There are 11 notebooks in this repo. These are created as part of Tensorflow Deep Learning course on Udemy
Description:
| Notebook | Topics covered |
|---|---|
| 00_tensorflow_fundamentals | tf.constant, tf.Variable, .assign, random class, numpy to tensor, shape, ndim, Other ways to create tensors, indexing, Math operations : +,-,/,*, Matrix multiplication, reshape, transpose, Aggregations, GPU configuration |
| 01_neural_network_regression_with_tensorflow | Regression using tensorlow for insurance dataset, Model building using Sequential API, model.compile (adding loss, optimizer and metrics), model.fit, model.summary, model.evaluate, model.predict, Download files from google colab, one-hot encoding using get_dummies in pandas, tf.keras.utils.plot_model, plotting model history |
| 02-neural_network_classification_with_tensorflow | Introduction to classification, losses - BinaryCrossentropy(), CategoricalCrossentropy(), SparseCategoricalCrossentropy(), output activations - sigmoid, softmax, Classification evaluation metrics, finding best learning rate using LearningRateScheduler callback, Multiclass classification for Fashion MNIST with 10 classes, plot confusion matrix for multi class classification |
| 03-introduction_to_computer_vision_with_tensrflow | Pizza 🍕 and Steak image classification from Food101 dataset. Use os.walk, visualize images, normalization (Rescale), read images from directory using tensorflow.keras.preprocessing.image.ImageDataGenerator, TinyVGG architecture with Conv2D, MaxPool2D and Flatten layers, Data Augmentation using ImageDataGenerator, Regularization using Dropout layer, Testing Tiny VGG on 10 classes of foods |
| 04-transfer_learning_with_tensorflow_part1_features_extraction | Types of Transfer Learning - As is, Feature extraction and Fine tuning, read images using tf.keras.utils.image_dataset_from_directory class, Create TensorBoard callback, explore EfficientNet_B0 and Resnet_v2_50 from tensorflow_hub, upload records to tensorboard dev |
| 05-transfer_learning_with_tensorflow_part2_fine_tuning | Feature extraction from tf.keras.applications, ModelCheckpoint, unfreeze layers and fine tune with reduced learning rate and setting initial_epoch as the last epoch of feature extractor model |
| 06-transfer_learning_with_tensorflow_part3_scaling_up | FoodVision 101 -> Scaling up to 101 classes, Data Augmentation using layers, feature extractor and fine-tuned EfficientNetB0 model, model.save and tf.keras.models.load_model, using sklearn classification_report, finding most wrong predictions |
| 07-milestone_project_1_food_101 | mixed_precision training, layer dtype and dtype_policy, making use of tf.data.Dataset API, Batch and prefetch data, Beat DeepFood paper results with fine tuning EfficientNetB0 😄 |
| 08-introduction_to_nlp_with_tensorflow | Disaster tweet classification, TextVectorization, word Embedding, LSTM, GRU, Bidirectional LSTM, Conv1D, feature extraction from Universal Sentence Encoder, speed/score tradeoff |
| 09_skimlit_milestone_project2 | Replicating this and this papers for multi class classification of PubMed 20k RCT data, feature extractor model, word embedding based model, character embedding based model, multimodal model - pretrained token embedding + character embedding, positional embedding of line_number and total_lines features |
| 10_time_series_fundamentals_with_tensorflow | Bitcoin price prediction 💰 using TimeSeries, naive model, windows and horizon, creating windows using numpy and timeseries_dataset_from_array APIs, Conv1D layer with causal padding, using Lambda layer, LSTM, N-BEATS algorithm, creating custom layers using subclassing, using Add and Subtract layers to create residual connections, ensemble models, uncertaining estimates (confidence interval), coconut and subway uncertainity discussion, future predictions, Turkey 🦃 problem in forecasting |