def evaluate(params):
# set up loader
data_json = osp.join(
'cache/prepro', params['dataset_splitBy'], 'data.json')
data_h5 = osp.join('cache/prepro', params['dataset_splitBy'], 'data.h5')
loader = GtMRCNLoader(data_h5=data_h5, data_json=data_json)
# load mode info
model_prefix = osp.join('output', params['dataset_splitBy'], params['id'])
infos = json.load(open(model_prefix+'.json'))
model_opt = infos['opt']
model_path = model_prefix + '.pth'
model = load_model(model_path, model_opt)
# loader's feats
feats_dir = '%s_%s_%s' % (
model_opt['net_name'], model_opt['imdb_name'], model_opt['tag'])
args.imdb_name = model_opt['imdb_name']
args.net_name = model_opt['net_name']
args.tag = model_opt['tag']
args.iters = model_opt['iters']
loader.prepare_mrcn(head_feats_dir=osp.join('cache/feats/', model_opt['dataset_splitBy'], 'mrcn', feats_dir),
args=args)
ann_feats = osp.join('cache/feats', model_opt['dataset_splitBy'], 'mrcn',
'%s_%s_%s_ann_feats.h5' % (model_opt['net_name'], model_opt['imdb_name'], model_opt['tag']))
loader.loadFeats({'ann': ann_feats})
# check model_info and params
assert model_opt['dataset'] == params['dataset']
assert model_opt['splitBy'] == params['splitBy']
# evaluate on the split,
# predictions = [{sent_id, sent, gd_ann_id, pred_ann_id, pred_score, sub_attn, loc_attn, weights}]
split = params['split']
model_opt['num_sents'] = params['num_sents']
model_opt['verbose'] = params['verbose']
crit = None
val_loss, acc, predictions, overall = eval_utils.eval_split(
loader, model, crit, split, model_opt)
print('Comprehension on %s\'s %s (%s sents) is %.2f%%' %
(params['dataset_splitBy'], params['split'], len(predictions), acc*100.))
print('attribute precision : %.2f%%' % (overall['precision']*100.0))
print('attribute recall : %.2f%%' % (overall['recall']*100.0))
print('attribute f1 : %.2f%%' % (overall['f1']*100.0))
# save
out_dir = osp.join('cache', 'results', params['dataset_splitBy'], 'easy')
if not osp.isdir(out_dir):
os.makedirs(out_dir)
out_file = osp.join(out_dir, params['id']+'_'+params['split']+'.json')
with open(out_file, 'w') as of:
json.dump({'predictions': predictions,
'acc': acc, 'overall': overall}, of)
# write to results.txt
f = open('experiments/easy_results.txt', 'a')
f.write('[%s][%s], id[%s]\'s acc is %.2f%%\n' %
(params['dataset_splitBy'], params['split'], params['id'], acc*100.0))
def main(args):
opt = vars(args)
tb_logger.configure('tb_logs/'+opt['id'], flush_secs=2)
# initialize
opt['dataset_splitBy'] = opt['dataset'] + '_' + opt['splitBy']
if opt['dataset'] == 'refcocog':
opt['unk_token'] = 3346
elif opt['dataset'] == 'refcoco':
opt['unk_token'] = 1996
elif opt['dataset'] == 'refcoco+':
opt['unk_token'] = 2629
checkpoint_dir = osp.join(opt['checkpoint_path'], opt['dataset_splitBy'])
if not osp.isdir(checkpoint_dir): os.makedirs(checkpoint_dir)
# set random seed
torch.manual_seed(opt['seed'])
random.seed(opt['seed'])
# set up loader
data_json = osp.join('cache/prepro', opt['dataset_splitBy'], 'data.json')
data_h5 = osp.join('cache/prepro', opt['dataset_splitBy'], 'data.h5')
loader = GtMRCNLoader(data_h5=data_h5, data_json=data_json)
# prepare feats
feats_dir = '%s_%s_%s' % (args.net_name, args.imdb_name, args.tag)
head_feats_dir=osp.join('cache/feats/', opt['dataset_splitBy'], 'mrcn', feats_dir)
loader.prepare_mrcn(head_feats_dir, args)
ann_feats = osp.join('cache/feats', opt['dataset_splitBy'], 'mrcn',
'%s_%s_%s_ann_feats.h5' % (opt['net_name'], opt['imdb_name'], opt['tag']))
loader.loadFeats({'ann': ann_feats})
