def main(**kwargs):
exp_base_dir = coco_paths['exp_dir']
if kwargs['dataset'] == 'flickr':
exp_base_dir = flickr_paths['exp_dir']
exp_const = ExpConstants(kwargs['exp_name'], exp_base_dir)
exp_const.model_dir = os.path.join(exp_const.exp_dir, 'models')
exp_const.seed = 0
exp_const.contextualize = not kwargs['no_context']
exp_const.random_lang = kwargs['random_lang']
data_const = FlickrDatasetConstants(kwargs['subset'])
model_const = Constants()
model_const.model_num = kwargs['model_num']
model_const.object_encoder = ObjectEncoderConstants()
model_const.object_encoder.context_layer.output_attentions = True
model_const.object_encoder.object_feature_dim = 2048
model_const.cap_encoder = CapEncoderConstants()
model_const.cap_encoder.output_attentions = True
model_const.cap_info_nce_layers = kwargs['cap_info_nce_layers']
if model_const.model_num == -100:
filename = os.path.join(exp_const.exp_dir, f'results_val_best.json')
results = io.load_json_object(filename)
model_const.model_num = results['model_num']
print('Selected model num:', model_const.model_num)
model_const.object_encoder_path = os.path.join(
exp_const.model_dir, f'object_encoder_{model_const.model_num}')
model_const.lang_sup_criterion_path = os.path.join(
exp_const.model_dir, f'lang_sup_criterion_{model_const.model_num}')
if exp_const.random_lang is True:
model_const.cap_encoder_path = os.path.join(
exp_const.model_dir, f'cap_encoder_{model_const.model_num}')
eval_flickr_phrase_loc.main(exp_const, data_const, model_const)
def exp_train():
exp_name = 'EXP_NAME'
out_base_dir = os.path.join(os.getcwd(), 'symlinks/exp/EXP_GROUP')
exp_const = ExpConstants(exp_name, out_base_dir)
exp_const.model_dir = os.path.join(exp_const.exp_dir, 'models')
exp_const.log_dir = os.path.join(exp_const.exp_dir, 'log')
exp_const.vis_dir = os.path.join(exp_const.exp_dir, 'vis')
exp_const.log_step = 10
exp_const.model_save_step = 1000
exp_const.val_step = 1000
exp_const.num_val_samples = 1000
exp_const.batch_size = 32
exp_const.num_epochs = 1000
exp_const.lr = 0.01
exp_const.momentum = 0.9
exp_const.num_workers = 5
exp_const.optimizer = 'SGD'
exp_const.subset = {'training': 'train', 'validation': 'val'}
data_const = DATASET_CONSTANTS()
model_const = Constants()
model_const.model_num = None
model_const.net = NET_CONSTANTS()
model_const.net_path = os.path.join(exp_const.model_dir,
f'net_{model_const.model_num}')
train.main(exp_const, data_const, model_const)
def main(**kwargs):
exp_base_dir = coco_paths['exp_dir']
if kwargs['dataset'] == 'flickr':
exp_base_dir = flickr_paths['exp_dir']
exp_const = ExpConstants(kwargs['exp_name'], exp_base_dir)
exp_const.log_dir = os.path.join(exp_const.exp_dir, 'logs')
exp_const.model_dir = os.path.join(exp_const.exp_dir, 'models')
exp_const.vis_dir = os.path.join(exp_const.exp_dir, 'vis')
exp_const.dataset = kwargs['dataset']
exp_const.optimizer = 'Adam'
exp_const.lr = kwargs['lr']
exp_const.momentum = None
exp_const.num_epochs = 10
exp_const.log_step = 20
# Save models approx. twice every epoch
exp_const.model_save_step = 400000 // (2 * kwargs['train_batch_size']
) # 4000=400000/(2*50)
if exp_const.dataset == 'flickr':
exp_const.model_save_step = 150000 // (2 * kwargs['train_batch_size'])
val_freq_factor = 2
if kwargs['val_frequently'] is True:
val_freq_factor = 1
exp_const.val_step = val_freq_factor * exp_const.model_save_step # set to 1*model_save_step for plotting mi vs perf
