AustingDong
commited on
Commit
·
f59a9b2
1
Parent(s):
369c141
add visual method changing: functions and pooling aggregated
Browse files- app.py +60 -41
- demo/model_utils.py +1 -1
- demo/modified_attn.py +221 -0
- demo/visualization.py +25 -22
app.py
CHANGED
@@ -5,6 +5,7 @@ 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
|
10 |
import matplotlib.pyplot as plt
|
@@ -53,7 +54,7 @@ def clean():
|
|
53 |
@spaces.GPU(duration=120)
|
54 |
def multimodal_understanding(model_type,
|
55 |
activation_map_method,
|
56 |
-
|
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
|
@@ -83,7 +84,7 @@ def multimodal_understanding(model_type,
|
|
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,
|
87 |
cam = cam.to("cpu")
|
88 |
cam = [generate_gradcam(cam, image, size=(224, 224))]
|
89 |
grad_cam.remove_hooks()
|
@@ -144,7 +145,7 @@ def multimodal_understanding(model_type,
|
|
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,
|
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])
|
@@ -159,7 +160,7 @@ def multimodal_understanding(model_type,
|
|
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,
|
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])
|
@@ -167,7 +168,7 @@ def multimodal_understanding(model_type,
|
|
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,
|
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):
|
@@ -190,13 +191,11 @@ def multimodal_understanding(model_type,
|
|
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 |
-
|
196 |
-
|
197 |
-
|
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)
|
@@ -218,36 +217,58 @@ def multimodal_understanding(model_type,
|
|
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 |
-
|
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 |
-
|
231 |
-
|
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 |
-
|
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 |
-
|
248 |
-
|
249 |
-
)
|
250 |
-
return res
|
251 |
|
252 |
elif model_type.split('-')[0] == "LLaVA":
|
253 |
|
@@ -259,32 +280,28 @@ def model_slider_change(model_type):
|
|
259 |
language_model_max_layer = 32 if version == "1.5" else 28
|
260 |
language_model_best_layer = 10
|
261 |
|
262 |
-
|
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 |
-
|
267 |
-
|
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 |
-
|
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 |
-
|
285 |
-
|
286 |
-
)
|
287 |
-
return res
|
288 |
|
289 |
|
290 |
|
@@ -362,7 +379,8 @@ with gr.Blocks() as demo:
|
|
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 |
-
|
|
|
366 |
|
367 |
|
368 |
visualization_layers_min = gr.Slider(minimum=1, maximum=12, value=12, step=1, label="visualization layers min")
|
@@ -377,10 +395,11 @@ with gr.Blocks() as demo:
|
|
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 |
|
@@ -492,7 +511,7 @@ with gr.Blocks() as demo:
|
|
492 |
|
493 |
understanding_button.click(
|
494 |
multimodal_understanding,
|
495 |
-
inputs=[model_selector, activation_map_method,
|
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 |
)
|
|
|
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 |
+
from demo.modified_attn import ModifiedLlamaAttention, ModifiedGemmaAttention
|
9 |
|
10 |
import numpy as np
|
11 |
import matplotlib.pyplot as plt
|
|
|
54 |
@spaces.GPU(duration=120)
|
55 |
def multimodal_understanding(model_type,
|
56 |
activation_map_method,
|
57 |
+
visual_method,
|
58 |
image, question, seed, top_p, temperature, target_token_idx,
|
59 |
