coinstac-dinunet 2.1.2

Distributed Neural Network implementation on COINSTAC.

coinstac-dinunet

Distributed Neural Network implementation on COINSTAC.

pip install coinstac-dinunet

Specify supported packages like pytorch & torchvision in a requirements.txt file

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.
6. We can integrate any custom reduction and learning mechanism by extending coinstac_dinunet.distrib.reducer/learner.
7. Realtime profiling each sites by specifying in compspec file(see dinune_fsv example below for details). 
...

---

Working examples:

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

Development guide - Add a new NN computation to COINSTAC:

imports

from coinstac_dinunet import COINNDataset, COINNTrainer, 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. Add entries to:

• Local node entry point CPU, GPU
• Aggregator node point CPU, GPU
• compspec.json file CPU, GPU

---

Define custom metrics:

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

Define custom DataHandle