def test_direction(cossim_pow, inceptionv1):
mixed_4a_depth = 508
random_direction = np.random.random((mixed_4a_depth))
objective = objectives.direction("mixed4a_pre_relu",
random_direction,
cossim_pow=cossim_pow)
assert_gradient_ascent(objective, inceptionv1)
def render_activation_grid_very_naive(self,
img,
layer="mixed4d",
W=42,
n_steps=256):
# Get the activations
with tf.Graph().as_default(), tf.Session() as sess:
t_input = tf.placeholder("float32", [None, None, None, 3])
T = render.import_model(self.model, t_input, t_input)
acts = T(layer).eval({t_input: img[None]})[0]
acts_flat = acts.reshape([-1] + [acts.shape[2]])
# Render an image for each activation vector
def param_f():
return param.image(W, batch=acts_flat.shape[0])
obj = objectives.Objective.sum([
objectives.direction(layer, v, batch=n)
for n, v in enumerate(acts_flat)
])
thresholds = (n_steps // 2, n_steps)
vis_imgs = render.render_vis(self.model,
obj,
param_f,
thresholds=thresholds)[-1]
# Combine the images and display the resulting grid
vis_imgs_ = vis_imgs.reshape(list(acts.shape[:2]) + [W, W, 3])
vis_imgs_cropped = vis_imgs_[:, :, 2:-2, 2:-2, :]
show(np.hstack(np.hstack(vis_imgs_cropped)))
return vis_imgs_cropped
def neuron_groups(img, filename, layer, n_groups=10, attr_classes=None, filenumber=0):
# Compute activations
dirname = '../images/' + filename+'/'
if attr_classes is None:
attr_classes = []
with tf.Graph().as_default(), tf.Session():
t_input = tf.placeholder_with_default(img, [None, None, 3])
T = render.import_model(model, t_input, t_input)
acts = T(layer).eval()
# We'll use ChannelReducer (a wrapper around scikit learn's factorization tools)
# to apply Non-Negative Matrix factorization (NMF).
nmf = ChannelReducer(n_groups, "NMF")
spatial_factors = nmf.fit_transform(acts)[0].transpose(2, 0, 1).astype("float32")
channel_factors = nmf._reducer.components_.astype("float32")
# Let's organize the channels based on their horizontal position in the image
x_peak = np.argmax(spatial_factors.max(1), 1)
ns_sorted = np.argsort(x_peak)
spatial_factors = spatial_factors[ns_sorted]
channel_factors = channel_factors[ns_sorted]
# And create a feature visualziation of each group
param_f = lambda: param.image(80, batch=n_groups)
obj = sum(objectives.direction(layer, channel_factors[i], batch=i)
for i in range(n_groups))
group_icons = render.render_vis(model, obj, param_f, verbose=False)[-1]
# We'd also like to know about attribution
# First, let's turn each group into a vector over activations
group_vecs = [spatial_factors[i, ..., None] * channel_factors[i]
for i in range(n_groups)]
attrs = np.asarray([raw_class_group_attr(img, layer, attr_class, group_vecs)
for attr_class in attr_classes])
print(
attrs
)
try:
os.mkdir(dirname )
except Exception as e:
print(e)
# Let's render the visualization!
finally:
with open(dirname + '/attrs.txt', 'w') as f_w:
f_w.write(str(attrs))
for index, icon in enumerate(group_icons):
imgdata=to_image_url(icon)
print(imgdata)
imgdata = base64.b64decode(str(imgdata))
print(imgdata)
with open(dirname + str(index) + '.png', 'wb') as f_jpg:
f_jpg.write(imgdata)
def neuron_groups(model, img, layer, n_groups=6, attr_classes=[]):
# Compute activations
with tf.Graph().as_default(), tf.Session():
t_input = tf.placeholder_with_default(img, [None, None, 3])
T = render.import_model(model, t_input, t_input)
acts = T(layer).eval()
# We'll use ChannelReducer (a wrapper around scikit learn's factorization tools)
# to apply Non-Negative Matrix factorization (NMF).
nmf = ChannelReducer(n_groups, "PCA")
print(layer, n_groups)
spatial_factors = nmf.fit_transform(acts)[0].transpose(2, 0, 1).astype("float32")
channel_factors = nmf._reducer.components_.astype("float32")
# Let's organize the channels based on their horizontal position in the image
x_peak = np.argmax(spatial_factors.max(1), 1)
ns_sorted = np.argsort(x_peak)
spatial_factors = spatial_factors[ns_sorted]
channel_factors = channel_factors[ns_sorted]
# And create a feature visualziation of each group
param_f = lambda: param.image(80, batch=n_groups)
obj = sum(objectives.direction(layer, channel_factors[i], batch=i)
for i in range(n_groups))
group_icons = render.render_vis(model, obj, param_f, verbose=False)[-1]
# We'd also like to know about attribution
#
# First, let's turn each group into a vector over activations
group_vecs = [spatial_factors[i, ..., None] * channel_factors[i]
for i in range(n_groups)]
attrs = np.asarray([raw_class_group_attr(img, layer, attr_class, model, group_vecs)
for attr_class in attr_classes])
gray_scale_groups = [skimage.color.rgb2gray(icon) for icon in group_icons]
