def compute_deg_hist(ds):
max_deg = max([
max(degree(data.edge_index[1], num_nodes=data.num_nodes, dtype=torch.long)).item()
for data in ds
])
deg = torch.zeros(max_deg + 1, dtype=torch.long)
for data in ds:
d = degree(data.edge_index[1], num_nodes=data.num_nodes, dtype=torch.long)
deg += torch.bincount(d, minlength=deg.numel())
return deg
def fast_hist(a, b, n):
'''
a and b are predict and mask respectively
n is the number of classes
'''
# k = (a >= 0) & (a < n)
# return np.bincount(n * a[k].astype(int) + b[k], minlength=n ** 2).reshape(n, n)
a = a.flatten()
b = b.flatten()
return torch.bincount(n * a.int() + b.int(), minlength=n**2).view(n, n)
def update(self, a, b):
with torch.no_grad():
n = self.num_classes
if self.mat is None:
self.mat = torch.zeros((n, n),
dtype=torch.int64,
device=a.device)
k = (a >= 0) & (a < n)
inds = n * a[k].to(torch.int64) + b[k]
self.mat += torch.bincount(inds, minlength=n**2).reshape(n, n)
def save(x, y, dir):
for i, (x_i, y_i) in enumerate(zip(x, y)):
plt.imshow(x_i)
plt.axis("off")
y_encoded = torch.bincount(torch.tensor(y_i), minlength=10)
plt.title("Label: {}\nEncoded: {}".format(y_i, y_encoded))
plt.savefig(os.path.join(dir, "%s.png" % i))
if i > 10:
break
def iIoU_class(pred, target, num_classes, verbose=False):
"""[summary]
It is well-known that the global IoU measure is biased toward object instances that cover a large image area.
In street scenes with their strong scale variation this can be problematic.
Specifically for traffic participants, which are the key classes in our scenario,
we aim to evaluate how well the individual instances in the scene are represented in the labeling.
To address this, we additionally evaluate the semantic labeling using an
instance-level intersection-over-union metric iIoU = iTP ⁄ (iTP+FP+iFN).
Again iTP, FP, and iFN denote the numbers of true positive, false positive, and false negative pixels, respectively.
However, in contrast to the standard IoU measure, iTP and iFN are computed by weighting the contribution of each pixel by the ratio of the class’
average instance size to the size of the respective ground truth instance.
It is important to note here that unlike the instance-level task below,
we assume that the methods only yield a standard per-pixel semantic class labeling as output.
Therefore, the false positive pixels are not associated with any instance and thus do not require normalization.
The final scores, iIoUcategory and iIoUclass, are obtained as the means for the two semantic granularities.
Parameters
----------
pred : [torch.tensor]
BSxD1xD2xD3 , predict class for each pixel. No need to predict the -1 class! element of 0-(num_classes-1)
target : [torch.tensor]
BSxD1xD2xD3 , -1 for the VOID pixels that should not induce an error! element of -1-(num_classes-1)
num_classes : [int]
invalid class does not count as a class. So lets say targets takes values -1 - 19 then you have 20 classes
"""
BS = pred.shape[0]
