AustingDong
commited on
Commit
·
369c141
1
Parent(s):
e788822
renew app, formatted visualization
Browse files- app-old.py +501 -0
- app.py +9 -9
- demo/visualization.py +524 -0
app-old.py
ADDED
@@ -0,0 +1,501 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import torch
|
3 |
+
from transformers import AutoConfig, AutoModelForCausalLM
|
4 |
+
from janus.models import MultiModalityCausalLM, VLChatProcessor
|
5 |
+
from janus.utils.io import load_pil_images
|
6 |
+
from demo.cam import generate_gradcam, AttentionGuidedCAMJanus, AttentionGuidedCAMClip, AttentionGuidedCAMChartGemma, AttentionGuidedCAMLLaVA
|
7 |
+
from demo.model_utils import Clip_Utils, Janus_Utils, LLaVA_Utils, ChartGemma_Utils, add_title_to_image
|
8 |
+
|
9 |
+
import numpy as np
|
10 |
+
import matplotlib.pyplot as plt
|
11 |
+
import gc
|
12 |
+
import os
|
13 |
+
import spaces
|
14 |
+
from PIL import Image
|
15 |
+
|
16 |
+
def set_seed(model_seed = 42):
|
17 |
+
torch.manual_seed(model_seed)
|
18 |
+
np.random.seed(model_seed)
|
19 |
+
torch.cuda.manual_seed(model_seed) if torch.cuda.is_available() else None
|
20 |
+
|
21 |
+
set_seed()
|
22 |
+
clip_utils = Clip_Utils()
|
23 |
+
clip_utils.init_Clip()
|
24 |
+
model_utils, vl_gpt, tokenizer = None, None, None
|
25 |
+
model_name = "Clip"
|
26 |
+
language_model_max_layer = 24
|
27 |
+
language_model_best_layer = 8
|
28 |
+
vision_model_best_layer = 24
|
29 |
+
|
30 |
+
def clean():
|
31 |
+
global model_utils, vl_gpt, tokenizer, clip_utils
|
32 |
+
# Move models to CPU first (prevents CUDA references)
|
33 |
+
if 'vl_gpt' in globals() and vl_gpt is not None:
|
34 |
+
vl_gpt.to("cpu")
|
35 |
+
if 'clip_utils' in globals() and clip_utils is not None:
|
36 |
+
del clip_utils
|
37 |
+
|
38 |
+
# Delete all references
|
39 |
+
del model_utils, vl_gpt, tokenizer
|
40 |
+
model_utils, vl_gpt, tokenizer, clip_utils = None, None, None, None
|
41 |
+
gc.collect()
|
42 |
+
|
43 |
+
# Empty CUDA cache
|
44 |
+
if torch.cuda.is_available():
|
45 |
+
torch.cuda.empty_cache()
|
46 |
+
torch.cuda.ipc_collect() # Frees inter-process CUDA memory
|
47 |
+
|
48 |
+
# Empty MacOS Metal backend (if using Apple Silicon)
|
49 |
+
if torch.backends.mps.is_available():
|
50 |
+
torch.mps.empty_cache()
|
51 |
+
|
52 |
+
# Multimodal Understanding function
|
53 |
+
@spaces.GPU(duration=120)
|
54 |
+
def multimodal_understanding(model_type,
|
55 |
+
activation_map_method,
|
56 |
+
visual_pooling_method,
|
57 |
+
image, question, seed, top_p, temperature, target_token_idx,
|
58 |
+
visualization_layer_min, visualization_layer_max, focus, response_type, chart_type):
|
59 |
+
# Clear CUDA cache before generating
|
60 |
+
gc.collect()
|
61 |
+
if torch.cuda.is_available():
|
62 |
+
torch.cuda.empty_cache()
|
63 |
+
torch.cuda.ipc_collect()
|
64 |
+
|
65 |
+
# set seed
|
66 |
+
torch.manual_seed(seed)
|
67 |
+
np.random.seed(seed)
|
68 |
+
torch.cuda.manual_seed(seed) if torch.cuda.is_available() else None
|
69 |
+
|
70 |
+
input_text_decoded = ""
|
71 |
+
answer = ""
|
72 |
+
if model_name == "Clip":
|
73 |
+
|
74 |
+
inputs = clip_utils.prepare_inputs([question], image)
|
75 |
+
|
76 |
+
|
77 |
+
if activation_map_method == "GradCAM":
|
78 |
+
# Generate Grad-CAM
|
79 |
+
all_layers = [layer.layer_norm1 for layer in clip_utils.model.vision_model.encoder.layers]
|
80 |
+
|
81 |
+
if visualization_layer_min != visualization_layer_max:
|
82 |
+
target_layers = all_layers[visualization_layer_min-1 : visualization_layer_max-1]
|
83 |
+
else:
|
84 |
+
target_layers = [all_layers[visualization_layer_min-1]]
|
85 |
+
grad_cam = AttentionGuidedCAMClip(clip_utils.model, target_layers)
|
86 |
+
cam, outputs, grid_size = grad_cam.generate_cam(inputs, class_idx=0, visual_pooling_method=visual_pooling_method)
|
87 |
+
cam = cam.to("cpu")
|
88 |
+
cam = [generate_gradcam(cam, image, size=(224, 224))]
|
89 |
+
grad_cam.remove_hooks()
|
90 |
+
target_token_decoded = ""
|
91 |
+
|
92 |
+
|
93 |
+
|
94 |
+
else:
|
95 |
+
|
96 |
+
for param in vl_gpt.parameters():
|
97 |
+
param.requires_grad = True
|
98 |
+
|
99 |
+
|
100 |
+
prepare_inputs = model_utils.prepare_inputs(question, image)
|
101 |
+
|
102 |
+
if response_type == "answer + visualization":
|
103 |
+
if model_name.split('-')[0] == "Janus":
|
104 |
+
inputs_embeds = model_utils.generate_inputs_embeddings(prepare_inputs)
|
105 |
+
outputs = model_utils.generate_outputs(inputs_embeds, prepare_inputs, temperature, top_p)
|
106 |
+
else:
|
107 |
+
outputs = model_utils.generate_outputs(prepare_inputs, temperature, top_p)
|
108 |
+
|
109 |
+
sequences = outputs.sequences.cpu().tolist()
|
110 |
+
answer = tokenizer.decode(sequences[0], skip_special_tokens=True)
|
111 |
+
attention_raw = outputs.attentions
|
112 |
+
print("answer generated")
|
113 |
+
|
114 |
+
input_ids = prepare_inputs.input_ids[0].cpu().tolist()
|
115 |
+
input_ids_decoded = [tokenizer.decode([input_ids[i]]) for i in range(len(input_ids))]
|
116 |
+
|
117 |
+
if activation_map_method == "GradCAM":
|
118 |
+
# target_layers = vl_gpt.vision_model.vision_tower.blocks
|
119 |
+
if focus == "Visual Encoder":
|
120 |
+
if model_name.split('-')[0] == "Janus":
|
121 |
+
all_layers = [block.norm1 for block in vl_gpt.vision_model.vision_tower.blocks]
|
122 |
+
else:
|
123 |
+
all_layers = [block.layer_norm1 for block in vl_gpt.vision_tower.vision_model.encoder.layers]
