Safran/PointMLP
1
0
Fork 0
Code Issues Pull requests Packages Projects Releases Wiki Activity

update

Browse source
This commit is contained in:
Xu Ma 2021-10-04 12:39:33 -04:00
parent 8e5c90481a
commit c60695ab1b
3 changed files with 366 additions and 309 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

10
part_segmentation/data.py
View file

@ -1,13 +1,3 @@
"""
@Author: Yue Wang
@Contact: yuewangx@mit.edu
@File: data.py
@Time: 2018/10/13 6:21 PM
Modified by
@Author: Tiange Xiang
@Contact: txia7609@uni.sydney.edu.au
@Time: 2021/1/21 3:10 PM
"""
import os import os

615
part_segmentation/main.py
View file

@ -1,367 +1,440 @@
"""
Usage:
python main.py --model CurveNet --exp_name=demo1
@Author: An Tao
@Contact: ta19@mails.tsinghua.edu.cn
@File: main_partseg.py
@Time: 2019/12/31 11:17 AM
Modified by
@Author: Tiange Xiang
@Contact: txia7609@uni.sydney.edu.au
@Time: 2021/01/21 3:10 PM
"""
from __future__ import print_function from __future__ import print_function
import os import os
import datetime
import argparse import argparse
import torch import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim import torch.optim as optim
from torch.optim.lr_scheduler import CosineAnnealingLR, StepLR, MultiStepLR from torch.optim.lr_scheduler import CosineAnnealingLR, StepLR
from data import ShapeNetPart from util.data_util import PartNormalDataset
import models as models import torch.nn.functional as F
import torch.nn as nn
import model as models
import numpy as np import numpy as np
from torch.utils.data import DataLoader from torch.utils.data import DataLoader
from util import cal_loss, IOStream from util.util import to_categorical, compute_overall_iou, IOStream
import sklearn.metrics as metrics from tqdm import tqdm
from collections import defaultdict
from torch.autograd import Variable
import random
classes_str = ['aero','bag','cap','car','chair','ear','guitar','knife','lamp','lapt','moto','mug','Pistol','rock','stake','table']
seg_num = [4, 2, 2, 4, 4, 3, 3, 2, 4, 2, 6, 2, 3, 3, 3, 3]
index_start = [0, 4, 6, 8, 12, 16, 19, 22, 24, 28, 30, 36, 38, 41, 44, 47]
def _init_(): def _init_():
# fix random seed
if args.seed is not None:
torch.manual_seed(args.seed)
np.random.seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
torch.cuda.manual_seed(args.seed)
torch.set_printoptions(10)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
os.environ['PYTHONHASHSEED'] = str(args.seed)
# prepare file structures
if not os.path.exists('checkpoints'): if not os.path.exists('checkpoints'):
os.makedirs('checkpoints') os.makedirs('checkpoints')
if not os.path.exists('checkpoints/' + args.exp_name): if not os.path.exists('checkpoints/' + args.exp_name):
os.makedirs('checkpoints/' + args.exp_name) os.makedirs('checkpoints/' + args.exp_name)
if not os.path.exists('checkpoints/'+args.exp_name+'/'+'models'):
os.makedirs('checkpoints/'+args.exp_name+'/'+'models')
def calculate_shape_IoU(pred_np, seg_np, label, class_choice, eva=False): def weight_init(m):
label = label.squeeze() if isinstance(m, torch.nn.Linear):
shape_ious = [] torch.nn.init.xavier_normal_(m.weight)
category = {} if m.bias is not None:
for shape_idx in range(seg_np.shape[0]): torch.nn.init.constant_(m.bias, 0)
if not class_choice: elif isinstance(m, torch.nn.Conv2d):
start_index = index_start[label[shape_idx]] torch.nn.init.xavier_normal_(m.weight)
num = seg_num[label[shape_idx]] if m.bias is not None:
parts = range(start_index, start_index + num) torch.nn.init.constant_(m.bias, 0)
else: elif isinstance(m, torch.nn.Conv1d):
parts = range(seg_num[label[0]]) torch.nn.init.xavier_normal_(m.weight)
part_ious = [] if m.bias is not None:
for part in parts: torch.nn.init.constant_(m.bias, 0)
I = np.sum(np.logical_and(pred_np[shape_idx] == part, seg_np[shape_idx] == part)) elif isinstance(m, torch.nn.BatchNorm2d):
U = np.sum(np.logical_or(pred_np[shape_idx] == part, seg_np[shape_idx] == part)) torch.nn.init.constant_(m.weight, 1)
if U == 0: torch.nn.init.constant_(m.bias, 0)
iou = 1 # If the union of groundtruth and prediction points is empty, then count part IoU as 1 elif isinstance(m, torch.nn.BatchNorm1d):