exp_const.num_val_samples = None
exp_const.train_batch_size = kwargs['train_batch_size']
exp_const.val_batch_size = 20
exp_const.num_workers = 10
exp_const.seed = 0
exp_const.neg_noun_loss_wt = kwargs['neg_noun_loss_wt']
exp_const.self_sup_loss_wt = kwargs['self_sup_loss_wt']
exp_const.lang_sup_loss_wt = kwargs['lang_sup_loss_wt']
exp_const.contextualize = not kwargs['no_context']
exp_const.random_lang = kwargs['random_lang']
DatasetConstants = CocoDatasetConstants
if exp_const.dataset == 'flickr':
DatasetConstants = FlickrDatasetConstants
data_const = {
'train': DatasetConstants('train'),
'val': DatasetConstants('val'),
}
model_const = Constants()
model_const.model_num = kwargs['model_num']
model_const.object_encoder = ObjectEncoderConstants()
model_const.object_encoder.context_layer.output_attentions = True
model_const.object_encoder.object_feature_dim = 2048
model_const.cap_encoder = CapEncoderConstants()
model_const.cap_encoder.output_attentions = True
model_const.cap_info_nce_layers = kwargs['cap_info_nce_layers']
model_const.object_encoder_path = os.path.join(
exp_const.model_dir, f'object_encoder_{model_const.model_num}')
model_const.self_sup_criterion_path = os.path.join(
exp_const.model_dir, f'self_sup_criterion_{model_const.model_num}')
model_const.lang_sup_criterion_path = os.path.join(
exp_const.model_dir, f'lang_sup_criterion_{model_const.model_num}')
train(exp_const, data_const, model_const)
def _exp_top_boxes_per_hoi(out_base_dir, data_const):
args = parser.parse_args()
not_specified_args = manage_required_args(
args,
parser,
required_args=['model_num'],
optional_args=[
'verb_given_appearance',
'verb_given_human_appearance',
'verb_given_object_appearance',
'verb_given_boxes_and_object_label',
'verb_given_human_pose',
'rcnn_det_prob'])
exp_name = 'factors'
if args.rcnn_det_prob:
exp_name += '_rcnn_det_prob'
if args.verb_given_appearance:
exp_name += '_appearance'
if args.verb_given_human_appearance:
exp_name += '_human_appearance'
if args.verb_given_object_appearance:
exp_name += '_object_appearance'
if args.verb_given_boxes_and_object_label:
exp_name += '_boxes_and_object_label'
if args.verb_given_human_pose:
exp_name += '_human_pose'
exp_const = ExpConstants(
exp_name=exp_name,
out_base_dir=out_base_dir)
exp_const.model_dir = os.path.join(exp_const.exp_dir,'models')
exp_const.num_to_vis = 10
data_const.pred_hoi_dets_h5py = os.path.join(
exp_const.exp_dir,
f'pred_hoi_dets_test_{args.model_num}.hdf5')
hoi_cand_dir = os.path.join(
os.getcwd(),
'data_symlinks/hico_exp/hoi_candidates')
data_const.human_pose_feats_hdf5 = os.path.join(
hoi_cand_dir,
'human_pose_feats_test.hdf5')
data_const.num_pose_keypoints = 18
model_const = Constants()
model_const.model_num = args.model_num
model_const.hoi_classifier = HoiClassifierConstants()
model_const.hoi_classifier.verb_given_appearance = args.verb_given_appearance
model_const.hoi_classifier.verb_given_boxes_and_object_label = args.verb_given_boxes_and_object_label
model_const.hoi_classifier.verb_given_human_pose = args.verb_given_human_pose
model_const.hoi_classifier.rcnn_det_prob = args.rcnn_det_prob
model_const.hoi_classifier.model_pth = os.path.join(
exp_const.model_dir,
f'hoi_classifier_{model_const.model_num}')
vis_top_boxes_per_hoi.main(exp_const, data_const, model_const)
