_e = scorelist_e[0:1].detach().cpu().numpy()
boxlist_e, boxlist_g, scorelist_e, scorelist_g = utils.evaluate.drop_invalid_boxes(
boxlist_e, boxlist_g, scorelist_e, scorelist_g)
ious = [0.3, 0.4, 0.5, 0.6, 0.7]
maps, precisions_avg, scores_pred_val, ious_found = utils.evaluate.get_mAP(
boxlist_e, scorelist_e, boxlist_g, ious)
results['maps'] = maps
results['ious'] = ious
for ind, overlap in enumerate(ious):
summ_writer.summ_scalar('ap/%.2f_iou' % overlap, maps[ind])
summ_writer.summ_scalar('precision/%.2f_iou' % overlap,
precisions_avg[ind])
# do all of this computation if validation time is active
# validation is only called after 50 step or validate after number
if self.config.do_generate_data or (self.config.do_validation
and feed['set_num'] == 1):
# means I am executing the validation part of the code
# Here I have the box in reference and memory is also in ref_frame
# I should enter here while generating data
bbox_in_ref_cam = feed['bbox_in_ref_cam']
# based on the batch size this would be B, N, 3
min_bounds = bbox_in_ref_cam[:, :, 0]
max_bounds = bbox_in_ref_cam[:, :, -1]
lengths = torch.abs(max_bounds - min_bounds)
center = (max_bounds + min_bounds) * 0.5
# now form the box and then covert to lrt list
B, N = self.config.B, 1 # since i have only one box
# 9 is cx, cy, cz, lx, ly, lz, rx, ry, rz
boxlist = torch.zeros(B, N, 9)
# NOTE: Note: I am assuming here that N = 1 !!!!!!
boxlist[:, :, :3] = center #.unsqueeze(1)
boxlist[:, :, 3:6] = lengths #.unsqueeze(1)
# convert it to lrt list, it contains box length and rt to go
# from box coordinates to ref coordinate system.
box_lrtlist = utils.geom.convert_boxlist_to_lrtlist(boxlist)
# now this is already in the ref coordinate system which was not
# the case with my previous use of the crop_zoom_from_mem func.
# Hence I had previously included camR_T_camXs which is not req here
presumably_object_tensor = utils.vox.crop_zoom_from_mem(
featRs_with_target_view, self.feat_mem_coord_Rs,
box_lrtlist[:, 0, :], 32, 32, 32)
# NOTE: As of now I am not doing backprop through this Tensor so
# no need to keep it in gpu anymore
results['object_tensor'] = presumably_object_tensor.detach().cpu()
results['record_name'] = feed['record']
# Add the plot of this to tensorboard, and also think how can you
# visualize if the correct thing is being returned to you.
summ_writer.summ_feats('crop_feats_val/object_tensor',
tuple([presumably_object_tensor]),
pca=True)
# if hyp.do_metric_learning:
# B, _, _, _, _ = presumably_object_tensor.shape
# assert B >= 2, "Metric learner requires one positive and atleast one negative example to train"
# metric_loss, _ = self.metric_learner(presumably_object_tensor,feed["object_id"])
# total_loss += metric_loss
# summ_writer.summ_scalar('metric_learn/metric_loss', metric_loss.cpu().item())
summ_writer.summ_scalar('loss', total_loss.cpu().item())
return total_loss, results