def __init__(self, args, config):
self.args = args
self.config = config
self.cache_dir = utils.get_cache_dir(config)
self.model_dir = utils.get_model_dir(config)
self.category = utils.get_category(config, self.cache_dir if os.path.exists(self.cache_dir) else None)
self.draw_bbox = utils.visualize.DrawBBox(self.category, colors=args.colors, thickness=args.thickness)
self.anchors = torch.from_numpy(utils.get_anchors(config)).contiguous()
self.height, self.width = tuple(map(int, config.get('image', 'size').split()))
self.path, self.step, self.epoch = utils.train.load_model(self.model_dir)
state_dict = torch.load(self.path, map_location=lambda storage, loc: storage)
self.dnn = utils.parse_attr(config.get('model', 'dnn'))(model.ConfigChannels(config, state_dict), self.anchors, len(self.category))
self.dnn.load_state_dict(state_dict)
self.inference = model.Inference(config, self.dnn, self.anchors)
self.inference.eval()
if torch.cuda.is_available():
self.inference.cuda()
logging.info(humanize.naturalsize(sum(var.cpu().numpy().nbytes for var in self.inference.state_dict().values())))
self.create_cap()
self.create_cap_size()
self.writer = self.create_writer()
self.keys = set(args.keys)
self.resize = transform.parse_transform(config, config.get('transform', 'resize_test'))
self.transform_image = transform.get_transform(config, config.get('transform', 'image_test').split())
self.transform_tensor = transform.get_transform(config, config.get('transform', 'tensor').split())
def get_loader(self):
paths = [
os.path.join(self.cache_dir, phase + '.pkl')
for phase in self.config.get('eval', 'phase').split()
]
dataset = utils.data.Dataset(utils.data.load_pickles(paths))
logging.info('num_examples=%d' % len(dataset))
size = tuple(map(int, self.config.get('image', 'size').split()))
try:
workers = self.config.getint('data', 'workers')
except configparser.NoOptionError:
workers = multiprocessing.cpu_count()
collate_fn = utils.data.Collate(
transform.parse_transform(
self.config, self.config.get('transform', 'resize_eval')),
[size],
transform_image=transform.get_transform(
self.config,
self.config.get('transform', 'image_test').split()),
transform_tensor=transform.get_transform(
self.config,
self.config.get('transform', 'tensor').split()),
)
return torch.utils.data.DataLoader(dataset,
batch_size=self.args.batch_size,
num_workers=workers,
collate_fn=collate_fn)
def main():
args = make_args()
config = configparser.ConfigParser()
utils.load_config(config, args.config)
for cmd in args.modify:
utils.modify_config(config, cmd)
with open(os.path.expanduser(os.path.expandvars(args.logging)), 'r') as f:
logging.config.dictConfig(yaml.load(f))
model_dir = utils.get_model_dir(config)
height, width = tuple(map(int, config.get('image', 'size').split()))
resize = transform.parse_transform(config, config.get('transform', 'resize_test'))
transform_image = transform.get_transform(config, config.get('transform', 'image_test').split())
transform_tensor = transform.get_transform(config, config.get('transform', 'tensor').split())
# load image
image_bgr = cv2.imread('image.jpg')
image_resized = resize(image_bgr, height, width)
image = transform_image(image_resized)
tensor = transform_tensor(image).unsqueeze(0)
# Caffe2
init_net = caffe2_pb2.NetDef()
with open(os.path.join(model_dir, 'init_net.pb'), 'rb') as f:
init_net.ParseFromString(f.read())
predict_net = caffe2_pb2.NetDef()
with open(os.path.join(model_dir, 'predict_net.pb'), 'rb') as f:
predict_net.ParseFromString(f.read())
p = workspace.Predictor(init_net, predict_net)
results = p.run([tensor.numpy()])
logging.info(utils.abs_mean(results[0]))
logging.info(hashlib.md5(results[0].tostring()).hexdigest())
def __init__(self, args, config):
self.args = args
self.config = config
self.device = torch.device(
'cuda' if torch.cuda.is_available() else 'cpu')
self.cache_dir = utils.get_cache_dir(config)
self.model_dir = utils.get_model_dir(config)
_, self.num_parts = utils.get_dataset_mappers(config)
self.limbs_index = utils.get_limbs_index(config)
if args.debug is None:
self.draw_cluster = utils.visualize.DrawCluster(
colors=args.colors, thickness=args.thickness)
else:
self.draw_feature = utils.visualize.DrawFeature()
s = re.search('(-?[0-9]+)([a-z]+)(-?[0-9]+)', args.debug)
stage = int(s.group(1))
name = s.group(2)
channel = int(s.group(3))
self.get_feature = lambda outputs: outputs[stage][name][0][channel]
self.height, self.width = tuple(
map(int,
config.get('image', 'size').split()))
if args.caffe:
init_net = caffe2_pb2.NetDef()
with open(os.path.join(self.model_dir, 'init_net.pb'), 'rb') as f:
init_net.ParseFromString(f.read())
predict_net = caffe2_pb2.NetDef()
with open(os.path.join(self.model_dir, 'predict_net.pb'),
'rb') as f:
predict_net.ParseFromString(f.read())
p = workspace.Predictor(init_net, predict_net)
self.inference = lambda tensor: [{
'parts': torch.from_numpy(parts),
'limbs': torch.from_numpy(limbs)
} for parts, limbs in zip(