else: torch.nn.init.constant_(m.weight, 1)
iou = I / float(U) torch.nn.init.constant_(m.bias, 0)
part_ious.append(iou)
shape_ious.append(np.mean(part_ious))
if label[shape_idx] not in category:
category[label[shape_idx]] = [shape_ious[-1]]
else:
category[label[shape_idx]].append(shape_ious[-1])
if eva:
return shape_ious, category
else:
return shape_ious
def train(args, io): def train(args, io):
train_dataset = ShapeNetPart(partition='trainval', num_points=args.num_points, class_choice=args.class_choice)
if (len(train_dataset) < 100):
drop_last = False
else:
drop_last = True
train_loader = DataLoader(train_dataset, num_workers=8, batch_size=args.batch_size, shuffle=True, drop_last=drop_last, pin_memory=True)
test_loader = DataLoader(ShapeNetPart(partition='test', num_points=args.num_points, class_choice=args.class_choice),
num_workers=8, batch_size=args.test_batch_size, shuffle=False, drop_last=False, pin_memory=True)
# ============= Model ===================
num_part = 50
device = torch.device("cuda" if args.cuda else "cpu") device = torch.device("cuda" if args.cuda else "cpu")
io.cprint("Let's use " + str(torch.cuda.device_count()) + " GPUs!")
seg_num_all = train_loader.dataset.seg_num_all model = models.__dict__[args.model](num_part).to(device)
seg_start_index = train_loader.dataset.seg_start_index
# create model
model = models.__dict__[args.model]().to(device)
io.cprint(str(model)) io.cprint(str(model))
model = nn.DataParallel(model)
model.apply(weight_init)
model = nn.DataParallel(model)
print("Let's use", torch.cuda.device_count(), "GPUs!")
'''Resume or not'''
if args.resume:
state_dict = torch.load("checkpoints/%s/best_insiou_model.pth" % args.exp_name,
map_location=torch.device('cpu'))['model']
for k in state_dict.keys():
if 'module' not in k:
from collections import OrderedDict
new_state_dict = OrderedDict()
for k in state_dict:
new_state_dict['module.' + k] = state_dict[k]
state_dict = new_state_dict
break
model.load_state_dict(state_dict)
print("Resume training model...")
print(torch.load("checkpoints/%s/best_insiou_model.pth" % args.exp_name).keys())
else:
print("Training from scratch...")
# =========== Dataloader =================
train_data = PartNormalDataset(npoints=2048, split='trainval', normalize=False)
print("The number of training data is:%d", len(train_data))
test_data = PartNormalDataset(npoints=2048, split='test', normalize=False)
print("The number of test data is:%d", len(test_data))
train_loader = DataLoader(train_data, batch_size=args.batch_size, shuffle=True, num_workers=args.workers,
drop_last=True)
test_loader = DataLoader(test_data, batch_size=args.test_batch_size, shuffle=False, num_workers=args.workers,
drop_last=False)
# ============= Optimizer ================
if args.use_sgd: if args.use_sgd:
print("Use SGD") print("Use SGD")
opt = optim.SGD(model.parameters(), lr=args.lr*100, momentum=args.momentum, weight_decay=1e-4) opt = optim.SGD(model.parameters(), lr=args.lr*100, momentum=args.momentum, weight_decay=0)
else: else:
print("Use Adam") print("Use Adam")
opt = optim.Adam(model.parameters(), lr=args.lr, weight_decay=1e-4) opt = optim.Adam(model.parameters(), lr=args.lr, betas=(0.9, 0.999), eps=1e-08, weight_decay=0)
if args.scheduler == 'cos': if args.scheduler == 'cos':
if args.use_sgd: print("Use CosLR")
eta_min = args.lr/5.0 scheduler = CosineAnnealingLR(opt, args.epochs, eta_min=args.lr if args.use_sgd else args.lr / 100)
else: else:
eta_min = args.lr/100.0 print("Use StepLR")
scheduler = CosineAnnealingLR(opt, args.epochs, eta_min=eta_min) scheduler = StepLR(opt, step_size=args.step, gamma=0.5)
elif args.scheduler == 'step':
scheduler = MultiStepLR(opt, [140, 180], gamma=0.1) # ============= Training =================
criterion = cal_loss best_acc = 0
best_class_iou = 0
best_instance_iou = 0
num_part = 50
num_classes = 16
best_test_iou = 0
for epoch in range(args.epochs): for epoch in range(args.epochs):
####################
# Train train_epoch(train_loader, model, opt, scheduler, epoch, num_part, num_classes, io)
####################