# add 1 so the index ranges from 0 to NUM_CLASSES
pred = pred.type(torch.int) + 1
target = target.type(torch.int) + 1
# NOW class=0 should not induce a loss
# Set pixels that are predicted but no label is available to 0. These pixels dont enduce a loss.
# Neither does the IoU of class 0 nor do these pixels count to the UNION for the other classes if predicted wrong.
pred = pred * (target > 0).type(pred.dtype)
iou_per_image = torch.zeros( (BS), device=pred.device)
# we have to do this calculation for each image.
for b in range(BS):
weight = torch.bincount(target[b].flatten())[1:]
weight = weight/ weight.sum()
w = torch.zeros( (num_classes), device=target.device, dtype=weight.dtype )
w[:weight.shape[0]] = weight
TPS, FPS, TNS, FNS, _ = stat_scores_multiple_classes(pred[b], target[b], num_classes+1)
if verbose:
print(f'TPS:{TPS}, \nFPS:{FPS}, \nFNS:{FNS}, \nTNS:{TNS}')
print(F'Inter: {TPS},\nUnion: {TPS+FPS+FNS}')
IoU = (TPS[1:]*w) / ((TPS[1:]*w) + FPS[1:] + (FNS[1:]*w) )
mIoU = (IoU[torch.isnan(IoU)==False]).mean()
iou_per_image[b] = mIoU
return torch.mean( iou_per_image ) #returns mean over batch
def test(args, model, device, test_loader, criterion):
# Switch model to evaluation mode. This is necessary for layers like dropout, batchnorm etc which behave differently in training and evaluation mode
model.eval()
test_loss = 0
total = 0
correct = 0
best_loss = 1
num_classes = 23
class_correct = list(0. for i in range(num_classes))
class_total = list(0. for i in range(num_classes))
example_images = []
with torch.no_grad():
for data, target in test_loader:
# Load the input features and labels from the test dataset
data, target = data.to(device), target.to(device)
# Make predictions: Pass image data from test dataset, make predictions about class image belongs to (0-9 in this case)
output = model(data)
# Compute the loss sum up batch loss
test_loss += criterion(output, target).item()
#test_loss += F.nll_loss(output, target, reduction='sum').item()
# Get the index of the max log-probability
pred = output.max(1, keepdim=True)[1][:,0]
per_class_count = torch.bincount(target, minlength=num_classes)
#correct += pred.eq(target.view_as(pred)).sum().item()
for i in range(num_classes):
class_total[i] += per_class_count[i]
mask = target == i
p = pred[mask]
t = target[mask]
class_correct[i] += p.eq(t.view_as(p)).sum().item()
total += pred.eq(pred).sum().item()
correct += pred.eq(target.view_as(pred)).sum().item()
# WandB Log images in your test dataset automatically, along with predicted and true labels by passing pytorch tensors with image data into wandb.Image
example_images.append(wandb.Image(
data[0], caption="Pred: {} Truth: {}".format(pred[0].item(), target[0])))
# WandB wandb.log(a_dict) logs the keys and values of the dictionary passed in and associates the values with a step.
# You can log anything by passing it to wandb.log, including histograms, custom matplotlib objects, images, video, text, tables, html, pointclouds and other 3D objects.
# Here we use it to log test accuracy, loss and some test images (along with their true and predicted labels).
#wandb.log({
# "Examples": example_images,
# "Test Accuracy": 100. * correct / len(test_loader.dataset),
# "Test Loss": test_loss})
log_dict = {
"Examples": example_images,
"Test Accuracy": 100. * correct / len(test_loader.dataset),
"Test Loss": test_loss}
for i in range(num_classes):
log_dict["Test Accuracy - " + list(test_loader.dataset.categories.keys())[i]] = 100. * class_correct[i] / class_total[i]
#print(log_dict)
wandb.log(log_dict)
def compute_present_locations(args, corpus, cache_filename,
model, segmenter, classnum, full_sample):
