def get_distance_to_performance(df, variable):
"""
plots the
:param df: pandas.DataFrame where the columns are performances of models on the different patients
:param variable: 'vel' if we are plotting velocity, else 'absVel' for absolute velocity
:return: dictionary where the performances are ordered based on the size of the receptive field of the networks
"""
distance_performance_dict = {}
for column in df.columns:
if 'k4_d2' not in column:
if ('Unnamed' not in column) and (variable in column):
print(column)
kernels, dilations = get_kernels_from_name(column)
max_k, _ = get_num_of_predictions(kernels, dilations)
distance = int((input_time_length - max_k + 1) / 2)
if ('k3_d3' in column) and ('sbp1' not in column):
distance = int(522 / 2)
performance = df[column].to_numpy()
print(column, distance, np.median(performance))
distance_performance_dict[distance] = [
np.median(performance),
scipy.stats.sem(performance)
]
ordered_distance_performance_dict = collections.OrderedDict(
sorted(distance_performance_dict.items()))
return ordered_distance_performance_dict
def get_gradients_for_intermediate_layers_from_np_arrays(
file, prefix, modules, train_mode, eval_mode, shift_by=None):
amp_grads, amp_grads_mch, amp_grads_nch, phase_grads, phase_grads_mch, phase_grads_nch = [], [], [], [], [], []
kernel_size, dilation = get_kernel_and_dilation_from_long_name(file)
X_reshaped_list = []
for module_name in modules:
if shift_by is not None:
shift_by_str = f'/shift_{shift_by}/'
else:
shift_by_str = ''
max_k, max_l = get_num_of_predictions(kernel_size,
dilation,
layer=module_name)
X_reshaped_list.append(
np.zeros([1, 1, input_time_length - max_k + 1, 1]))
print(module_name, prefix)
amp_grads.append(
np.load(
f'{home}/outputs/{gradient_save_dir}/{file}/{shift_by_str}{prefix}/amps/{module_name}/amps_avg_{file}_{train_mode}_{eval_mode}_ALLCH.npy'
))
amp_grads_mch.append(
np.load(
f'{home}/outputs/{gradient_save_dir}/{file}/{shift_by_str}{prefix}/amps/{module_name}/amps_avg_{file}_{train_mode}_{eval_mode}_MCH.npy'
))
amp_grads_nch.append(
np.load(
f'{home}/outputs/{gradient_save_dir}/{file}/{shift_by_str}{prefix}/amps/{module_name}/amps_avg_{file}_{train_mode}_{eval_mode}_NCH.npy'
))
# phase_grads.append(np.load(
# f'{home}/outputs/all_layer_gradients/{file}/{shift_by_str}{prefix}/phase/{module_name}/phase_avg_{file}_{train_mode}_{eval_mode}_ALLCH.npy'))
# phase_grads_mch.append(np.load(
# f'{home}/outputs/all_layer_gradients/{file}/{shift_by_str}{prefix}/phase/{module_name}/phase_avg_{file}_{train_mode}_{eval_mode}_MCH.npy'))
# phase_grads_nch.append(np.load(
# f'{home}/outputs/all_layer_gradients/{file}/{shift_by_str}{prefix}/phase/{module_name}/phase_avg_{file}_{train_mode}_{eval_mode}_NCH.npy'))
return X_reshaped_list, amp_grads, amp_grads_mch, amp_grads_nch, phase_grads, phase_grads_mch, phase_grads_nch
def plot_multiple_gradients_from_ndarrays(trained_mode, eval_mode, prefixes,
file, titles, grad_type):
plt.clf()
fig_s, ax_s = plt.subplots(2, 2, sharey='row', figsize=(20, 10))
output = f'{output_dir}/multigraphs/{file}/{grad_type}/{grad_type}_avg_{file}_{trained_mode}_{eval_mode}_ALLCH_sw.png'
kernels = None
indices = [(0, 0), (0, 1), (1, 0), (1, 1)]
kernels, dilations = get_kernel_and_dilation_from_long_name(file)
for i, prefix in enumerate(prefixes):
gradients = np.load(
f'{home}/outputs/all_layer_gradients/{file}/{prefix}/{grad_type}/conv_classifier/{grad_type}_avg_{file}_{train_mode}_{eval_mode}_ALLCH.npy'
)
gradients_mch = np.load(
f'{home}/outputs/all_layer_gradients/{file}/{prefix}/{grad_type}/conv_classifier/{grad_type}_avg_{file}_{train_mode}_{eval_mode}_MCH.npy'
)
gradients_nch = np.load(
f'{home}/outputs/all_layer_gradients/{file}/{prefix}/{grad_type}/conv_classifier/{grad_type}_avg_{file}_{train_mode}_{eval_mode}_NCH.npy'
)
gradients_sem = np.mean(scipy.stats.sem(np.abs(gradients), axis=1),
axis=0)
gradients_mch_sem = np.mean(scipy.stats.sem(np.abs(gradients_mch),
axis=1),
axis=0)
gradients_nch_sem = np.mean(scipy.stats.sem(np.abs(gradients_nch),
axis=1),
axis=0)
output_shape = get_num_of_predictions(kernels, dilations)
batch = np.zeros([1, 1, input_time_length - output_shape[0] + 1, 1])
create_multi_graph(
[gradients, gradients_mch, gradients_nch],
[gradients_sem, gradients_mch_sem, gradients_nch_sem], batch,
ax_s[indices[i]], titles[i])
plt.legend()
plt.xlabel('Frequency [Hz]')
plt.ylabel('Gradient')
plt.tight_layout()
print('saving figures:', output)
fig_s.savefig(output)
plt.show()
plt.close(fig_s)
def set_gradient_df_index(gradient_df, layer, file, shifts=False):
kernel_size, dilations = get_kernels_from_name(file)
max_k, max_l = get_num_of_predictions(kernel_size, dilations, layer=None)
print(file, max_k)
if 'k3_d3' in file:
max_k = 521
shape = 521 * 2
else:
shape = min((input_time_length - max_k) * 2, 1200)
# shape = 522
y = np.around(np.fft.rfftfreq(shape, 1 / 250.0), 0)
# index = np.linspace(0, 125, len(y))
if shifts:
new_columns = [
int(1000 * (int(column) / 250)) for column in gradient_df.columns
]
gradient_df.columns = new_columns
# y = [str(ypsilon).split('.')[0] for ypsilon in y]
gradient_df = gradient_df.set_index(pandas.Index(y), drop=True)
return gradient_df