*[iter(p.run([tensor.detach().cpu().numpy()]))] * 2)]
else:
self.step, self.epoch, self.dnn, self.stages = self.load()
self.inference = model.Inference(config, self.dnn, self.stages)
self.inference.eval()
if torch.cuda.is_available():
self.inference.cuda()
logging.info(
humanize.naturalsize(
sum(var.cpu().numpy().nbytes
for var in self.inference.state_dict().values())))
self.cap = self.create_cap()
self.keys = set(args.keys)
self.resize = transform.parse_transform(
config, config.get('transform', 'resize_test'))
self.transform_image = transform.get_transform(
config,
config.get('transform', 'image_test').split())
self.transform_tensor = transform.get_transform(
config,
config.get('transform', 'tensor').split())
def main():
args = make_args()
config = configparser.ConfigParser()
utils.load_config(config, args.config)
for cmd in args.modify:
utils.modify_config(config, cmd)
with open(os.path.expanduser(os.path.expandvars(args.logging)), 'r') as f:
logging.config.dictConfig(yaml.load(f))
cache_dir = utils.get_cache_dir(config)
model_dir = utils.get_model_dir(config)
category = utils.get_category(
config, cache_dir if os.path.exists(cache_dir) else None)
anchors = utils.get_anchors(config)
anchors = torch.from_numpy(anchors).contiguous()
path, step, epoch = utils.train.load_model(model_dir)
state_dict = torch.load(path, map_location=lambda storage, loc: storage)
dnn = utils.parse_attr(config.get('model', 'dnn'))(model.ConfigChannels(
config, state_dict), anchors, len(category))
dnn.load_state_dict(state_dict)
height, width = tuple(map(int, config.get('image', 'size').split()))
resize = transform.parse_transform(config,
config.get('transform', 'resize_test'))
transform_image = transform.get_transform(
config,
config.get('transform', 'image_test').split())
transform_tensor = transform.get_transform(
config,
config.get('transform', 'tensor').split())
# load image
image_bgr = cv2.imread('image.jpg')
image_resized = resize(image_bgr, height, width)
image = transform_image(image_resized)
tensor = transform_tensor(image).unsqueeze(0)
# Checksum
for key, var in dnn.state_dict().items():
a = var.cpu().numpy()
print('\t'.join(
map(str, [
key, a.shape,
utils.abs_mean(a),
hashlib.md5(a.tostring()).hexdigest()
])))
output = dnn(torch.autograd.Variable(tensor, volatile=True)).data
for key, a in [
('image_bgr', image_bgr),
('image_resized', image_resized),
('tensor', tensor.cpu().numpy()),
('output', output.cpu().numpy()),
]:
print('\t'.join(
map(str, [
key, a.shape,
utils.abs_mean(a),
hashlib.md5(a.tostring()).hexdigest()
])))
def get_stereo_transform():
endoscope_chesspts = list(read_camera.load_all('../calibration/endoscope_chesspts.p'))
camera_info = list(read_camera.load_all('../camera_data/camera_info.p'))
left_chesspts = np.matrix(list(read_camera.load_all('../camera_data/left_chesspts'))[0])
right_chesspts = np.matrix(list(read_camera.load_all('../camera_data/right_chesspts'))[0])
z = np.zeros((25, 1))
left_chesspts = np.hstack((left_chesspts, z))
right_chesspts = np.hstack((right_chesspts, z))
TL_R = transform.get_transform("Left Camera", "Right Camera", left_chesspts, right_chesspts, verbose=False)
return TL_R
def main():
# train on the GPU or on the CPU, if a GPU is not available
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
# our dataset has two classes only - background and person
num_classes = 2
# use our dataset and defined transformations
dataset = PennFudanDataset('C:\\zhulei\\maskRcnn\\data\\train',
get_transform(train=True))
dataset_test = PennFudanDataset('C:\\zhulei\\maskRcnn\\data\\test',
get_transform(train=False))
# split the dataset in train and test set
# indices = torch.randperm(len(dataset)).tolist()
# dataset = torch.utils.data.Subset(dataset, indices[:-50])
# dataset_test = torch.utils.data.Subset(dataset_test, indices[-50:])
# define training and validation data loaders
data_loader = torch.utils.data.DataLoader(dataset,
batch_size=2,
shuffle=True,
num_workers=4,
collate_fn=utils.collate_fn)
data_loader_test = torch.utils.data.DataLoader(dataset_test,
batch_size=1,
shuffle=False,
num_workers=4,
collate_fn=utils.collate_fn)
# get the model using our helper function
model = get_model_instance_segmentation(num_classes)
# move model to the right device
model.to(device)
# construct an optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.005,
momentum=0.9,
weight_decay=0.0005)
# and a learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=3,
gamma=0.1)
# let's train it for 10 epochs
num_epochs = 10
save_dir = "C:\\zhulei\\maskRcnn"
for epoch in range(num_epochs):
# train for one epoch, printing every 10 iterations
train_one_epoch(model,
optimizer,
data_loader,
device,
epoch,
print_freq=10)
# update the learning rate
lr_scheduler.step()
# save the model
torch.save(
model.state_dict(),
os.path.join(save_dir, 'models', '_epoch-' + str(epoch) + '.pth'))
# evaluate on the test dataset
evaluate(model, data_loader_test, device=device)
print("That's it!")
