def main():
# Seed all sources of randomness to 0 for reproducibility
np.random.seed(0)
torch.manual_seed(0)
torch.cuda.manual_seed(0)
random.seed(0)
opt = Opts().parse()
# Create data loaders
train_loader, test_loader = create_data_loaders(opt)
# Create nn
model, criterion_hm, criterion_paf, latest_inx = create_model(opt)
model = model.cuda()
criterion_hm = criterion_hm.cuda()
criterion_paf = criterion_paf.cuda()
# Create optimizer
optimizer = create_optimizer(opt, model)
# Other params
n_epochs = opt.nEpoch
to_train = opt.train
drop_lr = opt.dropLR
val_interval = opt.valInterval
learn_rate = opt.LR
visualize_out = opt.vizOut
# train/ test
train_net(train_loader, test_loader, model, criterion_hm, criterion_paf,
optimizer, n_epochs, val_interval, learn_rate, drop_lr,
opt.saveDir, visualize_out, latest_inx)
def main():
# Seed all sources of randomness to 0 for reproducibility
np.random.seed(0)
torch.manual_seed(0)
torch.cuda.manual_seed(0)
random.seed(0)
opt = Opts().parse()
# Create data loaders
train_loader, test_loader = create_data_loaders(opt)
# Create nn
model, criterion_hm, criterion_paf, latest_inx = create_model(opt)
# model = model.cuda()
# criterion_hm = criterion_hm.cuda()
# criterion_paf = criterion_paf.cuda()
# Create optimizer
optimizer = create_optimizer(opt, model)
# Other params
n_epochs = opt.nEpoch
to_train = opt.train
drop_lr = opt.dropLR
val_interval = opt.valInterval
learn_rate = opt.LR
visualize_out = opt.vizOut
# train/ test
img, heat_map, paf, ignore_mask, keypoints = test_loader.dataset.get_item_raw(0, False)
def main():
# Seed all sources of randomness to 0 for reproducibility
np.random.seed(0)
torch.manual_seed(0)
torch.cuda.manual_seed(0)
random.seed(0)
opt = Opts().parse()
os.environ["CUDA_VISIBLE_DEVICES"] = opt["env"]["device"]
print("Using GPU: {}".format(opt["env"]["device"]))
# Create data loaders
# Create data loaders
train_loader, test_loader, _ = create_data_loaders(opt)
# Create nn
model, criterion_hm, criterion_paf = create_model(opt)
model = torch.nn.DataParallel(model, device_ids=[int(index) for index in opt["env"]["device"].split(",")]).cuda() \
if "," in opt["env"]["device"] else model.cuda()
if opt["env"]["loadModel"] is not None and opt["typ"] == 'cpr':
model.load_state_dict(torch.load(opt["env"]["loadModel"]))
print('Loaded model from ' + opt["env"]["loadModel"])
criterion_hm = criterion_hm.cuda()
criterion_paf = criterion_paf.cuda()
# Create optimizer
optimizer = create_optimizer(opt, model)
# Other params
to_train = opt["to_train"]
visualize_out = opt["viz"]["vizOut"]
# train/ test
Processer = process(model)
if to_train:
Processer.train_net(train_loader, test_loader, criterion_hm, criterion_paf, optimizer, opt, viz_output=visualize_out)
else:
Processer.validate_net(test_loader, criterion_hm, criterion_paf, save_dir=opt["env"]["saveDir"], viz_output=visualize_out)
def main():
# Seed all sources of randomness to 0 for reproducibility
torch.manual_seed(0)
torch.cuda.manual_seed(0) if torch.cuda.is_available() else torch.manual_seed(0)
random.seed(0)
# Set cudnn.benchmark True to spped up training
if torch.cuda.is_available():
torch.backends.cudnn.enabled =True
torch.backends.cudnn.benchmark = True
opt = Opts().opt
# Create data loaders
train_loader, test_loader = create_data_loaders(opt)
# Create nn
model = create_model(opt).to(device())
# Create loss criterion
if opt.onlyAutoEncoder:
criterion = create_criterion(opt.criterionAutoEncoder).to(device())
else:
criterion = create_criterion(opt.criterionClassifier).to(device())