|
124 |
+
else:
|
125 |
+
all_layers = [layer.self_attn for layer in vl_gpt.language_model.model.layers]
|
126 |
+
|
127 |
+
print("layer values:", visualization_layer_min, visualization_layer_max)
|
128 |
+
if visualization_layer_min != visualization_layer_max:
|
129 |
+
print("multi layers")
|
130 |
+
target_layers = all_layers[visualization_layer_min-1 : visualization_layer_max]
|
131 |
+
else:
|
132 |
+
print("single layer")
|
133 |
+
target_layers = [all_layers[visualization_layer_min-1]]
|
134 |
+
|
135 |
+
|
136 |
+
if model_name.split('-')[0] == "Janus":
|
137 |
+
gradcam = AttentionGuidedCAMJanus(vl_gpt, target_layers)
|
138 |
+
elif model_name.split('-')[0] == "LLaVA":
|
139 |
+
gradcam = AttentionGuidedCAMLLaVA(vl_gpt, target_layers)
|
140 |
+
elif model_name.split('-')[0] == "ChartGemma":
|
141 |
+
gradcam = AttentionGuidedCAMChartGemma(vl_gpt, target_layers)
|
142 |
+
|
143 |
+
start = 0
|
144 |
+
cam = []
|
145 |
+
if focus == "Visual Encoder":
|
146 |
+
if target_token_idx != -1:
|
147 |
+
cam_tensors, grid_size, start = gradcam.generate_cam(prepare_inputs, tokenizer, temperature, top_p, target_token_idx, visual_pooling_method, focus)
|
148 |
+
cam_grid = cam_tensors.reshape(grid_size, grid_size)
|
149 |
+
cam_i = generate_gradcam(cam_grid, image)
|
150 |
+
cam_i = add_title_to_image(cam_i, input_ids_decoded[start + target_token_idx])
|
151 |
+
cam = [cam_i]
|
152 |
+
else:
|
153 |
+
i = 0
|
154 |
+
cam = []
|
155 |
+
while start + i < len(input_ids_decoded):
|
156 |
+
if model_name.split('-')[0] == "Janus":
|
157 |
+
gradcam = AttentionGuidedCAMJanus(vl_gpt, target_layers)
|
158 |
+
elif model_name.split('-')[0] == "LLaVA":
|
159 |
+
gradcam = AttentionGuidedCAMLLaVA(vl_gpt, target_layers)
|
160 |
+
elif model_name.split('-')[0] == "ChartGemma":
|
161 |
+
gradcam = AttentionGuidedCAMChartGemma(vl_gpt, target_layers)
|
162 |
+
cam_tensors, grid_size, start = gradcam.generate_cam(prepare_inputs, tokenizer, temperature, top_p, i, visual_pooling_method, focus)
|
163 |
+
cam_grid = cam_tensors.reshape(grid_size, grid_size)
|
164 |
+
cam_i = generate_gradcam(cam_grid, image)
|
165 |
+
cam_i = add_title_to_image(cam_i, input_ids_decoded[start + i])
|
166 |
+
cam.append(cam_i)
|
167 |
+
gradcam.remove_hooks()
|
168 |
+
i += 1
|
169 |
+
else:
|
170 |
+
cam_tensors, grid_size, start = gradcam.generate_cam(prepare_inputs, tokenizer, temperature, top_p, target_token_idx, visual_pooling_method, focus)
|
171 |
+
if target_token_idx != -1:
|
172 |
+
input_text_decoded = input_ids_decoded[start + target_token_idx]
|
173 |
+
for i, cam_tensor in enumerate(cam_tensors):
|
174 |
+
if i == target_token_idx:
|
175 |
+
cam_grid = cam_tensor.reshape(grid_size, grid_size)
|
176 |
+
cam_i = generate_gradcam(cam_grid, image)
|
177 |
+
cam = [add_title_to_image(cam_i, input_text_decoded)]
|
178 |
+
break
|
179 |
+
else:
|
180 |
+
cam = []
|
181 |
+
for i, cam_tensor in enumerate(cam_tensors):
|
182 |
+
cam_grid = cam_tensor.reshape(grid_size, grid_size)
|
183 |
+
cam_i = generate_gradcam(cam_grid, image)
|
184 |
+
cam_i = add_title_to_image(cam_i, input_ids_decoded[start + i])
|
185 |
+
|
186 |
+
cam.append(cam_i)
|
187 |
+
|
188 |
+
gradcam.remove_hooks()
|
189 |
+
|
190 |
+
|
191 |
+
# Collect Results
|
192 |
+
RESULTS_ROOT = "./results"
|
193 |
+
FILES_ROOT = f"{RESULTS_ROOT}/{model_name}/{focus}/{chart_type}/layer{visualization_layer_min}-{visualization_layer_max}"
|
194 |
+
os.makedirs(FILES_ROOT, exist_ok=True)
|
195 |
+
if focus == "Visual Encoder":
|
196 |
+
cam[0].save(f"{FILES_ROOT}/{visual_pooling_method}.png")
|
197 |
+
else:
|
198 |
+
for i, cam_p in enumerate(cam):
|
199 |
+
cam_p.save(f"{FILES_ROOT}/{i}.png")
|
200 |
+
|
201 |
+
with open(f"{FILES_ROOT}/input_text_decoded.txt", "w") as f:
|
202 |
+
f.write(input_text_decoded)
|
203 |
+
f.close()
|
204 |
+
|
205 |
+
with open(f"{FILES_ROOT}/answer.txt", "w") as f:
|
206 |
+
f.write(answer)
|
207 |
+
f.close()
|
208 |
+
|
209 |
+
|
210 |
+
|
211 |
+
return answer, cam, input_text_decoded
|
212 |
+
|
213 |
+
|
214 |
+
|
215 |
+
|
216 |
+
# Gradio interface
|
217 |
+
|
218 |
+
def model_slider_change(model_type):
|
219 |
+
global model_utils, vl_gpt, tokenizer, clip_utils, model_name, language_model_max_layer, language_model_best_layer, vision_model_best_layer
|
220 |
+
model_name = model_type
|
221 |
+
if model_type == "Clip":
|
222 |
+
clean()
|
223 |
+
set_seed()
|
224 |
+
clip_utils = Clip_Utils()
|
225 |
+
clip_utils.init_Clip()
|
226 |
+
res = (
|
227 |
+
gr.Dropdown(choices=["Visualization only"], value="Visualization only", label="response_type"),
|
228 |
+
gr.Slider(minimum=1, maximum=12, value=12, step=1, label="visualization layers min"),
|
229 |
+
gr.Slider(minimum=1, maximum=12, value=12, step=1, label="visualization layers max"),
|
230 |
+
gr.Dropdown(choices=["Visual Encoder"], value="Visual Encoder", label="focus"),
|
231 |
+
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type")
|
232 |
+
)
|
233 |
+
return res
|
234 |
+
elif model_type.split('-')[0] == "Janus":
|
235 |
+
|
236 |
+
clean()
|
237 |
+
set_seed()
|
238 |
+
model_utils = Janus_Utils()
|
239 |
+
vl_gpt, tokenizer = model_utils.init_Janus(model_type.split('-')[-1])
|
240 |
+
language_model_max_layer = 24
|
241 |
+
language_model_best_layer = 8
