valid_resnet.py

import tensorflow as tf

config = tf.ConfigProto()
config.gpu_options.allow_growth = True
import numpy as np
import scipy
import os
from scipy.spatial.distance import cdist
import scipy.io
import sys

try:
    import cPickle
except:
    import _pickle as cPickle
# Syspath for the folder with the utils files
# sys.path.insert(0, "/data/sylvestre")

import utils_data
import utils_cifar

######### Modifiable Settings ##########
batch_size = 128  # Batch size
nb_cl = 100  # Classes per group
nb_groups = 10  # Number of groups
top = 5  # Choose to evaluate the top X accuracy
is_cumul = 'cumul'  # Evaluate on the cumul of classes if 'cumul', otherwise on the first classes
gpu = '0'  # Used GPU
########################################

######### Paths  ##########
# Working station
train_path = '/data/datasets/imagenets72'
save_path = '/data/srebuffi/backup/'

###########################

# Load ResNet settings
str_mixing = str(nb_cl) + 'mixing.pickle'
with open(str_mixing, 'rb') as fp:
    mixing = cPickle.load(fp)

str_settings_resnet = str(nb_cl) + 'settings_resnet.pickle'
with open(str_settings_resnet, 'rb') as fp:
    order = cPickle.load(fp)
    files_valid = cPickle.load(fp)
    files_train = cPickle.load(fp)

# Load class means
str_class_means = str(nb_cl) + 'class_means.pickle'
with open(str_class_means, 'rb') as fp:
    class_means = cPickle.load(fp)

# Loading the labels
labels_dic, label_names, validation_ground_truth = utils_data.parse_devkit_meta(devkit_path)

# Initialization
acc_list = np.zeros((nb_groups, 3))

for itera in range(nb_groups):
    print("Processing network after {} increments\t".format(itera))
    # Evaluation on cumul of classes or original classes
    if is_cumul == 'cumul':
        eval_groups = np.array(range(itera + 1))
    else:
        eval_groups = [0]

    print("Evaluation on batches {} \t".format(eval_groups))
    # Load the evaluation files
    files_from_cl = []
    for i in eval_groups:
        files_from_cl.extend(files_valid[i])

    inits, scores, label_batch, loss_class, file_string_batch, op_feature_map = utils_icarl.reading_data_and_preparing_network(
        files_from_cl, gpu, itera, batch_size, train_path, labels_dic, mixing, nb_groups, nb_cl, save_path)

    with tf.Session(config=config) as sess:
        # Launch the prefetch system
        coord = tf.train.Coordinator()
        threads = tf.train.start_queue_runners(coord=coord)
        sess.run(inits)

        # Evaluation routine
        stat_hb1 = []
        stat_icarl = []
        stat_ncm = []

        for i in range(int(np.ceil(len(files_from_cl) / batch_size))):
            sc, l, loss, files_tmp, feat_map_tmp = sess.run(
                [scores, label_batch, loss_class, file_string_batch, op_feature_map])
            mapped_prototypes = feat_map_tmp[:, 0, 0, :]
            pred_inter = (mapped_prototypes.T) / np.linalg.norm(mapped_prototypes.T, axis=0)
            sqd_icarl = -cdist(class_means[:, :, 0, itera].T, pred_inter.T, 'sqeuclidean').T
            sqd_ncm = -cdist(class_means[:, :, 1, itera].T, pred_inter.T, 'sqeuclidean').T
            stat_hb1 += ([ll in best for ll, best in zip(l, np.argsort(sc, axis=1)[:, -top:])])
            stat_icarl += ([ll in best for ll, best in zip(l, np.argsort(sqd_icarl, axis=1)[:, -top:])])
            stat_ncm += ([ll in best for ll, best in zip(l, np.argsort(sqd_ncm, axis=1)[:, -top:])])

        coord.request_stop()
        coord.join(threads)

    print('Increment: %i' % itera)
    print('Hybrid 1 top ' + str(top) + ' accuracy: %f' % np.average(stat_hb1))
    print('iCaRL top ' + str(top) + ' accuracy: %f' % np.average(stat_icarl))
    print('NCM top ' + str(top) + ' accuracy: %f' % np.average(stat_ncm))
    acc_list[itera, 0] = np.average(stat_icarl)
    acc_list[itera, 1] = np.average(stat_hb1)
    acc_list[itera, 2] = np.average(stat_ncm)

    # Reset the graph to compute the numbers ater the next increment
    tf.reset_default_graph()

np.save('results_top' + str(top) + '_acc_' + is_cumul + '_cl' + str(nb_cl), acc_list)