cb_run.py

###############################################################################
import os
import time
import sys
import multiprocessing as mp
from multiprocessing import dummy as multiprocessing
import tensorflow as tf
import logging
import PIL
import matplotlib.pyplot as plt
import numpy as np
import tensorflow as tf
import socket
from inception_utils import *
import inception
import h5py
slim = tf.contrib.slim


DATA_DIR = './imagenet'
MEM_DIR = './results'
CHECKPOINT = 'inception_v3.ckpt'
BATCH_SIZE = 128
TIME_START = time.time()
NUM_CLASSES = 1000
SAVE_FREQ = 10


def remove_players(model, players):
    '''Remove selected players (filters) in the Inception-v3 network.'''
    if isinstance(players, str):
        players = [players]
    for player in players:
        variables = layer_dic['_'.join(player.split('_')[:-1])]
        var_vals = model.sess.run(variables)
        for var, var_val in zip(variables, var_vals):
            if 'variance' in var.name:
                var_val[..., int(player.split('_')[-1])] = 1.
            elif 'beta' in var.name:
                pass
            else:
                var_val[..., int(player.split('_')[-1])] = 0.
            var.load(var_val, model.sess)
        

def return_player_output(model, player):
    '''The output of a filter.'''
    layer = '_'.join(player.split('_')[:-1])
    layer = '/'.join(layer.split('/')[1:])
    unit = int(player.split('_')[-1])
    return model.ends[layer][..., unit]


def one_iteration(
    model, 
    players,
    images, 
    labels, 
    chosen_players=None,
    c=None, 
    metric='accuracy',
    truncation=None
):
    '''One iteration of Neuron-Shapley algoirhtm.'''
    model.restore(CHECKPOINT)
    # Original performance of the model with all players present.
    init_val = value(model, images, labels, metric)
    if c is None:
        c = {i: np.array([i]) for i in range(len(players))}
    elif not isinstance(c, dict):
        c = {i: np.where(c==i)[0] for i in set(c)}
    if truncation is None:
        truncation = len(c.keys())
    if chosen_players is None:
        chosen_players = np.arange(len(c.keys()))
    # A random ordering of players
    idxs = np.random.permutation(len(c.keys()))
    # -1 default value for players that have already converged
    marginals = -np.ones(len(c.keys()))
    marginals[chosen_players] = 0.
    t = time.time()
    truncation_counter = 0
    old_val = init_val.copy()
    for n, idx in enumerate(idxs[::-1]):
        if idx in chosen_players:
            if old_val is None:
                old_val = value(model, images, labels, metric)
            remove_players(model, players[c[idx]])
            new_val = value(model, images, labels, metric)
            marginals[c[idx]] = (old_val - new_val) / len(c[idx])
            old_val = new_val
            if isinstance(truncation, int):
                if n >= truncation:
                    break
            else:
                if n%10 == 0:
                    print(n, time.time() - t, new_val)
                val_diff = new_val - base_value
                if metric == 'accuracy' and val_diff <= truncation:
                    truncation_counter += 1
                elif (metric == 'xe_loss' and
                      val_diff <= truncation * np.abs(base_value)):
                    truncation_counter += 1
                elif metric == 'binary' and val_diff <= truncation:
                    truncation_counter += 1
                elif metric == 'logit' and new_val <= truncation * np.abs(init_val):
                    truncation_counter += 1
                else:
                    truncation_counter = 0
                if truncation_counter > 5:
                    break
        else:
            old_val = None
            remove_players(model, players[c[idx]])        
    return idxs.reshape((1, -1)), marginals.reshape((1, -1))


def sess_run(model, variable, images, labels=None, batch_size=BATCH_SIZE):
    '''Divides inputs into smaller chunks and performs sess.run'''
    output = []    
    num_batches = int(np.ceil(len(images) / batch_size))
    for batch in range(num_batches):
        batch_idxs = np.arange(
            batch * batch_size, min((batch+1) * batch_size, len(images))
        )
        input_dic = {model.input: images[batch_idxs]}
        if labels is not None:
            input_dic[model.y_input] = labels[batch_idxs]
        output.append(model.sess.run(variable, input_dic))
    try:
        return np.concatenate(output, 0)
    except:
        return np.array(output)
    
