nefesi/create_nefesimodel_hpeu.py at Nefesi2022 · CVC-Color/nefesi · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
"""
This file contains a toy examples to have a first contact with Nefesi, and Keras.
This file has been created with tensorflow (and tensorflow-gpu) 1.8.0, keras 2.2.0, and python 3.6 (with anaconda3 interpreter)
"""
import cv2
import torch
from torchvision import transforms
import numpy as np
import functools
from functions.network_data2 import NetworkData
import types
import functions.GPUtil as gpu
BATCH_SIZE = 100
from HPEU.module_hpeu_gf import DeepGuidedFilterGuidedMapConvGF
from  functions.image import ImageDataset
from functions.read_activations import get_activations
import interface_DeepFramework.DeepFramework as DeepF


def preproces_hpeu_img( imgs_hr,small_size=(64,64)):

    imgs_hr=np.array(imgs_hr)/256
    imgs_lr=cv2.resize(np.array(imgs_hr),small_size)


    gray = 1 - (0.299 * (imgs_lr[:, :, 0] + 1) + 0.587 * (imgs_lr[:, :, 1] + 1) + 0.114 * (imgs_lr[:, :, 2] + 1)) / 2
    gray = [np.stack((gray,) * 3, axis=-1)]

    imgs_lr = [imgs_lr]
    imgs = imgs_lr + gray
    tnsr = [transforms.ToTensor()(img) for img in imgs]


    return tnsr


def main():

    #Here You can find all the funcitons that allow us to visualize and quantify a trained Neural Network.
    # To perform the calculation taking into acount the negative activations instead of positives, uncoment line 53 in the file read_activations


    # Load the Model with your weigths first
    model_path= "C:/Users/g84194584/Desktop/Code/nefesi/HPEU/mse_net_latest.pth"
    device = torch.device('cuda:{}'.format(0))
    model= DeepGuidedFilterGuidedMapConvGF()
    model.load_state_dict(torch.load(model_path))
    model = model.lr
    model = DeepF.deep_model(model)


    # Create a list with the layers that you want to analyze and 0 if they are encoding or 1 if they are decoding
    layers_interest = [['conv_in', 0], ['leu1', 0], ['srm_down1', 0] ,['down1', 0],['leu2', 0], ['srm_down2' , 0],['down2', 0],['leu3', 0], ['srm_down3', 0] ,['down3', 0],['leu4', 0], ['srm_down4', 0] ,['down4', 0], ['leu5', 0],[ 'srm_down5', 0] ,['up0', 1],['srm_up1', 1] ,['up1', 1],['srm_up2', 1],[ 'up2', 1], ['srm_up3', 1],['up3', 1], ['srm_up4', 1],['leuwithoutup4', 1],['srm_up5', 1]]


    # Create the DatasetLoader: select your imagepath and your preprocessing functon (in case you have one)
    Path_images='C:/Users/g84194584/Desktop/Imagenet64'
    preproces_function=preproces_hpeu_img
    dataset = ImageDataset(src_dataset=Path_images,target_size=(64,64),preprocessing_function=preproces_function,color_mode='rgb')

    # Path where you will save your results
    save_path= 'C:/Users/g84194584/Desktop/Code/nefesi/nefesi/hpeu_imagenet_minus'


    Nefesimodel= NetworkData(model=model,layer_data=layers_interest,save_path = save_path, dataset=dataset,default_file_name = 'HPEU_imagenet',input_shape=[(1,3,64,64),(1,3,64,64)])
    Nefesimodel.generate_neuron_data()

    # calculate the top scoring images
    Nefesimodel.eval_network()

    print('Activation Calculus done!')

    # calculate the Neuron feature of each neuron (weighted average of top scoring images)
    Nefesimodel.calculateNF()
    print('NF done!')


    # calculate the Color selectivity of each neuron
    dataset = Nefesimodel.dataset
    for layer in Nefesimodel.get_layers_name():
        layer_data = Nefesimodel.get_layer_by_name(layer)
        print(layer)
        for n in range(Nefesimodel.get_len_neurons_of_layer(layer)):
            neurona = Nefesimodel.get_neuron_of_layer(layer, n)
            neurona.color_selectivity_idx_new(Nefesimodel, layer_data, dataset)
    Nefesimodel.save_to_disk('color_indx')

    # calculate the Similarity Index of each neuron in the same layer
    for layer in Nefesimodel.get_layers_name():
        print(layer)
        Nefesimodel.get_layer_by_name(layer).similarity_index = None
        x=Nefesimodel.similarity_idx(layer)
        print(x)

    Nefesimodel.save_to_disk('similarity')


if __name__ == '__main__':
    main()