Python-based Video Classification with Deep Learning

Download and extract the video classification dataset, upload to Google Drive, and use the train.csv, predict.csv, video weights.h5, and classification code.ipynb in a Python-based deep learning workflow in Google Colab.

Inspect the train folder and train.csv to assign labels to videos, creating the training dataset for deep learning video classification.

Build a transformer-based video classification model from python code in classification code.ipynb by extracting frames from training videos, preprocessing them, and saving video Weights.h5 to classify actions in predict folder.

Discover how the video_weights.h5 file stores a pre-trained model with learned weights and biases, and learn to train, save, and predict video classes with your own model.

Use a trained .h5 model to classify your custom videos. Manage the predict folder and predict.csv by adding video names for the prediction program.

Learn how to run your Python-based video classification notebook in Google Colab and activate the GPU via runtime type to speed up computation.

Check for a GPU in a Colab notebook using TensorFlow's gpu_device_name() function; print the GPU name or raise an error for debugging, with CPU fallback.

Connect Google Colab with Google Drive by mounting drive using drive.mount(content/drive) to access dataset images, store model weights, and note it's a one-time setup.

Install TensorFlow Docs from a GitHub repository using pip in a Jupyter cell, with quiet mode (-q), to access the TensorFlow API documentation.

Import python libraries for building a deep learning video classification model in TensorFlow, including Keras, tf.keras layers, tf docs.vis for visualization, NumPy, Pandas, and Matplotlib.

Load the training dataset from a csv on Google Drive using pandas and store it in the train df for inspection. Count total videos for dataset size.

Displays a random sample of ten videos from the training dataset using train df, revealing data structure, label distribution, and guiding data processing, model architecture, and training strategy decisions.

Configure the string lookup layer to map string labels to integer indices for training, using the unique train_df['tag'] vocabulary and zero out-of-vocabulary indices.

The crop_center function crops each video frame to a 128x128 square with center_crop_layer, reshapes frames to add a batch dimension, converts to numpy, and squeezes the batch for video classification.

Load and process video frames using OpenCV, cropping to 128x128 squares, collecting into a numpy array up to max frames, and releasing the capture to support consistent deep learning classification.

Create a feature extractor based on DenseNet 121 with ImageNet weights, returning a Keras model that outputs fixed-length feature vectors from 128x128x3 frames for video classification.

Process video data for a sequence model by loading frames, padding to a fixed length, and extracting per-frame features with a feature extractor. Return features and labels as numpy arrays.

Create a transformer-based model for video classification by using a positional embedding to capture frame order and a transformer encoder with multi-head self-attention and feedforward networks.

Define a compiled transformer-based model for video classification using the Keras functional API, with sequence_length, embed_dim, dense_dim, and num_heads, plus positional embedding and global max pooling 1D.

Train a transformer-based model for video classification over 13 epochs with a 15% validation split, using early stopping and saving only best weights to load and return the trained model.

Visualize the model architecture by inspecting layer details, parameter counts, and output shapes with trained_model.summary, and generate a graphical diagram with plot_model to assess complexity and guide fine-tuning.

Read test video paths from Predict.CSV, prepare frames with a 3D feature array using padding, extract features with a pre-trained extractor, and predict actions with the trained model, printing probabilities.