RuntimeError: value cannot be converted to type int without overflow

[bugramurat]

🐛 Bug

I'm having an error while trying to calculate mAP for my Faster-RCNN object detection model with custom dataset in pytorch while testing.

To Reproduce

Error message:

Traceback (most recent call last):
  File "c:\Users\lemon\Desktop\ap_py_2\inference.py", line 409, in <module> print(metric.compute())
  File "C:\Users\lemon\miniconda3\envs\cnn-env-03\lib\site-packages\torchmetrics\metric.py", line 531, in wrapped_func value = compute(*args, **kwargs)
  File "c:\Users\lemon\Desktop\ap_py_2\mean_ap.py", line 861, in compute precisions, recalls = self._calculate(classes)
  File "c:\Users\lemon\Desktop\ap_py_2\mean_ap.py", line 715, in _calculate eval_imgs = [
  File "c:\Users\lemon\Desktop\ap_py_2\mean_ap.py", line 716, in <listcomp> self._evaluate_image(img_id, class_id, area, max_detections, ious)
  File "c:\Users\lemon\Desktop\ap_py_2\mean_ap.py", line 563, in _evaluate_image return self.__evaluate_image_preds_no_gt(det, idx, det_label_mask, max_det, area_range, nb_iou_thrs)
  File "c:\Users\lemon\Desktop\ap_py_2\mean_ap.py", line 518, in __evaluate_image_preds_no_gt det_ignore_area = (det_areas < area_range[0]) | (det_areas > area_range[1])
RuntimeError: value cannot be converted to type int without overflow

Code sample (inference.py)

import numpy as np
import cv2
import torch
import glob as glob
import os
import time
from copy import deepcopy
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
from model import create_model
from config import (
    NUM_CLASSES, DEVICE, CLASSES, RESIZE_TO, NUM_WORKERS
)

COLORS = np.random.uniform(130, 255, size=(len(CLASSES), 3))
from datasets import (
    create_test_dataset, create_test_loader
)

from mean_ap import MeanAveragePrecision


model = create_model(num_classes=NUM_CLASSES)
checkpoint = torch.load('C:\\Users\\lemon\\Desktop\\outputs_3\\last_model.pth', map_location=DEVICE)
model.load_state_dict(checkpoint['model_state_dict'])
model.to(DEVICE).eval()

DIR_TEST = 'C:\\Users\\lemon\\Desktop\\ap\\OPIXray\\map_test_org'
test_images = glob.glob(f"{DIR_TEST}\\*.jpg")
print(f"Test instances: {len(test_images)}")


detection_threshold = 0.2

frame_count = 0

total_fps = 0 


targets = {}
targets['boxes'] = []
targets['labels'] = []
targets['image_id'] = []
test_dataset = create_test_dataset()
test_data_loader = create_test_loader(test_dataset, NUM_WORKERS)
for data in test_data_loader.dataset:
    for j in range(0,4):
        data[1]['boxes'][0][j] = int(data[1]['boxes'][0][j])
        
    targets['boxes'].append(data[1]['boxes'])
    targets['labels'].append(data[1]['labels'])
    targets['image_id'].append(data[1]['image_id'])

pred_scores = []
pred_boxes = []
pred_classes = []

for i in range(len(test_images)):
    # get the image file name for saving output later on
    image_name = test_images[i].split(os.path.sep)[-1].split('.')[0]
    image = cv2.imread(test_images[i])
    orig_image = image.copy()
    # BGR to RGB
    image = cv2.cvtColor(orig_image, cv2.COLOR_BGR2RGB).astype(np.float32)

    image_resized = cv2.resize(image, (RESIZE_TO, RESIZE_TO))
    orig_image = image_resized.copy()

    # make the pixel range between 0 and 1
    image_resized /= 255.0
    # bring color channels to front
    image_resized = np.transpose(image_resized, (2, 0, 1)).astype(np.float32)
    # convert to tensor
    image_resized = torch.tensor(image_resized, dtype=torch.float).cuda()
    # add batch dimension
    image_resized = torch.unsqueeze(image_resized, 0)
    start_time = time.time()
    with torch.no_grad():
        outputs = model(image_resized.to(DEVICE))
    end_time = time.time()
    # get the current fps
    fps = 1 / (end_time - start_time)
    # add `fps` to `total_fps`
    total_fps += fps
    # increment frame count
    frame_count += 1
    # load all detection to CPU for further operations
    outputs = [{k: v.to('cpu') for k, v in t.items()} for t in outputs]
    # carry further only if there are detected boxes
    if len(outputs[0]['boxes']) != 0:
        boxes = outputs[0]['boxes'].data.numpy()
        scores = outputs[0]['scores'].data.numpy()
        classes = outputs[0]['labels'].data.numpy()
        # filter out boxes according to `detection_threshold`
        boxes = boxes[scores >= detection_threshold].astype(np.int32)
        draw_boxes = boxes.copy()
        # get all the predicited class names
        # pred_classes.append(np.array([CLASSES[i] for i in outputs[0]['labels'].cpu().numpy()]))
        classes = classes[scores >= detection_threshold]
        scores = scores[scores >= detection_threshold]
        
        pred_scores.append(scores)
        pred_boxes.append(boxes)
        pred_classes.append(classes)

        ground = targets['boxes']
        # print(pred)
        # print(scores)
        # print(ground[0].cpu().detach().numpy()) 


    else:
        pred_scores.append(np.array([0]))
        pred_boxes.append(np.array([[0,0,0,0]]))
        pred_classes.append(np.array([0]))
    print(f"Image {i+1} done...")
    print('-'*50)
print('\nTEST PREDICTIONS COMPLETE')

avg_fps = total_fps / frame_count
print(f"Average FPS: {avg_fps:.3f}\n")  
   
send_pred_map = []
send_ground_map = []

for i in range(15):
    a = torch.from_numpy(pred_scores[i])
    b = targets['labels'][i]
    c = torch.from_numpy(np.array(pred_boxes[i]))
    d = torch.from_numpy(targets['boxes'][i].numpy())
    e = torch.from_numpy(pred_classes[i])
    print(i)
    send_pred_map.append({"boxes": c, "scores": a, "labels": e})
    send_ground_map.append({"boxes": d, "labels": b})


metric = MeanAveragePrecision()
metric.update(send_pred_map, send_ground_map)
print(metric.compute())

Expected behavior

I just want to calculate mAP for my model in pytorch.

Environment

• TorchMetrics version: 0.11.3 with pip in conda env
• Python & PyTorch Version: 3.9.16 & 1.13.1
• OS: Windows 11

[github-actions]

Hi! thanks for your contribution!, great first issue!

[SkafteNicki]

Hi @bugramurat, thanks for reporting this issue.
Would it be possible for you to make dump of send_pred_map and send_ground_map (using torch.save for example) in your code and send it to us? That is the essential part that torchmetrics is evaluated on.
I cannot reproduce the error, because I am unable to run your script without model + data.