# set up model
opt['vocab_size']= loader.vocab_size
opt['fc7_dim'] = loader.fc7_dim
opt['pool5_dim'] = loader.pool5_dim
opt['num_atts'] = loader.num_atts
model = JointMatching(opt)
# resume from previous checkpoint
infos = {}
if opt['start_from'] is not None:
checkpoint = torch.load(os.path.join('output',opt['dataset_splitBy'],opt['start_from']+'.pth'))
model.load_state_dict(checkpoint['model'].state_dict())
infos = json.load(open(os.path.join('output',opt['dataset_splitBy'],opt['start_from']+'.json') ,'r'))
print('start from model %s, best val score %.2f%%\n' % (opt['start_from'], infos['best_val_score']*100))
if opt['resume']:
iter = infos.get('iter', 0)
epoch = infos.get('epoch', 0)
val_accuracies = infos.get('val_accuracies', [])
val_loss_history = infos.get('val_loss_history', {})
val_result_history = infos.get('val_result_history', {})
loss_history = infos.get('loss_history', {})
loader.iterators = infos.get('iterators', loader.iterators)
best_val_score = infos.get('best_val_score', None)
else:
iter = 0
epoch = 0
val_accuracies = []
val_loss_history = {}
val_result_history = {}
loss_history = {}
best_val_score = None
# set up criterion
if opt['erase_lang_weight'] > 0 or opt['erase_allvisual_weight'] > 0:
if opt['erase_allvisual_weight'] > 0:
mm_crit = MaxMarginEraseCriterion(opt['visual_rank_weight'], opt['lang_rank_weight'],
opt['erase_lang_weight'], opt['erase_allvisual_weight'], opt['margin'], opt['erase_margin'])
elif opt['erase_lang_weight'] > 0:
mm_crit = MaxMarginEraseCriterion(opt['visual_rank_weight'], opt['lang_rank_weight'],
opt['erase_lang_weight'], opt['erase_allvisual_weight'], opt['margin'], opt['erase_margin'])
else:
mm_crit = MaxMarginCriterion(opt['visual_rank_weight'], opt['lang_rank_weight'], opt['margin'])
att_crit = nn.BCEWithLogitsLoss(loader.get_attribute_weights())
# move to GPU
if opt['gpuid'] >= 0:
model.cuda()
mm_crit.cuda()
att_crit.cuda()
# set up optimizer
optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, model.parameters()),
lr=opt['learning_rate'],
betas=(opt['optim_alpha'], opt['optim_beta']),
eps=opt['optim_epsilon'])
# start training
data_time, model_time = 0, 0
lr = opt['learning_rate']
best_predictions, best_overall = None, None
if opt['shuffle']:
loader.shuffle('train')
while True:
# run one iteration
loss, loss1, loss2, T, wrapped = lossFun(loader, optimizer, model, mm_crit, att_crit, opt, iter)
data_time += T['data']
model_time += T['model']
# write the training loss summary
if iter % opt['losses_log_every'] == 0:
loss_history[iter] = loss
# print stats
log_toc = time.time()
print('iter[%s](epoch[%s]), train_loss=%.3f, lr=%.2E, data:%.2fs/iter, model:%.2fs/iter' \
% (iter, epoch, loss, lr, data_time/opt['losses_log_every'], model_time/opt['losses_log_every']))
# write tensorboard logger
tb_logger.log_value('epoch', epoch, step=iter)
tb_logger.log_value('iter', iter, step=iter)
tb_logger.log_value('training_loss', loss, step=iter)
tb_logger.log_value('training_loss1', loss1, step=iter)
tb_logger.log_value('training_loss2', loss2, step=iter)
tb_logger.log_value('learning_rate', lr, step=iter)
data_time, model_time = 0, 0
# decay the learning rates
if opt['learning_rate_decay_start'] > 0 and epoch > opt['learning_rate_decay_start']:
frac = (epoch - opt['learning_rate_decay_start']) / opt['learning_rate_decay_every']
decay_factor = 0.1 ** frac
lr = opt['learning_rate'] * decay_factor
# update optimizer's learning rate
model_utils.set_lr(optimizer, lr)