train_time_cost = datetime.datetime.now() test_metrics, total_per_cat_iou = test_epoch(test_loader, model, epoch, num_part, num_classes, io)
# 1. when get the best accuracy, save the model:
if test_metrics['accuracy'] > best_acc:
best_acc = test_metrics['accuracy']
io.cprint('Max Acc:%.5f' % best_acc)
state = {
'model': model.module.state_dict() if torch.cuda.device_count() > 1 else model.state_dict(),
'optimizer': opt.state_dict(), 'epoch': epoch, 'test_acc': best_acc}
torch.save(state, 'checkpoints/%s/best_acc_model.pth' % args.exp_name)
# 2. when get the best instance_iou, save the model:
if test_metrics['shape_avg_iou'] > best_instance_iou:
best_instance_iou = test_metrics['shape_avg_iou']
io.cprint('Max instance iou:%.5f' % best_instance_iou)
state = {
'model': model.module.state_dict() if torch.cuda.device_count() > 1 else model.state_dict(),
'optimizer': opt.state_dict(), 'epoch': epoch, 'test_instance_iou': best_instance_iou}
torch.save(state, 'checkpoints/%s/best_insiou_model.pth' % args.exp_name)
# 3. when get the best class_iou, save the model:
# first we need to calculate the average per-class iou
class_iou = 0
for cat_idx in range(16):
class_iou += total_per_cat_iou[cat_idx]
avg_class_iou = class_iou / 16
if avg_class_iou > best_class_iou:
best_class_iou = avg_class_iou
# print the iou of each class:
for cat_idx in range(16):
io.cprint(classes_str[cat_idx] + ' iou: ' + str(total_per_cat_iou[cat_idx]))
io.cprint('Max class iou:%.5f' % best_class_iou)
state = {
'model': model.module.state_dict() if torch.cuda.device_count() > 1 else model.state_dict(),
'optimizer': opt.state_dict(), 'epoch': epoch, 'test_class_iou': best_class_iou}
torch.save(state, 'checkpoints/%s/best_clsiou_model.pth' % args.exp_name)
# report best acc, ins_iou, cls_iou
io.cprint('Final Max Acc:%.5f' % best_acc)
io.cprint('Final Max instance iou:%.5f' % best_instance_iou)
io.cprint('Final Max class iou:%.5f' % best_class_iou)
# save last model
state = {
'model': model.module.state_dict() if torch.cuda.device_count() > 1 else model.state_dict(),
'optimizer': opt.state_dict(), 'epoch': args.epochs - 1, 'test_iou': best_instance_iou}
torch.save(state, 'checkpoints/%s/model_ep%d.pth' % (args.exp_name, args.epochs))
def train_epoch(train_loader, model, opt, scheduler, epoch, num_part, num_classes, io):
train_loss = 0.0 train_loss = 0.0
count = 0.0 count = 0.0
accuracy = []
shape_ious = 0.0
metrics = defaultdict(lambda: list())
model.train() model.train()
train_true_cls = []
train_pred_cls = [] for batch_id, (points, label, target, norm_plt) in tqdm(enumerate(train_loader), total=len(train_loader), smoothing=0.9):
train_true_seg = [] batch_size, num_point, _ = points.size()
train_pred_seg = [] points, label, target, norm_plt = Variable(points.float()), Variable(label.long()), Variable(target.long()), \
train_label_seg = [] Variable(norm_plt.float())
for data, label, seg in train_loader: points = points.transpose(2, 1)
seg = seg - seg_start_index norm_plt = norm_plt.transpose(2, 1)
label_one_hot = np.zeros((label.shape[0], 16)) points, label, target, norm_plt = points.cuda(non_blocking=True), label.squeeze(1).cuda(non_blocking=True), \
for idx in range(label.shape[0]): target.cuda(non_blocking=True), norm_plt.cuda(non_blocking=True)
label_one_hot[idx, label[idx]] = 1 # target: b,n
label_one_hot = torch.from_numpy(label_one_hot.astype(np.float32)) seg_pred = model(points, norm_plt, to_categorical(label, num_classes)) # seg_pred: b,n,50
data, label_one_hot, seg = data.to(device), label_one_hot.to(device), seg.to(device) loss = F.nll_loss(seg_pred.contiguous().view(-1, num_part), target.view(-1, 1)[:, 0])
data = data.permute(0, 2, 1)
batch_size = data.size()[0] # instance iou without considering the class average at each batch_size:
batch_shapeious = compute_overall_iou(seg_pred, target, num_part) # list of of current batch_iou:[iou1,iou2,...,iou#b_size]
# total iou of current batch in each process:
batch_shapeious = seg_pred.new_tensor([np.sum(batch_shapeious)], dtype=torch.float64) # same device with seg_pred!!!
