-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathdissect.py
156 lines (136 loc) · 6.81 KB
/
dissect.py
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
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
import os
import argparse
import torch
import tqdm
from yacs.config import CfgNode as CN
import seaborn as sns
import matplotlib.pyplot as plt
import datetime
from data import NYUv2_Dataloader, KITTI_Dataloader
from model import MFF_Model, BTS_Model
from utils import get_num_units_bins_and_nums_for_plt, get_sid_thresholds, discretize_depths, count_discrete_depth, \
add_fmaps_on_discrete_depths, compute_selectivity
def main():
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
cfg = CN()
cfg.MODEL_WEIGHTS_FILE = "model_weights/mff_senet_asn"
cfg.LAYERS = 'D_MFF'
cfg.ON_TRAINING_DATA = False
# ---------------
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument("opts", help="Modify configs using the command-line", default=None, nargs=argparse.REMAINDER)
args = parser.parse_args()
cfg.merge_from_list(args.opts)
# ---------------
cfg.LAYERS = cfg.LAYERS.split('_')
if 'mff_resnet' in cfg.MODEL_WEIGHTS_FILE:
cfg.MODEL_NAME = 'MFF_resnet'
cfg.DATASET = 'nyuv2'
elif 'mff_senet' in cfg.MODEL_WEIGHTS_FILE:
cfg.MODEL_NAME = 'MFF_senet'
cfg.DATASET = 'nyuv2'
elif 'bts_nyu' in cfg.MODEL_WEIGHTS_FILE:
cfg.MODEL_NAME = 'BTS_nyu'
cfg.DATASET = 'nyuv2'
elif 'bts_kitti' in cfg.MODEL_WEIGHTS_FILE:
cfg.MODEL_NAME = 'BTS_kitti'
cfg.DATASET = 'kitti'
else:
raise NotImplementedError
if cfg.DATASET == 'nyuv2':
if cfg.ON_TRAINING_DATA:
if 'BTS' in cfg.MODEL_NAME:
cfg.DATASET_FILE = 'dataset/nyuv2/bts_train/bts_nyuv2_train.csv'
else:
cfg.DATASET_FILE = 'dataset/nyuv2/nyuv2_train.csv'
else:
cfg.DATASET_FILE = 'dataset/nyuv2/nyuv2_test.csv'
cfg.MAX_DEPTH_EVAL = 10.
elif cfg.DATASET == 'kitti':
if cfg.ON_TRAINING_DATA:
cfg.DATASET_FILE = 'dataset/kitti/eigen_train_files_with_gt.txt'
else:
cfg.DATASET_FILE = 'dataset/kitti/eigen_test_files_with_gt_652.txt'
cfg.MAX_DEPTH_EVAL = 80.
else:
raise NotImplementedError
# cfg.OUTPUT_DIR = f'outputs/dissect/bts_nyu_upconv2_eval'
cfg.OUTPUT_DIR = f'outputs/dissect/{cfg.MODEL_WEIGHTS_FILE.split("/")[-1]}_{"train" if cfg.ON_TRAINING_DATA else "eval"}-[{(datetime.datetime.now()).strftime("%m%d%H%M%S")}]'
print(cfg)
if cfg.DATASET == 'nyuv2':
dataloader = NYUv2_Dataloader('dissect', cfg.DATASET_FILE, cfg.MODEL_NAME, 1)
elif cfg.DATASET == 'kitti':
dataloader = KITTI_Dataloader('dissect', cfg.DATASET_FILE, cfg.MODEL_NAME, 1)
else:
raise NotImplementedError
if 'MFF' in cfg.MODEL_NAME:
model = MFF_Model(cfg.MODEL_NAME, cfg.MODEL_WEIGHTS_FILE, device).to(device)
elif 'BTS' in cfg.MODEL_NAME:
model = BTS_Model(cfg.DATASET, cfg.MAX_DEPTH_EVAL, cfg.MODEL_WEIGHTS_FILE).to(device)
else:
raise NotImplementedError
dissect(model, cfg.MODEL_NAME, dataloader, cfg.DATASET, cfg.LAYERS, cfg.OUTPUT_DIR, device)
def dissect(model, model_name, dataloader, dataset_name, layers, output_dir, device):
if output_dir:
os.makedirs(output_dir, exist_ok=True)
model.eval()