# update iter and epoch
iter += 1
#wrapped = True # for debugging validation phase
if wrapped:
if opt['shuffle']:
loader.shuffle('train')
epoch += 1
# eval loss and save checkpoint
val_loss, acc, predictions, overall = eval_utils.eval_split(loader, model, None, 'val', opt)
val_loss_history[iter] = val_loss
val_result_history[iter] = {'loss': val_loss, 'accuracy': acc}
val_accuracies += [(iter, acc)]
print('val loss: %.2f' % val_loss)
print('val acc : %.2f%%\n' % (acc*100.0))
print('val precision : %.2f%%' % (overall['precision']*100.0))
print('val recall : %.2f%%' % (overall['recall']*100.0))
print('val f1 : %.2f%%' % (overall['f1']*100.0))
# write tensorboard logger
tb_logger.log_value('val_loss', val_loss, step=iter)
tb_logger.log_value('val_acc', acc, step=iter)
tb_logger.log_value('val precision', overall['precision']*100.0, step=iter)
tb_logger.log_value('val recall', overall['recall']*100.0, step=iter)
tb_logger.log_value('val f1', overall['f1']*100.0, step=iter)
# save model if best
current_score = acc
if best_val_score is None or current_score > best_val_score:
best_val_score = current_score
best_predictions = predictions
best_overall = overall
checkpoint_path = osp.join(checkpoint_dir, opt['id'] + '.pth')
checkpoint = {}
checkpoint['model'] = model
checkpoint['opt'] = opt
torch.save(checkpoint, checkpoint_path)
print('model saved to %s' % checkpoint_path)
# write json report
infos['iter'] = iter
infos['epoch'] = epoch
infos['iterators'] = loader.iterators
infos['loss_history'] = loss_history
infos['val_accuracies'] = val_accuracies
infos['val_loss_history'] = val_loss_history
infos['best_val_score'] = best_val_score
infos['best_predictions'] = predictions if best_predictions is None else best_predictions
infos['best_overall'] = overall if best_overall is None else best_overall
infos['opt'] = opt
infos['val_result_history'] = val_result_history
infos['word_to_ix'] = loader.word_to_ix
infos['att_to_ix'] = loader.att_to_ix
with open(osp.join(checkpoint_dir, opt['id']+'.json'), 'wb') as io:
json.dump(infos, io)
if epoch >= opt['max_epochs'] and opt['max_epochs'] > 0:
break
def main(args):
opt = vars(args)
# initialize
opt['dataset_splitBy'] = opt['dataset'] + '_' + opt['splitBy']
checkpoint_dir = osp.join(opt['checkpoint_path'], opt['dataset_splitBy'])
if not osp.isdir(checkpoint_dir): os.makedirs(checkpoint_dir)
# set random seed
torch.manual_seed(opt['seed'])
random.seed(opt['seed'])
# set up loader
data_json = osp.join('cache/prepro', opt['dataset_splitBy'], 'data.json')
data_h5 = osp.join('cache/prepro', opt['dataset_splitBy'], 'data.h5')
loader = GtMRCNLoader(data_h5=data_h5, data_json=data_json)
# prepare feats
feats_dir = '%s_%s_%s' % (args.net_name, args.imdb_name, args.tag)
head_feats_dir = osp.join('cache/feats/', opt['dataset_splitBy'], 'mrcn',
feats_dir)
loader.prepare_mrcn(head_feats_dir, args)
ann_feats = osp.join(
'cache/feats', opt['dataset_splitBy'], 'mrcn',
'%s_%s_%s_ann_feats.h5' %
(opt['net_name'], opt['imdb_name'], opt['tag']))
loader.loadFeats({'ann': ann_feats})
# set up model
opt['vocab_size'] = loader.vocab_size
opt['fc7_dim'] = loader.fc7_dim
opt['pool5_dim'] = loader.pool5_dim
opt['num_atts'] = loader.num_atts
model = JointMatching(opt)
# resume from previous checkpoint
infos = {}
if opt['start_from'] is not None:
pass
iter = infos.get('iter', 0)
epoch = infos.get('epoch', 0)
val_accuracies = infos.get('val_accuracies', [])
val_loss_history = infos.get('val_loss_history', {})