# Choose to train or to test the model
if opt.toTrain:
# Create optimizer
optimizer = create_optimizer(opt, model)
train_net(opt, train_loader, test_loader, model, criterion, optimizer, opt.nEpoch, opt.valInterval, opt.LR, opt.dropLR)
# Test classifier or AutoEncoder
if not opt.onlyAutoEncoder:
# Testing classifier
predictList = []
labelsList = []
model.eval()
with torch.no_grad():
for data in test_loader:
images, labels = data
images = images.to(device())
labels = labels.to(device())
outputs = model(images)
_, predicted = torch.max(outputs, 1)
predictList.extend(predicted.cpu().numpy().tolist())
labelsList.extend(labels.cpu().numpy().tolist())
# ===================confusionMatrixGeneration====================
confusionMatrix=confusion_matrix(labelsList, predictList)
print('ConfusionMatrix:\n{}\n{}'.format(list(CLASSES.values()),confusionMatrix))
np.save(os.path.join(opt.saveDir,'confusionMatrix.npy'),confusionMatrix)
# ===================classificaitonReport: Precision, Recall and f1 score====================
classReport = classification_report(labelsList, predictList, digits=2)
df = pandas.DataFrame(classification_report(labelsList, predictList, digits=2, output_dict=True)).transpose()
df.to_csv(os.path.join(opt.saveDir, 'my_csv_file.csv'))
print(classReport)
# ===================heatmapGeneration====================
if opt.toGenerateHeatmap:
model.to(torch.device('cpu'))
#Currently 33 testing images in the folder test_images are chosen for heaimap generation
target_example = 33
for t in range(target_example):
# Get params for heatmap generation
(original_image, prep_img, target_class, file_name_to_export) = get_example_params(t)
# Grad cam, choose which block and which layer in the block for heatmap generation
grad_cam = GradCam(model, target_block=2, target_layer=9)
# Generate cam mask
cam = grad_cam.generate_cam(prep_img)
# Save mask
save_class_activation_images(opt, original_image, cam, file_name_to_export)
print('Grad cam completed')
else:
# Test AutoEncoder to plot original input images and reconstructed images
model.eval()
dataiter = iter(test_loader)
# Generate images in the first 8 iteration
for i in range(8):
images, labels = dataiter.next()
print('GroundTruth: ', ' '.join('%5s' % CLASSES[labels[j].item()] for j in range(opt.batchSize)))
# ===================showGroundTruthImages====================
imshow(torchvision.utils.make_grid(images))
images_ = images.to(device())
# ===================forward=====================
decoded_imgs = model(images_)
# ===================showReconstructedImages====================
imshow(torchvision.utils.make_grid(decoded_imgs.data))
def main():
# Seed all sources of randomness to 0 for reproducibility
torch.manual_seed(0)
torch.cuda.manual_seed(
0) if torch.cuda.is_available() else torch.manual_seed(0)
random.seed(0)
# Set cudnn.benchmark True to spped up training
if torch.cuda.is_available():
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
opt = Opts().opt
# Create train and val data loaders
train_loader, val_loader = create_data_loaders(
opt, 'train'), create_data_loaders(opt, 'val')
# Create nn
model = create_model(opt)
if opt.toCuda:
model = model.to(device())
# Choose to train or to test the model
if opt.toTrain:
# Create optimizer
optimizer = create_optimizer(opt, model)
train_net(opt, train_loader, val_loader, model, optimizer, opt.nEpoch,
opt.valInterval, opt.LR, opt.dropLR)
else:
# Change ASPP image pooling
output_stride = 32