|
242 |
+
|
243 |
+
res = (
|
244 |
+
gr.Dropdown(choices=["Visualization only", "answer + visualization"], value="answer + visualization", label="response_type"),
|
245 |
+
gr.Slider(minimum=1, maximum=24, value=24, step=1, label="visualization layers min"),
|
246 |
+
gr.Slider(minimum=1, maximum=24, value=24, step=1, label="visualization layers max"),
|
247 |
+
gr.Dropdown(choices=["Visual Encoder", "Language Model"], value="Visual Encoder", label="focus"),
|
248 |
+
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type")
|
249 |
+
)
|
250 |
+
return res
|
251 |
+
|
252 |
+
elif model_type.split('-')[0] == "LLaVA":
|
253 |
+
|
254 |
+
clean()
|
255 |
+
set_seed()
|
256 |
+
model_utils = LLaVA_Utils()
|
257 |
+
version = model_type.split('-')[1]
|
258 |
+
vl_gpt, tokenizer = model_utils.init_LLaVA(version=version)
|
259 |
+
language_model_max_layer = 32 if version == "1.5" else 28
|
260 |
+
language_model_best_layer = 10
|
261 |
+
|
262 |
+
res = (
|
263 |
+
gr.Dropdown(choices=["Visualization only", "answer + visualization"], value="answer + visualization", label="response_type"),
|
264 |
+
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer, step=1, label="visualization layers min"),
|
265 |
+
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer, step=1, label="visualization layers max"),
|
266 |
+
gr.Dropdown(choices=["Language Model"], value="Language Model", label="focus"),
|
267 |
+
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type")
|
268 |
+
)
|
269 |
+
return res
|
270 |
+
|
271 |
+
elif model_type.split('-')[0] == "ChartGemma":
|
272 |
+
clean()
|
273 |
+
set_seed()
|
274 |
+
model_utils = ChartGemma_Utils()
|
275 |
+
vl_gpt, tokenizer = model_utils.init_ChartGemma()
|
276 |
+
language_model_max_layer = 18
|
277 |
+
vision_model_best_layer = 19
|
278 |
+
language_model_best_layer = 15
|
279 |
+
|
280 |
+
res = (
|
281 |
+
gr.Dropdown(choices=["Visualization only", "answer + visualization"], value="answer + visualization", label="response_type"),
|
282 |
+
gr.Slider(minimum=1, maximum=language_model_best_layer, value=language_model_best_layer, step=1, label="visualization layers min"),
|
283 |
+
gr.Slider(minimum=1, maximum=language_model_best_layer, value=language_model_best_layer, step=1, label="visualization layers max"),
|
284 |
+
gr.Dropdown(choices=["Visual Encoder", "Language Model"], value="Language Model", label="focus"),
|
285 |
+
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type")
|
286 |
+
)
|
287 |
+
return res
|
288 |
+
|
289 |
+
|
290 |
+
|
291 |
+
|
292 |
+
def focus_change(focus):
|
293 |
+
global model_name, language_model_max_layer
|
294 |
+
if model_name == "Clip":
|
295 |
+
res = (
|
296 |
+
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type"),
|
297 |
+
gr.Slider(minimum=1, maximum=12, value=12, step=1, label="visualization layers min"),
|
298 |
+
gr.Slider(minimum=1, maximum=12, value=12, step=1, label="visualization layers max")
|
299 |
+
)
|
300 |
+
return res
|
301 |
+
|
302 |
+
if focus == "Language Model":
|
303 |
+
if response_type.value == "answer + visualization":
|
304 |
+
res = (
|
305 |
+
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type"),
|
306 |
+
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer, step=1, label="visualization layers min"),
|
307 |
+
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer, step=1, label="visualization layers max")
|
308 |
+
)
|
309 |
+
return res
|
310 |
+
else:
|
311 |
+
res = (
|
312 |
+
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type"),
|
313 |
+
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer, step=1, label="visualization layers min"),
|
314 |
+
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer, step=1, label="visualization layers max")
|
315 |
+
)
|
316 |
+
return res
|
317 |
+
|
318 |
+
else:
|
319 |
+
if model_name.split('-')[0] == "ChartGemma":
|
320 |
+
res = (
|
321 |
+
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type"),
|
322 |
+
gr.Slider(minimum=1, maximum=26, value=vision_model_best_layer, step=1, label="visualization layers min"),
|
323 |
+
gr.Slider(minimum=1, maximum=26, value=vision_model_best_layer, step=1, label="visualization layers max")
|
324 |
+
)
|
325 |
+
return res
|
326 |
+
|
327 |
+
else:
|
328 |
+
res = (
|
329 |
+
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type"),
|
330 |
+
gr.Slider(minimum=1, maximum=24, value=24, step=1, label="visualization layers min"),
|
331 |
+
gr.Slider(minimum=1, maximum=24, value=24, step=1, label="visualization layers max")
|
332 |
+
)
|
333 |
+
return res
|
334 |
+
|
335 |
+
|
336 |
+
|
337 |
+
|
338 |
+
|
339 |
+
with gr.Blocks() as demo:
|
340 |
+
gr.Markdown(value="# Multimodal Understanding")
|
341 |
+
|
342 |
+
with gr.Row():
|
343 |
+
image_input = gr.Image(height=500, label="Image")
|
344 |
+
activation_map_output = gr.Gallery(label="Visualization", height=500, columns=1, preview=True)
|
345 |
+
|
346 |
+
with gr.Row():
|
347 |
+
chart_type = gr.Textbox(label="Chart Type")
|
348 |
+
understanding_output = gr.Textbox(label="Answer")
|
349 |
+
|
350 |
+
with gr.Row():
|
351 |
+
|
352 |
+
with gr.Column():
|
353 |
+
model_selector = gr.Dropdown(choices=["Clip", "ChartGemma-3B", "Janus-Pro-1B", "Janus-Pro-7B", "LLaVA-1.5-7B"], value="Clip", label="model")
|
354 |
+