# Loss backward
loss = torch.mean(loss)
opt.zero_grad() opt.zero_grad()
seg_pred = model(data, label_one_hot)
seg_pred = seg_pred.permute(0, 2, 1).contiguous()
loss = criterion(seg_pred.view(-1, seg_num_all), seg.view(-1,1).squeeze())
loss.backward() loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1)
opt.step() opt.step()
pred = seg_pred.max(dim=2)[1] # (batch_size, num_points)
count += batch_size # accuracy
seg_pred = seg_pred.contiguous().view(-1, num_part) # b*n,50
target = target.view(-1, 1)[:, 0] # b*n
pred_choice = seg_pred.contiguous().data.max(1)[1] # b*n
correct = pred_choice.eq(target.contiguous().data).sum() # torch.int64: total number of correct-predict pts
# sum
shape_ious += batch_shapeious.item() # count the sum of ious in each iteration
count += batch_size # count the total number of samples in each iteration
train_loss += loss.item() * batch_size train_loss += loss.item() * batch_size
seg_np = seg.cpu().numpy() # (batch_size, num_points) accuracy.append(correct.item()/(batch_size * num_point)) # append the accuracy of each iteration
pred_np = pred.detach().cpu().numpy() # (batch_size, num_points)
train_true_cls.append(seg_np.reshape(-1)) # (batch_size * num_points) # Note: We do not need to calculate per_class iou during training
train_pred_cls.append(pred_np.reshape(-1)) # (batch_size * num_points)
train_true_seg.append(seg_np)
train_pred_seg.append(pred_np)
train_label_seg.append(label.reshape(-1))
if args.scheduler == 'cos': if args.scheduler == 'cos':
scheduler.step() scheduler.step()
elif args.scheduler == 'step': elif args.scheduler == 'step':
if opt.param_groups[0]['lr'] > 1e-5: if opt.param_groups[0]['lr'] > 0.9e-5:
scheduler.step() scheduler.step()
if opt.param_groups[0]['lr'] < 1e-5: if opt.param_groups[0]['lr'] < 0.9e-5:
for param_group in opt.param_groups: for param_group in opt.param_groups:
param_group['lr'] = 1e-5 param_group['lr'] = 0.9e-5
train_true_cls = np.concatenate(train_true_cls) io.cprint('Learning rate: %f' % opt.param_groups[0]['lr'])
train_pred_cls = np.concatenate(train_pred_cls)
train_acc = metrics.accuracy_score(train_true_cls, train_pred_cls)
avg_per_class_acc = metrics.balanced_accuracy_score(train_true_cls, train_pred_cls)
train_true_seg = np.concatenate(train_true_seg, axis=0)
train_pred_seg = np.concatenate(train_pred_seg, axis=0)
train_label_seg = np.concatenate(train_label_seg)
train_ious = calculate_shape_IoU(train_pred_seg, train_true_seg, train_label_seg, args.class_choice)
train_time_cost = int((datetime.datetime.now() - train_time_cost).total_seconds())
outstr = 'Train %d, loss: %.6f, train acc: %.6f, train avg acc: %.6f, train iou: %.6f' % (epoch,
train_loss*1.0/count,
train_acc,
avg_per_class_acc,
np.mean(train_ious))
io.cprint(outstr)
io.cprint(f"Training time: {train_time_cost} seconds.")