val_result_history = infos.get('val_result_history', {})
loss_history = infos.get('loss_history', {})
loader.iterators = infos.get('iterators', loader.iterators)
if opt['load_best_score'] == 1:
best_val_score = infos.get('best_val_score', None)
# set up criterion
mm_crit = MaxMarginCriterion(opt['visual_rank_weight'],
opt['lang_rank_weight'], opt['margin'])
att_crit = nn.BCEWithLogitsLoss(loader.get_attribute_weights())
# move to GPU
if opt['gpuid'] >= 0:
model.cuda()
mm_crit.cuda()
att_crit.cuda()
# set up optimizer
optimizer = torch.optim.Adam(model.parameters(),
lr=opt['learning_rate'],
betas=(opt['optim_alpha'], opt['optim_beta']),
eps=opt['optim_epsilon'])
# start training
data_time, model_time = 0, 0
lr = opt['learning_rate']
best_predictions, best_overall = None, None
while True:
# run one iteration
loss, T, wrapped = lossFun(loader, optimizer, model, mm_crit, att_crit,
opt, iter)
data_time += T['data']
model_time += T['model']
# write the training loss summary
if iter % opt['losses_log_every'] == 0:
loss_history[iter] = loss
# print stats
log_toc = time.time()
print('iter[%s](epoch[%s]), train_loss=%.3f, lr=%.2E, data:%.2fs/iter, model:%.2fs/iter' \
% (iter, epoch, loss, lr, data_time/opt['losses_log_every'], model_time/opt['losses_log_every']))
data_time, model_time = 0, 0
# decay the learning rates
if opt['learning_rate_decay_start'] > 0 and iter > opt[
'learning_rate_decay_start']:
frac = (iter - opt['learning_rate_decay_start']
) / opt['learning_rate_decay_every']
decay_factor = 0.1**frac
lr = opt['learning_rate'] * decay_factor
# update optimizer's learning rate
model_utils.set_lr(optimizer, lr)
# eval loss and save checkpoint
if iter % opt['save_checkpoint_every'] == 0 or iter == opt['max_iters']:
val_loss, acc, predictions, overall = eval_utils.eval_split(
loader, model, None, 'val', opt)
val_loss_history[iter] = val_loss
val_result_history[iter] = {'loss': val_loss, 'accuracy': acc}
val_accuracies += [(iter, acc)]
print('validation loss: %.2f' % val_loss)
print('validation acc : %.2f%%\n' % (acc * 100.0))
print('validation precision : %.2f%%' %
(overall['precision'] * 100.0))
print('validation recall : %.2f%%' %
(overall['recall'] * 100.0))
print('validation f1 : %.2f%%' % (overall['f1'] * 100.0))
# save model if best
current_score = acc
if best_val_score is None or current_score > best_val_score:
best_val_score = current_score
best_predictions = predictions
best_overall = overall
checkpoint_path = osp.join(checkpoint_dir, opt['id'] + '.pth')
checkpoint = {}
checkpoint['model'] = model
checkpoint['opt'] = opt
torch.save(checkpoint, checkpoint_path)
print('model saved to %s' % checkpoint_path)
# write json report
infos['iter'] = iter
infos['epoch'] = epoch
infos['iterators'] = loader.iterators
infos['loss_history'] = loss_history
infos['val_accuracies'] = val_accuracies
infos['val_loss_history'] = val_loss_history
infos['best_val_score'] = best_val_score
infos[
'best_predictions'] = predictions if best_predictions is None else best_predictions
infos[
'best_overall'] = overall if best_overall is None else best_overall
infos['opt'] = opt
infos['val_result_history'] = val_result_history
infos['word_to_ix'] = loader.word_to_ix
infos['att_to_ix'] = loader.att_to_ix
#with open(osp.join(checkpoint_dir, opt['id']+'.json'), 'wb') as io:
with open(osp.join(checkpoint_dir, opt['id'] + '.json'),
'w') as io:
json.dump(infos, io)
# update iter and epoch
iter += 1
if wrapped:
epoch += 1
if iter >= opt['max_iters'] and opt['max_iters'] > 0:
break