eval.py

from data import COCODetection, MEANS, COLORS, COCO_CLASSES
from yolact import Yolact
from utils.augmentations import BaseTransform, FastBaseTransform, Resize
from utils.functions import MovingAverage, ProgressBar
from layers.box_utils import jaccard, center_size
from utils import timer
from utils.functions import SavePath
from layers.output_utils import postprocess, undo_image_transformation
import pycocotools

from data import cfg, set_cfg, set_dataset

import numpy as np
import torch
import torch.backends.cudnn as cudnn
from torch.autograd import Variable
import argparse
import time
import random
import cProfile
import pickle
import json
import os
from pathlib import Path
from collections import OrderedDict
from PIL import Image

import matplotlib.pyplot as plt
import cv2

def str2bool(v):
    if v.lower() in ('yes', 'true', 't', 'y', '1'):
        return True
    elif v.lower() in ('no', 'false', 'f', 'n', '0'):
        return False
    else:
        raise argparse.ArgumentTypeError('Boolean value expected.')

def get_args():
    parser = argparse.ArgumentParser(description='YOLACT COCO Evaluation')
    parser.add_argument('--trained_model',
                        default='weights/ssd300_mAP_77.43_v2.pth', type=str,
                        help='Trained state_dict file path to open. If "interrupt", this will open the interrupt file.')
    parser.add_argument('--top_k', default=81, type=int,
                        help='Further restrict the number of predictions to parse')
    parser.add_argument('--cuda', default=True, type=str2bool,
                        help='Use cuda to evaulate model')
    parser.add_argument('--fast_nms', default=True, type=str2bool,
                        help='Whether to use a faster, but not entirely correct version of NMS.')
    parser.add_argument('--display_masks', default=True, type=str2bool,
                        help='Whether or not to display masks over bounding boxes')
    parser.add_argument('--display_bboxes', default=True, type=str2bool,
                        help='Whether or not to display bboxes around masks')
    parser.add_argument('--display_text', default=True, type=str2bool,
                        help='Whether or not to display text (class [score])')
    parser.add_argument('--display_scores', default=True, type=str2bool,
                        help='Whether or not to display scores in addition to classes')
    parser.add_argument('--display', dest='display', action='store_true',
                        help='Display qualitative results instead of quantitative ones.')
    parser.add_argument('--shuffle', dest='shuffle', action='store_true',
                        help='Shuffles the images when displaying them. Doesn\'t have much of an effect when display is off though.')
    parser.add_argument('--ap_data_file', default='results/ap_data.pkl', type=str,
                        help='In quantitative mode, the file to save detections before calculating mAP.')
    parser.add_argument('--resume', dest='resume', action='store_true',
                        help='If display not set, this resumes mAP calculations from the ap_data_file.')
    parser.add_argument('--max_images', default=-1, type=int,
                        help='The maximum number of images from the dataset to consider. Use -1 for all.')
    parser.add_argument('--output_coco_json', dest='output_coco_json', action='store_true',
                        help='If display is not set, instead of processing IoU values, this just dumps detections into the coco json file.')
    parser.add_argument('--bbox_det_file', default='results/bbox_detections.json', type=str,
                        help='The output file for coco bbox results if --coco_results is set.')
    parser.add_argument('--mask_det_file', default='results/mask_detections.json', type=str,
                        help='The output file for coco mask results if --coco_results is set.')
    parser.add_argument('--config', default=None,
                        help='The config object to use.')
    parser.add_argument('--output_web_json', dest='output_web_json', action='store_true',
                        help='If display is not set, instead of processing IoU values, this dumps detections for usage with the detections viewer web thingy.')
    parser.add_argument('--web_det_path', default='web/dets/', type=str,
                        help='If output_web_json is set, this is the path to dump detections into.')
    parser.add_argument('--no_bar', dest='no_bar', action='store_true',
                        help='Do not output the status bar. This is useful for when piping to a file.')
    parser.add_argument('--display_lincomb', default=False, type=str2bool,
                        help='If the config uses lincomb masks, output a visualization of how those masks are created.')
    parser.add_argument('--benchmark', default=False, dest='benchmark', action='store_true',
                        help='Equivalent to running display mode but without displaying an image.')
    parser.add_argument('--no_sort', default=False, dest='no_sort', action='store_true',
                        help='Do not sort images by hashed image ID.')
    parser.add_argument('--seed', default=None, type=int,
                        help='The seed to pass into random.seed. Note: this is only really for the shuffle and does not (I think) affect cuda stuff.')
    parser.add_argument('--mask_proto_debug', default=False, dest='mask_proto_debug', action='store_true',
                        help='Outputs stuff for scripts/compute_mask.py.')
    parser.add_argument('--no_crop', default=False, dest='crop', action='store_false',
                        help='Do not crop output masks with the predicted bounding box.')
    parser.add_argument('--image', default=None, type=str,
                        help='A path to an image to use for display.')
    parser.add_argument('--images', default=None, type=str,
                        help='An input folder of images and output folder to save detected images. Should be in the format input->output.')
    parser.add_argument('--video', default=None, type=str,
                        help='A path to a video to evaluate on.')
    parser.add_argument('--video_multiframe', default=1, type=int,
                        help='The number of frames to evaluate in parallel to make videos play at higher fps.')
    parser.add_argument('--score_threshold', default=0.15, type=float,
                        help='Detections with a score under this threshold will not be considered. This currently only works in display mode.')
    parser.add_argument('--dataset', default=None, type=str,
                        help='If specified, override the dataset specified in the config with this one (example: coco2017_dataset).')
    parser.add_argument('--detect', default=False, dest='detect', action='store_true',
                        help='Don\'t evauluate the mask branch at all and only do object detection. This only works for --display and --benchmark.')

