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

# Generic SdA optimization loop, adapted from the deeplearning.net tutorial

import numpy 
import theano
import time
import theano.tensor as T
import copy
import sys

from jobman import DD
import jobman, jobman.sql

from stacked_dae import SdA

def shared_dataset(data_xy):
    data_x, data_y = data_xy
    #shared_x = theano.shared(numpy.asarray(data_x, dtype=theano.config.floatX))
    #shared_y = theano.shared(numpy.asarray(data_y, dtype=theano.config.floatX))
    #shared_y = T.cast(shared_y, 'int32')
    shared_x = theano.shared(data_x)
    shared_y = theano.shared(data_y)
    return shared_x, shared_y

class SdaSgdOptimizer:
    def __init__(self, dataset, hyperparameters, n_ins, n_outs, input_divider=1.0,\
                job_tree=False, results_db=None,\
                experiment="",\
                num_hidden_layers_to_try=[1,2,3], \
                finetuning_lr_to_try=[0.1, 0.01, 0.001, 0.0001, 0.00001]):

        self.dataset = dataset
        self.hp = copy.copy(hyperparameters)
        self.n_ins = n_ins
        self.n_outs = n_outs
        self.input_divider = numpy.asarray(input_divider, dtype=theano.config.floatX)

        self.job_tree = job_tree
        self.results_db = results_db
        self.experiment = experiment
        if self.job_tree:
            assert(not results_db is None)
            # these hp should not be there, so we insert default values
            # we use 3 hidden layers as we'll iterate through 1,2,3
            self.hp.finetuning_lr = 0.1 # dummy value, will be replaced anyway
            cl = self.hp.corruption_levels
            nh = self.hp.hidden_layers_sizes
            self.hp.corruption_levels = [cl,cl,cl]
            self.hp.hidden_layers_sizes = [nh,nh,nh]
            
        self.num_hidden_layers_to_try = num_hidden_layers_to_try
        self.finetuning_lr_to_try = finetuning_lr_to_try

        self.printout_frequency = 1000

        self.rng = numpy.random.RandomState(1234)

        self.init_datasets()
        self.init_classifier()
     
    def init_datasets(self):
        print "init_datasets"
        train_set, valid_set, test_set = self.dataset
        self.test_set_x, self.test_set_y = shared_dataset(test_set)
        self.valid_set_x, self.valid_set_y = shared_dataset(valid_set)
        self.train_set_x, self.train_set_y = shared_dataset(train_set)

        # compute number of minibatches for training, validation and testing
        self.n_train_batches = self.train_set_x.value.shape[0] / self.hp.minibatch_size
        self.n_valid_batches = self.valid_set_x.value.shape[0] / self.hp.minibatch_size
        self.n_test_batches  = self.test_set_x.value.shape[0]  / self.hp.minibatch_size

    def init_classifier(self):
        print "Constructing classifier"
        # construct the stacked denoising autoencoder class
        self.classifier = SdA( \
                          train_set_x= self.train_set_x, \
                          train_set_y = self.train_set_y,\
                          batch_size = self.hp.minibatch_size, \
                          n_ins= self.n_ins, \
                          hidden_layers_sizes = self.hp.hidden_layers_sizes, \
                          n_outs = self.n_outs, \
                          corruption_levels = self.hp.corruption_levels,\
                          rng = self.rng,\
                          pretrain_lr = self.hp.pretraining_lr, \
                          finetune_lr = self.hp.finetuning_lr,\
                          input_divider = self.input_divider )

    def train(self):
        self.pretrain()
        if not self.job_tree:
            # if job_tree is True, finetuning was already performed
            self.finetune()

    def pretrain(self):
        print "STARTING PRETRAINING"

        printout_acc = 0.0
        last_error = 0.0

        start_time = time.clock()  
        ## Pre-train layer-wise 
        for i in xrange(self.classifier.n_layers):
            # go through pretraining epochs 
            for epoch in xrange(self.hp.pretraining_epochs_per_layer):
                # go through the training set
                for batch_index in xrange(self.n_train_batches):
                    c = self.classifier.pretrain_functions[i](batch_index)

                    printout_acc += c / self.printout_frequency
                    if (batch_index+1) % self.printout_frequency == 0:
                        print batch_index, "reconstruction cost avg=", printout_acc
                        last_error = printout_acc
                        printout_acc = 0.0
                        
                print 'Pre-training layer %i, epoch %d, cost '%(i,epoch),c

            self.job_splitter(i+1, time.clock()-start_time, last_error)
     
        end_time = time.clock()

        print ('Pretraining took %f minutes' %((end_time-start_time)/60.))