visualization_layer_min, visualization_layer_max, focus, response_type, chart_type):
|
60 |
# Clear CUDA cache before generating
|
|
|
84 |
else:
|
85 |
target_layers = [all_layers[visualization_layer_min-1]]
|
86 |
grad_cam = VisualizationClip(clip_utils.model, target_layers)
|
87 |
+
cam, outputs, grid_size = grad_cam.generate_cam(inputs, target_token_idx=0, visual_method=visual_method)
|
88 |
cam = cam.to("cpu")
|
89 |
cam = [generate_gradcam(cam, image, size=(224, 224))]
|
90 |
grad_cam.remove_hooks()
|
|
|
145 |
cam = []
|
146 |
if focus == "Visual Encoder":
|
147 |
if target_token_idx != -1:
|
148 |
+
cam_tensors, grid_size, start = gradcam.generate_cam(prepare_inputs, tokenizer, temperature, top_p, target_token_idx, visual_method, focus)
|
149 |
cam_grid = cam_tensors.reshape(grid_size, grid_size)
|
150 |
cam_i = generate_gradcam(cam_grid, image)
|
151 |
cam_i = add_title_to_image(cam_i, input_ids_decoded[start + target_token_idx])
|
|
|
160 |
gradcam = VisualizationLLaVA(vl_gpt, target_layers)
|
161 |
elif model_name.split('-')[0] == "ChartGemma":
|
162 |
gradcam = VisualizationChartGemma(vl_gpt, target_layers)
|
163 |
+
cam_tensors, grid_size, start = gradcam.generate_cam(prepare_inputs, tokenizer, temperature, top_p, i, visual_method, focus)
|
164 |
cam_grid = cam_tensors.reshape(grid_size, grid_size)
|
165 |
cam_i = generate_gradcam(cam_grid, image)
|
166 |
cam_i = add_title_to_image(cam_i, input_ids_decoded[start + i])
|
|
|
168 |
gradcam.remove_hooks()
|
169 |
i += 1
|
170 |
else:
|
171 |
+
cam_tensors, grid_size, start = gradcam.generate_cam(prepare_inputs, tokenizer, temperature, top_p, target_token_idx, visual_method, focus)
|
172 |
if target_token_idx != -1:
|
173 |
input_text_decoded = input_ids_decoded[start + target_token_idx]
|
174 |
for i, cam_tensor in enumerate(cam_tensors):
|
|
|
191 |
|
192 |
# Collect Results
|
193 |
RESULTS_ROOT = "./results"
|
194 |
+
FILES_ROOT = f"{RESULTS_ROOT}/{model_name}/{focus}/{visual_method}/{chart_type}/layer{visualization_layer_min}-{visualization_layer_max}/{'all_tokens' if target_token_idx == -1 else f'--{input_ids_decoded[start + target_token_idx]}--'}"
|
195 |
os.makedirs(FILES_ROOT, exist_ok=True)
|
196 |
+
|
197 |
+
for i, cam_p in enumerate(cam):
|
198 |
+
cam_p.save(f"{FILES_ROOT}/{i}.png")
|
|
|
|
|
199 |
|
200 |
with open(f"{FILES_ROOT}/input_text_decoded.txt", "w") as f:
|
201 |
f.write(input_text_decoded)
|
|
|
217 |
def model_slider_change(model_type):
|
218 |
global model_utils, vl_gpt, tokenizer, clip_utils, model_name, language_model_max_layer, language_model_best_layer, vision_model_best_layer
|
219 |
model_name = model_type
|
220 |
+
|
221 |
+
|
222 |
+
encoder_only_res = [
|
223 |
+
gr.Dropdown(choices=["Visualization only"], value="Visualization only", label="response_type"),
|
224 |
+
gr.Dropdown(choices=["Visual Encoder"], value="Visual Encoder", label="focus"),
|
225 |
+
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type"),
|
226 |
+
gr.Dropdown(choices=["CLS", "max", "avg"], value="CLS", label="visual pooling method")
|
227 |
+
]
|
228 |
+
|
229 |
+
visual_res = [
|
230 |
+
gr.Dropdown(choices=["Visualization only", "answer + visualization"], value="Visualization only", label="response_type"),
|
231 |
+
gr.Dropdown(choices=["Visual Encoder"], value="Visual Encoder", label="focus"),
|
232 |
+
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type"),
|
233 |
+
gr.Dropdown(choices=["softmax", "sigmoid"], value="softmax", label="activation function")
|
234 |
+
]
|
235 |
+
|