    parser.set_defaults(no_bar=False, display=False, resume=False, output_coco_json=False, output_web_json=False, shuffle=False,
                        benchmark=False, no_sort=False, no_hash=False, mask_proto_debug=False, crop=True, detect=False)

    global args
    args = parser.parse_args('--trained_model=weights/yolact_base_54_800000.pth --score_threshold=0.3 --top_k=100 --image=my_image.png'.split())
    
    if args.output_web_json:
        args.output_coco_json = True

    if args.seed is not None:
        random.seed(args.seed)

    if args.config is not None:
        set_cfg(args.config)

    if args.detect:
        cfg.eval_mask_branch = True

iou_thresholds = [x / 100 for x in range(50, 100, 5)]
coco_cats = [] # Call prep_coco_cats to fill this
coco_cats_inv = {}

def prep_display(dets_out, img, gt, gt_masks, h, w, undo_transform=True, class_color=False):
    """
    Note: If undo_transform=False then im_h and im_w are allowed to be None.
    gt and gt_masks are also allowed to be none (until I reimplement that functionality).
    """
    if undo_transform:
        img_numpy = undo_image_transformation(img, w, h)
        img_gpu = torch.Tensor(img_numpy).cuda()
    else:
        img_gpu = img / 255.0
        h, w, _ = img.shape
    
    with timer.env('Postprocess'):
        t = postprocess(dets_out, w, h, visualize_lincomb=args.display_lincomb, crop_masks=args.crop, score_threshold=args.score_threshold)
        torch.cuda.synchronize()

    with timer.env('Copy'):
        if cfg.eval_mask_branch:
            masks = t[3][:args.top_k] # We'll need this later
        classes, scores, boxes = [x[:args.top_k].cpu().numpy() for x in t[:3]]
    
    if classes.shape[0] == 0:
        return (img_gpu * 255).byte().cpu().numpy()

    def get_color(j):
        color = COLORS[(classes[j] * 5 if class_color else j * 5) % len(COLORS)]
        if not undo_transform:
            color = (color[2], color[1], color[0])
        return color

    # Draw masks first on the gpu
    if args.display_masks and cfg.eval_mask_branch:
        for j in reversed(range(min(args.top_k, classes.shape[0]))):
            if scores[j] >= args.score_threshold:
                color = get_color(j)

                mask = masks[j, :, :, None]
                mask_color = mask @ (torch.Tensor(color).view(1, 3) / 255.0)
                mask_alpha = 0.45

                # Alpha only the region of the image that contains the mask
                img_gpu = img_gpu * (1 - mask) \
                        + img_gpu * mask * (1-mask_alpha) + mask_color * mask_alpha
        
