def plot_more_than_three_canals(input_, prediction, groundtruth, labels):
plt.figure(figsize=(40, 16))
line_number = input_.shape[0]
col_number = groundtruth.shape[-1] + 3
for i, (input_thumb, pred_thumb,
truth_thumb) in enumerate(zip(input_, prediction, groundtruth)):
plt.subplot(line_number, col_number, i * col_number + 1)
imshow(input_thumb, cmap='gray')
if not i and labels is not None:
plt.title(labels[0])
plt.subplot(line_number, col_number, i * col_number + 2)
imshow(pred_thumb)
if not i and labels is not None:
plt.title(labels[1])
plt.subplot(line_number, col_number, i * col_number + 3)
imshow(truth_thumb)
if not i and labels is not None:
plt.title(labels[2])
for canal in range(pred_thumb.shape[2]):
plt.subplot(line_number, col_number, i * col_number + canal + 4)
imshow(pred_thumb[:, :, canal], cmap=COLORMAPS[canal])
plt.colorbar()
if not i and labels is not None and len(labels) > 3 + canal:
plt.title(f"Prediction: {labels[3 + canal]}")
def plot_regressor_distribution(examples, truths, predictions, attributes,
means, stds):
plt.figure(figsize=(6 * (len(attributes) + 1), 4 * 4))
for i, example in zip(range(3), examples):
if i < len(example):
plt.subplot(3, len(attributes) + 1, 1 + i * (len(attributes) + 1))
imshow(example, vmax=1)
prediction = np.round(predictions[i] * stds + means, 2)
truth = np.round(truths[i] * stds + means, 2)
plt.title(f'Prediction: {prediction}\nTruth: {truth}',
fontdict={'horizontalalignment': 'right'})
for i, attribute in enumerate(attributes):
mean = means[i]
std = stds[i]
plt.subplot(3, len(attributes) + 1, i + 2)
plt.hist(predictions[:, i] * std + mean,
bins=20,
range=(-std + mean, std + mean),
density=True)
plt.title(attribute)
if not i:
plt.ylabel('Predictions')
plt.subplot(3, len(attributes) + 1, i + 2 + len(attributes) + 1)
plt.hist(truths[:, i] * std + mean,
bins=20,
range=(-std + mean, std + mean),
density=True)
if not i:
plt.ylabel('Truth')
plt.subplot(3, len(attributes) + 1, i + 2 + 2 * (len(attributes) + 1))
plt.hist(abs(truths[:, i] - predictions[:, i]),
bins=20,
range=(0, 2),
density=True)
if not i:
plt.ylabel('Normalized Absolute Error')
def plot_example(example, prediction, labels, limit=LIMIT):
input_, truth = example
input_ = input_[:limit]
n = input_.shape[0]
plt.figure(figsize=(27, 12))
plt.tight_layout()
for i, (im, classes) in enumerate(zip(input_, prediction)):
if i != 0:
tick_label = [' ' for _ in labels]
else:
tick_label = labels
plt.subplot(2, n, i + 1)
imshow(im, vmax=1)
plt.subplot(2, n, i + 1 + n)
plt.barh(labels, truth[i], tick_label=tick_label, color='C1')
plt.barh(labels, classes, tick_label=tick_label, color='C0')
plt.xlim(0, 1)
plt.yticks()
plt.tight_layout()
def plot_less_than_three_canals(input_, prediction, groundtruth, labels):
plt.figure(figsize=(max(256, input_.shape[2]) / 100 * 9,
max(256, input_.shape[1]) / 100 * 3.5))
if groundtruth is None:
line_number = 2
else:
line_number = 3
col_number = input_.shape[0]
for i, im in enumerate(input_):
plt.subplot(line_number, col_number, 1 + i)
imshow(input_[i], vmin=0, vmax=255)
if not i and labels is not None:
plt.title(labels[0])
plt.subplot(line_number, col_number, i + 1 + col_number)
imshow(prediction[i], vmin=0, vmax=255, cmap='hot')
if not i and labels is not None:
plt.title(labels[1])
if line_number == 3:
plt.subplot(line_number, col_number, i + 1 + col_number * 2)
imshow(groundtruth[i], vmin=0, vmax=255)
if not i and labels is not None:
plt.title(labels[2])
def plot_regressor_on_color(examples, truths, predictions, means, stds):
plt.figure(figsize=(4 * 4, 3 * 4))
for i, (example, truth,
prediction) in enumerate(zip(examples[:12], truths, predictions)):
plt.subplot(3, 4, i + 1)
imshow(example, vmax=1)
prediction = prediction * stds + means
truth = truth * stds + means
if means[0] > 1:
prediction = prediction / 255
truth = truth / 255
plt.text(example.shape[0] / 2 * 1.05,
-example.shape[1] / 12,
'████',
color=np.clip(prediction, 0, 1))
plt.text(example.shape[0] / 2 * 1.05,
-example.shape[1] / 35,
'████',
color=np.clip(truth, 0, 1))
plt.title(f'\nPrediction:\nTruth:',
fontdict={'horizontalalignment': 'right'})