ogrisel

"""Empirical evaluation of the extended feature Gram matrix of a ReLU MLP

Here we try to estimate the spectrum of the H^\infty matrix as defined in:

Gradient Descent Provably Optimizes Over-parameterized Neural Networks (2018)
Simon S. Du, Xiyu Zhai, Barnabas Poczos, Aarti Singh
https://arxiv.org/abs/1810.02054

Theorem 4.1 relies on the assumption that H^\infty has a strictly positive
minimum eigenvalue. The following computes an estimate of this eigenvalue

ogrisel

from sklearn.datasets import make_blobs
from sklearn.cluster import KMeans
from sklearn.externals import joblib


m = joblib.Memory(cachedir='/tmp/joblib')
make_blobs = m.cache(make_blobs)
data, labels = make_blobs(n_samples=10**5, n_features=50, cluster_std=100,
                          centers=10, random_state=777)

ogrisel

from pickle import Pickler, load

try:
    from pickle import PickleBuffer
except ImportError:
    PickleBuffer = None
import copyreg
import os
import numpy as np
import time

ogrisel

from pickle import Pickler, _Pickler, Unpickler, _Unpickler, HIGHEST_PROTOCOL
import os
import time
import sys
import gc
from multiprocessing import get_context

PROTOCOL = HIGHEST_PROTOCOL

ctx = get_context('spawn')

ogrisel

distributed.worker - WARNING - Worker at 72 percent memory usage. Trigger GC. Process memory: 723.75 MB -- Worker memory limit: 1000.00 MB
distributed.worker - WARNING - Worker at 66 percent memory usage. After GC. Process memory: 660.93 MB -- Worker memory limit: 1000.00 MB
distributed.worker - WARNING - Worker at 73 percent memory usage. Trigger GC. Process memory: 732.79 MB -- Worker memory limit: 1000.00 MB
distributed.worker - WARNING - Worker at 73 percent memory usage. After GC. Process memory: 732.79 MB -- Worker memory limit: 1000.00 MB
distributed.core - WARNING - Event loop was unresponsive for 1.01s.  This is often caused by long-running GIL-holding functions or moving large chunks of data. This can cause timeouts and instability.
distributed.worker - WARNING - Worker at 70 percent memory usage. Trigger GC. Process memory: 705.26 MB -- Worker memory limit: 1000.00 MB
distributed.worker - WARNING - Worker at 67 percent memory usage. After GC. Process memory: 670.00 MB -- Worker memory limit: 1000.0

ogrisel

import os
import os.path as op
from time import time
import dask.dataframe as ddf
import dask.array as da
from distributed import Client


def make_categorical_data(n_samples=int(1e7), n_features=10, n_partitions=100):
    """Generate some random categorical data

ogrisel

#
# XXX: do not use this code, it's broken!
# Use: https://gist.github.com/ogrisel/b6a97ed87939e3b559568ac2f6599cba
#
# See comments.

import os
import os.path as op
from time import time
import dask.dataframe as ddf

ogrisel

__pycache__
*.json
*.png

ogrisel

_________________ TestsProcessPoolLokyExecutor.test_max_depth __________________

self = <tests.test_process_executor_loky.TestsProcessPoolLokyExecutor instance at 0x10392ed88>

    def test_max_depth(self):

        from loky.process_executor import MAX_DEPTH

        if self.context.get_start_method() == 'fork':

ogrisel