def conv_eeg_signal_time(input_eeg_list,
vec_idx=np.arange(0, EEG_SIGNAL_SIZE),
kernelrow=roi_property.BIOSEMI_CONV,
concat_dim1=2,
is_rep=False):
'''
The function will use a bunch of image tensor and concat them with the multiplication of kernel size and from the
time dimension as concat_dim1=2. The arrangement should be every eeg width (time) contains 5 kernel indexes.
Args:
input_eeg_list: input image tensor list that obtained from split_image.py
vec_idx: the length of the vector used for mapping
kernelrow: the kernel size
concat_dim1: concat dimension 2, the time dimension
is_rep: only used in testing local roi on spatial domain
Returns: the concated kernel tensor and the tensor input shape for verification
'''
if not isinstance(input_eeg_list, list):
print("input eeg split signal is not a list of tensor.")
if len(input_eeg_list) > 0:
input_shape = input_eeg_list[0].get_shape().as_list()
else:
print("input eeg split not exist.")
return
# get the size of the vec1
vec1shape = np.shape(vec_idx)
vec1row = vec1shape[0] # 256/128/64/32/16 here
# input_shape = [10, 1, 256, 1]
input_shape[concat_dim1] *= kernelrow
input_shape[concat_dim1] *= vec1row
# input_shape = [10, 5*256, 256, 1]
if is_rep:
vec_idx = np.reshape(vec_idx, (vec1row, 1))
image_kernel_idx = np.repeat(vec_idx, kernelrow, axis=1)
else:
image_kernel_idx = image1tokrow.image_1tok_kernel(vec_idx, kernelrow)
curr_kernel_tensor = []
for kernel_idx in image_kernel_idx: # go for every kernel => 256 kernels
# for each kernel => 5 index
# concat on the first dimension => concat_dim1 = 1
curr_kernel_tensor.append(
tf.concat(concat_dim1,
[input_eeg_list[idx] for idx in kernel_idx]))
kernel_tensor = tf.concat(concat_dim1, curr_kernel_tensor)
return kernel_tensor, input_shape
def conv_eeg_signal_time(
input_eeg_list,
vec_idx=np.arange(0, EEG_SIGNAL_SIZE),
kernelrow=roi_property.BIOSEMI_CONV,
concat_dim1=2, is_rep=False):
'''
The function will use a bunch of image tensor and concat them with the multiplication of kernel size and from the
time dimension as concat_dim1=2. The arrangement should be every eeg width (time) contains 5 kernel indexes.
Args:
input_eeg_list: input image tensor list that obtained from split_image.py
vec_idx: the length of the vector used for mapping
kernelrow: the kernel size
concat_dim1: concat dimension 2, the time dimension
is_rep: only used in testing local roi on spatial domain
Returns: the concated kernel tensor and the tensor input shape for verification
'''
if not isinstance(input_eeg_list, list):
print("input eeg split signal is not a list of tensor.")
if len(input_eeg_list) > 0:
input_shape = input_eeg_list[0].get_shape().as_list()
else:
print("input eeg split not exist.")
return
# get the size of the vec1
vec1shape = np.shape(vec_idx)
vec1row = vec1shape[0] # 256/128/64/32/16 here
# input_shape = [10, 1, 256, 1]
input_shape[concat_dim1] *= kernelrow
input_shape[concat_dim1] *= vec1row
# input_shape = [10, 5*256, 256, 1]
if is_rep:
vec_idx = np.reshape(vec_idx, (vec1row, 1))
image_kernel_idx = np.repeat(vec_idx, kernelrow, axis=1)
else:
image_kernel_idx = image1tokrow.image_1tok_kernel(vec_idx, kernelrow)
curr_kernel_tensor = []
for kernel_idx in image_kernel_idx: # go for every kernel => 256 kernels
# for each kernel => 5 index
# concat on the first dimension => concat_dim1 = 1
curr_kernel_tensor.append(tf.concat(concat_dim1, [input_eeg_list[idx] for idx in kernel_idx]))
kernel_tensor = tf.concat(concat_dim1, curr_kernel_tensor)
return kernel_tensor, input_shape
def concat_digit_image(input_image_list,
vec_idx=np.arange(0, 28),
kernelrow=5,
concat_dim1=1,
concat_dim2=3):
'''
The function will use a bunch of image tensor and concat them with the kernel size as concat_dim1 and the image
height as concat_dim2
Args:
input_image_list: input image tensor list that obtained from split_image.py
vec_idx: the length of the vector used for mapping
kernelrow: the kernel size
concat_dim1: concat dimension 1
concat_dim2: concat dimension 2
Returns: the concated kernel tensor and the tensor input shape for verification
'''
if not isinstance(input_image_list, list):
print("input image is not a list.")
if len(input_image_list) > 0:
input_shape = input_image_list[0].get_shape().as_list()
else:
print("input image not exist.")
return
# get the size of the vec1
vec1shape = np.shape(vec_idx)
vec1row = vec1shape[0] # 28 here
# input_shape = [10, 1, 28, 1]
input_shape[concat_dim1] *= kernelrow
input_shape[concat_dim2] *= vec1row
# input_shape = [10, 5, 28, 28]
image_kernel_idx = image1tokrow.image_1tok_kernel(vec_idx, kernelrow)
curr_kernel_tensor = []
for kernel_idx in image_kernel_idx: # go for every kernel => 28 kernels
# for each kernel => 5 index
# concat on the first dimension => concat_dim1 = 1
curr_kernel_tensor.append(
tf.concat(concat_dim1,
[input_image_list[idx] for idx in kernel_idx]))
kernel_tensor = tf.concat(concat_dim2, curr_kernel_tensor)
return kernel_tensor, input_shape
def concat_digit_image(input_image_list, vec_idx=np.arange(0, 28), kernelrow=5, concat_dim1=1, concat_dim2=3):
'''
The function will use a bunch of image tensor and concat them with the kernel size as concat_dim1 and the image
height as concat_dim2
Args:
input_image_list: input image tensor list that obtained from split_image.py
vec_idx: the length of the vector used for mapping
kernelrow: the kernel size
concat_dim1: concat dimension 1
concat_dim2: concat dimension 2
Returns: the concated kernel tensor and the tensor input shape for verification
'''
if not isinstance(input_image_list, list):
print("input image is not a list.")
if len(input_image_list) > 0:
input_shape = input_image_list[0].get_shape().as_list()
else:
print("input image not exist.")
return
# get the size of the vec1
vec1shape = np.shape(vec_idx)
vec1row = vec1shape[0] # 28 here
# input_shape = [10, 1, 28, 1]
input_shape[concat_dim1] *= kernelrow
input_shape[concat_dim2] *= vec1row
# input_shape = [10, 5, 28, 28]
image_kernel_idx = image1tokrow.image_1tok_kernel(vec_idx, kernelrow)
curr_kernel_tensor = []
for kernel_idx in image_kernel_idx: # go for every kernel => 28 kernels
# for each kernel => 5 index
# concat on the first dimension => concat_dim1 = 1
curr_kernel_tensor.append(tf.concat(concat_dim1, [input_image_list[idx] for idx in kernel_idx]))
kernel_tensor = tf.concat(concat_dim2, curr_kernel_tensor)
return kernel_tensor, input_shape
