Spaces:
Sleeping
Sleeping
File size: 21,383 Bytes
ef2badf 1b72673 ef2badf |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 |
'''
this is a combined script that implements DETR object detection with interpretability methods
using Grad-CAM, Grad-CAM++, Integrated Gradients, and Monte Carlo Dropout for uncertainty estimation.
It provides a Gradio-based web interface for users to upload images, select detected objects
and visualize explanations and uncertainty maps.
How to run it:
```python
python detr_and_interp.py
```
'''
import torch, requests, numpy as np
import matplotlib.pyplot as plt
import matplotlib.patches as patches
from PIL import Image, ImageFilter
import gradio as gr
from transformers import DetrImageProcessor, DetrForObjectDetection
from torchvision.transforms.functional import resize
from captum.attr import IntegratedGradients
import torch.nn.functional as F
import logging
import os
from datetime import datetime
# ---------- Logging Setup ----------
log_dir = os.path.join(os.path.dirname(__file__), "logs")
os.makedirs(log_dir, exist_ok=True)
log_file = os.path.join(log_dir, f"detr_interp_{datetime.now().strftime('%Y%m%d_%H%M%S')}.log")
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s - %(levelname)s - %(funcName)s:%(lineno)d - %(message)s',
handlers=[
logging.FileHandler(log_file),
logging.StreamHandler()
]
)
logger = logging.getLogger(__name__)
logger.info("Starting DETR Interpretability Dashboard")
device = "cuda" if torch.cuda.is_available() else "cpu"
logger.info(f"Using device: {device}")
model_name = "facebook/detr-resnet-50"
logger.info(f"Loading model: {model_name}")
model = DetrForObjectDetection.from_pretrained(model_name).to(device)
extractor = DetrImageProcessor.from_pretrained(model_name)
model.eval()
logger.info("Model loaded and set to evaluation mode")
# ---------- Grad-CAM / Grad-CAM++ ----------
def gradcam(img, det_idx, keep, pixel_values, use_pp=False):
"""
Compute Grad-CAM (or Grad-CAM++) heatmap for a selected detection.
What it computes:
- Captures feature-map activations from a late conv layer and the gradients of the
detection score w.r.t. those activations. Channel-wise weights are computed from
gradients and used to combine feature maps into a spatial heatmap.
Why this matters:
- Highlights which spatial regions the model used to make the prediction. Useful to
check whether the detector is attending to the object vs irrelevant background.
How to interpret results:
- High values in the returned heatmap indicate regions that contributed positively to
the detection score. Grad-CAM++ (use_pp=True) computes a refined weighting that often
yields sharper, better-localized maps when multiple instances overlap.
Caveats & tips:
- Choosing a layer too early will give fine-grained but semantically weak maps; too late
will be coarse. We pick a late backbone conv block (layer4[-1]) as a sensible default.
- Hooks must be removed after use to avoid memory leaks; we do that below.
References:
- Selvaraju et al., Grad-CAM (2017): https://arxiv.org/abs/1610.02391
"""
logger.info(f"Running {'Grad-CAM++' if use_pp else 'Grad-CAM'} for detection {det_idx}")
try:
# pick a late conv layer that still retains spatial info
conv_layer = model.model.backbone.conv_encoder.model.layer4[-1]
activations, gradients = {}, {}
def fwd(m, i, o):
activations["v"] = o.detach()
def bwd(m, gi, go):
gradients["v"] = go[0].detach()
h1 = conv_layer.register_forward_hook(fwd)
h2 = conv_layer.register_full_backward_hook(bwd) if hasattr(conv_layer, "register_full_backward_hook") else conv_layer.register_backward_hook(bwd)
logger.debug("Hooks registered for Grad-CAM")
outputs_for_attr = model(pixel_values)
logits = outputs_for_attr.logits
labels = logits.argmax(-1).squeeze(0)
label_id = labels[keep.nonzero()[det_idx]].item()
score = logits[0, keep.nonzero()[det_idx], label_id]
logger.debug(f"Target label_id: {label_id}, score: {score.item():.4f}")
model.zero_grad()
score.backward()
acts = activations["v"].squeeze(0)
grads = gradients["v"].squeeze(0)
logger.debug(f"Activations shape: {acts.shape}, Gradients shape: {grads.shape}")
if use_pp: # Grad-CAM++
weights = (grads ** 2).mean(dim=(1, 2)) / (2 * (grads ** 2).mean(dim=(1, 2)) + (acts * grads ** 3).mean(dim=(1, 2)) + 1e-8)
else: # vanilla Grad-CAM
weights = grads.mean(dim=(1, 2))
cam = torch.relu((weights[:, None, None] * acts).sum(0))
cam = cam / (cam.max() + 1e-8)
cam_resized = resize(cam.unsqueeze(0).unsqueeze(0), img.size[::-1])[0, 0].cpu().numpy()
h1.remove(); h2.remove()
logger.info(f"{'Grad-CAM++' if use_pp else 'Grad-CAM'} completed successfully")
return cam_resized
except Exception as e:
logger.error(f"Error in gradcam: {str(e)}", exc_info=True)
raise
# ---------- Integrated Gradients ----------
def integrated_grad(img, det_idx, keep, outputs_for_attr, pixel_values, baseline="black"):
"""
Compute Integrated Gradients attribution map for a detection's logit.