#################### metrics['accuracy'] = np.mean(accuracy)
# Test metrics['shape_avg_iou'] = shape_ious * 1.0 / count
####################
test_time_cost = datetime.datetime.now() outstr = 'Train %d, loss: %f, train acc: %f, train ins_iou: %f' % (epoch+1, train_loss * 1.0 / count,
metrics['accuracy'], metrics['shape_avg_iou'])
io.cprint(outstr)
def test_epoch(test_loader, model, epoch, num_part, num_classes, io):
test_loss = 0.0 test_loss = 0.0
count = 0.0 count = 0.0
accuracy = []
shape_ious = 0.0
final_total_per_cat_iou = np.zeros(16).astype(np.float32)
final_total_per_cat_seen = np.zeros(16).astype(np.int32)
metrics = defaultdict(lambda: list())
model.eval() model.eval()
test_true_cls = []
test_pred_cls = [] # label_size: b, means each sample has one corresponding class
test_true_seg = [] for batch_id, (points, label, target, norm_plt) in tqdm(enumerate(test_loader), total=len(test_loader), smoothing=0.9):
test_pred_seg = [] batch_size, num_point, _ = points.size()
test_label_seg = [] points, label, target, norm_plt = Variable(points.float()), Variable(label.long()), Variable(target.long()), \
for data, label, seg in test_loader: Variable(norm_plt.float())
seg = seg - seg_start_index points = points.transpose(2, 1)
label_one_hot = np.zeros((label.shape[0], 16)) norm_plt = norm_plt.transpose(2, 1)
for idx in range(label.shape[0]): points, label, target, norm_plt = points.cuda(non_blocking=True), label.squeeze(1).cuda(non_blocking=True), \
label_one_hot[idx, label[idx]] = 1 target.cuda(non_blocking=True), norm_plt.cuda(non_blocking=True)
label_one_hot = torch.from_numpy(label_one_hot.astype(np.float32)) seg_pred = model(points, norm_plt, to_categorical(label, num_classes)) # b,n,50
data, label_one_hot, seg = data.to(device), label_one_hot.to(device), seg.to(device)
data = data.permute(0, 2, 1) # instance iou without considering the class average at each batch_size:
batch_size = data.size()[0] batch_shapeious = compute_overall_iou(seg_pred, target, num_part) # [b]
seg_pred = model(data, label_one_hot) # per category iou at each batch_size:
seg_pred = seg_pred.permute(0, 2, 1).contiguous()
loss = criterion(seg_pred.view(-1, seg_num_all), seg.view(-1,1).squeeze()) for shape_idx in range(seg_pred.size(0)): # sample_idx
pred = seg_pred.max(dim=2)[1] cur_gt_label = label[shape_idx] # label[sample_idx], denotes current sample belongs to which cat
count += batch_size final_total_per_cat_iou[cur_gt_label] += batch_shapeious[shape_idx] # add the iou belongs to this cat
final_total_per_cat_seen[cur_gt_label] += 1 # count the number of this cat is chosen
# total iou of current batch in each process:
batch_ious = seg_pred.new_tensor([np.sum(batch_shapeious)], dtype=torch.float64) # same device with seg_pred!!!
# prepare seg_pred and target for later calculating loss and acc:
seg_pred = seg_pred.contiguous().view(-1, num_part)
target = target.view(-1, 1)[:, 0]
# Loss
loss = F.nll_loss(seg_pred.contiguous(), target.contiguous())
# accuracy:
pred_choice = seg_pred.data.max(1)[1] # b*n
correct = pred_choice.eq(target.data).sum() # torch.int64: total number of correct-predict pts
loss = torch.mean(loss)
shape_ious += batch_ious.item() # count the sum of ious in each iteration
count += batch_size # count the total number of samples in each iteration
test_loss += loss.item() * batch_size test_loss += loss.item() * batch_size
seg_np = seg.cpu().numpy() accuracy.append(correct.item() / (batch_size * num_point)) # append the accuracy of each iteration
pred_np = pred.detach().cpu().numpy()
test_true_cls.append(seg_np.reshape(-1)) for cat_idx in range(16):
test_pred_cls.append(pred_np.reshape(-1)) if final_total_per_cat_seen[cat_idx] > 0: # indicating this cat is included during previous iou appending
test_true_seg.append(seg_np) final_total_per_cat_iou[cat_idx] = final_total_per_cat_iou[cat_idx] / final_total_per_cat_seen[cat_idx] # avg class iou across all samples
