Deep Learning Tutorials · PyTorch on Colab

Learn Python, data analysis, and neural networks through fifteen self-contained notebooks. Lessons are written in English; deep-learning examples use PyTorch.

Start learning

Click a badge to open a notebook in Colab. No local Python installation or repository clone is needed.

First-time Colab use: You may see “This notebook was not authored by Google.” This is a notebook-source notice, not a Python error. Confirm that you opened the intended course repository and trust its code, then choose Run anyway.

Runtime availability: Work through one tutorial at a time and disconnect unused sessions. If Colab reports too many active sessions, use Manage sessions to terminate an unused runtime after downloading any files you want to keep. Closing a browser tab alone may leave its runtime active. Free-tier limits are dynamic rather than a guaranteed fixed number; see the Colab FAQ.

Tutorial	What you will learn	Open
01 Colab and Python Basics	Python types, collections, control flow, functions, and a synthetic energy summary
02 NumPy, Pandas, and Visualization	Array operations, Series/MultiIndex, data cleaning, Matplotlib styling, Seaborn plots, and Plotly maps
03 Regression and Classification	Auto MPG regression and Fashion-MNIST classification with PyTorch
04 Convolutional Neural Networks	MNIST dense/CNN comparison and CIFAR-10 regularization and augmentation
05 CNNs in Practice	Transfer learning, feature maps, Grad-CAM, optional neural style transfer, and a visual ViT image-to-token walkthrough
06 Transfer Learning	Pretrained inference, frozen features versus fine-tuning, parameter audits, and five-class flower classification
07 Object Detection and Segmentation	YOLO11 boxes, instance masks, NMS/threshold experiments, exports, and a tiny detection training workflow
08 Text and Word Embeddings	Tokenization, masked-pooling sentiment, TF-IDF, Skip-gram negative sampling, embedding visualization and export
09 Encoder–Decoder Transformers	Positions, self/cross-attention, causal masks, Spanish–English training and autoregressive translation
10 Autoencoders and VAEs	Reconstruction, denoising, ECG anomaly scoring, VAE latent exploration, and shared-encoder adversarial training
11 GANs and Image Translation	DCGAN, fixed-noise animation, latent inversion, paired U-Net/PatchGAN, and unpaired cycle consistency
12 Deep Reinforcement Learning	CartPole actor–critic, Pendulum DDPG, constrained toy materials control, baselines, and policy animations
13 Graph Neural Networks	Message passing, GCN/GraphSAGE on Cora, GIN graph classification, and QM9 dipole regression
14 LLM-Based Agents	Local Qwen inference, RDKit tools, bounded recovery, prompt selection, and auditable traces
15 Retrieval-Augmented Generation	Web loading, chunk overlap, MiniLM/TF-IDF retrieval, local Qwen answers, citations, and index reload

1. Open a notebook and connect to a runtime. A CPU is sufficient for the default examples.
2. Optionally choose File → Save a copy in Drive to keep your changes.
3. Choose Runtime → Run all, or work through the cells in order. Setup cells install missing packages inside Colab.
4. Read the explanations and try the exercises after the initial run. Each lesson can run independently; following 01 → 15 is recommended for beginners.

Runtime and data

• 01: Python standard library only; no package installation or external data.

• 02: NumPy, Pandas, Matplotlib, Seaborn, and Plotly; embedded synthetic energy data plus automatic downloads of historical tips and county unemployment data.

• 03: Auto MPG and Fashion-MNIST download automatically. The default quick run uses 6,000 Fashion-MNIST training images.

• 04: MNIST and CIFAR-10 download automatically (CIFAR-10 archive about 163 MB). Quick mode uses small subsets and three epochs per model. The full archive is still downloaded; the official server can take 20 minutes or more.

• 05: Ants/bees archive (about 45 MB) and pretrained MobileNetV2 weights (about 14 MB) download automatically. Frozen features keep the core lesson practical on CPU. Neural style transfer is off by default: enabling RUN_STYLE_TRANSFER downloads VGG19 weights (about 550 MB) and performs a separate optimization. The tiny Vision Transformer section compares images, patch sequences, numeric embeddings, positional information, and untrained attention; it does not train a classifier.