236 |
+
language_res = [
|
237 |
+
gr.Dropdown(choices=["Visualization only", "answer + visualization"], value="answer + visualization", label="response_type"),
|
238 |
+
gr.Dropdown(choices=["Language Model"], value="Language Model", label="focus"),
|
239 |
+
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type"),
|
240 |
+
gr.Dropdown(choices=["softmax", "sigmoid"], value="softmax", label="activation function")
|
241 |
+
]
|
242 |
+
|
243 |
+
|
244 |
if model_type == "Clip":
|
245 |
clean()
|
246 |
set_seed()
|
247 |
clip_utils = Clip_Utils()
|
248 |
clip_utils.init_Clip()
|
249 |
+
sliders = [
|
|
|
250 |
gr.Slider(minimum=1, maximum=12, value=12, step=1, label="visualization layers min"),
|
251 |
gr.Slider(minimum=1, maximum=12, value=12, step=1, label="visualization layers max"),
|
252 |
+
]
|
253 |
+
return tuple(encoder_only_res + sliders)
|
254 |
+
|
|
|
255 |
elif model_type.split('-')[0] == "Janus":
|
256 |
|
257 |
clean()
|
258 |
set_seed()
|
259 |
model_utils = Janus_Utils()
|
260 |
vl_gpt, tokenizer = model_utils.init_Janus(model_type.split('-')[-1])
|
261 |
+
for layer in vl_gpt.language_model.model.layers:
|
262 |
+
layer.self_attn = ModifiedLlamaAttention(layer.self_attn)
|
263 |
+
|
264 |
language_model_max_layer = 24
|
265 |
language_model_best_layer = 8
|
266 |
+
|
267 |
+
sliders = [
|
|
|
268 |
gr.Slider(minimum=1, maximum=24, value=24, step=1, label="visualization layers min"),
|
269 |
gr.Slider(minimum=1, maximum=24, value=24, step=1, label="visualization layers max"),
|
270 |
+
]
|
271 |
+
return tuple(visual_res + sliders)
|
|
|
|
|
272 |
|
273 |
elif model_type.split('-')[0] == "LLaVA":
|
274 |
|
|
|
280 |
language_model_max_layer = 32 if version == "1.5" else 28
|
281 |
language_model_best_layer = 10
|
282 |
|
283 |
+
sliders = [
|
|
|
284 |
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer, step=1, label="visualization layers min"),
|
285 |
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer, step=1, label="visualization layers max"),
|
286 |
+
]
|
287 |
+
return tuple(language_res + sliders)
|
|
|
|
|
288 |
|
289 |
elif model_type.split('-')[0] == "ChartGemma":
|
290 |
clean()
|
291 |
set_seed()
|
292 |
model_utils = ChartGemma_Utils()
|
293 |
vl_gpt, tokenizer = model_utils.init_ChartGemma()
|
294 |
+
for layer in vl_gpt.language_model.model.layers:
|
295 |
+
layer.self_attn = ModifiedGemmaAttention(layer.self_attn)
|
296 |
language_model_max_layer = 18
|
297 |
vision_model_best_layer = 19
|
298 |
language_model_best_layer = 15
|
299 |
|
300 |
+
sliders = [
|
|
|
301 |
gr.Slider(minimum=1, maximum=language_model_best_layer, value=language_model_best_layer, step=1, label="visualization layers min"),
|
302 |
gr.Slider(minimum=1, maximum=language_model_best_layer, value=language_model_best_layer, step=1, label="visualization layers max"),
|
303 |
+
]
|
304 |
+
return tuple(language_res + sliders)
|
|
|
|
|
305 |
|
306 |
|
307 |
|
|
|
379 |
response_type = gr.Dropdown(choices=["Visualization only"], value="Visualization only", label="response_type")
|
380 |
focus = gr.Dropdown(choices=["Visual Encoder"], value="Visual Encoder", label="focus")
|
381 |
activation_map_method = gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="visualization type")
|
382 |
+
# activation_function = gr.Dropdown(choices=["softmax", "sigmoid"], value="softmax", label="activation function")
|
383 |
+
visual_method = gr.Dropdown(choices=["CLS", "max", "avg"], value="CLS", label="visual pooling method")
|
384 |
|
