train_netvlad_RGB_real_supervise.py

import datetime
#import torch
#device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
import argparse
import importlib
import math
import os
import socket
import sys

import numpy as np
from sklearn.neighbors import KDTree, NearestNeighbors
import generating_queries.generate_training_tuples_RGB_real_supervise as generate_dataset_tt
import generating_queries.generate_test_RGB_real_supervise_sets as generate_dataset_eval

import config as cfg
import evaluate
import loss.pointnetvlad_loss as PNV_loss
import models.Verification_RGB_real as VFC
import models.ImageNetVlad as INV
import torch
import torch.nn as nn
from loading_pointclouds import *
from tensorboardX import SummaryWriter
from torch.autograd import Variable
from torch.backends import cudnn
import scipy.io as sio

BASE_DIR = os.path.dirname(os.path.abspath(__file__))
sys.path.append(BASE_DIR)



cudnn.enabled = True

parser = argparse.ArgumentParser()
parser.add_argument('--log_dir', default='log/', help='Log dir [default: log]')
parser.add_argument('--results_dir', default='results/',
                    help='results dir [default: results]')
parser.add_argument('--positives_per_query', type=int, default=2,
                    help='Number of potential positives in each training tuple [default: 2]')
parser.add_argument('--negatives_per_query', type=int, default=18,
                    help='Number of definite negatives in each training tuple [default: 18]')
parser.add_argument('--max_epoch', type=int, default=30,
                    help='Epoch to run [default: 100]')
parser.add_argument('--batch_num_queries', type=int, default=2,
                    help='Batch Size during training [default: 2]')
parser.add_argument('--learning_rate', type=float, default=0.000005,
                    help='Initial learning rate [default: 0.000005]')
parser.add_argument('--momentum', type=float, default=0.9,
                    help='Initial learning rate [default: 0.9]')
parser.add_argument('--optimizer', default='adam',
                    help='adam or momentum [default: adam]')
parser.add_argument('--decay_step', type=int, default=200000,
                    help='Decay step for lr decay [default: 200000]')
parser.add_argument('--decay_rate', type=float, default=0.7,
                    help='Decay rate for lr decay [default: 0.7]')
parser.add_argument('--margin_1', type=float, default=0.5,
                    help='Margin for hinge loss [default: 0.5]')
parser.add_argument('--margin_2', type=float, default=0.2,
                    help='Margin for hinge loss [default: 0.2]')
parser.add_argument('--loss_function', default='triplet', choices=[
                    'triplet', 'quadruplet'], help='triplet or quadruplet [default: quadruplet]')
parser.add_argument('--loss_not_lazy', action='store_false',
                    help='If present, do not use lazy variant of loss')
parser.add_argument('--loss_ignore_zero_batch', action='store_true',
                    help='If present, mean only batches with loss > 0.0')
parser.add_argument('--triplet_use_best_positives', action='store_true',
                    help='If present, use best positives, otherwise use hardest positives')
parser.add_argument('--resume', action='store_true',
                    help='If present, restore checkpoint and resume training')
parser.add_argument('--dataset_folder', default='/mnt/NAS/data/cc_data/2D_RGB_real_full3',
                    help='PointNetVlad Dataset Folder')

FLAGS = parser.parse_args()
#cfg.EVAL_BATCH_SIZE = 12
cfg.GRID_X = 1080
cfg.GRID_Y = 1920
cfg.BASE_LEARNING_RATE = FLAGS.learning_rate
cfg.MOMENTUM = FLAGS.momentum
cfg.OPTIMIZER = FLAGS.optimizer
cfg.DECAY_STEP = FLAGS.decay_step
cfg.DECAY_RATE = FLAGS.decay_rate
cfg.MARGIN1 = FLAGS.margin_1
cfg.MARGIN2 = FLAGS.margin_2