    
def value(model, images, labels, metric='accuracy', batch_size=BATCH_SIZE):
    '''The performance of the model on given image-label pairs.'''
    if isinstance(labels, str):
        labels = np.array([model.label_to_id(labels)] * len(images))
    elif isinstance(labels, int):
        labels = np.array([labels] * len(images))
    num_batches = int(np.ceil(len(images) / batch_size))
    val = 0.
    if metric == 'accuracy':
        val = np.mean(sess_run(
            model, model.accuracy, images, labels, batch_size=batch_size))
    elif metric=='xe_loss':
        probs = sess_run(
            model, model.probs, images, labels, batch_size=batch_size)
        val = np.mean(np.log(probs[np.arange(len(probs)), labels]))
    elif metric=='binary':
        probs = sess_run(
            model, model.probs, images, labels, batch_size=batch_size)
        preds = np.argmax(probs, -1)
        key_labels = np.expand_dims(list(set(labels[labels != -1])), 0)
        corrects_1 = np.sum(labels[labels != -1] == preds[labels != -1])
        corrects_2 = np.sum(1 - np.equal(preds[labels == -1], key_labels))
        val = (corrects_1 + corrects_2) * 1. / len(preds)
    elif metric=='logit':
        logits = sess_run(
            model, model.ends['logits'], images, labels, batch_size=batch_size)
        class_logits = np.mean(logits[np.arange(len(logits)), labels])
        return class_logits
    else:
        raise ValueError('Invalid metric!')
    return val


def adversarial_attack(model, images, target_label, epsilon=16./255, iters=30, 
               norm='l_inf', perturb=False, delta=16./255/20,
               minval=0., maxval=1., batch_size=16):
    '''Creates iterative adversarial attacks with PGD.'''
    
    if isinstance(target_label, int):
        target_label = np.array([target_label] * len(images))
        
    def batch_attack(x, y, gradient):
        
        if not epsilon:
            return x
        x_hat = x.copy()
        if perturb:
            if norm == 'l_2':
                r = np.random.normal(size=x.shape)
                r_norm =  np.sqrt(np.sum(r ** 2, axis=tuple(np.arange(1, len(x.shape))), keepdims=True))
                x_hat += r / r_norm * delta
            elif norm == 'l_inf':
                x_hat += delta * np.sign(np.random.random(x.shape))
        for nu in range(iters):
            grd = model.sess.run(gradient, {
                model.input: x_hat, 
                model.y_input: y
            })
            if norm == 'l_2':
                grd_norm = np.sqrt(np.sum(grd ** 2, axis=tuple(np.arange(1, len(x.shape))), 
                                      keepdims=True))
                grd /= grd_norm + 1e-8
                prtrb = x_hat - grd * delta - x 
                prtrb_norm = np.sqrt(np.sum(prtrb ** 2, axis=tuple(np.arange(1, len(x.shape))),
                                            keepdims=True))
                prj_coef = (prtrb_norm <= epsilon) * 1. + (prtrb_norm > epsilon) * prtrb_norm / epsilon
                prtrb /= prj_coef
            elif norm == 'l_inf':
                #print(nu, np.max(np.abs(x - x_hat).reshape((-1, np.prod(x.shape[1:]))), -1),
                     #model.sess.run(model.loss, {model.input: x_hat, model.y_input: y}))
                grd = np.sign(grd)
                prtrb = x_hat - grd * delta - x 
                prtrb = np.clip(prtrb, -epsilon, epsilon)
            else:
                raise ValueError('Invalid Norm {}'.format(norm))
            x_hat = np.clip(x + prtrb, minval, maxval)
        return x_hat
    
    gradient = tf.gradients(model.loss, model.input)[0]
    x = []
    num_batches = int(np.ceil(len(images) / batch_size))
    for batch in range(num_batches):
        print(batch)
        batch_idxs = np.arange(
            batch * batch_size, min((batch+1) * batch_size, len(images))
        )
        x.append(batch_attack(images[batch_idxs], target_label[batch_idxs], gradient))
    return np.concatenate(x, 0)


