###############################################
# Action 1) Get max activation for a slection of feat maps
###############################################
get_max_act = False
if get_max_act:
if not model:
model = load_model('./Data/vgg16_weights.h5')
if not Dec:
Dec = KerasDeconv.DeconvNet(model)
d_act_path = './Data/dict_top9_mean_act.pickle'
d_act = {"convolution2d_13": {},
"convolution2d_10": {}
}
for feat_map in range(10):
d_act["convolution2d_13"][feat_map] = find_top9_mean_act(
data, Dec, "convolution2d_13", feat_map, batch_size=32)
d_act["convolution2d_10"][feat_map] = find_top9_mean_act(
data, Dec, "convolution2d_10", feat_map, batch_size=32)
with open(d_act_path, 'w') as f:
pickle.dump(d_act, f)
###############################################
# Action 2) Get deconv images of images that maximally activate
# the feat maps selected in the step above
###############################################
deconv_img = False
if deconv_img:
d_act_path = './Data/dict_top9_mean_act.pickle'
d_deconv_path = './Data/dict_top9_deconv.pickle'
if not model:
model = load_model('./Data/vgg16_weights.h5')
###############################################
# Action 1) Get max activation for a slection of feat maps
###############################################
get_max_act = False
if get_max_act:
if not model:
model = load_model('./Data/vgg16_weights.h5')
if not Dec:
Dec = KerasDeconv.DeconvNet(model)
d_act_path = './Data/dict_top9_mean_act.pickle'
d_act = {"convolution2d_13": {},
"convolution2d_10": {}
}
for feat_map in range(10):
d_act["convolution2d_13"][feat_map] = find_top9_mean_act(
data, Dec, "convolution2d_13", feat_map, batch_size=32)
d_act["convolution2d_10"][feat_map] = find_top9_mean_act(
data, Dec, "convolution2d_10", feat_map, batch_size=32)
with open(d_act_path, 'w') as f:
pickle.dump(d_act, f)
###############################################
# Action 2) Get deconv images of images that maximally activate
# the feat maps selected in the step above
###############################################
deconv_img = False
if deconv_img:
d_act_path = './Data/dict_top9_mean_act.pickle'
d_deconv_path = './Data/dict_top9_deconv.pickle'
if not model:
model = load_model('./Data/vgg16_weights.h5')
deconv_model = deconv_models.Conv(pretrained=True)
out = model.predict(data)
#print(model.summary())
out = deconv_model.predict(out)
raise ValueError
if not Dec:
Dec = KerasDeconv.DeconvNet(model)
if get_max_act:
layers = ['block1_conv1', 'block1_conv2', 'pool1', 'block2_conv1', 'pool2', 'block2_conv2', 'pool3']
layer = layers[4]
d_act_path = './data/dict_top9_mean_act.pickle'
d_act = {layer: {},
}
## Filters to check
for feat_map in [3, 4]:
d_act[layer][feat_map] = find_top9_mean_act(
data, Dec, layer, feat_map, batch_size=32)
with open(d_act_path, 'w') as f:
pickle.dump(d_act, f)
###############################################
# Action 2) Get deconv images of images that maximally activate
# the feat maps selected in the step above
###############################################
deconv_img = True
if deconv_img:
d_act_path = './data/dict_top9_mean_act.pickle'
d_deconv_path = './data/dict_top9_deconv.pickle'
#get_deconv_images(d_act_path, d_deconv_path, data, Dec)
raise ValueError