question_input = gr.Textbox(label="Input Prompt")
|
355 |
+
und_seed_input = gr.Number(label="Seed", precision=0, value=42)
|
356 |
+
top_p = gr.Slider(minimum=0, maximum=1, value=0.95, step=0.05, label="top_p")
|
357 |
+
temperature = gr.Slider(minimum=0, maximum=1, value=0.1, step=0.05, label="temperature")
|
358 |
+
target_token_idx = gr.Number(label="target_token_idx (-1 means all)", precision=0, value=-1)
|
359 |
+
|
360 |
+
|
361 |
+
with gr.Column():
|
362 |
+
response_type = gr.Dropdown(choices=["Visualization only"], value="Visualization only", label="response_type")
|
363 |
+
focus = gr.Dropdown(choices=["Visual Encoder"], value="Visual Encoder", label="focus")
|
364 |
+
activation_map_method = gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="visualization type")
|
365 |
+
visual_pooling_method = gr.Dropdown(choices=["CLS", "max", "avg"], value="CLS", label="visual pooling method")
|
366 |
+
|
367 |
+
|
368 |
+
visualization_layers_min = gr.Slider(minimum=1, maximum=12, value=12, step=1, label="visualization layers min")
|
369 |
+
visualization_layers_max = gr.Slider(minimum=1, maximum=12, value=12, step=1, label="visualization layers max")
|
370 |
+
|
371 |
+
|
372 |
+
|
373 |
+
|
374 |
+
|
375 |
+
model_selector.change(
|
376 |
+
fn=model_slider_change,
|
377 |
+
inputs=model_selector,
|
378 |
+
outputs=[
|
379 |
+
response_type,
|
380 |
+
visualization_layers_min,
|
381 |
+
visualization_layers_max,
|
382 |
+
focus,
|
383 |
+
activation_map_method
|
384 |
+
]
|
385 |
+
)
|
386 |
+
|
387 |
+
focus.change(
|
388 |
+
fn = focus_change,
|
389 |
+
inputs = focus,
|
390 |
+
outputs=[
|
391 |
+
activation_map_method,
|
392 |
+
visualization_layers_min,
|
393 |
+
visualization_layers_max,
|
394 |
+
]
|
395 |
+
)
|
396 |
+
|
397 |
+
# response_type.change(
|
398 |
+
# fn = response_type_change,
|
399 |
+
# inputs = response_type,
|
400 |
+
# outputs = [activation_map_method]
|
401 |
+
# )
|
402 |
+
|
403 |
+
|
404 |
+
|
405 |
+
understanding_button = gr.Button("Submit")
|
406 |
+
|
407 |
+
understanding_target_token_decoded_output = gr.Textbox(label="Target Token Decoded")
|
408 |
+
|
409 |
+
|
410 |
+
examples_inpainting = gr.Examples(
|
411 |
+
label="Multimodal Understanding examples",
|
412 |
+
examples=[
|
413 |
+
|
414 |
+
[
|
415 |
+
"LineChart",
|
416 |
+
"What was the price of a barrel of oil in February 2020?",
|
417 |
+
"images/LineChart.png"
|
418 |
+
],
|
419 |
+
|
420 |
+
[
|
421 |
+
"BarChart",
|
422 |
+
"What is the average internet speed in Japan?",
|
423 |
+
"images/BarChart.png"
|
424 |
+
],
|
425 |
+
|
426 |
+
[
|
427 |
+
"StackedBar",
|
428 |
+
"What is the cost of peanuts in Seoul?",
|
429 |
+
"images/StackedBar.png"
|
430 |
+
],
|
431 |
+
|
432 |
+
[
|
433 |
+
"100%StackedBar",
|
434 |
+
"Which country has the lowest proportion of Gold medals?",
|
435 |
+
"images/Stacked100.png"
|
436 |
+
],
|
437 |
+
|
438 |
+
[
|
439 |
+
"PieChart",
|
440 |
+
"What is the approximate global smartphone market share of Samsung?",
|
441 |
+
"images/PieChart.png"
|
442 |
+
],
|
443 |
+
|
444 |
+
[
|
445 |
+
"Histogram",
|
446 |
+
"What distance have customers traveled in the taxi the most?",
|
447 |
+
"images/Histogram.png"
|
448 |
+
],
|
449 |
+
|
450 |
+
[
|
451 |
+
"Scatterplot",
|
452 |
+
"True/False: There is a negative linear relationship between the height and the weight of the 85 males.",
|
453 |
+
"images/Scatterplot.png"
|
454 |
+
],
|
455 |
+
|
456 |
+
[
|
457 |
+
"AreaChart",
|
458 |
+
"What was the average price of pount of coffee beans in October 2019?",
|
459 |
+
"images/AreaChart.png"
|
460 |
+
],
|
461 |
+
|
462 |
+
[
|
463 |
+
"StackedArea",
|
464 |
+
"What was the ratio of girls named 'Isla' to girls named 'Amelia' in 2012 in the UK?",
|
465 |
+
"images/StackedArea.png"
|
466 |
+
],
|
467 |
+
|
468 |
+
[
|
469 |
+
"BubbleChart",
|
470 |
+
"Which city's metro system has the largest number of stations?",
|
471 |
+
"images/BubbleChart.png"
|
472 |
+
],
|
473 |
+
|
474 |
+
[
|
475 |
+
"Choropleth",
|
476 |
+
"True/False: In 2020, the unemployment rate for Washington (WA) was higher than that of Wisconsin (WI).",
|
477 |
+
"images/Choropleth_New.png"
|
478 |
+
],
|
479 |
+
|
480 |
+
[
|
481 |
+
"TreeMap",
|
482 |
+
"True/False: eBay is nested in the Software category.",
|
483 |
+
"images/TreeMap.png"
|
484 |
+
]
|
485 |
+
|
486 |
+
],
|
487 |
+
inputs=[chart_type, question_input, image_input],
|
488 |
+
)
|
489 |
+
|
490 |
+
|
491 |
+
|
492 |
+
|
493 |
+
understanding_button.click(
|
494 |
+
multimodal_understanding,
|
495 |
+
inputs=[model_selector, activation_map_method, visual_pooling_method, image_input, question_input, und_seed_input, top_p, temperature, target_token_idx,
|
496 |
+
visualization_layers_min, visualization_layers_max, focus, response_type, chart_type],
|
497 |
+
outputs=[understanding_output, activation_map_output, understanding_target_token_decoded_output]
|
498 |
+
)
|
499 |
+
|
500 |
+
demo.launch(share=True)
|
501 |
+
# demo.queue(concurrency_count=1, max_size=10).launch(server_name="0.0.0.0", server_port=37906, root_path="/path")
|
app.py
CHANGED
@@ -3,7 +3,7 @@ import torch
|
|
3 |
from transformers import AutoConfig, AutoModelForCausalLM
|
4 |
from janus.models import MultiModalityCausalLM, VLChatProcessor
|
5 |
from janus.utils.io import load_pil_images
|
6 |
-
from demo.
