code
stringlengths
17
6.64M
def corr_speckle_noise_sev_2(image): return corruption_dict['speckle_noise'](image, 1)
def corr_speckle_noise_sev_3(image): return corruption_dict['speckle_noise'](image, 2)
def corr_speckle_noise_sev_4(image): return corruption_dict['speckle_noise'](image, 3)
def corr_speckle_noise_sev_5(image): return corruption_dict['speckle_noise'](image, 4)
def corr_zoom_blur_sev_1(image): return corruption_dict['zoom_blur'](image, 0)
def corr_zoom_blur_sev_2(image): return corruption_dict['zoom_blur'](image, 1)
def corr_zoom_blur_sev_3(image): return corruption_dict['zoom_blur'](image, 2)
def corr_zoom_blur_sev_4(image): return corruption_dict['zoom_blur'](image, 3)
def corr_zoom_blur_sev_5(image): return corruption_dict['zoom_blur'](image, 4)
def gen_corrupt_batch_gpu(corruption, severity): def corrupt_batch_gpu(images, model): for i in range(images.size(0)): corr_func = corruption_dict[corruption] images[i] = corr_func(images[i], severity, gpu=True) return images return corrupt_batch_gpu
def accuracy_topk_subselected(logits, targets): targets = torch.tensor([class_sublist.index(x) for x in targets]) return accuracy_topk(logits, targets)
def accuracy_topk_subselected(logits, targets): targets = torch.tensor(list(map((lambda x: class_sublist_1_8.index(x)), targets))) return accuracy_topk(logits, targets)
def crop_image(image, border=2): return PIL.ImageOps.crop(image, border=border)
def objectnet_accuracy(logits, targets, image_paths, using_class_sublist=True): if using_class_sublist: folder_map = {k: [class_sublist.index(x) for x in v] for (k, v) in folder_to_ids.items()} else: folder_map = folder_to_ids preds = logits.argmax(dim=1) (num_correct, num_total) = (0, 0) for (pred, image_path) in zip(preds, image_paths): folder = image_path.split('/')[0] if (folder in folder_map): num_total += 1 if (pred in folder_map[folder]): num_correct += 1 return {'top1': ((num_correct / num_total) * 100)}
def accuracy_topk_subselected(logits, targets): targets = torch.tensor([class_sublist.index(x) for x in targets]) return accuracy_topk(logits, targets)
def pgd_style_attack(d, images, model): eps = d['eps'] if (d['norm'] == 2): eps = ((eps * images.size(2)) / 224) return pgd(model, images, eps, d['step_size'], d['num_steps'], d['norm'], targeted=d['targeted'])
def gen_attack_fn(d): return (lambda images, model: pgd_style_attack(d, images, model))
def accuracy_topk_subselected(logits, targets): targets = torch.tensor([class_sublist_1_8.index(x) for x in targets]) return accuracy_topk(logits, targets)
def main(args): py_model = registry.get_model(args.model) py_eval_setting = registry.get_eval_setting(args.eval_setting) if (args.db and utils.evaluation_completed(py_model, py_eval_setting)): print(f'Evaluation for {py_model.name} x {py_eval_setting.name} already found. Skipping...') return args.num_gpus = torch.cuda.device_count() results_dict = mp.Manager().dict() mp.spawn(main_worker, nprocs=args.num_gpus, args=(args, results_dict)) (idx_sorted, idx_map) = torch.cat([results_dict[i]['idxs'] for i in range(args.num_gpus)]).sort() assert idx_sorted.eq(idx_sorted.unique()).all(), 'Error collecting results' assert idx_sorted.eq(torch.tensor(list(range(idx_sorted.size(0))))).all(), 'Error collecting results' logits = torch.cat([results_dict[i]['logits'] for i in range(args.num_gpus)])[idx_map] targets = torch.cat([results_dict[i]['targets'] for i in range(args.num_gpus)])[idx_map] image_paths = np.concatenate([results_dict[i]['image_paths'] for i in range(args.num_gpus)])[idx_map] metrics = py_eval_setting.get_metrics(logits, targets, image_paths, py_model) with open(join(args.logdir, 'metrics.json'), 'w') as outfile: json.dump(metrics, outfile) if args.db: utils.store_evaluation(py_model, py_eval_setting, metrics, logits) print('Uploaded to db') utils.close_db_connection() print('************************************') print(f'RESULT {args.model} on {args.eval_setting} - {metrics}') print('************************************')
def main_worker(gpu, args, results_dict): dist.init_process_group(backend=args.backend, init_method=args.dist_url, world_size=args.num_gpus, rank=gpu) torch.cuda.set_device(gpu) registry.load_full_registry() py_model = registry.get_model(args.model) py_eval_setting = registry.get_eval_setting(args.eval_setting) model = py_model.generate_classifier(py_eval_setting) model = model.cuda() model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[gpu]) batch_size = py_model.get_batch_size(py_eval_setting) gpu_perturbation_fn = py_eval_setting.get_perturbation_fn_gpu(py_model) torch.set_grad_enabled((py_eval_setting.adversarial_attack is not None)) setting_transform = ([py_eval_setting.transform] if (py_eval_setting.transform is not None) else []) val_dataset = CustomImageFolder(root=py_eval_setting.get_dataset_root(), transform=transforms.Compose((setting_transform + [py_model.transform])), perturbation_fn=py_eval_setting.get_perturbation_fn_cpu(py_model), idx_subsample_list=py_eval_setting.get_idx_subsample_list(py_model)) val_sampler = DistributedSampler(val_dataset, num_replicas=args.num_gpus, rank=gpu, shuffle=False) val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False, num_workers=args.workers, pin_memory=True, sampler=val_sampler) (logits, targets, image_paths, idxs) = validate(gpu, args, val_loader, model, gpu_perturbation_fn) results_dict[gpu] = {'logits': logits, 'targets': targets, 'image_paths': image_paths, 'idxs': idxs}
def validate(gpu, args, val_loader, model, gpu_perturbation_fn): model.eval() (all_logits, all_targets, all_idxs, all_image_paths) = ([], [], [], []) if (gpu == 0): val_loader = tqdm(val_loader, desc='Validating') for (idxs, image_paths, images, target) in val_loader: images = images.cuda() if (gpu_perturbation_fn is not None): images = gpu_perturbation_fn(images, model) output = model(images) all_logits.append(output.detach().cpu()) all_targets.append(target) all_idxs.append(idxs) all_image_paths.append(image_paths) all_logits = torch.cat(all_logits, dim=0) all_targets = torch.cat(all_targets, dim=0) all_idxs = torch.cat(all_idxs, dim=0) all_image_paths = [image_path for batch in all_image_paths for image_path in batch] return (all_logits, all_targets, all_image_paths, all_idxs)
def download_db(): if (not exists(join(s3_utils.default_cache_root_path, 'robustness_evaluation.db'))): print('downloading database dump...') subprocess.run(['wget', '-P', s3_utils.default_cache_root_path, DB_DUMP_URL, '--no-check-certificate'], check=True)
def gen_short_uuid(num_chars=None): num = uuid.uuid4().int alphabet = '23456789ABCDEFGHJKLMNPQRSTUVWXYZabcdefghijkmnopqrstuvwxyz' res = [] while (num > 0): (num, digit) = divmod(num, len(alphabet)) res.append(alphabet[digit]) res2 = ''.join(reversed(res)) if (num_chars is None): return res2 else: return res2[:num_chars]
def get_logdir_key(model_id): return 'logdir/{}'.format(model_id)
def get_checkpoint_data_key(checkpoint_id): return 'checkpoints/{}_data.bytes'.format(checkpoint_id)
def get_dataset_data_key(dataset_id): return 'datasets/{}_data.bytes'.format(dataset_id)
def get_evaluation_setting_extra_data_key(evaluation_setting_id): return 'evaluation_settings/{}_extra_data.bytes'.format(evaluation_setting_id)
def get_evaluation_setting_processed_dataset_key(evaluation_setting_id): return 'evaluation_settings/{}_processed_dataset.bytes'.format(evaluation_setting_id)
def get_raw_input_data_key(raw_input_id): return 'raw_inputs/{}_data.bytes'.format(raw_input_id)
def get_evaluation_extra_data_key(evaluation_id): return 'evaluations/{}_data.bytes'.format(evaluation_id)
def get_evaluation_logits_data_key(evaluation_id): return 'evaluations/{}_logits_data.bytes'.format(evaluation_id)
def get_evaluation_chunk_extra_data_key(evaluation_chunk_id): return 'evaluation_chunks/{}_data.bytes'.format(evaluation_chunk_id)
def get_evaluation_chunk_logits_data_key(evaluation_chunk_id): return 'evaluation_chunks/{}_logits_data.bytes'.format(evaluation_chunk_id)
def get_evaluation_chunk_indices_data_key(evaluation_chunk_id): return 'evaluation_chunks/{}_indices_data.bytes'.format(evaluation_chunk_id)
class Model(sqlalchemy_base): __tablename__ = 'models' uuid = sqla.Column(sqla.String, primary_key=True) name = sqla.Column(sqla.String, unique=True) description = sqla.Column(sqla.String) username = sqla.Column(sqla.String) creation_time = sqla.Column(sqla.DateTime(timezone=False), server_default=sqla.sql.func.now()) extra_info = sqla.Column(sqla.JSON) checkpoints = sqla.orm.relationship('Checkpoint', back_populates='model', cascade='all, delete, delete-orphan', foreign_keys='Checkpoint.model_uuid') final_checkpoint_uuid = sqla.Column(sqla.String, sqla.ForeignKey('checkpoints.uuid'), nullable=True) final_checkpoint = sqla.orm.relationship('Checkpoint', foreign_keys=[final_checkpoint_uuid], uselist=False) completed = sqla.Column(sqla.Boolean) hidden = sqla.Column(sqla.Boolean) logdir_filepaths = sqla.Column(sqla.JSON) def __repr__(self): return f'<Model(uuid="{self.uuid}", name="{self.name}")>' def __hash__(self): return hash((hash(self.uuid) + hash(self.name))) def __eq__(self, other): return (self.__hash__() == hash(other))
class Checkpoint(sqlalchemy_base): __tablename__ = 'checkpoints' uuid = sqla.Column(sqla.String, primary_key=True) name = sqla.Column(sqla.String, unique=True) creation_time = sqla.Column(sqla.DateTime(timezone=False), server_default=sqla.sql.func.now()) model_uuid = sqla.Column(sqla.String, sqla.ForeignKey('models.uuid'), nullable=False) model = sqla.orm.relationship('Model', back_populates='checkpoints', foreign_keys=[model_uuid]) evaluations = sqla.orm.relationship('Evaluation', back_populates='checkpoint', cascade='all, delete, delete-orphan', foreign_keys='Evaluation.checkpoint_uuid') training_step = sqla.Column(sqla.BigInteger) epoch = sqla.Column(sqla.Float) username = sqla.Column(sqla.String) extra_info = sqla.Column(sqla.JSON) hidden = sqla.Column(sqla.Boolean) def __repr__(self): return f'<Checkpoint(uuid="{self.uuid}", model_uuid="{self.model_uuid}")>' def __hash__(self): return hash((hash(self.uuid) + hash(self.name))) def __eq__(self, other): return (self.__hash__() == hash(other))
class Dataset(sqlalchemy_base): __tablename__ = 'datasets' uuid = sqla.Column(sqla.String, primary_key=True) name = sqla.Column(sqla.String, unique=True, nullable=False) description = sqla.Column(sqla.String) username = sqla.Column(sqla.String) creation_time = sqla.Column(sqla.DateTime(timezone=False), server_default=sqla.sql.func.now()) size = sqla.Column(sqla.Integer) extra_info = sqla.Column(sqla.JSON) evaluation_settings = sqla.orm.relationship('EvaluationSetting', back_populates='dataset', cascade='all, delete, delete-orphan', foreign_keys='EvaluationSetting.dataset_uuid') hidden = sqla.Column(sqla.Boolean) def __repr__(self): return f'<Dataset(uuid="{self.uuid}", name="{self.name}")>' def __hash__(self): return hash((hash(self.uuid) + hash(self.name))) def __eq__(self, other): return (self.__hash__() == hash(other))
class EvaluationSetting(sqlalchemy_base): __tablename__ = 'evaluation_settings' uuid = sqla.Column(sqla.String, primary_key=True) name = sqla.Column(sqla.String, unique=True, nullable=False) description = sqla.Column(sqla.String) username = sqla.Column(sqla.String) creation_time = sqla.Column(sqla.DateTime(timezone=False), server_default=sqla.sql.func.now()) dataset_uuid = sqla.Column(sqla.String, sqla.ForeignKey('datasets.uuid'), nullable=False) dataset = sqla.orm.relationship('Dataset', back_populates='evaluation_settings', foreign_keys=[dataset_uuid]) evaluations = sqla.orm.relationship('Evaluation', back_populates='setting', cascade='all, delete, delete-orphan', foreign_keys='Evaluation.setting_uuid') raw_inputs = sqla.orm.relationship('RawInput', back_populates='setting', cascade='all, delete, delete-orphan', foreign_keys='RawInput.setting_uuid') extra_info = sqla.Column(sqla.JSON) hidden = sqla.Column(sqla.Boolean) def __repr__(self): return f'<EvaluationSetting(uuid="{self.uuid}", name="{self.name}")>' def __hash__(self): return hash((hash(self.uuid) + hash(self.name))) def __eq__(self, other): return (self.__hash__() == hash(other))
class RawInput(sqlalchemy_base): __tablename__ = 'raw_inputs' uuid = sqla.Column(sqla.String, primary_key=True) name = sqla.Column(sqla.String, unique=True) description = sqla.Column(sqla.String) username = sqla.Column(sqla.String) creation_time = sqla.Column(sqla.DateTime(timezone=False), server_default=sqla.sql.func.now()) setting_uuid = sqla.Column(sqla.String, sqla.ForeignKey('evaluation_settings.uuid'), nullable=False) setting = sqla.orm.relationship('EvaluationSetting', back_populates='raw_inputs', foreign_keys=[setting_uuid]) data_shape = sqla.Column(sqla.JSON) data_format = sqla.Column(sqla.String) evaluations = sqla.orm.relationship('Evaluation', back_populates='raw_input', cascade='all, delete, delete-orphan', foreign_keys='Evaluation.raw_input_uuid') extra_info = sqla.Column(sqla.JSON) hidden = sqla.Column(sqla.Boolean) def __repr__(self): return f'<RawInput(uuid="{self.uuid}", name="{self.name}")>' def __hash__(self): return hash((hash(self.uuid) + hash(self.name))) def __eq__(self, other): return (self.__hash__() == hash(other))
class Evaluation(sqlalchemy_base): __tablename__ = 'evaluations' uuid = sqla.Column(sqla.String, primary_key=True) name = sqla.Column(sqla.String, unique=True) creation_time = sqla.Column(sqla.DateTime(timezone=False), server_default=sqla.sql.func.now()) checkpoint_uuid = sqla.Column(sqla.String, sqla.ForeignKey('checkpoints.uuid'), nullable=False) checkpoint = sqla.orm.relationship('Checkpoint', back_populates='evaluations', foreign_keys=[checkpoint_uuid]) setting_uuid = sqla.Column(sqla.String, sqla.ForeignKey('evaluation_settings.uuid'), nullable=True) setting = sqla.orm.relationship('EvaluationSetting', back_populates='evaluations', foreign_keys=[setting_uuid]) raw_input_uuid = sqla.Column(sqla.String, sqla.ForeignKey('raw_inputs.uuid'), nullable=True) raw_input = sqla.orm.relationship('RawInput', back_populates='evaluations', foreign_keys=[raw_input_uuid]) chunks = sqla.orm.relationship('EvaluationChunk', back_populates='evaluation', cascade='all, delete, delete-orphan', foreign_keys='EvaluationChunk.evaluation_uuid') username = sqla.Column(sqla.String) extra_info = sqla.Column(sqla.JSON) completed = sqla.Column(sqla.Boolean) hidden = sqla.Column(sqla.Boolean) def __repr__(self): return f'<Evaluation(uuid="{self.uuid}", checkpoint_uuid="{self.checkpoint_uuid}")>' def __hash__(self): return hash((hash(self.uuid) + hash(self.name))) def __eq__(self, other): return (self.__hash__() == hash(other))