def _exp_eval(out_base_dir, data_const):
args = parser.parse_args()
not_specified_args = manage_required_args(
args,
parser,
required_args=['model_num'],
optional_args=[
'verb_given_appearance',
'verb_given_human_appearance',
'verb_given_object_appearance',
'verb_given_boxes_and_object_label',
'verb_given_human_pose',
'rcnn_det_prob'])
exp_name = 'factors'
if args.rcnn_det_prob:
exp_name += '_rcnn_det_prob'
if args.verb_given_appearance:
exp_name += '_appearance'
if args.verb_given_human_appearance:
exp_name += '_human_appearance'
if args.verb_given_object_appearance:
exp_name += '_object_appearance'
if args.verb_given_boxes_and_object_label:
exp_name += '_boxes_and_object_label'
if args.verb_given_human_pose:
exp_name += '_human_pose'
exp_const = ExpConstants(
exp_name=exp_name,
out_base_dir=out_base_dir)
exp_const.model_dir = os.path.join(exp_const.exp_dir, 'models')
data_const.balanced_sampling = False
model_const = Constants()
model_const.model_num = args.model_num
model_const.hoi_classifier = HoiClassifierConstants()
model_const.hoi_classifier.verb_given_appearance = args.verb_given_appearance
model_const.hoi_classifier.verb_given_human_appearance = args.verb_given_human_appearance
model_const.hoi_classifier.verb_given_object_appearance = args.verb_given_object_appearance
model_const.hoi_classifier.verb_given_boxes_and_object_label = args.verb_given_boxes_and_object_label
model_const.hoi_classifier.verb_given_human_pose = args.verb_given_human_pose
model_const.hoi_classifier.rcnn_det_prob = args.rcnn_det_prob
model_const.hoi_classifier.model_pth = os.path.join(
exp_const.model_dir,
f'hoi_classifier_{model_const.model_num}')
if isinstance(data_const, FeatureConstantsVcoco):
data_sign = 'vcoco'
else:
data_sign = 'hico'
evaluate.main(exp_const, data_const, model_const, data_sign)
def exp_extract_embeddings():
args = parser.parse_args()
not_specified_args = manage_required_args(
args,
parser,
required_args=[
'embed_dim',
'xform',
'model_num',
'syn'])
exp_name = f'{args.xform}_{args.embed_dim}'
out_base_dir = os.path.join(
os.getcwd(),
'symlinks/exp/multi_sense_cooccur')
exp_const = ExpConstants(exp_name,out_base_dir)
exp_const.model_dir = os.path.join(exp_const.exp_dir,'models')
exp_const.cooccur_types = [
'syn',
'attr_attr',
'obj_attr',
'obj_hyp',
'context'
]
if args.syn==False:
exp_const.cooccur_types = exp_const.cooccur_types[1:]
data_const = MultiSenseCooccurDatasetConstants()
data_const.cooccur_csv = os.path.join(
os.getcwd(),
'symlinks/exp/multi_sense_cooccur/cooccurrences/merged_cooccur.csv')
model_const = Constants()
model_const.model_num = args.model_num
model_const.net = LogBilinearConstants()
model_const.net.num_words = 93553
model_const.net.embed_dims = args.embed_dim
model_const.net.two_embedding_layers = False
model_const.net.xform_type = args.xform
model_const.net.xform_num_layers = None
model_const.net.use_bias = True
model_const.net.use_fx = False
model_const.net.cooccur_types = copy.deepcopy(exp_const.cooccur_types)
model_const.net_path = os.path.join(
exp_const.model_dir,
f'net_{model_const.model_num}')
extract_embeddings.main(exp_const,data_const,model_const)
extract_embeddings_xformed.main(exp_const,data_const,model_const)
def main(**kwargs):
exp_base_dir = coco_paths['exp_dir']
if kwargs['train_dataset'] == 'flickr':
exp_base_dir = flickr_paths['exp_dir']
