jw3126

using Test
using Convex1D
using BenchmarkTools

function minimize_scaled_L1_diff(xs, ys)
    # find t::Number that minimizers f(t) = sum(abs, t*xs - ys)
    # f is convex with piecewise constant derivative given by
    # f'(t) = Σ xi * sign(t*xi - yi)
    # One of the points ti := yi/xi must be a minimizer (for some xi !=0. If all xi==0 then f == const anyway)
    # Based on remarks of Mathieu Tanneau in slack

jw3126

using PyCall
using Flux

function doit_keras(cfg)
    keras = pyimport("tensorflow.keras")
    inp = keras.layers.Input((nothing, 1))
    x = inp
    x = keras.layers.Conv1D(kernel_size=51, filters=50)(x)
    x = keras.layers.Conv1D(kernel_size=1, filters=1)(x)
    out = x

jw3126

# This reproduces a pytorch_lightning issue
# where gradient checkpointing + ddp results in nan loss
#
# * Run with gpus=1 and it works fine.
# * Run with gpus=4 and it loss becomes nan quickly
#
# See also https://forums.pytorchlightning.ai/t/gradient-checkpointing-ddp-nan/398
import torch
from torch import nn
from torch.nn import functional as F

jw3126