code
stringlengths
17
6.64M
class MultiCropTransform(): "Define multi crop transform that apply several sets of transform to the inputs.\n\n Args:\n set_transforms: List of Dictionary of sets of transforms specifying transforms and number of views per set.\n\n Example::\n\n set_transforms = [\n {'transform': [...], 'num_views': ...},\n {'transform': [...], 'num_views': ...},\n ...\n ]\n\n transform = MultiCropTransform(\n set_transforms\n )\n " def __init__(self, set_transforms: List[Any]) -> None: super().__init__() self.set_transforms = set_transforms transforms = [] for set_transform in self.set_transforms: if ('num_views' not in set_transform): set_transform['num_views'] = 1 transforms.extend(([set_transform['transform']] * set_transform['num_views'])) self.transforms = transforms def __call__(self, img: ((Image | Tensor) | Iterable[(Image | Tensor)])) -> Tensor: if (type(img) not in [Image, Tensor]): transformed_images = [transform(image) for (transform, image) in zip(self.transforms, img, strict=True)] else: transformed_images = [transform(img) for transform in self.transforms] return transformed_images def __repr__(self) -> str: format_string = self.__class__.__name__ for set_transform in self.set_transforms: format_string += '(\n' format_string += ' num views={}\n'.format(set_transform['num_views']) format_string += ' transforms={}'.format(set_transform['transform']) format_string += '\n)' return format_string
class OnlyInputListTransform(Compose): "Apply Transform to only the key ``'input'`` in a list of sample dictionary.\n\n Args:\n transform: The transform to apply.\n " def __init__(self, transform: Callable) -> None: transforms = [ApplyTransformInputKeyOnList(transform), DictKeepInputLabelIdx()] super().__init__(transforms=transforms)
class OnlyInputTransform(Compose): "Apply Transform to only the key ``'input'`` in a sample dictionary.\n\n Args:\n transform: The transform to apply.\n " def __init__(self, transform: Callable) -> None: transforms = [ApplyTransformInputKey(transform), DictKeepInputLabelIdx()] super().__init__(transforms=transforms)
class OnlyInputListSameTransform(Compose): "Apply the same transform to only the key ``'input'`` in a list of sample dictionary.\n\n Args:\n transform: The transform to apply.\n " def __init__(self, transform: Callable) -> None: transforms = [ApplySameTransformInputKeyOnList(transform), DictKeepInputLabelIdx()] super().__init__(transforms=transforms)
class OnlyInputTransformWithDictTransform(Compose): "Apply Transform to only the key ``'input'`` in a sample dictionary with a transformation on the dictionary\n afterwards.\n\n Args:\n transform: The transform to apply to the input.\n dict_transform: The transform to apply to the dictionary.\n first_dict: If ``True``, first apply the transformation on the dict, else, first apply the transformation on the input.\n " def __init__(self, transform: Callable, dict_transform: Callable, first_dict: bool=False) -> None: if first_dict: transforms = [ApplyTransformInputKey(transform), ApplyTransformOnDict(dict_transform), DictKeepInputLabelIdx()] else: transforms = [ApplyTransformInputKey(transform), ApplyTransformOnDict(dict_transform), DictKeepInputLabelIdx()] super().__init__(transforms=transforms)
class OnlyInputListTransformWithDictTransform(): "Apply Transform to only the key ``'input'`` in a list of sample dictionary with a transformation on the\n dictionary afterwards.\n\n Args:\n transform: The transform to apply to the input.\n dict_transform: The transform to apply to the dictionary.\n first_dict: If ``True``, first apply the transformation on the dict, else, first apply the transformation on the input.\n " def __init__(self, transform: Callable, dict_transform: Callable, first_dict: bool=False) -> None: self.transform = OnlyInputTransformWithDictTransform(transform, dict_transform, first_dict) def __call__(self, x: List[Dict[(str, Any)]]) -> List[Dict[(str, Any)]]: return [self.transform(sample) for sample in x] def __repr__(self) -> str: format_string = (self.__class__.__name__ + '(') format_string += '\n' format_string += f' {self.transform}' format_string += '\n)' return format_string
class RandomResizedCrop(transforms.RandomResizedCrop): def __init__(self, size: Union[(int, Iterable[int])], scale: Iterable[float]=[0.08, 1.0], ratio: Iterable[float]=[(3 / 4), (4 / 3)], interpolation: Union[(str, InterpolationMode)]='bilinear', antialias: bool=True, **kwargs) -> None: if (type(interpolation) is str): interpolation = _INTERPOLATION[interpolation] super().__init__(size, scale=scale, ratio=ratio, interpolation=interpolation, antialias=antialias, **kwargs)
class RemoveKey(Module): "Removes the given key from the input dict. Useful for removing modalities from a video clip that aren't\n needed.\n\n Args:\n key: The dictionary key to remove.\n " def __init__(self, key: str): super().__init__() self._key = key def __call__(self, x: Dict[(str, Tensor)]) -> Dict[(str, Tensor)]: if (self._key in x): del x[self._key] return x
class RemoveInputKey(RemoveKey): 'Remove video key from sample dictionary.' def __init__(self): super().__init__('input')
class RemoveAudioKey(RemoveKey): 'Remove audio key from sample dictionary.' def __init__(self): super().__init__('audio')
class RemoveTimeDim(nn.Module): 'Remove time dimension from tensor.\n\n Suppose the tensor shape is [C,T,H,W].\n ' def __init__(self) -> None: super().__init__() def forward(self, tensor: Tensor): (c, t, h, w) = tensor.shape return tensor.view((c * t), h, w) def __repr__(self): return f'{self.__class__.__name__}()'