exp_const = ExpConstants(kwargs['exp_name'], exp_base_dir)
exp_const.log_dir = os.path.join(exp_const.exp_dir, 'logs')
exp_const.model_dir = os.path.join(exp_const.exp_dir, 'models')
exp_const.train_dataset = kwargs['train_dataset']
exp_const.vis_dataset = kwargs['vis_dataset']
exp_const.vis_dir = os.path.join(exp_const.exp_dir,
f'vis/attention_{exp_const.vis_dataset}')
exp_const.num_vis_samples = 50
exp_const.seed = 0
exp_const.contextualize = not kwargs['no_context']
DatasetConstants = CocoDatasetConstants
if exp_const.vis_dataset == 'flickr':
DatasetConstants = FlickrDatasetConstants
data_const = DatasetConstants('val')
if exp_const.vis_dataset == 'coco':
data_const.image_dir = os.path.join(coco_paths['image_dir'],
data_const.subset_image_dirname)
data_const.read_neg_samples = False
data_const.read_noun_adj_tokens = False
model_const = Constants()
model_const.model_num = kwargs['model_num']
model_const.object_encoder = ObjectEncoderConstants()
model_const.object_encoder.object_feature_dim = 2048
model_const.cap_encoder = CapEncoderConstants()
if model_const.model_num == -100:
model_const.object_encoder_path = os.path.join(exp_const.model_dir,
f'best_object_encoder')
model_const.lang_sup_criterion_path = os.path.join(
exp_const.model_dir, f'best_lang_sup_criterion')
else:
model_const.object_encoder_path = os.path.join(
exp_const.model_dir, f'object_encoder_{model_const.model_num}')
model_const.lang_sup_criterion_path = os.path.join(
exp_const.model_dir, f'lang_sup_criterion_{model_const.model_num}')
if exp_const.vis_dataset == 'coco':
vis_att(exp_const, data_const, model_const)
else:
vis_att_flickr(exp_const, data_const, model_const)
def exp_eval():
exp_name = 'EXP_NAME'
out_base_dir = os.path.join(os.getcwd(), 'symlinks/exp/EXP_GROUP')
exp_const = ExpConstants(exp_name, out_base_dir)
exp_const.model_dir = os.path.join(exp_const.exp_dir, 'models')
exp_const.log_dir = os.path.join(exp_const.exp_dir, 'log')
exp_const.vis_dir = os.path.join(exp_const.exp_dir, 'vis')
exp_const.batch_size = 32
exp_const.num_workers = 5
data_const = DATASET_CONSTANTS()
data_const.subset = 'eval'
model_const = Constants()
model_const.model_num = None
model_const.net = NET_CONSTANTS()
model_const.net_path = os.path.join(exp_const.model_dir,
f'net_{model_const.model_num}')
evaluation.main(exp_const, data_const, model_const)
def main(**kwargs):
exp_const = ExpConstants(kwargs['exp_name'], kwargs['exp_base_dir'])
exp_const.log_dir = os.path.join(exp_const.exp_dir, 'logs')
exp_const.model_dir = os.path.join(exp_const.exp_dir, 'models')
exp_const.vis_dir = os.path.join(exp_const.exp_dir, 'vis')
exp_const.optimizer = 'Adam'
exp_const.lr = 1e-3
exp_const.momentum = None
exp_const.num_epochs = 100
exp_const.log_step = 100
exp_const.model_save_step = 1000
exp_const.val_step = 1000
exp_const.num_val_samples = None
data_const = {'train': Constants(), 'val': Constants()}
model_const = Constants()
model_const.model_num = kwargs['model_num']
model_const.net = Constants()
model_const.net_path = os.path.join(exp_const.model_dir,
f'net_{model_const.model_num}')
def exp_save_ae_combined_glove_and_visual_features():
exp_name = 'ae_glove_and_visual'
out_base_dir = os.path.join(
os.getcwd(),
'symlinks/exp/google_images/' + \
'normalized_resnet_embeddings_recon_loss_trained_on_google')
exp_const = ExpConstants(exp_name, out_base_dir)