num_units_list, num_bins_list, plt_row_list, plt_col_list, figsize_w_list, figsize_h_list = get_num_units_bins_and_nums_for_plt(layers, model_name)
depth_cnt_list, depth_sum_list, depth_avg_list = [], [], []
for layer_i in range(len(layers)):
depth_cnt_list.append(torch.zeros(num_bins_list[layer_i], dtype=torch.double).to(device))
depth_sum_list.append(torch.zeros((num_units_list[layer_i], num_bins_list[layer_i]), dtype=torch.double).to(device))
sid_thresholds_list = [get_sid_thresholds(num_bins_list[layer_i], dataset_name) for layer_i in range(len(layers))]
# -------------------- start batch iteration --------------------
for batch in tqdm.tqdm(dataloader, desc='Dissection'):
imgs = batch['image'].to(device)
if 'MFF' in model_name:
features_list = model.get_features(imgs, layers, also_get_output=False)
elif 'BTS' in model_name:
focals = batch['focal'].to(device) if 'focal' in batch.keys() else None
features_list = model.get_features(imgs, layers, focals)
else:
raise NotImplementedError
for layer_i in range(len(layers)):
depths = batch['depth']
discrete_concepts = discretize_depths(depths, dataset_name, thresholds=sid_thresholds_list[layer_i])
depth_cnt_list[layer_i] = count_discrete_depth(discrete_concepts, depth_cnt_list[layer_i])
depth_sum_list[layer_i] = add_fmaps_on_discrete_depths(discrete_concepts, features_list[layer_i],
depth_sum_list[layer_i])
# --------------- end of batch iteration --------------------
for layer_i in range(len(layers)):
depth_cnt_list[layer_i] = depth_cnt_list[layer_i].cpu().numpy()
depth_sum_list[layer_i] = depth_sum_list[layer_i].cpu().numpy()
depth_avg_list.append(depth_sum_list[layer_i] / (depth_cnt_list[layer_i] + 1e-15))
unit_max_bin_list, selectivity_index_list = [], []
for layer_i in range(len(layers)):
depth_avg = depth_avg_list[layer_i]
# ---- compute selectivity ----
depth_avg_abs = abs(depth_avg)
unit_max_bin = depth_avg_abs.argmax(-1)
selectivity_index = compute_selectivity(depth_avg_abs, unit_max_bin).numpy()
unit_max_bin_list.append(unit_max_bin)
selectivity_index_list.append(selectivity_index)
# --------------------------
# ---- plot ----
sns.set_theme(style="darkgrid")
fig, ax = plt.subplots(plt_row_list[layer_i], plt_col_list[layer_i],
figsize=(figsize_w_list[layer_i], figsize_h_list[layer_i]))
for i in range(plt_row_list[layer_i]):
for j in range(plt_col_list[layer_i]):
unit_i = i * plt_col_list[layer_i] + j
ax[i, j].set_title(
f'Unit {unit_i}, Layer {layers[layer_i]}, Selectivity: {selectivity_index[unit_i]:.04f}',
fontsize=12)
ax[i, j].bar(range(len(depth_cnt_list[layer_i])), depth_avg[unit_i], color='red', width=0.85)
plt.xticks(fontsize=12)
plt.yticks(fontsize=12)
plt.subplots_adjust(wspace=0.15)
selec_mean = selectivity_index.mean()
save_fig_name = os.path.join(output_dir,
f'{layers[layer_i]}-{selec_mean:.4f}.png')
fig.savefig(save_fig_name, bbox_inches='tight')
plt.close(fig)
print(f'selec_mean_{layers[layer_i]}', selectivity_index.mean())
return unit_max_bin_list, selectivity_index_list
if __name__ == '__main__':
main()