def load_model(config_path):
config = load_config(config_path)
if 'COLAB_GPU' in os.environ:
config.work_dir = '/content/drive/My Drive/kaggle_cloud/' + config.work_dir
elif 'KAGGLE_WORKING_DIR' in os.environ:
config.work_dir = '/kaggle/working/' + config.work_dir
if config.checkpoint_path == None:
config.checkpoint_path = config.work_dir + '/checkpoints/best.pth'
print(config.checkpoint_path)
if config.model.arch == 'Classification':
model = CustomNet(config.model.encoder,
config.data.num_classes,
pretrained=False)
else:
# create segmentation model with pre-trained encoder
model = getattr(smp, config.model.arch)(
encoder_name=config.model.encoder,
encoder_weights=None,
classes=config.data.num_classes,
activation=None,
)
model.to(config.device)
model.eval()
checkpoint = load_checkpoint(config.checkpoint_path)
model.load_state_dict(checkpoint['model_state_dict'])
return model
def _preprocess_model_for_stage(self, stage: str, model: Model):
stage_index = self.stages.index(stage)
if stage_index > 0:
checkpoint_path = f"{self.logdir}/checkpoints/best.pth"
checkpoint = utils.load_checkpoint(checkpoint_path)
utils.unpack_checkpoint(checkpoint, model=model)
return model
def run_cls(config_file_cls):
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
# ------------------------------------------------------------------------------------------------------------
# 1. classification inference
# ------------------------------------------------------------------------------------------------------------
config = load_config(config_file_cls)
validloader = make_loader(data_folder=config.data.train_dir,
df_path=config.data.train_df_path,
phase='valid',
batch_size=config.train.batch_size,
num_workers=config.num_workers,
idx_fold=config.data.params.idx_fold,
transforms=get_transforms(
config.transforms.test),
num_classes=config.data.num_classes,
task='cls')
model = CustomNet(config.model.encoder, config.data.num_classes)
model.to(config.device)
model.eval()
checkpoint = load_checkpoint(f"{config.work_dir}/checkpoints/best.pth")
model.load_state_dict(checkpoint['model_state_dict'])
all_predictions = []
all_targets = []
with torch.no_grad():
for i, (batch_images, batch_targets) in enumerate(tqdm(validloader)):
batch_images = batch_images.to(config.device)
batch_preds = predict_batch(model,
batch_images,
tta=config.test.tta,
task='cls')
all_targets.append(batch_targets)
all_predictions.append(batch_preds)
all_predictions = np.concatenate(all_predictions)
all_targets = np.concatenate(all_targets)
# evaluation
all_accuracy_scores = []
all_f1_scores = []
thresholds = np.linspace(0.1, 0.9, 9)
for th in thresholds:
accuracy = accuracy_score(all_targets > th, all_predictions > th)
f1 = f1_score(all_targets > th,
all_predictions > th,
average='samples')
all_accuracy_scores.append(accuracy)
all_f1_scores.append(f1)
for th, score in zip(thresholds, all_accuracy_scores):
print('validation accuracy for threshold {} = {}'.format(th, score))
for th, score in zip(thresholds, all_f1_scores):
print('validation f1 score for threshold {} = {}'.format(th, score))
np.save('valid_preds', all_predictions)
def predict_loader(
self,
*,
loader: DataLoader,
model: Model = None,
resume: str = None,
fp16: Union[Dict, bool] = None,
initial_seed: int = 42,
) -> Generator:
"""
Runs model inference on PyTorch Dataloader and returns
python generator with model predictions from `runner.predict_batch`.
Cleans up the experiment info to avoid possible collisions.
Sets `is_train_loader` and `is_valid_loader` to `False` while
keeping `is_infer_loader` as True. Moves model to evaluation mode.