|
7 |
from demo.model_utils import Clip_Utils, Janus_Utils, LLaVA_Utils, ChartGemma_Utils, add_title_to_image
|
8 |
|
9 |
import numpy as np
|
@@ -82,8 +82,8 @@ def multimodal_understanding(model_type,
|
|
82 |
target_layers = all_layers[visualization_layer_min-1 : visualization_layer_max-1]
|
83 |
else:
|
84 |
target_layers = [all_layers[visualization_layer_min-1]]
|
85 |
-
grad_cam =
|
86 |
-
cam, outputs, grid_size = grad_cam.generate_cam(inputs,
|
87 |
cam = cam.to("cpu")
|
88 |
cam = [generate_gradcam(cam, image, size=(224, 224))]
|
89 |
grad_cam.remove_hooks()
|
@@ -134,11 +134,11 @@ def multimodal_understanding(model_type,
|
|
134 |
|
135 |
|
136 |
if model_name.split('-')[0] == "Janus":
|
137 |
-
gradcam =
|
138 |
elif model_name.split('-')[0] == "LLaVA":
|
139 |
-
gradcam =
|
140 |
elif model_name.split('-')[0] == "ChartGemma":
|
141 |
-
gradcam =
|
142 |
|
143 |
start = 0
|
144 |
cam = []
|
@@ -154,11 +154,11 @@ def multimodal_understanding(model_type,
|
|
154 |
cam = []
|
155 |
while start + i < len(input_ids_decoded):
|
156 |
if model_name.split('-')[0] == "Janus":
|
157 |
-
gradcam =
|
158 |
elif model_name.split('-')[0] == "LLaVA":
|
159 |
-
gradcam =
|
160 |
elif model_name.split('-')[0] == "ChartGemma":
|
161 |
-
gradcam =
|
162 |
cam_tensors, grid_size, start = gradcam.generate_cam(prepare_inputs, tokenizer, temperature, top_p, i, visual_pooling_method, focus)
|
163 |
cam_grid = cam_tensors.reshape(grid_size, grid_size)
|
164 |
cam_i = generate_gradcam(cam_grid, image)
|
|
|
3 |
from transformers import AutoConfig, AutoModelForCausalLM
|
4 |
from janus.models import MultiModalityCausalLM, VLChatProcessor
|
5 |
from janus.utils.io import load_pil_images
|
6 |
+
from demo.visualization import generate_gradcam, VisualizationJanus, VisualizationClip, VisualizationChartGemma, VisualizationLLaVA
|
7 |
from demo.model_utils import Clip_Utils, Janus_Utils, LLaVA_Utils, ChartGemma_Utils, add_title_to_image
|
8 |
|
9 |
import numpy as np
|
|
|
82 |
target_layers = all_layers[visualization_layer_min-1 : visualization_layer_max-1]
|
83 |
else:
|
84 |
target_layers = [all_layers[visualization_layer_min-1]]
|
85 |
+
grad_cam = VisualizationClip(clip_utils.model, target_layers)
|
86 |
+
cam, outputs, grid_size = grad_cam.generate_cam(inputs, target_token_idx=0, visual_pooling_method=visual_pooling_method)
|
87 |
cam = cam.to("cpu")
|
88 |
cam = [generate_gradcam(cam, image, size=(224, 224))]
|
89 |
grad_cam.remove_hooks()
|
|
|
134 |
|
135 |
|
136 |
if model_name.split('-')[0] == "Janus":
|
137 |
+
gradcam = VisualizationJanus(vl_gpt, target_layers)
|
138 |
elif model_name.split('-')[0] == "LLaVA":
|
139 |
+
gradcam = VisualizationLLaVA(vl_gpt, target_layers)
|
140 |
elif model_name.split('-')[0] == "ChartGemma":
|
141 |
+
gradcam = VisualizationChartGemma(vl_gpt, target_layers)
|
142 |
|
143 |
start = 0
|
144 |
cam = []
|
|
|
154 |
cam = []
|
155 |
while start + i < len(input_ids_decoded):
|
156 |
if model_name.split('-')[0] == "Janus":
|
157 |
+
gradcam = VisualizationJanus(vl_gpt, target_layers)
|
158 |
elif model_name.split('-')[0] == "LLaVA":
|
159 |
+
gradcam = VisualizationLLaVA(vl_gpt, target_layers)
|
160 |
elif model_name.split('-')[0] == "ChartGemma":
|
161 |
+
gradcam = VisualizationChartGemma(vl_gpt, target_layers)
|
162 |
cam_tensors, grid_size, start = gradcam.generate_cam(prepare_inputs, tokenizer, temperature, top_p, i, visual_pooling_method, focus)
|
163 |
cam_grid = cam_tensors.reshape(grid_size, grid_size)
|
164 |
cam_i = generate_gradcam(cam_grid, image)
|
demo/visualization.py
ADDED
@@ -0,0 +1,524 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import cv2
|
2 |
+
import numpy as np
|
3 |
+
import types
|
4 |
+
import torch
|
5 |
+
import torch.nn.functional as F
|
6 |
+
import matplotlib.pyplot as plt
|
7 |
+
from PIL import Image
|
8 |
+
from torch import nn
|
9 |
+
import spaces
|
10 |
+
from demo.modify_llama import *
|
11 |
+
|
12 |
+
|
13 |
+
class Visualization:
|
14 |
+
def __init__(self, model, register=True):
|
15 |
+
self.model = model
|
16 |
+
self.gradients = []
|
17 |
+
self.activations = []
|
18 |
+
self.hooks = []
|
19 |
+
if register:
|
20 |
+
self._register_hooks()
|
21 |
+
|
22 |
+
def _register_hooks(self):
|
23 |
+
for layer in self.target_layers:
|
24 |
+
self.hooks.append(layer.register_forward_hook(self._forward_hook))
|
25 |
+
self.hooks.append(layer.register_backward_hook(self._backward_hook))
|
26 |
+
|
27 |
+
def _forward_hook(self, module, input, output):
|
28 |
+
self.activations.append(output)
|
29 |
+
|
30 |
+
def _backward_hook(self, module, grad_in, grad_out):
|
31 |
+
self.gradients.append(grad_out[0])
|
32 |
+
|
33 |
+
def _modify_layers(self):
|
34 |
+
for layer in self.target_layers:
|
35 |
+
setattr(layer, "attn_gradients", None)
|
36 |
+
setattr(layer, "attention_map", None)
|
37 |
+
|
38 |
+
layer.save_attn_gradients = types.MethodType(save_attn_gradients, layer)
|
39 |
+
layer.get_attn_gradients = types.MethodType(get_attn_gradients, layer)
|
40 |
+
layer.save_attn_map = types.MethodType(save_attn_map, layer)
|
41 |
+
layer.get_attn_map = types.MethodType(get_attn_map, layer)
|
42 |
+
|
43 |
+
def _forward_activate_hooks(self, module, input, output):
|
44 |
+
attn_output, attn_weights = output # Unpack outputs
|
45 |
+
print("attn_output shape:", attn_output.shape)
|
46 |
+
print("attn_weights shape:", attn_weights.shape)
|
47 |
+
module.save_attn_map(attn_weights)
|
48 |
+
attn_weights.register_hook(module.save_attn_gradients)
|
49 |
+
|
50 |
+
def _register_hooks_activations(self):
|
51 |
+
for layer in self.target_layers:
|
52 |
+
if hasattr(layer, "q_proj"): # is an attention layer
|
53 |
+
self.hooks.append(layer.register_forward_hook(self._forward_activate_hooks))
|
54 |
+
|
55 |
+
|
56 |
+
def remove_hooks(self):
|
57 |
+
for hook in self.hooks:
|
58 |
+
hook.remove()
|
59 |
+
|
60 |
+
def setup_grads(self):
|
61 |
+
torch.autograd.set_detect_anomaly(True)
|
62 |
+
for param in self.model.parameters():
|
63 |
+
param.requires_grad = False
|
64 |
+
|
65 |
+
for layer in self.target_layers:
|
66 |
+
for param in layer.parameters():
|
67 |
+
param.requires_grad = True
|
68 |
+
|
69 |
+
def forward_backward(self):
|
70 |
+
raise NotImplementedError
|
71 |
+
|
72 |
+
def grad_cam_vis(self):
|
73 |
+
self.model.zero_grad()
|
74 |
+
cam_sum = None
|
75 |
+
for act, grad in zip(self.activations, self.gradients):
|
76 |
+
|
77 |
+
act = F.relu(act[0])
|
78 |
+
|
79 |
+
grad_weights = grad.mean(dim=-1, keepdim=True)
|
80 |
+
|
81 |
+
print("act shape", act.shape)
|
82 |
+
print("grad_weights shape", grad_weights.shape)
|
83 |
+
|
84 |
+
# cam = (act * grad_weights).sum(dim=-1)
|
85 |
+
cam, _ = (act * grad_weights).max(dim=-1)
|
86 |
+
|
87 |
+
print("cam_shape: ", cam.shape)
|
88 |
+
|
89 |
+
# Sum across all layers
|
90 |
+
if cam_sum is None:
|
91 |
+
cam_sum = cam
|
92 |
+
else:
|
93 |
+
cam_sum += cam
|
94 |
+
|
95 |
+
cam_sum = F.relu(cam_sum)
|
96 |
+
return cam_sum
|
97 |
+
|
98 |
+
|
99 |
+
|
100 |
+
def grad_cam_llm(self, mean_inside=False):
|
101 |
+
|
102 |
+
cam_sum = None
|
103 |
+
for act, grad in zip(self.activations, self.gradients):
|
104 |
+
|
105 |
+
if mean_inside:
|
106 |
+
act = act.mean(dim=1)
|
107 |
+
grad = F.relu(grad.mean(dim=1))
|
108 |
+
cam = act * grad
|
109 |
+
else:
|
110 |
+
cam = act * grad
|
111 |
+
cam = act * grad.sum(dim=1)
|
112 |
+
|
113 |
+
print(cam.shape)
|
114 |
+
|
115 |
+
# Sum across all layers
|
116 |
+
if cam_sum is None:
|
117 |
+
cam_sum = cam
|
118 |
+
else:
|
119 |
+
cam_sum += cam
|
120 |
+
|
121 |
+
cam_sum = F.relu(cam_sum)
|
122 |
+
return cam_sum
|
123 |
+
|
124 |
+
def attention_map(self):
|
125 |
+
raise NotImplementedError
|
126 |
+
|
127 |
+
def attn_guided_cam(self):
|
128 |
+
|
129 |
+
cams = []
|
130 |
+
for act, grad in zip(self.activations, self.gradients):
|
131 |
+
print("act shape", act.shape)
|
132 |
+
print("grad shape", grad.shape)
|
133 |
+
|
134 |
+
grad = F.relu(grad)
|
135 |
+
|
136 |
+
# cam = grad
|
137 |
+
cam = act * grad # shape: [1, heads, seq_len, seq_len]
|
138 |
+
cam = cam.sum(dim=1) # shape: [1, seq_len, seq_len]
|
139 |
+
cam = cam.to(torch.float32).detach().cpu()
|
140 |
+
cams.append(cam)
|
141 |
+
return cams
|
142 |
+
|
143 |
+
|
144 |
+
def process(self, cam_sum, thresholding=True, remove_cls=True, normalize=True):
|
145 |
+
|
146 |
+
cam_sum = cam_sum.to(torch.float32)
|
147 |
+
|
148 |
+
# thresholding
|
149 |
+
if thresholding:
|
150 |
+
percentile = torch.quantile(cam_sum, 0.2) # Adjust threshold dynamically
|
151 |
+
cam_sum[cam_sum < percentile] = 0
|
152 |
+
|
153 |
+
# Remove CLS
|
154 |
+
if remove_cls:
|
155 |
+
cam_sum = cam_sum[0, 1:]
|
156 |
+
|
157 |
+
num_patches = cam_sum.shape[-1] # Last dimension of CAM output
|
158 |
+
grid_size = int(num_patches ** 0.5)
|
159 |
+
print(f"Detected grid size: {grid_size}x{grid_size}")
|
160 |
+
cam_sum = cam_sum.view(grid_size, grid_size).detach()
|
161 |
+
|
162 |
+
# Normalize
|
163 |
+
if normalize:
|
164 |
+
cam_sum = (cam_sum - cam_sum.min()) / (cam_sum.max() - cam_sum.min())
|
165 |
+
|
166 |
+
return cam_sum, grid_size
|
167 |
+
|
168 |
+
def process_multiple(self, cam_sum, start_idx, images_seq_mask, thresholding=True, normalize=True):
|
169 |
+
cam_sum = cam_sum.to(torch.float32)
|
170 |
+
# thresholding
|
171 |
+
if thresholding:
|
172 |
+
percentile = torch.quantile(cam_sum, 0.2) # Adjust threshold dynamically
|
173 |
+
cam_sum[cam_sum < percentile] = 0
|
174 |
+
|
175 |
+
|
176 |
+
# cam_sum shape: [1, seq_len, seq_len]
|
177 |
+
cam_sum_lst = []
|
178 |
+
cam_sum_raw = cam_sum
|
179 |
+
start = start_idx
|
180 |
+
for i in range(start, cam_sum_raw.shape[1]):
|
181 |
+
cam_sum = cam_sum_raw[:, i, :] # shape: [1: seq_len]
|
182 |
+
cam_sum = cam_sum[images_seq_mask].unsqueeze(0) # shape: [1, img_seq_len]
|
183 |
+
print("cam_sum shape: ", cam_sum.shape)
|
184 |
+
num_patches = cam_sum.shape[-1] # Last dimension of CAM output
|
185 |
+
grid_size = int(num_patches ** 0.5)
|
186 |
+
print(f"Detected grid size: {grid_size}x{grid_size}")
|
187 |
+
|
188 |
+
cam_sum = cam_sum.view(grid_size, grid_size)
|
189 |
+
if normalize:
|
190 |
+
cam_sum = (cam_sum - cam_sum.min()) / (cam_sum.max() - cam_sum.min())
|
191 |
+
cam_sum = cam_sum.detach().to("cpu")
|
192 |
+
cam_sum_lst.append(cam_sum)
|
193 |
+
return cam_sum_lst, grid_size
|
194 |
+
|
195 |
+
def process_multiple_withsum(self, cams, start_idx, images_seq_mask, normalize=False):
|
196 |
+
cam_sum_lst = []
|
197 |
+
for i in range(start_idx, cams[0].shape[1]):
|
198 |
+
cam_sum = None
|
199 |
+
for layer, cam_l in enumerate(cams):
|
200 |
+
cam_l_i = cam_l[0, i, :] # shape: [1: seq_len]
|
201 |
+
|
202 |
+
cam_l_i = cam_l_i[images_seq_mask].unsqueeze(0) # shape: [1, img_seq_len]
|
203 |
+
|
204 |
+
num_patches = cam_l_i.shape[-1] # Last dimension of CAM output
|
205 |
+
grid_size = int(num_patches ** 0.5)
|
206 |
+
# print(f"Detected grid size: {grid_size}x{grid_size}")
|
207 |
+
|
208 |
+
# Fix the reshaping step dynamically
|
209 |
+
cam_reshaped = cam_l_i.view(grid_size, grid_size)
|
210 |
+
|
211 |
+
if normalize:
|
212 |
+
cam_reshaped = (cam_reshaped - cam_reshaped.min()) / (cam_reshaped.max() - cam_reshaped.min())
