2. Deild Women Promotion Group stats & predictions

Odds Analysis: Understanding Bookmaker Margins and Value Bets [0]: #!/usr/bin/env python [1]: # -*- coding:utf-8 -*- [2]: # [3]: # Copyright (c) 2020 Intel Corporation [4]: # [5]: # Licensed under the Apache License, Version 2.0 (the "License"); [6]: # you may not use this file except in compliance with the License. [7]: # You may obtain a copy of the License at [8]: # [9]: # http://www.apache.org/licenses/LICENSE-2.0 [10]: # [11]: # Unless required by applicable law or agreed to in writing, [12]: # software distributed under the License is distributed on an "AS IS" BASIS, [13]: # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. [14]: # See the License for the specific language governing permissions [15]: # and limitations under the License. [16]: import os [17]: import sys [18]: import math [19]: import cv2 [20]: import numpy as np [21]: sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) [22]: from extensions.ops.hard_nms import hard_nms_cpu [23]: def postprocess(prediction, [24]: num_classes, [25]: conf_thres=0.7, [26]: nms_thres=0.45, [27]: use_soft_nms=False): [28]: """ [29]: Removes some detections from prediction using confidence thresholding, [30]: non-maximum suppression (NMS) and optionally soft-NMS. [31]: Returns detections with shape: [32]: (x1, y1, x2, y2, object_conf, class_score, class_pred) [33]: """ [34]: conf_mask = (prediction[:, :, 4] > conf_thres).float().unsqueeze(2) prediction = prediction * conf_mask box_corner = prediction.new(prediction.shape) box_corner[:, :, :2] = prediction[:, :, :2] - prediction[:, :, 2:] / 2 box_corner[:, :, 2:] = prediction[:, :, :2] + prediction[:, :, 2:] prediction[:, :, :4] = box_corner[:, :, :4] if use_soft_nms: nms_out_index = soft_nms_cpu( prediction.cpu().numpy(), num_classes=num_classes, nms_thresh=nms_thres, method=1, sigma=0.5, min_score=conf_thres) else: nms_out_index = hard_nms_cpu( prediction.cpu().numpy(), num_classes=num_classes, conf_thres=conf_thres, nms_thresh=nms_thres) return nms_out_index ***** Tag Data ***** ID: 1 description: Post-processing function which involves complex operations including confidence thresholding, non-maximum suppression (NMS), optional soft-NMS application, bounding box adjustments using advanced tensor operations. start line: 23 end line: 86 dependencies: - type: Function name: hard_nms_cpu start line: 22 end line: 22 context description: This function processes raw predictions from an object detection model by filtering out low-confidence detections using confidence thresholding, converting bounding box formats from center coordinates plus dimensions (cx,cy,w,h) format into corner coordinates format (x1,y1,x2,y2), applying NMS or soft-NMS based on user input. algorithmic depth: 5 algorithmic depth external: N obscurity: 5 advanced coding concepts: 5 interesting for students: 5 self contained: N ************ ## Challenging aspects ### Challenging aspects in above code 1. **Confidence Thresholding**: The process of applying confidence thresholds involves intricate manipulation of tensors using PyTorch operations which must be done efficiently without introducing errors that could affect downstream processes. 2. **Bounding Box Conversion**: Converting bounding boxes from center coordinates plus dimensions format `(cx,cy,w,h)` into corner coordinates format `(x1,y1,x2,y2)` requires careful tensor manipulation ensuring no loss of precision. 3. **Non-Maximum Suppression (NMS) Implementation**: Implementing both hard-NMS and soft-NMS requires understanding how these algorithms work internally — particularly how they handle overlapping boxes differently — along with efficient tensor operations. ### Extension 1. **Multi-Scale Detection Handling**: Extend functionality so that it can handle multi-scale predictions typically generated by models like YOLOv5. 2. **Batch Processing**: Modify code so that it processes multiple images at once instead of one at a time. 3. **Class-Agnostic NMS**: Implement class-agnostic NMS where detections are filtered regardless of class labels. ## Exercise ### Problem Statement: You are tasked with extending an object detection post-processing pipeline based on [SNIPPET]. The extension should incorporate additional functionalities while ensuring efficient execution. ### Requirements: 1. **Multi-Scale Detection Handling**: - Modify `postprocess` function so that it accepts multi-scale predictions. - Each scale should have its own set of predictions which need separate processing before being merged. 2. **Batch Processing**: - Extend `postprocess` function so that it can handle batched inputs where each batch consists of multiple images. - Ensure that outputs maintain batch dimensions correctly. 3. **Class-Agnostic NMS**: - Implement an option within `postprocess` function where NMS is performed class-agnostically. - Ensure that when class-agnostic mode is enabled all detections across classes are considered together before filtering. ### Additional Constraints: 1. Ensure minimal computational overhead when handling batches. 2. Maintain high precision during bounding box conversions. 3. Efficiently merge multi-scale predictions ensuring no redundant computations. python # [SNIPPET] def postprocess(prediction_batches, num_classes, conf_thres=0.7, nms_thres=0.45, use_soft_nms=False, class_agnostic=False): """ Removes some detections from batched multi-scale predictions using confidence thresholding, non-maximum suppression (NMS) optionally soft-NMS or class agnostic NMS. Returns detections with shape: (batch_size x num_detections x (x1, y1, x2, y2, object_conf, class_score(s), class_pred(s))) """ batch_size = prediction_batches.size(0) processed_batches = [] for batch_idx in range(batch_size): all_predictions = [] for scale_idx in range(prediction_batches.size(1)): prediction = prediction_batches[batch_idx][scale_idx] conf_mask = (prediction[:, :, 4] > conf_thres).float().unsqueeze(2) prediction = prediction * conf_mask box_corner = prediction.new(prediction.shape) box_corner[:, :, :2] = prediction[:, :, :2] - prediction[:, :, 2:] / 2 box_corner[:, :, 2:] = prediction[:, :, :2] + prediction[:, :, 2:] prediction[:, :, :4] = box_corner[:, :, :4] all_predictions.append(prediction) merged_predictions = torch.cat(all_predictions) if use_soft_nms: nms_out_index = soft_nms_cpu( merged_predictions.cpu().numpy(), num_classes=num_classes if not class_agnostic else -1, nms_thresh=nms_thres, method