def plot_filters_top_alphas(self,
n_filters,
save_path,
save_as_mat=False,
save_as_plot=True):
if not os.path.exists(save_path):
os.makedirs(save_path)
unwhitening_mat, a_mat, w_mat = self.load_filters_for_plotting()
top_alphas = np.argsort(a_mat)[::-1]
filter_ids = top_alphas[:n_filters]
rotated_w_mat = np.dot(w_mat[:, filter_ids].T, unwhitening_mat)
for idx, filter_id in enumerate(filter_ids):
flat_filter = rotated_w_mat[idx, :]
alpha = a_mat[filter_id]
plottable_filters = np.reshape(flat_filter,
[self.n_channels, self.ph, self.pw])
if save_as_mat:
file_name = 'filter_{}_alpha_{:.3e}.npy'.format(
filter_id, float(alpha))
np.save(save_path + file_name, plottable_filters)
if save_as_plot:
file_name = 'filter_{}_alpha_{:.3e}.png'.format(
filter_id, float(alpha))
plottable_filters -= np.min(plottable_filters)
plottable_filters /= np.max(plottable_filters)
plot_img_mats(plottable_filters,
rescale=False,
show=False,
save_path=save_path + file_name)
print('filter {} done'.format(filter_id))
def plot_filters_after_training(self, w_res, unwhiten_mat, n_vis):
comps = np.dot(w_res.T, unwhiten_mat)
comps -= np.min(comps)
comps /= np.max(comps)
co = np.reshape(comps[:n_vis, :], [n_vis, 3, self.ph, self.pw])
co = np.transpose(co, axes=[0, 2, 3, 1])
plot_img_mats(co, color=True, rescale=True)
def plot_filters_after_training(self, w_res, unwhiten_mat, n_vis):
comps = np.dot(w_res[0, :, :].T, unwhiten_mat[0, :, :])
print(comps.shape)
comps -= np.min(comps)
comps /= np.max(comps)
co = np.reshape(comps[:n_vis, :], [-1, self.ph, self.pw])
plot_img_mats(co, color=False, rescale=True)
def plot_channels_top_filters(self,
channel_ids,
save_path,
save_as_mat=False,
save_as_plot=True,
n_vis=144):
"""
visualizes filters of selected channels sorted descending by norm of the filter
saves these as one plot per channel.
:param channel_ids: collection of filter indices
:param save_path: location to save plots
:param save_as_mat: if true, saves each filter as channel x height x width matrix
:param save_as_plot: if true, saves each filter as image
:param n_vis: ...
:return: None
"""
if not os.path.exists(save_path):
os.makedirs(save_path)
unwhitening_mat, a_mat, w_mat = self.load_filters_for_plotting()
rotated_w_mat = np.dot(w_mat.T, unwhitening_mat)
for channel_id in channel_ids:
chan_start = self.ph * self.pw * channel_id
chan_end = self.ph * self.pw * (channel_id + 1)
flat_channel = rotated_w_mat[:, chan_start:chan_end]
w_min = np.min(flat_channel)
w_max = np.max(flat_channel)
w_norms = np.linalg.norm(flat_channel, axis=1)
norm_ids = np.argsort(w_norms)[-n_vis:][::-1]
flat_channel = flat_channel[norm_ids, :]
print(np.linalg.norm(flat_channel, axis=1))
plottable_channels = np.reshape(
flat_channel, [flat_channel.shape[0], self.ph, self.pw])
if save_as_mat:
file_name = 'channel_{}.npy'.format(channel_id)
np.save(save_path + file_name, plottable_channels)
if save_as_plot:
file_name = 'channel_{}.png'.format(channel_id)
plottable_channels -= w_min
plottable_channels /= (w_max - w_min)
plot_img_mats(plottable_channels,
rescale=False,
show=False,
save_path=save_path + file_name)
print('filter {} done'.format(channel_id))
def plot_filters_after_training(self, w_res, unwhiten_mat, n_vis):
w_res_t = w_res.T.reshape(
(self.n_components, self.n_channels, 1, self.ph * self.pw))
if self.channelwise_data:
comps = np.matmul(w_res_t, unwhiten_mat).squeeze()
comps -= np.min(comps)
comps /= np.max(comps)
co = np.reshape(comps[:n_vis, :, :],
[n_vis, self.n_channels, self.ph, self.pw])
else:
comps = np.dot(w_res.T, unwhiten_mat)
comps -= np.min(comps)
comps /= np.max(comps)
co = np.reshape(comps[:n_vis, :], [n_vis, 3, self.ph, self.pw])
co = np.transpose(co, axes=[0, 2, 3, 1])
plot_img_mats(co, color=True, rescale=True)
def plot_filters_all_channels(self,
filter_ids,
save_path,
save_as_mat=False,
save_as_plot=True):
"""
visualizes the patch for each channel of a trained filter and saves this as one plot.