def load_images(files_dir, filenames, num_workers=0):
    
    file_dirs = [os.path.join(files_dir, filename) for filename in filenames]
    if num_workers:
        pool = multiprocessing.Pool(num_workers)
        images = pool.map(lambda f: np.array(Image.open(f)), file_dirs)
        pool.close()
    else:
        images = [np.array(Image.open(f)) for f in file_dirs]
    return np.array(images)/255


def return_target_images(image_list, key):
    
    all_classes = list(set([image.split('/')[-2] for image in image_list]))
    if key == 'all' or key == 'rnd' or key == '-all':
        target_classes = all_classes
    elif key[0] == '-':
        target_classes = list(set(all_classes) - set(key[1:].split('+')))
    else:
        target_classes = key.split('+')
    target_images = [filename for filename in image_list
                     if filename.split('/')[-2] in target_classes]
    return target_images


def make_adv_images(key, images, model):
    
    adv_images, adv_labels = [], []
    if key == '-all':
        adv_labels = np.random.choice(np.arange(1001), len(images))
        adv_images = adversarial_attack(model, images, adv_labels)
        return np.array(adv_images), np.array(adv_labels)
    keys = key[1:].split('+')
    num_label_imgs = len(images) // len(keys)
    for i, k in enumerate(keys):
        key_id = model.label_to_id(k)
        adv_labels.extend(key_id * np.ones(num_label_imgs).astype(int))
        label_imgs = images[i * num_label_imgs: (i+1) * num_label_imgs]
        adv_images.extend(adversarial_attack(model, label_imgs, key_id))
    return np.array(adv_images), np.array(adv_labels)


def load_images_labels(key, num_images, max_sample_size, model, max_size=25000):

    image_list = open('val_images.txt').read().split('\n')[:max_size]
    val_images = return_target_images(image_list, key)
    num_images = min(num_images, max_sample_size, len(val_images))
    filenames = np.random.choice(val_images, num_images, replace=False)
    images = load_images(
        os.path.join(DATA_DIR),
        filenames,
        0)
    if key[0] == '-':
        return make_adv_images(key, images, model)
    labels = np.array([model.label_to_id(filename.split('/')[-2]) 
                       for filename in filenames])
    return images, labels



