def plot_and_save_expanded_data_for_samples(gate_expander, model, data_input,
sample_idxs_to_plot):
for sample_idx in sample_idxs_to_plot:
plotter = DataAndGatesPlotterDepthOne(model,
np.concatenate(data_input.x_tr))
plotter.plot_single_sample_with_gate(data_input.x_tr[sample_idx],
data_input.idxs_tr[sample_idx],
data_input.y_tr[sample_idx],
plt.gca(),
include_diagnostics=False)
expanded_data = gate_expander.expanded_data_per_sample[sample_idx]
if not (expanded_data.shape[0] == 0):
print(expanded_data.shape)
plotter.plot_single_sample_with_gate(
expanded_data,
data_input.idxs_tr[sample_idx],
data_input.y_tr[sample_idx],
plt.gca(),
color='r',
include_diagnostics=False)
clusters = gate_expander.clusterers_per_sample[
sample_idx].cluster_centers_
plt.gca().scatter(clusters[:, 0], clusters[:, 1], color='b', s=2)
plt.savefig('expanded_sample%d.png' % data_input.idxs_tr[sample_idx])
plt.clf()
def make_umap_plots_per_sample(model, data_input, sample_idxs_to_plot, plots_per_row=5, figlen=7, savename='plots_per_sample.png', background_data_to_plot=None, color='b', expanded_data_per_sample=None, sample_names_to_true_features=None, BALL=False):
if len(sample_idxs_to_plot) == 0:
print('idxs are empty!')
return None
vals_to_delete = []
for i, idx in enumerate(sample_idxs_to_plot):
if not (idx in data_input.sample_names_all):
print('Sample %d not in training data' %idx)
vals_to_delete.append(idx)
for val in vals_to_delete:
del sample_idxs_to_plot[sample_idxs_to_plot.index(val)]
plotter = DataAndGatesPlotterDepthOne(model, [])
idxs_in_data_input = [[1, data_input.idxs_tr.index(idx)] if idx in data_input.idxs_tr else [0, data_input.idxs_te.index(idx)] for idx in sample_idxs_to_plot]
n_samples_to_plot = len(sample_idxs_to_plot)
evenly_divides = not(n_samples_to_plot % plots_per_row)
n_rows = n_samples_to_plot//plots_per_row
if not(evenly_divides):
n_rows += 1
print(n_rows, plots_per_row)
fig, axes = plt.subplots(n_rows, plots_per_row, figsize=((figlen) * plots_per_row, (figlen) * n_rows),sharex=True, sharey=True)
#fig.suptitle('UMAP Embedding and Learned Gates per Sample')
axes = [axes] if len(axes.shape) == 1 else axes
if not (background_data_to_plot is None):
for i in range(n_rows):
for j in range(plots_per_row):
axes[i][j].scatter(
background_data_to_plot[:, 0],
background_data_to_plot[:, 1],
c='lightgrey', s=1/100, alpha=.5
)
axes[0][0].set_xlim(0, 1)
axes[0][0].set_ylim(0, 1)
fig.tight_layout(pad=1.3)
row_start_idx = 0
for i, row in enumerate(range(n_rows)):
sample_row_idxs = sample_idxs_to_plot[row_start_idx: row_start_idx + plots_per_row]
for j, sample_idx in enumerate(sample_row_idxs):
cur_axis = axes[i][j]
data_input_matching_idx = idxs_in_data_input[row_start_idx + j][1]
if idxs_in_data_input[j][0]:
sample = data_input.x_tr[data_input_matching_idx]
label = data_input.y_tr[data_input_matching_idx]
else:
sample = data_input.x_te[data_input_matching_idx]
label = data_input.y_te[data_input_matching_idx]
name = sample_idxs_to_plot[row_start_idx + j]
true_feature = None
if sample_names_to_true_features:
true_feature = sample_names_to_true_features[name]
if not BALL:
plotter.plot_single_sample_with_gate(
sample, name, label,
cur_axis, size=1, color='b',
true_feature=true_feature
)
else:
plotter.plot_single_sample_with_gate(
sample, name, label,
cur_axis, size=1, color='b',
true_feature=true_feature,
BALL=True
)
if not (expanded_data_per_sample is None):
expanded_data = expanded_data_per_sample[data_input_matching_idx]
cur_axis.scatter(expanded_data[:, 0], expanded_data[:, 1], s=1, color='r')
row_start_idx += plots_per_row
plt.savefig(savename)