|
213 |
+
if cam_sum == None:
|
214 |
+
cam_sum = cam_reshaped
|
215 |
+
else:
|
216 |
+
cam_sum += cam_reshaped
|
217 |
+
|
218 |
+
cam_sum = (cam_sum - cam_sum.min()) / (cam_sum.max() - cam_sum.min())
|
219 |
+
cam_sum_lst.append(cam_sum)
|
220 |
+
return cam_sum_lst, grid_size
|
221 |
+
|
222 |
+
def generate_cam(self, input_tensor, target_token_idx=None):
|
223 |
+
raise NotImplementedError
|
224 |
+
|
225 |
+
|
226 |
+
|
227 |
+
|
228 |
+
class VisualizationClip(Visualization):
|
229 |
+
def __init__(self, model, target_layers):
|
230 |
+
self.target_layers = target_layers
|
231 |
+
super().__init__(model)
|
232 |
+
|
233 |
+
@spaces.GPU(duration=120)
|
234 |
+
def forward_backward(self, input_tensor, visual_pooling_method, target_token_idx):
|
235 |
+
output_full = self.model(**input_tensor)
|
236 |
+
|
237 |
+
if target_token_idx is None:
|
238 |
+
target_token_idx = torch.argmax(output_full.logits, dim=1).item()
|
239 |
+
|
240 |
+
if visual_pooling_method == "CLS":
|
241 |
+
output = output_full.image_embeds
|
242 |
+
elif visual_pooling_method == "avg":
|
243 |
+
output = self.model.visual_projection(output_full.vision_model_output.last_hidden_state).mean(dim=1)
|
244 |
+
else:
|
245 |
+
output, _ = self.model.visual_projection(output_full.vision_model_output.last_hidden_state).max(dim=1)
|
246 |
+
|
247 |
+
|
248 |
+
output.backward(output_full.text_embeds[target_token_idx:target_token_idx+1], retain_graph=True)
|
249 |
+
return output_full
|
250 |
+
|
251 |
+
|
252 |
+
@spaces.GPU(duration=120)
|
253 |
+
def generate_cam(self, input_tensor, target_token_idx=None, visual_pooling_method="CLS"):
|
254 |
+
""" Generates Grad-CAM heatmap for ViT. """
|
255 |
+
self.setup_grads()
|
256 |
+
# Forward Backward pass
|
257 |
+
output_full = self.forward_backward(input_tensor, visual_pooling_method, target_token_idx)
|
258 |
+
|
259 |
+
cam_sum = self.grad_cam_vis()
|
260 |
+
cam_sum, grid_size = self.process(cam_sum)
|
261 |
+
|
262 |
+
return cam_sum, output_full, grid_size
|
263 |
+
|
264 |
+
|
265 |
+
|
266 |
+
|
267 |
+
|
268 |
+
|
269 |
+
|
270 |
+
|
271 |
+
|
272 |
+
|
273 |
+
|
274 |
+
|
275 |
+
|
276 |
+
|
277 |
+
|
278 |
+
|
279 |
+
|
280 |
+
|
281 |
+
|
282 |
+
|
283 |
+
|
284 |
+
|
285 |
+
|
286 |
+
|
287 |
+
class VisualizationJanus(Visualization):
|
288 |
+
def __init__(self, model, target_layers):
|
289 |
+
self.target_layers = target_layers
|
290 |
+
super().__init__(model)
|
291 |
+
self._modify_layers()
|
292 |
+
self._register_hooks_activations()
|
293 |
+
|
294 |
+
def forward_backward(self, input_tensor, tokenizer, temperature, top_p, target_token_idx=None, visual_pooling_method="CLS", focus="Visual Encoder"):
|
295 |
+
# Forward
|
296 |
+
image_embeddings, inputs_embeddings, outputs = self.model(input_tensor, tokenizer, temperature, top_p)
|
297 |
+
input_ids = input_tensor.input_ids
|
298 |
+
|
299 |
+
if focus == "Visual Encoder":
|
300 |
+
|
301 |
+
start_idx = 620
|
302 |
+
self.model.zero_grad()
|
303 |
+
|
304 |
+
loss = outputs.logits.max(dim=-1).values[0, start_idx + target_token_idx]
|
305 |
+
loss.backward()
|
306 |
+
|
307 |
+
elif focus == "Language Model":
|
308 |
+
self.model.zero_grad()
|
309 |
+
loss = outputs.logits.max(dim=-1).values.sum()
|
310 |
+
loss.backward()
|
311 |
+
|
312 |
+
self.activations = [layer.get_attn_map() for layer in self.target_layers]
|
313 |
+
self.gradients = [layer.get_attn_gradients() for layer in self.target_layers]
|
314 |
+
|
315 |
+
@spaces.GPU(duration=120)
|
316 |
+
def generate_cam(self, input_tensor, tokenizer, temperature, top_p, target_token_idx=None, visual_pooling_method="CLS", focus="Visual Encoder"):
|
317 |
+
|
318 |
+
self.setup_grads()
|
319 |
+
|
320 |
+
# Forward Backward pass
|
321 |
+
self.forward_backward(input_tensor, tokenizer, temperature, top_p, target_token_idx, visual_pooling_method, focus)
|
322 |
+
|
323 |
+
start_idx = 620
|
324 |
+
if focus == "Visual Encoder":
|
325 |
+
|
326 |
+
cam_sum = self.grad_cam_vis()
|
327 |
+
cam_sum, grid_size = self.process(cam_sum)
|
328 |
+
return cam_sum, grid_size, start_idx
|
329 |
+
|
330 |
+
elif focus == "Language Model":
|
331 |
+
|
332 |
+
cam_sum = self.grad_cam_llm(mean_inside=True)
|
333 |
+
|
334 |
+
images_seq_mask = input_tensor.images_seq_mask
|
335 |
+
|
336 |
+
cam_sum_lst, grid_size = self.process_multiple(cam_sum, start_idx, images_seq_mask)
|
337 |
+
|
338 |
+
return cam_sum_lst, grid_size, start_idx
|
339 |
+
|
340 |
+
|
341 |
+
|
342 |
+
|
343 |
+
|
344 |
+
|
345 |
+
|
346 |
+
|
347 |
+
|
348 |
+
class VisualizationLLaVA(Visualization):
|
349 |
+
def __init__(self, model, target_layers):
|
350 |
+
self.target_layers = target_layers
|
351 |
+
super().__init__(model, register=False)
|
352 |
+
self._modify_layers()
|
353 |
+
self._register_hooks_activations()
|
354 |
+
|
355 |
+
def forward_backward(self, inputs):
|
356 |
+
# Forward pass
|
357 |
+
outputs_raw = self.model(**inputs)
|
358 |
+
|
359 |
+
self.model.zero_grad()
|
360 |
+
print("outputs_raw", outputs_raw)
|
361 |
+
|
362 |
+
loss = outputs_raw.logits.max(dim=-1).values.sum()
|
363 |
+
loss.backward()
|
364 |
+
self.activations = [layer.get_attn_map() for layer in self.target_layers]
|
365 |
+
self.gradients = [layer.get_attn_gradients() for layer in self.target_layers]
|
366 |
+
|
367 |
+
@spaces.GPU(duration=120)
|
368 |
+
def generate_cam(self, inputs, tokenizer, temperature, top_p, target_token_idx=None, visual_pooling_method="CLS", focus="Visual Encoder"):
|
369 |
+
|
370 |
+
self.setup_grads()
|
371 |
+
self.forward_backward(inputs)
|
372 |
+
|
373 |
+
