def load_obj(filename_obj, normalization=False, load_texture=False, texture_res=4, texture_type='surface'):
"""
Load Wavefront .obj file.
This function only supports vertices (v x x x) and faces (f x x x).
"""
assert texture_type in ['surface', 'vertex']
# load vertices
vertices = []
with open(filename_obj) as f:
lines = f.readlines()
for line in lines:
if len(line.split()) == 0:
continue
if line.split()[0] == 'v':
vertices.append([float(v) for v in line.split()[1:4]])
vertices = jt.array(np.vstack(vertices)).float32()
# load faces
faces = []
for line in lines:
if len(line.split()) == 0:
continue
if line.split()[0] == 'f':
vs = line.split()[1:]
nv = len(vs)
v0 = int(vs[0].split('/')[0])
for i in range(nv - 2):
v1 = int(vs[i + 1].split('/')[0])
v2 = int(vs[i + 2].split('/')[0])
faces.append((v0, v1, v2))
faces = jt.array(np.vstack(faces).astype(np.int32)).float32() - 1
# load textures
if load_texture and texture_type == 'surface':
textures = None
for line in lines:
if line.startswith('mtllib'):
filename_mtl = os.path.join(os.path.dirname(filename_obj), line.split()[1])
textures = load_textures(filename_obj, filename_mtl, texture_res)
if textures is None:
raise Exception('Failed to load textures.')
elif load_texture and texture_type == 'vertex':
textures = []
for line in lines:
if len(line.split()) == 0:
continue
if line.split()[0] == 'v':
textures.append([float(v) for v in line.split()[4:7]])
textures = jt.array(np.vstack(textures)).float()
# normalize into a unit cube centered zero
if normalization:
vertices -= vertices.min(0)
vertices /= jt.abs(vertices).max()
vertices *= 2
vertices -= vertices.max(0) / 2
if load_texture:
return vertices, faces, textures
else:
return vertices, faces
def smooth_l1_loss(input, target, beta=1. / 9, size_average=True):
diff = jt.abs(input - target)
less_than_one = (diff<1.0).float32()
loss = (less_than_one * 0.5 * diff**2) + (1 - less_than_one) * (diff - 0.5)
if size_average:
return loss.mean()
return loss.sum()
def l2_loss(input, target):
"""
very similar to the smooth_l1_loss , but with
the extra beta parameter
"""
pos_inds = jt.nonzero(target > 0.0).squeeze(1)
if pos_inds.shape[0] > 0:
cond = jt.abs(input[pos_inds] - target[pos_inds])
loss = 0.5 * cond**2 / pos_inds.shape[0]
else:
loss = input * 0.0
return loss.sum()
def test_unary_op(self):
assert jt.float64(1).data.dtype == "float64"
assert (jt.abs(-1) == 1).data.all()
assert (abs(-jt.float64(1)) == 1).data.all()
a = np.array([-1,2,3,0])
check("abs", a)
check("negative", a)
check("logical_not", a)
check("bitwise_not", a)
b = np.array([1.1, 2.2, 3.3, 4.4, -1, 0])
check("log", a)
check("exp", a)
check("sqrt", a)
def execute(self, pred, true):
loss = self.loss_fcn(pred, true)
pred_prob = jt.sigmoid(pred) # prob from logits
alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
modulating_factor = jt.abs(true - pred_prob)**self.gamma
loss *= alpha_factor * modulating_factor
if self.reduction == 'mean':
return loss.mean()
elif self.reduction == 'sum':
return loss.sum()
else: # 'none'
return loss
def smooth_l1_loss(y_true, y_pred,reduction="mean"):
"""Implements Smooth-L1 loss.
y_true and y_pred are typically: [N, 4], but could be any shape.