class EvaluationChunk(sqlalchemy_base): __tablename__ = 'evaluation_chunks' uuid = sqla.Column(sqla.String, primary_key=True) creation_time = sqla.Column(sqla.DateTime(timezone=False), server_default=sqla.sql.func.now()) evaluation_uuid = sqla.Column(sqla.String, sqla.ForeignKey('evaluations.uuid'), nullable=False) evaluation = sqla.orm.relationship('Evaluation', back_populates='chunks', foreign_keys=[evaluation_uuid]) username = sqla.Column(sqla.String) extra_info = sqla.Column(sqla.JSON) hidden = sqla.Column(sqla.Boolean) def __repr__(self): return f'<EvaluationChunk(uuid="{self.uuid}", evaluation_uuid="{self.evaluation_uuid}")>' def __hash__(self): return hash((hash(self.uuid) + hash(self.name))) def __eq__(self, other): return (self.hash() == hash(other))
class ModelRepository(): def __init__(self, mode=s3_utils.DB_CONNECTION_MODE, sql_verbose=False, download_database=True): self.sql_verbose = sql_verbose if (mode == 'sqlite'): if download_database: download_db() self.db_connection_string = s3_utils.DB_CONNECTION_STRING_SQLITE self.engine = sqla.create_engine(self.db_connection_string, echo=self.sql_verbose) elif (mode == 'rds'): self.db_connection_string = s3_utils.DB_CONNECTION_STRING_RDS self.engine = sqla.create_engine(self.db_connection_string, echo=self.sql_verbose, pool_pre_ping=True) else: assert False if (not database_exists(self.engine.url)): create_database(self.engine.url) self.sessionmaker = sqla.orm.sessionmaker(bind=self.engine, expire_on_commit=False) self.cache_root_path = s3_utils.default_cache_root_path self.s3wrapper = s3_utils.S3Wrapper(bucket='robustness-eval', cache_root_path=self.cache_root_path, verbose=False) if (mode == 'sqlite'): self.s3wrapper.put = (lambda *args, **kwargs: None) self.uuid_length = 10 self.pickle_protocol = 4 def dispose(self): self.engine.dispose() @contextlib.contextmanager def session_scope(self): session = self.sessionmaker() try: (yield session) session.commit() except: session.rollback() raise finally: session.close() def gen_short_uuid(self): new_id = gen_short_uuid(num_chars=self.uuid_length) return new_id def gen_checkpoint_uuid(self): return gen_short_uuid(num_chars=None) def run_query_with_optional_session(self, query, session=None): if (session is None): with self.session_scope() as sess: return query(sess) else: return query(session) def run_get(self, get_fn, session=None, assert_exists=True): def query(sess): result = get_fn(sess) assert (len(result) <= 1) if assert_exists: assert (len(result) == 1) if (len(result) == 0): return None else: return result[0] return self.run_query_with_optional_session(query, session) def get_model(self, *, uuid=None, name=None, session=None, assert_exists=True, load_final_checkpoint=False, load_all_checkpoints=False, load_evaluations=False): if (uuid is not None): assert (type(uuid) is str) if (name is not None): assert (type(name) is str) def get_fn(sess): return self.get_models(uuids=([uuid] if (uuid is not None) else None), names=([name] if (name is not None) else None), session=sess, load_final_checkpoint=load_final_checkpoint, load_all_checkpoints=load_all_checkpoints, load_evaluations=load_evaluations, show_hidden=True) return self.run_get(get_fn, session=session, assert_exists=assert_exists) def get_checkpoint(self, uuid=None, *, session=None, assert_exists=True, load_parents=False, load_evaluations=False): if (uuid is not None): assert (type(uuid) is str) def get_fn(sess): return self.get_checkpoints(uuids=[uuid], session=sess, load_parents=load_parents, load_evaluations=load_evaluations, show_hidden=True) return self.run_get(get_fn, session=session, assert_exists=assert_exists) def get_dataset(self, *, uuid=None, name=None, session=None, assert_exists=True, load_evaluation_settings=False): if (uuid is not None): assert (type(uuid) is str) if (name is not None): assert (type(name) is str) def get_fn(sess): return self.get_datasets(uuids=([uuid] if (uuid is not None) else None), names=([name] if (name is not None) else None), session=sess, load_evaluation_settings=load_evaluation_settings, show_hidden=True) return self.run_get(get_fn, session=session, assert_exists=assert_exists) def get_evaluation_setting(self, *, uuid=None, name=None, session=None, assert_exists=True, load_parents=False, load_evaluations=False, load_raw_inputs=False): if (uuid is not None): assert (type(uuid) is str) if (name is not None): assert (type(name) is str) def get_fn(sess): return self.get_evaluation_settings(uuids=([uuid] if (uuid is not None) else None), names=([name] if (name is not None) else None), session=sess, load_parents=load_parents, load_evaluations=load_evaluations, load_raw_inputs=load_raw_inputs, show_hidden=True) return self.run_get(get_fn, session=session, assert_exists=assert_exists) def get_raw_input(self, uuid=None, *, session=None, assert_exists=True, load_parents=False, load_evaluations=False): if (uuid is not None): assert (type(uuid) is str) def get_fn(sess): return self.get_raw_inputs(uuids=[uuid], session=sess, load_parents=load_parents, load_evaluations=load_evaluations, show_hidden=True) return self.run_get(get_fn, session=session, assert_exists=assert_exists) def get_evaluation(self, uuid=None, *, session=None, assert_exists=True, load_parents=False, load_chunks=True): if (uuid is not None): assert (type(uuid) is str) def get_fn(sess): return self.get_evaluations(uuids=[uuid], session=sess, load_parents=load_parents, load_chunks=load_chunks, show_hidden=True) return self.run_get(get_fn, session=session, assert_exists=assert_exists) def get_evaluation_chunk(self, *, uuid=None, session=None, assert_exists=True, load_parents=False): if (uuid is not None): assert (type(uuid) is str) def get_fn(sess): return self.get_evaluation_chunks(uuids=[uuid], session=sess, load_parents=load_parents, show_hidden=True) return self.run_get(get_fn, session=session, assert_exists=assert_exists) def model_uuid_exists(self, uuid, session=None): return (self.get_model(uuid=uuid, assert_exists=False, session=session) is not None) def checkpoint_uuid_exists(self, uuid, session=None): return (self.get_checkpoint(uuid=uuid, assert_exists=False, session=session) is not None) def dataset_uuid_exists(self, uuid, session=None): return (self.get_dataset(uuid=uuid, assert_exists=False, session=session) is not None) def evaluation_setting_uuid_exists(self, uuid, session=None): return (self.get_evaluation_setting(uuid=uuid, assert_exists=False, session=session) is not None) def raw_input_uuid_exists(self, uuid, session=None): return (self.get_raw_input(uuid=uuid, assert_exists=False, session=session) is not None) def evaluation_uuid_exists(self, uuid, session=None): return (self.get_evaluation(uuid=uuid, assert_exists=False, session=session) is not None) def evaluation_chunk_uuid_exists(self, uuid, session=None): return (self.get_evaluation_chunk(uuid=uuid, assert_exists=False, session=session) is not None) def get_checkpoints(self, uuids=None, *, session=None, load_parents=True, load_evaluations=False, show_hidden=False): cur_options = [] if load_parents: cur_options.append(sqla.orm.subqueryload(Checkpoint.model)) if load_evaluations: cur_options.append(sqla.orm.subqueryload(Checkpoint.evaluations)) filter_list = [] if (not show_hidden): filter_list.append((Checkpoint.hidden == False)) if (uuids is not None): filter_list.append(Checkpoint.uuid.in_(uuids)) def query(sess): return sess.query(Checkpoint).options(cur_options).filter(*filter_list).all() return self.run_query_with_optional_session(query, session) def get_datasets(self, *, uuids=None, names=None, session=None, load_evaluation_settings=True, show_hidden=False): cur_options = [] if load_evaluation_settings: cur_options.append(sqla.orm.subqueryload(Dataset.evaluation_settings)) filter_list = [] if (not show_hidden): filter_list.append((Dataset.hidden == False)) if (uuids is not None): filter_list.append(Dataset.uuid.in_(uuids)) if (names is not None): filter_list.append(Dataset.name.in_(names)) def query(sess): return sess.query(Dataset).options(cur_options).filter(*filter_list).all() return self.run_query_with_optional_session(query, session) def get_evaluation_settings(self, *, uuids=None, names=None, session=None, load_parents=True, load_evaluations=False, load_raw_inputs=False, show_hidden=False): cur_options = [] if load_parents: cur_options.append(sqla.orm.subqueryload(EvaluationSetting.dataset)) if load_evaluations: cur_options.append(sqla.orm.subqueryload(EvaluationSetting.evaluations).subqueryload(Evaluation.checkpoint).subqueryload(Checkpoint.model)) if load_raw_inputs: cur_options.append(sqla.orm.subqueryload(EvaluationSetting.raw_inputs)) filter_list = [] if (not show_hidden): filter_list.append((EvaluationSetting.hidden == False)) if (uuids is not None): filter_list.append(EvaluationSetting.uuid.in_(uuids)) if (names is not None): filter_list.append(EvaluationSetting.name.in_(names)) def query(sess): return sess.query(EvaluationSetting).options(cur_options).filter(*filter_list).all() return self.run_query_with_optional_session(query, session) def get_raw_inputs(self, uuids=None, *, session=None, load_parents=True, load_evaluations=False, show_hidden=False): cur_options = [] if load_parents: cur_options.append(sqla.orm.subqueryload(RawInput.setting).subqueryload(EvaluationSetting.dataset)) if load_evaluations: cur_options.append(sqla.orm.subqueryload(RawInput.evaluations).subqueryload(Evaluation.checkpoint).subqueryload(Checkpoint.model)) filter_list = [] if (not show_hidden): filter_list.append((RawInput.hidden == False)) if (uuids is not None): filter_list.append(RawInput.uuid.in_(uuids)) def query(sess): return sess.query(RawInput).options(cur_options).filter(*filter_list).all() return self.run_query_with_optional_session(query, session) def get_evaluations(self, uuids=None, *, session=None, load_parents=True, load_chunks=True, show_hidden=False): cur_options = [] if load_parents: cur_options.append(sqla.orm.subqueryload(Evaluation.checkpoint).subqueryload(Checkpoint.model)) cur_options.append(sqla.orm.subqueryload(Evaluation.raw_input)) cur_options.append(sqla.orm.subqueryload(Evaluation.setting).subqueryload(EvaluationSetting.dataset)) if load_chunks: cur_options.append(sqla.orm.subqueryload(Evaluation.chunks)) filter_list = [] if (not show_hidden): filter_list.append((Evaluation.hidden == False)) if (uuids is not None): filter_list.append(Evaluation.uuid.in_(uuids)) def query(sess): return sess.query(Evaluation).options(cur_options).filter(*filter_list).all() return self.run_query_with_optional_session(query, session) def get_evaluation_chunks(self, *, uuids=None, session=None, load_parents=False, show_hidden=False): cur_options = [] if load_parents: cur_options.append(sqla.orm.subqueryload(EvaluationChunk.evaluation)) filter_list = [] if (not show_hidden): filter_list.append((EvaluationChunk.hidden == False)) if (uuids is not None): filter_list.append(EvaluationChunk.uuid.in_(uuids)) def query(sess): return sess.query(EvaluationChunk).options(cur_options).filter(*filter_list).all() return self.run_query_with_optional_session(query, session) def get_models(self, *, uuids=None, names=None, session=None, load_parents=True, load_final_checkpoint=True, load_all_checkpoints=False, load_evaluations=False, show_hidden=False): cur_options = [] checkpoint_nodes = [] if load_final_checkpoint: cur_options.append(sqla.orm.subqueryload(Model.final_checkpoint)) checkpoint_nodes.append(cur_options[(- 1)]) if load_all_checkpoints: cur_options.append(sqla.orm.subqueryload(Model.checkpoints)) checkpoint_nodes.append(cur_options[(- 1)]) if load_evaluations: for opt in checkpoint_nodes: opt.subqueryload(Checkpoint.evaluations) filter_list = [] if (not show_hidden): filter_list.append((Model.hidden == False)) if (uuids is not None): filter_list.append(Model.uuid.in_(uuids)) if (names is not None): filter_list.append(Model.name.in_(names)) def query(sess): return sess.query(Model).options(cur_options).filter(*filter_list).all() return self.run_query_with_optional_session(query, session) def create_model(self, extra_info=None, name=None, description=None, verbose=False, completed=False): with self.session_scope() as session: new_id = self.gen_short_uuid() username = getpass.getuser() new_model = Model(uuid=new_id, name=name, description=description, username=username, extra_info=extra_info, hidden=False, completed=completed, logdir_filepaths={}, final_checkpoint_uuid=None) session.add(new_model) return self.get_model(uuid=new_id, assert_exists=True) def rename_model(self, model_uuid, new_name): with self.session_scope() as session: model = self.get_model(uuid=model_uuid, session=session, assert_exists=True) old_name = model.name model.name = new_name return old_name def hide_model(self, model_uuid): with self.session_scope() as session: model = self.get_model(uuid=model_uuid, session=session, assert_exists=True) model.hidden = True def get_latest_model_checkpoint_data(self, model_uuid, verbose=False, allow_non_final_checkpoint=True): with self.session_scope() as session: model = self.get_model(uuid=model_uuid, session=session, assert_exists=True) if (len(model.checkpoints) == 0): return (None, None) if allow_non_final_checkpoint: cur_checkpoints = sorted(model.checkpoints, key=(lambda x: x.training_step)) checkpoint_to_load = cur_checkpoints[(- 1)] else: assert (model.final_checkpoint is not None) checkpoint_to_load = model.final_checkpoint checkpoint_uuid = checkpoint_to_load.uuid key = get_checkpoint_data_key(checkpoint_uuid) if self.s3wrapper.exists(key): data = self.s3wrapper.get(key, verbose=verbose) else: data = None return (data, checkpoint_to_load) def mark_model_as_completed(self, model_uuid): with self.session_scope() as session: model = self.get_model(uuid=model_uuid, session=session, assert_exists=True) model.completed = True def set_final_model_checkpoint(self, model_uuid, checkpoint_uuid): with self.session_scope() as session: model = self.get_model(uuid=model_uuid, session=session, assert_exists=True) checkpoint = self.get_checkpoint(checkpoint_uuid, session=session, assert_exists=True) assert (checkpoint.model_uuid == model_uuid) model.final_checkpoint_uuid = checkpoint_uuid def store_logdir(self, model_uuid, logdir, verbose=False): with self.session_scope() as session: model = self.get_model(uuid=model_uuid, session=session, assert_exists=True) logdir_path = pathlib.Path(logdir).resolve() assert logdir_path.is_dir() tmp_filepaths = [x for x in logdir_path.glob('**/*') if x.is_file()] all_data = {} base_key = (get_logdir_key(model_uuid) + '/') cur_logdir_files = {} for cur_filepath in tmp_filepaths: cur_filepath_resolved = cur_filepath.resolve() with open(cur_filepath_resolved, 'rb') as f: cur_data = f.read() cur_relative_path = str(cur_filepath.relative_to(logdir_path)) assert (cur_relative_path not in cur_logdir_files) cur_logdir_files[cur_relative_path] = {'size': cur_filepath_resolved.stat().st_size, 'mtime': cur_filepath_resolved.stat().st_mtime} cur_key = (base_key + cur_relative_path) all_data[cur_key] = cur_data self.s3wrapper.put_multiple(all_data, verbose=verbose) model.logdir_filepaths = cur_logdir_files sqla.orm.attributes.flag_modified(model, 'logdir_filepaths') def create_checkpoint(self, *, model_uuid, training_step=None, epoch=None, name=None, data_bytes=None, extra_info=None, verbose=False): with self.session_scope() as session: assert self.model_uuid_exists(model_uuid, session=session) new_id = self.gen_checkpoint_uuid() username = getpass.getuser() new_checkpoint = Checkpoint(uuid=new_id, model_uuid=model_uuid, username=username, extra_info=extra_info, name=name, training_step=training_step, epoch=epoch, hidden=False) if (data_bytes is not None): key = get_checkpoint_data_key(new_id) self.s3wrapper.put(data_bytes, key, verbose=verbose) session.add(new_checkpoint) return self.get_checkpoint(uuid=new_id, assert_exists=True) def get_checkpoint_data(self, checkpoint_uuid, verbose=False): with self.session_scope() as session: assert self.checkpoint_uuid_exists(checkpoint_uuid, session=session) key = get_checkpoint_data_key(checkpoint_uuid) if self.s3wrapper.exists(key): return self.s3wrapper.get(key, verbose=verbose) else: return None def create_evaluation(self, *, checkpoint_uuid, setting_uuid, name=None, logits_data_bytes=None, extra_data_bytes=None, raw_input_uuid=None, extra_info=None, completed=False, verbose=False): with self.session_scope() as session: assert self.checkpoint_uuid_exists(checkpoint_uuid, session=session) assert self.evaluation_setting_uuid_exists(setting_uuid, session=session) if (raw_input_uuid is not None): assert self.raw_input_uuid_exists(raw_input_uuid, session=session) new_id = self.gen_short_uuid() username = getpass.getuser() new_evaluation = Evaluation(uuid=new_id, checkpoint_uuid=checkpoint_uuid, setting_uuid=setting_uuid, raw_input_uuid=raw_input_uuid, username=username, extra_info=extra_info, name=name, completed=completed, hidden=False) if (extra_data_bytes is not None): key = get_evaluation_extra_data_key(new_id) self.s3wrapper.put(extra_data_bytes, key, verbose=verbose) if (logits_data_bytes is not None): key = get_evaluation_logits_data_key(new_id) self.s3wrapper.put(logits_data_bytes, key, verbose=verbose) session.add(new_evaluation) return self.get_evaluation(uuid=new_id, assert_exists=True) def hide_evaluation(self, evaluation_uuid): with self.session_scope() as session: evaluation = self.get_evaluation(evaluation_uuid, session=session, assert_exists=True) evaluation.hidden = True def rename_evaluation(self, evaluation_uuid, new_name): with self.session_scope() as session: evaluation = self.get_evaluation(evaluation_uuid, session=session, assert_exists=True) old_name = evaluation.name evaluation.name = new_name return old_name def mark_evaluation_as_completed(self, evaluation_uuid): with self.session_scope() as session: evaluation = self.get_evaluation(uuid=evaluation_uuid, session=session, assert_exists=True) evaluation.completed = True def get_evaluation_extra_data(self, evaluation_uuid, verbose=False): with self.session_scope() as session: assert self.evaluation_uuid_exists(evaluation_uuid, session=session) key = get_evaluation_extra_data_key(evaluation_uuid) if self.s3wrapper.exists(key): return self.s3wrapper.get(key, verbose=verbose) else: return None def get_evaluation_logits_data(self, evaluation_uuid, verbose=False): with self.session_scope() as session: assert self.evaluation_uuid_exists(evaluation_uuid, session=session) key = get_evaluation_logits_data_key(evaluation_uuid) if self.s3wrapper.exists(key): return self.s3wrapper.get(key, verbose=verbose) else: return None def has_evaluation_logits_data(self, evaluation_uuid): with self.session_scope() as session: assert self.evaluation_uuid_exists(evaluation_uuid, session=session) key = get_evaluation_logits_data_key(evaluation_uuid) return self.s3wrapper.exists(key) def put_evaluation_extra_data(self, evaluation_uuid, extra_data_bytes, verbose=False): with self.session_scope() as session: assert self.evaluation_uuid_exists(evaluation_uuid, session=session) key = get_evaluation_extra_data_key(evaluation_uuid) self.s3wrapper.put(extra_data_bytes, key, verbose=verbose) def put_evaluation_logits_data(self, evaluation_uuid, logits_data_bytes, verbose=False): with self.session_scope() as session: assert self.evaluation_uuid_exists(evaluation_uuid, session=session) key = get_evaluation_logits_data_key(evaluation_uuid) self.s3wrapper.put(logits_data_bytes, key, verbose=verbose) def create_dataset(self, *, name, size, description=None, data_bytes=None, data_filename=None, extra_info=None, verbose=False): assert (name is not None) assert (size is not None) assert (type(size) is int) assert ((data_bytes is None) or (data_filename is None)) assert ((data_bytes is not None) or (data_filename is not None)) with self.session_scope() as session: new_id = self.gen_short_uuid() username = getpass.getuser() new_dataset = Dataset(uuid=new_id, name=name, description=description, username=username, size=size, extra_info=extra_info, hidden=False) key = get_dataset_data_key(new_id) if (data_bytes is not None): self.s3wrapper.put(data_bytes, key, verbose=verbose) else: assert (data_filename is not None) self.s3wrapper.upload_file(data_filename, key, verbose=verbose) session.add(new_dataset) return self.get_dataset(uuid=new_id, assert_exists=True) def get_dataset_data(self, dataset_uuid, verbose=False): with self.session_scope() as session: assert self.dataset_uuid_exists(dataset_uuid, session=session) key = get_dataset_data_key(dataset_uuid) if self.s3wrapper.exists(key): return self.s3wrapper.get(key, verbose=verbose) else: return None def download_dataset_data(self, dataset_uuid, target_filename, verbose=False): with self.session_scope() as session: assert self.dataset_uuid_exists(dataset_uuid, session=session) key = get_dataset_data_key(dataset_uuid) if self.s3wrapper.exists(key): return self.s3wrapper.download_file(key, target_filename, verbose=verbose) def rename_dataset(self, dataset_uuid, new_name): with self.session_scope() as session: dataset = self.get_dataset(uuid=dataset_uuid, session=session, assert_exists=True) old_name = dataset.name dataset.name = new_name return old_name def hide_dataset(self, dataset_uuid): with self.session_scope() as session: dataset = self.get_dataset(uuid=dataset_uuid, session=session, assert_exists=True) dataset.hidden = True def create_evaluation_setting(self, *, name, dataset_uuid=None, description=None, extra_info=None, processed_dataset_bytes=None, processed_dataset_filename=None, extra_data_bytes=None, verbose=False): assert (name is not None) assert ((processed_dataset_filename is None) or (processed_dataset_bytes is None)) with self.session_scope() as session: if (dataset_uuid is not None): assert self.dataset_uuid_exists(dataset_uuid, session=session) new_id = self.gen_short_uuid() username = getpass.getuser() new_setting = EvaluationSetting(uuid=new_id, name=name, description=description, username=username, dataset_uuid=dataset_uuid, extra_info=extra_info, hidden=False) if (extra_data_bytes is not None): key = get_evaluation_setting_extra_data_key(new_id) self.s3wrapper.put(extra_data_bytes, key, verbose=verbose) key = get_evaluation_setting_processed_dataset_key(new_id) if (processed_dataset_bytes is not None): self.s3wrapper.put(processed_dataset_bytes, key, verbose=verbose) elif (processed_dataset_filename is not None): self.s3wrapper.upload_file(processed_dataset_filename, key, verbose=verbose) session.add(new_setting) return self.get_evaluation_setting(uuid=new_id, assert_exists=True) def hide_evaluation_setting(self, evaluation_setting_uuid): with self.session_scope() as session: evaluation_setting = self.get_evaluation_setting(uuid=evaluation_setting_uuid, session=session, assert_exists=True) evaluation_setting.hidden = True def rename_evaluation_setting(self, evaluation_setting_uuid, new_name): with self.session_scope() as session: evaluation_setting = self.get_evaluation_setting(uuid=evaluation_setting_uuid, session=session, assert_exists=True) old_name = evaluation_setting.name evaluation_setting.name = new_name return old_name def get_evaluation_setting_extra_data(self, evaluation_setting_uuid, verbose=False): with self.session_scope() as session: assert self.evaluation_setting_uuid_exists(evaluation_setting_uuid, session=session) key = get_evaluation_setting_extra_data_key(evaluation_setting_uuid) if self.s3wrapper.exists(key): return self.s3wrapper.get(key, verbose=verbose) else: return None def download_evaluation_setting_processed_dataset_data(self, evaluation_setting_uuid, target_filename, verbose=False): with self.session_scope() as session: assert self.evaluation_setting_uuid_exists(evaluation_setting_uuid, session=session) key = get_evaluation_setting_processed_dataset_key(evaluation_setting_uuid) if self.s3wrapper.exists(key): return self.s3wrapper.download_file(key, target_filename, verbose=verbose) def get_evaluation_setting_processed_dataset_data(self, evaluation_setting_uuid, verbose=False): with self.session_scope() as session: assert self.evaluation_setting_uuid_exists(evaluation_setting_uuid, session=session) key = get_evaluation_setting_processed_dataset_key(evaluation_setting_uuid) if self.s3wrapper.exists(key): return self.s3wrapper.get(key, verbose=verbose) else: return None def create_raw_input(self, *, name, evaluation_setting_uuid, data_shape, data_format, description=None, extra_info=None, data_bytes=None, data_filename=None, verbose=False): assert (name is not None) assert (evaluation_setting_uuid is not None) assert ((data_bytes is None) or (data_filename is None)) assert ((data_bytes is not None) or (data_filename is not None)) assert (data_format in ['float32', 'float64', 'uint8']) assert (type(data_shape) is list) for x in data_shape: assert (type(x) is int) with self.session_scope() as session: assert self.evaluation_setting_uuid_exists(evaluation_setting_uuid, session=session) new_id = self.gen_short_uuid() username = getpass.getuser() new_raw_input = RawInput(uuid=new_id, name=name, description=description, username=username, data_shape=data_shape, data_format=data_format, setting_uuid=evaluation_setting_uuid, extra_info=extra_info, hidden=False) key = get_raw_input_data_key(new_id) if (data_bytes is not None): self.s3wrapper.put(data_bytes, key, verbose=verbose) else: assert (data_filename is not None) self.s3wrapper.upload_file(data_filename, key, verbose=verbose) session.add(new_raw_input) return self.get_raw_input(uuid=new_id, assert_exists=True) def hide_raw_input(self, raw_input_uuid): with self.session_scope() as session: raw_input = self.get_raw_input(raw_input_uuid, session=session, assert_exists=True) raw_input.hidden = True def rename_raw_input(self, raw_input_uuid, new_name): with self.session_scope() as session: raw_input = self.get_raw_input(raw_input_uuid, session=session, assert_exists=True) old_name = raw_input.name raw_input.name = new_name return old_name def download_raw_input_data(self, raw_input_uuid, target_filename, verbose=False): with self.session_scope() as session: assert self.raw_input_uuid_exists(raw_input_uuid, session=session) key = get_raw_input_data_key(raw_input_uuid) if self.s3wrapper.exists(key): return self.s3wrapper.download_file(key, target_filename, verbose=verbose) def get_raw_input_data(self, raw_input_uuid, verbose=False): with self.session_scope() as session: assert self.raw_input_uuid_exists(raw_input_uuid, session=session) key = get_raw_input_data_key(raw_input_uuid) if self.s3wrapper.exists(key): return self.s3wrapper.get(key, verbose=verbose) else: return None def create_evaluation_chunk(self, *, evaluation_uuid, logits_data_bytes=None, indices=None, indices_bytes=None, extra_data_bytes=None, extra_info=None, verbose=False): if (indices is not None): assert (indices_bytes is None) indices_bytes = pickle.dumps(indices) else: assert (indices_bytes is not None) with self.session_scope() as session: assert self.evaluation_uuid_exists(evaluation_uuid, session=session) new_id = self.gen_short_uuid() username = getpass.getuser() new_chunk = EvaluationChunk(uuid=new_id, evaluation_uuid=evaluation_uuid, username=username, extra_info=extra_info, hidden=False) if (extra_data_bytes is not None): key = get_evaluation_chunk_extra_data_key(new_id) self.s3wrapper.put(extra_data_bytes, key, verbose=verbose) if (logits_data_bytes is not None): key = get_evaluation_chunk_logits_data_key(new_id) self.s3wrapper.put(logits_data_bytes, key, verbose=verbose) if (indices_bytes is not None): key = get_evaluation_chunk_indices_data_key(new_id) self.s3wrapper.put(indices_bytes, key, verbose=verbose) session.add(new_chunk) return self.get_evaluation_chunk(uuid=new_id, assert_exists=True) def hide_evaluation_chunk(self, evaluation_chunk_uuid): with self.session_scope() as session: evaluation_chunk = self.get_evaluation_chunk(uuid=evaluation_chunk_uuid, session=session, assert_exists=True) evaluation_chunk.hidden = True def get_evaluation_chunk_extra_data(self, evaluation_chunk_uuid, verbose=False): with self.session_scope() as session: assert self.evaluation_chunk_uuid_exists(evaluation_chunk_uuid, session=session) key = get_evaluation_chunk_extra_data_key(evaluation_chunk_uuid) if self.s3wrapper.exists(key): return self.s3wrapper.get(key, verbose=verbose) else: return None def get_evaluation_chunk_logits_data(self, evaluation_chunk_uuid, verbose=False): with self.session_scope() as session: assert self.evaluation_chunk_uuid_exists(evaluation_chunk_uuid, session=session) key = get_evaluation_chunk_logits_data_key(evaluation_chunk_uuid) if self.s3wrapper.exists(key): return self.s3wrapper.get(key, verbose=verbose) else: return None def get_evaluation_chunk_indices_data(self, evaluation_chunk_uuid, verbose=False, unpickle=False): with self.session_scope() as session: assert self.evaluation_chunk_uuid_exists(evaluation_chunk_uuid, session=session) key = get_evaluation_chunk_indices_data_key(evaluation_chunk_uuid) if self.s3wrapper.exists(key): data = self.s3wrapper.get(key, verbose=verbose) if unpickle: return pickle.loads(data) else: return data else: return None