    # Save time by reusing intermediate results
    def job_splitter(self, current_pretraining_layer, pretraining_time, last_error):

        state_copy = None
        original_classifier = None

        if self.job_tree and current_pretraining_layer in self.num_hidden_layers_to_try:
            for lr in self.finetuning_lr_to_try:
                sys.stdout.flush()
                sys.stderr.flush()

                state_copy = copy.copy(self.hp)

                self.hp.update({'num_hidden_layers':current_pretraining_layer, \
                            'finetuning_lr':lr,\
                            'pretraining_time':pretraining_time,\
                            'last_reconstruction_error':last_error})

                original_classifier = self.classifier
                print "ORIGINAL CLASSIFIER MEANS",original_classifier.get_params_means()
                self.classifier = SdA.copy_reusing_lower_layers(original_classifier, current_pretraining_layer, new_finetuning_lr=lr)
                
                self.finetune()
            
                self.insert_finished_job()

                print "NEW CLASSIFIER MEANS AFTERWARDS",self.classifier.get_params_means()
                print "ORIGINAL CLASSIFIER MEANS AFTERWARDS",original_classifier.get_params_means()
                self.classifier = original_classifier
                self.hp = state_copy

    def insert_finished_job(self):
        job = copy.copy(self.hp)
        job[jobman.sql.STATUS] = jobman.sql.DONE
        job[jobman.sql.EXPERIMENT] = self.experiment

        # don,t try to store arrays in db
        job['hidden_layers_sizes'] = job.hidden_layers_sizes[0]
        job['corruption_levels'] = job.corruption_levels[0]

        print "Will insert finished job", job
        jobman.sql.insert_dict(jobman.flatten(job), self.results_db)

    def finetune(self):
        print "STARTING FINETUNING"

        index   = T.lscalar()    # index to a [mini]batch 
        minibatch_size = self.hp.minibatch_size

        # create a function to compute the mistakes that are made by the model
        # on the validation set, or testing set
        test_model = theano.function([index], self.classifier.errors,
                 givens = {
                   self.classifier.x: self.test_set_x[index*minibatch_size:(index+1)*minibatch_size] / self.input_divider,
                   self.classifier.y: self.test_set_y[index*minibatch_size:(index+1)*minibatch_size]})

        validate_model = theano.function([index], self.classifier.errors,
                givens = {
                   self.classifier.x: self.valid_set_x[index*minibatch_size:(index+1)*minibatch_size] / self.input_divider,
                   self.classifier.y: self.valid_set_y[index*minibatch_size:(index+1)*minibatch_size]})


        # early-stopping parameters
        patience              = 10000 # look as this many examples regardless
        patience_increase     = 2.    # wait this much longer when a new best is 
                                      # found
        improvement_threshold = 0.995 # a relative improvement of this much is 
                                      # considered significant
        validation_frequency  = min(self.n_train_batches, patience/2)
                                      # go through this many 
                                      # minibatche before checking the network 
                                      # on the validation set; in this case we 
                                      # check every epoch 

        best_params          = None
        best_validation_loss = float('inf')
        test_score           = 0.
        start_time = time.clock()

        done_looping = False
        epoch = 0

        printout_acc = 0.0

        if not self.hp.has_key('max_finetuning_epochs'):
            self.hp.max_finetuning_epochs = 1000

        while (epoch < self.hp.max_finetuning_epochs) and (not done_looping):
            epoch = epoch + 1
            for minibatch_index in xrange(self.n_train_batches):

                cost_ij = self.classifier.finetune(minibatch_index)
                iter    = epoch * self.n_train_batches + minibatch_index

                printout_acc += cost_ij / float(self.printout_frequency * minibatch_size)
                if (iter+1) % self.printout_frequency == 0:
                    print iter, "cost avg=", printout_acc
                    printout_acc = 0.0

                if (iter+1) % validation_frequency == 0: 
                    
                    validation_losses = [validate_model(i) for i in xrange(self.n_valid_batches)]
                    this_validation_loss = numpy.mean(validation_losses)
                    print('epoch %i, minibatch %i/%i, validation error %f %%' % \
                           (epoch, minibatch_index+1, self.n_train_batches, \
                            this_validation_loss*100.))


                    # if we got the best validation score until now
                    if this_validation_loss < best_validation_loss:

                        #improve patience if loss improvement is good enough
                        if this_validation_loss < best_validation_loss *  \
                               improvement_threshold :
                            patience = max(patience, iter * patience_increase)

                        # save best validation score and iteration number
                        best_validation_loss = this_validation_loss
                        best_iter = iter

                        # test it on the test set
                        test_losses = [test_model(i) for i in xrange(self.n_test_batches)]
                        test_score = numpy.mean(test_losses)
                        print(('     epoch %i, minibatch %i/%i, test error of best '
                              'model %f %%') % 
                                     (epoch, minibatch_index+1, self.n_train_batches,
                                      test_score*100.))


            if patience <= iter :
                done_looping = True
                break

        end_time = time.clock()
        self.hp.update({'finetuning_time':end_time-start_time,\
                    'best_validation_error':best_validation_loss,\
                    'test_score':test_score,
                    'num_finetuning_epochs':epoch})
        print(('Optimization complete with best validation score of %f %%,'
               'with test performance %f %%') %  
                     (best_validation_loss * 100., test_score*100.))
        print ('The finetuning ran for %f minutes' % ((end_time-start_time)/60.))