# Phase 1. Identify a set of locations where there are doorways.
# Segment the image and find featuremap pixels that maximize the number
# of doorway pixels under the featuremap pixel.
if all(k in corpus for k in ['present_indices',
'object_present_sample', 'object_present_location',
'object_location_popularity', 'weighted_mean_present_feature']):
return
progress = default_progress()
feature_shape = model.feature_shape[args.layer][2:]
num_locations = numpy.prod(feature_shape).item()
num_units = model.feature_shape[args.layer][1]
with torch.no_grad():
weighted_feature_sum = torch.zeros(num_units).cuda()
object_presence_scores = []
for [zbatch] in progress(
torch.utils.data.DataLoader(TensorDataset(full_sample),
batch_size=args.inference_batch_size, num_workers=10,
pin_memory=True),
desc="Object pool"):
zbatch = zbatch.cuda()
tensor_image = model(zbatch)
segmented_image = segmenter.segment_batch(tensor_image,
downsample=2)
mask = (segmented_image == classnum).max(1)[0]
score = torch.nn.functional.adaptive_avg_pool2d(
mask.float(), feature_shape)
object_presence_scores.append(score.cpu())
feat = model.retained_layer(args.layer)
weighted_feature_sum += (feat * score[:,None,:,:]).view(
feat.shape[0],feat.shape[1], -1).sum(2).sum(0)
object_presence_at_feature = torch.cat(object_presence_scores)
object_presence_at_image, object_location_in_image = (
object_presence_at_feature.view(args.search_size, -1).max(1))
best_presence_scores, best_presence_images = torch.sort(
-object_presence_at_image)
all_present_indices = torch.sort(
best_presence_images[:(args.train_size+args.eval_size)])[0]
corpus.present_indices = all_present_indices[:args.train_size]
corpus.object_present_sample = full_sample[corpus.present_indices]
corpus.object_present_location = object_location_in_image[
corpus.present_indices]
corpus.object_location_popularity = torch.bincount(
corpus.object_present_location,
minlength=num_locations)
corpus.weighted_mean_present_feature = (weighted_feature_sum.cpu() / (
1e-20 + object_presence_at_feature.view(-1).sum()))
corpus.eval_present_indices = all_present_indices[-args.eval_size:]
corpus.eval_present_sample = full_sample[corpus.eval_present_indices]
corpus.eval_present_location = object_location_in_image[
corpus.eval_present_indices]
if cache_filename:
numpy.savez(cache_filename, **corpus)
def build_latent_distribution(self, alpha: int = 1):
"""Two passes:num_images_pixels
1. we calculate the minimum latent value across our entire training
distribution,
2. we then add |min| to the latent values such that they are all >= 0 and
then use torch.bincount to get discrete value counts
-> which we then laplace smooth and convert into a CDF.
If a code is in multiple parts, e.g lateral FPN features, they are flattened and concatenated.
"""
self.cdf = dict()
self.min_val = torch.tensor(0.0).to(self.device).long()
num_images = 0
self.model.eval()
if self.negative_codes:
for batch in self.train_loader:
with torch.no_grad():
out_dict = self.model(batch)
for code_feat in self.code_feats:
self.min_val = torch.min(
self.min_val,
out_dict[code_feat].long().min(),
)
num_images += len(batch)
if num_images > self.num_train_images:
break
self.min_val = self.min_val.abs()
self.bins = torch.tensor([0.0]).to(self.device).long()
