Faster RCNN Inference Pipeline
Section 6 of 6 in the Faster RCNN from-scratch series We load the checkpoint saved in section 05, run inference on COCO validation images streamed from Hugging Face, and visualise detections. Topics covered:- Loading and verifying a checkpoint
- Running
model.eval()forward pass (proposal generation + postprocessing) - Visualising class-agnostic proposals and final detections
- Measuring per-image inference latency
- Pseudo-evaluation: counting detections above a confidence threshold
Expected outcome after 5-step training: the checkpoint from section 05 is a
sanity-check model, not a trained detector. Expect zero or very few final
detections above the 0.05 score threshold, the ROI classifier has seen only
5 images and its weights are essentially random. What you should see is a healthy
set of RPN proposals (cyan boxes), confirming the proposal pipeline works end-to-end.
A properly trained COCO model (≥12 epochs) would show labelled red boxes on every image.
import sys, os, pathlib
# Locate frcnn_common.py, works whether run via papermill or interactively
_nb_candidates = [
pathlib.Path.cwd().parent, # interactive: cwd is the notebook dir
pathlib.Path.cwd() / 'notebooks' / 'scene-understanding' / 'object-detection' / 'faster-rcnn' / 'pytorch', # papermill: cwd is repo root
]
for _p in _nb_candidates:
if (_p / 'frcnn_common.py').exists():
sys.path.insert(0, str(_p))
break
import torch
import torchvision.transforms.functional as TF
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.patches as patches
from datasets import load_dataset
import time
from frcnn_common import (
IMG_SIZE, NUM_CLASSES, DEVICE,
IMAGENET_MEAN, IMAGENET_STD,
COCO_NAMES,
FasterRCNN,
)
print(f"Device: {DEVICE}")
/workspaces/eng-ai-agents/.venv/lib/python3.11/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html
from .autonotebook import tqdm as notebook_tqdm
Device: cuda
# All model components are imported from frcnn_common.
# FasterRCNN assembles: ResNet50 + FPN + RPN + ROI heads.
print(f"Model classes imported. NUM_CLASSES={NUM_CLASSES}, IMG_SIZE={IMG_SIZE}")
Model classes imported. NUM_CLASSES=81, IMG_SIZE=400
# ─── Load checkpoint ──────────────────────────────────────────────────────────
CKPT_PATH = 'checkpoints/faster_rcnn_demo.pth'
assert os.path.exists(CKPT_PATH), f"Checkpoint not found at {CKPT_PATH}. Run notebook 05 first."
model = FasterRCNN(num_classes=NUM_CLASSES).to(DEVICE)
ckpt = torch.load(CKPT_PATH, map_location=DEVICE)
model.load_state_dict(ckpt['model_state_dict'])
model.eval()
print(f"Checkpoint loaded: {CKPT_PATH}")
print(f" Trained for {ckpt['steps_trained']} steps")
print(f" Final losses: { {k: f'{v:.4f}' for k,v in ckpt['final_losses'].items()} }")
total = sum(p.numel() for p in model.parameters())
print(f" Parameters: {total/1e6:.1f}M")
Checkpoint loaded: checkpoints/faster_rcnn_demo.pth
Trained for 5 steps
Final losses: {'rpn_cls': '0.6011', 'rpn_box': '0.1212', 'roi_cls': '1.6939', 'roi_box': '0.0611', 'total': '2.4773'}
Parameters: 41.8M
# ─── Inference on 4 validation images ─────────────────────────────────────────
val_ds = load_dataset('detection-datasets/coco', split='val', streaming=True)
images_pil, results_list, latencies = [], [], []
NUM_IMAGES = 4
with torch.no_grad():
for i, sample in enumerate(val_ds):
if i >= NUM_IMAGES:
break
img_pil = sample['image'].convert('RGB')
images_pil.append(img_pil)
t = ((TF.to_tensor(img_pil.resize((IMG_SIZE, IMG_SIZE))) - IMAGENET_MEAN) / IMAGENET_STD)
t = t.unsqueeze(0).to(DEVICE)
t0 = time.perf_counter()
dets, proposals = model(t)
if DEVICE.type == 'cuda': torch.cuda.synchronize()
