tensorboard-pytorch

Write tensorboard events with simple command.

including scalar, image, histogram, audio, text, graph and embedding.

see demo (result of demo.py and some images generated by BEGAN)

install

#tested on anaconda2/anaconda3, tensorflow 1.2.1, pytorch 0.1.12, torchvision 0.1.8

pip install tensorboard-pytorch

pip install tensorflow or pip install tensorflow-gpu

As this time (v0.1.12), to use the graph drawing feature, you need to build pytorch from source.

usage

import torch
import torchvision.utils as vutils
import numpy as np
import torchvision.models as models
from datetime import datetime
from tensorboard import SummaryWriter
resnet18 = models.resnet18(True)
writer = SummaryWriter('runs/'+datetime.now().strftime('%B%d  %H:%M:%S'))
sample_rate = 44100
freqs = [262, 294, 330, 349, 392, 440, 440, 440, 440, 440, 440]
for n_iter in range(100):
    M_global = torch.rand(1) # value to keep
    writer.add_scalar('M_global', M_global[0], n_iter)
    x = torch.rand(32, 3, 64, 64) # output from network
    if n_iter%10==0:
        x = vutils.make_grid(x, normalize=True, scale_each=True)   
        writer.add_image('Image', x, n_iter)
        x = torch.zeros(sample_rate*2)
        for i in range(x.size(0)):
            x[i] = np.cos(freqs[n_iter//10]*np.pi*float(i)/float(sample_rate)) # sound amplitude should in [-1, 1]
        writer.add_audio('Audio', x, n_iter)
        writer.add_text('Text', 'testtext', n_iter)
        for name, param in resnet18.named_parameters():
            writer.add_histogram(name, param.clone().cpu().data.numpy(), n_iter)
writer.close()

python demo.py

tensorboard --logdir runs

scalar

image

audio

distribution

histogram

text

tweaks

To show more images in tensorboard's image tab, just modify the hardcoded event_accumulator in ~/anaconda3/lib/python3.6/site-packages/tensorflow/tensorboard/backend/application.py as you wish.

API

class SummaryWriter(str log_dir)

parameter:

• log_dir (str) - the location of the log folder.

add_scalar(str tag, float value, int global_step=None)

parameter:

• value (float) - the value to keep.

add_image(str tag, torch.Tensor t, int global_step=None)

parameter:

• t (torch.Tensor) - torch tensor of size (3,H,W). I suggest use torchvision.utils.make_grid() to prepare it.

add_audio(str tag, torch.Tensor t, int global_step=None)

parameter:

• t (torch.Tensor) - one dimensional torch tensor. The value should between [-1, 1]. The sample rate is currently fixed at 44100 KHz.

add_histogram(str tag, numpy.ndarray values, int global_step=None, bins='tensorflow')

parameter:

• values (numpy.ndarray) - one dimensional numpy array.
• bins (str) - one of {'tensorflow', numpy_arguments}, determines how the bins are made.

add_text(str tag, str text_to_save, int global_step=None)

parameter:

• text_to_save (str) - the string to keep.

add_graph(torch.nn.Module model, torch.autograd.Variable lastnode)

parameter:

• model (torch.nn.Module) - the model to draw.
• lastnode (torch.autograd.Variable) - the output variable of the model.

see demo_graph.py

add_embedding(torch.Tensor t, str save_path, list metadata=None, torch.Tensor label_img=None)

parameter:

• t (torch.Tensor) - A two dimensional tensor. Each row corresponds to one feature vector.
• save_path (str) - save path
• metadata (list) - a list of label composed of number or string. len(metadata) should equal the number of rows of t
• label_img (torch.Tensor) - 4-D tensor (NCHW). label_img.size(0) should equal the number of rows of t

see demo_embedding.py

Common parameters:

• tag (str) - values with same tag group together.
• global_step (int) - logs the training step on which the value is saved.

TODO

audio sample rate

reference:

https://github.com/TeamHG-Memex/tensorboard_logger

https://github.com/dmlc/tensorboard