from dataloader import PennFudanDataset
from PIL import Image
import os
import numpy as np
import matplotlib as mpl
import numpy
from models.detection.rpn import AnchorGenerator
import models
from model import get_instance_segmentation_model
from transform import get_transform
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
device=torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# use the PennFudan dataset and defined transformations
dataset = PennFudanDataset(root='./patch_data/', transforms=get_transform(train=True))
dataset_test = PennFudanDataset(root='./patch_data_test/same_tissue/', transforms=get_transform(train=False))
# define training and validation data loaders
data_loader = torch.utils.data.DataLoader(
dataset, batch_size=4, shuffle=True, num_workers=0,
collate_fn=utils.collate_fn)
data_loader_test = torch.utils.data.DataLoader(
dataset_test, batch_size=1, shuffle=True, num_workers=0,
collate_fn=utils.collate_fn)
# the dataset has two classes only - background and person
num_classes = 2
# get the model using the helper function
#bone: 'resnet50'/'mobilenet_v2'/'googlenet'/'densenet121'/'resnet50'/'shufflenet_v2_x1_0'/'inception_v3'/'squeezenet1_0'/
def main():
""" main function
"""
### header
parser = argparse.ArgumentParser()
# path
parser.add_argument('--root-path', default=CFG.root_path, help="root path")
parser.add_argument('--log-path', default=CFG.log_path, help="log path")
parser.add_argument('--model-path',
default=CFG.model_path,
help="model path")
parser.add_argument('--pretrained-path', help='pretrained path')
# image
parser.add_argument('--transform-version',
default=0,
type=int,
help="image transform version ex) 0, 1, 2 ...")
parser.add_argument('--image-size',
default=64,
type=int,
help="image size(64)")
# model
parser.add_argument('--model-name',
default=CFG.model_name,
help=f"model name({CFG.model_name})")
parser.add_argument('--backbone-name',
default=CFG.backbone_name,
help=f"backbone name({CFG.backbone_name})")
# learning
parser.add_argument('--batch-size',
default=CFG.batch_size,
type=int,
help=f"batch size({CFG.batch_size})")
parser.add_argument('--learning-rate',
default=CFG.learning_rate,
type=float,
help=f"learning rate({CFG.learning_rate})")
parser.add_argument('--num-epochs',
default=CFG.num_epochs,
type=int,
help=f"number of epochs({CFG.num_epochs})")
# etc
parser.add_argument("--seed",
default=CFG.seed,
type=int,
help=f"seed({CFG.seed})")
parser.add_argument("--workers",
default=CFG.workers,
type=int,
help=f"number of workers({CFG.workers})")
parser.add_argument("--debug", action="store_true", help="debug mode")
parser.add_argument("--val-fold",
default=CFG.val_fold,
choices=[list(range(0, CFG.n_folds))],
help=f"fold number for validation({CFG.val_fold})")
args = parser.parse_args()
# path
CFG.root_path = args.root_path
CFG.model_path = args.model_path
CFG.log_path = args.log_path
CFG.pretrained_path = args.pretrained_path
# image
CFG.transform_version = args.transform_version
CFG.image_size = args.image_size
# model
CFG.model_name = args.model_name
CFG.backbone_name = args.backbone_name
# learning
CFG.batch_size = args.batch_size
CFG.learning_rate = args.learning_rate
CFG.num_epochs = args.num_epochs
# etc
CFG.seed = args.seed
CFG.workers = args.workers
CFG.debug = args.debug
# get device
CFG.device = get_device()
# get version
_, version, _ = sys.argv[0].split('/')
CFG.version = version
# update log path
if not CFG.debug:
CFG.log_path = os.path.join(CFG.log_path, CFG.version)
os.makedirs(CFG.log_path, exist_ok=True)
CFG.log_path = os.path.join(
CFG.log_path, f'exp_{get_exp_id(CFG.log_path, prefix="exp_")}')
os.makedirs(CFG.log_path, exist_ok=True)
else:
CFG.log_path = os.path.join(CFG.log_path, "debug")
os.makedirs(CFG.log_path, exist_ok=True)
CFG.log_path = os.path.join(CFG.log_path, "debug")
os.makedirs(CFG.log_path, exist_ok=True)
# update model path
if not CFG.debug:
CFG.model_path = os.path.join(CFG.model_path, version)
os.makedirs(CFG.model_path, exist_ok=True)
CFG.model_path = os.path.join(
CFG.model_path, f'exp_{get_exp_id(CFG.model_path, prefix="exp_")}')
os.makedirs(CFG.model_path, exist_ok=True)
else:
CFG.model_path = os.path.join(CFG.model_path, "debug")
os.makedirs(CFG.model_path, exist_ok=True)
CFG.model_path = os.path.join(CFG.model_path, "debug")
os.makedirs(CFG.model_path, exist_ok=True)
pprint({k: v for k, v in dict(CFG.__dict__).items() if '__' not in k})
json.dump({k: v
for k, v in dict(CFG.__dict__).items() if '__' not in k},
open(os.path.join(CFG.log_path, 'CFG.json'), "w"))
print()
### Seed all
seed_everything(CFG.seed)
### Data Related
# load raw data
print("Load Raw Data")
train_df, test_df, ss_df = load_data(CFG)
# preprocess data
print("Preprocess Data")
train_df = preprocess_data(CFG, train_df)
# split data
print("Split Data")
train_df, valid_df = split_data(CFG, train_df)
# get transform
print("Get Transforms")
