Spaces:
Sleeping
Sleeping
Create generic.py
Browse files- generic.py +137 -0
generic.py
ADDED
@@ -0,0 +1,137 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import pylab
|
2 |
+
from lxmert.src.modeling_frcnn import GeneralizedRCNN
|
3 |
+
import lxmert.src.vqa_utils as utils
|
4 |
+
from lxmert.src.processing_image import Preprocess
|
5 |
+
from transformers import LxmertTokenizer
|
6 |
+
from lxmert.src.huggingface_lxmert import LxmertForQuestionAnswering
|
7 |
+
from lxmert.src.lxmert_lrp import LxmertForQuestionAnswering as LxmertForQuestionAnsweringLRP
|
8 |
+
from tqdm import tqdm
|
9 |
+
from lxmert.src.ExplanationGenerator import GeneratorOurs, GeneratorBaselines, GeneratorOursAblationNoAggregation
|
10 |
+
import random
|
11 |
+
import numpy as np
|
12 |
+
import cv2
|
13 |
+
import torch
|
14 |
+
import matplotlib.pyplot as plt
|
15 |
+
from PIL import Image
|
16 |
+
import torchvision.transforms as transforms
|
17 |
+
from captum.attr import visualization
|
18 |
+
import requests
|
19 |
+
|
20 |
+
OBJ_URL = "./lxmert/unc-nlp/raw.githubusercontent.com_airsplay_py-bottom-up-attention_master_demo_data_genome_1600-400-20_objects_vocab.txt"
|
21 |
+
ATTR_URL = "./lxmert/unc-nlp/raw.githubusercontent.com_airsplay_py-bottom-up-attention_master_demo_data_genome_1600-400-20_attributes_vocab.txt"
|
22 |
+
VQA_URL = "./lxmert/unc-nlp/raw.githubusercontent.com_airsplay_lxmert_master_data_vqa_trainval_label2ans.json"
|
23 |
+
|
24 |
+
|
25 |
+
class ModelUsage:
|
26 |
+
def __init__(self, use_lrp=False):
|
27 |
+
self.vqa_answers = utils.get_data(VQA_URL)
|
28 |
+
|
29 |
+
# load models and model components
|
30 |
+
self.frcnn_cfg = utils.Config.from_pretrained("./lxmert/unc-nlp/frcnn-vg-finetuned")
|
31 |
+
self.frcnn_cfg.MODEL.DEVICE = "cpu"
|
32 |
+
|
33 |
+
self.frcnn = GeneralizedRCNN.from_pretrained("./lxmert/unc-nlp/frcnn-vg-finetuned", config=self.frcnn_cfg)
|
34 |
+
|
35 |
+
self.image_preprocess = Preprocess(self.frcnn_cfg)
|
36 |
+
|
37 |
+
self.lxmert_tokenizer = LxmertTokenizer.from_pretrained("./lxmert/unc-nlp/lxmert-base-uncased")
|
38 |
+
|
39 |
+
if use_lrp:
|
40 |
+
self.lxmert_vqa = LxmertForQuestionAnsweringLRP.from_pretrained("./lxmert/unc-nlp/lxmert-vqa-uncased")
|
41 |
+
else:
|
42 |
+
self.lxmert_vqa = LxmertForQuestionAnswering.from_pretrained("./lxmert/unc-nlp/lxmert-vqa-uncased")
|
43 |
+
|
44 |
+
self.lxmert_vqa.eval()
|
45 |
+
self.model = self.lxmert_vqa
|
46 |
+
|
47 |
+
# self.vqa_dataset = vqa_data.VQADataset(splits="valid")
|
48 |
+
|
49 |
+
def forward(self, item):
|
50 |
+
URL, question = item
|
51 |
+
|
52 |
+
self.image_file_path = URL
|
53 |
+
|
54 |
+
# run frcnn
|
55 |
+
images, sizes, scales_yx = self.image_preprocess(URL)
|
56 |
+
output_dict = self.frcnn(
|
57 |
+
images,
|
58 |
+
sizes,
|
59 |
+
scales_yx=scales_yx,
|
60 |
+
padding="max_detections",
|
61 |
+