exp_const.model_dir = os.path.join(exp_const.exp_dir, 'models')
exp_const.batch_size = 1000
concat_embeddings_dir = os.path.join(
os.getcwd(),
'symlinks/exp/google_images/' + \
'normalized_resnet_embeddings_recon_loss_trained_on_google/' + \
'concat_glove_and_visual')
data_const = ConcatEmbedDatasetConstants(concat_embeddings_dir)
model_const = Constants()
model_const.model_num = 400
model_const.encoder = EncoderConstants()
model_const.decoder = DecoderConstants()
save_ae_embeddings.main(exp_const, data_const, model_const)
def exp_eval():
args = parser.parse_args()
not_specified_args = manage_required_args(
args,
parser,
required_args=['fappend', 'model_num'],
optional_args=[
'verb_given_appearance', 'verb_given_human_appearance',
'verb_given_object_appearance',
'verb_given_boxes_and_object_label', 'verb_given_human_pose',
'rcnn_det_prob'
])
exp_name = 'factors'
exp_name += '_' + args.fappend
out_base_dir = os.path.join(os.getcwd(),
'data_symlinks/hico_exp/hoi_classifier')
exp_const = ExpConstants(exp_name=exp_name, out_base_dir=out_base_dir)
exp_const.model_dir = os.path.join(exp_const.exp_dir, 'models')
data_const = FeatureConstants(subset='test')
data_const.balanced_sampling = False
model_const = Constants()
model_const.model_num = args.model_num
model_const.hoi_classifier = HoiClassifierConstants()
model_const.hoi_classifier.verb_given_appearance = args.verb_given_appearance
model_const.hoi_classifier.verb_given_human_appearance = args.verb_given_human_appearance
model_const.hoi_classifier.verb_given_object_appearance = args.verb_given_object_appearance
model_const.hoi_classifier.verb_given_boxes_and_object_label = args.verb_given_boxes_and_object_label
model_const.hoi_classifier.verb_given_human_pose = args.verb_given_human_pose
model_const.hoi_classifier.rcnn_det_prob = args.rcnn_det_prob
model_const.hoi_classifier.model_pth = os.path.join(
exp_const.model_dir, f'hoi_classifier_{model_const.model_num}')
evaluate.main(exp_const, data_const, model_const)
def exp_save_ae_visual_features():
exp_name = 'ae_visual_features'
out_base_dir = os.path.join(
os.getcwd(),
'symlinks/exp/google_images/' + \
'normalized_resnet_features_recon_loss_trained_on_google')
exp_const = ExpConstants(exp_name, out_base_dir)
exp_const.model_dir = os.path.join(exp_const.exp_dir, 'models')
exp_const.batch_size = 1000
features_dir = os.path.join(
os.getcwd(),
'symlinks/exp/google_images/' + \
'normalized_resnet_features_recon_loss_trained_on_google')
data_const = VisualFeaturesDatasetConstants(features_dir)
model_const = Constants()
model_const.model_num = 990
model_const.encoder = EncoderConstants()
model_const.encoder.output_dims = 300
model_const.decoder = DecoderConstants()
model_const.decoder.input_dims = 300
save_ae_visual_features.main(exp_const, data_const, model_const)
def main(**kwargs):
exp_base_dir = coco_paths['exp_dir']
if kwargs['dataset'] == 'flickr':
exp_base_dir = flickr_paths['exp_dir']
exp_const = ExpConstants(kwargs['exp_name'], exp_base_dir)
exp_const.model_dir = os.path.join(exp_const.exp_dir, 'models')
exp_const.seed = 0
exp_const.contextualize = not kwargs['no_context']
exp_const.random_lang = kwargs['random_lang']
data_const = FlickrDatasetConstants(kwargs['subset'])
model_const = Constants()
model_const.object_encoder = ObjectEncoderConstants()
model_const.object_encoder.context_layer.output_attentions = True