def mix_spotting(x: Tensor, mix_value: Tensor, permutation: Tensor, labels: Tensor, has_label: Tensor, ignore_class: Tensor): 'Make mixup of the batch for action spotting.\n\n Args:\n x: The batch values to mix.\n mix_value: Value coefficients for mixing.\n permutation: Permutation to perform mix.\n labels: Labels of the timestamps in the batch.\n has_label: Whether timestamps have label.\n ignore_class: Whether class in the batch should be ignored.\n\n Returns:\n Tuple containing:\n - The mixed input.\n - The mixed class labels.\n - The `ignore_class` of the mixed elements.\n - The concatenated `mix_value` of the mixed elements.\n ' x_permuted = x[permutation] labels_permuted = labels[permutation] has_label_permuted = has_label[permutation] ignore_class_permuted = ignore_class[permutation] labels_cat = torch.cat((labels, labels_permuted)) has_label_cat = torch.cat((has_label, has_label_permuted)) ignore_class_cat = torch.cat((ignore_class, ignore_class_permuted)) mix_value_x = mix_value.view([(- 1), *([1] * (x.ndim - 1))]) one_minus_mix_value_x = (1 - mix_value_x) mix_value_label = mix_value.view([(- 1), *([1] * (labels.ndim - 1))]) one_minus_mix_value_label = (1 - mix_value_label) x_mixed = ((mix_value_x * x) + (one_minus_mix_value_x * x_permuted)) mixed_weights = torch.cat((mix_value_label, one_minus_mix_value_label)) return (x_mixed, labels_cat, has_label_cat, ignore_class_cat, mixed_weights)
class SpottingMixup(Module): 'Make mixup for spotting for labels.\n\n Args:\n alpha: Alpha value for the beta distribution of mixup.\n ' def __init__(self, alpha: float=0.5) -> None: super().__init__() self.alpha = alpha self.mix_sampler = torch.distributions.Beta(alpha, alpha) def forward(self, batch: Dict[(str, Any)]): (x, labels, has_label, ignore_class) = (batch['input'], batch['labels'], batch['has_label'], batch['ignore_class']) (device, dtype) = (x.device, x.dtype) batch_size = x.shape[0] with torch.inference_mode(): mix_value = self.mix_sampler.sample((batch_size,)).to(device=device, dtype=dtype) mix_value = mix_value.clone() permutation = torch.randperm(batch_size, device=device) (x_mixed, labels_after_mix, has_label_after_mix, ignore_class_after_mix, mixed_weights) = mix_spotting(x=x, mix_value=mix_value, permutation=permutation, labels=labels, has_label=has_label, ignore_class=ignore_class) new_batch = {'input': x_mixed, 'labels': labels_after_mix, 'ignore_class': ignore_class_after_mix, 'has_label': has_label_after_mix, 'mixup_weights': mixed_weights} return new_batch def __repr__(self): return f'{__class__.__name__}(alpha={self.alpha})'
def get_matching_files_in_dir(dir: str, file_pattern: str) -> List[Path]: 'Retrieve files in directory matching a pattern.\n\n Args:\n dir: Directory path.\n file_pattern: Pattern for the files.\n\n Raises:\n NotADirectoryError: If `dir` does not exist or is not a directory.\n\n Returns:\n List of files matching the pattern\n ' dir = Path(dir) if (dir.exists() and dir.is_dir()): files = list(dir.glob(file_pattern)) return files else: raise NotADirectoryError(f'Directory "{dir}" does not exist or is not a directory')
def get_ckpts_in_dir(dir: str, ckpt_pattern: str='*.ckpt') -> List[Path]: 'Get all checkpoints in a directory.\n\n Args:\n dir: Directory path containing the checkpoints.\n ckpt_pattern: Checkpoint glob pattern.\n\n Returns:\n List of checkpoints paths in directory.\n ' try: files = get_matching_files_in_dir(dir, ckpt_pattern) except NotADirectoryError: warnings.warn(f'No checkpoint found in: {dir}', category=RuntimeWarning) files = [] return files
def get_last_ckpt_in_dir(dir: str, ckpt_pattern: str='*.ckpt', key_sort: Callable=(lambda x: x.stat().st_mtime)) -> Optional[Path]: 'Get last ckpt in directory following a sorting function.\n\n Args:\n dir: Directory path containing the checkpoints.\n ckpt_pattern: Checkpoint glob pattern.\n key_sort: Function to sort the checkpoints.\n\n Returns:\n Last checkpoint in `dir`, if it exists, according to `key_sort`.\n ' ckpts = get_ckpts_in_dir(dir, ckpt_pattern) if (ckpts == []): return None ckpts.sort(key=key_sort, reverse=False) return ckpts[(- 1)]
def get_last_ckpt_in_path_or_dir(checkpoint_file: Optional[str]=None, checkpoint_dir: Optional[str]=None, ckpt_pattern: str='*.ckpt', key_sort: Callable=(lambda x: x.stat().st_mtime)) -> Optional[Path]: 'Get checkpoint from file or from last checkpoint in directory following a sorting function.\n\n Args:\n checkpoint_file: Checkpoint file path containing the checkpoint.\n checkpoint_dir: Directory path containing the checkpoints.\n ckpt_pattern: Checkpoint glob pattern.\n key_sort: Function to sort the checkpoints.\n\n Returns:\n Checkpoint file if it exists or last checkpoint in `dir` according to `key_sort`.\n ' if (checkpoint_file is not None): checkpoint_file_path = Path(checkpoint_file) if (checkpoint_file_path.exists() and checkpoint_file_path.is_file()): return checkpoint_file_path else: warnings.warn(f'{checkpoint_file} is not a file or do not exist.', category=RuntimeWarning) if (checkpoint_dir is not None): return get_last_ckpt_in_dir(checkpoint_dir, ckpt_pattern=ckpt_pattern, key_sort=key_sort) return None
def get_ckpt_by_callback_mode(checkpoint_path: str, checkpoint_mode: str) -> List[Path]: "Get checkpoint from ModelCheckpoint callback based on the mode: ``'best'``, ``'last'``, or ``'both'``.\n\n Args:\n checkpoint_path: Checkpoint file path containing the callback checkpoint.\n checkpoint_mode: Mode to read the callback checkpoint. Can be either ``'best'``, ``'last'`` or ``'both'``.\n\n Returns:\n Checkpoint paths based on the mode.\n " checkpoint = torch.load(checkpoint_path, map_location='cpu') model_checkpoint_str = str(checkpoint['callbacks']) paths: List[Path] = [] if ((checkpoint_mode == 'best') or (checkpoint_mode == 'both')): regex = "'best_model_path':\\s'([a-zA-Z/0-9=_\\-\\.]+\\.ckpt)'" paths.append(Path(re.search(regex, model_checkpoint_str).group(1))) elif ((checkpoint_mode == 'last') or (checkpoint_mode == 'both')): regex = "'last_model_path':\\s'([a-zA-Z/0-9=_\\-\\.]+\\.ckpt)'" paths.append(Path(re.search(regex, model_checkpoint_str).group(1))) else: raise NotImplementedError(f"Checkpoint mode '{checkpoint_mode}' not supported.") new_paths = [] checkpoint_dir = Path(checkpoint_path).parent for path in paths: if path.exists(): new_paths.append(path) else: new_path = (checkpoint_dir / path.name) assert new_path.exists(), f'The checkpoint {path} is not available and not found at {new_path}.' new_paths.append(new_path) return new_paths
