Basic Tensorflow example

This example demonstrates using Simvue to track a simple Python code, in particular:

• Collecting some metadata,
• Saving the Python script,
• Logging some metrics while the training is running,
• Adding some additional metadata containing the final values of the metrics.

The code is based on a dynamic recurrent neural network example - find the original neural network code on Github.

Running the code

To run this example, firstly clone the Simvue client GitHub repository:

git clone https://github.com/simvue-io/client
cd examples/Tensorflow

Create a virtual environment:

python3 -m venv venv
source venv/bin/activate

Install the required dependencies:

pip install --upgrade pip
pip install tensorflow simvue

Ensure that Simvue is configured properly, for example by creating a .simvue.toml file in the Tensorflow directory. Click on Create new run in the UI for more information.

Run the code:

python3 dynamic_rnn.py

Explanation

To begin with we import the required class:

from simvue import Run

We next initialise the run and specify metadata:

run = Run()
run.init(metadata={'dataset.num_classes': num_classes,
                   'dataset.seq_max_len': seq_max_len,
                   'dataset.seq_min_len': seq_min_len,
                   'dataset.masking_val': masking_val,
                   'training.learning_rate': learning_rate,
                   'training.training_steps': training_steps,
                   'training.batch_size': batch_size,
                   'network.num_units': num_units})

The Python code itself is saved:

run.save_file('dynamic_rnn.py', 'code')

During the part of the code which carries out the training we log metrics:

run.log_metrics({'loss': float(loss), 'accuracy': float(acc)})

Once the training has completed we add some metadata specifying the final values of the metrics:

run.update_metadata({'loss': float(loss), 'accuracy': float(acc)})

Finally we finish the run:

run.close()

Results

As well as having a record of metadata associated with the training run you can visualize the metrics in real-time, for example loss: