view scripts/stacked_dae/nist_sda.py @ 141:3346fcd3818b

Changé un True pour False, mais je le commit car c'est ça que j'utilise pour lancer mes jobs.
author fsavard
date Mon, 22 Feb 2010 14:22:07 -0500
parents 7d8366fb90bf
children
line wrap: on
line source

#!/usr/bin/python
# coding: utf-8

import numpy 
import theano
import time
import theano.tensor as T
from theano.tensor.shared_randomstreams import RandomStreams
import copy

import sys
import os.path

from sgd_optimization import SdaSgdOptimizer

from jobman import DD
import jobman, jobman.sql
from pylearn.io import filetensor

from utils import produit_croise_jobs

TEST_CONFIG = False

NIST_ALL_LOCATION = '/data/lisa/data/nist/by_class/all'

JOBDB = 'postgres://ift6266h10@gershwin/ift6266h10_db/'
REDUCE_TRAIN_TO = None
MAX_FINETUNING_EPOCHS = 1000
if TEST_CONFIG:
    JOBDB = 'postgres://ift6266h10@gershwin/ift6266h10_sandbox_db/'
    REDUCE_TRAIN_TO = 1000
    MAX_FINETUNING_EPOCHS = 2

JOBDB_JOBS = JOBDB + 'fsavard_sda1_jobs'
JOBDB_RESULTS = JOBDB + 'fsavard_sda1_results'
EXPERIMENT_PATH = "ift6266.scripts.stacked_dae.nist_sda.jobman_entrypoint"

# There used to be
# 'finetuning_lr': [0.00001, 0.0001, 0.001, 0.01, 0.1]
# and
#  'num_hidden_layers':[1,2,3]
# but this is now handled by a special mechanism in SgdOptimizer
# to reuse intermediate results (for the same training of lower layers,
# we can test many finetuning_lr)
JOB_VALS = {'pretraining_lr': [0.1, 0.01, 0.001],#, 0.0001],
        'pretraining_epochs_per_layer': [10,20],
        'hidden_layers_sizes': [300,800],
        'corruption_levels': [0.1,0.2],
        'minibatch_size': [20],
        'max_finetuning_epochs':[MAX_FINETUNING_EPOCHS]}
FINETUNING_LR_VALS = [0.1, 0.01, 0.001]#, 0.0001]
NUM_HIDDEN_LAYERS_VALS = [1,2,3]

# Just useful for tests... minimal number of epochs
DEFAULT_HP_NIST = DD({'finetuning_lr':0.01,
                       'pretraining_lr':0.01,
                       'pretraining_epochs_per_layer':1,
                       'max_finetuning_epochs':1,
                       'hidden_layers_sizes':[1000],
                       'corruption_levels':[0.2],
                       'minibatch_size':20})

def jobman_entrypoint(state, channel):
    state = copy.copy(state)

    print "Will load NIST"
    nist = NIST(20)
    print "NIST loaded"

    rtt = None
    if state.has_key('reduce_train_to'):
        rtt = state['reduce_train_to']
    elif REDUCE_TRAIN_TO:
        rtt = REDUCE_TRAIN_TO

    if rtt:
        print "Reducing training set to ", rtt, " examples"
        nist.reduce_train_set(rtt)

    train,valid,test = nist.get_tvt()
    dataset = (train,valid,test)

    n_ins = 32*32
    n_outs = 62 # 10 digits, 26*2 (lower, capitals)

    db = jobman.sql.db(JOBDB_RESULTS)
    optimizer = SdaSgdOptimizer(dataset, state, n_ins, n_outs,\
                    input_divider=255.0, job_tree=True, results_db=db, \
                    experiment=EXPERIMENT_PATH, \
                    finetuning_lr_to_try=FINETUNING_LR_VALS, \
                    num_hidden_layers_to_try=NUM_HIDDEN_LAYERS_VALS)
    optimizer.train()

    return channel.COMPLETE

def estimate_pretraining_time(job):
    job = DD(job)
    # time spent on pretraining estimated as O(n^2) where n=num hidens
    # no need to multiply by num_hidden_layers, as results from num=1 
    # is reused for num=2, or 3, so in the end we get the same time
    # as if we were training 3 times a single layer
    # constants:
    # - 20 mins to pretrain a layer with 1000 units (per 1 epoch)
    # - 12 mins to finetune (per 1 epoch)
    # basically the job_tree trick gives us a 5 times speedup on the
    # pretraining time due to reusing for finetuning_lr
    # and gives us a second x2 speedup for reusing previous layers
    # to explore num_hidden_layers
    return (job.pretraining_epochs_per_layer * 20 / (1000.0*1000) \
            * job.hidden_layer_sizes * job.hidden_layer_sizes)