def get_sub_state_dict_from_pl_ckpt(checkpoint_path: str, pattern: str='^(trunk\\.)') -> Dict[(Any, Any)]: 'Retrieve sub state dict from a pytorch lightning checkpoint.\n\n Args:\n checkpoint_path: Pytorch lightning checkpoint path.\n pattern: Pattern to filter the keys for the sub state dictionary.\n If value is ``""`` keep all keys.\n\n Returns:\n Sub state dict from the checkpoint following the pattern.\n ' model = torch.load(checkpoint_path) if ('state_dict' in model): state_dict = {k: v for (k, v) in model['state_dict'].items() if ((pattern == '') or re.match(pattern, k))} else: state_dict = {k: v for (k, v) in model.items() if ((pattern == '') or re.match(pattern, k))} return state_dict
def remove_pattern_in_keys_from_dict(d: Dict[(Any, Any)], pattern: str) -> Dict[(Any, Any)]: 'Remove the pattern from keys in a dictionary.\n\n Args:\n d: The dictionary.\n pattern: Pattern to remove from the keys.\n If value is ``""`` keep all keys.\n\n Returns:\n Input dictionary with updated keys.\n ' if (pattern == ''): return d return {re.sub(pattern, '', k): v for (k, v) in d.items()}
def mask_tube_in_sequence(mask_ratio: float, tube_size: int, len_sequence: int, device: (str | torch.device)='cpu'): 'Generate indices to mask tubes from a sequence.\n\n Args:\n mask_ratio: Ratio for the masking.\n tube_size: Tube size for the masking.\n len_sequence (int): Length of the sequence to mask.\n device: Device for the mask.\n\n Returns:\n Tuple:\n - The indices to mask.\n - The indices to keep.\n - The reversed order for temporal masking.\n - The number of tokens masked.\n ' num_masked = floor((len_sequence * mask_ratio)) indices_permuted = ((torch.randperm((len_sequence // tube_size), device=device) * tube_size).repeat_interleave(tube_size) + torch.arange(tube_size, device=device).repeat((len_sequence // tube_size))) indices_not_kept: torch.Tensor = indices_permuted[:num_masked].sort()[0] indices_kept: torch.Tensor = indices_permuted[num_masked:].sort()[0] indices = torch.cat((indices_not_kept, indices_kept)) inversed_temporal_masked_indices = torch.argsort(indices) return (indices_not_kept, indices_kept, inversed_temporal_masked_indices, num_masked)
def batch_mask_tube_in_sequence(mask_ratio: float, tube_size: int, len_sequence: int, batch_size: int, device: (str | torch.device)='cpu'): 'Generate indices to mask tubes from a batch of sequences.\n\n Args:\n mask_ratio: Ratio for the masking.\n tube_size: Tube size for the masking.\n len_sequence: Length of the sequence to mask.\n batch_size: The size of the batch.\n device: Device for the mask.\n\n Returns:\n Tuple:\n - The indices to mask.\n - The indices to keep.\n - The reversed order for temporal masking.\n - The number of tokens masked.\n ' tot_indices_not_kept = [None for i in range(batch_size)] tot_indices_kept = [None for i in range(batch_size)] tot_inversed_temporal_masked_indices = [None for i in range(batch_size)] tot_num_masked = 0 expected_num_masked = floor((mask_ratio * len_sequence)) tot_indices_not_kept = torch.empty((batch_size, expected_num_masked), device=device, dtype=torch.long) tot_indices_kept = torch.empty((batch_size, (len_sequence - expected_num_masked)), device=device, dtype=torch.long) tot_inversed_temporal_masked_indices = torch.empty((batch_size, len_sequence), device=device, dtype=torch.long) for i in range(batch_size): (indices_not_kept, indices_kept, inversed_temporal_masked_indices, num_masked) = mask_tube_in_sequence(mask_ratio, tube_size, len_sequence, device) tot_indices_not_kept[i] = indices_not_kept tot_indices_kept[i] = indices_kept tot_inversed_temporal_masked_indices[i] = inversed_temporal_masked_indices tot_num_masked += num_masked return (tot_indices_not_kept, tot_indices_kept, tot_inversed_temporal_masked_indices, tot_num_masked)
def get_global_batch_size_in_trainer(local_batch_size: int, trainer: Trainer) -> int: 'Get global batch size used by a trainer based on the local batch size.\n\n Args:\n local_batch_size: The local batch size used by the trainer.\n trainer: The trainer used.\n\n Raises:\n AttributeError: The strategy is not supported.\n\n Returns:\n The global batch size.\n ' strategy = get_trainer_strategy(trainer) devices = trainer.num_devices num_nodes = trainer.num_nodes if (not any([isinstance(strategy, supported_strategy) for supported_strategy in supported_strategies])): raise AttributeError(f'Strategy {strategy} not supported.') elif any([isinstance(strategy, tpu_strategy) for tpu_strategy in tpu_strategies]): return (local_batch_size * devices) elif any([isinstance(strategy, process_independent_strategy) for process_independent_strategy in process_independent_strategies]): return ((local_batch_size * devices) * num_nodes) elif any([isinstance(strategy, fully_dependent_strategy) for fully_dependent_strategy in fully_dependent_strategies]): return local_batch_size else: raise AttributeError(f'Strategy {strategy} not supported.')
def get_local_batch_size_in_trainer(global_batch_size: int, trainer: Trainer) -> int: 'Get local batch size used by a trainer based on the global batch size.\n\n Args:\n global_batch_size: The global batch size used by the trainer.\n strategy: The trainer used.\n\n Raises:\n AttributeError: The strategy is not supported.\n\n Returns:\n The local batch size.\n ' strategy = get_trainer_strategy(trainer) devices = trainer.num_devices num_nodes = trainer.num_nodes if (not any([isinstance(strategy, supported_strategy) for supported_strategy in supported_strategies])): raise AttributeError(f'Strategy {strategy} not supported.') elif any([isinstance(strategy, tpu_strategy) for tpu_strategy in tpu_strategies]): return (global_batch_size // devices) elif any([isinstance(strategy, process_independent_strategy) for process_independent_strategy in process_independent_strategies]): return ((global_batch_size // devices) // num_nodes) elif any([isinstance(strategy, fully_dependent_strategy) for fully_dependent_strategy in fully_dependent_strategies]): return global_batch_size else: raise AttributeError(f'Strategy {strategy} not supported.')