test_pred_seg.append(pred_np)
test_label_seg.append(label.reshape(-1)) metrics['accuracy'] = np.mean(accuracy)
test_true_cls = np.concatenate(test_true_cls) metrics['shape_avg_iou'] = shape_ious * 1.0 / count
test_pred_cls = np.concatenate(test_pred_cls)
test_acc = metrics.accuracy_score(test_true_cls, test_pred_cls) outstr = 'Test %d, loss: %f, test acc: %f test ins_iou: %f' % (epoch + 1, test_loss * 1.0 / count,
avg_per_class_acc = metrics.balanced_accuracy_score(test_true_cls, test_pred_cls) metrics['accuracy'], metrics['shape_avg_iou'])
test_true_seg = np.concatenate(test_true_seg, axis=0)
test_pred_seg = np.concatenate(test_pred_seg, axis=0)
test_label_seg = np.concatenate(test_label_seg)
test_ious = calculate_shape_IoU(test_pred_seg, test_true_seg, test_label_seg, args.class_choice)
test_time_cost = int((datetime.datetime.now() - test_time_cost).total_seconds())
outstr = 'Test %d, loss: %.6f, test acc: %.6f, test avg acc: %.6f, test iou: %.6f, best iou %.6f' % (epoch,
test_loss*1.0/count,
test_acc,
avg_per_class_acc,
np.mean(test_ious), best_test_iou)
io.cprint(outstr) io.cprint(outstr)
io.cprint(f"Testing time: {test_time_cost} seconds.")
if np.mean(test_ious) >= best_test_iou: return metrics, final_total_per_cat_iou
best_test_iou = np.mean(test_ious)
torch.save(model.state_dict(), 'checkpoints/%s/models/model.t7' % args.exp_name)
def test(args, io): def test(args, io):
test_loader = DataLoader(ShapeNetPart(partition='test', num_points=args.num_points, class_choice=args.class_choice), # Dataloader
batch_size=args.test_batch_size, shuffle=True, drop_last=False) test_data = PartNormalDataset(npoints=2048, split='test', normalize=False)
print("The number of test data is:%d", len(test_data))
device = torch.device("cuda" if args.cuda else "cpu") test_loader = DataLoader(test_data, batch_size=args.test_batch_size, shuffle=False, num_workers=args.workers,
drop_last=False)
# Try to load models # Try to load models
seg_start_index = test_loader.dataset.seg_start_index num_part = 50
model = models.__dict__[args.model]().to(device) device = torch.device("cuda" if args.cuda else "cpu")
model = nn.DataParallel(model)
model.load_state_dict(torch.load(args.model_path))
model = model.eval() model = models.__dict__[args.model](num_part).to(device)
test_acc = 0.0 io.cprint(str(model))
test_true_cls = []
test_pred_cls = []
test_true_seg = []
test_pred_seg = []
test_label_seg = []
category = {}
for data, label, seg in test_loader:
seg = seg - seg_start_index
label_one_hot = np.zeros((label.shape[0], 16))
for idx in range(label.shape[0]):
label_one_hot[idx, label[idx]] = 1
label_one_hot = torch.from_numpy(label_one_hot.astype(np.float32))
data, label_one_hot, seg = data.to(device), label_one_hot.to(device), seg.to(device)
data = data.permute(0, 2, 1)
seg_pred = model(data, label_one_hot)
seg_pred = seg_pred.permute(0, 2, 1).contiguous()
pred = seg_pred.max(dim=2)[1]
seg_np = seg.cpu().numpy()
pred_np = pred.detach().cpu().numpy()
test_true_cls.append(seg_np.reshape(-1))
test_pred_cls.append(pred_np.reshape(-1))
test_true_seg.append(seg_np)
test_pred_seg.append(pred_np)
test_label_seg.append(label.reshape(-1))
test_true_cls = np.concatenate(test_true_cls) from collections import OrderedDict
test_pred_cls = np.concatenate(test_pred_cls) state_dict = torch.load("checkpoints/%s/best_%s_model.pth" % (args.exp_name, args.model_type),
test_acc = metrics.accuracy_score(test_true_cls, test_pred_cls) map_location=torch.device('cpu'))['model']
avg_per_class_acc = metrics.balanced_accuracy_score(test_true_cls, test_pred_cls)
test_true_seg = np.concatenate(test_true_seg, axis=0) new_state_dict = OrderedDict()
test_pred_seg = np.concatenate(test_pred_seg, axis=0) for layer in state_dict:
test_label_seg = np.concatenate(test_label_seg) new_state_dict[layer.replace('module.', '')] = state_dict[layer]
