Convert GCN project to pytorch

[zhangyu2ustc]

deep learning on graph

pytorch has very nice tools for gcns, including:
DGL: https://github.com/dmlc/dgl
PyTorch Geometric : https://github.com/rusty1s/pytorch_geometric

I want to transfer my GCN models using these new tools

[zhangyu2ustc]

First, I need to rewrite a dataset class which loads fmri data and event data:
check the notebook for examples: pytorch_hcp_task_fmri.ipynb

'''
#######
params = {'batch_size': 10,
'shuffle': True,
'num_workers': 2}

ind = np.loadtxt("motor_perm_indices.txt", comments="#", delimiter=",", unpack=False).astype(int)

fmri_matrix = np.array(subjects_tc_matrix)
label_matrix = np.array(subjects_trial_label_matrix)
fmri_dataset = HCP_taskfmri_matrix_datasets(pathout,fmri_matrix,label_matrix,target_name)
#fmri_dataset = HCP_taskfmri_matrix_datasets(pathout,fmri_files,label_matrix,target_name)

fmri_loader = DataLoader(fmri_dataset, collate_fn=fmri_samples_collate_fn, **params)
data,label = next(iter(fmri_loader))
print(data.size(),label.size())
'''

[zhangyu2ustc]

working with batch-size=1, but the training process is too slow.
Need to fix the mini-batch of gcn

Update: fixed the problem.
there are two ways to doing the batches. the silly way is using a for loop, which is straightforward but very slow.
the other way is the do permutate and view in pytorch

[zhangyu2ustc]

https://github.com/meliketoy/graph-cnn.pytorch

[zhangyu2ustc]

add LSTM as one of the baseline methods:

In pytorch, the performance of LSTM (5 stacked layers followed by 3 linear layers) are quite good.
for motor classification (epochs=100):
train_acc: 0.99
test_acc: 0.95

[zhangyu2ustc]

After several experiments on GCN and GAT in pytorch. I find it quite strange that the performance is slightly lower in pytorch implementation.

for motor classification (test_acc):
GCN ChebNet GAT
tensorflow: 0.98 0.97
pytorch: 0.88 0.92

we need to figure out what was wrong in the pytorch version

[zhangyu2ustc]

transfer learning with different TRs: there are two ways to do it

1. sampling the TR in HCP, which means take the mean of 3 volumes each step
2. interpolation fmri time series in IBC