385 |
|
386 |
visualization_layers_min = gr.Slider(minimum=1, maximum=12, value=12, step=1, label="visualization layers min")
|
|
|
395 |
inputs=model_selector,
|
396 |
outputs=[
|
397 |
response_type,
|
|
|
|
|
398 |
focus,
|
399 |
+
activation_map_method,
|
400 |
+
visual_method,
|
401 |
+
visualization_layers_min,
|
402 |
+
visualization_layers_max
|
403 |
]
|
404 |
)
|
405 |
|
|
|
511 |
|
512 |
understanding_button.click(
|
513 |
multimodal_understanding,
|
514 |
+
inputs=[model_selector, activation_map_method, visual_method, image_input, question_input, und_seed_input, top_p, temperature, target_token_idx,
|
515 |
visualization_layers_min, visualization_layers_max, focus, response_type, chart_type],
|
516 |
outputs=[understanding_output, activation_map_output, understanding_target_token_decoded_output]
|
517 |
)
|
demo/model_utils.py
CHANGED
@@ -204,7 +204,7 @@ class ChartGemma_Utils(Model_Utils):
|
|
204 |
self.vl_gpt = PaliGemmaForConditionalGeneration.from_pretrained(
|
205 |
model_path,
|
206 |
torch_dtype=torch.float16,
|
207 |
-
attn_implementation="
|
208 |
output_attentions=True
|
209 |
)
|
210 |
self.vl_gpt, self.dtype, self.cuda_device = set_dtype_device(self.vl_gpt)
|
|
|
204 |
self.vl_gpt = PaliGemmaForConditionalGeneration.from_pretrained(
|
205 |
model_path,
|
206 |
torch_dtype=torch.float16,
|
207 |
+
attn_implementation="eager",
|
208 |
output_attentions=True
|
209 |
)
|
210 |
self.vl_gpt, self.dtype, self.cuda_device = set_dtype_device(self.vl_gpt)
|
demo/modified_attn.py
ADDED
@@ -0,0 +1,221 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import sys
|
3 |
+
import typing
|
4 |
+
import typing_extensions
|
5 |
+
from torch import nn
|
6 |
+
from typing import Callable, List, Optional, Tuple, Union
|
7 |
+
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
|
8 |
+
from transformers.cache_utils import Cache
|
9 |
+
from transformers.models.llama.configuration_llama import LlamaConfig
|
10 |
+
from transformers.models.llama.modeling_llama import LlamaAttention
|
11 |
+
from transformers.models.gemma.modeling_gemma import GemmaAttention
|
12 |
+
|
13 |
+
# from transformers.models.paligemma.modeling_paligemma import
|
14 |
+
|
15 |
+
if sys.version_info >= (3, 11):
|
16 |
+
Unpack = typing.Unpack
|
17 |
+
else:
|
18 |
+
Unpack = typing_extensions.Unpack
|
19 |
+
|
20 |
+
|
21 |
+
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
22 |
+
"""
|
23 |
+
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
24 |
+
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
25 |
+
"""
|
26 |
+
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
27 |
+
if n_rep == 1:
|
28 |
+
return hidden_states
|
29 |
+
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
30 |
+
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
31 |
+
|
32 |
+
def rotate_half(x):
|
33 |
+
"""Rotates half the hidden dims of the input."""
|
34 |
+
x1 = x[..., : x.shape[-1] // 2]
|
35 |
+
x2 = x[..., x.shape[-1] // 2 :]
|
36 |
+
return torch.cat((-x2, x1), dim=-1)
|
37 |
+
|
38 |
+
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
|
39 |
+
"""Applies Rotary Position Embedding to the query and key tensors.
|
40 |
+
|
41 |
+
Args:
|
42 |
+
q (`torch.Tensor`): The query tensor.
|
43 |
+
k (`torch.Tensor`): The key tensor.
|
44 |
+
cos (`torch.Tensor`): The cosine part of the rotary embedding.
|
45 |
+
sin (`torch.Tensor`): The sine part of the rotary embedding.
|
46 |
+
position_ids (`torch.Tensor`, *optional*):
|
47 |
+
Deprecated and unused.
