Lucid Modelzoo¶
If you want to study techniques for visualizing and understanding neural networks, it's important to be able to try your experiments on multiple models.
Lucid is a library for visualizing neural networks. As of lucid v0.3, we provide a consistent API for interacting with 27 different vision models.
General Setup¶
In [0]:
# Expanded modelzoo is only available as of lucid v0.3
!pip install --quiet lucid==0.3
import numpy as np
import tensorflow as tf
from lucid.misc.io import show, load
import lucid.optvis.objectives as objectives
import lucid.optvis.param as param
import lucid.optvis.render as render
import lucid.optvis.transform as transform
Import Modelzoo¶
In [0]:
# Lucid's modelzoo can be accessed as classes in vision_models
import lucid.modelzoo.vision_models as models
# ... or throguh a more systematic factory API
import lucid.modelzoo.nets_factory as nets
List Models¶
As of lucid v0.3
In [3]:
print ""
print "Model".ljust(27), " ", "Dataset"
print ""
for name, Model in nets.models_map.iteritems():
print name.ljust(27), " ", Model.dataset
Model Dataset PnasnetMobile_slim ImageNet PnasnetLarge_slim ImageNet InceptionV1_caffe_Places365 Places365 MobilenetV1_025_slim ImageNet AlexNet_caffe_Places365 Places365 NasnetMobile_slim ImageNet InceptionV2_slim ImageNet ResnetV1_101_slim ImageNet InceptionV1_caffe ImageNet InceptionResnetV2_slim ImageNet InceptionV3_slim ImageNet InceptionV1_caffe_Places205 Places205 InceptionV1_slim ImageNet ResnetV1_50_slim ImageNet CaffeNet_caffe ImageNet InceptionV4_slim ImageNet VGG19_caffe ImageNet ResnetV1_152_slim ImageNet MobilenetV1_050_slim ImageNet NasnetLarge_slim ImageNet AlexNet ImageNet ResnetV2_101_slim ImageNet ResnetV2_152_slim ImageNet MobilenetV1_slim ImageNet VGG16_caffe ImageNet InceptionV1 ImageNet ResnetV2_50_slim ImageNet
List Model Layers¶
In [4]:
models.InceptionV4_slim.layers
Out[4]:
[{'name': 'InceptionV4/InceptionV4/Conv2d_1a_3x3/Relu',
'size': 32,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Conv2d_2a_3x3/Relu',
'size': 32,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Conv2d_2b_3x3/Relu',
'size': 64,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_3a/concat',
'size': 160,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_4a/concat',
'size': 192,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_5a/concat',
'size': 384,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_5b/concat',
'size': 384,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_5c/concat',
'size': 384,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_5d/concat',
'size': 384,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_5e/concat',
'size': 384,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_6a/concat',
'size': 1024,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_6b/concat',
'size': 1024,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_6c/concat',
'size': 1024,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_6d/concat',
'size': 1024,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_6e/concat',
'size': 1024,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_6f/concat',
'size': 1024,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_6g/concat',
'size': 1024,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_6h/concat',
'size': 1024,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_7a/concat',
'size': 1536,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_7b/concat',
'size': 1536,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_7c/concat',
'size': 1536,
'type': 'conv'},
{'name': 'InceptionV4/InceptionV4/Mixed_7d/concat',
'size': 1536,
'type': 'conv'},
{'name': 'InceptionV4/Logits/Predictions', 'size': 1001, 'type': 'dense'}]
Show Model Graph¶
In [5]:
model = models.InceptionV4_slim()
model.load_graphdef()
model.show_graph()
Visualize Neuron¶
See the lucid tutorial to learn more.
We pick InceptionV4/InceptionV4/Mixed_6b/concat from above, and chose to focus on unit 0.
In [9]:
model = models.InceptionV4_slim()
model.load_graphdef()
_ = render.render_vis(model, "InceptionV4/InceptionV4/Mixed_6b/concat:0")
512 12.305752
Caricature¶
See the inversion and caricature notebook to learn more.
In [8]:
from lucid.recipes.caricature import feature_inversion
img = load("https://storage.googleapis.com/lucid-static/building-blocks/examples/dog_cat.png")
model = models.InceptionV4_slim()
model.load_graphdef()
result = feature_inversion(img, model, "InceptionV4/InceptionV4/Mixed_6b/concat", n_steps=512, cossim_pow=0.0)
show(result)
WARNING:lucid.misc.io.showing:Show only supports numpy arrays so far. Using repr().
None