cfg.TRIPLET_USE_BEST_POSITIVES = FLAGS.triplet_use_best_positives
cfg.LOSS_LAZY = FLAGS.loss_not_lazy
cfg.LOSS_IGNORE_ZERO_BATCH = FLAGS.loss_ignore_zero_batch

cfg.TRAIN_FILE = 'generating_queries/train_pickle/training_queries_baseline_0.pickle'
cfg.TEST_FILE = 'generating_queries/train_pickle/test_queries_baseline_0.pickle'
cfg.DB_FILE = 'generating_queries/train_pickle/db_queries_baseline_0.pickle'

cfg.LOG_DIR = FLAGS.log_dir
if not os.path.exists(cfg.LOG_DIR):
    os.mkdir(cfg.LOG_DIR)
LOG_FOUT = open(os.path.join(cfg.LOG_DIR, 'log_train.txt'), 'w')
LOG_FOUT.write(str(FLAGS) + '\n')

cfg.RESULTS_FOLDER = FLAGS.results_dir
print("cfg.RESULTS_FOLDER:"+str(cfg.RESULTS_FOLDER))

# Load dictionary of training queries
TRAINING_QUERIES = get_queries_dict(cfg.TRAIN_FILE)
TEST_QUERIES = get_queries_dict(cfg.TEST_FILE)
DB_QUERIES = get_queries_dict(cfg.DB_FILE)

cfg.BN_INIT_DECAY = 0.5
cfg.BN_DECAY_DECAY_RATE = 0.5
BN_DECAY_DECAY_STEP = float(cfg.DECAY_STEP)
cfg.BN_DECAY_CLIP = 0.99

HARD_NEGATIVES = {}
TRAINING_LATENT_VECTORS = []

TOTAL_ITERATIONS = 0

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
cfg.margin = 0.1

def get_bn_decay(batch):
    bn_momentum = cfg.BN_INIT_DECAY * \
        (cfg.BN_DECAY_DECAY_RATE **
         (batch * cfg.BATCH_NUM_QUERIES // BN_DECAY_DECAY_STEP))
    return min(cfg.BN_DECAY_CLIP, 1 - bn_momentum)


def log_string(out_str):
    LOG_FOUT.write(out_str + '\n')
    LOG_FOUT.flush()
    print(out_str)

# learning rate halfed every 5 epoch

def get_learning_rate(epoch):
    learning_rate = cfg.BASE_LEARNING_RATE * ((0.9) ** (epoch // 5))
    learning_rate = max(learning_rate, 0.00001)  # CLIP THE LEARNING RATE!
    return learning_rate


def train():
    global HARD_NEGATIVES, TOTAL_ITERATIONS, TRAINING_QUERIES
    bn_decay = get_bn_decay(0)
    
    TRAINING_QUERIES = get_queries_dict(cfg.TRAIN_FILE)
    TEST_QUERIES = get_queries_dict(cfg.TEST_FILE)
    DB_QUERIES = get_queries_dict(cfg.DB_FILE)

    cfg.RESULTS_FOLDER = os.path.join("results/", 'scene2')
    if not os.path.isdir(cfg.RESULTS_FOLDER):
        os.mkdir(cfg.RESULTS_FOLDER)
    if cfg.LOSS_FUNCTION_RGB == 'quadruplet':
        loss_function = PNV_loss.quadruplet_loss
    elif cfg.LOSS_FUNCTION_RGB == 'triplet_RI':
        loss_function = PNV_loss.triplet_loss_RI
    else:
        loss_function = PNV_loss.triplet_loss
    learning_rate = get_learning_rate(0)
    
    train_writer = SummaryWriter(os.path.join(cfg.LOG_DIR, 'train'))

    model = INV.ImageNetVlad(global_feat=True, feature_transform=True,
                             max_pool=False, output_dim=cfg.FEATURE_OUTPUT_DIM, grid_x = cfg.GRID_X, grid_y = cfg.GRID_Y)
    model = model.to(device)

    parameters = filter(lambda p: p.requires_grad, model.parameters())

