early-stopping-callback

feature-engineering-helper

feature-importance-analyzer

gradient-clipping-helper

hyperparameter-tuner

learning-rate-scheduler

mlflow-tracking-setup

model-checkpoint-manager

model-explainability-tool

optuna-study-creator

pytorch-model-trainer

tensorflow-model-trainer

wandb-experiment-logger

This skill helps you systematically assess where Bitcoin sits in its market cycle — from extreme fear (accumulation opportunity) to extreme greed (distribution/exit signal). Through a weighted evaluat

Alibi explainability skill for counterfactual explanations, anchors, and trust scores.

LIME-based local explanation skill for individual predictions across tabular, text, and image data.

PyTorch model training skill with custom training loops, gradient management, and GPU optimization.

TensorFlow/Keras model training skill with callbacks, distributed strategies, and TensorBoard integration.

Agent-based modeling skill for simulating complex adaptive systems with heterogeneous interacting agents

Driver-based budgeting and forecasting skill with rolling forecast support and variance analysis

ASC 606 five-step model implementation skill for revenue recognition analysis and documentation

Analyze skill gaps and prioritize learning investments across the organization

Advanced econometric modeling for marketing effectiveness and budget optimization

Media interview preparation and crisis simulation tools for executive readiness and spokesperson development

Electric motor and drive unit design and optimization expertise

Multi-sensor fusion algorithms for perception in autonomous driving

Deep integration with vehicle dynamics simulation tools for handling, ride, and stability analysis

DeepVariant deep learning variant calling skill for high-accuracy SNV and indel detection

Deep integration with finite element analysis tools for structural simulation across static, dynamic, and nonlinear domains

Apply LDA, NMF, and other computational methods to discover patterns in large text corpora with appropriate parameter tuning

AI perception skill for sight, hearing, and threat detection systems.

Создание класса U-Net и функций encoder_block/decoder_block в PyTorch, соответствующих конкретным требованиям к структуре слоев (Conv->ReLU->Pool/Upsample) и логике skip-connections, чтобы пройти заданные проверки (assertions).

Defines a custom loss function in TensorFlow/Keras where predicting 1 as 0 (False Negative) has zero cost, while predicting 0 as 1 (False Positive) has a cost of 1.

Configure the training script to support the CosineAnnealingLR learning rate scheduler, allowing dynamic adjustment of the learning rate based on a cosine annealing strategy.

在PyTorch中实现基于中间特征图的知识蒸馏，通过添加一个可学习的网络层将学生模型的特征映射到教师模型特征空间，并在训练过程中联合优化学生模型和映射层的权重。

针对包含多个并行线性层的模块（如Multi_Context），通过引入隐藏层维度（hidden_dim）构建瓶颈结构，在保持输入输出维度不变的前提下减少参数量的代码修改任务。

针对引入新模块或调整模型结构后出现的训练损失下降缓慢、收敛困难及过拟合问题，提供学习率调整、预热策略、正则化配置及动态调度器使用的系统性解决方案。

针对树莓派等边缘设备，设计基于PyTorch的轻量级CNN模型，用于将5帧RAW图像融合为RGB图像。要求采用类UNet结构，集成注意力机制，并确保推理时延低于30ms。

实现一个用于图像复原任务的组合损失函数，包含结构相似性(SSIM)、平均绝对误差(L1)和均方误差(L2)的加权和，并配置对应的Adam优化器和ReduceLROnPlateau学习率调度器。

提供在Windows环境下使用TensorFlow for Java进行GPU计算的完整配置指南，包括Maven依赖、代码示例、CUDA/cuDNN版本兼容性检查，以及针对JDK 17的DLL加载路径设置方案。

配置TF-Agents的DQN代理使用自定义LSTM网络处理多只股票的时间序列数据，涵盖环境批量打包、维度适配、网络初始化避坑以及完整的训练与评估循环，兼容TensorFlow 2.10.1。

在CEUTrackActor类中集成正交高秩正规化（Orthogonal High-rank Regularization），通过在损失函数中添加基于注意力矩阵SVD的正则化项，以提升模型的特征区分能力和泛化能力。

实现一个PyTorch模块，用于动态融合RGB和Event特征，并包含正则化项以防止模型过度偏向某一模态。

Act as WagerGPT to analyze sports games using stats, injuries, and models to predict point spreads and over/under totals with win probabilities for ROI maximization.

Rephrases input text in the model's own words while preserving tone and intent, without adding any conversational filler, summaries, expansions, or introductory phrases.

Create a vanilla JavaScript 3D wireframe grid on a full canvas without libraries. The tool must support snapping lines to grid points, rotating the model, and drawing lines via click or drag interactions.

Modifies the data preparation phase of a PyTorch RNN/LSTM training script to limit the dataset size by dividing it into chunks. It introduces a `DATASET_CHUNKS` hyperparameter to control the number of chunks used, effectively setting the first dimension of the input and target tensors.

Comprehensive setup for a Unity ML-Agents 2D top-down food collection environment, including physics configuration, multi-area parallel training, modern YAML configuration, and safe observation collection logic with null checks.

Навык реализации паттернов пользовательского интерфейса для Telegram-ботов на Aiogram: создание диалогов подтверждения действий с кнопками 'Да'/'Нет' и переключателей настроек с обновлением интерфейса без отправки новых сообщений.

Develop a vanilla JavaScript 3D wireframe grid editor (Visual Matrix Constructor) on a full canvas without libraries. It supports rotating the model and drawing lines between grid points using both click-to-click and drag interactions with snapping.