PyTorch implementation of beam search decoding for seq2seq models

Beam search decoding with Pytorch

The implemention of Beam-Search decoding based on pytorch

An end-to-end speech recognition system built with PyTorch that utilizes Connectionist Temporal Classification (CTC) loss for training and beam search decoding for inference. This project implements a bidirectional LSTM architecture to transform audio features into text transcriptions.

FAQ CHATBOT using pytorch LSTM Encoder-Decoder model along with beam search and greedy search

CUDA Beam Search implementation for CTC Decoding with PyTorch

Scratch implementation of greedy and beam search decoding for PyTorch's encoder-decoder model.

Diverse CTC Beam Search Decoder for PyTorch based on Paddle Paddle's implementation

character level decoder only transformer implemented in pytorch uses beam search to generate the output text

A Transformer-based seq2seq model for French-to-English translation using PyTorch. Includes tokenization, training, beam search decoding, and BLEU score evaluation.

A Seq2Seq chatbot using LSTM encoder-decoder with attention, pretrained GloVe embeddings, and beam search decoding. Built on the DailyDialog dataset as a proof-of-concept for dialog modeling with PyTorch.

Transformer architecture implemented from scratch using PyTorch — Multi-Head Attention, Encoder-Decoder, Beam Search, Mixed Precision, LR Scheduler, SST-2 fine-tuning.

PyTorch re-implementation of "Generating Sentences from a Continuous Space" by Bowman et al. (2015), using BART encoder and GRU decoder with beam search.

An end-to-end deep learning image captioning system using ResNet-50 encoder and LSTM decoder, trained on 31,000+ Flickr30k images. Features greedy & beam search decoding with an interactive Streamlit web app. Built with PyTorch.

From-scratch PyTorch implementation of the encoder-decoder Transformer (Vaswani et al., 2017) for English→Italian neural machine translation. Features shared BPE tokenisation, paper-exact training schedule, mixed-precision training, and beam search decoding with BLEU evaluation.

A deep learning-based image captioning system built with Vision Transformer (ViT) and Transformer Decoder in PyTorch, capable of generating natural language descriptions for images using beam search.

A text summarization system built using T5 and PyTorch, trained on the CNN/DailyMail news dataset. The model generates concise summaries from long articles using beam search decoding and is evaluated using ROUGE metrics. This project demonstrates practical experience with encoder-decoder Transformers

Neural Storyteller: An end-to-end image captioning system using a ResNet-50 CNN encoder and LSTM decoder in PyTorch. Features BLEU-4 evaluation and Greedy/Beam search inference.

Full re-implementation of core LLaMA 2 building blocks in PyTorch with clear, well-commented code and math notes. Includes multiple decoding strategies (greedy, beam search, temperature, top-k, nucleus/top-p, random sampling). PDF slides included.

MiniBART Transformer model for generating text conclusions from descriptions. Includes data preprocessing, class balancing with SMOTE, tokenization, training, beam search decoding, evaluation with BLEU/ROUGE/BERTScore, and attention visualization. Requires PyTorch for BERTScore and TensorFlow for the model.

Fine-tuned Facebook’s mBART-large-50 for high-accuracy translation using PyTorch, Transformers, Pandas. Pre-processing included lowercasing, punctuation/digit stripping, length filtering, and BPE tokenization with language markers. Training utilized Adam with warm-up + decay, early stopping, and beam-search decoding.

Xây dựng từ đầu hệ thống Neural Machine Translation (English → French) bằng PyTorch với kiến trúc Encoder–Decoder LSTM và Luong Attention. Triển khai pipeline NLP gồm tokenization (spaCy), xây dựng vocabulary, batching với padding động, training bằng teacher forcing, beam search decoding và đánh giá bằng BLEU score trên tập test

Image captioning with a CNN encoder and LSTM decoder trained on Flickr30k. ResNet-50 extracts visual features; a beam-search-guided LSTM generates natural-language descriptions. Built with PyTorch and deployed as a Streamlit web app.

This PyTorch-based image captioning model uses ResNet-50 encoder and Transformer decoder to generate descriptive captions from Flickr8k images. Features include data augmentation (image transforms, synonym replacement), training with AdamW and early stopping, and inference via greedy, beam search (k=3,5,10), or nucleus sampling.