def tar_directory(dir_name, target_filename): subprocess.run(['tar', '-cf', str(target_filename), str(dir_name)], check=True)
def untar_directory(tar_filename, target_dir, strip=None, one_top_level=False): cmd = ['tar', '-xf', str(tar_filename)] if strip: cmd += [f'--strip={strip}'] cmd += ['-C', str(target_dir)] if one_top_level: cmd += ['--one-top-level'] subprocess.run(cmd, check=True) subprocess.run(['rm', tar_filename], check=True)
def get_s3_client(): if (DB_CONNECTION_MODE == 'rds'): if (default_profile in boto3.Session()._session.available_profiles): session = boto3.Session(profile_name=default_profile) else: session = boto3.Session() client = session.client('s3', endpoint_url='https://vasa.millennium.berkeley.edu:9000', aws_access_key_id='robustness-eval', aws_secret_access_key='rtB_HizvjHVl59_HgKjOBYZJZTbXjNRHbIsBEj5D4g4', config=Config(connect_timeout=250, read_timeout=250), verify=(pathlib.Path(__file__).parent / 'vasa_chain.cer').resolve(), region_name='us-east-1') return client elif (DB_CONNECTION_MODE == 'sqlite'): client = boto3.client('s3', endpoint_url='https://vasa.millennium.berkeley.edu:9000', config=Config(signature_version=botocore.UNSIGNED), verify=(pathlib.Path(__file__).parent / 'vasa_chain.cer').resolve(), region_name='us-east-1') return client
def key_exists(bucket, key): client = get_s3_client() try: client.head_object(Bucket=bucket, Key=key) return True except botocore.exceptions.ClientError as exc: if (exc.response['Error']['Code'] != '404'): raise return False except: raise
def get_s3_object_bytes_parallel(keys, *, bucket, cache_on_local_disk=True, cache_root_path=None, verbose=False, special_verbose=True, max_num_threads=90, num_tries=5, initial_delay=1.0, delay_factor=math.sqrt(2.0), download_callback=None, skip_modification_time_check=False): if cache_on_local_disk: assert (cache_root_path is not None) cache_root_path = pathlib.Path(cache_root_path).resolve() missing_keys = [] existing_keys = [] for key in keys: local_filepath = (cache_root_path / key) if (not local_filepath.is_file()): missing_keys.append(key) local_filepath.parent.mkdir(parents=True, exist_ok=True) else: existing_keys.append(key) keys_to_download = missing_keys.copy() if skip_modification_time_check: if verbose: print(f'Skipping the file modification time check for {len(existing_keys)} keys that have local copies.') for key in existing_keys: if download_callback: download_callback(1) else: if verbose: print(f'Getting metadata for {len(existing_keys)} keys that have local copies ... ', end='') metadata_start = timer() metadata = get_s3_object_metadata_parallel(existing_keys, bucket=bucket, verbose=False, max_num_threads=max_num_threads, num_tries=num_tries, initial_delay=initial_delay, delay_factor=delay_factor, download_callback=None) metadata_end = timer() if verbose: print(f'took {(metadata_end - metadata_start):.3f} seconds') for key in existing_keys: local_filepath = (cache_root_path / key) assert local_filepath.is_file local_time = datetime.datetime.fromtimestamp(local_filepath.stat().st_mtime, datetime.timezone.utc) remote_time = metadata[key]['LastModified'] if (local_time <= remote_time): if verbose: print(f'Local copy of key "{key}" is outdated') keys_to_download.append(key) elif download_callback: download_callback(1) tl = threading.local() def cur_download_file(key): local_filepath = (cache_root_path / key) if (verbose or special_verbose): print('{} not available locally or outdated, downloading from S3 ... '.format(key)) download_s3_file_with_backoff(key, str(local_filepath), bucket=bucket, num_tries=num_tries, initial_delay=initial_delay, delay_factor=delay_factor, thread_local=tl) return local_filepath.is_file() if (len(keys_to_download) > 0): download_start = timer() with concurrent.futures.ThreadPoolExecutor(max_workers=max_num_threads) as executor: future_to_key = {executor.submit(cur_download_file, key): key for key in keys_to_download} for future in concurrent.futures.as_completed(future_to_key): key = future_to_key[future] try: success = future.result() assert success if download_callback: download_callback(1) except Exception as exc: print('Key {} generated an exception: {}'.format(key, exc)) raise exc download_end = timer() if verbose: print('Downloading took {:.3f} seconds'.format((download_end - download_start))) result = {} for key in keys: local_filepath = (cache_root_path / key) if verbose: print('Reading from local file {} ... '.format(local_filepath), end='') with open(local_filepath, 'rb') as f: result[key] = f.read() if verbose: print('done') else: tl = threading.local() def cur_get_object_bytes(key): if verbose: print('Loading {} from S3 ... '.format(key)) return get_s3_object_bytes_with_backoff(key, bucket=bucket, num_tries=num_tries, initial_delay=initial_delay, delay_factor=delay_factor, thread_local=tl)[0] download_start = timer() result = {} with concurrent.futures.ThreadPoolExecutor(max_workers=max_num_threads) as executor: future_to_key = {executor.submit(cur_get_object_bytes, key): key for key in keys} for future in concurrent.futures.as_completed(future_to_key): key = future_to_key[future] try: result[key] = future.result() if download_callback: download_callback(1) except Exception as exc: print('Key {} generated an exception: {}'.format(key, exc)) raise exc download_end = timer() if verbose: print('Getting object bytes took {} seconds'.format((download_end - download_start))) return result
def get_s3_object_bytes_with_backoff(key, *, bucket, num_tries=5, initial_delay=1.0, delay_factor=math.sqrt(2.0), num_replicas=1, thread_local=None): if (thread_local is None): client = get_s3_client() else: if (not hasattr(thread_local, 'get_object_client')): thread_local.get_object_client = get_s3_client() client = thread_local.get_object_client delay = initial_delay num_tries_left = num_tries if (num_replicas > 1): replicas_counter_len = len(str(num_replicas)) format_string = '_replica{{:0{}d}}-{{}}'.format(replicas_counter_len) while (num_tries_left >= 1): try: if (num_replicas > 1): cur_replica = random.randint(1, num_replicas) cur_key = (key + format_string.format(cur_replica, num_replicas)) else: cur_key = key read_bytes = client.get_object(Key=cur_key, Bucket=bucket)['Body'].read() return (read_bytes, cur_key) except: if (num_tries_left == 1): raise Exception(((('get backoff failed ' + key) + ' ') + str(delay))) else: time.sleep(delay) delay *= delay_factor num_tries_left -= 1
def get_s3_object_metadata_with_backoff(key, *, bucket, num_tries=5, initial_delay=1.0, delay_factor=math.sqrt(2.0), thread_local=None): if (thread_local is None): client = get_s3_client() else: if (not hasattr(thread_local, 'get_object_client')): thread_local.get_object_client = get_s3_client() client = thread_local.get_object_client delay = initial_delay num_tries_left = num_tries while (num_tries_left >= 1): try: metadata = client.head_object(Key=key, Bucket=bucket) return metadata except: if (num_tries_left == 1): raise Exception(((('get backoff failed ' + key) + ' ') + str(delay))) else: time.sleep(delay) delay *= delay_factor num_tries_left -= 1
def get_s3_object_metadata_parallel(keys, bucket, verbose=False, max_num_threads=20, num_tries=5, initial_delay=1.0, delay_factor=math.sqrt(2.0), download_callback=None): tl = threading.local() def cur_get_object_metadata(key): if verbose: print('Loading metadata for {} from S3 ... '.format(key)) return get_s3_object_metadata_with_backoff(key, bucket=bucket, num_tries=num_tries, initial_delay=initial_delay, delay_factor=delay_factor, thread_local=tl) download_start = timer() result = {} with concurrent.futures.ThreadPoolExecutor(max_workers=max_num_threads) as executor: future_to_key = {executor.submit(cur_get_object_metadata, key): key for key in keys} for future in concurrent.futures.as_completed(future_to_key): key = future_to_key[future] try: result[key] = future.result() if download_callback: download_callback(1) except Exception as exc: print('Key {} generated an exception: {}'.format(key, exc)) raise exc download_end = timer() if verbose: print('Getting object metadata took {} seconds'.format((download_end - download_start))) return result
def put_s3_object_bytes_with_backoff(file_bytes, key, bucket, num_tries=10, initial_delay=1.0, delay_factor=2.0): client = get_s3_client() delay = initial_delay num_tries_left = num_tries while (num_tries_left >= 1): try: bio = io.BytesIO(file_bytes) client.upload_fileobj(bio, Key=key, Bucket=bucket, ExtraArgs={'ACL': 'bucket-owner-full-control'}) return except: if (num_tries_left == 1): print(('put backoff failed' + key)) raise Exception(((((('put backoff failed ' + key) + ' ') + str(len(file_bytes))) + ' ') + str(delay))) else: time.sleep(delay) delay *= delay_factor num_tries_left -= 1
def list_all_keys(client, bucket, prefix, max_keys=None): objects = client.list_objects(Bucket=bucket, Prefix=prefix, Delimiter=prefix) if (objects.get('Contents') == None): return [] keys = list(map((lambda x: x['Key']), objects.get('Contents', []))) truncated = objects['IsTruncated'] next_marker = objects.get('NextMarker') while truncated: objects = client.list_objects(Bucket=bucket, Prefix=prefix, Delimiter=prefix, Marker=next_marker) truncated = objects['IsTruncated'] next_marker = objects.get('NextMarker') keys += list(map((lambda x: x['Key']), objects['Contents'])) if ((max_keys is not None) and (len(keys) >= max_keys)): break return list(filter((lambda x: (len(x) > 0)), keys))
def download_s3_file_with_caching(key, local_filename, *, bucket, cache_on_local_disk=True, cache_root_path=None, verbose=False, special_verbose=True, num_tries=5, initial_delay=1.0, delay_factor=math.sqrt(2.0), num_replicas=1, skip_modification_time_check=False): if cache_on_local_disk: assert (cache_root_path is not None) cache_root_path = pathlib.Path(cache_root_path).resolve() currently_cached = False cache_filepath = (cache_root_path / key) if (not cache_filepath.is_file()): cache_filepath.parent.mkdir(parents=True, exist_ok=True) elif skip_modification_time_check: if verbose: print(f'Skipping the file modification time check the local copy in the cache.') currently_cached = True else: if verbose: print(f'Getting metadata to check the modification time compared to the local copy ... ', end='') metadata_start = timer() metadata = get_s3_object_metadata_with_backoff(key, bucket=bucket, num_tries=num_tries, initial_delay=initial_delay, delay_factor=delay_factor) metadata_end = timer() if verbose: print(f'took {(metadata_end - metadata_start):.3f} seconds') local_time = datetime.datetime.fromtimestamp(cache_filepath.stat().st_mtime, datetime.timezone.utc) remote_time = metadata['LastModified'] if (local_time <= remote_time): if verbose: print(f'Local copy of key "{key}" is outdated') else: currently_cached = True if (not currently_cached): if (verbose or special_verbose): print('{} not available locally or outdated, downloading from S3 ... '.format(key)) download_start = timer() download_s3_file_with_backoff(key, str(cache_filepath), bucket=bucket, initial_delay=initial_delay, delay_factor=delay_factor, num_replicas=num_replicas) download_end = timer() if verbose: print('Downloading took {:.3f} seconds'.format((download_end - download_start))) assert cache_filepath.is_file() if verbose: print(f'Copying to the target from the cache file {cache_filepath} ...') shutil.copy(cache_filepath, local_filename) else: if verbose: print('Loading {} from S3 ... '.format(key)) download_start = timer() download_s3_file_with_backoff(key, local_filename, bucket=bucket, initial_delay=initial_delay, delay_factor=delay_factor, num_replicas=num_replicas) download_end = timer() if verbose: print('Downloading took {:.3f} seconds'.format((download_end - download_start)))
def download_s3_file_with_backoff(key, local_filename, *, bucket, num_tries=5, initial_delay=1.0, delay_factor=math.sqrt(2.0), num_replicas=1, thread_local=None): if (thread_local is None): client = get_s3_client() else: if (not hasattr(thread_local, 's3_client')): thread_local.s3_client = get_s3_client() client = thread_local.s3_client delay = initial_delay num_tries_left = num_tries if (num_replicas > 1): replicas_counter_len = len(str(num_replicas)) format_string = '_replica{{:0{}d}}-{{}}'.format(replicas_counter_len) while (num_tries_left >= 1): try: if (num_replicas > 1): cur_replica = random.randint(1, num_replicas) cur_key = (key + format_string.format(cur_replica, num_replicas)) else: cur_key = key client.download_file(bucket, cur_key, local_filename) return cur_key except: if (num_tries_left == 1): raise Exception((((('download backoff failed ' + ' ') + str(key)) + ' ') + str(delay))) else: time.sleep(delay) delay *= delay_factor num_tries_left -= 1
def upload_file_to_s3_with_backoff(local_filename, key, *, bucket, num_tries=5, initial_delay=1.0, delay_factor=math.sqrt(2.0), thread_local=None): assert pathlib.Path(local_filename).is_file() if (thread_local is None): client = get_s3_client() else: if (not hasattr(thread_local, 's3_client')): thread_local.s3_client = get_s3_client() client = thread_local.s3_client delay = initial_delay num_tries_left = num_tries while (num_tries_left >= 1): try: client.upload_file(local_filename, bucket, key, ExtraArgs={'ACL': 'bucket-owner-full-control'}) return except: if (num_tries_left == 1): raise Exception((((('upload backoff failed ' + ' ') + str(key)) + ' ') + str(delay))) else: time.sleep(delay) delay *= delay_factor num_tries_left -= 1
def default_option_if_needed(*, user_option, default): if (user_option is None): return default else: return user_option