Args:
loader: loader to predict
model: model to use for prediction
resume: path to checkpoint to resume
fp16 (Union[Dict, bool]): fp16 usage flag
initial_seed: seed to use before prediction
Yields:
bathes with model predictions
"""
if isinstance(fp16, bool) and fp16:
fp16 = {"opt_level": "O1"}
if model is not None:
self.model = model
assert self.model is not None
if resume is not None:
checkpoint = utils.load_checkpoint(resume)
utils.unpack_checkpoint(checkpoint, model=self.model)
self.experiment = None
utils.set_global_seed(initial_seed)
(model, _, _, _, device) = utils.process_components( # noqa: WPS122
model=self.model,
distributed_params=fp16,
device=self.device,
)
self._prepare_inner_state(
stage="infer",
model=model,
device=device,
is_train_loader=False,
is_valid_loader=False,
is_infer_loader=True,
)
utils.maybe_recursive_call(self.model, "train", mode=False)
utils.set_global_seed(initial_seed)
for batch in loader:
yield self.predict_batch(batch)
def _get_optimizer(self, *, model_params, **params):
load_from_previous_stage = \
params.pop("load_from_previous_stage", False)
optimizer = OPTIMIZERS.get_from_params(**params, params=model_params)
if load_from_previous_stage:
checkpoint_path = f"{self.logdir}/checkpoints/best.pth"
checkpoint = utils.load_checkpoint(checkpoint_path)
utils.unpack_checkpoint(checkpoint, optimizer=optimizer)
for key, value in params.items():
for pg in optimizer.param_groups:
pg[key] = value
return optimizer
def predict_loader(
self,
*,
loader: DataLoader,
model: Model = None,
resume: str = None,
fp16: Union[Dict, bool] = None,
initial_seed: int = 42,
) -> Generator:
"""
Runs model inference on PyTorch Dataloader and returns
python Generator with model predictions from `runner.predict_batch`
Args:
loader (DataLoader): loader to predict
model (Model): model to use for prediction
resume (str): path to checkpoint to resume
fp16 (Union[Dict, bool]): fp16 usage flag
initial_seed (int): seed to use before prediction
Yields:
bathes with model predictions
"""
if isinstance(fp16, bool) and fp16:
fp16 = {"opt_level": "O1"}
if model is not None:
self.model = model
assert self.model is not None
if resume is not None:
checkpoint = utils.load_checkpoint(resume)
utils.unpack_checkpoint(checkpoint, model=self.model)
( # noqa: WPS122
self.model,
_,
_,
_,
self.device,
) = utils.process_components(
model=self.model,
distributed_params=fp16,
device=self.device,
)
utils.set_global_seed(initial_seed)
for batch in loader:
yield self.predict_batch(batch)
def run_cls(config_file_cls):
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
# ------------------------------------------------------------------------------------------------------------
# 1. classification inference
# ------------------------------------------------------------------------------------------------------------
config = load_config(config_file_cls)
model = CustomNet(config.model.encoder, config.data.num_classes)
testloader = make_loader(
data_folder=config.data.test_dir,
df_path=config.data.sample_submission_path,
phase='test',
batch_size=config.test.batch_size,
num_workers=config.num_workers,
transforms=get_transforms(config.transforms.test),
num_classes=config.data.num_classes,
)
model.to(config.device)
model.eval()
checkpoint = load_checkpoint(f"{config.work_dir}/checkpoints/best.pth")
model.load_state_dict(checkpoint['model_state_dict'])
all_fnames = []
all_predictions = []
with torch.no_grad():
for i, (batch_fnames, batch_images) in enumerate(tqdm(testloader)):
batch_images = batch_images.to(config.device)
batch_preds = predict_batch(model,
batch_images,
tta=config.test.tta,
task='cls')
all_fnames.extend(batch_fnames)
all_predictions.append(batch_preds)
all_predictions = np.concatenate(all_predictions)
np.save('all_preds', all_predictions)
df = pd.DataFrame(data=all_predictions, index=all_fnames)
df.to_csv('cls_preds.csv', index=False)
df.to_csv(f"{config.work_dir}/cls_preds.csv", index=False)
def load_checkpoint(*, filename, state: RunnerState):
if os.path.isfile(filename):
print(f"=> loading checkpoint {filename}")
checkpoint = utils.load_checkpoint(filename)
state.epoch = checkpoint["epoch"]
utils.unpack_checkpoint(checkpoint,
model=state.model,
criterion=state.criterion,
optimizer=state.optimizer,
scheduler=state.scheduler)
print(
f"loaded checkpoint {filename} (epoch {checkpoint['epoch']})")
else:
raise Exception(f"No checkpoint found at {filename}")
def load_model(config_path, checkpoint_path, fold=0):
config = load_config(config_path)
if not 'fold' in config.work_dir:
config.work_dir = config.work_dir + '_fold{}'.format(fold)
model = CenterNetFPN(
slug=config.model.encoder,
num_classes=len(config.data.features),
)
model.to(config.device)
model.eval()
checkpoint = load_checkpoint(checkpoint_path)
print('load model from {}'.format(checkpoint_path))
model.load_state_dict(checkpoint['model_state_dict'])
return model
def load_model(config_path):
config = load_config(config_path)
print(config.checkpoint_path)
if config.model.arch == 'Classification':
model = CustomNet(config.model.encoder, config.data.num_classes, pretrained=False)
else:
# create segmentation model with pre-trained encoder
model = getattr(smp, config.model.arch)(
encoder_name=config.model.encoder,
encoder_weights=None,
classes=config.data.num_classes,
activation=None,
)
model.to(config.device)
model.eval()
checkpoint = load_checkpoint(config.checkpoint_path)
model.load_state_dict(checkpoint['model_state_dict'])
return model
def _get_optimizer(self, *, model_params, **params):
load_from_previous_stage = \
params.pop("load_from_previous_stage", False)
optimizer = OPTIMIZERS.get_from_params(**params, params=model_params)
if load_from_previous_stage:
checkpoint_path = f"{self.logdir}/checkpoints/best_full.pth"
checkpoint = utils.load_checkpoint(checkpoint_path)
utils.unpack_checkpoint(checkpoint, optimizer=optimizer)
# move optimizer to device
device = get_device()
for param in model_params:
param = param["params"][0]
state = optimizer.state[param]
for key, value in state.items():
state[key] = any2device(value, device)
# update optimizer params
for key, value in params.items():
for pg in optimizer.param_groups:
pg[key] = value
return optimizer
def _get_optimizer(self, stage: str, model: Union[Model, Dict[str, Model]],
**params) -> Optimizer:
# @TODO 1: refactoring; this method is too long
# @TODO 2: load state dicts for schedulers & criterion
layerwise_params = params.pop("layerwise_params", OrderedDict())
no_bias_weight_decay = params.pop("no_bias_weight_decay", True)
# linear scaling rule from https://arxiv.org/pdf/1706.02677.pdf
lr_scaling_params = params.pop("lr_linear_scaling", None)
if lr_scaling_params:
data_params = dict(self.stages_config[stage]["data_params"])
batch_size = data_params.get("batch_size")
per_gpu_scaling = data_params.get("per_gpu_scaling", False)
distributed_rank = utils.get_rank()
distributed = distributed_rank > -1
if per_gpu_scaling and not distributed:
num_gpus = max(1, torch.cuda.device_count())
batch_size *= num_gpus
base_lr = lr_scaling_params.get("lr")
base_batch_size = lr_scaling_params.get("base_batch_size", 256)
lr_scaling = batch_size / base_batch_size
params["lr"] = base_lr * lr_scaling # scale default lr
else:
lr_scaling = 1.0
# getting model parameters
model_key = params.pop("_model", None)
if model_key is None:
assert isinstance(
model, nn.Module
), "model is key-value, but optimizer has no specified model"
model_params = utils.process_model_params(model, layerwise_params,
no_bias_weight_decay,
lr_scaling)
elif isinstance(model_key, str):
model_params = utils.process_model_params(
model[model_key],
layerwise_params,
no_bias_weight_decay,
lr_scaling,
)
elif isinstance(model_key, (list, tuple)):
model_params = []
for model_key_ in model_key:
model_params_ = utils.process_model_params(
model[model_key_],
layerwise_params,
no_bias_weight_decay,
lr_scaling,
)
model_params.extend(model_params_)