What it computes:
- Integrates gradients along a path from a baseline input to the real input in embedding
space, producing per-pixel (or per-channel) attributions.
Why baseline choice matters:
- The baseline defines what the model should consider as 'no signal'. Common choices:
black (zeros), a blurred version of the image, or a neutral/mean image. Different
baselines highlight different aspects of the input.
How to read the output:
- Values > 0 indicate pixels that increase the detection logit vs baseline; values < 0
reduce it. We normalize the result to [0,1] for visualization convenience.
Tips:
- Increase n_steps for smoother attributions (costlier). Check convergence_delta to
validate IG's completeness property.
References:
- Distill article on baselines: https://distill.pub/2020/attribution-baselines
- Captum IntegratedGradients docs: https://captum.ai/api/integrated_gradients.html
"""
logger.info(f"Running Integrated Gradients with {baseline} baseline for detection {det_idx}")
try:
logits = outputs_for_attr.logits
labels = logits.argmax(-1).squeeze(0)
label_id = labels[keep.nonzero()[det_idx]].item()
logger.debug(f"IG target label_id: {label_id}")
# Baselines
if baseline == "black":
base = torch.zeros_like(pixel_values)
logger.debug("Using black baseline")
elif baseline == "blur":
blur = img.filter(ImageFilter.GaussianBlur(radius=15))
base = extractor(images=blur, return_tensors="pt")["pixel_values"].to(device)
logger.debug("Using blurred baseline")
else:
base = torch.zeros_like(pixel_values)
logger.debug("Defaulting to black baseline")
def forward_func(pix):
return model(pix).logits[:, keep.nonzero()[det_idx], label_id]
ig = IntegratedGradients(forward_func)
attr, _ = ig.attribute(pixel_values, baselines=base, n_steps=25, return_convergence_delta=True)
arr = attr.squeeze().mean(0).cpu().detach().numpy()
logger.info(f"Integrated Gradients with {baseline} baseline completed")
return (arr - arr.min()) / (arr.max() - arr.min() + 1e-8)
except Exception as e:
logger.error(f"Error in integrated_grad: {str(e)}", exc_info=True)
raise
# ---------- Monte Carlo Dropout Uncertainty ----------
def mc_dropout_uncertainty(img, det_idx, keep, pixel_values, n_samples=20, dropout_p=0.1):
"""
Estimate uncertainty by running multiple stochastic forward passes with dropout active.
What it computes:
- Runs the model multiple times with dropout enabled and computes a CAM per run.
- Returns the per-pixel mean and standard deviation across CAMs. High std indicates
the model's focus is unstable across stochastic perturbations.
Why this helps:
- If heatmaps vary a lot, the interpretability output is less reliable. Use this to flag
detections where explanations may not be trustworthy.
Practical tips:
- Increasing n_samples reduces variance in the estimate but increases runtime.
- Temporarily sets the model to train mode to activate dropout modules; restores eval mode.