latencies.append((time.perf_counter() - t0) * 1000)
results_list.append(dets[0])
print(f"Mean latency: {sum(latencies)/len(latencies):.1f} ms ({IMG_SIZE}x{IMG_SIZE} input)")
print("Detections per image:", [len(r['boxes']) for r in results_list])
print("(Detections are random, model trained only 5 steps)")
Mean latency: 262.9 ms (400x400 input)
Detections per image: [0, 0, 0, 0]
(Detections are random, model trained only 5 steps)
# ─── Visualise proposals + detections ─────────────────────────────────────────
fig, axes = plt.subplots(2, NUM_IMAGES, figsize=(5*NUM_IMAGES, 10))
TOP_K_PROPS = 30
with torch.no_grad():
for col, sample in enumerate(load_dataset('detection-datasets/coco',
split='val', streaming=True)):
if col >= NUM_IMAGES: break
img_pil = sample['image'].convert('RGB')
img_res = img_pil.resize((IMG_SIZE, IMG_SIZE))
t = ((TF.to_tensor(img_res) - IMAGENET_MEAN) / IMAGENET_STD).unsqueeze(0).to(DEVICE)
_, proposals = model(t)
props = proposals[0].cpu()[:TOP_K_PROPS]
dets = results_list[col]
# Row 0: RPN proposals
ax = axes[0][col]
ax.imshow(img_res); ax.axis('off')
ax.set_title(f'Image {col+1}: top-{TOP_K_PROPS} proposals', fontsize=9)
for box in props.tolist():
x1,y1,x2,y2=box
ax.add_patch(patches.Rectangle((x1,y1),x2-x1,y2-y1,
linewidth=1,edgecolor='cyan',facecolor='none'))
# Row 1: final detections
ax = axes[1][col]
ax.imshow(img_res); ax.axis('off')
n_det = len(dets['boxes'])
ax.set_title(f'Image {col+1}: {n_det} detections', fontsize=9)
for box,score,label in zip(dets['boxes'].tolist(),
dets['scores'].tolist(),
dets['labels'].tolist()):
x1,y1,x2,y2=box
cls_name = COCO_NAMES[label] if label < len(COCO_NAMES) else str(label)
ax.add_patch(patches.Rectangle((x1,y1),x2-x1,y2-y1,
linewidth=1.5,edgecolor='red',facecolor='none'))
ax.text(x1,y1-2,f'{cls_name} {score:.2f}',
color='white',fontsize=6,backgroundcolor='red')
plt.suptitle('Row 1: RPN proposals | Row 2: Final detections (5-step model)', y=1.01)
plt.tight_layout()
os.makedirs('images', exist_ok=True)
plt.savefig('images/inference_results.png', dpi=100, bbox_inches='tight')
plt.show()
Series complete. Faster RCNN from scratch, all 6 notebooks executed.
Pseudo-evaluation: detection counts above threshold
This is not mAP evaluation. It counts detections that survive the score
threshold (0.05) and NMS, a rough proxy for “did the model produce any confident
predictions?” on a small sample. For real evaluation use the
COCO evaluation API with a fully trained checkpoint.
# ─── Pseudo-evaluation: detection counts (NOT mAP) ────────────────────────────
SCORE_THRESHOLDS = [0.05, 0.1, 0.3, 0.5]
EVAL_IMAGES = NUM_IMAGES # reuse the 4 images already run above
print(f"Detection counts across {EVAL_IMAGES} validation images")
print(f"{'Threshold':>12} {'Total dets':>12} {'Images with ≥1 det':>20}")
print("-" * 50)
for thr in SCORE_THRESHOLDS:
total_dets = 0
images_with_dets = 0
for r in results_list:
high_conf = (r['scores'] >= thr).sum().item()
total_dets += high_conf
if high_conf > 0:
images_with_dets += 1
print(f"{thr:>12.2f} {total_dets:>12d} {images_with_dets:>20d}/{EVAL_IMAGES}")
print()
print("Note: 0 detections at all thresholds is expected for a 5-step checkpoint.")
print("A converged model typically produces 10–50 detections per image on COCO val.")
Detection counts across 4 validation images
Threshold Total dets Images with ≥1 det
--------------------------------------------------
0.05 0 0/4
0.10 0 0/4
0.30 0 0/4
0.50 0 0/4
Note: 0 detections at all thresholds is expected for a 5-step checkpoint.
A converged model typically produces 10–50 detections per image on COCO val.
Found an error on this page? Report an issue.