else:
raise ValueError("unknown type of model_params")
load_from_previous_stage = params.pop("load_from_previous_stage",
False)
optimizer_key = params.pop("optimizer_key", None)
optimizer = OPTIMIZERS.get_from_params(**params, params=model_params)
if load_from_previous_stage and self.stages.index(stage) != 0:
checkpoint_path = f"{self.logdir}/checkpoints/best_full.pth"
checkpoint = utils.load_checkpoint(checkpoint_path)
dict2load = optimizer
if optimizer_key is not None:
dict2load = {optimizer_key: optimizer}
utils.unpack_checkpoint(checkpoint, optimizer=dict2load)
# move optimizer to device
device = utils.get_device()
for param in model_params:
param = param["params"][0]
state = optimizer.state[param]
for key, value in state.items():
state[key] = utils.any2device(value, device)
# update optimizer params
for key, value in params.items():
for pg in optimizer.param_groups:
pg[key] = value
return optimizer
def run_seg(config_file_seg):
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
# ------------------------------------------------------------------------------------------------------------
# 2. segmentation inference
# ------------------------------------------------------------------------------------------------------------
config = load_config(config_file_seg)
validloader = make_loader(
data_folder=config.data.train_dir,
df_path=config.data.train_df_path,
phase='valid',
batch_size=config.train.batch_size,
num_workers=config.num_workers,
idx_fold=config.data.params.idx_fold,
transforms=get_transforms(config.transforms.test),
num_classes=config.data.num_classes,
)
# create segmentation model with pre-trained encoder
model = getattr(smp, config.model.arch)(
encoder_name=config.model.encoder,
encoder_weights=config.model.pretrained,
classes=config.data.num_classes,
activation=None,
)
model.to(config.device)
model.eval()
checkpoint = load_checkpoint(f"{config.work_dir}/checkpoints/best.pth")
model.load_state_dict(checkpoint['model_state_dict'])
all_dice = {}
min_sizes = [100, 300, 500, 750, 1000, 1500, 2000, 3000]
for min_size in min_sizes:
all_dice[min_size] = {}
for cls in range(config.data.num_classes):
all_dice[min_size][cls] = []
with torch.no_grad():
for i, (batch_images, batch_masks) in enumerate(tqdm(validloader)):
batch_images = batch_images.to(config.device)
batch_preds = predict_batch(model,
batch_images,
tta=config.test.tta)
batch_masks = batch_masks.cpu().numpy()
for masks, preds in zip(batch_masks, batch_preds):
for cls in range(config.data.num_classes):
for min_size in min_sizes:
pred, _ = post_process(preds[cls, :, :],
config.test.best_threshold,
min_size)
mask = masks[cls, :, :]
all_dice[min_size][cls].append(dice_score(pred, mask))
for cls in range(config.data.num_classes):
for min_size in min_sizes:
all_dice[min_size][cls] = sum(all_dice[min_size][cls]) / len(
all_dice[min_size][cls])
dict_to_json(all_dice, config.work_dir + '/threshold_search.json')
if config.data.num_classes == 4:
defect_class = cls + 1
else:
defect_class = cls
print('average dice score for class{} for min_size {}: {}'.format(
defect_class, min_size, all_dice[min_size][cls]))
def trace_model_from_checkpoint(
logdir: Path,
method_name: str,
checkpoint_name: str,
stage: str = None,
loader: Union[str, int] = None,
mode: str = "eval",
requires_grad: bool = False,
opt_level: str = None,
device: Device = "cpu",
):
"""Traces model using created experiment and runner.
Args:
logdir (Union[str, Path]): Path to Catalyst logdir with model
checkpoint_name (str): Name of model checkpoint to use
stage (str): experiment's stage name
loader (Union[str, int]): experiment's loader name or its index
method_name (str): Model's method name that will be
used as entrypoint during tracing
mode (str): Mode for model to trace (``train`` or ``eval``)
requires_grad (bool): Flag to use grads
opt_level (str): AMP FP16 init level
device (str): Torch device
Returns:
the traced model
"""
config_path = logdir / "configs" / "_config.json"