model_const.object_encoder.object_feature_dim = 2048
model_const.cap_encoder = CapEncoderConstants()
model_const.cap_encoder.output_attentions = True
model_const.cap_info_nce_layers = kwargs['cap_info_nce_layers']
model_nums = find_all_model_numbers(exp_const.model_dir)
for num in model_nums:
continue
if num <= 3000:
continue
model_const.model_num = num
model_const.object_encoder_path = os.path.join(
exp_const.model_dir, f'object_encoder_{model_const.model_num}')
model_const.lang_sup_criterion_path = os.path.join(
exp_const.model_dir, f'lang_sup_criterion_{model_const.model_num}')
if exp_const.random_lang is True:
model_const.cap_encoder_path = os.path.join(
exp_const.model_dir, f'cap_encoder_{model_const.model_num}')
filename = os.path.join(exp_const.exp_dir,
f'results_{data_const.subset}_{num}.json')
if os.path.exists(filename):
print(io.load_json_object(filename))
continue
eval_flickr_phrase_loc.main(exp_const, data_const, model_const)
best_model_num = -1
best_pt_recall = 0
best_results = None
for num in model_nums:
filename = os.path.join(exp_const.exp_dir,
f'results_{data_const.subset}_{num}.json')
if not os.path.exists(filename):
continue
results = io.load_json_object(filename)
results['model_num'] = num
print(results)
if results['pt_recall'] >= best_pt_recall:
best_results = results
best_pt_recall = results['pt_recall']
best_model_num = num
print('-' * 80)
best_results['model_num'] = best_model_num
print(best_results)
filename = os.path.join(exp_const.exp_dir,
f'results_{data_const.subset}_best.json')
io.dump_json_object(best_results, filename)
def exp_train():
args = parser.parse_args()
# create experiments directory and required folders
out_base_dir = os.path.join(os.getcwd(), f'exp/{args.dataset_type}')
exp_const = ExpConstants(args.run_name, out_base_dir)
exp_const.model_dir = os.path.join(exp_const.exp_dir, 'models')
exp_const.log_dir = os.path.join(exp_const.exp_dir, 'log')
exp_const.vis_dir = os.path.join(exp_const.exp_dir, 'vis')
use_cuda = torch.cuda.is_available()
exp_const.device = "cuda:0" if use_cuda else "cpu"
# tranining params
exp_const.optimizer = args.optimizer
exp_const.num_epochs = args.num_epochs
exp_const.batch_size = args.batch_size
exp_const.lr = args.lr
exp_const.momentum = args.momentum
exp_const.num_workers = args.num_workers
# logging, saving
exp_const.log_step = args.log_step
exp_const.model_save_epoch = args.model_save_epoch
exp_const.val_epoch = args.val_epoch
exp_const.subset = {'training': 'train', 'test': 'test'}
# dataset
data_const = DatasetConstants(root=args.dataroot,
download=args.download_dataset,
train=True)
data_const.dataset_type = args.dataset_type
# model (resnet and attribute embeddings)
model_const = Constants()
model_const.model_num = None
model_const.sim_loss = args.sim_loss
model_const.ce_loss_warmup = args.ce_loss_warmup
model_const.net = ResnetConstants()
if args.dataset_type == 'Cifar100':
model_const.net.num_layers = "cifar100" # a custom resnet for cifar100, to adjust the dimensions of the feature maps
model_const.net.num_classes = 100
else:
model_const.net.num_layers = args.num_layers
if args.dataset_type == "Imagenet":
model_const.net.num_classes = 1000
elif args.dataset_type == "VOC":
model_const.net.num_classes = 20
elif args.dataset_type == "STL10":
model_const.net.num_layers = 'cifar100' # TODO: deeper resnets does not work on STL10.