# get image masks
|
374 |
+
images_seq_mask = []
|
375 |
+
last = 0
|
376 |
+
for i in range(inputs["input_ids"].shape[1]):
|
377 |
+
decoded_token = tokenizer.decode(inputs["input_ids"][0][i].item())
|
378 |
+
if (decoded_token == "<image>"):
|
379 |
+
images_seq_mask.append(True)
|
380 |
+
last = i
|
381 |
+
else:
|
382 |
+
images_seq_mask.append(False)
|
383 |
+
|
384 |
+
|
385 |
+
# Aggregate activations and gradients from ALL layers
|
386 |
+
start_idx = last + 1
|
387 |
+
cams = self.attn_guided_cam()
|
388 |
+
cam_sum_lst, grid_size = self.process_multiple_withsum(cams, start_idx, images_seq_mask)
|
389 |
+
|
390 |
+
return cam_sum_lst, grid_size, start_idx
|
391 |
+
|
392 |
+
|
393 |
+
|
394 |
+
|
395 |
+
|
396 |
+
|
397 |
+
class VisualizationChartGemma(Visualization):
|
398 |
+
def __init__(self, model, target_layers):
|
399 |
+
self.target_layers = target_layers
|
400 |
+
super().__init__(model, register=True)
|
401 |
+
self._modify_layers()
|
402 |
+
self._register_hooks_activations()
|
403 |
+
|
404 |
+
def forward_backward(self, inputs, focus, start_idx, target_token_idx):
|
405 |
+
outputs_raw = self.model(**inputs, output_hidden_states=True)
|
406 |
+
if focus == "Visual Encoder":
|
407 |
+
|
408 |
+
self.model.zero_grad()
|
409 |
+
|
410 |
+
loss = outputs_raw.logits.max(dim=-1).values[0, start_idx + target_token_idx]
|
411 |
+
loss.backward()
|
412 |
+
|
413 |
+
elif focus == "Language Model":
|
414 |
+
self.model.zero_grad()
|
415 |
+
if target_token_idx == -1:
|
416 |
+
loss = outputs_raw.logits.max(dim=-1).values.sum()
|
417 |
+
else:
|
418 |
+
loss = outputs_raw.logits.max(dim=-1).values[0, start_idx + target_token_idx]
|
419 |
+
loss.backward()
|
420 |
+
self.activations = [layer.get_attn_map() for layer in self.target_layers]
|
421 |
+
self.gradients = [layer.get_attn_gradients() for layer in self.target_layers]
|
422 |
+
|
423 |
+
@spaces.GPU(duration=120)
|
424 |
+
def generate_cam(self, inputs, tokenizer, temperature, top_p, target_token_idx=None, visual_pooling_method="CLS", focus="Visual Encoder"):
|
425 |
+
|
426 |
+
# Forward pass
|
427 |
+
self.setup_grads()
|
428 |
+
|
429 |
+
# get image masks
|
430 |
+
images_seq_mask = []
|
431 |
+
last = 0
|
432 |
+
for i in range(inputs["input_ids"].shape[1]):
|
433 |
+
decoded_token = tokenizer.decode(inputs["input_ids"][0][i].item())
|
434 |
+
if (decoded_token == "<image>"):
|
435 |
+
images_seq_mask.append(True)
|
436 |
+
last = i
|
437 |
+
else:
|
438 |
+
images_seq_mask.append(False)
|
439 |
+
start_idx = last + 1
|
440 |
+
|
441 |
+
|
442 |
+
self.forward_backward(inputs, focus, start_idx, target_token_idx)
|
443 |
+
if focus == "Visual Encoder":
|
444 |
+
|
445 |
+
cam_sum = self.grad_cam_vis()
|
446 |
+
cam_sum, grid_size = self.process(cam_sum, remove_cls=False)
|
447 |
+
|
448 |
+
return cam_sum, grid_size, start_idx
|
449 |
+
|
450 |
+
elif focus == "Language Model":
|
451 |
+
|
452 |
+
cams = self.attn_guided_cam()
|
453 |
+
cam_sum_lst, grid_size = self.process_multiple_withsum(cams, start_idx, images_seq_mask)
|
454 |
+
|
455 |
+
# cams shape: [layers, 1, seq_len, seq_len]
|
456 |
+
|
457 |
+
|
458 |
+
|
459 |
+
return cam_sum_lst, grid_size, start_idx
|
460 |
+
|
461 |
+
|
462 |
+
|
463 |
+
|
464 |
+
|
465 |
+
|
466 |
+
|
467 |
+
|
468 |
+
|
469 |
+
|
470 |
+
def generate_gradcam(
|
471 |
+
cam,
|
472 |
+
image,
|
473 |
+
size = (384, 384),
|
474 |
+
alpha=0.5,
|
475 |
+
colormap=cv2.COLORMAP_JET,
|
476 |
+
aggregation='mean',
|
477 |
+
normalize=False
|
478 |
+
):
|
479 |
+
"""
|
480 |
+
Generates a Grad-CAM heatmap overlay on top of the input image.
|
481 |
+
|
482 |
+
Parameters:
|
483 |
+
attributions (torch.Tensor): A tensor of shape (C, H, W) representing the
|
484 |
+
intermediate activations or gradients at the target layer.
|
485 |
+
image (PIL.Image): The original image.
|
486 |
+
alpha (float): The blending factor for the heatmap overlay (default 0.5).
|
487 |
+
colormap (int): OpenCV colormap to apply (default cv2.COLORMAP_JET).
|
488 |
+
aggregation (str): How to aggregate across channels; either 'mean' or 'sum'.
|
489 |
+
|
490 |
+
Returns:
|
491 |
+
PIL.Image: The image overlaid with the Grad-CAM heatmap.
|
492 |
+
"""
|
493 |
+
# print("Generating Grad-CAM with shape:", cam.shape)
|
494 |
+
|
495 |
+
if normalize:
|
496 |
+
cam_min, cam_max = cam.min(), cam.max()
|
497 |
+
cam = cam - cam_min
|
498 |
+
cam = cam / (cam_max - cam_min)
|
499 |
+
# Convert tensor to numpy array
|
500 |
+
cam = torch.nn.functional.interpolate(cam.unsqueeze(0).unsqueeze(0), size=size, mode='bilinear').squeeze()
|
501 |
+
cam_np = cam.squeeze().detach().cpu().numpy()
|
502 |
+
|
503 |
+
# Apply Gaussian blur for smoother heatmaps
|
504 |
+
cam_np = cv2.GaussianBlur(cam_np, (5,5), sigmaX=0.8)
|
505 |
+
|
506 |
+
# Resize the cam to match the image size
|
507 |
+
width, height = size
|
508 |
+
cam_resized = cv2.resize(cam_np, (width, height))
|
509 |
+
|
510 |
+
# Convert the normalized map to a heatmap (0-255 uint8)
|
511 |
+
heatmap = np.uint8(255 * cam_resized)
|
512 |
+
heatmap = cv2.applyColorMap(heatmap, colormap)
|
513 |
+
# OpenCV produces heatmaps in BGR, so convert to RGB for consistency
|
514 |
+
heatmap = cv2.cvtColor(heatmap, cv2.COLOR_BGR2RGB)
|
515 |
+
|
516 |
+
# Convert original image to a numpy array
|
517 |
+
image_np = np.array(image)
|
518 |
+
image_np = cv2.resize(image_np, (width, height))
|
519 |
+
|
520 |
+
# Blend the heatmap with the original image
|
521 |
+
overlay = cv2.addWeighted(image_np, 1 - alpha, heatmap, alpha, 0)
|
522 |
+
|
523 |
+
return Image.fromarray(overlay)
|
524 |
+
|