tensorboardX

Write TensorBoard events with simple function call.

• Support scalar, image, figure, histogram, audio, text, graph, onnx_graph, embedding, pr_curve, mesh, hyper-parameters and video summaries.

• requirement for demo_graph.py is tensorboardX>=1.9 and pytorch>=1.2

• FAQ

Install

Tested on anaconda2 / anaconda3, with PyTorch 1.1.0 / torchvision 0.3 / tensorboard 1.13.0

pip install tensorboardX

or build from source:

git clone https://github.com/lanpa/tensorboardX && cd tensorboardX && python setup.py install

You can optionally install crc32c to speed up saving a large amount of data.

Example

• Run the demo script: python examples/demo.py

• Use TensorBoard with tensorboard --logdir runs (needs to install TensorFlow)

# demo.pyimport torchimport torchvision.utils as vutilsimport numpy as npimport torchvision.models as modelsfrom torchvision import datasetsfrom tensorboardX import SummaryWriter

resnet18 = models.resnet18(False)
writer = SummaryWriter()
sample_rate = 44100freqs = [262, 294, 330, 349, 392, 440, 440, 440, 440, 440, 440]for n_iter in range(100):

    dummy_s1 = torch.rand(1)
    dummy_s2 = torch.rand(1)    # data grouping by `slash`
    writer.add_scalar('data/scalar1', dummy_s1[0], n_iter)
    writer.add_scalar('data/scalar2', dummy_s2[0], n_iter)

    writer.add_scalars('data/scalar_group', {'xsinx': n_iter * np.sin(n_iter),                                             'xcosx': n_iter * np.cos(n_iter),                                             'arctanx': np.arctan(n_iter)}, n_iter)

    dummy_img = torch.rand(32, 3, 64, 64)  # output from network
    if n_iter % 10 == 0:
        x = vutils.make_grid(dummy_img, normalize=True, scale_each=True)
        writer.add_image('Image', x, n_iter)

        dummy_audio = torch.zeros(sample_rate * 2)        for i in range(x.size(0)):            # amplitude of sound should in [-1, 1]
            dummy_audio[i] = np.cos(freqs[n_iter // 10] * np.pi * float(i) / float(sample_rate))
        writer.add_audio('myAudio', dummy_audio, n_iter, sample_rate=sample_rate)

        writer.add_text('Text', 'text logged at step:' + str(n_iter), n_iter)        for name, param in resnet18.named_parameters():
            writer.add_histogram(name, param.clone().cpu().data.numpy(), n_iter)        # needs tensorboard 0.4RC or later
        writer.add_pr_curve('xoxo', np.random.randint(2, size=100), np.random.rand(100), n_iter)

dataset = datasets.MNIST('mnist', train=False, download=True)
images = dataset.test_data[:100].float()
label = dataset.test_labels[:100]

features = images.view(100, 784)
writer.add_embedding(features, metadata=label, label_img=images.unsqueeze(1))# export scalar data to JSON for external processingwriter.export_scalars_to_json("./all_scalars.json")
writer.close()

Screenshots