Spaces:
Sleeping
Sleeping
added dff
Browse files
yolov5.py
CHANGED
@@ -82,90 +82,166 @@ def xai_yolov5(image):
|
|
82 |
return Image.fromarray(final_image), caption
|
83 |
|
84 |
|
85 |
-
|
86 |
-
import yaml
|
87 |
-
import torch
|
88 |
-
import warnings
|
89 |
-
warnings.filterwarnings('ignore')
|
90 |
-
from PIL import Image
|
91 |
import numpy as np
|
92 |
-
import
|
93 |
-
import cv2
|
94 |
import torch
|
95 |
-
|
96 |
-
from
|
97 |
-
from
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
98 |
|
99 |
-
# Check if CUDA is available
|
100 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
129 |
img = cv2.resize(img, (640, 640))
|
130 |
-
rgb_img_float = np.float32(img) /255.0
|
131 |
-
input_tensor = torch.from_numpy(rgb_img_float).permute(2, 0, 1).unsqueeze(0).to(device)
|
132 |
-
return img, rgb_img_float, input_tensor
|
133 |
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
target_layer = model.model.model.model[-lyr_idx] # Select a feature extraction layer
|
139 |
-
|
140 |
-
dff = DeepFeatureFactorization(model=model.model, target_layer=target_layer)
|
141 |
-
|
142 |
-
# Run DFF on the input tensor
|
143 |
-
concepts, batch_explanations = dff(input_tensor, n_components)
|
144 |
-
|
145 |
-
# Softmax normalization
|
146 |
-
concept_outputs = torch.softmax(torch.from_numpy(concepts), axis=-1).numpy()
|
147 |
-
concept_label_strings = create_labels(concept_outputs, top_k=top_k)
|
148 |
-
|
149 |
-
# Visualize explanations
|
150 |
-
visualization = show_factorization_on_image(rgb_img_float,
|
151 |
-
batch_explanations[0],
|
152 |
-
image_weight=0.2,
|
153 |
-
concept_labels=concept_label_strings)
|
154 |
|
155 |
-
|
156 |
-
plt.imshow(visualization)
|
157 |
-
plt.savefig("test" + str(lyr_idx) + ".png")
|
158 |
-
result = np.hstack((img, visualization))
|
159 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
160 |
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
166 |
|
167 |
-
# Test with images
|
168 |
-
for indx in range(2,12):
|
169 |
-
Image.fromarray(visualize_image(model,
|
170 |
-
"https://github.com/jacobgil/pytorch-grad-cam/blob/master/examples/both.png?raw=true", lyr_idx = indx))
|
171 |
-
"""
|
|
|
82 |
return Image.fromarray(final_image), caption
|
83 |
|
84 |
|
85 |
+
|
|
|
|
|
|
|
|
|
|
|
86 |
import numpy as np
|
87 |
+
from PIL import Image
|
|
|
88 |
import torch
|
89 |
+
import cv2
|
90 |
+
from typing import Callable, List, Tuple, Optional
|
91 |
+
from sklearn.decomposition import NMF
|
92 |
+
from pytorch_grad_cam.activations_and_gradients import ActivationsAndGradients
|
93 |
+
from pytorch_grad_cam.utils.image import scale_cam_image, create_labels_legend, show_factorization_on_image
|
94 |
+
import matplotlib.pyplot as plt
|
95 |
+
from pytorch_grad_cam.utils.image import show_factorization_on_image
|
96 |
+
import requests
|
97 |
+
import yaml
|
98 |
+
import matplotlib.patches as patches
|
99 |
|
|
|
100 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
101 |
+
def dff_l(activations, model, n_components):
|
102 |
+
batch_size, channels, h, w = activations.shape
|
103 |
+
print('activation', activations.shape)
|
104 |
+
target_layer_index = 4
|
105 |
+
reshaped_activations = activations.transpose((1, 0, 2, 3))
|
106 |
+
reshaped_activations[np.isnan(reshaped_activations)] = 0
|
107 |
+
reshaped_activations = reshaped_activations.reshape(
|
108 |
+
reshaped_activations.shape[0], -1)
