def e2n_generator(net_name,
h,
w,
n_injuries,
max_iter=10000,
test_interval=50,
snapshot=1000):
# Specify the architecture using a list of dictionaries.
e2n_arch = [
[
'e2n', # e2n layer
{
'n_filters': 130, # 130 feature maps
'kernel_h': h,
'kernel_w':
w # Cross filter of size h x 1 by 1 x w (non-sliding, only on diagonal)
}
],
['dropout', {
'dropout_ratio': 0.5
}], # Dropout with 0.5 dropout rate.
['relu', {
'negative_slope': 0.33
}], # Very leaky ReLU.
['fc', {
'n_filters': 30
}], # Fully connected/dense (Node-to-Graph when after e2n) layer
['relu', {
'negative_slope': 0.33
}], # Very leaky ReLU
['out', {
'n_filters': n_injuries
}] # Output with 2 nodes that correspond to 2 injuries.
]
# Create BrainNetCNN model
E2Nnet_sml = BrainNetCNN(
net_name, # Unique model name.
e2n_arch, # List of dictionaries specifying the architecture.
hardware='cpu', # Or 'cpu'.
dir_data='./generated_synthetic_data', # Where to write the data to.
)
#set pars
E2Nnet_sml.pars[
'max_iter'] = max_iter # Train the model for 1000 iterations. (note this should be run for much longer!)
E2Nnet_sml.pars[
'test_interval'] = test_interval # Check the valid data every 50 iterations.
E2Nnet_sml.pars[
'snapshot'] = snapshot # Save the model weights every 1000 iterations.
return E2Nnet_sml
['fc', {'n_filters': 30}], # Fully connected (n2g) layer with 30 filters.
['relu', {'negative_slope': 0.33}],
['out', {'n_filters': 1}]] # Output layer with num_outs nodes as outputs.
e2e_arch = [
['e2e', # e2e layer
{'n_filters': 32, # 32 feature maps
'kernel_h': x_train.shape[2], 'kernel_w': x_train.shape[3] # Sliding cross filter of size h x 1 by 1 x w
}
],
['e2e', # e2e layer
{'n_filters': 32, # 32 feature maps
'kernel_h': x_train.shape[2], 'kernel_w': x_train.shape[3] # Sliding cross filter of size h x 1 by 1 x w
}
],
['e2n', {'n_filters': 64, 'kernel_h': x_train.shape[2], 'kernel_w': x_train.shape[3]}],
['dropout', {'dropout_ratio': 0.5}],
['relu', {'negative_slope': 0.33}],
['fc', {'n_filters': 30}],
['relu', {'negative_slope': 0.33}],
['out', {'n_filters': 1}]
]
hello_net = BrainNetCNN('e2e2', e2e_arch) # Create BrainNetCNN model
hello_net.fit(x_train, y_train[:, 0], x_valid, y_valid[:, 0]) # Train (regress only on class 0)
preds = hello_net.predict(x_test) # Predict labels of test data
print("Correlation:", pearsonr(preds, y_test[:, 0])[0])
hello_net.plot_iter_metrics()
print("Correlation:", pearsonr(preds, y_test[:, 0])[0])
}], # Very leaky ReLU.
['fc', {
'n_filters': 30
}], # Fully connected/dense (Node-to-Graph when after e2n) layer
['relu', {
'negative_slope': 0.33
}], # Very leaky ReLU
['out', {
'n_filters': n_injuries
}] # Output with 2 nodes that correspond to 2 injuries.
]
# Create BrainNetCNN model
E2Nnet_sml = BrainNetCNN(
net_name, # Unique model name.
e2n_arch, # List of dictionaries specifying the architecture.
hardware='cpu', # Or 'cpu'.
dir_data='./generated_synthetic_data', # Where to write the data to.
)
#set pars
E2Nnet_sml.pars[
'max_iter'] = 100 # Train the model for 1000 iterations. (note this should be run for much longer!)
E2Nnet_sml.pars[
'test_interval'] = 50 # Check the valid data every 50 iterations.