num_images = 0
for batch in self.train_loader:
with torch.no_grad():
out_dict = self.model(batch)
flat_codes = []
for code_feat in self.code_feats:
flat_codes.append(out_dict[code_feat].long().flatten() +
self.min_val)
batch_bins = torch.bincount(torch.cat(flat_codes))
if len(batch_bins) > len(self.bins):
batch_bins[:len(self.bins)] += self.bins
self.bins = batch_bins
elif len(self.bins) > len(batch_bins):
self.bins[:len(batch_bins)] += batch_bins
else:
self.bins += batch_bins
num_images += len(batch)
if num_images > self.num_train_images:
break
bins = self.bins.float()
bins_smooth = ((bins + alpha) / (bins.sum() + len(bins) * alpha)
).cpu() # additive smooth counts using alpha
self.cdf = rc.prob_to_cum_freq(bins_smooth,
resolution=2 *
len(bins_smooth)) # convert pdf -> cdf
self.model.train()
def fuse_mask(n_mask, r_mask):
base = torch.where(n_mask > 0,
torch.tensor(1).cuda(cuda_id),
torch.tensor(0).cuda(cuda_id)).float()
areas = torch.max(n_mask)
#for i in range(1,torch.max(r_mask).long()+1):
i = 1
shift = torch.where(r_mask == i,
torch.tensor(1).cuda(cuda_id),
torch.tensor(0).cuda(cuda_id)).float()
non_overlap = torch.where(base - shift == -1,
torch.tensor(1).cuda(cuda_id),
torch.tensor(0).cuda(cuda_id)).float()
overlap = shift - non_overlap
if torch.sum(non_overlap) / torch.sum(shift) > 0.4:
areas += 1
n_mask = torch.where(non_overlap == 1, areas, n_mask)
base = torch.where(n_mask > 0,
torch.tensor(1).cuda(cuda_id),
torch.tensor(0).cuda(cuda_id)).float()
#print(areas)
else:
area_num = torch.argmax(
torch.bincount(
torch.where(
overlap.long() == 1, n_mask.long(),
torch.tensor(0).cuda(cuda_id)).view(-1))[1:]).float() + 1
n_mask = torch.where(non_overlap == 1, area_num, n_mask)
base = torch.where(n_mask > 0,
torch.tensor(1).cuda(cuda_id),
torch.tensor(0).cuda(cuda_id)).float()
#print(areas)
# areas_nums=torch.tensor(1).float().cuda(cuda_id)
# for i in range(1,torch.max(n_mask).long()+1):
# region=torch.where(n_mask==i,torch.tensor(1).cuda(cuda_id),torch.tensor(0).cuda(cuda_id)).float()
# pixels=region.nonzero()
# if pixels.shape[0]>0:
# minx=torch.min(pixels[:,0])
# maxx=torch.max(pixels[:,0])
# miny=torch.min(pixels[:,1])
# maxy=torch.max(pixels[:,1])
# for i in range(1,torch.ceil((maxx-minx).float()/80).int()+1):
# for j in range(1,torch.ceil((maxy-miny).float()/80).int()+1):
# if torch.sum(region[minx+80*(i-1):minx+80*i,miny+80*(j-1):miny+80*j])>400:
# region[minx+80*(i-1):minx+80*i,miny+80*(j-1):miny+80*j]*=i*j
# areas=torch.unique(region).sort()[0]
# for i in range(1,len(areas)):
# region=torch.where(region==areas[i],-areas_nums,region)
# areas_nums+=1
# n_mask=torch.where(n_mask==i,region,n_mask)
# n_mask=-n_mask
return n_mask
def forward(self, data):
joints = data.y
joints_norepeat = []
joints_batch = []
joints_sample_1 = []
joints_sample_2 = []
pair_batch = []
label = []
accumulate_start_pair = 0
for i in range(len(torch.unique(data.batch))):
joints_sample = joints[data.batch == i, :]
joints_sample = joints_sample[:data.num_joint[i], :]
joints_norepeat.append(joints_sample)
joints_batch.append(data.batch.new_full((data.num_joint[i],), i))
pair_idx = data.pairs[accumulate_start_pair: accumulate_start_pair + data.num_pair[i]]
accumulate_start_pair += data.num_pair[i]