Args:
y_true - ground truth
y_pred - predictions
reduction - the mode of cal loss which must be in ['mean','sum','none']
"""
diff = jt.abs(y_true - y_pred)
less_than_one = (diff<1.0).float32()
loss = (less_than_one * 0.5 * diff.sqr()) + (1 - less_than_one) * (diff - 0.5)
if reduction=="mean":
return loss.mean()
elif reduction=="sum":
return loss.sum()
elif reduction=="none":
return loss
else:
raise ValueError(f'not support {reduction}')
def execute(self, input, target):
ret = jt.abs(input - target)
if self.reduction != None:
ret = jt.mean(ret) if self.reduction == 'mean' else jt.sum(ret)
return ret
for i, (imgs, _) in enumerate(train_loader):
# Configure input
real_imgs = jt.array(imgs)
# -----------------
# Train Generator
# -----------------
# Sample noise as generator input
z = jt.array(np.random.normal(0, 1, (imgs.shape[0], opt.latent_dim))).float32()
# Generate a batch of images
gen_imgs = generator(z)
# Loss measures generator's ability to fool the discriminator
g_loss = jt.mean(jt.abs(discriminator(gen_imgs) - gen_imgs))
optimizer_G.step(g_loss)
# ---------------------
# Train Discriminator
# ---------------------
# Measure discriminator's ability to classify real from generated samples
d_real = discriminator(real_imgs)
d_fake = discriminator(gen_imgs.stop_grad())
d_loss_real = jt.mean(jt.abs(d_real - real_imgs))
d_loss_fake = jt.mean(jt.abs(d_fake - gen_imgs.stop_grad()))
d_loss = d_loss_real - k * d_loss_fake
optimizer_D.step(d_loss)
def execute(self, net, predictions, targets, masks, num_crowds):
"""Multibox Loss
Args:
predictions (tuple): A tuple containing loc preds, conf preds,
mask preds, and prior boxes from SSD net.
loc shape: jt.size(batch_size,num_priors,4)
conf shape: jt.size(batch_size,num_priors,num_classes)
masks shape: jt.size(batch_size,num_priors,mask_dim)
priors shape: jt.size(num_priors,4)
proto* shape: jt.size(batch_size,mask_h,mask_w,mask_dim)
targets (list<tensor>): Ground truth boxes and labels for a batch,
shape: [batch_size][num_objs,5] (last idx is the label).
masks (list<tensor>): Ground truth masks for each object in each image,
shape: [batch_size][num_objs,im_height,im_width]
num_crowds (list<int>): Number of crowd annotations per batch. The crowd
annotations should be the last num_crowds elements of targets and masks.
* Only if mask_type == lincomb
"""
loc_data = predictions['loc']
conf_data = predictions['conf']
mask_data = predictions['mask']
priors = predictions['priors']
if cfg.mask_type == mask_type.lincomb:
proto_data = predictions['proto']
score_data = predictions['score'] if cfg.use_mask_scoring else None
inst_data = predictions['inst'] if cfg.use_instance_coeff else None
labels = [None] * len(targets) # Used in sem segm loss
batch_size = loc_data.shape[0]
num_priors = priors.shape[0]
num_classes = self.num_classes
# Match priors (default boxes) and ground truth boxes
# These tensors will be created with the same device as loc_data
loc_t = jt.empty((batch_size, num_priors, 4),dtype=loc_data.dtype)
gt_box_t = jt.empty((batch_size, num_priors, 4),dtype=loc_data.dtype)
conf_t = jt.empty((batch_size, num_priors)).int32()
idx_t = jt.empty((batch_size, num_priors)).int32()
if cfg.use_class_existence_loss:
class_existence_t = jt.empty((batch_size, num_classes-1),dtype=loc_data.dtype)
# jt.sync(list(predictions.values()))
for idx in range(batch_size):
truths = targets[idx][:, :-1]
labels[idx] = targets[idx][:, -1].int32()
if cfg.use_class_existence_loss:
# Construct a one-hot vector for each object and collapse it into an existence vector with max
# Also it's fine to include the crowd annotations here
class_existence_t[idx,:] = jt.eye(num_classes-1)[labels[idx]].max(dim=0)[0]
# Split the crowd annotations because they come bundled in
cur_crowds = num_crowds[idx]
if cur_crowds > 0:
split = lambda x: (x[-cur_crowds:], x[:-cur_crowds])
crowd_boxes, truths = split(truths)
# We don't use the crowd labels or masks
_, labels[idx] = split(labels[idx])
_, masks[idx] = split(masks[idx])
else:
crowd_boxes = None
match(self.pos_threshold, self.neg_threshold,
truths, priors, labels[idx], crowd_boxes,
loc_t, conf_t, idx_t, idx, loc_data[idx])
gt_box_t[idx,:,:] = truths[idx_t[idx]]
# wrap targets
loc_t.stop_grad()
conf_t.stop_grad()
idx_t.stop_grad()
pos = conf_t > 0
num_pos = pos.sum(dim=1, keepdims=True)
# Shape: [batch,num_priors,4]
pos_idx = pos.unsqueeze(pos.ndim).expand_as(loc_data)
losses = {}
# Localization Loss (Smooth L1)
if cfg.train_boxes:
loc_p = loc_data[pos_idx].view(-1, 4)
loc_t = loc_t[pos_idx].view(-1, 4)
# print(loc_t)
losses['B'] = nn.smooth_l1_loss(loc_p, loc_t, reduction='sum') * cfg.bbox_alpha
if cfg.train_masks:
if cfg.mask_type == mask_type.direct:
if cfg.use_gt_bboxes:
pos_masks = []
for idx in range(batch_size):
pos_masks.append(masks[idx][idx_t[idx, pos[idx]]])
masks_t = jt.contrib.concat(pos_masks, 0)
masks_p = mask_data[pos, :].view(-1, cfg.mask_dim)
losses['M'] = nn.bce_loss(jt.clamp(masks_p, 0, 1), masks_t, size_average=False) * cfg.mask_alpha
else:
losses['M'] = self.direct_mask_loss(pos_idx, idx_t, loc_data, mask_data, priors, masks)
elif cfg.mask_type == mask_type.lincomb:
ret = self.lincomb_mask_loss(pos, idx_t, loc_data, mask_data, priors, proto_data, masks, gt_box_t, score_data, inst_data, labels)
if cfg.use_maskiou:
loss, maskiou_targets = ret
else:
loss = ret
losses.update(loss)
if cfg.mask_proto_loss is not None:
if cfg.mask_proto_loss == 'l1':
losses['P'] = jt.mean(jt.abs(proto_data)) / self.l1_expected_area * self.l1_alpha
elif cfg.mask_proto_loss == 'disj':
losses['P'] = -jt.mean(jt.max(nn.log_softmax(proto_data, dim=-1), dim=-1)[0])
# Confidence loss
if cfg.use_focal_loss:
if cfg.use_sigmoid_focal_loss:
losses['C'] = self.focal_conf_sigmoid_loss(conf_data, conf_t)
elif cfg.use_objectness_score:
losses['C'] = self.focal_conf_objectness_loss(conf_data, conf_t)
else:
losses['C'] = self.focal_conf_loss(conf_data, conf_t)
else:
if cfg.use_objectness_score:
losses['C'] = self.conf_objectness_loss(conf_data, conf_t, batch_size, loc_p, loc_t, priors)
else:
losses['C'] = self.ohem_conf_loss(conf_data, conf_t, pos, batch_size)
# Mask IoU Loss
if cfg.use_maskiou and maskiou_targets is not None:
losses['I'] = self.mask_iou_loss(net, maskiou_targets)
# These losses also don't depend on anchors
if cfg.use_class_existence_loss:
losses['E'] = self.class_existence_loss(predictions['classes'], class_existence_t)
if cfg.use_semantic_segmentation_loss:
losses['S'] = self.semantic_segmentation_loss(predictions['segm'], masks, labels)
# Divide all losses by the number of positives.
# Don't do it for loss[P] because that doesn't depend on the anchors.
total_num_pos = num_pos.sum().float()
for k in losses:
if k not in ('P', 'E', 'S'):
losses[k] /= total_num_pos
else:
losses[k] /= batch_size
# Loss Key:
# - B: Box Localization Loss
# - C: Class Confidence Loss
# - M: Mask Loss
# - P: Prototype Loss
# - D: Coefficient Diversity Loss
# - E: Class Existence Loss
# - S: Semantic Segmentation Loss
return losses
def L1(self, A, B):
diff = A.get_mask_tensor() - B.get_mask_tensor()
diff = jt.sum(jt.abs(diff.float())).item()
return diff