def get_num_devices_in_trainer(trainer: Trainer) -> int: 'Get the number of devices used by the trainer.\n\n Args:\n trainer: The trainer.\n\n Raises:\n AttributeError: The strategy used by trainer is not supported\n\n Returns:\n The number of devices used by trainer.\n ' strategy = get_trainer_strategy(trainer) if (not any([isinstance(strategy, supported_strategy) for supported_strategy in supported_strategies])): raise AttributeError(f'Strategy {strategy} not supported.') elif any([isinstance(strategy, tpu_strategy) for tpu_strategy in tpu_strategies]): return trainer.num_devices elif any([isinstance(strategy, process_independent_strategy) for process_independent_strategy in process_independent_strategies]): return (trainer.num_devices * trainer.num_nodes) elif any([isinstance(strategy, fully_dependent_strategy) for fully_dependent_strategy in fully_dependent_strategies]): return 1 else: raise AttributeError(f'Strategy {strategy} not supported.')
def get_trainer_strategy(trainer: Trainer) -> Any: 'Retrieve the strategy from a trainer.\n\n Args:\n trainer: The trainer.\n\n Returns:\n The strategy.\n ' if (pl.__version__ < '1.6.0'): return trainer.training_type_plugin else: return trainer.strategy
def is_strategy_ddp(strategy: Any) -> bool: 'Test if strategy is ddp.\n\n Args:\n strategy: The strategy.\n\n Returns:\n ``True`` if strategy is ddp.\n ' return any([isinstance(strategy, process_strategy) for process_strategy in process_independent_strategies])
def is_strategy_tpu(strategy: Any) -> bool: 'Test if strategy is tpu.\n\n Args:\n strategy: The strategy.\n\n Returns:\n ``True`` if strategy is tpu.\n ' return any([isinstance(strategy, tpu_strategy) for tpu_strategy in tpu_strategies])
def compile_model(model: LightningModule, do_compile: bool=False, fullgraph: bool=False, dynamic: bool=False, backend: Union[(str, Callable)]='inductor', mode: Union[(str, None)]=None, options: Optional[Dict[(str, Union[(str, int, bool)])]]=None, disable: bool=False): "If torch version is greater than `'2.0.0'` and users ask for it, compile the model.\n\n Args:\n model: Model to compile.\n do_compile: Whether to compile the model.\n fullgraph: Defined by `torch.compile`.\n dynamic: Defined by `torch.compile`.\n backend: Defined by `torch.compile`.\n mode: Defined by `torch.compile`.\n passes: Defined by `torch.compile`.\n\n Returns:\n The compiled model if available else the model.\n " if ((version.parse(torch.__version__) >= version.parse('2.0.0.dev')) and do_compile): rank_zero_info(f'Compiling model {model.__class__.__name__}.') return torch.compile(model=model, fullgraph=fullgraph, dynamic=dynamic, backend=backend, mode=mode, options=options, disable=disable) else: rank_zero_info(f'Not compiling model {model.__class__.__name__}.') return model
def get_default_seed(default_seed: int=0) -> int: 'Get the default seed if pytorch lightning did not initialize one.\n\n Args:\n default_seed: The default seed.\n\n Returns:\n Pytorch lightning seed or the default one.\n ' return int(os.getenv('PL_GLOBAL_SEED', default_seed))
def get_global_rank() -> int: 'Get global rank of the process.' if (dist.is_available() and dist.is_initialized()): return dist.get_rank() if (('RANK' in os.environ) and ('WORLD_SIZE' in os.environ)): rank = int(os.environ['RANK']) elif (int(os.environ.get('SLURM_NPROCS', 1)) > 1): rank = int(os.environ['SLURM_PROCID']) else: rank = 0 return rank
def get_local_rank() -> int: 'Get local rank of the process.' if (('RANK' in os.environ) and ('WORLD_SIZE' in os.environ)): local_rank = int(os.environ['LOCAL_RANK']) elif (int(os.environ.get('SLURM_NPROCS', 1)) > 1): local_rank = int(os.environ['SLURM_LOCALID']) else: local_rank = 0 return local_rank
def get_world_size() -> int: 'Get world size or number of the processes.' if (dist.is_available() and dist.is_initialized()): return dist.get_world_size() if (('RANK' in os.environ) and ('WORLD_SIZE' in os.environ)): world_size = int(os.environ['WORLD_SIZE']) elif (int(os.environ.get('SLURM_NPROCS', 1)) > 1): world_size = int(os.environ['SLURM_NPROCS']) else: world_size = 1 return world_size
def get_local_world_size() -> int: 'Get local world size or number of processes on the node.' if (dist.is_available() and dist.is_initialized()): return torch.cuda.device_count() else: return 1
def is_only_one_condition_true(*conditions: List[bool]) -> bool: 'Test if only one of the conditions is True.' a = conditions[0] b = conditions[0] for condition in conditions[1:]: a = (a ^ condition) b = (b & condition) return (a & (~ b))
def all_false(*conditions: List[bool]) -> bool: 'Test that all conditions are False.' return all([(~ condition) for condition in conditions])
def warmup_value(initial_value: float, final_value: float, step: int=0, max_step: int=0) -> float: 'Apply warmup to a value.\n\n Args:\n initial_value: Initial value.\n final_value: Final value.\n step: Current step.\n max_step: Max step for warming up.\n\n Returns:\n The value at current warmup step.\n ' if (step >= max_step): return final_value else: return (initial_value + ((step * (final_value - initial_value)) / max_step))
def scheduler_value(scheduler: Optional[str], initial_value: float, final_value: float, step: int=0, max_step: int=0) -> float: 'Apply scheduler to a value.\n\n Args:\n scheduler: The type of the scheduler.\n initial_value: The initial value.\n final_value: The final value.\n step: Current step.\n max_step: Maximum step for scheduler.\n\n Returns:\n The value at current step.\n ' if (scheduler is None): return initial_value elif (scheduler == 'linear'): if (final_value < initial_value): return (initial_value + ((step * (final_value - initial_value)) / max_step)) else: return (final_value + ((step * (initial_value - final_value)) / max_step)) elif (scheduler == 'cosine'): if (final_value > initial_value): return (initial_value + ((0.5 * (1.0 + math.cos((math.pi + ((math.pi * step) / max_step))))) * (final_value - initial_value))) else: return (initial_value - ((0.5 * (1.0 + math.cos((math.pi + ((math.pi * step) / max_step))))) * (initial_value - final_value)))