class S3Wrapper(): def __init__(self, bucket, cache_on_local_disk=True, cache_root_path=default_cache_root_path, verbose=False, max_num_threads=90, num_tries=5, initial_delay=1.0, delay_factor=math.sqrt(2.0), skip_modification_time_check=False): self.bucket = bucket self.cache_on_local_disk = cache_on_local_disk self.client = get_s3_client() if self.cache_on_local_disk: assert (cache_root_path is not None) self.cache_root_path = pathlib.Path(cache_root_path).resolve() self.cache_root_path.mkdir(parents=True, exist_ok=True) assert self.cache_root_path.is_dir() else: self.cache_root_path = None self.verbose = verbose self.max_num_threads = max_num_threads self.num_tries = num_tries self.initial_delay = initial_delay self.delay_factor = delay_factor self.skip_modification_time_check = skip_modification_time_check def list_keys(self, prefix, max_keys=None): return list_all_keys(self.client, self.bucket, prefix, max_keys) def put(self, bytes_to_store, key, verbose=None): cur_verbose = default_option_if_needed(user_option=verbose, default=self.verbose) put_s3_object_bytes_with_backoff(bytes_to_store, key, bucket=self.bucket, num_tries=self.num_tries, initial_delay=self.initial_delay, delay_factor=self.delay_factor) if cur_verbose: print('Stored {} bytes under key {}'.format(len(bytes_to_store), key)) def put_multiple(self, data, verbose=None, callback=None): for (key, bytes_to_store) in data.items(): self.put(bytes_to_store, key, verbose) def upload_file(self, filename, key, verbose=None): cur_verbose = default_option_if_needed(user_option=verbose, default=self.verbose) upload_file_to_s3_with_backoff(filename, key, bucket=self.bucket, num_tries=self.num_tries, initial_delay=self.initial_delay, delay_factor=self.delay_factor, thread_local=None) def download_file(self, key, filename, verbose=None, skip_modification_time_check=None): cur_verbose = default_option_if_needed(user_option=verbose, default=self.verbose) cur_skip_time_check = default_option_if_needed(user_option=skip_modification_time_check, default=self.skip_modification_time_check) download_s3_file_with_caching(key, filename, bucket=self.bucket, cache_on_local_disk=self.cache_on_local_disk, cache_root_path=self.cache_root_path, num_tries=self.num_tries, initial_delay=self.initial_delay, delay_factor=self.delay_factor, skip_modification_time_check=cur_skip_time_check, verbose=cur_verbose) def get(self, key, verbose=None, skip_modification_time_check=None): return self.get_multiple([key], verbose=verbose, skip_modification_time_check=skip_modification_time_check)[key] def get_multiple(self, keys, verbose=None, callback=None, skip_modification_time_check=None): if (verbose is None): cur_verbose = self.verbose else: cur_verbose = verbose cur_verbose = default_option_if_needed(user_option=verbose, default=self.verbose) cur_skip_time_check = default_option_if_needed(user_option=skip_modification_time_check, default=self.skip_modification_time_check) return get_s3_object_bytes_parallel(keys, bucket=self.bucket, cache_on_local_disk=self.cache_on_local_disk, cache_root_path=self.cache_root_path, verbose=cur_verbose, max_num_threads=self.max_num_threads, num_tries=self.num_tries, initial_delay=self.initial_delay, delay_factor=self.delay_factor, download_callback=callback, skip_modification_time_check=cur_skip_time_check) def exists(self, key): return key_exists(self.bucket, key)
def evaluation_completed(py_model, py_eval_setting): assert (py_model.name in MODEL_NAMES), (f'Model {py_model.name} is not recognized as an existing model in the' + ' server. Did you run the db script?') assert (py_eval_setting.name in EVAL_SETTING_NAMES), (f'Model {py_eval_setting.name} is not recognized as an existing eval setting in the' + ' server. Did you run the db script?') checkpoint = m_repo.get_model(name=py_model.name, load_final_checkpoint=True, load_evaluations=True).final_checkpoint setting_uuids = [m_repo.get_evaluation_setting(name=py_eval_setting.name).uuid] if (py_eval_setting.parent_eval_setting is not None): assert (py_eval_setting.parent_eval_setting in EVAL_SETTING_NAMES), (f'Eval setting {py_eval_setting.parent_eval_setting} is not recognized as an existing eval setting in the' + ' server. Did you run the db script?') setting_uuids += [m_repo.get_evaluation_setting(name=py_eval_setting.parent_eval_setting).uuid] for e in m_repo.get_evaluations([x.uuid for x in checkpoint.evaluations]): if ((e.setting_uuid in setting_uuids) and e.completed): return True return False
def store_evaluation(py_model, py_eval_setting, metrics, logits): assert (py_model.name in MODEL_NAMES), (f'Model {py_model.name} is not recognized as an existing model in the' + ' server. Did you run the db script?') assert (py_eval_setting.name in EVAL_SETTING_NAMES), (f'Model {py_eval_setting.name} is not recognized as an existing eval setting in the' + ' server. Did you run the db script?') bio = io.BytesIO() torch.save(logits.cpu(), bio) model_uuid = m_repo.get_model(name=py_model.name).final_checkpoint_uuid setting_uuid = m_repo.get_evaluation_setting(name=py_eval_setting.name).uuid eval = m_repo.create_evaluation(checkpoint_uuid=model_uuid, setting_uuid=setting_uuid, extra_info=metrics, logits_data_bytes=bio.getvalue(), completed=True)
def download_dataset(dataset): dataset = m_repo.get_dataset(name=dataset) filedir = join(default_cache_root_path, f'datasets/{dataset.name}') if (not exists(filedir)): filename = (filedir + '.tar') m_repo.download_dataset_data(dataset_uuid=dataset.uuid, target_filename=filename) if ('format-val' in dataset.name): (strip, one_top_level) = (2, False) elif ('imagenet-c' in dataset.name): (strip, one_top_level) = (6, True) elif (dataset.name in ['imagenetv2-matched-frequency', 'imagenetv2-topimages', 'imagenetv2-threshold0.7', 'val']): (strip, one_top_level) = (3, False) else: (strip, one_top_level) = (1, True) untar_directory(filename, join(default_cache_root_path, 'datasets'), strip=strip, one_top_level=one_top_level) return filedir
def load_model_checkpoint_bytes(model_name): r_model = m_repo.get_model(name=model_name) data = m_repo.get_checkpoint_data(r_model.final_checkpoint_uuid) bio = io.BytesIO(data) return bio
def load_model_state_dict(model, name): bio = load_model_checkpoint_bytes(name) state_dict = torch.load(bio, map_location=f'cpu') if ('state_dict' in state_dict): state_dict = state_dict['state_dict'] model.load_state_dict(state_dict)
def add_model_shell(model_name): r_model = m_repo.create_model(name=model_name, completed=True) checkpoint = m_repo.create_checkpoint(model_uuid=r_model.uuid) m_repo.set_final_model_checkpoint(r_model.uuid, checkpoint.uuid)
def hide_rename_model(model_name): model = m_repo.get_model(name=model_name, load_final_checkpoint=True, load_evaluations=True) for e in m_repo.get_evaluations([x.uuid for x in model.final_checkpoint.evaluations]): m_repo.hide_evaluation(e.uuid) new_name = (model_name + f'_hidden_{random.randint(0, 10000)}') m_repo.rename_model(model.uuid, new_name) m_repo.hide_model(model.uuid)
def create_eval_setting(eval_setting_name): m_repo.create_evaluation_setting(name=eval_setting_name)
def hide_rename_eval_setting(eval_setting_name): setting = m_repo.get_evaluation_setting(name=eval_setting_name, load_evaluations=True) for e in m_repo.get_evaluations([x.uuid for x in setting.evaluations]): m_repo.hide_evaluation(e.uuid) new_name = (eval_setting_name + f'_hidden_{random.randint(0, 10000)}') m_repo.rename_evaluation_setting(setting.uuid, new_name) m_repo.hide_evaluation_setting(setting.uuid)
def rename_model(model_name, new_model_name): uuid = m_repo.get_model(name=model_name).uuid m_repo.rename_model(uuid, new_model_name)
def rename_eval_setting(eval_setting_name, new_eval_setting_name): uuid = m_repo.get_evaluation_setting(name=eval_setting_name).uuid m_repo.rename_evaluation_setting(uuid, new_eval_setting_name)
def hide_evaluation(model_name, eval_setting_name): flag = False for e in m_repo.get_evaluations(): if ((e.checkpoint.model.name == model_name) and (e.setting.name == eval_setting_name)): m_repo.hide_evaluation(e.uuid) flag = True return flag
def get_eval_extra_info(model_name, eval_setting_name): for e in m_repo.get_evaluations(): if ((e.checkpoint.model.name == model_name) and (e.setting.name == eval_setting_name) and e.completed): return e.extra_info
def close_db_connection(): m_repo.dispose()
def classifier_loader(): model = torch_models.resnet50() load_model_state_dict(model, 'resnet50_adv-train-free') return model
def gen_classifier_loader(name, d): def classifier_loader(): model = getattr(models_lpf, d['arch'])(filter_size=d['filter_size']) load_model_state_dict(model, name) return model return classifier_loader
def classifier_loader(): model = torch_models.resnet50() load_model_state_dict(model, 'resnet50_augmix') return model
class StdConv2d(nn.Conv2d): def forward(self, x): w = self.weight (v, m) = torch.var_mean(w, dim=[1, 2, 3], keepdim=True, unbiased=False) w = ((w - m) / torch.sqrt((v + 1e-10))) return F.conv2d(x, w, self.bias, self.stride, self.padding, self.dilation, self.groups)
def conv3x3(cin, cout, stride=1, groups=1, bias=False): return StdConv2d(cin, cout, kernel_size=3, stride=stride, padding=1, bias=bias, groups=groups)
def conv1x1(cin, cout, stride=1, bias=False): return StdConv2d(cin, cout, kernel_size=1, stride=stride, padding=0, bias=bias)
def tf2th(conv_weights): 'Possibly convert HWIO to OIHW.' if (conv_weights.ndim == 4): conv_weights = conv_weights.transpose([3, 2, 0, 1]) return torch.from_numpy(conv_weights)
class PreActBottleneck(nn.Module): 'Pre-activation (v2) bottleneck block.\n Follows the implementation of "Identity Mappings in Deep Residual Networks":\n https://github.com/KaimingHe/resnet-1k-layers/blob/master/resnet-pre-act.lua\n Except it puts the stride on 3x3 conv when available.\n ' def __init__(self, cin, cout=None, cmid=None, stride=1): super().__init__() cout = (cout or cin) cmid = (cmid or (cout // 4)) self.gn1 = nn.GroupNorm(32, cin) self.conv1 = conv1x1(cin, cmid) self.gn2 = nn.GroupNorm(32, cmid) self.conv2 = conv3x3(cmid, cmid, stride) self.gn3 = nn.GroupNorm(32, cmid) self.conv3 = conv1x1(cmid, cout) self.relu = nn.ReLU(inplace=True) if ((stride != 1) or (cin != cout)): self.downsample = conv1x1(cin, cout, stride) def forward(self, x): out = self.relu(self.gn1(x)) residual = x if hasattr(self, 'downsample'): residual = self.downsample(out) out = self.conv1(out) out = self.conv2(self.relu(self.gn2(out))) out = self.conv3(self.relu(self.gn3(out))) return (out + residual) def load_from(self, weights, prefix=''): convname = 'standardized_conv2d' with torch.no_grad(): self.conv1.weight.copy_(tf2th(weights[f'{prefix}a/{convname}/kernel'])) self.conv2.weight.copy_(tf2th(weights[f'{prefix}b/{convname}/kernel'])) self.conv3.weight.copy_(tf2th(weights[f'{prefix}c/{convname}/kernel'])) self.gn1.weight.copy_(tf2th(weights[f'{prefix}a/group_norm/gamma'])) self.gn2.weight.copy_(tf2th(weights[f'{prefix}b/group_norm/gamma'])) self.gn3.weight.copy_(tf2th(weights[f'{prefix}c/group_norm/gamma'])) self.gn1.bias.copy_(tf2th(weights[f'{prefix}a/group_norm/beta'])) self.gn2.bias.copy_(tf2th(weights[f'{prefix}b/group_norm/beta'])) self.gn3.bias.copy_(tf2th(weights[f'{prefix}c/group_norm/beta'])) if hasattr(self, 'downsample'): w = weights[f'{prefix}a/proj/{convname}/kernel'] self.downsample.weight.copy_(tf2th(w))
class ResNetV2(nn.Module): 'Implementation of Pre-activation (v2) ResNet mode.' def __init__(self, block_units, width_factor, head_size=21843, zero_head=False): super().__init__() wf = width_factor self.root = nn.Sequential(OrderedDict([('conv', StdConv2d(3, (64 * wf), kernel_size=7, stride=2, padding=3, bias=False)), ('pad', nn.ConstantPad2d(1, 0)), ('pool', nn.MaxPool2d(kernel_size=3, stride=2, padding=0))])) self.body = nn.Sequential(OrderedDict([('block1', nn.Sequential(OrderedDict(([('unit01', PreActBottleneck(cin=(64 * wf), cout=(256 * wf), cmid=(64 * wf)))] + [(f'unit{i:02d}', PreActBottleneck(cin=(256 * wf), cout=(256 * wf), cmid=(64 * wf))) for i in range(2, (block_units[0] + 1))])))), ('block2', nn.Sequential(OrderedDict(([('unit01', PreActBottleneck(cin=(256 * wf), cout=(512 * wf), cmid=(128 * wf), stride=2))] + [(f'unit{i:02d}', PreActBottleneck(cin=(512 * wf), cout=(512 * wf), cmid=(128 * wf))) for i in range(2, (block_units[1] + 1))])))), ('block3', nn.Sequential(OrderedDict(([('unit01', PreActBottleneck(cin=(512 * wf), cout=(1024 * wf), cmid=(256 * wf), stride=2))] + [(f'unit{i:02d}', PreActBottleneck(cin=(1024 * wf), cout=(1024 * wf), cmid=(256 * wf))) for i in range(2, (block_units[2] + 1))])))), ('block4', nn.Sequential(OrderedDict(([('unit01', PreActBottleneck(cin=(1024 * wf), cout=(2048 * wf), cmid=(512 * wf), stride=2))] + [(f'unit{i:02d}', PreActBottleneck(cin=(2048 * wf), cout=(2048 * wf), cmid=(512 * wf))) for i in range(2, (block_units[3] + 1))]))))])) self.zero_head = zero_head self.head = nn.Sequential(OrderedDict([('gn', nn.GroupNorm(32, (2048 * wf))), ('relu', nn.ReLU(inplace=True)), ('avg', nn.AdaptiveAvgPool2d(output_size=1)), ('conv', nn.Conv2d((2048 * wf), head_size, kernel_size=1, bias=True))])) def forward(self, x): x = self.head(self.body(self.root(x))) assert (x.shape[(- 2):] == (1, 1)) return x[(..., 0, 0)] def load_from(self, weights, prefix='resnet/'): with torch.no_grad(): self.root.conv.weight.copy_(tf2th(weights[f'{prefix}root_block/standardized_conv2d/kernel'])) self.head.gn.weight.copy_(tf2th(weights[f'{prefix}group_norm/gamma'])) self.head.gn.bias.copy_(tf2th(weights[f'{prefix}group_norm/beta'])) if self.zero_head: nn.init.zeros_(self.head.conv.weight) nn.init.zeros_(self.head.conv.bias) else: self.head.conv.weight.copy_(tf2th(weights[f'{prefix}head/conv2d/kernel'])) self.head.conv.bias.copy_(tf2th(weights[f'{prefix}head/conv2d/bias'])) for (bname, block) in self.body.named_children(): for (uname, unit) in block.named_children(): unit.load_from(weights, prefix=f'{prefix}{bname}/{uname}/')
def gen_classifier_loader(name, d): def classifier_loader(): model = KNOWN_MODELS[d['arch']](head_size=(1000 if ('head_size' not in d) else d['head_size'])) load_model_state_dict(model, name) return model return classifier_loader
def load_weights(weight_file): if (weight_file == None): return try: weights_dict = np.load(weight_file, allow_pickle=True).item() except: weights_dict = np.load(weight_file, encoding='bytes').item() return weights_dict