# Let's render the visualization!
data = {
"img": _image_url(img),
"n_groups": n_groups,
"spatial_factors": [_image_url(factor[..., None] / np.percentile(spatial_factors, 99) * [1, 0, 0]) for factor in
spatial_factors],
"group_icons": [_image_url(icon) for icon in gray_scale_groups]
}
# with open('ng.pickle', 'wb') as handle:
# pickle.dump(data, handle, protocol=pickle.HIGHEST_PROTOCOL)
# with open('./svelte_python/ng.pickle', 'rb') as p_file:
# data = pickle.load(p_file)
generate_html('neuron_groups', data)
def render_activation_grid_less_naive(self,
img,
layer="mixed4d",
W=42,
n_groups=6,
subsample_factor=1,
n_steps=256):
# Get the activations
with tf.Graph().as_default(), tf.Session() as sess:
t_input = tf.placeholder("float32", [None, None, None, 3])
T = render.import_model(self.model, t_input, t_input)
acts = T(layer).eval({t_input: img[None]})[0]
acts_flat = acts.reshape([-1] + [acts.shape[2]])
N = acts_flat.shape[0]
# The trick to avoiding "decoherence" is to recognize images that are
# for similar activation vectors and
if n_groups > 0:
reducer = ChannelReducer(n_groups, "NMF")
groups = reducer.fit_transform(acts_flat)
groups /= groups.max(0)
else:
groups = np.zeros([])
print(groups.shape)
# The key trick to increasing memory efficiency is random sampling.
# Even though we're visualizing lots of images, we only run a small
# subset through the network at once. In order to do this, we'll need
# to hold tensors in a tensorflow graph around the visualization process.
with tf.Graph().as_default() as graph, tf.Session() as sess:
# Using the groups, create a paramaterization of images that
# partly shares paramters between the images for similar activation
# vectors. Each one still has a full set of unique parameters, and could
# optimize to any image. We're just making it easier to find solutions
# where things are the same.
group_imgs_raw = param.fft_image([n_groups, W, W, 3])
unique_imgs_raw = param.fft_image([N, W, W, 3])
opt_imgs = param.to_valid_rgb(tf.stack([
0.7 * unique_imgs_raw[i] +
0.5 * sum(groups[i, j] * group_imgs_raw[j]
for j in range(n_groups)) for i in range(N)
]),
decorrelate=True)
# Construct a random batch to optimize this step
batch_size = 64
rand_inds = tf.random_uniform([batch_size], 0, N, dtype=tf.int32)
pres_imgs = tf.gather(opt_imgs, rand_inds)
pres_acts = tf.gather(acts_flat, rand_inds)
obj = objectives.Objective.sum([
objectives.direction(layer, pres_acts[n], batch=n)
for n in range(batch_size)
])
# Actually do the optimization...
T = render.make_vis_T(self.model, obj, param_f=pres_imgs)
tf.global_variables_initializer().run()
for i in range(n_steps):
T("vis_op").run()
if (i + 1) % (n_steps // 2) == 0:
show(pres_imgs.eval()[::4])
vis_imgs = opt_imgs.eval()
# Combine the images and display the resulting grid
print("")
vis_imgs_ = vis_imgs.reshape(list(acts.shape[:2]) + [W, W, 3])
vis_imgs_cropped = vis_imgs_[:, :, 2:-2, 2:-2, :]
show(np.hstack(np.hstack(vis_imgs_cropped)))
return vis_imgs_cropped
def neuron_groups(imglist, filenamelist, layer, n_groups=6, attr_classes=None):
# Compute activations
filename = ''
for f in filenamelist:
filename += f
with open('result/' + filename + '.html', 'a') as f:
f.write('''<!DOCTYPE html>
<html>
<head >
<title>%s</title>
<script src='GroupWidget_1cb0e0d.js'></script>
</head>
<body>''' % (filename))
for key, img in enumerate(imglist):
if attr_classes is None:
attr_classes = []
with tf.Graph().as_default(), tf.Session():
t_input = tf.placeholder_with_default(img, [None, None, 3])
T = render.import_model(model, t_input, t_input)
acts = T(layer).eval()
# We'll use ChannelReducer (a wrapper around scikit learn's factorization tools)
# to apply Non-Negative Matrix factorization (NMF).
nmf = ChannelReducer(n_groups, "NMF")
spatial_factors = nmf.fit_transform(acts)[0].transpose(
2, 0, 1).astype("float32")
channel_factors = nmf._reducer.components_.astype("float32")
# Let's organize the channels based on their horizontal position in the image
x_peak = np.argmax(spatial_factors.max(1), 1)
ns_sorted = np.argsort(x_peak)
spatial_factors = spatial_factors[ns_sorted]
channel_factors = channel_factors[ns_sorted]
# And create a feature visualziation of each group
param_f = lambda: param.image(80, batch=n_groups)
obj = sum(
objectives.direction(layer, channel_factors[i], batch=i)
for i in range(n_groups))
group_icons = render.render_vis(model, obj, param_f, verbose=False)[-1]
# We'd also like to know about attribution
# First, let's turn each group into a vector over activations
group_vecs = [
spatial_factors[i, ..., None] * channel_factors[i]
for i in range(n_groups)
]
attrs = np.asarray([
raw_class_group_attr(img, layer, attr_class, group_vecs)
for attr_class in attr_classes
])
print(attrs)
# Let's render the visualization!
with open('result/' + filename + '.html', 'a') as f:
f.write(''' <main%s></main%s>
<script>
var app = new GroupWidget_1cb0e0d({
target: document.querySelector( 'main%s' ),''' %
(key, key, key))
f.write('''data: {''')
f.write('"img":"%s",\n' % str(_image_url(img)))
f.write('"n_groups"' + ":" + str(n_groups) + ',\n')
f.write('"spatial_factors"' + ":" + str([
_image_url(factor[..., None] /
np.percentile(spatial_factors, 99) * [1, 0, 0])
for factor in spatial_factors
]) + ',\n')
f.write('"group_icons"' + ":" +
str([_image_url(icon) for icon in group_icons]) + ',\n')
f.write('''} });''')
f.write('''</script>''')
with open('result/' + filename + '.html', 'a') as f:
f.write('''</body></html >''')
print(filename)