checkpoint_path = logdir / "checkpoints" / f"{checkpoint_name}.pth"
print("Load config")
config: Dict[str, dict] = utils.load_config(config_path)
runner_params = config.get("runner_params", {}) or {}
# Get expdir name
config_expdir = Path(config["args"]["expdir"])
# We will use copy of expdir from logs for reproducibility
expdir = Path(logdir) / "code" / config_expdir.name
print("Import experiment and runner from logdir")
ExperimentType, RunnerType = utils.import_experiment_and_runner(expdir)
experiment: Experiment = ExperimentType(config)
print(f"Load model state from checkpoints/{checkpoint_name}.pth")
if stage is None:
stage = list(experiment.stages)[0]
model = experiment.get_model(stage)
checkpoint = utils.load_checkpoint(checkpoint_path)
utils.unpack_checkpoint(checkpoint, model=model)
runner: RunnerType = RunnerType(**runner_params)
runner.model, runner.device = model, device
if loader is None:
loader = 0
batch = experiment.get_native_batch(stage, loader)
print("Tracing")
traced = trace.trace_model(
model=model,
runner=runner,
batch=batch,
method_name=method_name,
mode=mode,
requires_grad=requires_grad,
opt_level=opt_level,
device=device,
)
print("Done")
return traced
def run_seg(config_file_seg):
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
# ------------------------------------------------------------------------------------------------------------
# 2. segmentation inference
# ------------------------------------------------------------------------------------------------------------
config = load_config(config_file_seg)
model = getattr(smp, config.model.arch)(
encoder_name=config.model.encoder,
encoder_weights=config.model.pretrained,
classes=config.data.num_classes,
activation=None,
)
if os.path.exists('cls_preds.csv'):
testloader = make_loader(data_folder=config.data.test_dir,
df_path='cls_preds.csv',
phase='filtered_test',
batch_size=config.test.batch_size,
num_workers=config.num_workers,
transforms=get_transforms(
config.transforms.test))
else:
testloader = make_loader(data_folder=config.data.test_dir,
df_path=config.data.sample_submission_path,
phase='test',
batch_size=config.test.batch_size,
num_workers=config.num_workers,
transforms=get_transforms(
config.transforms.test))
model.to(config.device)
model.eval()
checkpoint = load_checkpoint(f"{config.work_dir}/checkpoints/best.pth")
model.load_state_dict(checkpoint['model_state_dict'])
if os.path.exists(config.work_dir + '/threshold_search.json'):
with open(config.work_dir + '/threshold_search.json') as json_file:
data = json.load(json_file)
df = pd.DataFrame(data)
min_sizes = list(df.T.idxmax().values.astype(int))
print('load best threshold from validation:', min_sizes)
else:
min_sizes = config.test.min_size
print('load default threshold:', min_sizes)
predictions = []
with torch.no_grad():
for i, (batch_fnames, batch_images) in enumerate(tqdm(testloader)):
batch_images = batch_images.to(config.device)
batch_preds = predict_batch(model,
batch_images,
tta=config.test.tta)
for fname, preds in zip(batch_fnames, batch_preds):
if config.data.num_classes == 4:
for cls in range(preds.shape[0]):
mask = preds[cls, :, :]
mask, num = post_process(mask,
config.test.best_threshold,
min_sizes[cls])
rle = mask2rle(mask)
name = fname + f"_{cls + 1}"
predictions.append([name, rle])
else: # == 5
for cls in range(1, 5):
mask = preds[cls, :, :]
mask, num = post_process(mask,
config.test.best_threshold,
min_sizes[cls])
rle = mask2rle(mask)
name = fname + f"_{cls}"
predictions.append([name, rle])
# ------------------------------------------------------------------------------------------------------------
# submission
# ------------------------------------------------------------------------------------------------------------
df = pd.DataFrame(predictions,
columns=['ImageId_ClassId', 'EncodedPixels'])
df.to_csv(config.work_dir + "/submission.csv", index=False)
def main(args, _=None):
"""Run the ``catalyst-data text2embeddings`` script."""