def apply_several_transforms(images: Iterable[Tensor], transforms: Iterable[Module]) -> List[List[Tensor]]: 'Apply several transformations to a list of images.\n\n Args:\n images: The images.\n transforms: The transformations to apply to the images.\n\n Returns:\n List of list of transformed images.\n ' transformed_images = [[transform(image) for image in images] for transform in transforms] return transformed_images
def apply_several_video_transforms(videos: Iterable[Dict[(str, Any)]], transforms: Iterable[Module]) -> List[List[Tensor]]: 'Apply several transformations to a list of videos.\n\n Args:\n videos: The videos.\n transforms: The transformations to apply to the videos.\n\n Returns:\n List of list of transformed videos.\n ' transformed_images = [[transform(copy.deepcopy(video)) for video in videos] for transform in transforms] return transformed_images
def make_grid_from_several_transforms(sets_images: Iterable[Iterable[Tensor]], n_images_per_row: int=8) -> Tensor: 'Make a grid of images by aligning images from several transformations vertically.\n\n Args:\n sets_images: Sets of transformed images aligned, base_image(sets_images[0][?]) == ... == base_images(sets_images[-1][?]).\n n_images_per_row: Number of images displayed per row.\n\n Returns:\n Grid of images.\n ' n_images = len(sets_images[0]) for set_image in sets_images: assert (len(set_image) == n_images) if (n_images_per_row == (- 1)): nrow = n_images all_images = torch.cat(*sets_images, dim=0) else: nrow = n_images_per_row all_images = [] number_of_rows = math.ceil((n_images // n_images_per_row)) for row_idx in range(number_of_rows): for set_image in sets_images: for i in range((row_idx * n_images_per_row), min(n_images, ((row_idx + 1) * n_images_per_row))): all_images.append(set_image[i]) grid = torchvision.utils.make_grid(all_images, nrow=nrow) return grid
def make_several_transforms_from_config(cfg_transforms: Mapping[(Any, Any)]) -> List[Module]: "Make several transformations from a configuration dictionary.\n\n Args:\n cfg_transforms: Configuration of transformations in the form {'transform1': {...}, 'transform2': {...}}.\n\n Returns:\n List of transformations.\n " transforms = [hydra.utils.instantiate(conf_transform) for (_, conf_transform) in cfg_transforms.items()] return transforms
def show_images(imgs: Union[(Iterable[Tensor], Tensor)], figsize: Iterable[float]=[6.4, 4.8]): 'Show images from a tensor or a list of tensor.\n\n Args:\n imgs: Images to display.\n figsize: Figure size for the images.\n ' if (not isinstance(imgs, list)): imgs = [imgs] (_, axs) = plt.subplots(ncols=len(imgs), squeeze=False, figsize=figsize) for (i, img) in enumerate(imgs): img = img.detach() img = np.asarray(F.to_pil_image(img)) axs[(0, i)].imshow(np.asarray(img)) axs[(0, i)].set(xticklabels=[], yticklabels=[], xticks=[], yticks=[]) plt.show()
def show_video(video: Tensor) -> animation.ArtistAnimation: 'Show a video thanks to animation from matplotlib.\n\n Args:\n video: The raw video to display.\n\n Returns:\n The animation to show.\n ' video = video.long() video = np.asarray(video.permute(1, 2, 3, 0)) (fig, ax) = plt.subplots() fig.patch.set_visible(False) frames = [[ax.imshow(video[i], animated=True)] for i in range(len(video))] anim = animation.ArtistAnimation(fig, frames) plt.axis('off') plt.show() return anim
def main(): args = parser.parse_args() soccernet_dataset(data_path=args.data_path, transform=None, video_path_prefix=args.path_prefix, decoder_args={'fps': args.fps}, features_args=None, label_args={'radius_label': args.radius_label, 'cache_dir': args.cache_dir}, task=SoccerNetTask(args.task))
def get_video_duration(video_file): cmd = ['ffmpeg', '-i', str(video_file), '-f', 'null', '-'] try: output = subprocess.check_output(cmd, stderr=subprocess.STDOUT) except subprocess.CalledProcessError as err: print(video_file, err.output) return (- 1) try: output_decoded = output.decode() result_all = ffmpeg_duration_template.findall(output_decoded) except Exception as err: print(video_file, err, 'chose to carry on decoding video') return 1 if result_all: result = result_all[(- 1)] duration = (((((float(result[0]) * 60) * 60) + (float(result[1]) * 60)) + float(result[2])) + (float(result[3]) * (10 ** (- len(result[3]))))) else: duration = (- 1) return duration
def has_video_stream(video_file): cmd = ['ffprobe', '-i', str(video_file), '-show_streams', '-select_streams', 'v', '-loglevel', 'error'] try: output = subprocess.check_output(cmd, stderr=subprocess.STDOUT) except subprocess.CalledProcessError as err: print(video_file, err.output) return False return (output != '')
def is_video_empty(video_file): return ((get_video_duration(video_file) <= 0) or (not has_video_stream(video_file)))