    # Then draw the stuff that needs to be done on the cpu
    # Note, make sure this is a uint8 tensor or opencv will not anti alias text for whatever reason
    img_numpy = (img_gpu * 255).byte().cpu().numpy()
    
    if args.display_text or args.display_bboxes:
        for j in reversed(range(min(args.top_k, classes.shape[0]))):
            score = scores[j]

            if scores[j] >= args.score_threshold:
                x1, y1, x2, y2 = boxes[j, :]
                color = get_color(j)

                if args.display_bboxes:
                    cv2.rectangle(img_numpy, (x1, y1), (x2, y2), color, 1)

                if args.display_text:
                    _class = COCO_CLASSES[classes[j]]
                    text_str = '%s: %.2f' % (_class, score) if args.display_scores else _class

                    font_face = cv2.FONT_HERSHEY_DUPLEX
                    font_scale = 0.6
                    font_thickness = 1

                    text_w, text_h = cv2.getTextSize(text_str, font_face, font_scale, font_thickness)[0]

                    text_pt = (x1, y1 - 3)
                    text_color = [255, 255, 255]

                    cv2.rectangle(img_numpy, (x1, y1), (x1 + text_w, y1 - text_h - 4), color, -1)
                    cv2.putText(img_numpy, text_str, text_pt, font_face, font_scale, text_color, font_thickness, cv2.LINE_AA)
    
    return img_numpy

def prep_benchmark(dets_out, h, w):
    with timer.env('Postprocess'):
        t = postprocess(dets_out, w, h, crop_masks=args.crop, score_threshold=args.score_threshold)

    with timer.env('Copy'):
        classes, scores, boxes, masks = [x[:args.top_k].cpu().numpy() for x in t]
    
    with timer.env('Sync'):
        # Just in case
        torch.cuda.synchronize()

def prep_coco_cats(cats):
    """ Prepare inverted table for category id lookup given a coco cats object. """
    name_lookup = {}

    for _id, cat_obj in cats.items():
        name_lookup[cat_obj['name']] = _id

    # Bit of a roundabout way to do this but whatever
    for i in range(len(COCO_CLASSES)):
        coco_cats.append(name_lookup[COCO_CLASSES[i]])
        coco_cats_inv[coco_cats[-1]] = i


def get_coco_cat(transformed_cat_id):
    """ transformed_cat_id is [0,80) as indices in COCO_CLASSES """
    return coco_cats[transformed_cat_id]

def get_transformed_cat(coco_cat_id):
    """ transformed_cat_id is [0,80) as indices in COCO_CLASSES """
    return coco_cats_inv[coco_cat_id]


class Detections:

    def __init__(self):
        self.bbox_data = []
        self.mask_data = []

    def add_bbox(self, image_id:int, category_id:int, bbox:list, score:float):
        """ Note that bbox should be a list or tuple of (x1, y1, x2, y2) """
        bbox = [bbox[0], bbox[1], bbox[2]-bbox[0], bbox[3]-bbox[1]]

        # Round to the nearest 10th to avoid huge file sizes, as COCO suggests
        bbox = [round(float(x)*10)/10 for x in bbox]

        self.bbox_data.append({
            'image_id': int(image_id),
            'category_id': get_coco_cat(int(category_id)),
            'bbox': bbox,
            'score': float(score)
        })

    def add_mask(self, image_id:int, category_id:int, segmentation:np.ndarray, score:float):
        """ The segmentation should be the full mask, the size of the image and with size [h, w]. """
        rle = pycocotools.mask.encode(np.asfortranarray(segmentation.astype(np.uint8)))
        rle['counts'] = rle['counts'].decode('ascii') # json.dump doesn't like bytes strings

        self.mask_data.append({
            'image_id': int(image_id),
            'category_id': get_coco_cat(int(category_id)),
            'segmentation': rle,
            'score': float(score)
        })
    
    def dump(self):
        dump_arguments = [
            (self.bbox_data, args.bbox_det_file),
            (self.mask_data, args.mask_det_file)
        ]

        for data, path in dump_arguments:
            with open(path, 'w') as f:
                json.dump(data, f)
    
    def dump_web(self):
        """ Dumps it in the format for my web app. Warning: bad code ahead! """
        config_outs = ['preserve_aspect_ratio', 'use_prediction_module',
                        'use_yolo_regressors', 'use_prediction_matching',
                        'train_masks']

        output = {
            'info' : {
                'Config': {key: getattr(cfg, key) for key in config_outs},
            }
        }

        image_ids = list(set([x['image_id'] for x in self.bbox_data]))
        image_ids.sort()
        image_lookup = {_id: idx for idx, _id in enumerate(image_ids)}

        output['images'] = [{'image_id': image_id, 'dets': []} for image_id in image_ids]

        # These should already be sorted by score with the way prep_metrics works.
        for bbox, mask in zip(self.bbox_data, self.mask_data):
            image_obj = output['images'][image_lookup[bbox['image_id']]]
            image_obj['dets'].append({
                'score': bbox['score'],
                'bbox': bbox['bbox'],
                'category': COCO_CLASSES[get_transformed_cat(bbox['category_id'])],
                'mask': mask['segmentation'],
            })

        with open(os.path.join(args.web_det_path, '%s.json' % cfg.name), 'w') as f:
            json.dump(output, f)
        

        

def mask_iou(mask1, mask2, iscrowd=False):
    """
    Inputs inputs are matricies of size _ x N. Output is size _1 x _2.
    Note: if iscrowd is True, then mask2 should be the crowd.
    """
    timer.start('Mask IoU')

    intersection = torch.matmul(mask1, mask2.t())
    area1 = torch.sum(mask1, dim=1).view(1, -1)
    area2 = torch.sum(mask2, dim=1).view(1, -1)
    union = (area1.t() + area2) - intersection

    if iscrowd:
        # Make sure to brodcast to the right dimension
        ret = intersection / area1.t()
    else:
        ret = intersection / union
    timer.stop('Mask IoU')
    return ret.cpu()

def bbox_iou(bbox1, bbox2, iscrowd=False):
    with timer.env('BBox IoU'):
        ret = jaccard(bbox1, bbox2, iscrowd)
    return ret.cpu()

def prep_metrics(ap_data, dets, img, gt, gt_masks, h, w, num_crowd, image_id, detections:Detections=None):
    """ Returns a list of APs for this image, with each element being for a class  """
    if not args.output_coco_json:
        with timer.env('Prepare gt'):
            gt_boxes = torch.Tensor(gt[:, :4])
            gt_boxes[:, [0, 2]] *= w
            gt_boxes[:, [1, 3]] *= h
            gt_classes = list(gt[:, 4].astype(int))
            gt_masks = torch.Tensor(gt_masks).view(-1, h*w)

            if num_crowd > 0:
                split = lambda x: (x[-num_crowd:], x[:-num_crowd])
                crowd_boxes  , gt_boxes   = split(gt_boxes)
                crowd_masks  , gt_masks   = split(gt_masks)
                crowd_classes, gt_classes = split(gt_classes)

    with timer.env('Postprocess'):
        classes, scores, boxes, masks = postprocess(dets, w, h, crop_masks=args.crop, score_threshold=args.score_threshold)

        if classes.size(0) == 0:
            return

        classes = list(classes.cpu().numpy().astype(int))
        scores = list(scores.cpu().numpy().astype(float))
        masks = masks.view(-1, h*w).cuda()
        boxes = boxes.cuda()


    if args.output_coco_json:
        with timer.env('JSON Output'):
            boxes = boxes.cpu().numpy()
            masks = masks.view(-1, h, w).cpu().numpy()
            for i in range(masks.shape[0]):
                # Make sure that the bounding box actually makes sense and a mask was produced
                if (boxes[i, 3] - boxes[i, 1]) * (boxes[i, 2] - boxes[i, 0]) > 0:
                    detections.add_bbox(image_id, classes[i], boxes[i,:],   scores[i])
                    detections.add_mask(image_id, classes[i], masks[i,:,:], scores[i])
            return
    
    with timer.env('Eval Setup'):
        num_pred = len(classes)
        num_gt   = len(gt_classes)

        mask_iou_cache = mask_iou(masks, gt_masks)
        bbox_iou_cache = bbox_iou(boxes.float(), gt_boxes.float())

        if num_crowd > 0:
            crowd_mask_iou_cache = mask_iou(masks, crowd_masks, iscrowd=True)
            crowd_bbox_iou_cache = bbox_iou(boxes.float(), crowd_boxes.float(), iscrowd=True)
        else:
            crowd_mask_iou_cache = None
            crowd_bbox_iou_cache = None

        iou_types = [
            ('box',  lambda i,j: bbox_iou_cache[i, j].item(), lambda i,j: crowd_bbox_iou_cache[i,j].item()),
            ('mask', lambda i,j: mask_iou_cache[i, j].item(), lambda i,j: crowd_mask_iou_cache[i,j].item())
        ]

    timer.start('Main loop')
    for _class in set(classes + gt_classes):
        ap_per_iou = []
        num_gt_for_class = sum([1 for x in gt_classes if x == _class])
        
        for iouIdx in range(len(iou_thresholds)):
            iou_threshold = iou_thresholds[iouIdx]

            for iou_type, iou_func, crowd_func in iou_types:
                gt_used = [False] * len(gt_classes)
                
                ap_obj = ap_data[iou_type][iouIdx][_class]
                ap_obj.add_gt_positives(num_gt_for_class)

                for i in range(num_pred):
                    if classes[i] != _class:
                        continue
                    
                    max_iou_found = iou_threshold
                    max_match_idx = -1
                    for j in range(num_gt):
                        if gt_used[j] or gt_classes[j] != _class:
                            continue
                            
                        iou = iou_func(i, j)

                        if iou > max_iou_found:
                            max_iou_found = iou
                            max_match_idx = j
                    
                    if max_match_idx >= 0:
                        gt_used[max_match_idx] = True
                        ap_obj.push(scores[i], True)
                    else:
                        # If the detection matches a crowd, we can just ignore it
                        matched_crowd = False

                        if num_crowd > 0:
                            for j in range(len(crowd_classes)):
                                if crowd_classes[j] != _class:
                                    continue
                                
                                iou = crowd_func(i, j)

                                if iou > iou_threshold:
                                    matched_crowd = True
                                    break

                        # All this crowd code so that we can make sure that our eval code gives the
                        # same result as COCOEval. There aren't even that many crowd annotations to
                        # begin with, but accuracy is of the utmost importance.
                        if not matched_crowd:
                            ap_obj.push(scores[i], False)
    timer.stop('Main loop')


class APDataObject:
    """
    Stores all the information necessary to calculate the AP for one IoU and one class.
    Note: I type annotated this because why not.
    """

    def __init__(self):
        self.data_points = []
        self.num_gt_positives = 0

    def push(self, score:float, is_true:bool):
        self.data_points.append((score, is_true))
    
    def add_gt_positives(self, num_positives:int):
        """ Call this once per image. """
        self.num_gt_positives += num_positives

    def is_empty(self) -> bool:
        return len(self.data_points) == 0 and self.num_gt_positives == 0

    def get_ap(self) -> float:
        """ Warning: result not cached. """

        if self.num_gt_positives == 0:
            return 0

        # Sort descending by score
        self.data_points.sort(key=lambda x: -x[0])

        precisions = []
        recalls    = []
        num_true  = 0
        num_false = 0

        # Compute the precision-recall curve. The x axis is recalls and the y axis precisions.
        for datum in self.data_points:
            # datum[1] is whether the detection a true or false positive
            if datum[1]: num_true += 1
            else: num_false += 1
            
            precision = num_true / (num_true + num_false)
            recall    = num_true / self.num_gt_positives

            precisions.append(precision)
            recalls.append(recall)

        # Smooth the curve by computing [max(precisions[i:]) for i in range(len(precisions))]
        # Basically, remove any temporary dips from the curve.
        # At least that's what I think, idk. COCOEval did it so I do too.
        for i in range(len(precisions)-1, 0, -1):
            if precisions[i] > precisions[i-1]:
                precisions[i-1] = precisions[i]

        # Compute the integral of precision(recall) d_recall from recall=0->1 using fixed-length riemann summation with 101 bars.
        y_range = [0] * 101 # idx 0 is recall == 0.0 and idx 100 is recall == 1.00
        x_range = np.array([x / 100 for x in range(101)])
        recalls = np.array(recalls)

        # I realize this is weird, but all it does is find the nearest precision(x) for a given x in x_range.
        # Basically, if the closest recall we have to 0.01 is 0.009 this sets precision(0.01) = precision(0.009).
        # I approximate the integral this way, because that's how COCOEval does it.
        indices = np.searchsorted(recalls, x_range, side='left')
        for bar_idx, precision_idx in enumerate(indices):
            if precision_idx < len(precisions):
                y_range[bar_idx] = precisions[precision_idx]

        # Finally compute the riemann sum to get our integral.
        # avg([precision(x) for x in 0:0.01:1])
        return sum(y_range) / len(y_range)

def badhash(x):
    """
    Just a quick and dirty hash function for doing a deterministic shuffle based on image_id.

