TensorFlow is a Python-based open source machine learning framework developed primarily by Google. It supplies low-level tools for tensor computation and gradient-based training, while its wider ecosystem supports workflows from research to production deployment. section 3.3 section 3.2
A linear classifier from scratch
Start with two classes of points in a 2D plane, stored as 2,000 input rows with corresponding 0-or-1 targets. A classifier computes prediction = tf.matmul(inputs, W) + b. Its trainable variables W and b define a line separating the classes. Training minimizes mean squared error; predictions above 0.5 can then be plotted to inspect the learned separation. section 3.3.2
The training step puts TensorFlow’s core operations together:
- Use tensors and tensor operations to calculate predictions and loss. section 3.3.2
- Use tf.GradientTape to obtain gradients of the loss with respect to W and b, then update those variables with assign_sub. section 3.3.2
- Repeat the step in a training loop and watch the loss; the example runs 40 steps. section 3.3.2
Listing 3.20 The training-step function
learning_rate = 0.1 @tf.function(jit_compile=True) #1 def training_step(inputs, targets, W, b): with tf.GradientTape() as tape: #2 predictions = model(inputs, W, b) #2 loss = mean_squared_error(predictions, targets) #2 grad_loss_wrt_W, grad_loss_wrt_b = tape.gradient(loss, [W, b]) #3 W.assign_sub(grad_loss_wrt_W * learning_rate) #4 b.assign_sub(grad_loss_wrt_b * learning_rate) #4 return loss
The training-step listing shows the gradient calculation and variable updates in one place. It also wraps the step in tf.function with XLA compilation enabled. section 3.3.2
TensorFlow can also serve as a Keras backend. In that pairing, TensorFlow provides lower-level tensors, operations, and gradients; Keras supplies high-level layers, models, losses, optimizers, metrics, and training-loop APIs. Use the lower-level approach when you need to see or control each training step. section 3.2 section 3.3.2
For broader work, TensorFlow offers tf.data for preprocessing, string and ragged tensors for specialized inputs, and deployment components including TF-Serving and TFLite. Graph mode and XLA compilation are performance strengths. The trade-offs are a sprawling API, occasional differences from NumPy’s numerical API, and less Hugging Face support for some recent generative AI models. section 3.3.3 section 3.3