    if cfg.OPTIMIZER == 'momentum':
        optimizer = torch.optim.SGD(
            parameters, learning_rate, momentum=cfg.MOMENTUM)
    elif cfg.OPTIMIZER == 'adam':
        optimizer = torch.optim.Adam(parameters, learning_rate)
    else:
        optimizer = None
        exit(0)

    if FLAGS.resume:
        resume_filename = cfg.LOG_DIR + "checkpoint.pth.tar"
        print("Resuming From ", resume_filename)
        checkpoint = torch.load(resume_filename)
        saved_state_dict = checkpoint['state_dict']
        starting_epoch = checkpoint['epoch']
        print("starting_epoch:"+str(starting_epoch))
        #starting_epoch = starting_epoch +1
        TOTAL_ITERATIONS = starting_epoch * len(TRAINING_QUERIES)
        starting_epoch = starting_epoch +1
        model.load_state_dict(saved_state_dict)
        optimizer.load_state_dict(checkpoint['optimizer'])

    else:
        starting_epoch = 0


    LOG_FOUT.write(cfg.cfg_str())
    LOG_FOUT.write("\n")
    LOG_FOUT.flush()

    try:
        potential_positives
    except NameError:
        potential_positives = None
        potential_distributions = None
        trusted_positives = None
    
    for epoch in range(starting_epoch, cfg.MAX_EPOCH):
        print(epoch)
        print()

        log_string('**** EPOCH %03d ****' % (epoch))
        sys.stdout.flush()
        
        train_one_epoch(model, optimizer, train_writer, loss_function, epoch, TRAINING_QUERIES, TEST_QUERIES, DB_QUERIES)
        
        log_string('EVALUATING...')
        cfg.OUTPUT_FILE = os.path.join(cfg.RESULTS_FOLDER , 'results_' + str(epoch) + '.txt')
        eval_recall_1, eval_recall_5, eval_recall_10 = evaluate.evaluate_model_RGB_real(model,optimizer,epoch,True)

        log_string('EVAL RECALL_1: %s' % str(eval_recall_1))
        log_string('EVAL RECALL_5: %s' % str(eval_recall_5))
        log_string('EVAL RECALL_10: %s' % str(eval_recall_10))


def train_one_epoch(model, optimizer, train_writer, loss_function, epoch, TRAINING_QUERIES, TEST_QUERIES, DB_QUERIES):
    global HARD_NEGATIVES
    global TRAINING_LATENT_VECTORS, TOTAL_ITERATIONS

    is_training = True
    sampled_neg = 4000
    # number of hard negatives in the training tuple
    # which are taken from the sampled negatives
    num_to_take = 10

    # Shuffle train files
    train_file_idxs = np.arange(0, len(TRAINING_QUERIES.keys()))
    np.random.shuffle(train_file_idxs)
    
    for i in range(len(train_file_idxs)//cfg.BATCH_NUM_QUERIES):
    # for i in range(1):
        batch_keys = train_file_idxs[i *
                                     cfg.BATCH_NUM_QUERIES:(i+1)*cfg.BATCH_NUM_QUERIES]
        q_tuples = []
        
        faulty_tuple = False
        no_other_neg = False
        for j in range(cfg.BATCH_NUM_QUERIES):
            if (len(DB_QUERIES[batch_keys[j]]["positives"]) < cfg.TRAIN_POSITIVES_PER_QUERY):
                faulty_tuple = True
                break
            # no cached feature vectors
            if (len(TRAINING_LATENT_VECTORS) == 0):
                q_tuples.append(
                    get_query_tuple_RGB_real_supervise(DB_QUERIES[batch_keys[j]], cfg.TRAIN_POSITIVES_PER_QUERY, cfg.TRAIN_NEGATIVES_PER_QUERY,
                                    DB_QUERIES, hard_neg=[], other_neg=True))
                #print("q_tuples:"+str(q_tuples))

