Discrepancy between image from hv and using Datashader directly

On the left is the saved image from datashader, on the right is the layout / plot from holoviews, datashader, and bokeh. The only difference that I’ve been able to discern between my two methods is that I’m using the line method to plot with datashader and the curve method with holoviews: left:

    for column in range(1,4):
        agg = cvs.line(df, '0', str(column), ds.count())
        img = tf.shade(agg, how='eq_hist')
        imgs.append(img)

    stacked = tf.stack(*imgs)
    export_image(stacked, data)

right:

    curvesx = hv.Curve(df.iloc[:, [0,1]])
    curvesy = hv.Curve(df.iloc[:, [0,2]])
    curvesz = hv.Curve(df.iloc[:, [0,3]])
    layout = curvesx * curvesy * curvesz
    renderer = hv.renderer('bokeh').instance(fig='html')
    plot = renderer.get_plot(datashade(layout, dynamic = False)).state
    plot.plot_width = 800
    plot.plot_height = 500
    show(plot)

Not only does the data look a little different, (note the bottom of the graph, on the left there is data that is an outlier which doesn’t exist on the right). I’ve tried different aspect ratios, sizes etc.

In addition, I believe that I’m not generating a large enough image from holoviews, the right side graph seems much lower quality then the saved image, almost as if it is created small and then stretched out.

Any insight?

Thank you.

[jbednar]

Maybe you want something like this?

import pandas as pd, numpy as np
df = pd.DataFrame([[1, 10, 100], [2, 20, 200]],
                   columns=['0', '1', '2'])
df

def concat_columns(df, separator=np.NaN):
    x = df.columns[0]
    df = df.append({x:separator}, ignore_index=True)
    return pd.concat([df[[x,c]].rename(columns={x:'x', c:'y'}) for c in df.columns[1:]])

df2 = concat_columns(df)
df2

You should then be able to datashade df2 directly.

[adamhrosenberg]

I’ll look into it thank you