model_const.net.num_classes = 10
model_const.net.pretrained = False
model_const.net_path = os.path.join(exp_const.model_dir,
f'net_{model_const.model_num}')
model_const.attr_embed = AttributeEmbeddingsConstants()
model_const.attr_embed_path = os.path.join(
exp_const.model_dir, f'attr_embed_{model_const.model_num}')
model_const.attr_embed.glove_dim = 300
model_const.attr_embed.num_classes = model_const.net.num_classes
# attribute embedding dimensions
if args.embed_type == 'vico_linear':
model_const.attr_embed.no_glove = True # Zero out the glove component
model_const.attr_embed.embed_dims = 300 + args.vico_dim
embed_dir = os.path.join(
os.getcwd(),
'data/pretrained-embeddings/' + \
f'glove_300_vico_linear_100/')
model_const.attr_embed.embed_h5py = os.path.join(
embed_dir, 'visual_word_vecs.h5py')
model_const.attr_embed.embed_word_to_idx_json = os.path.join(
embed_dir, 'visual_word_vecs_idx.json')
elif args.embed_type == 'vico_select':
model_const.attr_embed.no_glove = True # Zero out the glove component
model_const.attr_embed.hypernym = args.hypernym
model_const.attr_embed.embed_dims = 300 + args.vico_dim
embed_dir = os.path.join(
os.getcwd(),
'data/pretrained-embeddings/' + \
f'glove_300_vico_select_200/')
model_const.attr_embed.embed_h5py = os.path.join(
embed_dir, 'visual_word_vecs.h5py')
model_const.attr_embed.embed_word_to_idx_json = os.path.join(
embed_dir, 'visual_word_vecs_idx.json')
else:
err_str = f'{args.embed_type} is currently not implemented in the runner'
assert (False), err_str
# pass all constants to training method
train.main(exp_const, data_const, model_const)
def exp_train():
args = parser.parse_args()
not_specified_args = manage_required_args(args,
parser,
required_args=[
'held_classes', 'embed_type',
'glove_dim', 'vico_dim',
'run'
],
optional_args=[])
exp_name = \
args.embed_type + '_' + \
str(args.glove_dim) + '_' + \
str(args.vico_dim) + '_' + \
'held_classes_' + str(args.held_classes)
out_base_dir = os.path.join(os.getcwd(),
f'symlinks/exp/cifar100/zero_shot_{args.run}')
exp_const = ExpConstants(exp_name, out_base_dir)
exp_const.model_dir = os.path.join(exp_const.exp_dir, 'models')
exp_const.log_dir = os.path.join(exp_const.exp_dir, 'log')
exp_const.vis_dir = os.path.join(exp_const.exp_dir, 'vis')
exp_const.log_step = 200
exp_const.model_save_step = 1000
exp_const.val_step = 1000
exp_const.batch_size = 128
exp_const.num_epochs = 50 #100
exp_const.lr = 0.01
exp_const.momentum = 0.9
exp_const.num_workers = 5
exp_const.optimizer = 'Adam'
exp_const.feedforward = False
exp_const.subset = {'training': 'train', 'test': 'test'}
data_const = Cifar100DatasetConstants()
data_const.num_held_out_classes = args.held_classes
model_const = Constants()
model_const.model_num = None
model_const.net = ResnetConstants()
model_const.net.num_layers = 32
model_const.net.num_classes = 100
model_const.net.pretrained = False
model_const.net_path = os.path.join(exp_const.model_dir,
f'net_{model_const.model_num}')
model_const.embed2class = Embed2ClassConstants()
model_const.embed2class.linear = True
model_const.embed2class_path = os.path.join(
exp_const.model_dir, f'embed2class_{model_const.model_num}')
model_const.embed2class.glove_dim = args.glove_dim
# Dimensions
if args.embed_type == 'glove':
model_const.embed2class.embed_dims = args.glove_dim
model_const.embed2class.embed_h5py = os.path.join(
os.getcwd(),
f'symlinks/data/glove/proc/glove_6B_{args.glove_dim}d.h5py')
model_const.embed2class.embed_word_to_idx_json = os.path.join(
os.getcwd(),
f'symlinks/data/glove/proc/glove_6B_{args.glove_dim}d_word_to_idx.json'