batch_size = args.batch_size
num_workers = args.num_workers
max_length = args.max_length
pooling_groups = args.pooling.split(",")
utils.set_global_seed(args.seed)
utils.prepare_cudnn(args.deterministic, args.benchmark)
if hasattr(args, "in_huggingface"):
model_config = BertConfig.from_pretrained(args.in_huggingface)
model_config.output_hidden_states = args.output_hidden_states
model = BertModel.from_pretrained(args.in_huggingface,
config=model_config)
tokenizer = BertTokenizer.from_pretrained(args.in_huggingface)
else:
model_config = BertConfig.from_pretrained(args.in_config)
model_config.output_hidden_states = args.output_hidden_states
model = BertModel(config=model_config)
tokenizer = BertTokenizer.from_pretrained(args.in_vocab)
if hasattr(args, "in_model"):
checkpoint = utils.load_checkpoint(args.in_model)
checkpoint = {"model_state_dict": checkpoint}
utils.unpack_checkpoint(checkpoint=checkpoint, model=model)
model = model.eval()
model, _, _, _, device = utils.process_components(model=model)
df = pd.read_csv(args.in_csv)
df = df.dropna(subset=[args.txt_col])
df.to_csv(f"{args.out_prefix}.df.csv", index=False)
df = df.reset_index().drop("index", axis=1)
df = list(df.to_dict("index").values())
num_samples = len(df)
open_fn = LambdaReader(
input_key=args.txt_col,
output_key=None,
lambda_fn=partial(
tokenize_text,
strip=args.strip,
lowercase=args.lowercase,
remove_punctuation=args.remove_punctuation,
),
tokenizer=tokenizer,
max_length=max_length,
)
dataloader = utils.get_loader(
df,
open_fn,
batch_size=batch_size,
num_workers=num_workers,
)
features = {}
dataloader = tqdm(dataloader) if args.verbose else dataloader
with torch.no_grad():
for idx, batch in enumerate(dataloader):
batch = utils.any2device(batch, device)
bert_output = model(**batch)
mask = (batch["attention_mask"].unsqueeze(-1)
if args.mask_for_max_length else None)
if utils.check_ddp_wrapped(model):
# using several gpu
hidden_size = model.module.config.hidden_size
hidden_states = model.module.config.output_hidden_states
else:
# using cpu or one gpu
hidden_size = model.config.hidden_size
hidden_states = model.config.output_hidden_states
features_ = process_bert_output(
bert_output=bert_output,
hidden_size=hidden_size,
output_hidden_states=hidden_states,
pooling_groups=pooling_groups,
mask=mask,
)
# create storage based on network output
if idx == 0:
for key, value in features_.items():
name_ = key if isinstance(key, str) else f"{key:02d}"
_, embedding_size = value.shape
features[name_] = np.memmap(
f"{args.out_prefix}.{name_}.npy",
dtype=np.float32,
mode="w+",
shape=(num_samples, embedding_size),
)
indices = np.arange(idx * batch_size,
min((idx + 1) * batch_size, num_samples))
for key, value in features_.items():
name_ = key if isinstance(key, str) else f"{key:02d}"
features[name_][indices] = _detach(value)
def main(args, _=None):
batch_size = args.batch_size
num_workers = args.num_workers
max_length = args.max_length
pooling_groups = args.pooling.split(",")
utils.set_global_seed(args.seed)
utils.prepare_cudnn(args.deterministic, args.benchmark)
model_config = BertConfig.from_pretrained(args.in_config)
model_config.output_hidden_states = args.output_hidden_states
model = BertModel(config=model_config)
checkpoint = utils.load_checkpoint(args.in_model)
checkpoint = {"model_state_dict": checkpoint}
utils.unpack_checkpoint(checkpoint=checkpoint, model=model)
model = model.eval()
model, _, _, _, device = utils.process_components(model=model)
tokenizer = BertTokenizer.from_pretrained(args.in_vocab)
df = pd.read_csv(args.in_csv)
df = df.dropna(subset=[args.txt_col])
df.to_csv(f"{args.out_prefix}.df.csv", index=False)
df = df.reset_index().drop("index", axis=1)
df = list(df.to_dict("index").values())
num_samples = len(df)
open_fn = LambdaReader(
input_key=args.txt_col,
output_key=None,
lambda_fn=get_features,
tokenizer=tokenizer,
max_length=max_length,
)
dataloader = utils.get_loader(
df,
open_fn,
batch_size=batch_size,