def process(row, folder_path, output_path, args): classname = row[0] videoname = row[1] videostem = row[2] inname = ((folder_path / classname) / videoname) if is_video_empty(inname): print(f'{inname} is empty.') return (False, f'{inname} is empty.') output_folder = (output_path / classname) if (os.path.isdir(output_folder) is False): try: os.mkdir(output_folder) except: print(f"{output_folder} can't be created.") if args.downscale: downscaled_cmd = f"""-c:v libx264 -filter:v "scale='if(gt(ih,iw),{args.downscale_size},trunc(oh*a/2)*2):if(gt(ih,iw),trunc(ow/a/2)*2,{args.downscale_size})'" -c:a copy""" else: downscaled_cmd = '' if args.frames: outfile = '%08d.jpg' outfolder = (output_folder / videostem) outfolder.mkdir(exist_ok=False, parents=False) outname = (outfolder / outfile) frames_cmd = f'-q:v {args.video_quality}' else: outname = (output_folder / videoname) frames_cmd = '' if (args.fps > 0): fps_cmd = f'-r {args.fps}' else: fps_cmd = '' status = False inname = ('"%s"' % inname) outname = ('"%s"' % outname) command = f'ffmpeg -loglevel panic -i {inname} {downscaled_cmd} {frames_cmd} {fps_cmd} {outname}' try: subprocess.run(command, shell=True, check=True, stdout=subprocess.PIPE).stdout except subprocess.CalledProcessError as err: print(inname, outname, status, err.output) if args.frames: os.rmdir(outfolder) return (status, err.output) status = os.path.exists(outname) return (status, 'Process')
@hydra.main(config_path='../eztorch/configs/run/evaluation/feature_extractor/resnet3d50', config_name='resnet3d50_ucf101') def main(config: DictConfig) -> None: rundir = Path(to_absolute_path(config.dir.run)) rundir.mkdir(parents=True, exist_ok=True) os.chdir(rundir) rank_zero_info(f'Run directory: {rundir}') hydradir = (rundir / 'config/') hydradir.mkdir(parents=True, exist_ok=True) config_file = (hydradir / 'extract_features.yaml') resolved_config = OmegaConf.to_yaml(config, resolve=True) with config_file.open(mode='w') as f: f.write(resolved_config) if config.get('seed'): hydra.utils.instantiate(config.seed) else: warnings.warn('No seed fixed, the results are not reproducible.') datamodule: LightningDataModule = hydra.utils.instantiate(config.datamodule) if (not config.model.get('_target_')): with open_dict(config): config.model._target_ = 'eztorch.evaluation.FeatureExtractor' config.model._recursive_ = False model: Module = hydra.utils.instantiate(config.model) callbacks = [] if config.get('callbacks'): for (_, callback_cfg) in config.callbacks.items(): callback: Callback = hydra.utils.instantiate(callback_cfg) callbacks.append(callback) trainer: Trainer = hydra.utils.instantiate(config.trainer, callbacks=callbacks, devices=1, strategy='auto') rank_zero_info(resolved_config) rank_zero_info(model) if config.datamodule.get('train'): trainer.fit(model, datamodule=datamodule) elif config.datamodule.get('val'): trainer.validate(model, datamodule=datamodule) if config.datamodule.get('test'): trainer.test(model, datamodule=datamodule)
@hydra.main(config_path='../eztorch/configs/run/evaluation/linear_classifier/sce/resnet50', config_name='resnet50_imagenet_mocov3') def main(config: DictConfig) -> None: rundir = Path(to_absolute_path(config.dir.run)) rundir.mkdir(parents=True, exist_ok=True) os.chdir(rundir) rank_zero_info(f'Run directory: {rundir}') hydradir = (rundir / 'config') hydradir.mkdir(parents=True, exist_ok=True) config_file = (hydradir / 'linear_classifier_evaluation.yaml') resolved_config = OmegaConf.to_yaml(config, resolve=True) with config_file.open(mode='w') as f: f.write(resolved_config) if config.get('seed'): hydra.utils.instantiate(config.seed) else: warnings.warn('No seed fixed, the results are not reproducible.') callbacks = [] if config.get('callbacks'): for (_, callback_cfg) in config.callbacks.items(): callback: Callback = hydra.utils.instantiate(callback_cfg) callbacks.append(callback) trainer: Trainer = hydra.utils.instantiate(config.trainer, callbacks=callbacks) datamodule: LightningDataModule = hydra.utils.instantiate(config.datamodule) model: LightningModule = hydra.utils.instantiate(config.model) model_ckpt_dirpath = (config.callbacks.model_checkpoint.dirpath if config.callbacks.get('model_checkpoint') else None) ckpt_path = get_last_ckpt_in_path_or_dir(config.ckpt_path, model_ckpt_dirpath) if (ckpt_path is not None): warnings.warn(f'A checkpoint has been found and loaded from this file: {ckpt_path}', category=RuntimeWarning) rank_zero_info(resolved_config) rank_zero_info(model) model = compile_model(model, **config.get('compile', {})) trainer.fit(model, datamodule=datamodule, ckpt_path=ckpt_path) if config.get('test'): if config.test.get('ckpt_by_callback_mode'): ckpt_paths = get_ckpt_by_callback_mode(config.test.ckpt_path, config.test.ckpt_by_callback_mode) else: ckpt_paths = [config.test.ckpt_path] for ckpt_path in ckpt_paths: trainer.test(model, ckpt_path=ckpt_path, datamodule=datamodule)
@hydra.main(config_path='../eztorch/configs/run/pretrain/sce/resnet50', config_name='resnet50_imagenet') def main(config: DictConfig) -> None: rundir = Path(to_absolute_path(config.dir.run)) rundir.mkdir(parents=True, exist_ok=True) os.chdir(rundir) rank_zero_info(f'Run directory: {rundir}') hydradir = (rundir / 'config/') hydradir.mkdir(parents=True, exist_ok=True) config_file = (hydradir / 'pretrain.yaml') resolved_config = OmegaConf.to_yaml(config, resolve=True) with config_file.open(mode='w') as f: f.write(resolved_config) if config.get('seed'): hydra.utils.instantiate(config.seed) else: warnings.warn('No seed fixed, the results are not reproducible.') callbacks = [] if config.get('callbacks'): for (_, callback_cfg) in config.callbacks.items(): callback: Callback = hydra.utils.instantiate(callback_cfg) callbacks.append(callback) trainer: Trainer = hydra.utils.instantiate(config.trainer, callbacks=callbacks) datamodule: LightningDataModule = hydra.utils.instantiate(config.datamodule) model: LightningModule = hydra.utils.instantiate(config.model) model_ckpt_dirpath = (config.callbacks.model_checkpoint.dirpath if config.callbacks.get('model_checkpoint') else None) ckpt_path = get_last_ckpt_in_path_or_dir(config.ckpt_path, model_ckpt_dirpath) if (ckpt_path is not None): warnings.warn(f'A checkpoint has been found and loaded from this file: {ckpt_path}', category=RuntimeWarning) rank_zero_info(resolved_config) rank_zero_info(model) model = compile_model(model, **config.get('compile', {})) trainer.fit(model, datamodule=datamodule, ckpt_path=ckpt_path)