)
elif args.embed_type == 'glove_vico_linear':
model_const.embed2class.embed_dims = args.glove_dim + args.vico_dim
embed_dir = os.path.join(
os.getcwd(),
'symlinks/exp/multi_sense_cooccur/' + \
f'linear_100/concat_with_glove_{args.glove_dim}')
model_const.embed2class.embed_h5py = os.path.join(
embed_dir, 'visual_word_vecs.h5py')
model_const.embed2class.embed_word_to_idx_json = os.path.join(
embed_dir, 'visual_word_vecs_idx.json')
elif args.embed_type == 'vico_linear':
model_const.embed2class.no_glove = True # Zero out the glove component
model_const.embed2class.embed_dims = args.glove_dim + args.vico_dim
embed_dir = os.path.join(
os.getcwd(),
'symlinks/exp/multi_sense_cooccur/' + \
f'linear_100/concat_with_glove_{args.glove_dim}')
model_const.embed2class.embed_h5py = os.path.join(
embed_dir, 'visual_word_vecs.h5py')
model_const.embed2class.embed_word_to_idx_json = os.path.join(
embed_dir, 'visual_word_vecs_idx.json')
elif args.embed_type == 'glove_vico_select':
model_const.embed2class.embed_dims = args.glove_dim + args.vico_dim
embed_dir = os.path.join(
os.getcwd(),
'symlinks/exp/multi_sense_cooccur/' + \
f'select_200/concat_with_glove_{args.glove_dim}')
model_const.embed2class.embed_h5py = os.path.join(
embed_dir, 'visual_word_vecs.h5py')
model_const.embed2class.embed_word_to_idx_json = os.path.join(
embed_dir, 'visual_word_vecs_idx.json')
else:
err_str = f'{args.embed_type} is currently not implemented in the runner'
assert (False), err_str
train.main(exp_const, data_const, model_const)
def exp_train():
args = parser.parse_args()
not_specified_args = manage_required_args(
args,
parser,
required_args=[
'embed_dim',
'xform',
'model_num',
'syn'])
exp_name = f'{args.xform}_{args.embed_dim}'
out_base_dir = os.path.join(
os.getcwd(),
'symlinks/exp/multi_sense_cooccur')
exp_const = ExpConstants(exp_name,out_base_dir)
exp_const.model_dir = os.path.join(exp_const.exp_dir,'models')
exp_const.log_dir = os.path.join(exp_const.exp_dir,'log')
exp_const.vis_dir = os.path.join(exp_const.exp_dir,'vis')
exp_const.log_step = 100
exp_const.model_save_step = 10000
exp_const.batch_size = 1000
exp_const.num_epochs = 10
exp_const.lr = 0.01
exp_const.momentum = 0.9 # used only when optimizer is set to 'SGD'
exp_const.num_workers = 5
# First train with Adam then finetune with Adagrad
if args.model_num==-1:
exp_const.optimizer = 'Adam'
else:
exp_const.optimizer = 'Adagrad'
exp_const.weight_decay = 0
exp_const.cooccur_weights = {
'syn': 1,
'attr_attr': 1,
'obj_attr': 1,
'obj_hyp': 1,
'context': 1,
}
if args.syn==False:
del exp_const.cooccur_weights['syn']
exp_const.use_neg = True
data_const = MultiSenseCooccurDatasetConstants()
data_const.cooccur_csv = os.path.join(
os.getcwd(),
'symlinks/exp/multi_sense_cooccur/cooccurrences/merged_cooccur.csv')
data_const.use_self_count = True
model_const = Constants()
if args.model_num==-1:
model_const.model_num = None
else:
model_const.model_num = args.model_num
model_const.net = LogBilinearConstants()
model_const.net.num_words = 93553
model_const.net.embed_dims = args.embed_dim
model_const.net.two_embedding_layers = False
model_const.net.xform_type = args.xform
model_const.net.xform_num_layers = None
model_const.net.use_bias = True
model_const.net.use_fx = False
model_const.net.cooccur_types = [
'syn',
'attr_attr',
'obj_attr',
'obj_hyp',
'context'
]
if args.syn==False:
model_const.net.cooccur_types = model_const.net.cooccur_types[1:]
model_const.net_path = os.path.join(
exp_const.model_dir,
f'net_{model_const.model_num}')
train.main(exp_const,data_const,model_const)