E2Nnet_sml.pars[
'snapshot'] = 20 # Save the model weights every 1000 iterations.
# In[32]:
# %%
# Train (optimize) the network.
# WARNING: If you have a high max_iter and no GPU, this could take awhile...
def e2e(net_name,
h,
w,
n_injuries,
max_iter=10000,
test_interval=100,
snapshot=1000):
# Specify the architecture.
e2e_arch = [
[
'e2e', # e2e layer
{
'n_filters': 32, # 32 feature maps
'kernel_h': h,
'kernel_w': w # Sliding cross filter of size h x 1 by 1 x w
}
],
['e2n', {
'n_filters': 64,
'kernel_h': h,
'kernel_w': w
}],
['dropout', {
'dropout_ratio': 0.5
}],
['relu', {
'negative_slope': 0.33
}],
['fc', {
'n_filters': 30
}],
['relu', {
'negative_slope': 0.33
}],
['out', {
'n_filters': n_injuries
}]
]
# Create BrainNetCNN model
E2Enet_sml = BrainNetCNN(
net_name,
e2e_arch,
hardware='cpu', # Or 'cpu'.
dir_data='./generated_synthetic_data', # Where to write the data to.
)
# Overwrite default parameters.
# ann4brains.nets.get_default_hyper_params() shows the hyper-parameters that can be overwritten.
#E2Enet_sml.pars['max_iter'] = 100000 # Train the model for 100K iterations.
#E2Enet_sml.pars['test_interval'] = 500 # Check the valid data every 500 iterations.
#E2Enet_sml.pars['snapshot'] = 10000 # Save the model weights every 10000 iterations.
# NOTE using the above parameters takes awhile for the model to train (~2 hours ish? on a GPU)
# If you want to do some simple fast experiments to start, use these settings instead.
E2Enet_sml.pars[
'max_iter'] = max_iter # Train the model for 100 iterations (note this should be run for much longer!)
E2Enet_sml.pars[
'test_interval'] = test_interval # Check the valid data every 20 iterations.
E2Enet_sml.pars[
'snapshot'] = snapshot # Save the model weights every 100 iterations.
return E2Enet_sml
[
'e2n',
{
'n_filters': 16, # e2n layer with 16 filters.
'kernel_h': x_train.shape[2],
'kernel_w': x_train.shape[3]
}
], # Same dimensions as spatial inputs.
['dropout', {
'dropout_ratio': 0.5
}], # Dropout at 0.5
['relu', {
'negative_slope': 0.33
}], # For leaky-ReLU
['fc', {
'n_filters': 30
}], # Fully connected (n2g) layer with 30 filters.
['relu', {
'negative_slope': 0.33
}],
['out', {
'n_filters': 1
}]
] # Output layer with num_outs nodes as outputs.
hello_net = BrainNetCNN('hello_world', hello_arch) # Create BrainNetCNN model
hello_net.fit(x_train, y_train[:, 0], x_valid,
y_valid[:, 0]) # Train (regress only on class 0)
preds = hello_net.predict(x_test) # Predict labels of test data
print("Correlation:", pearsonr(preds, y_test[:, 0])[0])
'negative_slope': 0.33
}],
['out', {
'n_filters': 1
}]
]
exp_arch = [['fc', {
'n_filters': 2
}], ['relu', {
'negative_slope': 0.33
}], ['out', {
'n_filters': 1
}]]
siam_A = BrainNetCNN('ann_siam', exp_arch) # Create BrainNetCNN model
siam_A.pars[
'max_iter'] = epochs * num_batch # running for correct number of iterations
siam_A.pars['snapshot'] = epochs * num_batch
siam_A.pars['train_batch_size'] = batch_size
siam_A.pars['test_batch_size'] = batch_size
#siam_A.fit_siamese(x_train, y_train, x_valid, y_valid)
ann_train = siam_A.fit_siamese(x_train, y_train, x_valid, y_valid)
print("Finished executing ANN")
## SIAMESE GCN
## Starting with implementing siamese structure for GCN