def estimate_total_time():
    jobs = produit_croise_jobs(JOB_VALS)
    sumtime = 0.0
    sum_without = 0.0
    for job in jobs:
        sumtime += estimate_pretraining_time(job)
        # 12 mins per epoch * 30 epochs
        # 5 finetuning_lr per pretraining combination
    sum_without = (12*20*len(jobs) + sumtime*2) * len(FINETUNING_LR_VALS)
    sumtime += len(FINETUNING_LR_VALS) * len(jobs) * 12 * 20
    print "num jobs=", len(jobs)
    print "estimate", sumtime/60, " hours"
    print "estimate without tree optimization", sum_without/60, "ratio", sumtime / sum_without

def jobman_insert_nist():
    jobs = produit_croise_jobs(JOB_VALS)

    db = jobman.sql.db(JOBDB_JOBS)
    for job in jobs:
        job.update({jobman.sql.EXPERIMENT: EXPERIMENT_PATH})
        jobman.sql.insert_dict(job, db)

    print "inserted"

class NIST:
    def __init__(self, minibatch_size, basepath=None, reduce_train_to=None):
        global NIST_ALL_LOCATION

        self.minibatch_size = minibatch_size
        self.basepath = basepath and basepath or NIST_ALL_LOCATION

        self.set_filenames()

        # arrays of 2 elements: .x, .y
        self.train = [None, None]
        self.test = [None, None]

        self.load_train_test()

        self.valid = [[], []]
        self.split_train_valid()
        if reduce_train_to:
            self.reduce_train_set(reduce_train_to)

    def get_tvt(self):
        return self.train, self.valid, self.test

    def set_filenames(self):
        self.train_files = ['all_train_data.ft',
                                'all_train_labels.ft']

        self.test_files = ['all_test_data.ft',
                            'all_test_labels.ft']

    def load_train_test(self):
        self.load_data_labels(self.train_files, self.train)
        self.load_data_labels(self.test_files, self.test)

    def load_data_labels(self, filenames, pair):
        for i, fn in enumerate(filenames):
            f = open(os.path.join(self.basepath, fn))
            pair[i] = filetensor.read(f)
            f.close()

    def reduce_train_set(self, max):
        self.train[0] = self.train[0][:max]
        self.train[1] = self.train[1][:max]

        if max < len(self.test[0]):
            for ar in (self.test, self.valid):
                ar[0] = ar[0][:max]
                ar[1] = ar[1][:max]

    def split_train_valid(self):
        test_len = len(self.test[0])
        
        new_train_x = self.train[0][:-test_len]
        new_train_y = self.train[1][:-test_len]

        self.valid[0] = self.train[0][-test_len:]
        self.valid[1] = self.train[1][-test_len:]

        self.train[0] = new_train_x
        self.train[1] = new_train_y

def test_load_nist():
    print "Will load NIST"

    import time
    t1 = time.time()
    nist = NIST(20)
    t2 = time.time()

    print "NIST loaded. time delta = ", t2-t1

    tr,v,te = nist.get_tvt()

    print "Lenghts: ", len(tr[0]), len(v[0]), len(te[0])

    raw_input("Press any key")

# hp for hyperparameters
def sgd_optimization_nist(hp=None, dataset_dir='/data/lisa/data/nist'):
    global DEFAULT_HP_NIST
    hp = hp and hp or DEFAULT_HP_NIST

    print "Will load NIST"

    import time
    t1 = time.time()
    nist = NIST(20, reduce_train_to=100)
    t2 = time.time()

    print "NIST loaded. time delta = ", t2-t1

    train,valid,test = nist.get_tvt()
    dataset = (train,valid,test)

    print train[0][15]
    print type(train[0][1])


    print "Lengths train, valid, test: ", len(train[0]), len(valid[0]), len(test[0])

    n_ins = 32*32
    n_outs = 62 # 10 digits, 26*2 (lower, capitals)

    optimizer = SdaSgdOptimizer(dataset, hp, n_ins, n_outs, input_divider=255.0)
    optimizer.train()

if __name__ == '__main__':

    import sys

    args = sys.argv[1:]

    if len(args) > 0 and args[0] == 'load_nist':
        test_load_nist()

    elif len(args) > 0 and args[0] == 'jobman_insert':
        jobman_insert_nist()
    elif len(args) > 0 and args[0] == 'test_job_tree':
        # dont forget to comment out sql.inserts and make reduce_train_to=100
        print "TESTING JOB TREE"
        chanmock = {'COMPLETE':0}
        hp = copy.copy(DEFAULT_HP_NIST)
        hp.update({'reduce_train_to':100})
        jobman_entrypoint(hp, chanmock)
    elif len(args) > 0 and args[0] == 'estimate':
        estimate_total_time()
    else:
        sgd_optimization_nist()