• 06: Flowers archive (about 218 MB) and MobileNetV2 weights (about 14 MB) download automatically. Quick mode uses 200/50/50 train/validation/test images, 20 cached-feature head epochs, and two image-based fine-tuning epochs.

• 07: Pinned Ultralytics/YOLO11 setup, bundled example image, about 12 MB of nano model weights, and a roughly 7 MB COCO128 download. The default includes five CPU-compatible detection training epochs on ten training images; set RUN_TRAINING=False to skip them. COCO128 is a workflow smoke test, not an independent benchmark.

• 08: IMDb archive (about 84 MB), with balanced subsets and exact-text deduplication. Eight sentiment epochs and three Skip-gram epochs run by default; includes TF-IDF comparison, embedding plots, TSV exports, and checkpoint reload.

• 09: A 2.6 MB Spanish–English archive; 3,000/400/400 short pairs and six Transformer training epochs in quick mode. Includes token/position/attention plots, teacher-forced and greedy evaluation, and CPU checkpoint reload.

• 10: Fashion-MNIST, MNIST, and ECG downloads (~50 MB). CPU-friendly AE/denoising, normal-only anomaly fitting, a 2D VAE, and a two-epoch shared-encoder AAE demonstration. Includes ten figures, a fixed-prior training GIF, and five checkpoint reloads.

• 11: MNIST (~12 MB), twelve compact DCGAN epochs, four paired edge-to-digit epochs, and four unpaired normal/inverted-digit epochs. Includes fixed-noise animation, latent interpolation/inversion, update-isolation checks, and eight inference checkpoint reloads.

• 12: Gymnasium classic control; 240 CartPole episodes, 6,000 Pendulum steps, and 4,000 synthetic materials steps. Includes independent evaluation seeds, seven plots, two policy GIFs, and nine network reloads. No dataset or Atari ROM download.

• 13: PyG/RDKit, Cora, IMDB-BINARY, and a roughly 30 MB QM9 CSV. Quick mode uses 120 Cora epochs per model, 35 IMDB epochs per GIN, and 45 regression epochs on 3,000 molecules. Includes message/batch diagrams, baselines, and six checkpoint reloads.

• 14: Public Qwen2.5-0.5B-Instruct (~1 GB download), PyTorch and RDKit. No API key or paid inference calls. Two prompt candidates, an explicit six-molecule catalogue, and bounded tool calls. Default CPU code passed; held-out routing/end-to-end accuracy was 4/8, with failure cases retained.

• 15: MiniLM embeddings (~91 MB) and Qwen2.5-0.5B-Instruct (~1 GB), a live web article, and embedded fictional materials records. No API key or paid calls. Includes lexical/dense comparison, separate retrieval/content/citation/abstention diagnostics, four figures, and index/manifest reload.

For 03–06 and 08–13, QUICK_RUN = True is the default. Set it to False for larger experiments. DEVICE_PREFERENCE = "auto" uses a GPU when available and otherwise CPU; select "cpu" to force CPU. Tutorials 14–15 use the same device setting; their quick flags control generated answer length rather than training size. Downloads require internet access; no API key or Google Drive mount is required. Colab runtime files are temporary: download anything you want to keep from outputs/tutorial_XX/ before disconnecting.

The notebooks retain installed Colab packages and install only missing dependencies. They print their environment versions. All fifteen default code paths passed sequential execution in Colab CPU runtimes. See validation evidence and limits for the exact tests performed; quick-mode results are teaching examples, not benchmark claims.

Attribution and license

License information is available in LICENSE. TensorFlow versions are outside this series' scope.

Data and pretrained model sources are attributed in the notebooks. These include UCI Auto MPG, Fashion-MNIST, MNIST, CIFAR-10, the PyTorch-hosted ants/bees archive, Google flower photos, ManyThings/Tatoeba bilingual pairs, the Google-hosted ECG CSV, COCO128, Stanford IMDb, torchvision model weights, Ultralytics YOLO11 weights/example assets, Cora/Planetoid, TU IMDB-BINARY, and QM9 via DeepChem. The code license does not replace their respective terms. Datasets and model weights are downloaded from their sources and are not committed here.