def make_unique_synth(set_t, images_set):
'''Generates synthetic data for 191 channels using the seen images from the training map.
set_t: Input the parameter set_t for indicating training or testing.
images_set: Can be either 'synthetic' or 'CIFAR'.
Saves an np.array of the following shape: (n_images, n_electrodes, width, height)
where n_images is based on the unique images in the test or training set (test: 889, train: 7998)
n_electrodes is 191, and img witdh x height = 240 x 240.
If you want to separate by layers (V1 & V4 respectively), then
v1 = inputs191_list[:, :97].sum(1)
v4 = weighted_gaus[:, 97:].sum(1)
inputs_2 = np.stack((v1,v4), axis=0)
'''
gaus = module.load_gausdata()
nn_seen = module.load_ydata(images_set)
data_indices = np.load(f'{set_t}/index_{set_t}_LFP_split.npy').astype(
np.int)
data_indices_unique = np.unique(data_indices)
if set_t == 'training':
slices = [slice(x * 1000, (x + 1) * 1000) for x in range(8)]
else:
slices = [slice(None, None)]
for batch_i, batch_slice in enumerate(slices):
inputs191_list = []
for data_index in tqdm(data_indices_unique[batch_slice]):
# data_index = data_indices[0]
synth_signal = nn_seen[data_index][:, :, :,
np.newaxis].repeat(191, axis=3)
# gaus_expand_to_batch = gaus.expand([191, 240, 240])
synth_signal_sum4 = synth_signal.transpose(3, 0, 1, 2).sum(
3) # merge color dimensions
dot_number = (gaus * synth_signal_sum4).sum((1, 2))
# weighted_gaus = gaus_expand_to_batch * dot_number[:, np.newaxis, np.newaxis].repeat((1, 240,240))
inputs191_list.append(dot_number)
inputs191_list = np.stack(inputs191_list)
np.save(f'{set_t}/{set_t}_{images_set[:5]}191batch{batch_i}',
inputs191_list)
lossFunction = module.VGGLoss(device)
optimizer = optim.Adam(model.parameters(), 0.1)
hori_means, verti_means, std_avg = RF.extract_means_std()
# -----
# Inputs:
# Will be dot number times the gaus
# ------
gaus = module.load_gausdata()
# ------
# Targets:
# are the masked nn_seen_torch, correspponding to the dot_number.
# ------
nn_seen_torch = torch.from_numpy(module.load_ydata())
# ------
# Training
# ------
dot_numbers_train = np.load(f'training/training_synth191final.npy')
training_iterator = module.make_iterator_unique(dot_numbers_train,
'training',
batch_size,
shuffle=True)
# ------
# Testing
# ------
dot_numbers_test = np.load(f'testing/testing_synth191final.npy')
testing_iterator = module.make_iterator_unique(dot_numbers_test,
lossFunction = nn.BCELoss()
vgg_lossFunction = module.VGGLoss(device)
if in_channels == 2:
inputtype = 'V1_V4'
if in_channels == 191:
inputtype = 'all_channels'
# -----
# RF gaus maps
# ------
gaus = module.load_gausdata(size= '96')
seen_images = module.load_ydata(None, size='96')
seen_images_torch = torch.from_numpy(seen_images)
# ------
# Training
# ------
dot_numbers_train = np.load(f'training/training_{images_set[:5]}191final.npy')
training_iterator = module.make_iterator_unique(dot_numbers_train, 'training', batch_size, shuffle = True)
# ------
# Testing
# ------
dot_numbers_test = np.load(f'testing/testing_{images_set[:5]}191final.npy')
testing_iterator = module.make_iterator_unique(dot_numbers_test, 'testing', batch_size, shuffle = False)
if __name__ == '__main__':
if device >= 0:
model.cuda(device)
# lossFunction = module.LossFunction(device)
lossFunction = module.VGGLoss(device)
optimizer = optim.Adam(model.parameters(), )
hori_means, verti_means, std_avg = RF.extract_means_std()
# -----
# RF gaus maps
# ------
gaus = module.load_gausdata()
seen_images = module.load_ydata()
# ------
# Training
# ------
nn_training = module.load_LFPdata('training')
training_iterator = module.make_iterator(nn_training, 'training', batch_size, shuffle = True)
# ------
# Testing
# ------
nn_testing = module.load_LFPdata('testing')
# testing_targets = module.load_ydata('testing')
testing_iterator = module.make_iterator(nn_testing, 'testing', batch_size, shuffle = False)
# -----