does so for the filter of each given index
:param filter_ids: collection of filter indices
:param save_path: location to save plots
:param save_as_mat: if true, saves each filter as channel x height x width matrix
:param save_as_plot: if true, saves each filter as image
:return: None
"""
if not os.path.exists(save_path):
os.makedirs(save_path)
unwhitening_mat, a_mat, w_mat = self.load_filters_for_plotting()
rotated_w_mat = np.dot(w_mat[:, filter_ids].T, unwhitening_mat)
for idx, filter_id in enumerate(filter_ids):
flat_filter = rotated_w_mat[idx, :]
alpha = a_mat[filter_id]
plottable_filters = np.reshape(flat_filter,
[self.n_channels, self.ph, self.pw])
if save_as_mat:
file_name = 'filter_{}_alpha_{:.3e}.npy'.format(
filter_id, float(alpha))
np.save(save_path + file_name, plottable_filters)
if save_as_plot:
file_name = 'filter_{}_alpha_{:.3e}.png'.format(
filter_id, float(alpha))
plottable_filters -= np.min(plottable_filters)
plottable_filters /= np.max(plottable_filters)
plot_img_mats(plottable_filters,
rescale=False,
show=False,
save_path=save_path + file_name)
print('filter {} done'.format(filter_id))
def plot_channels_all_filters(self,
channel_ids,
save_path,
save_as_mat=False,
save_as_plot=True,
n_vis=144):
"""
visualizes all filters of selected channels and saves these as one plot per channel.
:param channel_ids: ...
:param save_path: location to save plots
:param save_as_mat: if true, saves each filter as channel x height x width matrix
:param save_as_plot: if true, saves each filter as image
:param n_vis: ...
:return: None
"""
if not os.path.exists(save_path):
os.makedirs(save_path)
unwhitening_mat, a_mat, w_mat = self.load_filters_for_plotting()
rotated_w_mat = np.dot(w_mat[:, :n_vis].T, unwhitening_mat)
for channel_id in channel_ids:
chan_start = self.ph * self.pw * channel_id
chan_end = self.ph * self.pw * (channel_id + 1)
flat_channel = rotated_w_mat[:, chan_start:chan_end]
plottable_channels = np.reshape(
flat_channel, [flat_channel.shape[0], self.ph, self.pw])
if save_as_mat:
file_name = 'channel_{}.npy'.format(channel_id)
np.save(save_path + file_name, plottable_channels)
if save_as_plot:
file_name = 'channel_{}.png'.format(channel_id)
plottable_channels -= np.min(plottable_channels)
plottable_channels /= np.max(plottable_channels)
plot_img_mats(plottable_channels,
rescale=False,
show=False,
save_path=save_path + file_name)
print('filter {} done'.format(channel_id))
def plot_filters_all_channels(self,
channel_ids,
save_path,
save_as_mat=False,
save_as_plot=True):
"""
visualizes the patch for each channel of a trained filter and saves this as one plot.
does so for the filter of each given index
:param channel_ids: collection of channel indices
:param save_path: location to save plots
:param save_as_mat: if true, saves each filter as channel x height x width matrix
:param save_as_plot: if true, saves each filter as image
:return None
"""
if not os.path.exists(save_path):
os.makedirs(save_path)
unwhitening_mat, a_mat, w_mat = self.load_filters_for_plotting()
w_mat_select = w_mat[channel_ids, :, :].transpose((0, 2, 1))
rotated_w_mat = w_mat_select @ unwhitening_mat[channel_ids, :, :]
for idx, channel_id in enumerate(channel_ids):
chan_filters = rotated_w_mat[idx, :, :]
plottable_filters = np.reshape(
chan_filters, [self.n_components, self.ph, self.pw])
if save_as_mat:
file_name = 'channel_{}_filters.npy'.format(channel_id)
np.save(save_path + file_name, plottable_filters)
if save_as_plot:
file_name = 'channel_{}_filters.png'.format(channel_id)
plottable_filters -= np.min(plottable_filters)
plottable_filters /= np.max(plottable_filters)
plot_img_mats(plottable_filters,
rescale=False,
show=False,
save_path=save_path + file_name)
print('channel {} done'.format(channel_id))