|
109 |
+
offset = reshaped_activations.min(axis=-1)
|
110 |
+
reshaped_activations = reshaped_activations - offset[:, None]
|
111 |
+
model = NMF(n_components=n_components, init='random', random_state=0)
|
112 |
+
W = model.fit_transform(reshaped_activations)
|
113 |
+
H = model.components_
|
114 |
+
concepts = W + offset[:, None]
|
115 |
+
explanations = H.reshape(n_components, batch_size, h, w)
|
116 |
+
explanations = explanations.transpose((1, 0, 2, 3))
|
117 |
+
return concepts, explanations
|
118 |
+
|
119 |
+
class DeepFeatureFactorization:
|
120 |
+
def __init__(self,
|
121 |
+
model: torch.nn.Module,
|
122 |
+
target_layer: torch.nn.Module,
|
123 |
+
reshape_transform: Callable = None,
|
124 |
+
computation_on_concepts=None
|
125 |
+
):
|
126 |
+
self.model = model
|
127 |
+
self.computation_on_concepts = computation_on_concepts
|
128 |
+
self.activations_and_grads = ActivationsAndGradients(
|
129 |
+
self.model, [target_layer], reshape_transform)
|
130 |
+
|
131 |
+
def __call__(self,
|
132 |
+
input_tensor: torch.Tensor,
|
133 |
+
model: torch.nn.Module,
|
134 |
+
n_components: int = 16):
|
135 |
+
if isinstance(input_tensor, np.ndarray):
|
136 |
+
input_tensor = torch.from_numpy(input_tensor) # Convert NumPy array
|
137 |
+
|
138 |
+
batch_size, channels, h, w = input_tensor.size()
|
139 |
+
_ = self.activations_and_grads(input_tensor)
|
140 |
+
|
141 |
+
with torch.no_grad():
|
142 |
+
activations = self.activations_and_grads.activations[0].cpu(
|
143 |
+
).numpy()
|
144 |
+
|
145 |
+
concepts, explanations = dff_l(activations, model, n_components=n_components)
|
146 |
+
processed_explanations = []
|
147 |
+
|
148 |
+
for batch in explanations:
|
149 |
+
processed_explanations.append(scale_cam_image(batch, (w, h)))
|
150 |
+
|
151 |
+
if self.computation_on_concepts:
|
152 |
+
with torch.no_grad():
|
153 |
+
concept_tensors = torch.from_numpy(
|
154 |
+
np.float32(concepts).transpose((1, 0)))
|
155 |
+
concept_outputs = self.computation_on_concepts(
|
156 |
+
concept_tensors).cpu().numpy()
|
157 |
+
return concepts, processed_explanations, concept_outputs
|
158 |
+
else:
|
159 |
+
return concepts, processed_explanations, explanations
|
160 |
+
|
161 |
+
def __del__(self):
|
162 |
+
self.activations_and_grads.release()
|
163 |
+
|
164 |
+
def __exit__(self, exc_type, exc_value, exc_tb):
|
165 |
+
self.activations_and_grads.release()
|
166 |
+
if isinstance(exc_value, IndexError):
|
167 |
+
# Handle IndexError here...
|
168 |
+
print(
|
169 |
+
f"An exception occurred in ActivationSummary with block: {exc_type}. Message: {exc_value}")
|
170 |
+
return True
|
171 |
+
|
172 |
+
|
173 |
+
def dff_nmf(image, target_lyr, n_components):
|
174 |
+
mean = [0.485, 0.456, 0.406] # Mean for RGB channels
|
175 |
+
std = [0.229, 0.224, 0.225] # Standard deviation for RGB channels
|
176 |
+
img, rgb_img_float, input_tensor = image.to(device)
|
177 |
img = cv2.resize(img, (640, 640))
|
178 |
+
rgb_img_float = np.float32(img) / 255.0
|
179 |
+
input_tensor = torch.from_numpy(rgb_img_float).permute(2, 0, 1).unsqueeze(0).to(device)
|
|
|
180 |
|
181 |
+
model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True).to(device)
|
182 |
+
dff= DeepFeatureFactorization(model=model,
|
183 |
+
target_layer=model.model.model.model[int(target_lyr)],
|
184 |
+
computation_on_concepts=None)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
185 |
|
