rustonaut/keras-plot

live ploting with bokeh, port of blocks-extras.extensions.Plot to keras

keras-plot

A mostly strait forward port of blocks-extras.extensions.Plot to a keras callback, including most of the original drawback and benefits.

Note: Require bokeh version 0.10!!, like the blocks version

Note: this is mostly a fast straight forward port, with some little goodies added in and a bit of refactoring do not expect production quality or maintance for this package

Installation

No special inatallation requirements, but because it relies on a old version of bokeh I won't upload it on PyPI, you can use e.g.: pip install git+https://github.com/dathinab/keras-plot.git --user

Changes compared to blocks-extras.extensions.Plot

• some refactoring making unessesary complex internal code mor simple and readable
• removed start_server option, as it should not have existed at all
• changed the default server_url behaviour (there is not simple mapping from the config mechanism from blocks to keras)
• added some small nice goodies:

  • you can just pass in the same functions to the plots channel parameter, as you can pass to the fit functions metrics parameter. Be aware that many functions are implemented both in keras.objectives and keras.metrics and that e.g. a mean_squared_error objective is not the same as a mean_squared_error metric! The Plot class intentionally rejects keras.objective functions as this is mostly likely a bug, which can lead to non obvious error messages, if the objective is also used as a metric in the training functionality.

  • Also both 'accuracy' and 'acc' will map to 'acc' as this is how keras behaves.

Example

(note: it's quite a nonsensical example)

Before running this example make sure to start a bokeh server of version 0.10 on the localhost, port 5006 (wich is the default for it) or add parameters to Plot constructor.

from keras_plot import Plot
import numpy

from keras.models import Sequential
from keras.layers import Dense, Activation

model = Sequential()
model.add(Dense(output_dim=1, input_dim=100))
model.add(Activation("relu"))

from keras.optimizers import SGD
import keras.objectives as objectives
from keras.metrics import mean_absolute_error, binary_accuracy

model.compile(
    loss=objectives.mean_squared_error,
    optimizer=SGD(),
    metrics=[binary_accuracy, mean_absolute_error]
)

X_train = numpy.random.random(size=(300, 100))
Y_train = numpy.random.randint(0, 2, size=(300, 1))

model.fit(X_train, Y_train, nb_epoch=50, batch_size=10, callbacks=[
    Plot(
        document_name="test4231",
        channels=[
            # the first plot (not that MSE+Accuracy in same plot make hardly any sense)
            ['loss', binary_accuracy],
            # the second plot
            [mean_absolute_error]
        ],
    )
])

Collection of possible improvements

• switch to newer bokeh version
• when plotting after every batch also add a (samely) colored graph ploted after every epoch, which is kind of the smothed version of "after every batch"
• possible mix "after every batch" and "after every epoch" for different metrics
• possible add filters like "moving window smoothing"
• change API to accept bokeh Session instead of a url
  • for login, configuarability etc.
  • possible add some load/create from config function for it