            elif (len(HARD_NEGATIVES.keys()) == 0):
                query = get_feature_representation(
                    DB_QUERIES[batch_keys[j]]['query'], model)
                random.shuffle(DB_QUERIES[batch_keys[j]]['negatives'])
                negatives = DB_QUERIES[batch_keys[j]
                                             ]['negatives'][0:sampled_neg]
                hard_negs = get_random_hard_negatives(
                    query, negatives, num_to_take)
                q_tuples.append(
                    get_query_tuple_RGB_real_supervise(DB_QUERIES[batch_keys[j]], cfg.TRAIN_POSITIVES_PER_QUERY, cfg.TRAIN_NEGATIVES_PER_QUERY,
                                    DB_QUERIES, hard_negs, other_neg=True))
            else:
                query = get_feature_representation(
                    DB_QUERIES[batch_keys[j]]['query'], model)
                random.shuffle(DB_QUERIES[batch_keys[j]]['negatives'])
                negatives = DB_QUERIES[batch_keys[j]
                                             ]['negatives'][0:sampled_neg]
                hard_negs = get_random_hard_negatives(
                    query, negatives, num_to_take)
                hard_negs = list(set().union(
                    HARD_NEGATIVES[batch_keys[j]], hard_negs))
                q_tuples.append(
                    get_query_tuple_RGB_real_supervise(DB_QUERIES[batch_keys[j]], cfg.TRAIN_POSITIVES_PER_QUERY, cfg.TRAIN_NEGATIVES_PER_QUERY,
                                    DB_QUERIES, hard_negs, other_neg=True))
            
            if (q_tuples[j][3].shape[2] != 3):
                no_other_neg = True
                break

        if(faulty_tuple):
            log_string('----' + str(i) + '-----')
            log_string('----' + 'FAULTY TUPLE' + '-----')
            continue

        if(no_other_neg):
            log_string('----' + str(i) + '-----')
            log_string('----' + 'NO OTHER NEG' + '-----')
            continue
        
        queries = []
        positives = []
        negatives = []
        other_neg = []
        for k in range(len(q_tuples)):
            queries.append(q_tuples[k][0])
            positives.append(q_tuples[k][1])
            negatives.append(q_tuples[k][2])
            other_neg.append(q_tuples[k][3])

        queries = np.array(queries, dtype=np.float32)
        queries = np.expand_dims(queries, axis=1)
        other_neg = np.array(other_neg, dtype=np.float32)
        other_neg = np.expand_dims(other_neg, axis=1)
        positives = np.array(positives, dtype=np.float32)
        negatives = np.array(negatives, dtype=np.float32)

        log_string('----' + str(i) + '-----')
        if (len(queries.shape) != 5):
            log_string('----' + 'FAULTY QUERY' + '-----')
            continue

        model.train()
        optimizer.zero_grad()
        
        output_queries, output_positives, output_negatives, output_other_neg = run_model(
            model, queries, positives, negatives, other_neg)
        
        loss = loss_function(output_queries, output_positives, output_negatives,  0.1,  use_min=cfg.TRIPLET_USE_BEST_POSITIVES, lazy=cfg.LOSS_LAZY, ignore_zero_loss=cfg.LOSS_IGNORE_ZERO_BATCH)
        #loss = loss_function(output_queries, output_positives, output_negatives,  rot_output_queries, rot_output_positives, rot_output_negatives, 0.1,  use_min=cfg.TRIPLET_USE_BEST_POSITIVES, lazy=False, ignore_zero_loss=cfg.LOSS_IGNORE_ZERO_BATCH)
        loss.backward()
        optimizer.step()

        log_string('batch loss: %f' % loss)
        train_writer.add_scalar("Loss", loss.cpu().item(), TOTAL_ITERATIONS)
        TOTAL_ITERATIONS += cfg.BATCH_NUM_QUERIES

def get_feature_representation(filename, model):
    model.eval()
    queries = load_image_files([filename],False)
    queries = np.expand_dims(queries, axis=1)

