def graphEval(X, truth_spec, true_sp, true_cc, true_dd, true_bc):
sp_emd_cur = []
cc_emd_cur = []
dd_emd_cur = []
assorts_cur = []
spec_l2_cur = []
spec_l2_lin_cur = []
bc_emd_cur = []
for j in range(20):
A = gnp(X)
G = nx.from_numpy_matrix(A)
print(nx.is_connected(G))
if not nx.is_connected(G):
Gc = max(nx.connected_component_subgraphs(G), key=len)
print(len(Gc.nodes()))
sp = utils.sp(A)
cc = utils.cc(A)
dd = utils.degree_sequence(A)
spec_weight_l2 = l2_exp_weight(truth_spec, utils.spectrum(A))
spec_weight_l2_lin = l2_lin_weight(truth_spec, utils.spectrum(A))
bc = sorted(nx.betweenness_centrality(G).values())
sp_emd_cur.append(utils.emd(sp, true_sp))
cc_emd_cur.append(utils.emd(cc, true_cc))
dd_emd_cur.append(utils.emd(dd, true_dd))
assorts_cur.append(nx.degree_assortativity_coefficient(G))
spec_l2_cur.append(spec_weight_l2)
spec_l2_lin_cur.append(spec_weight_l2_lin)
bc_emd_cur.append(utils.emd(bc, true_bc))
return np.mean(sp_emd_cur), np.mean(cc_emd_cur), np.mean(
dd_emd_cur), np.mean(assorts_cur), np.mean(spec_l2_cur), np.mean(
bc_emd_cur), np.mean(spec_l2_lin_cur)
def global_comp(A, X, sps_truth, spec):
sp_emds = []
specs = []
ccs = []
for i in range(100):
sampled_graph = gnp(X)
sps_gen = utils.sp(sampled_graph)
sp_emds.append(utils.emd(sps_truth, sps_gen))
specs.append((spec - utils.specGap(sampled_graph))**2)
ccs.append(3 * utils.statistics_triangle_count(sampled_graph) /
utils.statistics_claw_count(sampled_graph))
stats = {}
stats['shortest path emds'] = sp_emds
stats['spec gaps'] = specs
stats['clustering coefficients'] = ccs
print('Shortest path')
print(np.mean(sp_emds))
print(np.std(sp_emds))
print('Specs')
print(np.mean(specs))
print(np.std(specs))
print('Clusterig Coefficient')
print(np.mean(ccs))
print(np.std(ccs))
print('True CC')
tc = 3 * utils.statistics_triangle_count(
np.array(A)) / utils.statistics_claw_count(np.array(A))
print(tc)
print('ABS diff')
print(tc - np.mean(ccs))
return stats
def global_comp(A,X,sps_truth,spec):
sp_emds = []
specs = []
for i in range(100):
sampled_graph = gnp(X)
sps_gen = utils.sp(sampled_graph)
sp_emds.append(utils.emd(sps_truth,sps_gen))
specs.append((spec-utils.specGap(sampled_graph))**2)
return np.mean(sp_emds), np.std(sp_emds), np.mean(specs), np.std(specs)
def print_evaluate(A_matrix,sampled_graph): stats = utils.compute_graph_statistics(sampled_graph) sps_truth = utils.sp(A_matrix) sps_gen = utils.sp(sampled_graph) stats['sp_emd']= utils.emd(sps_truth,sps_gen) s = utils.specGap(A_matrix) stats['spec_gap']=(s-utils.specGap(sampled_graph))**2 print(stats)
def gen_graphs(X_from_walks, target):
X = genExpected_fromWalks(X_from_walks, target.sum())
print(X)
sps_truth = utils.sp(target)
sp_emds = []
sgl2s = []
for i in range(20):
sampled_graph = gnp(X)
sps_gen = utils.sp(sampled_graph)
sp_emd = utils.emd(sps_truth, sps_gen)
sgl2 = (utils.specGap(target) - utils.specGap(sampled_graph))**2
sp_emds.append(sp_emd)
sgl2s.append(sgl2)
return sp_emds, sgl2s
cc_emd_cur = []
dd_emd_cur = []
assorts_cur = []
spec_l2_cur = []
bc_emd_cur = []
for j in range(20):
A = gnp(X)
G = nx.from_numpy_matrix(A)
sp = utils.sp(A)
cc = utils.cc(A)
dd = utils.degree_sequence(A)
spec_weight_l2 = l2_exp_weight(truth_spec, utils.spectrum(A))
bc = sorted(nx.betweenness_centrality(G).values())
sp_emd_cur.append(utils.emd(sp, true_sp))
cc_emd_cur.append(utils.emd(cc, true_cc))
dd_emd_cur.append(utils.emd(dd, true_dd))
assorts_cur.append(nx.degree_assortativity_coefficient(G))
spec_l2_cur.append(spec_weight_l2)
bc_emd_cur.append(utils.emd(bc, true_bc))
sp_emds.append(np.mean(sp_emd_cur))
cc_emds.append(np.mean(cc_emd_cur))
dd_emds.append(np.mean(dd_emd_cur))
assorts.append(np.mean(assorts_cur))
true_assorts.append(true_assort)
spectrum_weighted_distances.append(np.mean(spec_l2_cur))
bc_emds.append(np.mean(bc_emd_cur))
index = [i * step for i in range(start, k + 1)]
def run(net, loader, optimizer, train=False, epoch=0, pool=None):
writer = train_writer
if train:
net.train()
prefix = "train"
torch.set_grad_enabled(True)
else:
net.eval()
prefix = "test"
torch.set_grad_enabled(False)
if args.export_dir:
true_export = []
pred_export = []
iters_per_epoch = len(loader)
loader = tqdm(
loader,
ncols=0,
desc="{1} E{0:02d}".format(epoch, "train" if train else "test "),
)
for i, sample in enumerate(loader, start=epoch * iters_per_epoch):
# input is either a set or an image
input, target_set, target_mask = map(lambda x: x.cuda(), sample)
# forward evaluation through the network
(progress, masks, evals,
gradn), (y_enc, y_label) = net(input, target_set, target_mask)
progress_only = progress
# if using mask as feature, concat mask feature into progress
if args.mask_feature:
target_set = torch.cat(
[target_set, target_mask.unsqueeze(dim=1)], dim=1)
progress = [
torch.cat([p, m.unsqueeze(dim=1)], dim=1)
for p, m in zip(progress, masks)
]
if args.loss == "chamfer":
# dim 0 is over the inner iteration steps
# target set is broadcasted over dim 0
set_loss = utils.chamfer_loss(torch.stack(progress),
target_set.unsqueeze(0))
elif args.loss == "hungarian":
set_loss = utils.hungarian_loss(progress[-1],
target_set,
thread_pool=pool).unsqueeze(0)
elif args.loss == "emd":
set_loss = utils.emd(progress[-1], target_set).unsqueeze(0)
# Only use representation loss with DSPN and when doing general
# supervised prediction, not when auto-encoding
if args.supervised and not args.baseline:
repr_loss = args.huber_repr * F.smooth_l1_loss(y_enc, y_label)
loss = set_loss.mean() + repr_loss.mean()
else:
loss = set_loss.mean()
# restore progress variable to not contain masks for correct
# exporting
progress = progress_only
# Outer optim step
if train:
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Tensorboard tracking of metrics for debugging
tracked_last = tracker.update(f"{prefix}_last",
set_loss[-1].item())
tracked_loss = tracker.update(f"{prefix}_loss", loss.item())
if train:
writer.add_scalar("metric/set-loss",
loss.item(),
global_step=i)
writer.add_scalar("metric/set-last",
set_loss[-1].mean().item(),
global_step=i)
if not args.baseline:
writer.add_scalar("metric/eval-first",
evals[0].mean().item(),
global_step=i)
writer.add_scalar("metric/eval-last",
evals[-1].mean().item(),
global_step=i)
writer.add_scalar("metric/max-inner-grad-norm",
max(g.item() for g in gradn),
global_step=i)
writer.add_scalar("metric/mean-inner-grad-norm",
sum(g.item()
for g in gradn) / len(gradn),
global_step=i)
if args.supervised:
writer.add_scalar("metric/repr_loss",
repr_loss.item(),
global_step=i)
# Print current progress to progress bar
fmt = "{:.6f}".format
loader.set_postfix(last=fmt(tracked_last),
loss=fmt(tracked_loss),
bad=fmt(evals[-1].detach().cpu().item() *
1000) if not args.baseline else 0)
if args.export_dir:
# export last inner optim of each input as csv
# (one input per row)
if args.export_csv:
# the second to last element are the last of the
# inner optim
for batch_i, p in enumerate(progress[-2]):
img_id = i * args.batch_size + batch_i
names.append(loader.iterable.dataset.get_fname(img_id))
m = masks[-2][batch_i]
m = m.cpu().detach().numpy().astype(bool)
p = p.cpu().detach().numpy()
p = p[:, m]
sample_preds = [
p[k % 2, k // 2] for k in range(p.shape[1] * 2)
]
# remove values according to mask and add zeros to the
# end in stead
sample_preds += [0] * (len(m) * 2 - len(sample_preds))
predictions.append(sample_preds)
true_mask = target_set[batch_i, 2, :].cpu().detach()
true_mask = true_mask.numpy().astype(bool)
trues = target_set[batch_i, :2, :]
trues = trues.cpu().detach().numpy()
t = trues[:, true_mask]
t = [t[k % 2, k // 2] for k in range(t.shape[1] * 2)]
t += [0] * (len(true_mask) * 2 - len(t))
export_targets.append(t)
# Store predictions to be exported
else:
if len(true_export) < args.export_n:
for p, m in zip(target_set, target_mask):
true_export.append(p.detach().cpu())
progress_steps = []
for pro, ms in zip(progress, masks):
