coinstac-dinunet 0.8.4

Distributed Neural Network implementation on COINSTAC.

coinstac-dinunet

Distributed Neural Network implementation on COINSTAC.

pip install coinstac-dinunet

Install supported pytorch & torchvision binaries in your device/docker ecosystem:

torch==1.5.1+cu92
torchvision==0.6.1+cu92

Highlights:

1. Handles multi-network/complex training schemes.
2. Automatic data splitting/k-fold cross validation.
3. Automatic model checkpointing.
4. GPU enabled local sites.
5. Customizable metrics(w/Auto serialization between nodes) to work with any schemes.
...

---

Pipeline for reducing gradients across sites.

Full working examples

1. FreeSurfer volumes classification.
2. VBM 3D images classification.

General use case:

imports

from coinstac_dinunet import COINNDataset, COINNTrainer, COINNRemote, COINNLocal
from coinstac_dinunet.metrics import COINNAverages, Prf1a

1. Define Data Loader

class MyDataset(COINNDataset):
    def __init__(self, **kw):
        super().__init__(**kw)
        self.labels = None

    def load_index(self, id, file):
        data_dir = self.path(id, 'data_dir') # data_dir comes from inputspecs.json
        ...
        self.indices.append([id, file])

    def __getitem__(self, ix):
        id, file = self.indices[ix]
        data_dir = self.path(id, 'data_dir') # data_dir comes from inputspecs.json
        label_dir = self.path(id, 'label_dir') # label_dir comes from inputspecs.json
        ...
        # Logic to load, transform single data item.
        ...
        return {'inputs':.., 'labels': ...}

2. Define Trainer

class MyTrainer(COINNTrainer):
    def __init__(self, **kw):
        super().__init__(**kw)

    def _init_nn_model(self):
        self.nn['model'] = MYModel(in_size=self.cache['input_size'], out_size=self.cache['num_class'])

    def iteration(self, batch):
        inputs, labels = batch['inputs'].to(self.device['gpu']).float(), batch['labels'].to(self.device['gpu']).long()

        out = F.log_softmax(self.nn['model'](inputs), 1)
        loss = F.nll_loss(out, labels)
        _, predicted = torch.max(out, 1)
        score = self.new_metrics()
        score.add(predicted, labels)
        val = self.new_averages()
        val.add(loss.item(), len(inputs))
        return {'out': out, 'loss': loss, 'averages': val,
                'metrics': score, 'prediction': predicted}

3. Supply to local node in local.py

if __name__ == "__main__":
    args = json.loads(sys.stdin.read())
    local = COINNLocal(cache=args['cache'], input=args['input'], state=args['state'])
    local.compute(MyDataset, MyTrainer)
    local.send()

4. Define remote node in remote.py

class MyRemote(COINNRemote):

    def _new_metrics(self):  #
        return coinstac_dinunet.metrics.Prf1a()

    def _new_averages(self):
        return coinstac_dinunet.metrics.COINNAverages()

    def _monitor_metric(self):
        return 'f1', 'maximize'


if __name__ == "__main__":
    args = json.loads(sys.stdin.read())
    remote = MyRemote(cache=args['cache'], input=args['input'], state=args['state'])
    remote.compute()
    remote.send()

Define custom metrics

• Extend coinstac_dinunet.metrics.COINNMetrics
• Example: coinstac_dinunet.metrics.Prf1a for Precision, Recall, F1, and Accuracy

Default arguments:

• task_name: str = None, Name of the task. [Required]
• mode: str = None, Eg. train/test [Required]
• batch_size: int = 4
• epochs: int = 21
• learning_rate: float = 0.001
• gpus: _List[int] = None, Eg. [0], [1], [0, 1]...
• pin_memory: bool = True, if cuda available
• num_workers: int = 0
• load_limit: int = float('inf'), Limit on dataset to load for debugging purpose.
• pretrained_path: str = None, Path to pretrained weights
• patience: int = 5, patience to end training by monitoring validation scores.
• load_sparse: bool = False, Load each data item in separate loader to reconstruct images from patches, if needed.
• num_folds: int = None, Number of k-folds.
• split_ratio: _List[float] = (0.6, 0.2, 0.2), Exclusive to num_folds.

Directly passed parameters in coinstac_dinunet.nodes.COINNLocal, args passed through inputspec will override the defaults in the same order.

Custom data splits can be provided in the path specified by split_dir for each sites in their respective inputspecs file. This is mutually exclusive to both num_folds and split_ratio.

---