test_ious,category = calculate_shape_IoU(test_pred_seg, test_true_seg, test_label_seg, args.class_choice, eva=True) model.load_state_dict(new_state_dict)
outstr = 'Test :: test acc: %.6f, test avg acc: %.6f, test iou: %.6f' % (test_acc,
avg_per_class_acc, model.eval()
np.mean(test_ious)) num_part = 50
num_classes = 16
metrics = defaultdict(lambda: list())
hist_acc = []
shape_ious = []
total_per_cat_iou = np.zeros((16)).astype(np.float32)
total_per_cat_seen = np.zeros((16)).astype(np.int32)
for batch_id, (points, label, target, norm_plt) in tqdm(enumerate(test_loader), total=len(test_loader), smoothing=0.9):
batch_size, num_point, _ = points.size()
points, label, target, norm_plt = Variable(points.float()), Variable(label.long()), Variable(target.long()), Variable(norm_plt.float())
points = points.transpose(2, 1)
norm_plt = norm_plt.transpose(2, 1)
points, label, target, norm_plt = points.cuda(non_blocking=True), label.squeeze().cuda(
non_blocking=True), target.cuda(non_blocking=True), norm_plt.cuda(non_blocking=True)
with torch.no_grad():
seg_pred = model(points, norm_plt, to_categorical(label, num_classes)) # b,n,50
# instance iou without considering the class average at each batch_size:
batch_shapeious = compute_overall_iou(seg_pred, target, num_part) # [b]
shape_ious += batch_shapeious # iou +=, equals to .append
# per category iou at each batch_size:
for shape_idx in range(seg_pred.size(0)): # sample_idx
cur_gt_label = label[shape_idx] # label[sample_idx]
total_per_cat_iou[cur_gt_label] += batch_shapeious[shape_idx]
total_per_cat_seen[cur_gt_label] += 1
# accuracy:
seg_pred = seg_pred.contiguous().view(-1, num_part)
target = target.view(-1, 1)[:, 0]
pred_choice = seg_pred.data.max(1)[1]
correct = pred_choice.eq(target.data).cpu().sum()
metrics['accuracy'].append(correct.item() / (batch_size * num_point))
hist_acc += metrics['accuracy']
metrics['accuracy'] = np.mean(hist_acc)
metrics['shape_avg_iou'] = np.mean(shape_ious)
for cat_idx in range(16):
if total_per_cat_seen[cat_idx] > 0:
total_per_cat_iou[cat_idx] = total_per_cat_iou[cat_idx] / total_per_cat_seen[cat_idx]
# First we need to calculate the iou of each class and the avg class iou:
class_iou = 0
for cat_idx in range(16):
class_iou += total_per_cat_iou[cat_idx]
io.cprint(classes_str[cat_idx] + ' iou: ' + str(total_per_cat_iou[cat_idx])) # print the iou of each class
avg_class_iou = class_iou / 16
outstr = 'Test :: test acc: %f test class mIOU: %f, test instance mIOU: %f' % (metrics['accuracy'], avg_class_iou, metrics['shape_avg_iou'])
io.cprint(outstr) io.cprint(outstr)
results = []
for key in category.keys():
results.append((int(key), np.mean(category[key]), len(category[key])))
results.sort(key=lambda x:x[0])
for re in results:
io.cprint('idx: %d mIoU: %.3f num: %d' % (re[0], re[1], re[2]))
if __name__ == "__main__": if __name__ == "__main__":
# Training settings # Training settings
parser = argparse.ArgumentParser(description='Point Cloud Part Segmentation') parser = argparse.ArgumentParser(description='3D Shape Part Segmentation')
parser.add_argument('--model', type=str, default='CurveNet') parser.add_argument('--model', type=str, default='PointMLP1')
parser.add_argument('--exp_name', type=str, default='exp', metavar='N', parser.add_argument('--exp_name', type=str, default='demo1', metavar='N',
help='Name of the experiment') help='Name of the experiment')
parser.add_argument('--dataset', type=str, default='shapenetpart', metavar='N',
choices=['shapenetpart'])
parser.add_argument('--class_choice', type=str, default=None, metavar='N',
choices=['airplane', 'bag', 'cap', 'car', 'chair',
'earphone', 'guitar', 'knife', 'lamp', 'laptop',
'motor', 'mug', 'pistol', 'rocket', 'skateboard', 'table'])
parser.add_argument('--batch_size', type=int, default=32, metavar='batch_size', parser.add_argument('--batch_size', type=int, default=32, metavar='batch_size',
help='Size of batch)') help='Size of batch)')