max_detections=self.frcnn_cfg.max_detections,
|
62 |
+
return_tensors="pt"
|
63 |
+
)
|
64 |
+
inputs = self.lxmert_tokenizer(
|
65 |
+
question,
|
66 |
+
truncation=True,
|
67 |
+
return_token_type_ids=True,
|
68 |
+
return_attention_mask=True,
|
69 |
+
add_special_tokens=True,
|
70 |
+
return_tensors="pt"
|
71 |
+
)
|
72 |
+
self.question_tokens = self.lxmert_tokenizer.convert_ids_to_tokens(inputs.input_ids.flatten())
|
73 |
+
self.text_len = len(self.question_tokens)
|
74 |
+
# Very important that the boxes are normalized
|
75 |
+
normalized_boxes = output_dict.get("normalized_boxes")
|
76 |
+
features = output_dict.get("roi_features")
|
77 |
+
self.image_boxes_len = features.shape[1]
|
78 |
+
self.bboxes = output_dict.get("boxes")
|
79 |
+
self.output = self.lxmert_vqa(
|
80 |
+
input_ids=inputs.input_ids),
|
81 |
+
attention_mask=inputs.attention_mask,
|
82 |
+
visual_feats=features,
|
83 |
+
visual_pos=normalized_boxes,
|
84 |
+
token_type_ids=inputs.token_type_ids,
|
85 |
+
return_dict=True,
|
86 |
+
output_attentions=False,
|
87 |
+
)
|
88 |
+
return self.output
|
89 |
+
|
90 |
+
|
91 |
+
def save_image_vis(image_file_path, bbox_scores):
|
92 |
+
bbox_scores = image_scores
|
93 |
+
_, top_bboxes_indices = bbox_scores.topk(k=1, dim=-1)
|
94 |
+
img = cv2.imread(image_file_path)
|
95 |
+
mask = torch.zeros(img.shape[0], img.shape[1])
|
96 |
+
for index in range(len(bbox_scores)):
|
97 |
+
[x, y, w, h] = model_lrp.bboxes[0][index]
|
98 |
+
curr_score_tensor = mask[int(y):int(h), int(x):int(w)]
|
99 |
+
new_score_tensor = torch.ones_like(curr_score_tensor) * bbox_scores[index].item()
|
100 |
+
mask[int(y):int(h), int(x):int(w)] = torch.max(new_score_tensor, mask[int(y):int(h), int(x):int(w)])
|
101 |
+
mask = (mask - mask.min()) / (mask.max() - mask.min())
|
102 |
+
mask = mask.unsqueeze_(-1)
|
103 |
+
mask = mask.expand(img.shape)
|
104 |
+
img = img * mask.cpu().data.numpy()
|
105 |
+
cv2.imwrite(
|
106 |
+
'lxmert/lxmert/experiments/paper/new.jpg', img)
|
107 |
+
return img
|
108 |
+
|
109 |
+
|
110 |
+
model_lrp = ModelUsage(use_lrp=True)
|
111 |
+
lrp = GeneratorOurs(model_lrp)
|
112 |
+
baselines = GeneratorBaselines(model_lrp)
|
113 |
+
vqa_answers = utils.get_data(VQA_URL)
|
114 |
+
|
115 |
+
image_ids = [
|
116 |
+
# giraffe
|
117 |
+
'COCO_val2014_000000185590',
|
118 |
+
# baseball
|
119 |
+
'COCO_val2014_000000127510',
|
120 |
+
# bath
|
121 |
+
'COCO_val2014_000000324266',
|
122 |
+
# frisbee
|
123 |
+
'COCO_val2014_000000200717'
|
124 |
+
]
|
125 |
+
|
126 |
+
test_questions_for_images = [
|
127 |
+
################## paper samples
|
128 |
+
# giraffe
|
129 |
+
"is the animal eating?",
|
130 |
+
# baseball
|
131 |
+
"did he catch the ball?",
|
132 |
+
# bath
|
133 |
+
"is the tub white ?",
|
134 |
+
# frisbee
|
135 |
+
"did the man just catch the frisbee?"
|
136 |
+
################## paper samples
|
137 |
+
]
|