|
48 |
+
unsqueeze_dim (`int`, *optional*, defaults to 1):
|
49 |
+
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
|
50 |
+
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
|
51 |
+
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
|
52 |
+
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
|
53 |
+
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
|
54 |
+
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
|
55 |
+
Returns:
|
56 |
+
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
|
57 |
+
"""
|
58 |
+
cos = cos.unsqueeze(unsqueeze_dim)
|
59 |
+
sin = sin.unsqueeze(unsqueeze_dim)
|
60 |
+
q_embed = (q * cos) + (rotate_half(q) * sin)
|
61 |
+
k_embed = (k * cos) + (rotate_half(k) * sin)
|
62 |
+
return q_embed, k_embed
|
63 |
+
|
64 |
+
def eager_attention_forward(
|
65 |
+
module: nn.Module,
|
66 |
+
query: torch.Tensor,
|
67 |
+
key: torch.Tensor,
|
68 |
+
value: torch.Tensor,
|
69 |
+
attention_mask: Optional[torch.Tensor],
|
70 |
+
scaling: float,
|
71 |
+
dropout: float = 0.0,
|
72 |
+
**kwargs,
|
73 |
+
):
|
74 |
+
key_states = repeat_kv(key, module.num_key_value_groups)
|
75 |
+
value_states = repeat_kv(value, module.num_key_value_groups)
|
76 |
+
|
77 |
+
attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
|
78 |
+
if attention_mask is not None:
|
79 |
+
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
|
80 |
+
attn_weights = attn_weights + causal_mask
|
81 |
+
|
82 |
+
attn_sigmoid_weights = nn.functional.sigmoid(attn_weights).to(query.dtype)
|
83 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
|
84 |
+
attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
|
85 |
+
attn_output = torch.matmul(attn_weights, value_states)
|
86 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
87 |
+
|
88 |
+
return attn_output, attn_weights, attn_sigmoid_weights
|
89 |
+
|
90 |
+
|
91 |
+
class ModifiedLlamaAttention(nn.Module):
|
92 |
+
def __init__(self, llama_attention_old: LlamaAttention):
|
93 |
+
super().__init__()
|
94 |
+
self.config = llama_attention_old.config
|
95 |
+
self.layer_idx = llama_attention_old.layer_idx
|
96 |
+
self.head_dim = llama_attention_old.head_dim
|
97 |
+
self.num_key_value_groups = llama_attention_old.num_key_value_groups
|
98 |
+
self.scaling = self.head_dim**-0.5
|
99 |
+
self.attention_dropout = llama_attention_old.attention_dropout
|
100 |
+
self.is_causal = True
|
101 |
+
|
102 |
+
self.q_proj = llama_attention_old.q_proj
|
103 |
+
self.k_proj = llama_attention_old.k_proj
|
104 |
+
self.v_proj = llama_attention_old.v_proj
|
105 |
+
self.o_proj = llama_attention_old.o_proj
|
106 |
+
|
107 |
+
|
108 |
+
def forward(
|
109 |
+
self,
|
110 |
+
hidden_states: torch.Tensor,
|
111 |
+
position_embeddings: Tuple[torch.Tensor, torch.Tensor],
|
112 |
+
attention_mask: Optional[torch.Tensor],
|
113 |
+
past_key_value: Optional[Cache] = None,
|
114 |
+
cache_position: Optional[torch.LongTensor] = None,
|
115 |
+
**kwargs: Unpack[FlashAttentionKwargs],
|
116 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
117 |
+
input_shape = hidden_states.shape[:-1]
|
118 |
+
hidden_shape = (*input_shape, -1, self.head_dim)
|
119 |
+
|
120 |
+
query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
|
121 |
+
key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
|
122 |
+
value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
|
123 |
+
|
124 |
+
cos, sin = position_embeddings
|
125 |
+
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
|
126 |
+
|
127 |
+
if past_key_value is not None:
|
128 |
+
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
129 |
+
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
130 |
+