"""
logger.info(f"Running MC Dropout uncertainty: samples={n_samples}, p={dropout_p}, detection={det_idx}")
try:
def enable_dropout(m):
if isinstance(m, torch.nn.Dropout):
m.train()
model.train()
model.apply(enable_dropout)
cams = []
conv_layer = model.model.backbone.conv_encoder.model.layer4[-1]
for i in range(n_samples):
outputs = model(pixel_values)
logits = outputs.logits
labels = logits.argmax(-1).squeeze(0)
label_id = labels[keep.nonzero()[det_idx]].item()
score = logits[0, keep.nonzero()[det_idx], label_id]
acts, grads = {}, {}
def fwd(m, i, o):
acts['v'] = o.detach()
def bwd(m, gi, go):
grads['v'] = go[0].detach()
h1 = conv_layer.register_forward_hook(fwd)
h2 = (conv_layer.register_full_backward_hook(bwd)
if hasattr(conv_layer, 'register_full_backward_hook')
else conv_layer.register_backward_hook(bwd))
model.zero_grad()
score.backward(retain_graph=False)
if 'v' not in acts:
logger.warning(f"No activations captured in sample {i}, using fallback zero map")
cam_resized = np.zeros((img.size[1], img.size[0]))
else:
act = acts['v'].squeeze(0)
grad = grads['v'].squeeze(0)
weights = grad.mean(dim=(1, 2))
cam = torch.relu((weights[:, None, None] * act).sum(0))
cam = cam / (cam.max() + 1e-8)
cam_resized = resize(cam.unsqueeze(0).unsqueeze(0), img.size[::-1])[0, 0].cpu().numpy()
cams.append(cam_resized)
h1.remove(); h2.remove()
model.eval()
if len(cams) == 0:
logger.error("No valid CAM maps generated")
return np.zeros((img.size[1], img.size[0])), np.zeros((img.size[1], img.size[0]))
cams_arr = np.stack(cams, axis=0)
mean_map = cams_arr.mean(0)
std_map = cams_arr.std(0)
mean_map = (mean_map - mean_map.min()) / (mean_map.max() - mean_map.min() + 1e-8)
std_map = (std_map - std_map.min()) / (std_map.max() - std_map.min() + 1e-8)
logger.info("MC Dropout uncertainty completed")
return mean_map, std_map
except Exception as e:
logger.error(f"Error in mc_dropout_uncertainty: {str(e)}", exc_info=True)
model.eval()
raise
# ---------- Full pipeline ----------
def interpret(img, det_choice, conf_thresh, cam_variant, mc_samples, dropout_p):
logger.info(f"Starting interpretation - detection: {det_choice}, threshold: {conf_thresh}, cam: {cam_variant}, mc_samples: {mc_samples}, dropout_p: {dropout_p}")
try:
inputs = extractor(images=img, return_tensors="pt").to(device)
with torch.no_grad(): outputs = model(**inputs)
pixel_values_attr = inputs["pixel_values"].clone().requires_grad_(True)
target_sizes = [img.size[::-1]]
results = extractor.post_process_object_detection(outputs, target_sizes=target_sizes, threshold=0.0)[0]
keep = results["scores"] > conf_thresh
labels, scores = results["labels"][keep], results["scores"][keep]
logger.info(f"Found {len(labels)} detections above threshold {conf_thresh}")
if len(labels) == 0:
logger.warning("No detections found above threshold")
return None, "No detections above threshold", None, ""
if det_choice is None:
det_idx = 0
else:
try: det_idx = int(str(det_choice).split(":")[0])
except: det_idx = 0
label = model.config.id2label[labels[det_idx].item()]
logger.info(f"Selected detection {det_idx}: {label}")
# Grad-CAM / Grad-CAM++ (single deterministic pass)
cam = gradcam(img, det_idx, keep, pixel_values_attr, use_pp=(cam_variant=="Grad-CAM++"))
fig1, ax1 = plt.subplots(); ax1.imshow(img); ax1.imshow(cam, cmap="jet", alpha=0.5); ax1.axis("off")
ax1.set_title(f"{cam_variant}: {label}"); plt.close(fig1)
logger.debug(f"{cam_variant} visualization created")
# MC Dropout Uncertainty analysis
mean_map, std_map = mc_dropout_uncertainty(img, det_idx, keep, pixel_values_attr, n_samples=int(mc_samples), dropout_p=float(dropout_p))
# Create a composite figure: mean map and std map side-by-side
fig2, axes = plt.subplots(1,2, figsize=(8,4))
axes[0].imshow(img); axes[0].imshow(mean_map, cmap='hot', alpha=0.5); axes[0].axis('off'); axes[0].set_title('Predictive Mean')
axes[1].imshow(img); axes[1].imshow(std_map, cmap='viridis', alpha=0.5); axes[1].axis('off'); axes[1].set_title('Predictive Std (Uncertainty)')
plt.close(fig2)
logger.debug("MC Dropout uncertainty visualization created")
exp1 = f"π {cam_variant}:\nGradient-weighted feature maps β highlights where DETR focused."
