def feature_net_cross_eval(datasets, outputPath, epochs, debug=False):
eval_scores = np.empty((8, 8))
for test_run_idx in range(8):
# create indices for train and test split
train_runs_indices = [j for j in range(8)]
train_runs_indices.pop(test_run_idx)
train_dataset_list = [datasets[j] for j in train_runs_indices]
# create combined training set
combined_train_dataset = ConcatDataset(train_dataset_list)
train_sampler = RandomSampler(combined_train_dataset)
train_loader = DataLoader(combined_train_dataset,
sampler=train_sampler,
batch_size=64,
num_workers=0,
pin_memory=True)
# create model, optimizer, error function
num_features = combined_train_dataset.__getitem__(0)[0].shape[0]
num_ffa_voxels = combined_train_dataset.__getitem__(0)[1].shape[0]
model = FeatureNet(num_features, num_ffa_voxels).cuda()
model.train()
criterion = torch.nn.MSELoss()
optimizer = torch.optim.SGD(model.parameters(),
lr=0.001,
momentum=0.9,
weight_decay=0.001)
# fit model
fit_feature_model(model,
criterion,
optimizer,
train_loader,
epochs=epochs,
debug=debug)
# Evaluate model
model.eval()
# Each column stores the individual run correlations for a given test set
for run_idx, run in enumerate(datasets):
eval_scores[run_idx,
test_run_idx] = featuresCorr(datasets[run_idx], model)
# # in sample error
# eval_scores[test_run_idx, 0] = featuresCorr(combined_train_dataset, model)
# # out of sample error
# test_dataset = datasets[test_run_idx]
# eval_scores[test_run_idx, 1] = featuresCorr(test_dataset, model)
# Save data
np.save(outputPath, eval_scores)
return eval_scores
class MultiwayDataset:
def __init__(self, pairs, tokenizer, dictionary):
self.datasets = [
ParallelDataset(fst, snd, tokenizer, dictionary)
for fst, snd in pairs
]
self.concatenated = ConcatDataset(self.datasets)
self.lengths = list(
itertools.chain(*[dataset.lengths for dataset in self.datasets]))
self.indices = np.argsort(self.lengths)
def __getitem__(self, idx):
# idy = self.indices[idx]
# print(idy, idx, self.lengths[idx])
return self.concatenated.__getitem__(idx)
def __len__(self):
return self.concatenated.__len__()
def __getitem__(self, index):
return ConcatDataset.__getitem__(self, index)
def __getitem__(self, index: int) -> tuple:
return ConcatDataset.__getitem__(self, index)