@hydra.main(config_path='../eztorch/configs/run/pretrain/moco', config_name='resnet18_cifar10') def main(config: DictConfig) -> None: rundir = Path(to_absolute_path(config.dir.run)) rundir.mkdir(parents=True, exist_ok=True) os.chdir(rundir) rank_zero_info(f'Run directory: {rundir}') hydradir = (rundir / 'config/') hydradir.mkdir(parents=True, exist_ok=True) config_file = (hydradir / 'dataloader.yaml') resolved_config = OmegaConf.to_yaml(config, resolve=True) with config_file.open(mode='w') as f: f.write(resolved_config) if config.get('seed'): hydra.utils.instantiate(config.seed) else: warnings.warn('No seed fixed, the results are not reproducible.') trainer: Trainer = hydra.utils.instantiate(config.trainer, callbacks=[]) datamodule: LightningDataModule = hydra.utils.instantiate(config.datamodule) if (config.model.get('input_shape') is None): raise AssertionError('input_shape should be specified in model config.') if config.model.get('transform'): transform = config.model.train_transform else: transform = None model: LightningModule = DummyModel(config.model.input_shape, transform=transform) rank_zero_info(config.datamodule) rank_zero_info(model) trainer.fit(model, datamodule=datamodule)
@hydra.main(config_path='../eztorch/configs/run/evaluation/retrieval_from_bank', config_name='default') def main(config: DictConfig) -> None: rundir = Path(to_absolute_path(config.dir.run)) rundir.mkdir(parents=True, exist_ok=True) os.chdir(rundir) rank_zero_info(f'Run directory: {rundir}') hydradir = (rundir / 'config/') hydradir.mkdir(parents=True, exist_ok=True) config_file = (hydradir / 'retrieval_train_from_test.yaml') resolved_config = OmegaConf.to_yaml(config, resolve=True) with config_file.open(mode='w') as f: f.write(resolved_config) rank_zero_info(resolved_config) ranks = config.ranks rank_zero_info('\nLoading query features and labels...') query_features = torch.load(config.query.features_path) query_labels = torch.load(config.query.labels_path) rank_zero_info(f'''Loaded query features and labels. shape of features is: {query_features.shape}. shape of labels is: {query_labels.shape}.''') rank_zero_info('\nLoading bank features and labels...') bank_features = torch.load(config.bank.features_path) bank_labels = torch.load(config.bank.labels_path) rank_zero_info(f'''Loaded bank features and labels. shape of features is: {bank_features.shape}. shape of labels is: {bank_labels.shape}.''') if torch.cuda.is_available(): rank_zero_info('\nCuda available, tensors put on GPU...') query_features = query_features query_labels = query_labels bank_features = bank_features bank_labels = bank_labels if config.query.center: rank_zero_info('\nQuery centering...') query_features = (query_features - query_features.mean(dim=0, keepdim=True)) rank_zero_info('Query centered...') if config.bank.center: rank_zero_info('\nBank centering...') bank_features = (bank_features - bank_features.mean(dim=0, keepdim=True)) rank_zero_info('Bank centered...') if config.query.normalize: rank_zero_info('\nQuery normalizing...') query_features = torch.nn.functional.normalize(query_features, p=2, dim=1) rank_zero_info('Query normalized...') if config.bank.normalize: rank_zero_info('\nBank normalizing...') bank_features = torch.nn.functional.normalize(bank_features, p=2, dim=1) rank_zero_info('Bank normalized...') rank_zero_info('\nComputing similarties...') sim = query_features.matmul(bank_features.t()) rank_zero_info('Computed similarities...') rank_zero_info('\nStart computing metrics:') for rank in ranks: (_, topkidx) = torch.topk(sim, rank, dim=1) acc = torch.any((bank_labels[topkidx] == query_labels.unsqueeze(1)), dim=1).float().mean().item() rank_zero_info(f'R @ {rank} = {acc:.4f}')
@hydra.main(config_path='../eztorch/configs/run/supervised/resnet3d18', config_name='kinetics200') def main(config: DictConfig) -> None: rundir = Path(to_absolute_path(config.dir.run)) rundir.mkdir(parents=True, exist_ok=True) os.chdir(rundir) rank_zero_info(f'Run directory: {rundir}') hydradir = (rundir / 'config/') hydradir.mkdir(parents=True, exist_ok=True) config_file = (hydradir / 'supervised.yaml') resolved_config = OmegaConf.to_yaml(config, resolve=True) with config_file.open(mode='w') as f: f.write(resolved_config) if config.get('seed'): hydra.utils.instantiate(config.seed) else: warnings.warn('No seed fixed, the results are not reproducible.') callbacks = [] if config.get('callbacks'): for (_, callback_cfg) in config.callbacks.items(): callback: Callback = hydra.utils.instantiate(callback_cfg) callbacks.append(callback) trainer: Trainer = hydra.utils.instantiate(config.trainer, callbacks=callbacks) datamodule: LightningDataModule = hydra.utils.instantiate(config.datamodule) model: LightningModule = hydra.utils.instantiate(config.model) model_ckpt_dirpath = (config.callbacks.model_checkpoint.dirpath if config.callbacks.get('model_checkpoint') else None) ckpt_path = get_last_ckpt_in_path_or_dir(config.ckpt_path, model_ckpt_dirpath) if (ckpt_path is not None): warnings.warn(f'A checkpoint has been found and loaded from this file: {ckpt_path}', category=RuntimeWarning) rank_zero_info(resolved_config) rank_zero_info(model) model = compile_model(model, **config.get('compile', {})) trainer.fit(model, datamodule=datamodule, ckpt_path=ckpt_path) if config.get('test'): if config.test.get('ckpt_by_callback_mode'): ckpt_paths = get_ckpt_by_callback_mode(config.test.ckpt_path, config.test.ckpt_by_callback_mode) else: ckpt_paths = [config.test.ckpt_path] for ckpt_path in ckpt_paths: trainer.test(model, ckpt_path=ckpt_path, datamodule=datamodule)