train_transforms, test_transforms = get_transform(CFG)
# get dataset
print("Get Dataset")
trn_data = DFDDataset(CFG, train_df, train_transforms)
val_data = DFDDataset(CFG, valid_df, test_transforms)
### Model related
# get learner
learner = Learner(CFG)
learner.name = f"model.fold_{CFG.val_fold}"
if CFG.pretrained_path:
print("Load Pretrained Model")
print(f"... Pretrained Info - {CFG.pretrained_path}")
learner.load(CFG.pretrained_path, f"model_state_dict")
model = learner.best_model.to(CFG.deivce)
else:
print(f"Load Model")
model = get_model(CFG)
model = model.to(CFG.device)
if torch.cuda.device_count() > 1:
model = nn.DataParallel(model)
# get optimizer
optimizer = optim.Adam(model.parameters(), lr=CFG.learning_rate)
# get scheduler
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
mode='min',
patience=2,
verbose=False,
factor=0.5)
### train related
# train model
learner.train(trn_data, val_data, model, optimizer, scheduler)
def get_transform(config, sequence):
return transform.get_transform(config, sequence)
def main():
parser = argparse.ArgumentParser()
# path
parser.add_argument('--root-path', default=CFG.root_path, help="root path")
parser.add_argument('--save-path', default=CFG.save_path, help="save path")
parser.add_argument('--sub-name',
default=CFG.sub_name,
help="submission name")
# learning
parser.add_argument('--batch-size',
default=CFG.batch_size,
type=int,
help=f"batch size({CFG.batch_size})")
parser.add_argument("--workers",
default=CFG.workers,
type=int,
help=f"number of workers({CFG.workers})")
parser.add_argument("--seed",
default=CFG.seed,
type=int,
help=f"seed({CFG.seed})")
# version
parser.add_argument('--version', type=int)
parser.add_argument('--exp-id', type=int)
# etc
parser.add_argument('--tta', action='store_true', default=False)
args = parser.parse_args()
CFG.root_path = args.root_path
CFG.save_path = args.save_path
CFG.sub_name = args.sub_name
CFG.batch_size = args.batch_size
CFG.workers = args.workers
CFG.seed = args.seed
CFG.model_path = f"./model/v{args.version}/exp_{args.exp_id}/"
CFG.log_path = f"./log/v{args.version}/exp_{args.exp_id}/"
CFG.tta = args.tta
# get device
CFG.device = get_device()
# load train environment
env = json.load(open(os.path.join(CFG.log_path, 'CFG.json'), 'r'))
for k, v in env.items():
setattr(CFG, k, v)
loss, metric = 0, 0
for fold in range(CFG.n_folds):
fn = os.path.join(CFG.log_path, f"log.fold_{fold}.csv")
score = pd.read_csv(fn).sort_values("val_metric",
ascending=False).iloc[0]
loss += score['val_loss'] #/ CFG.n_folds
metric += score['val_metric'] #/ CFG.n_folds
break
CFG.sub_name = f"submission." \
f"ver_{args.version}." \
f"exp_{args.exp_id}." \
f"loss_{loss:.4f}." \
f"metric_{metric:.4f}.csv"
if CFG.tta:
CFG.sub_name = "tta." + CFG.sub_name
pprint({k: v for k, v in dict(CFG.__dict__).items() if '__' not in k})
print()
### seed all
seed_everything(CFG.seed)
### Data related logic
# load data
print("Load Raw Data")
_, test_df, ss_df = load_data(CFG)
print()
# get transform
print("Get Transform")
_, test_transforms = get_transform(CFG)
print()
# dataset
tst_data = DFDDataset(CFG, test_df, test_transforms)
# if tta
if CFG.tta:
test_transforms = get_tta_transform(CFG)
tst_data = DFDDataset(CFG, test_df, test_transforms)
final_preds = np.zeros(test_df.shape[0])
# folds
for fold in range(CFG.n_folds):
print(f"========== Fold: {fold} ==========")
# load learner
print("Load Model")
model_name = f'model.fold_{fold}.best.pt'
learner = Learner(CFG)
learner.load(os.path.join(CFG.model_path, model_name),
f"model_state_dict")
# prediction
if not CFG.tta:
test_preds = torch.sigmoid(
learner.predict(tst_data).view(-1)).numpy()
else:
test_preds = np.zeros(test_df.shape[0])
for _ in range(4):
test_preds += torch.sigmoid(
learner.predict(tta_data).view(-1)).numpy() / 4
final_preds += test_preds #/ CFG.n_folds
print()
break
image_path = [
path.replace(".npy", "").replace("./input/test_face_margin/", "")
for path in test_df['image_path'].values
]
print(final_preds.max(), final_preds.min())
test_df = pd.DataFrame({"path": image_path, "y": np.round(final_preds)})
test_df = test_df.set_index("path")
ss_df = test_df.loc[ss_df['path']].reset_index()[['path', 'y']]
ss_df.to_csv(os.path.join(CFG.save_path, f"{CFG.sub_name}"), index=False)
print(ss_df.head())
test_df = pd.DataFrame({"path": image_path, "y": final_preds})
test_df = test_df.set_index("path")
ss_df = test_df.loc[ss_df['path']].reset_index()[['path', 'y']]
ss_df.to_csv(os.path.join(CFG.save_path, f"raw.{CFG.sub_name}"),
index=False)
print(ss_df.head())
from dataloader import PennFudanDataset
from PIL import Image
import os
import numpy as np
import matplotlib as mpl
import numpy
from color import pre_to_img
from transform import get_transform
from model import get_instance_segmentation_model
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# use the PennFudan dataset and defined transformations