class KitModel(nn.Module): def __init__(self, weight_file): super(KitModel, self).__init__() global __weights_dict __weights_dict = load_weights(weight_file) self.conv_conv1 = self.__conv(2, name='conv_conv1', in_channels=3, out_channels=96, kernel_size=(7, 7), stride=(2, 2), groups=1, bias=True) self.bn_conv1 = self.__batch_normalization(2, 'bn_conv1', num_features=96, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_conv2red = self.__conv(2, name='conv_conv2red', in_channels=96, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_conv2red = self.__batch_normalization(2, 'bn_conv2red', num_features=128, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_conv2 = self.__conv(2, name='conv_conv2', in_channels=128, out_channels=288, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_conv2 = self.__batch_normalization(2, 'bn_conv2', num_features=288, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_3a_1x1 = self.__conv(2, name='conv_3a_1x1', in_channels=288, out_channels=96, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_3a_3x3_reduce = self.__conv(2, name='conv_3a_3x3_reduce', in_channels=288, out_channels=96, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_3a_double_3x3_reduce = self.__conv(2, name='conv_3a_double_3x3_reduce', in_channels=288, out_channels=96, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_3a_1x1 = self.__batch_normalization(2, 'bn_3a_1x1', num_features=96, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_3a_3x3_reduce = self.__batch_normalization(2, 'bn_3a_3x3_reduce', num_features=96, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_3a_double_3x3_reduce = self.__batch_normalization(2, 'bn_3a_double_3x3_reduce', num_features=96, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_3a_proj = self.__conv(2, name='conv_3a_proj', in_channels=288, out_channels=48, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_3a_proj = self.__batch_normalization(2, 'bn_3a_proj', num_features=48, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_3a_3x3 = self.__conv(2, name='conv_3a_3x3', in_channels=96, out_channels=96, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.conv_3a_double_3x3_0 = self.__conv(2, name='conv_3a_double_3x3_0', in_channels=96, out_channels=144, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_3a_3x3 = self.__batch_normalization(2, 'bn_3a_3x3', num_features=96, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_3a_double_3x3_0 = self.__batch_normalization(2, 'bn_3a_double_3x3_0', num_features=144, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_3a_double_3x3_1 = self.__conv(2, name='conv_3a_double_3x3_1', in_channels=144, out_channels=144, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_3a_double_3x3_1 = self.__batch_normalization(2, 'bn_3a_double_3x3_1', num_features=144, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_3b_1x1 = self.__conv(2, name='conv_3b_1x1', in_channels=384, out_channels=96, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_3b_3x3_reduce = self.__conv(2, name='conv_3b_3x3_reduce', in_channels=384, out_channels=96, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_3b_double_3x3_reduce = self.__conv(2, name='conv_3b_double_3x3_reduce', in_channels=384, out_channels=96, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_3b_1x1 = self.__batch_normalization(2, 'bn_3b_1x1', num_features=96, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_3b_3x3_reduce = self.__batch_normalization(2, 'bn_3b_3x3_reduce', num_features=96, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_3b_double_3x3_reduce = self.__batch_normalization(2, 'bn_3b_double_3x3_reduce', num_features=96, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_3b_proj = self.__conv(2, name='conv_3b_proj', in_channels=384, out_channels=96, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_3b_proj = self.__batch_normalization(2, 'bn_3b_proj', num_features=96, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_3b_3x3 = self.__conv(2, name='conv_3b_3x3', in_channels=96, out_channels=144, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.conv_3b_double_3x3_0 = self.__conv(2, name='conv_3b_double_3x3_0', in_channels=96, out_channels=144, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_3b_3x3 = self.__batch_normalization(2, 'bn_3b_3x3', num_features=144, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_3b_double_3x3_0 = self.__batch_normalization(2, 'bn_3b_double_3x3_0', num_features=144, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_3b_double_3x3_1 = self.__conv(2, name='conv_3b_double_3x3_1', in_channels=144, out_channels=144, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_3b_double_3x3_1 = self.__batch_normalization(2, 'bn_3b_double_3x3_1', num_features=144, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_3c_3x3_reduce = self.__conv(2, name='conv_3c_3x3_reduce', in_channels=480, out_channels=192, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_3c_double_3x3_reduce = self.__conv(2, name='conv_3c_double_3x3_reduce', in_channels=480, out_channels=96, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_3c_3x3_reduce = self.__batch_normalization(2, 'bn_3c_3x3_reduce', num_features=192, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_3c_double_3x3_reduce = self.__batch_normalization(2, 'bn_3c_double_3x3_reduce', num_features=96, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_3c_3x3 = self.__conv(2, name='conv_3c_3x3', in_channels=192, out_channels=240, kernel_size=(3, 3), stride=(2, 2), groups=1, bias=True) self.conv_3c_double_3x3_0 = self.__conv(2, name='conv_3c_double_3x3_0', in_channels=96, out_channels=144, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_3c_3x3 = self.__batch_normalization(2, 'bn_3c_3x3', num_features=240, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_3c_double_3x3_0 = self.__batch_normalization(2, 'bn_3c_double_3x3_0', num_features=144, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_3c_double_3x3_1 = self.__conv(2, name='conv_3c_double_3x3_1', in_channels=144, out_channels=144, kernel_size=(3, 3), stride=(2, 2), groups=1, bias=True) self.bn_3c_double_3x3_1 = self.__batch_normalization(2, 'bn_3c_double_3x3_1', num_features=144, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4a_1x1 = self.__conv(2, name='conv_4a_1x1', in_channels=864, out_channels=224, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_4a_3x3_reduce = self.__conv(2, name='conv_4a_3x3_reduce', in_channels=864, out_channels=64, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_4a_double_3x3_reduce = self.__conv(2, name='conv_4a_double_3x3_reduce', in_channels=864, out_channels=96, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_4a_1x1 = self.__batch_normalization(2, 'bn_4a_1x1', num_features=224, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_4a_3x3_reduce = self.__batch_normalization(2, 'bn_4a_3x3_reduce', num_features=64, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_4a_double_3x3_reduce = self.__batch_normalization(2, 'bn_4a_double_3x3_reduce', num_features=96, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4a_proj = self.__conv(2, name='conv_4a_proj', in_channels=864, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_4a_proj = self.__batch_normalization(2, 'bn_4a_proj', num_features=128, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4a_3x3 = self.__conv(2, name='conv_4a_3x3', in_channels=64, out_channels=96, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.conv_4a_double_3x3_0 = self.__conv(2, name='conv_4a_double_3x3_0', in_channels=96, out_channels=128, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_4a_3x3 = self.__batch_normalization(2, 'bn_4a_3x3', num_features=96, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_4a_double_3x3_0 = self.__batch_normalization(2, 'bn_4a_double_3x3_0', num_features=128, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4a_double_3x3_1 = self.__conv(2, name='conv_4a_double_3x3_1', in_channels=128, out_channels=128, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_4a_double_3x3_1 = self.__batch_normalization(2, 'bn_4a_double_3x3_1', num_features=128, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4b_1x1 = self.__conv(2, name='conv_4b_1x1', in_channels=576, out_channels=192, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_4b_3x3_reduce = self.__conv(2, name='conv_4b_3x3_reduce', in_channels=576, out_channels=96, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_4b_double_3x3_reduce = self.__conv(2, name='conv_4b_double_3x3_reduce', in_channels=576, out_channels=96, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_4b_1x1 = self.__batch_normalization(2, 'bn_4b_1x1', num_features=192, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_4b_3x3_reduce = self.__batch_normalization(2, 'bn_4b_3x3_reduce', num_features=96, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_4b_double_3x3_reduce = self.__batch_normalization(2, 'bn_4b_double_3x3_reduce', num_features=96, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4b_proj = self.__conv(2, name='conv_4b_proj', in_channels=576, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_4b_proj = self.__batch_normalization(2, 'bn_4b_proj', num_features=128, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4b_3x3 = self.__conv(2, name='conv_4b_3x3', in_channels=96, out_channels=128, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.conv_4b_double_3x3_0 = self.__conv(2, name='conv_4b_double_3x3_0', in_channels=96, out_channels=128, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_4b_3x3 = self.__batch_normalization(2, 'bn_4b_3x3', num_features=128, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_4b_double_3x3_0 = self.__batch_normalization(2, 'bn_4b_double_3x3_0', num_features=128, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4b_double_3x3_1 = self.__conv(2, name='conv_4b_double_3x3_1', in_channels=128, out_channels=128, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_4b_double_3x3_1 = self.__batch_normalization(2, 'bn_4b_double_3x3_1', num_features=128, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4c_1x1 = self.__conv(2, name='conv_4c_1x1', in_channels=576, out_channels=160, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_4c_3x3_reduce = self.__conv(2, name='conv_4c_3x3_reduce', in_channels=576, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_4c_double_3x3_reduce = self.__conv(2, name='conv_4c_double_3x3_reduce', in_channels=576, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_4c_1x1 = self.__batch_normalization(2, 'bn_4c_1x1', num_features=160, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_4c_3x3_reduce = self.__batch_normalization(2, 'bn_4c_3x3_reduce', num_features=128, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_4c_double_3x3_reduce = self.__batch_normalization(2, 'bn_4c_double_3x3_reduce', num_features=128, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4c_proj = self.__conv(2, name='conv_4c_proj', in_channels=576, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_4c_proj = self.__batch_normalization(2, 'bn_4c_proj', num_features=128, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4c_3x3 = self.__conv(2, name='conv_4c_3x3', in_channels=128, out_channels=160, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.conv_4c_double_3x3_0 = self.__conv(2, name='conv_4c_double_3x3_0', in_channels=128, out_channels=160, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_4c_3x3 = self.__batch_normalization(2, 'bn_4c_3x3', num_features=160, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_4c_double_3x3_0 = self.__batch_normalization(2, 'bn_4c_double_3x3_0', num_features=160, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4c_double_3x3_1 = self.__conv(2, name='conv_4c_double_3x3_1', in_channels=160, out_channels=160, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_4c_double_3x3_1 = self.__batch_normalization(2, 'bn_4c_double_3x3_1', num_features=160, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4d_1x1 = self.__conv(2, name='conv_4d_1x1', in_channels=608, out_channels=96, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_4d_3x3_reduce = self.__conv(2, name='conv_4d_3x3_reduce', in_channels=608, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_4d_double_3x3_reduce = self.__conv(2, name='conv_4d_double_3x3_reduce', in_channels=608, out_channels=160, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_4d_1x1 = self.__batch_normalization(2, 'bn_4d_1x1', num_features=96, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_4d_3x3_reduce = self.__batch_normalization(2, 'bn_4d_3x3_reduce', num_features=128, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_4d_double_3x3_reduce = self.__batch_normalization(2, 'bn_4d_double_3x3_reduce', num_features=160, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4d_proj = self.__conv(2, name='conv_4d_proj', in_channels=608, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_4d_proj = self.__batch_normalization(2, 'bn_4d_proj', num_features=128, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4d_3x3 = self.__conv(2, name='conv_4d_3x3', in_channels=128, out_channels=192, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.conv_4d_double_3x3_0 = self.__conv(2, name='conv_4d_double_3x3_0', in_channels=160, out_channels=96, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_4d_3x3 = self.__batch_normalization(2, 'bn_4d_3x3', num_features=192, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_4d_double_3x3_0 = self.__batch_normalization(2, 'bn_4d_double_3x3_0', num_features=96, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4d_double_3x3_1 = self.__conv(2, name='conv_4d_double_3x3_1', in_channels=96, out_channels=96, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_4d_double_3x3_1 = self.__batch_normalization(2, 'bn_4d_double_3x3_1', num_features=96, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4e_3x3_reduce = self.__conv(2, name='conv_4e_3x3_reduce', in_channels=512, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_4e_double_3x3_reduce = self.__conv(2, name='conv_4e_double_3x3_reduce', in_channels=512, out_channels=192, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_4e_3x3_reduce = self.__batch_normalization(2, 'bn_4e_3x3_reduce', num_features=128, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_4e_double_3x3_reduce = self.