Visualize images without Trainer

[disktnk]

This PR provides summary.image method and summary.reporter to visualize images on browser. Current ChainerUI has already provided (undocumented)ImageReporter, the extension collects images cached by summary.image, split register phase and write phase, to reduce file writing cost. In many case, however, registering images and write tend to do in same time such as evaluation time, this design confuses.

Usage

from chainerui import summary
summary.image(x, out=out_path)
summary.image(x, out=out_path)

x, x2 are converted as PNG format and show on a browser. out_path is required to same as log file directory.

timestamp	image
2018-11-12T12:00:00	[x image]
2018-11-12T12:00:01	[x2 image]

from chainerui import summary
with summary.reporter(out=out_path) as r:
    r.image(x)
    r.image(x2)

x and x2 are converted as PNG format and show in same row.

timestamp	image_0	image_1
2018-11-12T12:00:00	[x image]	[x2 image]

Pros

1. more simplified usage
2. no dependency to Trainer

Cons

1. epoch and iteration are not set automatically
2. set out_put every time

To deal with Cons. 1, timestamp key is added automatically to distinguish images roughly. Additionally any key-pair discription can be set on the function, like summary.image(x, out_path, epoch=10)

To be ease Cons 2, prepare chainerui.summary.set_out(out_path) method to set out_path once.

[disktnk]

I put quick examples on https://github.com/disktnk/chainerui-demo-20180609/tree/master/vae

[makky3939]

LGTM. I tried this p-r using example scripts

[gky360]

LGTM 🙇