dataset_test = PennFudanDataset(root='./patch_data_test/same_tissue/',
transforms=get_transform(train=False))
data_loader_test = torch.utils.data.DataLoader(dataset_test,
batch_size=1,
shuffle=True,
num_workers=0,
collate_fn=utils.collate_fn)
# put the model in evaluation mode
def get_counts(sequence):
counts = {}
for i in range(sequence.shape[0]):
for x in sequence[i]:
if x in counts:
def main():
parser = argparse.ArgumentParser()
# path
parser.add_argument('--root-path', default=CFG.root_path, help="root path")
parser.add_argument('--save-path', default=CFG.save_path, help="save path")
parser.add_argument('--sub-name',
default=CFG.sub_name,
help="submission name")
# learning
parser.add_argument('--batch-size',
default=CFG.batch_size,
type=int,
help=f"batch size({CFG.batch_size})")
parser.add_argument("--workers",
default=CFG.workers,
type=int,
help=f"number of workers({CFG.workers})")
parser.add_argument("--seed",
default=CFG.seed,
type=int,
help=f"seed({CFG.seed})")
# version
parser.add_argument('--version', type=int)
parser.add_argument('--exp-id', type=int)
# etc
parser.add_argument('--tta', action='store_true', default=False)
args = parser.parse_args()
CFG.root_path = args.root_path
CFG.save_path = args.save_path
CFG.sub_name = args.sub_name
CFG.batch_size = args.batch_size
CFG.workers = args.workers
CFG.seed = args.seed
CFG.model_path = f"./model/v{args.version}/exp_{args.exp_id}/"
CFG.log_path = f"./log/v{args.version}/exp_{args.exp_id}/"
CFG.tta = args.tta
# get device
CFG.device = get_device()
# load train environment
env = json.load(open(os.path.join(CFG.log_path, 'CFG.json'), 'r'))
for k, v in env.items():
setattr(CFG, k, v)
loss, metric = 0, 0
for fold in range(CFG.n_folds):
fn = os.path.join(CFG.log_path, f"log.fold_{fold}.csv")
score = pd.read_csv(fn).sort_values("val_loss", ascending=True).iloc[0]
loss += score['val_loss'] / CFG.n_folds
metric += score['val_metric'] / CFG.n_folds
CFG.sub_name = f"submission." \
f"ver_{args.version}." \
f"exp_{args.exp_id}." \
f"loss_{loss:.4f}." \
f"metric_{metric:.4f}.csv"
if CFG.tta:
CFG.sub_name = "tta." + CFG.sub_name
pprint({k: v for k, v in dict(CFG.__dict__).items() if '__' not in k})
print()
### seed all
seed_everything(CFG.seed)
### Data related logic
# load data
print("Load Raw Data")
_, test_df = load_data(CFG)
print()
# preprocess data
print("Preprocess Data")
test_df = preprocess_data(CFG, test_df)
print()
# get transform
print("Get Transform")
_, test_transforms = get_transform(CFG)
print()
# dataset
tst_data = MelanomaDataset(CFG, test_df, test_transforms)
# if tta
tta_transforms = transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
tta_data = MelanomaDataset(CFG, test_df, tta_transforms)
final_preds = np.zeros(test_df.shape[0])
# folds
for fold in range(CFG.n_folds):
print(f"========== Fold: {fold} ==========")
# load learner
print("Load Model")
model_name = f'model.fold_{fold}.best.pt'
learner = Learner(CFG)
learner.load(os.path.join(CFG.model_path, model_name),
f"model_state_dict")
# prediction
if not CFG.tta:
test_preds = torch.sigmoid(
learner.predict(tst_data).view(-1)).numpy()
else:
test_preds = np.zeros(test_df.shape[0])
for _ in range(4):
test_preds += torch.sigmoid(
learner.predict(tta_data).view(-1)).numpy() / 4
final_preds += test_preds / CFG.n_folds
print()
ss_df = pd.read_csv(
os.path.join(CFG.root_path, "melanoma-external-malignant-256",
"sample_submission.csv"))
test_df['target'] = final_preds
test_df.set_index("image_name", inplace=True)
ss_df = test_df.loc[ss_df['image_name']].reset_index()[[
'image_name', 'target'
]]
ss_df.to_csv(os.path.join(CFG.save_path, f"{CFG.sub_name}"), index=False)
print(ss_df.head())
def main():
# train on the GPU or on the CPU, if a GPU is not available
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
# our dataset has two classes only - background and person
num_classes = 2
# get the model using our helper function
model = get_model_instance_segmentation(num_classes)
# load pretrain_dict
pretrain_dict = torch.load(
os.path.join("C:\\zhulei\\maskRcnn\\models", "_epoch-9.pth"))
model.load_state_dict(pretrain_dict)
# move model to the right device
model.to(device)
# use our dataset and defined transformations
dataset_test = PennFudanDataset('C:\\zhulei\\maskRcnn\\data\\test',
get_transform(train=False))
data_loader_test = torch.utils.data.DataLoader(dataset_test,
batch_size=1,
shuffle=False,
num_workers=1,
collate_fn=utils.collate_fn)
n_threads = torch.get_num_threads()
# FIXME remove this and make paste_masks_in_image run on the GPU
torch.set_num_threads(1)
cpu_device = torch.device("cpu")
model.eval()
metric_logger = utils.MetricLogger(delimiter=" ")
header = 'Test:'
coco = get_coco_api_from_dataset(data_loader_test.dataset)
iou_types = _get_iou_types(model)
coco_evaluator = CocoEvaluator(coco, iou_types)