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
131 |
+
|
132 |
+
attention_interface: Callable = eager_attention_forward
|
133 |
+
|
134 |
+
|
135 |
+
attn_output, attn_weights, attn_sigmoid_weights = attention_interface(
|
136 |
+
self,
|
137 |
+
query_states,
|
138 |
+
key_states,
|
139 |
+
value_states,
|
140 |
+
attention_mask,
|
141 |
+
dropout=0.0 if not self.training else self.attention_dropout,
|
142 |
+
scaling=self.scaling,
|
143 |
+
**kwargs,
|
144 |
+
)
|
145 |
+
|
146 |
+
self.attn_sigmoid_weights = attn_sigmoid_weights
|
147 |
+
|
148 |
+
attn_output = attn_output.reshape(*input_shape, -1).contiguous()
|
149 |
+
attn_output = self.o_proj(attn_output)
|
150 |
+
return attn_output, attn_weights
|
151 |
+
|
152 |
+
|
153 |
+
|
154 |
+
|
155 |
+
|
156 |
+
|
157 |
+
|
158 |
+
|
159 |
+
|
160 |
+
|
161 |
+
class ModifiedGemmaAttention(nn.Module):
|
162 |
+
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
163 |
+
|
164 |
+
def __init__(self, gemma_attention_old: GemmaAttention):
|
165 |
+
super().__init__()
|
166 |
+
self.config = gemma_attention_old.config
|
167 |
+
self.layer_idx = gemma_attention_old.layer_idx
|
168 |
+
self.head_dim = gemma_attention_old.head_dim
|
169 |
+
self.num_key_value_groups = gemma_attention_old.num_key_value_groups
|
170 |
+
self.scaling = gemma_attention_old.scaling
|
171 |
+
self.attention_dropout = gemma_attention_old.attention_dropout
|
172 |
+
self.is_causal = True
|
173 |
+
|
174 |
+
self.q_proj = gemma_attention_old.q_proj
|
175 |
+
self.k_proj = gemma_attention_old.k_proj
|
176 |
+
self.v_proj = gemma_attention_old.v_proj
|
177 |
+
self.o_proj = gemma_attention_old.o_proj
|
178 |
+
|
179 |
+
def forward(
|
180 |
+
self,
|
181 |
+
hidden_states: torch.Tensor,
|
182 |
+
position_embeddings: Tuple[torch.Tensor, torch.Tensor],
|
183 |
+
attention_mask: Optional[torch.Tensor],
|
184 |
+
past_key_value: Optional[Cache] = None,
|
185 |
+
cache_position: Optional[torch.LongTensor] = None,
|
186 |
+
**kwargs: Unpack[FlashAttentionKwargs],
|
187 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
188 |
+
input_shape = hidden_states.shape[:-1]
|
189 |
+
hidden_shape = (*input_shape, -1, self.head_dim)
|
190 |
+
|
191 |
+
query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
|
192 |
+
key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
|
193 |
+
value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
|
194 |
+
|
195 |
+
cos, sin = position_embeddings
|
196 |
+
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
|
197 |
+
|
198 |
+
if past_key_value is not None:
|
199 |
+
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
200 |
+
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
201 |
+
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
202 |
+
|
203 |
+
attention_interface: Callable = eager_attention_forward
|
204 |
+
|
205 |
+
attn_output, attn_weights, attn_sigmoid_weights = attention_interface(
|
206 |
+
self,
|
207 |
+
query_states,
|
208 |
+
key_states,
|
209 |
+
value_states,
|
210 |
+
attention_mask,
|
211 |
+
dropout=0.0 if not self.training else self.attention_dropout,
|
212 |
+
scaling=self.scaling,
|
213 |
+
**kwargs,
|
214 |
+
)
|
215 |
+
|
216 |
+
self.attn_sigmoid_weights = attn_sigmoid_weights
|
217 |
+
|
218 |
+
attn_output = attn_output.reshape(*input_shape, -1).contiguous()
|
219 |
+
attn_output = self.o_proj(attn_output)
|
220 |
+
return attn_output, attn_weights
|
221 |
+
|
demo/visualization.py
CHANGED
@@ -8,7 +8,7 @@ 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):
|
@@ -25,6 +25,7 @@ class Visualization:
|
|
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):
|
@@ -41,6 +42,9 @@ class Visualization:
|
|
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)
|
@@ -231,15 +235,15 @@ class VisualizationClip(Visualization):
|
|
231 |
super().__init__(model)
|
232 |
|
233 |
@spaces.GPU(duration=120)
|
234 |
-
def forward_backward(self, input_tensor,
|
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
|
241 |
output = output_full.image_embeds
|
242 |
-
elif
|
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)
|
@@ -250,11 +254,11 @@ class VisualizationClip(Visualization):
|
|
250 |
|
251 |
|
252 |
@spaces.GPU(duration=120)
|
253 |
-
def generate_cam(self, input_tensor, target_token_idx=None,
|
254 |
""" Generates Grad-CAM heatmap for ViT. """
|
255 |
self.setup_grads()
|
256 |
# Forward Backward pass
|
257 |
-
output_full = self.forward_backward(input_tensor,
|
258 |
|
259 |
cam_sum = self.grad_cam_vis()
|
260 |
cam_sum, grid_size = self.process(cam_sum)
|
@@ -291,34 +295,33 @@ class VisualizationJanus(Visualization):
|
|
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,
|
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 |
-
|
309 |
-
|
|
|
|
|
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,
|
317 |
|
318 |
self.setup_grads()
|
319 |
|
320 |
# Forward Backward pass
|
321 |
-
self.forward_backward(input_tensor, tokenizer, temperature, top_p, target_token_idx,
|
322 |
|
323 |
start_idx = 620
|
324 |
if focus == "Visual Encoder":
|
@@ -365,7 +368,7 @@ class VisualizationLLaVA(Visualization):
|
|
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,
|
369 |
|
370 |
self.setup_grads()
|
371 |
self.forward_backward(inputs)
|
@@ -401,7 +404,7 @@ class VisualizationChartGemma(Visualization):
|
|
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 |
|
@@ -417,11 +420,11 @@ class VisualizationChartGemma(Visualization):
|
|
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,
|
425 |
|
426 |
# Forward pass
|
427 |
self.setup_grads()
|
@@ -439,7 +442,7 @@ class VisualizationChartGemma(Visualization):
|
|
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()
|
|
|
8 |
from torch import nn
|
9 |
import spaces
|
10 |
from demo.modify_llama import *
|
11 |
+
from demo.modified_attn import ModifiedLlamaAttention
|
12 |
|
13 |
class Visualization:
|
14 |
def __init__(self, model, register=True):
|
|
|
25 |
self.hooks.append(layer.register_backward_hook(self._backward_hook))
|
26 |
|
27 |
def _forward_hook(self, module, input, output):
|
28 |
+
print("forward_hook: self_attn_input: ", input)
|
29 |
self.activations.append(output)
|
30 |
|
31 |
def _backward_hook(self, module, grad_in, grad_out):
|
|
|
42 |
layer.get_attn_map = types.MethodType(get_attn_map, layer)
|
43 |
|
44 |
def _forward_activate_hooks(self, module, input, output):
|
45 |
+
print("forward_activate_hool: module: ", module)
|
46 |
+
print("forward_activate_hook: self_attn_input: ", input)
|
47 |
+
|
48 |
attn_output, attn_weights = output # Unpack outputs
|
49 |
print("attn_output shape:", attn_output.shape)
|
50 |
print("attn_weights shape:", attn_weights.shape)
|
|
|
235 |
super().__init__(model)
|
236 |
|
237 |
@spaces.GPU(duration=120)
|
238 |
+
def forward_backward(self, input_tensor, visual_method, target_token_idx):
|
239 |
output_full = self.model(**input_tensor)
|
240 |
|
241 |
if target_token_idx is None:
|
242 |
target_token_idx = torch.argmax(output_full.logits, dim=1).item()
|
243 |
|
244 |
+
if visual_method == "CLS":
|
245 |
output = output_full.image_embeds
|
246 |
+
elif visual_method == "avg":
|
247 |
output = self.model.visual_projection(output_full.vision_model_output.last_hidden_state).mean(dim=1)
|
248 |
else:
|
249 |
output, _ = self.model.visual_projection(output_full.vision_model_output.last_hidden_state).max(dim=1)