num_workers=num_workers,
)
features = {}
poolings = {}
dataloader = tqdm(dataloader) if args.verbose else dataloader
with torch.no_grad():
for idx, batch in enumerate(dataloader):
batch = utils.any2device(batch, device)
features_ = model(**batch)
# create storage based on network output
if idx == 0:
# class
_, embedding_size = features_[1].shape
features["class"] = np.memmap(
f"{args.out_prefix}.class.npy",
dtype=np.float32,
mode="w+",
shape=(num_samples, embedding_size),
)
if args.output_hidden_states:
# all embeddings
for i, feature_ in enumerate(features_[2]):
name_ = f"embeddings_{i + 1:02d}"
_, _, embedding_size = feature_.shape
poolings[name_] = LamaPooling(
features_in=embedding_size,
groups=pooling_groups,
)
features[name_] = np.memmap(
f"{args.out_prefix}.{name_}.npy",
dtype=np.float32,
mode="w+",
shape=(num_samples, embedding_size),
)
else:
# last
_, _, embedding_size = features_[0].shape
poolings["last"] = LamaPooling(
features_in=embedding_size,
groups=pooling_groups,
)
features["last"] = np.memmap(
f"{args.out_prefix}.last.npy",
dtype=np.float32,
mode="w+",
shape=(num_samples, embedding_size),
)
indices = np.arange(idx * batch_size,
min((idx + 1) * batch_size, num_samples))
features["class"][indices] = _detach(features_[1])
if args.output_hidden_states:
# all embeddings
for i, feature_ in enumerate(features_[2]):
name_ = f"embeddings_{i + 1:02d}"
feature_ = poolings[name_](feature_)
features[name_][indices] = _detach(feature_)
else:
feature_ = poolings[name_](features_[0])
features["last"][indices] = _detach(feature_)
def main_kaggle_smp(path_dataset='/dataset/kaggle/understanding_cloud_organization',
ENCODER='resnet50',
ENCODER_WEIGHTS='imagenet',
num_workers=0,
batch_size=8,
epochs=19,
debug=False,
exec_catalyst=True,
logdir="/src/logs/segmentation",
pretrained=True
):
# below line is potential input args
# (name_dataset='eurosat', lr=0.0001, wd=0, ratio=0.9, batch_size=32, workers=4, epochs=15, num_gpus=1,
# resume=None, dir_weights='./weights'):
torch.backends.cudnn.benchmark = True
# Dataset
train, sub = get_meta_info_table(path_dataset)
train_ids, valid_ids, test_ids = prepare_dataset(train, sub)
preprocessing_fn = smp.encoders.get_preprocessing_fn(ENCODER, ENCODER_WEIGHTS)
train_dataset = CloudDataset(df=train, datatype='train', img_ids=train_ids, transforms=get_training_augmentation(),
preprocessing=get_preprocessing(preprocessing_fn), path=path_dataset)
valid_dataset = CloudDataset(df=train, datatype='valid', img_ids=valid_ids,
transforms=get_validation_augmentation(),
preprocessing=get_preprocessing(preprocessing_fn), path=path_dataset)
# DataLoader
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=num_workers)
valid_loader = DataLoader(valid_dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers)
loaders = {
"train": train_loader,
"valid": valid_loader
}
# todo: check how to used device in this case
DEVICE = 'cuda'
if debug:
device = 'cpu'
else:
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
ACTIVATION = None
model = smp.Unet(
encoder_name=ENCODER,
encoder_weights=ENCODER_WEIGHTS,
classes=4,
activation=ACTIVATION,
)
images, labels = next(iter(train_loader))
model.to(device)
print(model)
print(summary(model, input_size=tuple(images.shape[1:])))
# use smp epoch
# num_epochs = 19
# model, criterion, optimizer
optimizer = torch.optim.Adam([
{'params': model.decoder.parameters(), 'lr': 1e-2},
{'params': model.encoder.parameters(), 'lr': 1e-3},
])
scheduler = ReduceLROnPlateau(optimizer, factor=0.15, patience=2)
criterion = smp.utils.losses.DiceLoss(eps=1.) # smp.utils.losses.BCEDiceLoss(eps=1.)
if not pretrained:
# catalyst
if exec_catalyst:
device = utils.get_device()
runner = SupervisedRunner(device=device)
# train model
runner.train(
model=model,
criterion=criterion,
optimizer=optimizer,
scheduler=scheduler,
loaders=loaders,
callbacks=[DiceCallback(), EarlyStoppingCallback(patience=5, min_delta=0.001)],
logdir=logdir,
num_epochs=epochs,
verbose=True
)