if np.random.uniform() > 0.5:
joints_sample_1.append(joints_sample[pair_idx[:, 0].long()])
joints_sample_2.append(joints_sample[pair_idx[:, 1].long()])
else:
joints_sample_1.append(joints_sample[pair_idx[:, 1].long()])
joints_sample_2.append(joints_sample[pair_idx[:, 0].long()])
pair_batch.append(data.batch.new_full((data.num_pair[i],), i))
label.append(pair_idx[:, -1])
joints_norepeat = torch.cat(joints_norepeat, dim=0)
joints_batch = torch.cat(joints_batch).long()
pair_batch = torch.cat(pair_batch).long()
joints_sample_1 = torch.cat(joints_sample_1, dim=0)
joints_sample_2 = torch.cat(joints_sample_2, dim=0)
label = torch.cat(label, dim=0).unsqueeze(1)
joints_pair = torch.cat((joints_sample_1, joints_sample_2, data.pairs[:, 2:4]), dim=1)
pair_feature = self.expand_joint_feature(joints_pair)
joint_feature = self.joint_encoder(joints_norepeat, joints_batch)
joint_feature = torch.repeat_interleave(joint_feature, torch.bincount(pair_batch), dim=0)
shape_feature = self.shape_encoder(data)
shape_feature = torch.repeat_interleave(shape_feature, torch.bincount(pair_batch), dim=0)
pair_feature = torch.cat((shape_feature, joint_feature, pair_feature), dim=1)
pre_label = self.mix_transform(pair_feature)
return pre_label, label
def update(self, outputs, targets):
with torch.no_grad():
outputs = outputs.argmax(dim=1, keepdim=True)
outputs, targets = outputs.cpu(), targets.cpu()
for output, target in zip(outputs, targets):
output, target = output.flatten(), target.flatten()
mask = (target >= 0) * (target < self.n_classes)
self.confusion_matrix += torch.bincount(
self.n_classes * target[mask] + output[mask],
minlength=self.n_classes**2).reshape(
self.n_classes, self.n_classes)
def __init__(self, labels: Union[List[int], torch.Tensor], batch_size: int,
num_identities: int, num_iterations: int):
self.num_identities = num_identities
self.num_iterations = num_iterations
self.samples_per_id = batch_size // num_identities
self.labels = torch.as_tensor(labels, dtype=torch.long)
self.counts = torch.bincount(self.labels)
self.label_indices = [
torch.nonzero(self.labels == i).squeeze(1).tolist()
for i in range(len(self.counts))
]
def update(self, target, output):
n = self.num_classes
if self.mat is None:
self.mat = torch.zeros((n, n),
dtype=torch.int64,
device=target.device)
with torch.no_grad():
k = (target >= 0) & (target < n)
inds = n * target[k].to(torch.int64) + output[k]
self.mat += torch.bincount(inds, minlength=n**2).reshape(n, n)
def compute_mean_seg_in_images(batch_z, *args):
img = model(batch_z.cuda())
seg = segmodel.segment_batch(img, downsample=4)
seg_area = seg.shape[2] * seg.shape[3]
seg_counts = torch.bincount(
(seg + (num_seglabels * torch.arange(
seg.shape[0], dtype=seg.dtype,
device=seg.device)[:, None, None, None])).view(-1),
minlength=num_seglabels * seg.shape[0]).view(seg.shape[0], -1)
seg_fracs = seg_counts.float() / seg_area
return seg_fracs
def mode(self, list, dim):
out_mode = []
stack_list = tr.stack(list, dim)
for i in range(stack_list.size(0)):
sample_i = stack_list[i]
mode = tr.bincount(sample_i)
mode = tr.argmax(mode)
out_mode.append(mode)
out_mode = tr.stack(out_mode, 0)
return out_mode
def forward(model, i, data):
# 改成用 geometric的Data格式
items, targets, mask, batch, seq = data.x, data.y, data.sequence_mask, data.batch, data.sequence