key = sys.argv[1] #Class name. Use 'all' for overll performance.
model_scope = 'InceptionV3'
metric = sys.argv[2] #metric one of accuracy, binary, xe_loss.
num_images = int(sys.argv[3]) #Number of validation images.
bound = 'Bernstein'
truncation = 0.2
max_sample_size = 128
adversarial = (sys.argv[4] == 'True') #If True, computes contributions for adversarial setting.
time.sleep(10 * np.random.random())
## Experiment Directory
experiment_dir = os.path.join(
    MEM_DIR, 'NShap/inceptionv3/{}_{}_new'.format(metric, key))
if not tf.gfile.Exists(experiment_dir):
    tf.gfile.MakeDirs(experiment_dir)
## CB directory
if max_sample_size is None or max_sample_size > num_images:
    max_sample_size = num_images
experiment_name = 'cb_{}_{}_{}'.format(bound, truncation, max_sample_size)
if adversarial:
    experiment_name = 'ADV' + experiment_name
cb_dir = os.path.join(experiment_dir, experiment_name)
if not tf.gfile.Exists(cb_dir):
    tf.gfile.MakeDirs(cb_dir)
## Load Model and find all convolutional filters
tf.reset_default_graph()
model = inception.inpcetion_instance(checkpoint=CHECKPOINT)
model_variables = tf.global_variables(scope=model_scope)
convs = ['/'.join(k.name.split('/')[:-1]) for k in model_variables if 'weights'
         in k.name and 'Aux' not in k.name and 'Logits' not in k.name]
layer_dic = {conv: [var for var in model_variables if conv in var.name]
             for conv in convs}
## Load the list of all players (filters) else save
if tf.gfile.Exists(os.path.join(experiment_dir, 'players.txt')):
    players = open(os.path.join(
        experiment_dir, 'players.txt')).read().split(',')
    players = np.array(players)
else:
    players = []
    var_dic = {var.name: var for var in model_variables}
    for conv in layer_dic.keys():
        players.append(['{}_{}'.format(conv, i) for i in
                        range(var_dic[conv + '/weights:0'].shape[-1])])
    players = np.sort(np.concatenate(players))
    open(os.path.join(experiment_dir, 'players.txt'), 'w').write(
        ','.join(players))
## Load metric's base value (random performance)
if metric == 'accuracy':
    base_value = 1./NUM_CLASSES
elif metric == 'xe_loss':
    base_value = -np.log(NUM_CLASSES)
elif metric == 'binary':
    base_value = 0.5
elif metric == 'logit':
    base_value = 0
else:
    raise ValueError('Invalid metric!')
## Assign expriment number to this specific run of cb_run.py
results = [saved for saved in tf.gfile.ListDirectory(cb_dir)
           if 'agg' not in saved and '.h5' in saved]
experiment_number = 0
if len(results):
    results_experiment_numbers = [int(result.split('.')[-2].split('_')[-1][1:])
                                  for result in results]
    experiment_number += np.max(results_experiment_numbers) + 1
print(experiment_number)
save_dir = os.path.join(
    cb_dir, '{}.h5'.format('0' + str(experiment_number).zfill(5))
)
## Create placeholder for results in save ASAP to prevent having the 
## same expriment_number with other parallel cb_run.py scripts
mem_tmc = np.zeros((0, len(players)))
idxs_tmc = np.zeros((0, len(players))).astype(int)
with h5py.File(save_dir, 'w') as foo:
    foo.create_dataset("mem_tmc", data=mem_tmc, compression='gzip')
    foo.create_dataset("idxs_tmc", data=idxs_tmc, compression='gzip')
## Running CB-Shapley
c = None
if c is None:
    c = {i: np.array([i]) for i in range(len(players))}
elif not isinstance(c, dict):
    c = {i: np.where(np.array(c)==i)[0] for i in set(list(c))}

counter = 0
while True:
    ## Load the list of players (filters) that are determined to be not confident enough
    ## by the cb_aggregate.py running in parallel to this script
    if tf.gfile.Exists(os.path.join(cb_dir, 'chosen_players.txt')):
        chosen_players = open(os.path.join(
                cb_dir, 'chosen_players.txt')).read()
        chosen_players = np.array(chosen_players.split(',')).astype(int)
        if len(chosen_players) == 1:
            break
    else:
        chosen_players = None
        
    t_init = time.time()
    iter_images, iter_labels = load_images_labels(
        '-'+ key if adversarial else key,
        num_images,
        max_sample_size,
        model,
        max_size=25000)
    
    if metric == 'binary':
        rnd_images, _ = load_images_labels(
            'rnd',
            len(iter_images),
            max_sample_size,
            model,
            max_size=25000)
        iter_images = np.concatenate([iter_images, rnd_images])
        iter_labels = np.concatenate([
            iter_labels, 
            -np.ones(len(rnd_images)).astype(int)
        ])
        
    idxs, vals =  one_iteration(
        model=model,
        players=players,
        images=iter_images,
        labels=iter_labels,
        chosen_players=chosen_players,
        c=c,
        metric=metric,
        truncation=truncation
    )
    mem_tmc = np.concatenate([mem_tmc, vals])
    idxs_tmc = np.concatenate([idxs_tmc, idxs])
    ## Save results every SAVE_FREQ iterations
    if counter % SAVE_FREQ == SAVE_FREQ - 1:
        with h5py.File(save_dir, 'w') as foo:
            foo.create_dataset("mem_tmc", data=mem_tmc, compression='gzip')
            foo.create_dataset("idxs_tmc", data=idxs_tmc, compression='gzip')
            
    counter += 1
    print(time.time() - t_init, time.time() - TIME_START)
    if not tf.test.is_gpu_available():
        print('No gpu!')
        print(time.time() - TIME_START)
    else:
        print('There is a gpu!')
        print(time.time() - TIME_START)