'add_4': 'rectangle',
'add_6': 'rectangle',
'add_8': 'rectangle'
}
n_nodes_in_graph = 3
edge_label_font_size = '22'
header_font_size = '40'
add_heat_map_on_connections = True
calc_heat_map_on_connections = False
# Get the saved weights
list_of_weights = glob.glob(pjoin(EXP_PATH, 'weights.*.hdf5'))
weights_dir = list_of_weights[-1]
# Load config
config = load_from_file(EXP_PATH, ['config'])[0]
# Update model config
config['set_gmm_activation_layer_as_output'] = True
# Load model
model = GMM_CNN()
model.load_model(weights_dir=weights_dir, config=config)
layers_to_inference = []
for key in vis_option:
layers_to_inference.append(key)
layers_to_inference.append(model.keras_model.output_names[0])
# make clusters saving directory
vis_dir = pjoin(EXP_PATH, 'clusters_vis')
# get data
(_, _), (x_val, y_val) = cifar10.load_data()
import matplotlib.pyplot as plt
import numpy as np
from keras.datasets import cifar10
from keras.utils import to_categorical
from gmm_cnn import GMM_CNN
from receptivefield import ReceptiveField
from utils.file import load_from_file
from utils.file import makedir
from utils.vis_utils import crop_image, get_cifar10_labels
EXP_PATH = pjoin(
*['C:', os.environ["HOMEPATH"], 'Desktop', 'tmp', 'resnet20_cifar10'])
# Load config
config = load_from_file(EXP_PATH, ['config'])[0]
# Load model
model = GMM_CNN()
model.load_model(config=config)
# Visualization parameters
# How to vis the representatives images: 1. 'rectangle' around the origin image. 2. 'patches' draw only the rf
vis_option = 'rectangle'
label_fontsize = 26
dot_radius = 0.5
n_representatives = 6
# make clusters saving directory
vis_dir = pjoin(EXP_PATH, 'clusters_vis')
makedir(vis_dir)
def create_connections_heat_maps(gmm_model, exp_path, RF, data, connections_dict, saving_dir):
""" This script creates the heat map between two clusters connected in the inference tree"""
# TODO the data needs to be with mean subtraction
makedir(saving_dir)
clusters_rep_path = pjoin(exp_path, 'clusters_representatives.json')
if os.path.isfile(clusters_rep_path) and os.access(clusters_rep_path, os.R_OK):
# checks if file exists
print("File exists and is readable")
print("Load results...")
clusters_representatives = load_from_file(exp_path, ['clusters_representatives'])[0]
else:
raise ValueError("Either file is missing or is not readable")
for explained_key in connections_dict:
if 'class' in explained_key or 'fc' in explained_key:
continue
explained_layer = explained_key.split('_')[0]
for i in range(len(explained_key.split('_')) - 2):
explained_layer = explained_layer + '_' + explained_key.split('_')[i+1]
explained_cluster = explained_key.split('_')[-1]
explained_gmm = gmm_model.get_correspond_gmm_layer(explained_layer)
c_reps = clusters_representatives[explained_gmm][explained_cluster]
c_imgs_indice = np.asarray(c_reps['image'])
c_imgs = data[c_imgs_indice]
if len(connections_dict[explained_key]) != 0:
explaining_key = next(iter(connections_dict[explained_key]))
explaining_layer = explaining_key.split('_')[0] + '_' + explaining_key.split('_')[1]
explaining_gmm = gmm_model.get_correspond_gmm_layer(explaining_layer)
imgs_pred = gmm_model.predict(c_imgs)
explaining_pred = imgs_pred['GMM'][explaining_gmm]
for explaining_key in connections_dict[explained_key]:
explaining_layer = explaining_key.split('_')[0] + '_' + explaining_key.split('_')[1]
explaining_cluster = int(explaining_key.split('_')[2]) - 1
saving_plot_path = pjoin(saving_dir, explained_key + '_' + explaining_key + '.png')
if os.path.isfile(saving_plot_path) and os.access(saving_plot_path, os.R_OK):
# checks if heat map file exists
continue
_, _, origin_size = RF.target_neuron_rf(explained_layer, [0, 0],
rf_layer_name=explaining_layer, return_origin_size=True)
heat_map = np.zeros((origin_size[0], origin_size[1]))
for i in range(len(c_imgs_indice)):
h = c_reps['spatial_location']['row'][i]
w = c_reps['spatial_location']['col'][i]
size, center, UL_pos, origin_center = RF.target_neuron_rf(explained_layer, [h, w],
rf_layer_name=explaining_layer,
return_upper_left_pos=True,
return_origin_center=True)
row_offset = abs(min(0, origin_center[0] - np.floor_divide(origin_size[0], 2)))
col_offset = abs(min(0, origin_center[1] - np.floor_divide(origin_size[1], 2)))
upper_left_row = UL_pos[0]
upper_left_col = UL_pos[1]
rf_preds = explaining_pred[i, upper_left_row:upper_left_row + size[0],
upper_left_col:upper_left_col + size[1], explaining_cluster]
pad_rf = np.zeros((origin_size[0], origin_size[1]))
pad_rf[row_offset:rf_preds.shape[0] + row_offset, col_offset:rf_preds.shape[1] + col_offset] = rf_preds
heat_map += pad_rf
heat_map = heat_map / len(c_imgs_indice)
fig = plt.figure()
label = str(connections_dict[explained_key][explaining_key]['CN']) + ', ' +\
str(connections_dict[explained_key][explaining_key]['LR'])
ax = fig.add_subplot(111)
ax.matshow(heat_map, interpolation='hermite')
ax.axis('off')
ax.set_title(label=label, fontsize=70)
plt.tight_layout()
plt.savefig(saving_plot_path)
plt.close(fig)