seq = seq.view(targets.shape[0], -1)
mask = mask.view(targets.shape[0], -1)
A = []
# datas = data.to_data_list()
# graphs = [to_networkx(d) for d in datas]
# A = [nx.convert_matrix.to_pandas_adjacency(g).values for g in graphs] # 無向圖adj = in + out
# A_out = [g for g in graphs] # 有向圖的adj就是A_out
# todo 解決cpu usage高的問題
# global graph
gg = model.global_data
gg_edge_index = gg.edge_index
# 直接對 batch下所有node做NeighborSample
batch_nodes = seq.flatten()
# batch_nodes = torch.unique(batch_nodes) # 取unique node in batch sessions
# batch_nodes = batch_nodes[batch_nodes!=0] # 移除padding node id
# sample as whole batch, 從大graph中找session graph內的node id的鄰居
# subgraph_loaders = NeighborSampler(gg_edge_index, node_idx=batch_nodes, sizes=[-1], shuffle=False, num_workers=0, batch_size=batch_nodes.shape[0]) # all neighbors
# fixme 放全部node
subgraph_loaders = NeighborSampler(
gg_edge_index,
node_idx=batch_nodes,
sizes=[10, 5],
shuffle=False,
num_workers=0,
batch_size=batch_nodes.shape[0]) # 2 hop
hidden, pad, g_h = model(items, A, data.edge_index,
subgraph_loaders) # session graph node embeddings
# 推回原始序列
sections = torch.bincount(batch).cpu().numpy()
# split whole x back into graphs G_i
hidden = torch.split(hidden, tuple(sections))
# todo 增加不考慮padding的選項
mask_true = True
if mask_true:
leng = mask.shape[1] # padding完的session長度
alias_inputs = data.alias_inputs
s_len = data.sequence_len.cpu().numpy().tolist()
alias_inputs = torch.split(alias_inputs, s_len)
seq_hidden = torch.stack([
get(pad, i, hidden, alias_inputs, leng)
for i in torch.arange(len(alias_inputs)).long()
])
g_h = g_h.view([len(hidden), leng, -1])
else:
seq_hidden = hidden
seq_hidden += g_h
return targets, model.compute_scores(seq_hidden, mask, mask_true)
def forward(self, z, targets):
targets_ = targets.argmax(dim=1).to(z.device)
bc = torch.bincount(targets_)
freq_weights = bc.max() / bc.float()
freq_smp_weights = freq_weights[targets.argmax(dim=1)]
inputs = self.kappa * z
loss = (freq_smp_weights * self.loss(inputs, targets_)).mean()
return loss
def update(self,
val: Union[float, torch.Tensor, Sequence[torch.Tensor]] = None,
**kwargs):
y_pred, y_true = val
self._num_samples += len(y_pred)
y_pred = torch.argmax(y_pred, dim=1)
matrix_indices = self.num_classes * y_true + y_pred
m = torch.bincount(matrix_indices,
minlength=self.num_classes**2).reshape(
self.num_classes, self.num_classes)
self.confusion_matrix += m.to(self.confusion_matrix)
def quantization(raw_feat):
# Feature quantization
raw_feat = torch.Tensor(raw_feat).type(torch.float32)
c = torch.Tensor(cluster_centers[None, :, :]).type(torch.float32)
raw_feat = LazyTensor(raw_feat[:, None, :])
c = LazyTensor(c)
dist = ((raw_feat - c) ** 2).sum(-1)
preds = dist.argmin(dim=1).long().view(-1)
term_freq = torch.bincount(preds, minlength=num_clusters).type(torch.float32).numpy()
normed_term_freq = term_freq / np.linalg.norm(term_freq)
return normed_term_freq
def confusion_matrix(pred: torch.Tensor,
target: torch.Tensor,
num_classes: int = None) -> torch.Tensor:
num_classes = get_num_classes(pred, target, num_classes)
unique_labels = target.view(-1) * num_classes + pred.view(-1)
bins = torch.bincount(unique_labels, minlength=num_classes**2)