    Source:
    https://stackoverflow.com/questions/664014/what-integer-hash-function-are-good-that-accepts-an-integer-hash-key
    """
    x = (((x >> 16) ^ x) * 0x45d9f3b) & 0xFFFFFFFF
    x = (((x >> 16) ^ x) * 0x45d9f3b) & 0xFFFFFFFF
    x = ((x >> 16) ^ x) & 0xFFFFFFFF
    return x

def evalimage(net:Yolact, path:str, save_path:str=None):
    frame = torch.Tensor(cv2.imread(path)).cuda().float()
    batch = FastBaseTransform()(frame.unsqueeze(0))
    preds = net(batch)

    img_numpy = prep_display(preds, frame, None, None, None, None, undo_transform=False)
    
    if save_path is None:
        img_numpy = img_numpy[:, :, (2, 1, 0)]

    if save_path is None:
        plt.imshow(img_numpy)
        plt.title(path)
        plt.show()
    else:
        cv2.imwrite(save_path, img_numpy)

def evalimages(net:Yolact, input_folder:str, output_folder:str):
    if not os.path.exists(output_folder):
        os.mkdir(output_folder)

    print()
    for p in Path(input_folder).glob('*'): 
        path = str(p)
        name = os.path.basename(path)
        name = '.'.join(name.split('.')[:-1]) + '.png'
        out_path = os.path.join(output_folder, name)

        evalimage(net, path, out_path)
        print(path + ' -> ' + out_path)
    print('Done.')

from multiprocessing.pool import ThreadPool

def evalvideo(net:Yolact, path:str):
    vid = cv2.VideoCapture(path)
    transform = FastBaseTransform()
    frame_times = MovingAverage()
    fps = 0
    frame_time_target = 1 / vid.get(cv2.CAP_PROP_FPS)

    def cleanup_and_exit():
        print()
        pool.terminate()
        vid.release()
        cv2.destroyAllWindows()
        exit()

    def get_next_frame(vid):
        return [vid.read()[1] for _ in range(args.video_multiframe)]

    def transform_frame(frames):
        with torch.no_grad():
            frames = [torch.Tensor(frame).float().cuda() for frame in frames]
            return frames, transform(torch.stack(frames, 0))

    def eval_network(inp):
        with torch.no_grad():
            frames, imgs = inp
            return frames, net(imgs)

    def prep_frame(inp):
        with torch.no_grad():
            frame, preds = inp
            return prep_display(preds, frame, None, None, None, None, undo_transform=False, class_color=True)

    extract_frame = lambda x, i: (x[0][i], [x[1][i]])

    # Prime the network on the first frame because I do some thread unsafe things otherwise
    print('Initializing model... ', end='')
    eval_network(transform_frame(get_next_frame(vid)))
    print('Done.')

    # For each frame the sequence of functions it needs to go through to be processed (in reversed order)
    sequence = [prep_frame, eval_network, transform_frame]
    pool = ThreadPool(processes=len(sequence) + args.video_multiframe)

    active_frames = []

    print()
    while vid.isOpened():
        start_time = time.time()

        # Start loading the next frames from the disk
        next_frames = pool.apply_async(get_next_frame, args=(vid,))
        
        # For each frame in our active processing queue, dispatch a job
        # for that frame using the current function in the sequence
        for frame in active_frames:
            frame['value'] = pool.apply_async(sequence[frame['idx']], args=(frame['value'],))
        
        # For each frame whose job was the last in the sequence (i.e. for all final outputs)
        for frame in active_frames:
            if frame['idx'] == 0:
                # Wait here so that the frame has time to process and so that the video plays at the proper speed
                time.sleep(frame_time_target)

                cv2.imshow(path, frame['value'].get())
                if cv2.waitKey(1) == 27: # Press Escape to close
                    cleanup_and_exit()

        # Remove the finished frames from the processing queue
        active_frames = [x for x in active_frames if x['idx'] > 0]

        # Finish evaluating every frame in the processing queue and advanced their position in the sequence
        for frame in list(reversed(active_frames)):
            frame['value'] = frame['value'].get()
            frame['idx'] -= 1

            if frame['idx'] == 0:
                # Split this up into individual threads for prep_frame since it doesn't support batch size
                active_frames += [{'value': extract_frame(frame['value'], i), 'idx': 0} for i in range(1, args.video_multiframe)]
                frame['value'] = extract_frame(frame['value'], 0)

        
        # Finish loading in the next frames and add them to the processing queue
        active_frames.append({'value': next_frames.get(), 'idx': len(sequence)-1})
        
        # Compute FPS
        frame_times.add(time.time() - start_time)
        fps = args.video_multiframe / frame_times.get_avg()

        print('\rAvg FPS: %.2f     ' % fps, end='')
    
    cleanup_and_exit()

def savevideo(net:Yolact, in_path:str, out_path:str):

    vid = cv2.VideoCapture(in_path)

    target_fps   = round(vid.get(cv2.CAP_PROP_FPS))
    frame_width  = round(vid.get(cv2.CAP_PROP_FRAME_WIDTH))
    frame_height = round(vid.get(cv2.CAP_PROP_FRAME_HEIGHT))
    num_frames   = round(vid.get(cv2.CAP_PROP_FRAME_COUNT))
    
    out = cv2.VideoWriter(out_path, cv2.VideoWriter_fourcc(*"mp4v"), target_fps, (frame_width, frame_height))

    transform = FastBaseTransform()
    frame_times = MovingAverage()
    progress_bar = ProgressBar(30, num_frames)

    try:
        for i in range(num_frames):
            timer.reset()
            with timer.env('Video'):
                frame = torch.Tensor(vid.read()[1]).float().cuda()
                batch = transform(frame.unsqueeze(0))
                preds = net(batch)
                processed = prep_display(preds, frame, None, None, None, None, undo_transform=False, class_color=True)

                out.write(processed)
            
            if i > 1:
                frame_times.add(timer.total_time())
                fps = 1 / frame_times.get_avg()
                progress = (i+1) / num_frames * 100
                progress_bar.set_val(i+1)

                print('\rProcessing Frames  %s %6d / %6d (%5.2f%%)    %5.2f fps        '
                    % (repr(progress_bar), i+1, num_frames, progress, fps), end='')
    except KeyboardInterrupt:
        print('Stopping early.')
    
    vid.release()
    out.release()
    print()


def evaluate(net:Yolact, dataset, train_mode=False):
    net.detect.use_fast_nms = args.fast_nms
    cfg.mask_proto_debug = args.mask_proto_debug

    if args.image is not None:
        if ':' in args.image:
            inp, out = args.image.split(':')
            evalimage(net, inp, out)
        else:
            evalimage(net, args.image)
        return
    elif args.images is not None:
        inp, out = args.images.split(':')
        evalimages(net, inp, out)
        return
    elif args.video is not None:
        if ':' in args.video:
            inp, out = args.video.split(':')
            savevideo(net, inp, out)
        else:
            evalvideo(net, args.video)
        return

    frame_times = MovingAverage()
    dataset_size = len(dataset) if args.max_images < 0 else min(args.max_images, len(dataset))
    progress_bar = ProgressBar(30, dataset_size)

    print()

    if not args.display and not args.benchmark:
        # For each class and iou, stores tuples (score, isPositive)
        # Index ap_data[type][iouIdx][classIdx]
        ap_data = {
            'box' : [[APDataObject() for _ in COCO_CLASSES] for _ in iou_thresholds],
            'mask': [[APDataObject() for _ in COCO_CLASSES] for _ in iou_thresholds]
        }
        detections = Detections()
    else:
        timer.disable('Load Data')

    dataset_indices = list(range(len(dataset)))
    
    if args.shuffle:
        random.shuffle(dataset_indices)
    elif not args.no_sort:
        # Do a deterministic shuffle based on the image ids
        #
        # I do this because on python 3.5 dictionary key order is *random*, while in 3.6 it's
        # the order of insertion. That means on python 3.6, the images come in the order they are in
        # in the annotations file. For some reason, the first images in the annotations file are
        # the hardest. To combat this, I use a hard-coded hash function based on the image ids
        # to shuffle the indices we use. That way, no matter what python version or how pycocotools
        # handles the data, we get the same result every time.
        hashed = [badhash(x) for x in dataset.ids]
        dataset_indices.sort(key=lambda x: hashed[x])

    dataset_indices = dataset_indices[:dataset_size]

    try:
        # Main eval loop
        for it, image_idx in enumerate(dataset_indices):
            timer.reset()

            with timer.env('Load Data'):
                img, gt, gt_masks, h, w, num_crowd = dataset.pull_item(image_idx)