# pro and ms are one step of the inner optim
# score boxes contains the list of predicted
# elements for one step
score_boxes = []
for p, m in zip(pro.cpu().detach(),
ms.cpu().detach()):
score_box = torch.cat([m.unsqueeze(0), p],
dim=0)
score_boxes.append(score_box)
progress_steps.append(score_boxes)
for b in zip(*progress_steps):
pred_export.append(b)
# Plot predictions in Tensorboard
if args.show and epoch % args.show_skip == 0 and not train:
name = f"set/epoch-{epoch}/img-{i}"
# thresholded set
progress.append(progress[-1])
masks.append((masks[-1] > 0.5).float())
# target set
if args.mask_feature:
# target set is augmented with masks, so remove them
progress.append(target_set[:, :-1])
else:
progress.append(target_set)
masks.append(target_mask)
# intermediate sets
for j, (s, ms) in enumerate(zip(progress, masks)):
if args.dataset == "clevr-state":
continue
if args.dataset.startswith("clevr"):
threshold = 0.5
else:
threshold = None
s, ms = utils.scatter_masked(
s,
ms,
binned=args.dataset.startswith("clevr"),
threshold=threshold)
if j != len(progress) - 1:
tag_name = f"{name}"
else:
tag_name = f"{name}-target"
if args.dataset == "clevr-box":
img = input[0].detach().cpu()
writer.add_image_with_boxes(tag_name,
img,
s.transpose(0, 1),
global_step=j)
elif args.dataset == "cats" \
or args.dataset == "wflw" \
or args.dataset == "merged":
img = input[0].detach().cpu()
fig = plt.figure()
plt.scatter(s[0, :] * 128, s[1, :] * 128)
plt.imshow(np.transpose(img, (1, 2, 0)))
writer.add_figure(tag_name, fig, global_step=j)
else: # mnist
fig = plt.figure()
y, x = s
y = 1 - y
ms = ms.numpy()
rgba_colors = np.zeros((ms.size, 4))
rgba_colors[:, 2] = 1.0
rgba_colors[:, 3] = ms
plt.scatter(x, y, color=rgba_colors)
plt.axes().set_aspect("equal")
plt.xlim(0, 1)
plt.ylim(0, 1)
writer.add_figure(tag_name, fig, global_step=j)
# Export predictions
if args.export_dir and not args.export_csv:
os.makedirs(f"{args.export_dir}/groundtruths", exist_ok=True)
os.makedirs(f"{args.export_dir}/detections", exist_ok=True)
for i, (gt, dets) in enumerate(zip(true_export, pred_export)):
export_groundtruths_path = os.path.join(
args.export_dir, "groundtruths", f"{i}.txt")
with open(export_groundtruths_path, "w") as fd:
for box in gt.transpose(0, 1):
if (box == 0).all():
continue
s = "box " + " ".join(map(str, box.tolist()))
fd.write(s + "\n")
if args.export_progress:
for step, det in enumerate(dets):
export_progress_path = os.path.join(
args.export_dir, "detections",
f"{i}-step{step}.txt")
with open(export_progress_path, "w") as fd:
for sbox in det.transpose(0, 1):
s = f"box " + " ".join(map(str, sbox.tolist()))
fd.write(s + "\n")
export_path = os.path.join(args.export_dir, "detections",
f"{i}.txt")
with open(export_path, "w") as fd:
for sbox in dets[-1].transpose(0, 1):
s = f"box " + " ".join(map(str, sbox.tolist()))
fd.write(s + "\n")
.format(i * 100))
X_r = np.loadtxt(
'plots/barbell_sameDensity/barbell_walk_mixed/trainingIteration_{}_expectedGraph.txt'
.format(i * 100))
cc_emd_combo_iter = []
cc_emd_fmm_iter = []
cc_emd_reg_iter = []
for j in range(100):
sampled_graph_c = gnp(X_c)
sampled_graph_f = gnp(X_f)
sampled_graph_r = gnp(X_r)
cc_c = utils.cc(sampled_graph_c)
cc_f = utils.cc(sampled_graph_f)
cc_r = utils.cc(sampled_graph_r)
cc_emd_combo_iter.append(utils.emd(truth_cc, cc_c))
cc_emd_fmm_iter.append(utils.emd(truth_cc, cc_f))
cc_emd_reg_iter.append(utils.emd(truth_cc, cc_r))
cc_emd_combo.append(np.mean(cc_emd_combo_iter))
cc_emd_fmm.append(np.mean(cc_emd_fmm_iter))
cc_emd_reg.append(np.mean(cc_emd_reg_iter))
cc_emd_combo_std.append(np.std(cc_emd_combo_iter))
cc_emd_fmm_std.append(np.std(cc_emd_fmm_iter))
cc_emd_reg_std.append(np.std(cc_emd_reg_iter))
plt.errorbar(range(1, k), cc_emd_combo, yerr=cc_emd_combo_std, label='Combo')
plt.errorbar(range(1, k), cc_emd_fmm, yerr=cc_emd_fmm_std, label='FMM')
plt.errorbar(range(1, k), cc_emd_reg, yerr=cc_emd_reg_std, label='Reg Walk')
plt.legend(loc='best')
plt.savefig('plots/barbell_sameDensity/barbell_cc_comp.pdf')
plt.gcf().clear()