cm = bins.reshape(num_classes, num_classes).squeeze().float()
return cm
def __call__(self, samples):
l_idx = -1
if isinstance(samples, tuple):
samples = [samples]
maxlen = 0
for i, X, y, seq_id in samples:
if maxlen < X.shape[l_idx]:
maxlen = X.shape[l_idx]
X_ret = list()
y_ret = list()
idx_ret = list()
size_ret = list()
seq_id_ret = list()
for i, X, y, seq_id in samples:
dif = maxlen - X.shape[l_idx]
X_ = X
if dif > 0:
X_ = F.pad(X, (0, dif), value=self.padval)
X_ret.append(X_)
y_ret.append(y)
size_ret.append(X.shape[l_idx])
idx_ret.append(i)
seq_id_ret.append(seq_id)
# calculate tetranucleotide frequency
chunks = torch.stack(X_ret)
## 1. hash 4-mers
__seq = self.cmap[chunks]
i4mers = torch.stack([__seq[:, 0:-3], __seq[:, 1:-2], __seq[:, 2:-1], __seq[:, 3:]], axis=2)
mask = torch.any(i4mers < 0, axis=2)
h4mers = i4mers.matmul(self.bases) # hashed 4-mers
h4mers[mask] = 256 # use 257 to mark any 4-mers that had ambiguous nucleotides
## 2. count hashed 4-mers i.e. count integers from between 0-257 inclusive
tnf = torch.zeros((32, 257), dtype=float)
for i in range(tnf.shape[0]):
counts = torch.bincount(h4mers[i], minlength=257)
tnf[i] = counts/i4mers.shape[1]
## 3. merge canonical 4-mers
canon_tnf = torch.zeros((32, 136))
canon_tnf[:, :len(self.canonical)] = tnf[:, self.canonical] + tnf[:, self.noncanonical]
canon_tnf[:, len(self.canonical):] = tnf[:, self.palindromes]
X_ret = canon_tnf
y_ret = torch.stack(y_ret)
size_ret = torch.tensor(size_ret)
idx_ret = torch.tensor(idx_ret)
seq_id_ret = torch.tensor(seq_id_ret)
return (idx_ret, X_ret, y_ret, size_ret, seq_id_ret)
def batch_bincount(
batch_input: Union[torch.Tensor, List[torch.Tensor]],
n_batches: int,
weights: Optional[torch.Tensor] = None,
minlength=0,
) -> torch.Tensor:
return torch.stack([
torch.bincount(batch_input[batch_idx],
weights=weights,
minlength=minlength) for batch_idx in range(n_batches)
])
def update(self, a, b):
n = self.num_classes
if self.mat is None:
# 创建混淆矩阵
self.mat = torch.zeros((n, n), dtype=torch.int64, device=a.device)
with torch.no_grad():
# 寻找GT中为目标的像素索引
k = (a >= 0) & (a < n)
# 统计像素真实类别a[k]被预测成类别b[k]的个数(这里的做法很巧妙)
inds = n * a[k].to(torch.int64) + b[k]
self.mat += torch.bincount(inds, minlength=n**2).reshape(n, n)
def torch_compute_confusion_matrix(gts, preds):
gts = torch.stack(gts, dim=0)
preds = torch.stack(preds, dim=0)
batch_confusion = torch.zeros((no_of_classes, no_of_classes),
dtype=torch.int64)
num_classes = no_of_classes
def check_shape(y, y_pred):
if y_pred.ndimension() < 2:
raise ValueError(
"y_pred must have shape (batch_size, num_categories, ...), "
"but given {}".format(y_pred.shape))
if y_pred.shape[1] != num_classes:
raise ValueError(
"y_pred ({}) does not have correct number of categories: {} vs {}"
.format(y_pred.shape, y_pred.shape[1], num_classes))
if not (y.ndimension() + 1 == y_pred.ndimension()):
raise ValueError(
"y_pred must have shape (batch_size, num_categories, ...) and y must have "
"shape of (batch_size, ...), "
"but given {} vs {}.".format(y.shape, y_pred.shape))
y_shape = y.shape
y_pred_shape = y_pred.shape
if y.ndimension() + 1 == y_pred.ndimension():
y_pred_shape = (y_pred_shape[0], ) + y_pred_shape[2:]
if y_shape != y_pred_shape:
raise ValueError("y and y_pred must have compatible shapes.")