                # Test flag, do not upvote
                if cfg.mask_proto_debug:
                    with open('scripts/info.txt', 'w') as f:
                        f.write(str(dataset.ids[image_idx]))
                    np.save('scripts/gt.npy', gt_masks)

                batch = Variable(img.unsqueeze(0))
                if args.cuda:
                    batch = batch.cuda()

            with timer.env('Network Extra'):
                preds = net(batch)

            # Perform the meat of the operation here depending on our mode.
            if args.display:
                img_numpy = prep_display(preds, img, gt, gt_masks, h, w)
            elif args.benchmark:
                prep_benchmark(preds, h, w)
            else:
                prep_metrics(ap_data, preds, img, gt, gt_masks, h, w, num_crowd, dataset.ids[image_idx], detections)
            
            # First couple of images take longer because we're constructing the graph.
            # Since that's technically initialization, don't include those in the FPS calculations.
            if it > 1:
                frame_times.add(timer.total_time())
            
            if args.display:
                if it > 1:
                    print('Avg FPS: %.4f' % (1 / frame_times.get_avg()))
                plt.imshow(img_numpy)
                plt.title(str(dataset.ids[image_idx]))
                plt.show()
            elif not args.no_bar:
                if it > 1: fps = 1 / frame_times.get_avg()
                else: fps = 0
                progress = (it+1) / dataset_size * 100
                progress_bar.set_val(it+1)
                print('\rProcessing Images  %s %6d / %6d (%5.2f%%)    %5.2f fps        '
                    % (repr(progress_bar), it+1, dataset_size, progress, fps), end='')



        if not args.display and not args.benchmark:
            print()
            if args.output_coco_json:
                print('Dumping detections...')
                if args.output_web_json:
                    detections.dump_web()
                else:
                    detections.dump()
            else:
                if not train_mode:
                    print('Saving data...')
                    with open(args.ap_data_file, 'wb') as f:
                        pickle.dump(ap_data, f)

                return calc_map(ap_data)
        elif args.benchmark:
            print()
            print()
            print('Stats for the last frame:')
            timer.print_stats()
            avg_seconds = frame_times.get_avg()
            print('Average: %5.2f fps, %5.2f ms' % (1 / frame_times.get_avg(), 1000*avg_seconds))

    except KeyboardInterrupt:
        print('Stopping...')


def calc_map(ap_data):
    print('Calculating mAP...')
    aps = [{'box': [], 'mask': []} for _ in iou_thresholds]

    for _class in range(len(COCO_CLASSES)):
        for iou_idx in range(len(iou_thresholds)):
            for iou_type in ('box', 'mask'):
                ap_obj = ap_data[iou_type][iou_idx][_class]

                if not ap_obj.is_empty():
                    aps[iou_idx][iou_type].append(ap_obj.get_ap())

    all_maps = {'box': OrderedDict(), 'mask': OrderedDict()}

    # Looking back at it, this code is really hard to read :/
    for iou_type in ('box', 'mask'):
        all_maps[iou_type]['all'] = 0 # Make this first in the ordereddict
        for i, threshold in enumerate(iou_thresholds):
            mAP = sum(aps[i][iou_type]) / len(aps[i][iou_type]) * 100 if len(aps[i][iou_type]) > 0 else 0
            all_maps[iou_type][int(threshold*100)] = mAP
        all_maps[iou_type]['all'] = (sum(all_maps[iou_type].values()) / (len(all_maps[iou_type].values())-1))
    
    print_maps(all_maps)
    return all_maps

def print_maps(all_maps):
    # Warning: hacky 
    make_row = lambda vals: (' %5s |' * len(vals)) % tuple(vals)
    make_sep = lambda n:  ('-------+' * n)

    print()
    print(make_row([''] + [('.%d ' % x if isinstance(x, int) else x + ' ') for x in all_maps['box'].keys()]))
    print(make_sep(len(all_maps['box']) + 1))
    for iou_type in ('box', 'mask'):
        print(make_row([iou_type] + ['%.2f' % x for x in all_maps[iou_type].values()]))
    print(make_sep(len(all_maps['box']) + 1))
    print()