train_crop_h, train_crop_w = (1025, 2049)
scale = 1. / output_stride
pool_h = int((float(train_crop_h) - 1.0) * scale + 1.0)
pool_w = int((float(train_crop_w) - 1.0) * scale + 1.0)
model.set_image_pooling((pool_h, pool_w))
# Create test data loaders, change batch size to 1
opt.batchSize = 1
test_loader = create_data_loaders(opt, 'test')
# test_loader = create_data_loaders(opt, 'test')
panoptic_metric = CityscapesPanopticEvaluator(
output_dir=os.path.join(opt.saveDir, 'panoptic'),
train_id_to_eval_id=test_loader.dataset.train_id_to_eval_id(),
label_divisor=test_loader.dataset.label_divisor,
void_label=test_loader.dataset.label_divisor *
test_loader.dataset.ignore_label,
gt_dir=opt.data,
split=test_loader.dataset.split,
num_classes=test_loader.dataset.num_classes)
image_filename_list = [
os.path.splitext(os.path.basename(ann))[0]
for ann in test_loader.dataset.img_list
]
debug_out_dir = os.path.join(opt.saveDir, 'debug_test')
PathManager.mkdirs(debug_out_dir)
model.eval()
with torch.no_grad():
for i, data in enumerate(test_loader):
if opt.toCuda:
data = to_cuda(data, device())
image = data.pop('image')
out_dict = model(image)
# post-processing
semantic_pred = get_semantic_segmentation(out_dict['semantic'])
if 'foreground' in out_dict:
foreground_pred = get_semantic_segmentation(
out_dict['foreground'])
else:
foreground_pred = None
panoptic_pred, center_pred = get_panoptic_segmentation(
semantic_pred,
out_dict['center'],
out_dict['offset'],
thing_list=test_loader.dataset.thing_list,
label_divisor=test_loader.dataset.label_divisor,
stuff_area=2048,
void_label=(test_loader.dataset.label_divisor *
test_loader.dataset.ignore_label),
threshold=0.1,
nms_kernel=7,
top_k=200,
foreground_mask=foreground_pred)
# save predictions
semantic_pred = semantic_pred.squeeze(0).cpu().numpy()
panoptic_pred = panoptic_pred.squeeze(0).cpu().numpy()
# Crop padded regions.
image_size = data['size'].squeeze(0).cpu().numpy()
panoptic_pred = panoptic_pred[:image_size[0], :image_size[1]]
# Resize back to the raw image size.
raw_image_size = data['raw_size'].squeeze(0).cpu().numpy()
if raw_image_size[0] != image_size[0] or raw_image_size[
1] != image_size[1]:
semantic_pred = cv2.resize(
semantic_pred.astype(np.float),
(raw_image_size[1], raw_image_size[0]),
interpolation=cv2.INTER_NEAREST).astype(np.int32)
panoptic_pred = cv2.resize(
panoptic_pred.astype(np.float),
(raw_image_size[1], raw_image_size[0]),
interpolation=cv2.INTER_NEAREST).astype(np.int32)
# Optional: evaluates panoptic segmentation.
image_id = '_'.join(image_filename_list[i].split('_')[:3])
panoptic_metric.update(panoptic_pred,
image_filename=image_filename_list[i],
image_id=image_id)
# Processed outputs
# save_annotation(semantic_pred, debug_out_dir, 'semantic_pred_%d' % i,
# add_colormap=True, colormap=test_loader.dataset.create_label_colormap())
# pan_to_sem = panoptic_pred // test_loader.dataset.label_divisor
# save_annotation(pan_to_sem, debug_out_dir, 'pan_to_sem_pred_%d' % i,
# add_colormap=True, colormap=test_loader.dataset.create_label_colormap())
# ins_id = panoptic_pred % test_loader.dataset.label_divisor
# pan_to_ins = panoptic_pred.copy()
# pan_to_ins[ins_id == 0] = 0
# save_instance_annotation(pan_to_ins, debug_out_dir, 'pan_to_ins_pred_%d' % i)
save_panoptic_annotation(
panoptic_pred,
debug_out_dir,
'panoptic_pred_%d' % i,
label_divisor=test_loader.dataset.label_divisor,
colormap=test_loader.dataset.create_label_colormap())
print('1111111111111111111111')
results = panoptic_metric.evaluate()
print(results)