for image, targets in metric_logger.log_every(data_loader_test, 100,
header):
image = list(img.to(device) for img in image)
targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
torch.cuda.synchronize()
model_time = time.time()
outputs = model(image)
instance_segmentation_api(image[0], outputs)
# 可视化
# for img in image:
# Image.fromarray((img.mul(255).permute(1, 2, 0).byte().cpu().numpy())[0])
# print(outputs[0]['masks'].shape)
# for i in range(99):
# result = Image.fromarray(outputs[0]['masks'][i, 0].mul(255).byte().cpu().numpy())
# result.show()
outputs = [{k: v.to(cpu_device)
for k, v in t.items()} for t in outputs]
model_time = time.time() - model_time
res = {
target["image_id"].item(): output
for target, output in zip(targets, outputs)
}
evaluator_time = time.time()
coco_evaluator.update(res)
evaluator_time = time.time() - evaluator_time
metric_logger.update(model_time=model_time,
evaluator_time=evaluator_time)
# gather the stats from all processes
metric_logger.synchronize_between_processes()
print("Averaged stats:", metric_logger)
coco_evaluator.synchronize_between_processes()
# accumulate predictions from all images
coco_evaluator.accumulate()
coco_evaluator.summarize()
torch.set_num_threads(n_threads)
pos2 = PyKDL.Vector(-0.0972128, -0.0170138, -0.106974)
sideways = PyKDL.Rotation(-0.453413, 0.428549, -0.781513, -0.17203,
0.818259, 0.548505, 0.874541, 0.383143,
-0.297286)
""" Move to arbitrary start position (near upper left corner) & release anything gripper is
holding. """
home(psm2, pos2, sideways)
""" Get PSM and endoscope calibration data (25 corresponding chess points) """
psm2_calibration_data = list(
transform.load_all('utils/psm2_recordings.txt'))
psm2_calibration_matrix = transform.fit_to_plane(
transform.psm_data_to_matrix(psm2_calibration_data))
endoscope_calibration_matrix = transform.fit_to_plane(
np.matrix(
list(read_camera.load_all('camera_data/endoscope_chesspts.p'))[0]))
world = transform.generate_world()
TE_2 = transform.get_transform("Endoscope", "PSM2",
endoscope_calibration_matrix,
psm2_calibration_matrix)
psme_2 = transform.transform_data("Endoscope", "PSM2",
endoscope_calibration_matrix, TE_2,
psm2_calibration_matrix)
pprint.pprint(psme_2)
""" Move to chessboard corner, descend, come up,and go to next. """
move_to(psm2, psme_2.tolist(), z_upper)
home(psm2, pos, rot)
def main():
parser = argparse.ArgumentParser()
# path
parser.add_argument('--root-path', default=CFG.root_path, help="root path")
parser.add_argument('--save-path', default=CFG.save_path, help="save path")
parser.add_argument('--sub-name',
default=CFG.sub_name,
help="submission name")
# learning
parser.add_argument('--batch-size',
default=CFG.batch_size,
type=int,
help=f"batch size({CFG.batch_size})")
parser.add_argument("--workers",
default=CFG.workers,
type=int,
help=f"number of workers({CFG.workers})")
parser.add_argument("--seed",
default=CFG.seed,
type=int,
help=f"seed({CFG.seed})")
# version
parser.add_argument('--version', type=int)
parser.add_argument('--exp-id', type=int)
# etc
parser.add_argument('--tta', action='store_true', default=False)
parser.add_argument('--use-swa', action='store_true', default=False)
parser.add_argument('--use-snapshot', action='store_true', default=False)
args = parser.parse_args()
CFG.root_path = args.root_path
CFG.save_path = args.save_path
CFG.sub_name = args.sub_name
CFG.batch_size = args.batch_size
CFG.workers = args.workers
CFG.seed = args.seed
CFG.model_path = f"./model/v{args.version}/exp_{args.exp_id}/"
CFG.log_path = f"./log/v{args.version}/exp_{args.exp_id}/"
CFG.tta = args.tta
CFG.use_swa = args.use_swa
CFG.use_snapshot = args.use_snapshot
# get device
CFG.device = get_device()
# load train environment
env = json.load(open(os.path.join(CFG.log_path, 'CFG.json'), 'r'))
for k, v in env.items():
setattr(CFG, k, v)
loss, metric = 0, 0
for fold in range(CFG.n_folds):
fn = os.path.join(CFG.log_path, f"log.fold_{fold}.csv")
score = pd.read_csv(fn).sort_values("val_metric",
ascending=False).iloc[0]
loss += score['val_loss'] / CFG.n_folds
metric += score['val_metric'] / CFG.n_folds
CFG.sub_name = f"submission." \
f"ver_{args.version}." \
f"exp_{args.exp_id}." \
f"loss_{loss:.4f}." \
f"metric_{metric:.4f}.csv"
if CFG.tta:
CFG.sub_name = "tta." + CFG.sub_name
if CFG.use_swa:
CFG.sub_name = "swa." + CFG.sub_name
if CFG.use_snapshot:
CFG.sub_name = "snapshot." + CFG.sub_name
# CFG.batch_size = 1
pprint({k: v for k, v in dict(CFG.__dict__).items() if '__' not in k})
print()
### seed all
seed_everything(CFG.seed)
### Data related logic
# load data
print("Load Raw Data")
_, test_df = load_data(CFG)
print()
# preprocess data
print("Preprocess Data")
test_df = preprocess_data(CFG, test_df)
print()
# get transform
print("Get Transform")
if CFG.tta:
CFG.transform_version = 99
test_transforms = get_transform(CFG)
else:
_, test_transforms = get_transform(CFG)
print()
