Linear-Explanations/explain_neurons_no_gs.py at main · Trustworthy-ML-Lab/Linear-Explanations · 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
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
import argparse
import os
import math
import numpy as np
import pandas as pd
import torch

from tqdm import tqdm
import data_utils
import linear_explanation

#explains neurons without greedy search, just takes top5 concepts from relatively sparse layer.

def parse_arguments():
    parser = argparse.ArgumentParser(description='Linear explanations')
    parser.add_argument('--mode', type=str, default='label', help='which version to use label or siglip')
    parser.add_argument('--device', type=str, default='cuda', help='whether to use gpu')
    parser.add_argument('--dataset_name', type=str, default='imagenet_val', help='Dataset name')
    parser.add_argument('--batch_size', type=int, default=128, help='Batch size for clip and target model')
    parser.add_argument('--activations_dir', type=str, default='saved_activations', help='Save directory')
    parser.add_argument('--result_dir', type=str, default='results', help='Save directory')
    parser.add_argument('--clip_name', type=str, default='ViT-L-16-SigLIP-384', help='Which CLIP model to use')
    parser.add_argument('--concept_set', type=str, default=None, help='Concept set path, only used in siglip mode. If not provided uses default based on dataset.')
    parser.add_argument('--target_name', type=str, default='resnet50_imagenet', help='Target model')
    parser.add_argument('--target_layer', type=str, default='layer4', help='Target layer')
    parser.add_argument('--start_neuron', type=int, default=0, help='First neuron')
    parser.add_argument('--end_neuron', type=int, default=None, help='Last neuron, not included. Default is all neurons.')
    parser.add_argument('--description_length', type=int, default=5, help='Description length')
    parser.add_argument('--glm_neuron_batch', type=int, default=128, help="How many neurons we optimize for in one pass of GLM-Saga")
    parser.add_argument('--pool_mode', type=str, default='avg', help="Activation pooling function for i.e. CNN channel outputs. {avg, max, first}")
    return parser.parse_args()

def explain_neurons(args):
    if args.mode == "siglip":
        if args.concept_set == None:
            concept_set = 'data/concept_sets/combined_concepts_{}.txt'.format(args.dataset_name.split("_")[0])
        else:
            concept_set = args.concept_set
        concept_activations = linear_explanation.get_clip_feats(clip_name = args.clip_name, dataset_name = args.dataset_name,
                                                concept_set = concept_set,  save_dir = args.activations_dir,
                                                batch_size = args.batch_size, device = args.device)
        with open(concept_set, 'r') as f:
            concept_text = (f.read()).split('\n')

    elif args.mode == "label":
        concept_activations, concept_text = linear_explanation.get_onehot_labels(args.dataset_name, args.device)

    target_activations = linear_explanation.get_target_acts(target_name = args.target_name, dataset_name = args.dataset_name,
                                                                target_layer = args.target_layer, save_dir = args.activations_dir,
                                                                batch_size = args.batch_size, device = args.device,
                                                                start_neuron = args.start_neuron, end_neuron = args.end_neuron,
                                                                pool_mode = args.pool_mode)

    if not os.path.exists(args.result_dir):
        os.makedirs(args.result_dir)

    #deleting concepts that are not active enough, or have 0 standard devication
    concept_top5 = torch.mean(torch.topk(concept_activations, dim=0, k=5)[0], dim=0)
    top5_active = (concept_top5 >= 0.5)
    std_active = torch.std(concept_activations, dim=0) >= 1e-5
    active_concepts = top5_active*std_active
    concept_activations = concept_activations[:, active_concepts]
    concept_text = [concept for i, concept in enumerate(concept_text) if active_concepts[i]]
    #print(torch.sum(active_concepts))


    data_utils.save_train_test_split(args.dataset_name)
    train_ids = torch.load("data/data_splits/{}/train_ids.pt".format(args.dataset_name))
    val_ids = torch.load("data/data_splits/{}/val_ids.pt".format(args.dataset_name))

    top_concepts = []
    for i in tqdm(range(math.ceil(target_activations.shape[1]/args.glm_neuron_batch))):
        curr_target = target_activations[:, i*(args.glm_neuron_batch):(i+1)*(args.glm_neuron_batch)]
        train_data, val_data = linear_explanation.get_glm_datasets(concept_activations, curr_target, train_ids, val_ids)

        #train relatively_sparse model
        linear = linear_explanation.train_glm_model(train_data, val_data, args.device)
        vals, curr_top_concepts = torch.sort(linear.weight.detach(), dim=1, descending=True)
        top_concepts.append(curr_top_concepts)
    top_concepts = torch.cat(top_concepts, dim=0)

    result_df = {"layer":[], "unit":[], "val correlation":[]}

    for i in range(args.description_length):
        result_df["weight{}".format(i)] = []
        result_df["concept{}".format(i)] = []
    result_df["bias"] = []
    lengths = []
    end_neuron = args.start_neuron+target_activations.shape[1] #calculate it in case it is args.end_neuron = None
    for n_id, target_neuron in enumerate(range(args.start_neuron, end_neuron)):
        result_df["layer"].append(args.target_layer)
        result_df["unit"].append(target_neuron)
        print("Neuron: {}".format(target_neuron))

        train_target = target_activations[train_ids, n_id:n_id+1]
        val_target = target_activations[val_ids, n_id:n_id+1]
        concepts = top_concepts[n_id][:args.description_length]
        best_corr, new_linear = linear_explanation.train_linear_model(concept_activations[train_ids][:, concepts],
                                                                      concept_activations[val_ids][:, concepts],
                                                                       train_target, val_target, args.device)
        best_weight = new_linear.weight[0].detach()
        best_bias = float(new_linear.bias)

        finetune_texts = [concept_text[id] for id in concepts]
        result_df["val correlation"].append(best_corr)
        result_df["bias"].append(best_bias)

        new_vals, new_ids = torch.sort(best_weight, descending=True)
        lengths.append(len(new_vals))
        to_print = ""
        for i in range(args.description_length):
            try:
                id = new_ids[i]
                concept, weight = finetune_texts[id], new_vals[i].detach().cpu().numpy()
                result_df["concept{}".format(i)].append(concept)
                result_df["weight{}".format(i)].append(weight)
                to_print += "{:.2f}*{} + ".format(weight, concept)
            except(IndexError):
                result_df["concept{}".format(i)].append(0)
                result_df["weight{}".format(i)].append(0)
        pd_df = pd.DataFrame(result_df)
        pd_df.to_csv("{}/le_{}_{}_{}.csv".format(args.result_dir, args.mode, args.target_name, args.target_layer), index=False)
        to_print += "{:.2f}".format(best_bias)
        print(to_print)
        print("Val Correlation:{:.3f} \n".format(best_corr))

    print("Average val correlation: {:.4f}".format(np.mean(result_df["val correlation"])))
    print("Average explanation length: {:.2f}".format(np.mean(lengths)))


if __name__ == "__main__":
    args = parse_arguments()
    explain_neurons(args)