    with torch.no_grad():
        q = torch.from_numpy(queries).float()
        q = q.to(device)
        output = model(q)
    output = output.detach().cpu().numpy()
    output = np.squeeze(output)
    model.train()
    return output


def get_random_hard_negatives(query_vec, random_negs, num_to_take):
    global TRAINING_LATENT_VECTORS

    latent_vecs = []
    for j in range(len(random_negs)):
        latent_vecs.append(TRAINING_LATENT_VECTORS[random_negs[j]])

    latent_vecs = np.array(latent_vecs)
    nbrs = KDTree(latent_vecs)
    distances, indices = nbrs.query(np.array([query_vec]), k=num_to_take)
    hard_negs = np.squeeze(np.array(random_negs)[indices[0]])
    hard_negs = hard_negs.tolist()
    return hard_negs


def get_latent_vectors(model, dict_to_process):
    train_file_idxs = np.arange(0, len(dict_to_process.keys()))

    batch_num = cfg.BATCH_NUM_QUERIES * \
        (1 + cfg.TRAIN_POSITIVES_PER_QUERY + cfg.TRAIN_NEGATIVES_PER_QUERY + 1)
    q_output = []

    model.eval()

    for q_index in range(len(train_file_idxs)//batch_num):
        file_indices = train_file_idxs[q_index *
                                       batch_num:(q_index+1)*(batch_num)]
        file_names = []
        for index in file_indices:
            file_names.append(dict_to_process[index]["query"])
        queries = load_image_files(file_names,False)

        feed_tensor = torch.from_numpy(queries).float()
        feed_tensor = feed_tensor.unsqueeze(1)
        feed_tensor = feed_tensor.to(device)
        with torch.no_grad():
            out = model(feed_tensor)

        out = out.detach().cpu().numpy()
        out = np.squeeze(out)

        q_output.append(out)

    q_output = np.array(q_output)
    if(len(q_output) != 0):
        q_output = q_output.reshape(-1, q_output.shape[-1])

    # handle edge case
    for q_index in range((len(train_file_idxs) // batch_num * batch_num), len(dict_to_process.keys())):
        index = train_file_idxs[q_index]
        queries = load_image_files([dict_to_process[index]["query"]],False)
        queries = np.expand_dims(queries, axis=1)

        with torch.no_grad():
            queries_tensor = torch.from_numpy(queries).float()
            o1 = model(queries_tensor.to(device))

        output = o1.detach().cpu().numpy()
        output = np.squeeze(output)
        if (q_output.shape[0] != 0):
            q_output = np.vstack((q_output, output))
        else:
            q_output = output

    model.train()
    return q_output


def run_model(model, queries, positives, negatives, other_neg, require_grad=True):
    queries_tensor = torch.from_numpy(queries).float()
    positives_tensor = torch.from_numpy(positives).float()
    negatives_tensor = torch.from_numpy(negatives).float()
    other_neg_tensor = torch.from_numpy(other_neg).float()
    feed_tensor = torch.cat(
        (queries_tensor, positives_tensor, negatives_tensor, other_neg_tensor), 1)
    feed_tensor = feed_tensor.view((-1, 1, cfg.SIZED_GRID_X, cfg.SIZED_GRID_Y, 3))
    
    feed_tensor.requires_grad_(require_grad)
    feed_tensor = feed_tensor.to(device)
    if require_grad:
        output = model(feed_tensor)
    else:
        with torch.no_grad():
            output = model(feed_tensor)

    output = output.view(cfg.BATCH_NUM_QUERIES, -1, output.shape[-1])
    o1, o2, o3, o4 = torch.split(
        output, [1, cfg.TRAIN_POSITIVES_PER_QUERY, cfg.TRAIN_NEGATIVES_PER_QUERY, 1], dim=1)
    return o1, o2, o3, o4#, ro1, ro2, ro3, ro4 


if __name__ == "__main__":
    for i in range(1):
        train()