186 |
+
concepts, batch_explanations, explanations = dff(input_tensor, model, n_components)
|
|
|
|
|
|
|
187 |
|
188 |
+
|
189 |
+
yolov5_categories_url = \
|
190 |
+
"https://github.com/ultralytics/yolov5/raw/master/data/coco128.yaml" # URL to the YOLOv5 categories file
|
191 |
+
yaml_data = requests.get(yolov5_categories_url).text
|
192 |
+
labels = yaml.safe_load(yaml_data)['names'] # Parse the YAML file to get class names
|
193 |
+
num_classes = model.model.model.model[-1].nc
|
194 |
+
|
195 |
+
for indx in range( explanations[0].shape[0]):
|
196 |
+
upsampled_input = explanations[0][indx]
|
197 |
+
upsampled_input = torch.tensor(upsampled_input)
|
198 |
+
device = next(model.parameters()).device
|
199 |
+
input_tensor = upsampled_input.unsqueeze(0)
|
200 |
+
input_tensor = input_tensor.unsqueeze(1).repeat(1, 128, 1, 1)
|
201 |
+
detection_lyr = model.model.model.model[-1]
|
202 |
+
output1 = detection_lyr.m[0](input_tensor.to(device))
|
203 |
+
objectness = output1[..., 4] # Objectness score (index 4)
|
204 |
+
class_scores = output1[..., 5:] # Class scores (from index 5 onwards, representing 80 classes)
|
205 |
+
objectness = torch.sigmoid(objectness)
|
206 |
+
class_scores = torch.sigmoid(class_scores)
|
207 |
+
confidence_mask = objectness > 0.5
|
208 |
+
objectness = objectness[confidence_mask]
|
209 |
+
class_scores = class_scores[confidence_mask]
|
210 |
+
scores, class_ids = class_scores.max(dim=-1) # Get max class score per cell
|
211 |
+
scores = scores * objectness # Adjust scores by objectness
|
212 |
+
boxes = output1[..., :4] # First 4 values are x1, y1, x2, y2
|
213 |
+
boxes = boxes[confidence_mask] # Filter boxes by confidence mask
|
214 |
+
fig, ax = plt.subplots(1, figsize=(10, 10))
|
215 |
+
ax.imshow(torch.tensor(batch_explanations[0][indx]).cpu().numpy(), cmap="gray") # Display image
|
216 |
+
top_score_idx = scores.argmax(dim=0) # Get the index of the max score
|
217 |
+
top_score = scores[top_score_idx].item()
|
218 |
+
top_class_id = class_ids[top_score_idx].item()
|
219 |
+
top_box = boxes[top_score_idx].cpu().numpy()
|
220 |
+
scale_factor = 16
|
221 |
+
x1, y1, x2, y2 = top_box
|
222 |
+
x1, y1, x2, y2 = x1 * scale_factor, y1 * scale_factor, x2 * scale_factor, y2 * scale_factor
|
223 |
+
rect = patches.Rectangle(
|
224 |
+
(x1, y1), x2 - x1, y2 - y1,
|
225 |
+
linewidth=2, edgecolor='r', facecolor='none')
|
226 |
+
ax.add_patch(rect)
|
227 |
|
228 |
+
predicted_label = labels[top_class_id] # Map ID to label
|
229 |
+
ax.text(x1, y1, f"{predicted_label}: {top_score:.2f}",
|
230 |
+
color='r', fontsize=12, verticalalignment='top')
|
231 |
+
plt.show()
|
232 |
+
plt.savefig("test_" + str(indx) + ".png" )
|
233 |
+
plt.clf()
|
234 |
+
return rgb_img_float, explanations
|
235 |
+
|
236 |
+
|
237 |
+
def visualize_batch_explanations(rgb_img_float, batch_explanations, image_weight=0.7):
|
238 |
+
for i, explanation in enumerate(batch_explanations):
|
239 |
+
# Create visualization for each explanation
|
240 |
+
visualization = show_factorization_on_image(rgb_img_float, explanation, image_weight=image_weight)
|
241 |
+
plt.figure()
|
242 |
+
plt.imshow(visualization) # Correctly pass the visualization data
|
243 |
+
plt.title(f'Explanation {i + 1}') # Set the title for each plot
|
244 |
+
plt.axis('off') # Hide axes
|
245 |
+
plt.show() # Show the plot
|
246 |
+
plt.savefig("test_w.png")
|
247 |
|
|
|
|
|
|
|
|
|
|