check_shape(gts, preds)
# target is (batch_size, ...)
preds = torch.argmax(preds, dim=1).flatten()
gts = gts.flatten()
target_mask = (gts >= 0) & (gts < num_classes)
gts = gts[target_mask]
preds = preds[target_mask]
indices = num_classes * gts + preds
m = torch.bincount(indices, minlength=num_classes**2).reshape(
num_classes, num_classes)
batch_confusion += m.to(batch_confusion)
batch_confusion = batch_confusion.cpu().numpy().astype(np.float64)
inter = np.diag(batch_confusion)
union = np.sum(batch_confusion, 0) + np.sum(batch_confusion,
1) - np.diag(batch_confusion)
union = np.maximum(union, 1)
return inter / union, batch_confusion
def _confusion_matrix_update(
preds: torch.Tensor, target: torch.Tensor, num_classes: int, threshold: float = 0.5
) -> torch.Tensor:
preds, target, mode = _input_format_classification(preds, target, threshold)
if mode not in (DataType.BINARY, DataType.MULTILABEL):
preds = preds.argmax(dim=1)
target = target.argmax(dim=1)
unique_mapping = (target.view(-1) * num_classes + preds.view(-1)).to(torch.long)
bins = torch.bincount(unique_mapping, minlength=num_classes**2)
confmat = bins.reshape(num_classes, num_classes)
return confmat
def loss_weights(dataset, agg_mask: np.ndarray,
device: torch.device) -> torch.tensor:
""" These weights are designed to compensate for class imabalance in the dataset (negated effects if class has
undergone oversampling or undersampling) """
imb_wc = torch.bincount(
dataset.ndata["y"][agg_mask], minlength=int(
dataset.ndata["y"].max())).float().clamp(
min=1e-10, max=1e10) / dataset.ndata["y"][agg_mask].shape[0]
weights = (1 / imb_wc) / (sum(1 / imb_wc))
return weights.to(device)
def f_torch_radial_profile(img, center=(None,None)):
''' Module to compute radial profile of a 2D image
Bincount causes issues with backprop, so not using this code
'''
y,x=torch.meshgrid(torch.arange(0,img.shape[0]),torch.arange(0,img.shape[1])) # Get a grid of x and y values
if center[0]==None and center[1]==None:
center = torch.Tensor([(x.max()-x.min())/2.0, (y.max()-y.min())/2.0]) # compute centers
# get radial values of every pair of points
r = torch.sqrt((x - center[0])**2 + (y - center[1])**2)
r= r.int()
# print(r.shape,img.shape)
# Compute histogram of r values
tbin=torch.bincount(torch.reshape(r,(-1,)),weights=torch.reshape(img,(-1,)).type(torch.DoubleTensor))
nr = torch.bincount(torch.reshape(r,(-1,)))
radialprofile = tbin / nr
return radialprofile[1:-1]
def get_cc_partition(self) -> Partition:
labels = skimage.measure.label(self.index_matrix > 0,
connectivity=2,
background=0,
return_num=False)
membership = torch.tensor(
labels, dtype=torch.long,
device=self.device)[self.i_coordinate_fg_pixel,
self.j_coordinate_fg_pixel]
return Partition(membership=membership,
sizes=torch.bincount(membership))
def class_imbalance_sampler(targets, segmentation_threshold):
if len(targets.shape) > 1: # if posed as segmentation task
targets = targets.sum(axis=1) / targets.shape[1]
targets = targets > segmentation_threshold
targets = tensor(targets).long().squeeze()
class_count = torch.bincount(targets)
weighting = tensor(1.) / class_count.float()
weights = weighting[targets]
sampler = WeightedRandomSampler(weights, len(targets))
return sampler
def _generate_matrix(self, gt, pr):
target_mask = (gt >= 0) & (gt < self.num_classes)
gt = gt[target_mask]
pr = pr[target_mask]
gt = gt.long()
indices = self.num_classes * gt + pr
conf = torch.bincount(indices, minlength=self.num_classes**2)
conf = conf.reshape(self.num_classes, self.num_classes)
return conf.float()