# dataset
tst_data = MelanomaDataset(CFG, test_df, test_transforms)
final_preds = np.zeros(test_df.shape[0])
# folds
for fold in range(CFG.n_folds):
print(f"========== Fold: {fold} ==========")
if CFG.use_snapshot:
fn = os.path.join(CFG.log_path, f"log.fold_{fold}.csv")
log = pd.read_csv(fn).rename({
"Unnamed: 0": "epoch"
}, axis=1).sort_values("sub_1_score", ascending=False)
print(log.head(5), '\n')
epochs = log['epoch'].values.tolist()[:5]
test_preds = np.zeros(test_df.shape[0])
for epoch in epochs:
print(f"***** EPOCH: {epoch} *****")
model_name = f'model.fold_{fold}.epoch_{epoch}.pt'
learner = Learner(CFG)
learner.load(os.path.join(CFG.model_path, model_name),
f"model_state_dict")
preds, sub_1 = learner.predict(tst_data)
preds = torch.sigmoid(preds.view(-1)).numpy()
sub_1 = sub_1.view(-1).numpy()
# case 1
snapshot_preds = preds
# case 2
snapshot_preds = sub_1
test_preds += snapshot_preds / len(epochs)
else:
# load learner
print("Load Model")
if CFG.use_swa:
model_name = f'model.fold_{fold}.swa.pt'
else:
model_name = f'model.fold_{fold}.best.pt'
learner = Learner(CFG)
learner.load(os.path.join(CFG.model_path, model_name),
f"model_state_dict")
# prediction
if not CFG.tta:
preds, sub_1 = learner.predict(tst_data)
preds = torch.sigmoid(preds.view(-1)).numpy()
sub_1 = sub_1.view(-1).numpy()
# case 1
test_preds = preds
# case 2
test_preds = sub_1
# case 3
# test_preds = (preds + sub_1) / 2
else:
test_preds = np.zeros(test_df.shape[0])
for _ in range(4):
test_preds += torch.sigmoid(
learner.predict(tst_data)[0].view(-1)).numpy() / 4
final_preds += test_preds / CFG.n_folds
print()
ss_df = pd.read_csv(
os.path.join(CFG.root_path, "melanoma-external-malignant-256",
"sample_submission.csv"))
test_df['target'] = final_preds
test_df.set_index("image_name", inplace=True)
ss_df = test_df.loc[ss_df['image_name']].reset_index()[[
'image_name', 'target'
]]
ss_df.to_csv(os.path.join(CFG.save_path, f"{CFG.sub_name}"), index=False)
print(ss_df.head())
device = "cuda"
train_dir = "../data/train_20k"
test_dir = "../data/test_20k"
save_path = "../data/models/200_iter{}.bin"
best_path = "../data/models/200_best.bin"
model = QRN18(target_classes=target_classes)
model.double()
model.to(device)
# Lets add some transformation to make our model translation and scale invarient
to_tensor = torchvision.transforms.ToTensor()
train_dataset = QaidaDataset(train_dir,
transform=get_transform(mode="train"),
max_classes=400)
# Train dataloader should shuffle images
train_dataloader = DataLoader(train_dataset,
batch_size=train_batch_size,
shuffle=True,
num_workers=4)
# Test dataset do not need transformations
test_dataset = QaidaDataset(test_dir,
transform=get_transform(mode="test"),
max_classes=400)
# Test dataloader should not shuffle images
test_dataloader = DataLoader(test_dataset,
batch_size=test_batch_size,
shuffle=False,
def main():
""" main function
"""
### header
parser = argparse.ArgumentParser()
# path
parser.add_argument('--root-path', default=CFG.root_path, help="root path")
parser.add_argument('--log-path', default=CFG.log_path, help="log path")
parser.add_argument('--model-path',
default=CFG.model_path,
help="model path")
parser.add_argument('--pretrained-path', help='pretrained path')
# image
parser.add_argument('--transform-version',
default=0,
type=int,
help="image transform version ex) 0, 1, 2 ...")
parser.add_argument('--image-size',
default=256,
type=int,
help="image size(256)")
# model
parser.add_argument('--model-name',
default=CFG.model_name,
help=f"model name({CFG.model_name})")
parser.add_argument('--backbone-name',
default=CFG.backbone_name,
help=f"backbone name({CFG.backbone_name})")
parser.add_argument('--dropout',
default=CFG.dropout,
type=float,
help=f"dropout({CFG.dropout})")
parser.add_argument('--weight-decay',
default=CFG.weight_decay,
type=float,
help=f"weight decay({CFG.weight_decay})")
# learning
parser.add_argument('--batch-size',
default=CFG.batch_size,
type=int,
help=f"batch size({CFG.batch_size})")
parser.add_argument('--learning-rate',
default=CFG.learning_rate,
type=float,
help=f"learning rate({CFG.learning_rate})")
parser.add_argument('--num-epochs',
default=CFG.num_epochs,
type=int,
help=f"number of epochs({CFG.num_epochs})")
parser.add_argument("--swa",
action="store_true",
help="do stochastic weight averaging")
# etc
parser.add_argument("--seed",
default=CFG.seed,
type=int,
help=f"seed({CFG.seed})")
parser.add_argument("--workers",
default=CFG.workers,
type=int,
help=f"number of workers({CFG.workers})")
parser.add_argument("--debug", action="store_true", help="debug mode")
args = parser.parse_args()
# path
CFG.root_path = args.root_path
CFG.model_path = args.model_path
CFG.log_path = args.log_path
CFG.pretrained_path = args.pretrained_path
# image
CFG.transform_version = args.transform_version
CFG.image_size = args.image_size
# model
CFG.model_name = args.model_name
CFG.backbone_name = args.backbone_name