__batch_normalization(2, 'bn_4e_double_3x3_reduce', num_features=192, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4e_3x3 = self.__conv(2, name='conv_4e_3x3', in_channels=128, out_channels=192, kernel_size=(3, 3), stride=(2, 2), groups=1, bias=True) self.conv_4e_double_3x3_0 = self.__conv(2, name='conv_4e_double_3x3_0', in_channels=192, out_channels=256, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_4e_3x3 = self.__batch_normalization(2, 'bn_4e_3x3', num_features=192, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_4e_double_3x3_0 = self.__batch_normalization(2, 'bn_4e_double_3x3_0', num_features=256, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_4e_double_3x3_1 = self.__conv(2, name='conv_4e_double_3x3_1', in_channels=256, out_channels=256, kernel_size=(3, 3), stride=(2, 2), groups=1, bias=True) self.bn_4e_double_3x3_1 = self.__batch_normalization(2, 'bn_4e_double_3x3_1', num_features=256, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_5a_1x1 = self.__conv(2, name='conv_5a_1x1', in_channels=960, out_channels=352, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_5a_3x3_reduce = self.__conv(2, name='conv_5a_3x3_reduce', in_channels=960, out_channels=192, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_5a_double_3x3_reduce = self.__conv(2, name='conv_5a_double_3x3_reduce', in_channels=960, out_channels=160, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_5a_1x1 = self.__batch_normalization(2, 'bn_5a_1x1', num_features=352, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_5a_3x3_reduce = self.__batch_normalization(2, 'bn_5a_3x3_reduce', num_features=192, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_5a_double_3x3_reduce = self.__batch_normalization(2, 'bn_5a_double_3x3_reduce', num_features=160, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_5a_proj = self.__conv(2, name='conv_5a_proj', in_channels=960, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_5a_proj = self.__batch_normalization(2, 'bn_5a_proj', num_features=128, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_5a_3x3 = self.__conv(2, name='conv_5a_3x3', in_channels=192, out_channels=320, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.conv_5a_double_3x3_0 = self.__conv(2, name='conv_5a_double_3x3_0', in_channels=160, out_channels=224, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_5a_3x3 = self.__batch_normalization(2, 'bn_5a_3x3', num_features=320, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_5a_double_3x3_0 = self.__batch_normalization(2, 'bn_5a_double_3x3_0', num_features=224, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_5a_double_3x3_1 = self.__conv(2, name='conv_5a_double_3x3_1', in_channels=224, out_channels=224, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_5a_double_3x3_1 = self.__batch_normalization(2, 'bn_5a_double_3x3_1', num_features=224, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_5b_1x1 = self.__conv(2, name='conv_5b_1x1', in_channels=1024, out_channels=352, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_5b_3x3_reduce = self.__conv(2, name='conv_5b_3x3_reduce', in_channels=1024, out_channels=192, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.conv_5b_double_3x3_reduce = self.__conv(2, name='conv_5b_double_3x3_reduce', in_channels=1024, out_channels=192, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_5b_1x1 = self.__batch_normalization(2, 'bn_5b_1x1', num_features=352, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_5b_3x3_reduce = self.__batch_normalization(2, 'bn_5b_3x3_reduce', num_features=192, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_5b_double_3x3_reduce = self.__batch_normalization(2, 'bn_5b_double_3x3_reduce', num_features=192, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_5b_proj = self.__conv(2, name='conv_5b_proj', in_channels=1024, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True) self.bn_5b_proj = self.__batch_normalization(2, 'bn_5b_proj', num_features=128, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_5b_3x3 = self.__conv(2, name='conv_5b_3x3', in_channels=192, out_channels=320, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.conv_5b_double_3x3_0 = self.__conv(2, name='conv_5b_double_3x3_0', in_channels=192, out_channels=224, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_5b_3x3 = self.__batch_normalization(2, 'bn_5b_3x3', num_features=320, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.bn_5b_double_3x3_0 = self.__batch_normalization(2, 'bn_5b_double_3x3_0', num_features=224, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.conv_5b_double_3x3_1 = self.__conv(2, name='conv_5b_double_3x3_1', in_channels=224, out_channels=224, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True) self.bn_5b_double_3x3_1 = self.__batch_normalization(2, 'bn_5b_double_3x3_1', num_features=224, eps=9.999999747378752e-05, momentum=0.8999999761581421) self.fc1 = self.__dense(name='fc1', in_features=1024, out_features=21841, bias=True) def forward(self, x): conv_conv1_pad = F.pad(x, (3, 3, 3, 3)) conv_conv1 = self.conv_conv1(conv_conv1_pad) bn_conv1 = self.bn_conv1(conv_conv1) relu_conv1 = F.relu(bn_conv1) pool1 = F.max_pool2d(relu_conv1, kernel_size=(3, 3), stride=(2, 2), padding=0, ceil_mode=False) conv_conv2red = self.conv_conv2red(pool1) bn_conv2red = self.bn_conv2red(conv_conv2red) relu_conv2red = F.relu(bn_conv2red) conv_conv2_pad = F.pad(relu_conv2red, (1, 1, 1, 1)) conv_conv2 = self.conv_conv2(conv_conv2_pad) bn_conv2 = self.bn_conv2(conv_conv2) relu_conv2 = F.relu(bn_conv2) pool2 = F.max_pool2d(relu_conv2, kernel_size=(3, 3), stride=(2, 2), padding=0, ceil_mode=False) conv_3a_1x1 = self.conv_3a_1x1(pool2) conv_3a_3x3_reduce = self.conv_3a_3x3_reduce(pool2) conv_3a_double_3x3_reduce = self.conv_3a_double_3x3_reduce(pool2) avg_pool_3a_pool = F.avg_pool2d(pool2, kernel_size=(3, 3), stride=(1, 1), padding=(1,), ceil_mode=False, count_include_pad=False) bn_3a_1x1 = self.bn_3a_1x1(conv_3a_1x1) bn_3a_3x3_reduce = self.bn_3a_3x3_reduce(conv_3a_3x3_reduce) bn_3a_double_3x3_reduce = self.bn_3a_double_3x3_reduce(conv_3a_double_3x3_reduce) conv_3a_proj = self.conv_3a_proj(avg_pool_3a_pool) relu_3a_1x1 = F.relu(bn_3a_1x1) relu_3a_3x3_reduce = F.relu(bn_3a_3x3_reduce) relu_3a_double_3x3_reduce = F.relu(bn_3a_double_3x3_reduce) bn_3a_proj = self.bn_3a_proj(conv_3a_proj) conv_3a_3x3_pad = F.pad(relu_3a_3x3_reduce, (1, 1, 1, 1)) conv_3a_3x3 = self.conv_3a_3x3(conv_3a_3x3_pad) conv_3a_double_3x3_0_pad = F.pad(relu_3a_double_3x3_reduce, (1, 1, 1, 1)) conv_3a_double_3x3_0 = self.conv_3a_double_3x3_0(conv_3a_double_3x3_0_pad) relu_3a_proj = F.relu(bn_3a_proj) bn_3a_3x3 = self.bn_3a_3x3(conv_3a_3x3) bn_3a_double_3x3_0 = self.bn_3a_double_3x3_0(conv_3a_double_3x3_0) relu_3a_3x3 = F.relu(bn_3a_3x3) relu_3a_double_3x3_0 = F.relu(bn_3a_double_3x3_0) conv_3a_double_3x3_1_pad = F.pad(relu_3a_double_3x3_0, (1, 1, 1, 1)) conv_3a_double_3x3_1 = self.conv_3a_double_3x3_1(conv_3a_double_3x3_1_pad) bn_3a_double_3x3_1 = self.bn_3a_double_3x3_1(conv_3a_double_3x3_1) relu_3a_double_3x3_1 = F.relu(bn_3a_double_3x3_1) ch_concat_3a_chconcat = torch.cat((relu_3a_1x1, relu_3a_3x3, relu_3a_double_3x3_1, relu_3a_proj), 1) conv_3b_1x1 = self.conv_3b_1x1(ch_concat_3a_chconcat) conv_3b_3x3_reduce = self.conv_3b_3x3_reduce(ch_concat_3a_chconcat) conv_3b_double_3x3_reduce = self.conv_3b_double_3x3_reduce(ch_concat_3a_chconcat) avg_pool_3b_pool = F.avg_pool2d(ch_concat_3a_chconcat, kernel_size=(3, 3), stride=(1, 1), padding=(1,), ceil_mode=False, count_include_pad=False) bn_3b_1x1 = self.bn_3b_1x1(conv_3b_1x1) bn_3b_3x3_reduce = self.bn_3b_3x3_reduce(conv_3b_3x3_reduce) bn_3b_double_3x3_reduce = self.bn_3b_double_3x3_reduce(conv_3b_double_3x3_reduce) conv_3b_proj = self.conv_3b_proj(avg_pool_3b_pool) relu_3b_1x1 = F.relu(bn_3b_1x1) relu_3b_3x3_reduce = F.relu(bn_3b_3x3_reduce) relu_3b_double_3x3_reduce = F.relu(bn_3b_double_3x3_reduce) bn_3b_proj = self.bn_3b_proj(conv_3b_proj) conv_3b_3x3_pad = F.pad(relu_3b_3x3_reduce, (1, 1, 1, 1)) conv_3b_3x3 = self.conv_3b_3x3(conv_3b_3x3_pad) conv_3b_double_3x3_0_pad = F.pad(relu_3b_double_3x3_reduce, (1, 1, 1, 1)) conv_3b_double_3x3_0 = self.conv_3b_double_3x3_0(conv_3b_double_3x3_0_pad) relu_3b_proj = F.relu(bn_3b_proj) bn_3b_3x3 = self.bn_3b_3x3(conv_3b_3x3) bn_3b_double_3x3_0 = self.bn_3b_double_3x3_0(conv_3b_double_3x3_0) relu_3b_3x3 = F.relu(bn_3b_3x3) relu_3b_double_3x3_0 = F.relu(bn_3b_double_3x3_0) conv_3b_double_3x3_1_pad = F.pad(relu_3b_double_3x3_0, (1, 1, 1, 1)) conv_3b_double_3x3_1 = self.conv_3b_double_3x3_1(conv_3b_double_3x3_1_pad) bn_3b_double_3x3_1 = self.bn_3b_double_3x3_1(conv_3b_double_3x3_1) relu_3b_double_3x3_1 = F.relu(bn_3b_double_3x3_1) ch_concat_3b_chconcat = torch.cat((relu_3b_1x1, relu_3b_3x3, relu_3b_double_3x3_1, relu_3b_proj), 1) conv_3c_3x3_reduce = self.conv_3c_3x3_reduce(ch_concat_3b_chconcat) conv_3c_double_3x3_reduce = self.conv_3c_double_3x3_reduce(ch_concat_3b_chconcat) max_pool_3c_pool_pad = F.pad(ch_concat_3b_chconcat, (1, 1, 1, 1), value=float('-inf')) max_pool_3c_pool = F.max_pool2d(max_pool_3c_pool_pad, kernel_size=(3, 3), stride=(2, 2), padding=0, ceil_mode=False) bn_3c_3x3_reduce = self.bn_3c_3x3_reduce(conv_3c_3x3_reduce) bn_3c_double_3x3_reduce = self.bn_3c_double_3x3_reduce(conv_3c_double_3x3_reduce) relu_3c_3x3_reduce = F.relu(bn_3c_3x3_reduce) relu_3c_double_3x3_reduce = F.relu(bn_3c_double_3x3_reduce) conv_3c_3x3_pad = F.pad(relu_3c_3x3_reduce, (1, 1, 1, 1)) conv_3c_3x3 = self.conv_3c_3x3(conv_3c_3x3_pad) conv_3c_double_3x3_0_pad = F.pad(relu_3c_double_3x3_reduce, (1, 1, 1, 1)) conv_3c_double_3x3_0 = self.conv_3c_double_3x3_0(conv_3c_double_3x3_0_pad) bn_3c_3x3 = self.bn_3c_3x3(conv_3c_3x3) bn_3c_double_3x3_0 = self.bn_3c_double_3x3_0(conv_3c_double_3x3_0) relu_3c_3x3 = F.relu(bn_3c_3x3) relu_3c_double_3x3_0 = F.relu(bn_3c_double_3x3_0) conv_3c_double_3x3_1_pad = F.pad(relu_3c_double_3x3_0, (1, 1, 1, 1)) conv_3c_double_3x3_1 = self.conv_3c_double_3x3_1(conv_3c_double_3x3_1_pad) bn_3c_double_3x3_1 = self.bn_3c_double_3x3_1(conv_3c_double_3x3_1) relu_3c_double_3x3_1 = F.relu(bn_3c_double_3x3_1) ch_concat_3c_chconcat = torch.cat((relu_3c_3x3, relu_3c_double_3x3_1, max_pool_3c_pool), 1) conv_4a_1x1 = self.conv_4a_1x1(ch_concat_3c_chconcat) conv_4a_3x3_reduce = self.conv_4a_3x3_reduce(ch_concat_3c_chconcat) conv_4a_double_3x3_reduce = self.conv_4a_double_3x3_reduce(ch_concat_3c_chconcat) avg_pool_4a_pool = F.avg_pool2d(ch_concat_3c_chconcat, kernel_size=(3, 3), stride=(1, 1), padding=(1,), ceil_mode=False, count_include_pad=False) bn_4a_1x1 = self.bn_4a_1x1(conv_4a_1x1) bn_4a_3x3_reduce = self.bn_4a_3x3_reduce(conv_4a_3x3_reduce) bn_4a_double_3x3_reduce = self.bn_4a_double_3x3_reduce(conv_4a_double_3x3_reduce) conv_4a_proj = self.conv_4a_proj(avg_pool_4a_pool) relu_4a_1x1 = F.relu(bn_4a_1x1) relu_4a_3x3_reduce = F.relu(bn_4a_3x3_reduce) relu_4a_double_3x3_reduce = F.relu(bn_4a_double_3x3_reduce) bn_4a_proj = self.bn_4a_proj(conv_4a_proj) conv_4a_3x3_pad = F.pad(relu_4a_3x3_reduce, (1, 1, 1, 1)) conv_4a_3x3 = self.conv_4a_3x3(conv_4a_3x3_pad) conv_4a_double_3x3_0_pad = F.pad(relu_4a_double_3x3_reduce, (1, 1, 1, 1)) conv_4a_double_3x3_0 = self.conv_4a_double_3x3_0(conv_4a_double_3x3_0_pad) relu_4a_proj = F.relu(bn_4a_proj) bn_4a_3x3 = self.bn_4a_3x3(conv_4a_3x3) bn_4a_double_3x3_0 = self.bn_4a_double_3x3_0(conv_4a_double_3x3_0) relu_4a_3x3 = F.relu(bn_4a_3x3) relu_4a_double_3x3_0 = F.relu(bn_4a_double_3x3_0) conv_4a_double_3x3_1_pad = F.pad(relu_4a_double_3x3_0, (1, 1, 1, 1)) conv_4a_double_3x3_1 = self.conv_4a_double_3x3_1(conv_4a_double_3x3_1_pad) bn_4a_double_3x3_1 = self.bn_4a_double_3x3_1(conv_4a_double_3x3_1) relu_4a_double_3x3_1 = F.relu(bn_4a_double_3x3_1) ch_concat_4a_chconcat = torch.cat((relu_4a_1x1, relu_4a_3x3, relu_4a_double_3x3_1, relu_4a_proj), 1) conv_4b_1x1 = self.conv_4b_1x1(ch_concat_4a_chconcat) conv_4b_3x3_reduce = self.conv_4b_3x3_reduce(ch_concat_4a_chconcat) conv_4b_double_3x3_reduce = self.conv_4b_double_3x3_reduce(ch_concat_4a_chconcat) avg_pool_4b_pool = F.avg_pool2d(ch_concat_4a_chconcat, kernel_size=(3, 3), stride=(1, 1), padding=(1,), ceil_mode=False, count_include_pad=False) bn_4b_1x1 = self.bn_4b_1x1(conv_4b_1x1) bn_4b_3x3_reduce = self.bn_4b_3x3_reduce(conv_4b_3x3_reduce) bn_4b_double_3x3_reduce = self.bn_4b_double_3x3_reduce(conv_4b_double_3x3_reduce) conv_4b_proj = self.conv_4b_proj(avg_pool_4b_pool) relu_4b_1x1 = F.relu(bn_4b_1x1) relu_4b_3x3_reduce = F.relu(bn_4b_3x3_reduce) relu_4b_double_3x3_reduce = F.relu(bn_4b_double_3x3_reduce) bn_4b_proj = self.bn_4b_proj(conv_4b_proj) conv_4b_3x3_pad = F.pad(relu_4b_3x3_reduce, (1, 1, 1, 1)) conv_4b_3x3 = self.conv_4b_3x3(conv_4b_3x3_pad) conv_4b_double_3x3_0_pad = F.pad(relu_4b_double_3x3_reduce, (1, 1, 1, 1)) conv_4b_double_3x3_0 = self.conv_4b_double_3x3_0(conv_4b_double_3x3_0_pad) relu_4b_proj = F.relu(bn_4b_proj) bn_4b_3x3 = self.bn_4b_3x3(conv_4b_3x3) bn_4b_double_3x3_0 = self.bn_4b_double_3x3_0(conv_4b_double_3x3_0) relu_4b_3x3 = F.relu(bn_4b_3x3) relu_4b_double_3x3_0 = F.relu(bn_4b_double_3x3_0) conv_4b_double_3x3_1_pad = F.pad(relu_4b_double_3x3_0, (1, 1, 1, 1)) conv_4b_double_3x3_1 = self.conv_4b_double_3x3_1(conv_4b_double_3x3_1_pad) bn_4b_double_3x3_1 = self.bn_4b_double_3x3_1(conv_4b_double_3x3_1) relu_4b_double_3x3_1 = F.relu(bn_4b_double_3x3_1) ch_concat_4b_chconcat = torch.cat((relu_4b_1x1, relu_4b_3x3, relu_4b_double_3x3_1, relu_4b_proj), 1) conv_4c_1x1 = self.conv_4c_1x1(ch_concat_4b_chconcat) conv_4c_3x3_reduce = self.conv_4c_3x3_reduce(ch_concat_4b_chconcat) conv_4c_double_3x3_reduce = self.conv_4c_double_3x3_reduce(ch_concat_4b_chconcat) avg_pool_4c_pool = F.avg_pool2d(ch_concat_4b_chconcat, kernel_size=(3, 3), stride=(1, 1), padding=(1,), ceil_mode=False, count_include_pad=False) bn_4c_1x1 = self.bn_4c_1x1(conv_4c_1x1) bn_4c_3x3_reduce = self.bn_4c_3x3_reduce(conv_4c_3x3_reduce) bn_4c_double_3x3_reduce = self.bn_4c_double_3x3_reduce(conv_4c_double_3x3_reduce) conv_4c_proj = self.conv_4c_proj(avg_pool_4c_pool) relu_4c_1x1 = F.relu(bn_4c_1x1) relu_4c_3x3_reduce = F.relu(bn_4c_3x3_reduce) relu_4c_double_3x3_reduce = F.relu(bn_4c_double_3x3_reduce) bn_4c_proj = self.bn_4c_proj(conv_4c_proj) conv_4c_3x3_pad = F.pad(relu_4c_3x3_reduce, (1, 1, 1, 1)) conv_4c_3x3 = self.conv_4c_3x3(conv_4c_3x3_pad) conv_4c_double_3x3_0_pad = F.pad(relu_4c_double_3x3_reduce, (1, 1, 1, 1)) conv_4c_double_3x3_0 = self.conv_4c_double_3x3_0(conv_4c_double_3x3_0_pad) relu_4c_proj = F.relu(bn_4c_proj) bn_4c_3x3 = self.bn_4c_3x3(conv_4c_3x3) bn_4c_double_3x3_0 = self.bn_4c_double_3x3_0(conv_4c_double_3x3_0) relu_4c_3x3 = F.relu(bn_4c_3x3) relu_4c_double_3x3_0 = F.relu(bn_4c_double_3x3_0) conv_4c_double_3x3_1_pad = F.pad(relu_4c_double_3x3_0, (1, 1, 1, 1)) conv_4c_double_3x3_1 = self.conv_4c_double_3x3_1(conv_4c_double_3x3_1_pad) bn_4c_double_3x3_1 = self.bn_4c_double_3x3_1(conv_4c_double_3x3_1) relu_4c_double_3x3_1 = F.relu(bn_4c_double_3x3_1) ch_concat_4c_chconcat = torch.cat((relu_4c_1x1, relu_4c_3x3, relu_4c_double_3x3_1, relu_4c_proj), 1) conv_4d_1x1 = self.conv_4d_1x1(ch_concat_4c_chconcat) conv_4d_3x3_reduce = self.conv_4d_3x3_reduce(ch_concat_4c_chconcat) conv_4d_double_3x3_reduce = self.conv_4d_double_3x3_reduce(ch_concat_4c_chconcat) avg_pool_4d_pool = F.avg_pool2d(ch_concat_4c_chconcat, kernel_size=(3, 3), stride=(1, 1), padding=(1,), ceil_mode=False, count_include_pad=False) bn_4d_1x1 = self.bn_4d_1x1(conv_4d_1x1) bn_4d_3x3_reduce = self.bn_4d_3x3_reduce(conv_4d_3x3_reduce) bn_4d_double_3x3_reduce = self.bn_4d_double_3x3_reduce(conv_4d_double_3x3_reduce) conv_4d_proj = self.conv_4d_proj(avg_pool_4d_pool) relu_4d_1x1 = F.relu(bn_4d_1x1) relu_4d_3x3_reduce = F.relu(bn_4d_3x3_reduce) relu_4d_double_3x3_reduce = F.relu(bn_4d_double_3x3_reduce) bn_4d_proj = self.bn_4d_proj(conv_4d_proj) conv_4d_3x3_pad = F.pad(relu_4d_3x3_reduce, (1, 1, 1, 1)) conv_4d_3x3 = self.conv_4d_3x3(conv_4d_3x3_pad) conv_4d_double_3x3_0_pad = F.pad(relu_4d_double_3x3_reduce, (1, 1, 1, 1)) conv_4d_double_3x3_0 = self.conv_4d_double_3x3_0(conv_4d_double_3x3_0_pad) relu_4d_proj = F.relu(bn_4d_proj) bn_4d_3x3 = self.bn_4d_3x3(conv_4d_3x3) bn_4d_double_3x3_0 = self.bn_4d_double_3x3_0(conv_4d_double_3x3_0) relu_4d_3x3 = F.relu(bn_4d_3x3) relu_4d_double_3x3_0 = F.relu(bn_4d_double_3x3_0) conv_4d_double_3x3_1_pad = F.pad(relu_4d_double_3x3_0, (1, 1, 1, 1)) conv_4d_double_3x3_1 = self.conv_4d_double_3x3_1(conv_4d_double_3x3_1_pad) bn_4d_double_3x3_1 = self.bn_4d_double_3x3_1(conv_4d_double_3x3_1) relu_4d_double_3x3_1 = F.relu(bn_4d_double_3x3_1) ch_concat_4d_chconcat = torch.cat((relu_4d_1x1, relu_4d_3x3, relu_4d_double_3x3_1, relu_4d_proj), 1) conv_4e_3x3_reduce = self.conv_4e_3x3_reduce(ch_concat_4d_chconcat) conv_4e_double_3x3_reduce = self.conv_4e_double_3x3_reduce(ch_concat_4d_chconcat) max_pool_4e_pool_pad = F.pad(ch_concat_4d_chconcat, (1, 1, 1, 1), value=float('-inf')) max_pool_4e_pool = F.max_pool2d(max_pool_4e_pool_pad, kernel_size=(3, 3), stride=(2, 2), padding=0, ceil_mode=False) bn_4e_3x3_reduce = self.bn_4e_3x3_reduce(conv_4e_3x3_reduce) bn_4e_double_3x3_reduce = self.bn_4e_double_3x3_reduce(conv_4e_double_3x3_reduce) relu_4e_3x3_reduce = F.relu(bn_4e_3x3_reduce) relu_4e_double_3x3_reduce = F.relu(bn_4e_double_3x3_reduce) conv_4e_3x3_pad = F.pad(relu_4e_3x3_reduce, (1, 1, 1, 1)) conv_4e_3x3 = self.conv_4e_3x3(conv_4e_3x3_pad) conv_4e_double_3x3_0_pad = F.pad(relu_4e_double_3x3_reduce, (1, 1, 1, 1)) conv_4e_double_3x3_0 = self.conv_4e_double_3x3_0(conv_4e_double_3x3_0_pad) bn_4e_3x3 = self.bn_4e_3x3(conv_4e_3x3) bn_4e_double_3x3_0 = self.bn_4e_double_3x3_0(conv_4e_double_3x3_0) relu_4e_3x3 = F.relu(bn_4e_3x3) relu_4e_double_3x3_0 = F.relu(bn_4e_double_3x3_0) conv_4e_double_3x3_1_pad = F.pad(relu_4e_double_3x3_0, (1, 1, 1, 1)) conv_4e_double_3x3_1 = self.conv_4e_double_3x3_1(conv_4e_double_3x3_1_pad) bn_4e_double_3x3_1 = self.bn_4e_double_3x3_1(conv_4e_double_3x3_1) relu_4e_double_3x3_1 = F.relu(bn_4e_double_3x3_1) ch_concat_4e_chconcat = torch.cat((relu_4e_3x3, relu_4e_double_3x3_1, max_pool_4e_pool), 1) conv_5a_1x1 = self.conv_5a_1x1(ch_concat_4e_chconcat) conv_5a_3x3_reduce = self.conv_5a_3x3_reduce(ch_concat_4e_chconcat) conv_5a_double_3x3_reduce = self.conv_5a_double_3x3_reduce(ch_concat_4e_chconcat) avg_pool_5a_pool = F.avg_pool2d(ch_concat_4e_chconcat, kernel_size=(3, 3), stride=(1, 1), padding=(1,), ceil_mode=False, count_include_pad=False) bn_5a_1x1 = self.bn_5a_1x1(conv_5a_1x1) bn_5a_3x3_reduce = self.bn_5a_3x3_reduce(conv_5a_3x3_reduce) bn_5a_double_3x3_reduce = self.bn_5a_double_3x3_reduce(conv_5a_double_3x3_reduce) conv_5a_proj = self.conv_5a_proj(avg_pool_5a_pool) relu_5a_1x1 = F.relu(bn_5a_1x1) relu_5a_3x3_reduce = F.relu(bn_5a_3x3_reduce) relu_5a_double_3x3_reduce = F.relu(bn_5a_double_3x3_reduce) bn_5a_proj = self.bn_5a_proj(conv_5a_proj) conv_5a_3x3_pad = F.pad(relu_5a_3x3_reduce, (1, 1, 1, 1)) conv_5a_3x3 = self.conv_5a_3x3(conv_5a_3x3_pad) conv_5a_double_3x3_0_pad = F.pad(relu_5a_double_3x3_reduce, (1, 1, 1, 1)) conv_5a_double_3x3_0 = self.conv_5a_double_3x3_0(conv_5a_double_3x3_0_pad) relu_5a_proj = F.relu(bn_5a_proj) bn_5a_3x3 = self.bn_5a_3x3(conv_5a_3x3) bn_5a_double_3x3_0 = self.bn_5a_double_3x3_0(conv_5a_double_3x3_0) relu_5a_3x3 = F.relu(bn_5a_3x3) relu_5a_double_3x3_0 = F.relu(bn_5a_double_3x3_0) conv_5a_double_3x3_1_pad = F.pad(relu_5a_double_3x3_0, (1, 1, 1, 1)) conv_5a_double_3x3_1 = self.conv_5a_double_3x3_1(conv_5a_double_3x3_1_pad) bn_5a_double_3x3_1 = self.bn_5a_double_3x3_1(conv_5a_double_3x3_1) relu_5a_double_3x3_1 = F.relu(bn_5a_double_3x3_1) ch_concat_5a_chconcat = torch.cat((relu_5a_1x1, relu_5a_3x3, relu_5a_double_3x3_1, relu_5a_proj), 1) conv_5b_1x1 = self.conv_5b_1x1(ch_concat_5a_chconcat) conv_5b_3x3_reduce = self.conv_5b_3x3_reduce(ch_concat_5a_chconcat) conv_5b_double_3x3_reduce = self.conv_5b_double_3x3_reduce(ch_concat_5a_chconcat) max_pool_5b_pool_pad = F.pad(ch_concat_5a_chconcat, (1, 1, 1, 1), value=float('-inf')) max_pool_5b_pool = F.max_pool2d(max_pool_5b_pool_pad, kernel_size=(3, 3), stride=(1, 1), padding=0, ceil_mode=False) bn_5b_1x1 = self.bn_5b_1x1(conv_5b_1x1) bn_5b_3x3_reduce = self.bn_5b_3x3_reduce(conv_5b_3x3_reduce) bn_5b_double_3x3_reduce = self.bn_5b_double_3x3_reduce(conv_5b_double_3x3_reduce) conv_5b_proj = self.conv_5b_proj(max_pool_5b_pool) relu_5b_1x1 = F.relu(bn_5b_1x1) relu_5b_3x3_reduce = F.relu(bn_5b_3x3_reduce) relu_5b_double_3x3_reduce = F.relu(bn_5b_double_3x3_reduce) bn_5b_proj = self.bn_5b_proj(conv_5b_proj) conv_5b_3x3_pad = F.pad(relu_5b_3x3_reduce, (1, 1, 1, 1)) conv_5b_3x3 = self.conv_5b_3x3(conv_5b_3x3_pad) conv_5b_double_3x3_0_pad = F.pad(relu_5b_double_3x3_reduce, (1, 1, 1, 1)) conv_5b_double_3x3_0 = self.conv_5b_double_3x3_0(conv_5b_double_3x3_0_pad) relu_5b_proj = F.relu(bn_5b_proj) bn_5b_3x3 = self.bn_5b_3x3(conv_5b_3x3) bn_5b_double_3x3_0 = self.bn_5b_double_3x3_0(conv_5b_double_3x3_0) relu_5b_3x3 = F.relu(bn_5b_3x3) relu_5b_double_3x3_0 = F.relu(bn_5b_double_3x3_0) conv_5b_double_3x3_1_pad = F.pad(relu_5b_double_3x3_0, (1, 1, 1, 1)) conv_5b_double_3x3_1 = self.conv_5b_double_3x3_1(conv_5b_double_3x3_1_pad) bn_5b_double_3x3_1 = self.bn_5b_double_3x3_1(conv_5b_double_3x3_1) relu_5b_double_3x3_1 = F.relu(bn_5b_double_3x3_1) ch_concat_5b_chconcat = torch.cat((relu_5b_1x1, relu_5b_3x3, relu_5b_double_3x3_1, relu_5b_proj), 1) global_pool = F.avg_pool2d(ch_concat_5b_chconcat, kernel_size=(7, 7), stride=(1, 1), padding=(0,), ceil_mode=False, count_include_pad=False) flatten = global_pool.view(global_pool.size(0), (- 1)) fc1 = self.fc1(flatten) softmax = F.softmax(fc1) return softmax @staticmethod def __conv(dim, name, **kwargs): if (dim == 1): layer = nn.Conv1d(**kwargs) elif (dim == 2): layer = nn.Conv2d(**kwargs) elif (dim == 3): layer = nn.Conv3d(**kwargs) else: raise NotImplementedError() layer.state_dict()['weight'].copy_(torch.from_numpy(__weights_dict[name]['weights'])) if ('bias' in __weights_dict[name]): layer.state_dict()['bias'].copy_(torch.from_numpy(__weights_dict[name]['bias'])) return layer @staticmethod def __batch_normalization(dim, name, **kwargs): if ((dim == 0) or (dim == 1)): layer = nn.BatchNorm1d(**kwargs) elif (dim == 2): layer = nn.BatchNorm2d(**kwargs) elif (dim == 3): layer = nn.BatchNorm3d(**kwargs) else: raise NotImplementedError() if ('scale' in __weights_dict[name]): layer.state_dict()['weight'].copy_(torch.from_numpy(__weights_dict[name]['scale'])) else: layer.weight.data.fill_(1) if ('bias' in __weights_dict[name]): layer.state_dict()['bias'].copy_(torch.from_numpy(__weights_dict[name]['bias'])) else: layer.bias.data.fill_(0) layer.state_dict()['running_mean'].copy_(torch.from_numpy(__weights_dict[name]['mean'])) layer.state_dict()['running_var'].copy_(torch.from_numpy(__weights_dict[name]['var'])) return layer @staticmethod def __dense(name, **kwargs): layer = nn.Linear(**kwargs) layer.state_dict()['weight'].copy_(torch.from_numpy(__weights_dict[name]['weights'])) if ('bias' in __weights_dict[name]): layer.state_dict()['bias'].copy_(torch.from_numpy(__weights_dict[name]['bias'])) return layer
def classifier_loader(): return KitModel(load_model_checkpoint_bytes('bninception-imagenet21k'))
def gen_classifier_loader(name, d): def classifier_loader(): model = pretrainedmodels.__dict__[d['arch']](num_classes=1000, pretrained=None) load_model_state_dict(model, name) return model return classifier_loader
def gen_classifier_loader(path): def classifier_loader(): model = build_clip_imagenet_model(path)[0] return model return classifier_loader
def gen_classifier_loader(name, d): def classifier_loader(): model = torch_models.__dict__[d['arch']]() load_model_state_dict(model, name) return model return classifier_loader
def gen_classifier_loader(name, d): def classifier_loader(): model = torch_models.__dict__[d['arch']]() load_model_state_dict(model, name) return model return classifier_loader
def gen_classifier_loader(name, d): def classifier_loader(): model = torch_models.__dict__[d['arch']]() load_model_state_dict(model, name) return model return classifier_loader
def gen_classifier_loader(name, d): def classifier_loader(): model = torch_models.__dict__[d['arch']](pretrained=False) load_model_state_dict(model, name) return model return classifier_loader
def gen_classifier_loader(name, d): def classifier_loader(): model = EfficientNet.from_name(d['arch']) load_model_state_dict(model, name) return model return classifier_loader
def noisystudent_loader(): model = timm.create_model('tf_efficientnet_l2_ns', pretrained=False) load_model_state_dict(model, 'efficientnet-l2-noisystudent') return model
def store_logits_jft(images, model): global store_counter assert (store_filename_prefix is not '') images = images.cpu().permute([0, 2, 3, 1]).numpy() with open(f'{store_filename_prefix}_{store_counter}.npy', 'wb') as f: np.save(f, images) store_counter += 1 return torch.empty(images.shape[0], 1000).cuda()
def load_logits_jft(images, model): global load_counter assert (load_filename_prefix is not '') with open(f'{load_filename_prefix}_{load_counter}.npy', 'rb') as f: logits = torch.from_numpy(np.load(f)).cuda() load_counter += 1 return logits
def gen_classifier_loader(name, d): def classifier_loader(): model = torch.hub.load('facebookresearch/WSL-Images', (d['arch'] + '_wsl')) load_model_state_dict(model, name) return model return classifier_loader
def gen_classifier_loader(name, d): def classifier_loader(): model = torch_models.__dict__[d['arch']]() load_model_state_dict(model, name) return model return classifier_loader
def gen_classifier_loader(name, d): def classifier_loader(): model = torch_models.__dict__[d['arch']]() load_model_state_dict(model, name) return model return classifier_loader
class ConvertToPyTorchModel(nn.Module): def __init__(self, base_model, classify_fn_args, classify=None, normalization=None, class_sublist=None, adversarial_attack=None): super().__init__() if (normalization is not None): self.input_space = normalization.input_space self.mean = nn.Parameter(torch.tensor(normalization.mean).float().view(3, 1, 1)) self.std = nn.Parameter(torch.tensor(normalization.std).float().view(3, 1, 1)) self.base_model = base_model self.classify_fn_args = classify_fn_args self.classify = classify self.class_sublist = class_sublist self.adversarial_attack = adversarial_attack self.normalization = normalization def forward(self, x): if (self.normalization is not None): if (self.input_space == 'BGR'): x = x.flip(1) x = ((x - self.mean) / self.std) if (self.classify is None): x = self.base_model(x) else: kwargs = {'images': x, 'model': self.base_model} if ('class_sublist' in self.classify_fn_args): kwargs['class_sublist'] = self.class_sublist if ('adversarial_attack' in self.classify_fn_args): kwargs['adversarial_attack'] = self.adversarial_attack x = self.classify(**kwargs) if ((self.class_sublist is not None) and ('class_sublist' not in self.classify_fn_args)): x = x.t()[self.class_sublist].t() return x
class Model(): ADVERSARIAL_BATCH_SIZE_REDUCTION_FACTOR = 8 def __init__(self, name, transform, classifier_loader, eval_batch_size, arch=None, normalization=None, classify=None, adversarial_batch_size=None): super().__init__() self.name = name self.arch = (arch if (arch is not None) else 'NA') self.transform = transform self.classifier_loader = classifier_loader self.eval_batch_size = eval_batch_size self.adversarial_batch_size = adversarial_batch_size self.normalization = normalization self.classify = classify self.classify_fn_args = set() if (self.classify is not None): sig = list(signature(self.classify).parameters.keys()) assert (('images' in sig) and ('model' in sig)), ('Unrecognized metrics function ' + 'definition. Make sure function takes arguments "images" and "model"') for arg in ['images', 'model', 'class_sublist', 'adversarial_attack']: if (arg in sig): self.classify_fn_args.add(arg) def generate_classifier(self, py_eval_setting): self.classifier = self.classifier_loader() model = ConvertToPyTorchModel(self.classifier, self.classify_fn_args, self.classify, self.normalization, py_eval_setting.class_sublist, py_eval_setting.adversarial_attack) if (len(list(model.parameters())) == 0): model._dummy = nn.Parameter(torch.tensor(0.0)) return model def get_batch_size(self, py_eval_setting): if (not py_eval_setting.adversarial_attack): return self.eval_batch_size elif (py_eval_setting.adversarial_attack and (self.adversarial_batch_size is not None)): return self.adversarial_batch_size return max((self.eval_batch_size // self.ADVERSARIAL_BATCH_SIZE_REDUCTION_FACTOR), 1)
class StandardNormalization(): def __init__(self, mean, std, input_space='RGB'): assert (input_space in ['RGB', 'BGR']), f'Can only handle RGB or BGR channel input formats' self.mean = mean self.std = std self.input_space = input_space