Algorithms from Scratch in TensorFlow is a learning-oriented collection of computer-vision, natural-language-processing, and reinforcement-learning implementations built directly from TensorFlow operations and model components. The goal is to expose the mechanics behind an algorithm—its tensors, losses, updates, data flow, and runtime behaviour—rather than hide it behind a task-specific high-level API.
The repository is source-first. Datasets, checkpoints, TensorBoard events, generated figures, caches, and model exports are intentionally local-only.
TensorFlow is used here as a numerical and automatic-differentiation runtime:
tf.Tensor for vectorized computation, tf.data for input pipelines,
tf.GradientTape for custom learning steps, and tf.distribute where a
training family supports multiple devices. The implementations make it possible
to trace mathematical ideas into executable tensor programs and inspect their
practical trade-offs: batching, broadcasting, memory, stability, and compute.
flowchart TD
A[Algorithms from Scratch in TensorFlow]
A --> CV[Computer Vision]
A --> NLP[Natural Language Processing]
A --> RL[Reinforcement Learning]
CV --> CVD[Diffusion]
CV --> CVG[Generative Adversarial Networks]
CV --> CVC[Image Classification]
CV --> CVO[Object Detection]
NLP --> NLPE[Word Embeddings]
NLP --> NLPS[Sequence and Seq2Seq Models]
NLP --> NLPP[RedditStory Project]
RL --> RLP[Policy-gradient and REINFORCE]
RL --> RLQ[Vectorized DQN and Double DQN]
RL --> RLG[Game agents]
| Area | Algorithms and architectures | TensorFlow techniques visible in source | Documentation |
|---|---|---|---|
| Diffusion | Diffusion probabilistic models, DDPM, Improved DDPM, DDIM, GLIDE, Imagen, DALL-E 2, Stable Diffusion, ControlNet, consistency models | Custom tf.keras.Model/layers, GradientTape, tf.function, tf.data, optional distribution strategies |
Diffusion |
| GANs | Initial GAN, DCGAN, LSGAN, WGAN, WGAN-GP, Conditional GAN, CycleGAN, StarGAN, StyleGAN, SRGAN, Progressive GAN | Explicit adversarial losses and custom training steps | GANs |
| Image classification | LeNet, ZFNet, SqueezeNet, XNOR-Net, MobileNet, ShuffleNet, MLP-Mixer, Xception, AlexNet, VGG, Inception, ResNet, DenseNet, ViT, and more | Custom model assembly, datasets, callbacks, selected multi-GPU paths | Image classification |
| Detection and robustness | R-CNN through YOLOv3, SSD, Mask R-CNN; adversarial/saliency attack scripts | Custom heads/losses, tf.data, GradientTape, distributed detector runs |
Object detection · Robustness |
| NLP | One-hot/co-occurrence/PPMI-SVD/Word2Vec/GloVe/FastText, RNN/LSTM/GRU, attention seq2seq, copy/coverage, Transformer seq2seq | tf.data, custom layers/models, selected MirroredStrategy paths |
NLP |
| Language-model project | RedditStory BPE tokenizer, TFRecords, decoder-only Transformer, evaluation and sampling | tf.data, GradientTape, causal masking, KV cache, optional XLA/runtime controls |
RedditStory |
| Reinforcement learning | REINFORCE, policy-gradient/critic variants, vectorized DQN/Double DQN, Tic-Tac-Toe and Connect4 agents | GradientTape, tf.function, replay/trajectory buffers, TensorBoard summaries |
Reinforcement learning |
Some directories are intentionally documented as scaffolds when their
tracked source is empty or contains explicit NotImplementedError/pass
placeholders. They are not presented as complete experiments.
.
├── ComputerVision/
│ ├── Diffusion/
│ ├── GenerativeAdvesarialNetworks/
│ ├── ImageClassification/
│ └── ObjectDetection/
├── NaturalLanguageProcessing/
│ ├── WordEmbeddings/
│ ├── SequenceModels/
│ ├── Seq2Seq/
│ └── Project/RedditStory/
├── ReinforcementLearning/
│ ├── SingleEnv/
│ ├── Vectorized/
│ └── Project/
└── LICENSE
| Technique | Used for |
|---|---|
tf.keras.Model and custom layers |
Architectures, detector heads, attention blocks, embedding models, and game/RL networks |
tf.GradientTape |
GAN, diffusion, detector, adversarial-attack, and RL update steps |
tf.function |
Compiled high-frequency train/update paths in several GAN, diffusion, and RL implementations |
tf.data |
Streaming image/text/detection datasets and the RedditStory TFRecord pipeline |
tf.distribute.MirroredStrategy |
Multi-device capable paths in selected CV, NLP, and RL families |
tf.summary |
TensorBoard scalars in detector and reinforcement-learning runs |
GPU acceleration is most useful for batched convolution, attention, dense matrix operations, and vectorized environment updates. The code also contains CPU fallbacks in many entry points; device availability and local datasets remain runtime prerequisites.
The source implements adversarial minimax objectives, score/noise-prediction losses, cross-entropy classification, sequence likelihood, attention, co-occurrence and factorisation objectives, temporal-difference learning, and policy-gradient returns. Complexity depends on each family: attention is quadratic in sequence length, dense convolutional models scale with spatial activation volume, and replay-based RL scales with batch sampling and network updates. Technical pages state only the complexity justified by their code.
The repository includes executable training/evaluation scripts, some detector tests, and configuration-driven runs. It does not ship a consistent set of tracked benchmark reports or generated figures, so this README intentionally does not claim numerical accuracy, FID, mAP, reward, or runtime results. Run artifacts are local and ignored by Git.
No requirements.txt, Conda environment, or pinned package manifest is
tracked. Install the dependencies required by the implementation you want to
run in an isolated environment. Source imports show TensorFlow and NumPy as the
core requirements; image, data, RL, and project-specific paths additionally
use packages such as OpenCV, Pillow, pandas, tqdm, Gymnasium, PRAW, and W&B.
python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install tensorflow numpyThen consult the README in the target family before installing its optional
dataset or runtime dependencies. Dataset roots are commonly configured in the
folder-local config.py files.
Most runnable folders follow a local pattern such as:
cd ComputerVision/GenerativeAdvesarialNetworks/1_InitialGAN
python3 train_and_test.pyor, for the standalone language-model project:
cd NaturalLanguageProcessing/Project/RedditStory
./train.sh --prepare-onlyDo not assume every numbered directory is complete: read its local status and configuration before launching it.
Start with word embeddings, LeNet-style image classification, and REINFORCE; then move to recurrent/seq2seq models, DQN, residual/convolutional families, and GANs; finally explore detection, diffusion, Transformers, and the end-to-end RedditStory pipeline.
Keep additions self-contained, document the mathematical objective and runtime assumptions, avoid committing generated artifacts, and include a reproducible entry point or validation check where possible.
This project is available under the MIT License.