# # prediction
# encoded_pixels = []
# loaders = {"infer": valid_loader}
# runner.infer(
# model=model,
# loaders=loaders,
# callbacks=[
# CheckpointCallback(
# resume=f"{logdir}/checkpoints/best.pth"),
# InferCallback()
# ],
# )
# valid_masks = []
#
# # todo: where .pth?
# # todo: from here
# valid_num = valid_dataset.__len__()
# probabilities = np.zeros((valid_num * 4, 350, 525))
# for i, (batch, output) in enumerate(tqdm(zip(
# valid_dataset, runner.callbacks[0].predictions["logits"]))):
# image, mask = batch
# for m in mask:
# if m.shape != (350, 525):
# m = cv2.resize(m, dsize=(525, 350), interpolation=cv2.INTER_LINEAR)
# valid_masks.append(m)
#
# for j, probability in enumerate(output):
# if probability.shape != (350, 525):
# probability = cv2.resize(probability, dsize=(525, 350), interpolation=cv2.INTER_LINEAR)
# probabilities[valid_num * 4 + j, :, :] = probability
#
# # todo: from here
# class_params = {}
# for class_id in range(4):
# print(class_id)
# attempts = []
# for t in range(0, 100, 5):
# t /= 100
# for ms in [0, 100, 1200, 5000, 10000]:
# masks = []
# for i in range(class_id, len(probabilities), 4):
# probability = probabilities[i]
# predict, num_predict = post_process(sigmoid(probability), t, ms)
# masks.append(predict)
#
# d = []
# for i, j in zip(masks, valid_masks[class_id::4]):
# if (i.sum() == 0) & (j.sum() == 0):
# d.append(1)
# else:
# d.append(dice(i, j))
#
# attempts.append((t, ms, np.mean(d)))
#
# attempts_df = pd.DataFrame(attempts, columns=['threshold', 'size', 'dice'])
#
# attempts_df = attempts_df.sort_values('dice', ascending=False)
# print(attempts_df.head())
# best_threshold = attempts_df['threshold'].values[0]
# best_size = attempts_df['size'].values[0]
#
# class_params[class_id] = (best_threshold, best_size)
else:
for epoch in trange(epochs, desc="Epochs"):
metrics_train = train_epoch(model, train_loader, criterion, optimizer, device)
metrics_eval = eval_epoch(model, valid_loader, criterion, device)
scheduler.step(metrics_eval['valid_loss'])
print(f'epoch: {epoch} ', metrics_train, metrics_eval)
else:
if exec_catalyst:
device = utils.get_device()
checkpoint = utils.load_checkpoint(f'{logdir}/checkpoints/best_full.pth')
utils.unpack_checkpoint(checkpoint, model=model)
runner = SupervisedRunner(model=model)
# prediction with infer
encoded_pixels = []
loaders = {"infer": valid_loader}
runner.infer(
model=model,
loaders=loaders,
callbacks=[
CheckpointCallback(
resume=f"{logdir}/checkpoints/best.pth"),
InferCallback()
],
)
# todo: jupyterで確認中
valid_masks = []
valid_num = valid_dataset.__len__()
probabilities = np.zeros((valid_num * 4, 350, 525))
for i, (batch, output) in enumerate(tqdm(zip(
valid_dataset, runner.callbacks[0].predictions["logits"]))):
image, mask = batch
for m in mask:
if m.shape != (350, 525):
m = cv2.resize(m, dsize=(525, 350), interpolation=cv2.INTER_LINEAR)
valid_masks.append(m)
for j, probability in enumerate(output):
if probability.shape != (350, 525):
probability = cv2.resize(probability, dsize=(525, 350), interpolation=cv2.INTER_LINEAR)
probabilities[i * 4 + j, :, :] = probability
class_params = {}
for class_id in range(4):
print(class_id)
attempts = []
for t in range(0, 100, 5):
t /= 100
for ms in [0, 100, 1200, 5000, 10000]:
masks = []
for i in range(class_id, len(probabilities), 4):
probability = probabilities[i]
predict, num_predict = post_process(sigmoid(probability), t, ms)
masks.append(predict)
d = []
for i, j in zip(masks, valid_masks[class_id::4]):
if (i.sum() == 0) & (j.sum() == 0):
d.append(1)
else:
d.append(dice(i, j))
attempts.append((t, ms, np.mean(d)))
attempts_df = pd.DataFrame(attempts, columns=['threshold', 'size', 'dice'])
attempts_df = attempts_df.sort_values('dice', ascending=False)
print(attempts_df.head())
best_threshold = attempts_df['threshold'].values[0]
best_size = attempts_df['size'].values[0]
class_params[class_id] = (best_threshold, best_size)
# predictions
torch.cuda.empty_cache()
gc.collect()
test_dataset = CloudDataset(df=sub, datatype='test', img_ids=test_ids, transforms=get_validation_augmentation(),
preprocessing=get_preprocessing(preprocessing_fn))
test_loader = DataLoader(test_dataset, batch_size=8, shuffle=False, num_workers=0)
loaders = {"test": test_loader}
encoded_pixels = []
image_id = 0
for i, test_batch in enumerate(tqdm(loaders['test'])):
runner_out = runner.predict_batch({"features": test_batch[0].cuda()})['logits']
for i, batch in enumerate(runner_out):
for probability in batch:
probability = probability.cpu().detach().numpy()
if probability.shape != (350, 525):
probability = cv2.resize(probability, dsize=(525, 350), interpolation=cv2.INTER_LINEAR)
predict, num_predict = post_process(sigmoid(probability), class_params[image_id % 4][0],
class_params[image_id % 4][1])
if num_predict == 0:
encoded_pixels.append('')
else:
r = mask2rle(predict)
encoded_pixels.append(r)
image_id += 1
sub['EncodedPixels'] = encoded_pixels
sub.to_csv('data/kaggle_cloud_org/submission.csv', columns=['Image_Label', 'EncodedPixels'], index=False)
