Added.

conversion/torch2oonx.py

@@ -0,0 +1,100 @@
+import torch
+import torchio as tio
+from torch.utils.data import DataLoader
+from torchio.data import GridSampler, GridAggregator
+from rhtorch.utilities.config import UserConfig
+from rhtorch.utilities.modules import (
+    recursive_find_python_class,
+    find_best_checkpoint
+)
+import numpy as np
+from pathlib import Path
+import argparse
+import nibabel as nib
+import sys
+from tqdm import tqdm
+
+
+def infer_data_from_model(model, subject, ps=None, po=None, bs=1, GPU=True):
+    """Infer a full volume given a trained model for 1 patient
+
+    Args:
+        model (torch.nn.Module): trained pytorch model
+        subject (torchio.Subject): Subject instance from TorchIO library
+        ps (list, optional): Patch size (from config). Defaults to None.
+        po (int or list, optional): Patch overlap. Defaults to None.
+        bs (int, optional): batch_size (from_config). Defaults to 1.
+
+    Returns:
+        [np.ndarray]: Full volume inferred from model
+    """
+    grid_sampler = GridSampler(subject, ps, po)
+    patch_loader = DataLoader(grid_sampler, batch_size=bs)
+    aggregator = GridAggregator(grid_sampler, overlap_mode='average')
+    with torch.no_grad():
+        for patches_batch in patch_loader:
+            patch_x, _ = model.prepare_batch(patches_batch)
+            if GPU:
+                patch_x = patch_x.to('cuda')
+            locations = patches_batch[tio.LOCATION]
+            patch_y = model(patch_x)
+            aggregator.add_batch(patch_y, locations)
+    return aggregator.get_output_tensor()
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(
+        description='Infer new data from input model.')
+    parser.add_argument("-c", "--config",
+                        help="Config file of saved model",
+                        type=str, default='config.yaml')
+    parser.add_argument("--checkpoint",
+                        help="Choose specific checkpoint that overwrites the config",
+                        type=str, default=None)
+    parser.add_argument("-o", "--onnx",
+                        help="Output onnx path",
+                        type=str, default='model.onnx')
+
+    args = parser.parse_args()
+
+    # load configs in inference mode
+    user_configs = UserConfig(args, mode='infer')
+    model_dir = user_configs.rootdir
+    configs = user_configs.hparams
+    project_dir = Path(configs['project_dir'])
+    model_name = configs['model_name']
+    data_shape_in = configs['data_shape_in']
+    patch_size = configs['patch_size']
+    channels_in = data_shape_in[0]
+
+    input_sample = torch.randn([1,channels_in,]+patch_size)
+
+    # load the model
+    module_name = recursive_find_python_class(configs['module'])
+    model = module_name(configs, data_shape_in)
+
+    if args.checkpoint is None:
+        # Load the final (best) model
+        if 'best_model' in configs:
+            ckpt_path = Path(configs['best_model'])
+            epoch_suffix = ''
+        # Not done training. Load the most recent (best) ckpt
+        else:
+            ckpt_path = find_best_checkpoint(project_dir.joinpath('trained_models', model_name, 'checkpoints'))
+            epoch_suffix = None
+    else:
+        ckpt_path = args.checkpoint
+    ckpt = torch.load(ckpt_path)
+    model.load_state_dict(ckpt['state_dict'])
+
+    # Export the model
+    torch.onnx.export(model,                   # model being run
+                    input_sample,              # model input (or a tuple for multiple inputs)
+                    args.onnx,                 # where to save the model (can be a file or file-like object)
+                    export_params=True,        # store the trained parameter weights inside the model file
+                    opset_version=10,          # the ONNX version to export the model to
+                    do_constant_folding=True,  # whether to execute constant folding for optimization
+                    input_names = ['input'],   # the model's input names
+                    output_names = ['output'], # the model's output names
+                    dynamic_axes={'input' : {0 : 'batch_size'},    # variable length axes
+                                  'output' : {0 : 'batch_size'}})