|
|
|
254 |
|
255 |
|
256 |
@spaces.GPU(duration=120)
|
257 |
+
def generate_cam(self, input_tensor, target_token_idx=None, visual_method="CLS"):
|
258 |
""" Generates Grad-CAM heatmap for ViT. """
|
259 |
self.setup_grads()
|
260 |
# Forward Backward pass
|
261 |
+
output_full = self.forward_backward(input_tensor, visual_method, target_token_idx)
|
262 |
|
263 |
cam_sum = self.grad_cam_vis()
|
264 |
cam_sum, grid_size = self.process(cam_sum)
|
|
|
295 |
self._modify_layers()
|
296 |
self._register_hooks_activations()
|
297 |
|
298 |
+
def forward_backward(self, input_tensor, tokenizer, temperature, top_p, target_token_idx=None, visual_method="softmax", focus="Visual Encoder"):
|
299 |
# Forward
|
300 |
image_embeddings, inputs_embeddings, outputs = self.model(input_tensor, tokenizer, temperature, top_p)
|
301 |
input_ids = input_tensor.input_ids
|
302 |
+
start_idx = 620
|
303 |
+
self.model.zero_grad()
|
304 |
if focus == "Visual Encoder":
|
|
|
|
|
|
|
|
|
305 |
loss = outputs.logits.max(dim=-1).values[0, start_idx + target_token_idx]
|
306 |
loss.backward()
|
307 |
|
308 |
elif focus == "Language Model":
|
309 |
+
if target_token_idx == -1:
|
310 |
+
loss = outputs.logits.max(dim=-1).values.sum()
|
311 |
+
else:
|
312 |
+
loss = outputs.logits.max(dim=-1).values[0, start_idx + target_token_idx]
|
313 |
loss.backward()
|
314 |
|
315 |
+
self.activations = self.activations = [layer.attn_sigmoid_weights for layer in self.target_layers] if visual_method == "sigmoid" else [layer.get_attn_map() for layer in self.target_layers]
|
316 |
self.gradients = [layer.get_attn_gradients() for layer in self.target_layers]
|
317 |
|
318 |
@spaces.GPU(duration=120)
|
319 |
+
def generate_cam(self, input_tensor, tokenizer, temperature, top_p, target_token_idx=None, visual_method="softmax", focus="Visual Encoder"):
|
320 |
|
321 |
self.setup_grads()
|
322 |
|
323 |
# Forward Backward pass
|
324 |
+
self.forward_backward(input_tensor, tokenizer, temperature, top_p, target_token_idx, visual_method, focus)
|
325 |
|
326 |
start_idx = 620
|
327 |
if focus == "Visual Encoder":
|
|
|
368 |
self.gradients = [layer.get_attn_gradients() for layer in self.target_layers]
|
369 |
|
370 |
@spaces.GPU(duration=120)
|
371 |
+
def generate_cam(self, inputs, tokenizer, temperature, top_p, target_token_idx=None, visual_method="softmax", focus="Visual Encoder"):
|
372 |
|
373 |
self.setup_grads()
|
374 |
self.forward_backward(inputs)
|
|
|
404 |
self._modify_layers()
|
405 |
self._register_hooks_activations()
|
406 |
|
407 |
+
def forward_backward(self, inputs, focus, start_idx, target_token_idx, visual_method="softmax"):
|
408 |
outputs_raw = self.model(**inputs, output_hidden_states=True)
|
409 |
if focus == "Visual Encoder":
|
410 |
|
|
|
420 |
else:
|
421 |
loss = outputs_raw.logits.max(dim=-1).values[0, start_idx + target_token_idx]
|
422 |
loss.backward()
|
423 |
+
self.activations = [layer.attn_sigmoid_weights for layer in self.target_layers] if visual_method == "sigmoid" else [layer.get_attn_map() for layer in self.target_layers]
|
424 |
self.gradients = [layer.get_attn_gradients() for layer in self.target_layers]
|
425 |
|
426 |
@spaces.GPU(duration=120)
|
427 |
+
def generate_cam(self, inputs, tokenizer, temperature, top_p, target_token_idx=None, visual_method="softmax", focus="Visual Encoder"):
|
428 |
|
429 |
# Forward pass
|
430 |
self.setup_grads()
|
|
|
442 |
start_idx = last + 1
|
443 |
|
444 |
|
445 |
+
self.forward_backward(inputs, focus, start_idx, target_token_idx, visual_method)
|
446 |
if focus == "Visual Encoder":
|
447 |
|
448 |
cam_sum = self.grad_cam_vis()
|