parser.add_argument('--test_batch_size', type=int, default=16, metavar='batch_size', parser.add_argument('--test_batch_size', type=int, default=32, metavar='batch_size',
help='Size of batch)') help='Size of batch)')
parser.add_argument('--epochs', type=int, default=200, metavar='N', parser.add_argument('--epochs', type=int, default=350, metavar='N',
help='number of episode to train') help='number of episode to train')
parser.add_argument('--seed', type=int) parser.add_argument('--use_sgd', type=bool, default=False,
parser.add_argument('--use_sgd', type=bool, default=True,
help='Use SGD') help='Use SGD')
parser.add_argument('--lr', type=float, default=0.0005, metavar='LR', parser.add_argument('--scheduler', type=str, default='step',
help='learning rate (default: 0.001, 0.1 if using sgd)') help='lr scheduler')
parser.add_argument('--step', type=int, default=40,
help='lr decay step')
parser.add_argument('--lr', type=float, default=0.003, metavar='LR',
help='learning rate')
parser.add_argument('--momentum', type=float, default=0.9, metavar='M', parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
help='SGD momentum (default: 0.9)') help='SGD momentum (default: 0.9)')
parser.add_argument('--scheduler', type=str, default='step', metavar='N',
choices=['cos', 'step'],
help='Scheduler to use, [cos, step]')
parser.add_argument('--no_cuda', type=bool, default=False, parser.add_argument('--no_cuda', type=bool, default=False,
help='enables CUDA training') help='enables CUDA training')
parser.add_argument('--manual_seed', type=int, metavar='S',
help='random seed (default: 1)')
parser.add_argument('--eval', type=bool, default=False, parser.add_argument('--eval', type=bool, default=False,
help='evaluate the model') help='evaluate the model')
parser.add_argument('--num_points', type=int, default=2048, parser.add_argument('--num_points', type=int, default=2048,
help='num of points to use') help='num of points to use')
parser.add_argument('--model_path', type=str, default='', metavar='N', parser.add_argument('--workers', type=int, default=12)
help='Pretrained model path') parser.add_argument('--resume', type=bool, default=False,
help='Resume training or not')
parser.add_argument('--model_type', type=str, default='insiou',
help='choose to test the best insiou/clsiou/acc model (options: insiou, clsiou, acc)')
args = parser.parse_args() args = parser.parse_args()
time_str = str(datetime.datetime.now().strftime('-%Y%m%d%H%M%S'))
if args.exp_name is None:
args.exp_name = time_str
args.exp_name = args.model+"_"+args.exp_name args.exp_name = args.model+"_"+args.exp_name
_init_() _init_()
if args.eval: if not args.eval:
io = IOStream('checkpoints/' + args.exp_name + '/eval.log') io = IOStream('checkpoints/' + args.exp_name + '/%s_train.log' % (args.exp_name))
else: else:
io = IOStream('checkpoints/' + args.exp_name + '/run.log') io = IOStream('checkpoints/' + args.exp_name + '/%s_test.log' % (args.exp_name))
io.cprint(str(args)) io.cprint(str(args))
io.cprint('random seed is: ' + str(args.seed))
if args.manual_seed is not None:
random.seed(args.manual_seed)
np.random.seed(args.manual_seed)
torch.manual_seed(args.manual_seed)
args.cuda = not args.no_cuda and torch.cuda.is_available() args.cuda = not args.no_cuda and torch.cuda.is_available()
if args.cuda: if args.cuda:
io.cprint( io.cprint('Using GPU')
'Using GPU : ' + str(torch.cuda.current_device()) + ' from ' + str(torch.cuda.device_count()) + ' devices') if args.manual_seed is not None:
torch.cuda.manual_seed(args.manual_seed)
torch.cuda.manual_seed_all(args.manual_seed)
else: else:
io.cprint('Using CPU') io.cprint('Using CPU')
if not args.eval: if not args.eval:
train(args, io) train(args, io)
else: else:
with torch.no_grad():
test(args, io) test(args, io)

6
part_segmentation/util.py
View file

@ -1,9 +1,3 @@
"""
@Author: Yue Wang
@Contact: yuewangx@mit.edu
@File: util
@Time: 4/5/19 3:47 PM
"""
import numpy as np import numpy as np