wandb-tsbw 0.1.0

A compatibility layer for TensorBoard's SummaryWriter to log to Weights & Biases

wandb_tsbw

A compatibility layer for TensorBoard's SummaryWriter to log to Weights & Biases.

Description

wandb_tsbw provides a drop-in replacement for TensorBoard's SummaryWriter that logs to Weights & Biases (W&B) instead. This allows users to easily migrate from TensorBoard to W&B without changing their existing code.

Installation

pip install wandb_tsbw

Requirements

• Python >= 3.6
• wandb >= 0.12.0
• numpy >= 1.19.0

Usage

Simply replace your TensorBoard import with the wandb_tsbw import:

# Original TensorBoard code:
# from torch.utils.tensorboard import SummaryWriter

# New W&B-based code:
from wandb_tsbw import SummaryWriter

# Rest of your code remains the same
writer = SummaryWriter(log_dir='./runs/experiment_1')
writer.add_scalar('loss', 0.5, global_step=1)

# Pass W&B specific parameters through wandb_kwargs
writer = SummaryWriter(
    log_dir='./runs/experiment_1',
    wandb_kwargs={
        'project': 'my_project',
        'entity': 'my_team',
        'tags': ['experiment', 'classification'],
        'group': 'model_v1'
    }
)

Features

The SummaryWriter class implements the same interface as TensorBoard's SummaryWriter but sends all data to W&B. It supports:

• Scalar logging
• Multiple scalars logging
• Histogram logging
• Image logging
• Text logging
• Figure logging (Matplotlib)
• Hyperparameter logging
• PR curve logging
• Embedding visualization
• Audio logging
• Video logging

Differences from TensorBoard

While wandb_tsbw attempts to provide a drop-in replacement for TensorBoard's SummaryWriter, there are some subtle differences:

1. W&B requires authentication. Make sure you've run wandb login before using this package.
2. W&B has different visualization capabilities, so some visualizations may look different.
3. W&B automatically syncs data to their cloud servers, while TensorBoard keeps data locally.
4. Some advanced TensorBoard features (like debugging) may not be fully supported.

Example

import torch
import torch.nn.functional as F
from wandb_tsbw import SummaryWriter
import numpy as np

# Initialize the SummaryWriter
writer = SummaryWriter(log_dir='runs/experiment_1')

# Log scalars
for i in range(10):
    writer.add_scalar('Loss/train', np.random.random(), i)
    writer.add_scalar('Loss/test', np.random.random(), i)
    writer.add_scalar('Accuracy/train', np.random.random(), i)
    writer.add_scalar('Accuracy/test', np.random.random(), i)

# Log histogram
writer.add_histogram('Conv1.bias', torch.randn(100), 0)

# Log image
img = torch.zeros(3, 100, 100)  # C, H, W
img[0] = torch.arange(0, 10000).reshape(100, 100) / 10000
writer.add_image('Example Image', img, 0)

# Log text
writer.add_text('Text Example', 'Hello World!', 0)

# Log hyperparameters
writer.add_hparams(
    {'lr': 0.1, 'batch_size': 64},
    {'accuracy': 0.9, 'loss': 0.2}
)

# Close the writer
writer.close()

Debugging:

Option 1: Use pytest's command line option

Run pytest with the --trace option to drop into the debugger at the start of each test:

pytest-3 tests/test_writer.py::TestSummaryWriter::test_add_hparams --trace

Option 4: Debug on test failure

You can configure pytest to automatically start the debugger when a test fails:

pytest-3 tests/test_writer.py -v --pdb

License

MIT