exp2 = f"π MC Dropout Uncertainty:\nSamples={mc_samples}, dropout={dropout_p}. Shows predictive mean and per-pixel std as uncertainty."
logger.info("Interpretation completed successfully")
return fig1, exp1, fig2, exp2
except Exception as e:
logger.error(f"Error in interpret function: {str(e)}", exc_info=True)
return None, f"Error: {str(e)}", None, ""
# ---------- Gradio UI ----------
with gr.Blocks() as demo:
gr.Markdown("## π§ DETR Interpretability Dashboard with Controls")
gr.Markdown(
"""
**How to use this dashboard**
- Upload an image using the left panel. The model will run object detection and list detected objects. Try [imageNet](https://www.image-net.org/)
- Use the "Confidence Threshold" slider to filter detections by score. Detections below the threshold are hidden.
- Pick a detection from the dropdown to generate explanations for that object.
- Choose between `Grad-CAM` and `Grad-CAM++` (Grad-CAM++ often gives sharper, more localized maps).
- `MC Dropout Samples` controls how many stochastic forward passes are used to estimate prediction uncertainty. More samples give smoother estimates but take longer.
- `Dropout Probability` sets the dropout rate used during MC Dropout; higher values typically increase predicted uncertainty.
Tooltips are provided on each control (hover or focus) for quick hints.
"""
)
with gr.Row():
img_in = gr.Image(type="pil", label="Upload an image")
det_out = gr.Label(label="Detections")
det_fig = gr.Plot(label="Detections visualization")
det_choice = gr.Dropdown(label="Pick a detection for explanation")
with gr.Row():
conf_thresh = gr.Slider(0, 1, value=0.7, step=0.05, label="Confidence Threshold")
cam_variant = gr.Radio(["Grad-CAM", "Grad-CAM++"], value="Grad-CAM", label="Grad-CAM Variant")
mc_samples = gr.Slider(1, 100, value=20, step=1, label="MC Dropout Samples")
dropout_p = gr.Slider(0.0, 0.9, value=0.1, step=0.05, label="Dropout Probability")
btn = gr.Button("Explain")
gc_fig = gr.Plot(label="Grad-CAM / Grad-CAM++")
gc_txt = gr.Textbox(label="Explanation (Grad-CAM)")
unc_fig = gr.Plot(label="Uncertainty (MC Dropout)")
unc_txt = gr.Textbox(label="Explanation (Uncertainty)")
# Visible control tooltips section (for environments where hovering tooltips are not available)
gr.Markdown(
"""
**Control tooltips (quick reference)**
- Confidence Threshold: Filter out detections with confidence below this value.
- Grad-CAM Variant: Choose the gradient-based visualization method. Grad-CAM++ may highlight smaller regions more precisely.
- MC Dropout Samples: Number of stochastic forward passes for uncertainty estimation. Increase for more stable results.
- Dropout Probability: Dropout rate used during MC Dropout sampling. Higher values typically increase predictive variance.
- Pick a detection: Select which detected object to explain. Format shown as 'index: label (score)'.
"""
)
# ---------- Key interpretability choices (Feynman-style) ----------
gr.Markdown(
"""
**Key interpretability choices & why they matter**
- **Baseline (Integrated Gradients)**: Defines what 'no signal' looks like. Black (zeros) is simple, but blurred or neutral baselines may give more meaningful attributions.
- **Which conv layer for Grad-CAM**: Early layers give fine texture but low semantics; very late layers are coarse. A late backbone conv (default used) is a good compromise.
- **Number of MC Dropout samples**: More samples = smoother, more stable uncertainty estimates, but higher compute cost.
- **Grad-CAM vs Grad-CAM++**: Grad-CAM++ can be sharper and better for overlapping instances; vanilla Grad-CAM is faster and simpler.
"""
)
# ---------- Further reading / Feynman-style references ----------
# Add short, clickable references so users can read the original papers and deep-dive articles.
gr.Markdown(
"""
**Further reading (recommended)**
- [Grad-CAM β Selvaraju et al., 2017 (arXiv)](https://arxiv.org/abs/1610.02391) β the original Grad-CAM paper; explains the core idea of gradient-weighted localization.
- [Grad-CAM++ β Chattopadhay et al.](https://arxiv.org/abs/1710.11063) β an improved variant that often produces sharper maps and handles multiple instances better.
- [Visualizing the Impact of Feature Attribution Baselines (Distill)](https://distill.pub/2020/attribution-baselines) β an accessible deep dive on baseline choices for Integrated Gradients.
- [Captum docs β IntegratedGradients](https://captum.ai/api/integrated_gradients.html) β practical API notes for baseline, n_steps, and convergence delta.
- [Constructing sensible baselines for Integrated Gradients](https://arxiv.org/abs/2004.09627) β discussion and techniques for choosing baselines beyond a black image.
- [A New Baseline Assumption of Integrated Gradients Based on Shapley Values](https://arxiv.org/html/2310.04821v3) β recent research on improved baselines.
"""
)
# Helper: safe label getter in case model.config.id2label is missing or not a dict
def safe_label_lookup(idx):
try:
id2label = getattr(model.config, 'id2label', None)
if id2label is None:
return f"Class {idx}"
return id2label.get(int(idx), f"Class {idx}")
except Exception:
return f"Class {idx}"
def run_detect(img, conf_thresh):
logger.info(f"Running detection with confidence threshold: {conf_thresh}")
try:
inputs = extractor(images=img, return_tensors="pt").to(device)
with torch.no_grad(): outputs = model(**inputs)
target_sizes = [img.size[::-1]]
results = extractor.post_process_object_detection(outputs, target_sizes=target_sizes, threshold=0.0)[0]
keep = results["scores"] > conf_thresh
boxes, labels, scores = results["boxes"][keep], results["labels"][keep], results["scores"][keep]
logger.info(f"Detection found {len(labels)} objects above threshold")
det_list = [f"{i}: {safe_label_lookup(l.item())} ({s:.2f})" for i,(l,s) in enumerate(zip(labels,scores))]
fig, ax = plt.subplots(); ax.imshow(img); ax.axis("off")
for box,label,score in zip(boxes,labels,scores):
xmin,ymin,xmax,ymax = box
ax.add_patch(patches.Rectangle((xmin,ymin),xmax-xmin,ymax-ymin,fill=False,color="red",lw=2))
ax.text(xmin,ymin,f"{safe_label_lookup(label.item())}:{score:.2f}",color="black",
bbox=dict(facecolor="yellow",alpha=0.5))
plt.close(fig)
default_val = det_list[0] if len(det_list) > 0 else None
logger.debug("Detection visualization created")
return {det_out: str(det_list), det_fig: fig, det_choice: gr.update(choices=det_list, value=default_val)}
except Exception as e:
logger.error(f"Error in run_detect: {str(e)}", exc_info=True)
return {det_out: "Error in detection", det_fig: None, det_choice: gr.update(choices=[], value=None)}
img_in.change(run_detect, inputs=[img_in, conf_thresh], outputs=[det_out, det_fig, det_choice])
btn.click(interpret, inputs=[img_in, det_choice, conf_thresh, cam_variant, mc_samples, dropout_p],
outputs=[gc_fig, gc_txt, unc_fig, unc_txt])
logger.info("Gradio interface configured, launching demo")
demo.launch()
|