from torchray.attribution.excitation_backprop import contrastive_excitation_backprop
from torchray.benchmark import get_example_data, plot_example
# Obtain example data.
model, x, category_id, _ = get_example_data()
# Contrastive excitation backprop.
saliency = contrastive_excitation_backprop(
model,
x,
category_id,
saliency_layer='features.9',
contrast_layer='features.30',
classifier_layer='classifier.6',
)
# Plots.
plot_example(x, saliency, 'contrastive excitation backprop', category_id)
# Obtain example data.
model, x, category_id_1, category_id_2 = get_example_data()
# Run on GPU if available.
device = get_device()
model.to(device)
x = x.to(device)
# Extremal perturbation backprop.
masks_1, _ = extremal_perturbation(
model,
x,
category_id_1,
reward_func=contrastive_reward,
debug=True,
areas=[0.12],
)
masks_2, _ = extremal_perturbation(
model,
x,
category_id_2,
reward_func=contrastive_reward,
debug=True,
areas=[0.05],
)
# Plots.
plot_example(x, masks_1, 'extremal perturbation', category_id_1)
plot_example(x, masks_2, 'extremal perturbation', category_id_2)
from torchray.attribution.common import Probe, get_module from torchray.attribution.linear_approx import gradient_to_linear_approx_saliency from torchray.benchmark import get_example_data, plot_example # Obtain example data. model, x, category_id, _ = get_example_data() # Linear approximation. saliency_layer = get_module(model, 'features.29') probe = Probe(saliency_layer, target='output') y = model(x) z = y[0, category_id] z.backward() saliency = gradient_to_linear_approx_saliency(probe.data[0]) # Plots. plot_example(x, saliency, 'linear approx', category_id)
from torchray.attribution.grad_cam import grad_cam from torchray.benchmark import get_example_data, plot_example # Obtain example data. model, x, category_id, _ = get_example_data() # Grad-CAM backprop. saliency = grad_cam(model, x, category_id, saliency_layer='features.29') # Plots. plot_example(x, saliency, 'grad-cam backprop', category_id)
from torchray.attribution.common import gradient_to_saliency from torchray.benchmark import get_example_data, plot_example # Obtain example data. model, x, category_id, _ = get_example_data() # Gradient method. x.requires_grad_(True) y = model(x) z = y[0, category_id] z.backward() saliency = gradient_to_saliency(x) # Plots. plot_example(x, saliency, 'gradient', category_id)
from torchray.attribution.common import Probe, get_module
from torchray.attribution.excitation_backprop import ExcitationBackpropContext
from torchray.attribution.excitation_backprop import gradient_to_excitation_backprop_saliency
from torchray.benchmark import get_example_data, plot_example
# Obtain example data.
model, x, category_id, _ = get_example_data()
# Contrastive excitation backprop.
saliency_layer = get_module(model, 'features.9')
saliency_probe = Probe(saliency_layer, target='output')
with ExcitationBackpropContext():
y = model(x)
z = y[0, category_id]
z.backward()
saliency = gradient_to_excitation_backprop_saliency(saliency_probe.data[0])
saliency_probe.remove()
# Plots.
plot_example(x, saliency, 'excitation backprop', category_id)
from torchray.attribution.common import gradient_to_saliency
from torchray.attribution.deconvnet import DeConvNetContext
from torchray.benchmark import get_example_data, plot_example
# Obtain example data.
model, x, category_id, _ = get_example_data()
# DeConvNet method.
x.requires_grad_(True)
with DeConvNetContext():
y = model(x)
z = y[0, category_id]
z.backward()
saliency = gradient_to_saliency(x)
# Plots.
plot_example(x, saliency, 'deconvnet', category_id)
from torchray.attribution.guided_backprop import guided_backprop from torchray.benchmark import get_example_data, plot_example # Obtain example data. model, x, category_id, _ = get_example_data() # Guided backprop. saliency = guided_backprop(model, x, category_id) # Plots. plot_example(x, saliency, 'guided backprop', category_id)
from torchray.attribution.rise import rise from torchray.benchmark import get_example_data, plot_example from torchray.utils import get_device # Obtain example data. model, x, category_id, _ = get_example_data() # Run on GPU if available. device = get_device() model.to(device) x = x.to(device) # RISE method. saliency = rise(model, x) saliency = saliency[:, category_id].unsqueeze(0) # Plots. plot_example(x, saliency, 'RISE', category_id)