CFG.dropout = args.dropout
CFG.weight_decay = args.weight_decay
# learning
CFG.batch_size = args.batch_size
CFG.learning_rate = args.learning_rate
CFG.num_epochs = args.num_epochs
CFG.swa = args.swa
# etc
CFG.seed = args.seed
CFG.workers = args.workers
CFG.debug = args.debug
# get device
CFG.device = get_device()
# get version
_, version, _ = sys.argv[0].split('/')
CFG.version = version
# update log path
if not CFG.debug:
CFG.log_path = os.path.join(CFG.log_path, CFG.version)
os.makedirs(CFG.log_path, exist_ok=True)
CFG.log_path = os.path.join(
CFG.log_path, f'exp_{get_exp_id(CFG.log_path, prefix="exp_")}')
os.makedirs(CFG.log_path, exist_ok=True)
else:
CFG.log_path = os.path.join(CFG.log_path, "debug")
os.makedirs(CFG.log_path, exist_ok=True)
CFG.log_path = os.path.join(CFG.log_path, "debug")
os.makedirs(CFG.log_path, exist_ok=True)
# update model path
if not CFG.debug:
CFG.model_path = os.path.join(CFG.model_path, version)
os.makedirs(CFG.model_path, exist_ok=True)
CFG.model_path = os.path.join(
CFG.model_path, f'exp_{get_exp_id(CFG.model_path, prefix="exp_")}')
os.makedirs(CFG.model_path, exist_ok=True)
else:
CFG.model_path = os.path.join(CFG.model_path, "debug")
os.makedirs(CFG.model_path, exist_ok=True)
CFG.model_path = os.path.join(CFG.model_path, "debug")
os.makedirs(CFG.model_path, exist_ok=True)
pprint({k: v for k, v in dict(CFG.__dict__).items() if '__' not in k})
json.dump({k: v
for k, v in dict(CFG.__dict__).items() if '__' not in k},
open(os.path.join(CFG.log_path, 'CFG.json'), "w"))
print()
### seed all
seed_everything(CFG.seed)
### data related
# load data
print("Load Raw Data")
data_df, test_df = load_data(CFG)
# preprocess data
print("Preprocess Data")
data_df = preprocess_data(CFG, data_df)
# split data into train with valid
print("Split Data")
data_df = split_data(CFG, data_df)
# get transform
print("Get Transform")
train_transforms, test_transforms = get_transform(CFG)
# train test split
for fold in range(CFG.n_folds):
print(f"\nValidation Fold: {fold}")
train_df = data_df[data_df['fold'] != fold].reset_index(drop=True)
valid_df = data_df[data_df['fold'] == fold].reset_index(drop=True)
print(
f"... Train Shape: {train_df.shape}, Valid Shape: {valid_df.shape}"
)
# dataset
trn_data = MelanomaDataset(CFG, train_df, train_transforms)
val_data = MelanomaDataset(CFG, valid_df, test_transforms)
### model related
# get learner
learner = Learner(CFG)
learner.name = f"model.fold_{fold}"
if CFG.pretrained_path:
print("Load Pretrained Model")
print(f"... Pretrained Info - {CFG.pretrained_path}")
learner.load(CFG.pretrained_path, f"model_state_dict")
# get model
if CFG.pretrained_path:
print(f"Get Model")
model = learner.best_model.to(CFG.deivce)
else:
print(f"Get Model")
model = get_model(CFG)
model = model.to(CFG.device)
# get optimizer
module_names = list(model.state_dict())
no_decay = ['bias']
for m in module_names:
if 'running_mean' in m:
no_decay.append(m.split('.running_mean')[0])
param_optimizer = list(model.named_parameters())
optimizer_grouped_parameters = [{
'params': [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
CFG.weight_decay
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
optimizer = optim.AdamW(optimizer_grouped_parameters,
CFG.learning_rate)
model, optimizer = amp.initialize(model, optimizer, verbosity=0)
if torch.cuda.device_count() > 1:
model = nn.DataParallel(model)
# get optimizer
# optimizer = optim.Adam(model.parameters(), lr=CFG.learning_rate)
if CFG.swa:
optimizer = torchcontrib.optim.SWA(optimizer)
# get scheduler
# scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
# optimizer, mode='min', patience=1, verbose=False, factor=0.2)
# scheduler = CosineAnnealingLRWarmup(optimizer, T_min=int(CFG.num_epochs / 5), T_max=CFG.num_epochs)
T = 1
scheduler = CosineAnnealingLRWarmup(optimizer,
T_min=0,
T_max=CFG.num_epochs // T)
### train related
# train model
learner.train(trn_data, val_data, model, optimizer, scheduler)
print()
except EOFError:
return
def print_cache(lst, heading):
print(heading)
print('---')
pprint.pprint(lst)
if __name__ == '__main__':
endoscope_chesspts = list(load_all('camera_data/endoscope_chesspts.p'))
# camera_info = list(load_all('camera_data/camera_info.p'))
left_chesspts = np.matrix(list(load_all('camera_data/left_chesspts'))[0])
right_chesspts = np.matrix(list(load_all('camera_data/right_chesspts'))[0])
z = np.zeros((25, 1))
left_chesspts = np.hstack((left_chesspts, z))
right_chesspts = np.hstack((right_chesspts, z))
print_cache(endoscope_chesspts, "ENDOSCOPE CHESSPOINTS")
# print_cache(camera_info, "CAMERA INFO")
print_cache(left_chesspts, "LEFT CHESSPOINTS")
print_cache(right_chesspts, "RIGHT CHESSPOINTS")
TL_R = transform.get_transform("Left Camera", "Right Camera",
left_chesspts, right_chesspts)
L_R = transform.transform_data("Left Camera", "Right Camera",
left_chesspts, TL_R, right_chesspts)