@hydra.main(config_path='../eztorch/configs/run/supervised/resnet3d18', config_name='kinetics200') def main(config: DictConfig) -> None: rundir = Path(to_absolute_path(config.dir.run)) rundir.mkdir(parents=True, exist_ok=True) os.chdir(rundir) rank_zero_info(f'Run directory: {rundir}') hydradir = (rundir / 'config/') hydradir.mkdir(parents=True, exist_ok=True) config_file = (hydradir / 'supervised.yaml') resolved_config = OmegaConf.to_yaml(config, resolve=True) with config_file.open(mode='w') as f: f.write(resolved_config) if config.get('seed'): hydra.utils.instantiate(config.seed) else: warnings.warn('No seed fixed, the results are not reproducible.') callbacks = [] if config.get('callbacks'): for (_, callback_cfg) in config.callbacks.items(): callback: Callback = hydra.utils.instantiate(callback_cfg) callbacks.append(callback) schedule = torch.profiler.schedule(wait=0, warmup=50, active=2, repeat=0, skip_first=0) trace_handler = torch.profiler.tensorboard_trace_handler(dir_name=config.dir.root) profiler = PyTorchProfiler(dirpath=config.dir.root, export_to_chrome=False, with_stack=True, schedule=schedule, on_trace_ready=trace_handler) trainer: Trainer = hydra.utils.instantiate(config.trainer, callbacks=callbacks, profiler=profiler) datamodule: LightningDataModule = hydra.utils.instantiate(config.datamodule) model: LightningModule = hydra.utils.instantiate(config.model) model_ckpt_dirpath = (config.callbacks.model_checkpoint.dirpath if config.callbacks.get('model_checkpoint') else None) ckpt_path = get_last_ckpt_in_path_or_dir(config.ckpt_path, model_ckpt_dirpath) if (ckpt_path is not None): warnings.warn(f'A checkpoint has been found and loaded from this file: {ckpt_path}', category=RuntimeWarning) rank_zero_info(resolved_config) rank_zero_info(model) model = compile_model(model, **config.get('compile', {})) trainer.fit(model, datamodule=datamodule, ckpt_path=ckpt_path) if config.get('test'): if config.test.get('ckpt_by_callback_mode'): ckpt_paths = get_ckpt_by_callback_mode(config.test.ckpt_path, config.test.ckpt_by_callback_mode) else: ckpt_paths = [config.test.ckpt_path] for ckpt_path in ckpt_paths: trainer.test(model, ckpt_path=ckpt_path, datamodule=datamodule)
@hydra.main(config_path='../eztorch/configs/run/supervised/resnet3d50', config_name='ucf101') def main(config: DictConfig) -> None: rundir = Path(to_absolute_path(config.dir.run)) rundir.mkdir(parents=True, exist_ok=True) os.chdir(rundir) rank_zero_info(f'Run directory: {rundir}') hydradir = (rundir / 'config/') hydradir.mkdir(parents=True, exist_ok=True) config_file = (hydradir / 'test.yaml') resolved_config = OmegaConf.to_yaml(config, resolve=True) with config_file.open(mode='w') as f: f.write(resolved_config) if config.get('seed'): hydra.utils.instantiate(config.seed) else: warnings.warn('No seed fixed, the results are not reproducible.') callbacks = [] if config.get('callbacks'): for (_, callback_cfg) in config.callbacks.items(): callback: Callback = hydra.utils.instantiate(callback_cfg) callbacks.append(callback) datamodule: LightningDataModule = hydra.utils.instantiate(config.datamodule) model: LightningModule = hydra.utils.instantiate(config.model) model_ckpt_dirpath = (config.callbacks.model_checkpoint.dirpath if config.callbacks.get('model_checkpoint') else None) ckpt_path = get_last_ckpt_in_path_or_dir(config.ckpt_path, model_ckpt_dirpath) if (ckpt_path is not None): warnings.warn(f'A checkpoint has been found and loaded from this file: {ckpt_path}', category=RuntimeWarning) rank_zero_info(resolved_config) rank_zero_info(model) model = compile_model(model, **config.get('compile', {})) if config.test.get('ckpt_by_callback_mode'): ckpt_paths = get_ckpt_by_callback_mode(config.test.ckpt_path, config.test.ckpt_by_callback_mode) else: ckpt_paths = [config.test.ckpt_path] for ckpt_path in ckpt_paths: trainer: Trainer = hydra.utils.instantiate(config.trainer, callbacks=callbacks, devices=1, strategy='auto') trainer.test(model, ckpt_path=ckpt_path, datamodule=datamodule)
class TestFrameSoccerNetVideo(unittest.TestCase): def setUp(self) -> None: self.default_args = {'video_path': Path('/video/'), 'half_path': Path('/video/half1'), 'transform': None, 'video_frame_to_path_fn': get_video_to_frame_path_fn(zeros=8), 'num_threads_io': 0} def test_same_fps_get_timestamps_indices(self) -> None: video = FrameSoccerNetVideo(duration=2700, fps_video=2, fps=2, num_frames=5400, **self.default_args) (timestamps, frame_indices, fps_video_frame_indices) = video.get_timestamps_and_frame_indices(0.75, 15.02) expected_timestamps = torch.arange(0.5, 15, 0.5) expected_frame_indices = torch.arange(1, (15 * 2)) expected_fps_video_frame_indices = expected_frame_indices assert torch.allclose(timestamps, expected_timestamps) assert torch.allclose(frame_indices, expected_frame_indices) assert torch.allclose(fps_video_frame_indices, expected_fps_video_frame_indices) video = FrameSoccerNetVideo(duration=2700, fps_video=2, fps=2, num_frames=5400, **self.default_args) (timestamps, frame_indices, fps_video_frame_indices) = video.get_timestamps_and_frame_indices(0.75, 15.52) expected_timestamps = torch.arange(0.5, 15.5, 0.5) expected_frame_indices = torch.arange(1, ((15 * 2) + 1)) expected_fps_video_frame_indices = expected_frame_indices assert torch.allclose(timestamps, expected_timestamps) assert torch.allclose(frame_indices, expected_frame_indices) assert torch.allclose(fps_video_frame_indices, expected_fps_video_frame_indices) video = FrameSoccerNetVideo(duration=2699.5, fps_video=2, fps=2, num_frames=5399, **self.default_args) (timestamps, frame_indices, fps_video_frame_indices) = video.get_timestamps_and_frame_indices(0.0, 2699.5) expected_timestamps = torch.arange(0.0, 2699.5, 0.5) expected_frame_indices = torch.arange(0, ((2699 * 2) + 1)) expected_fps_video_frame_indices = expected_frame_indices assert torch.allclose(timestamps, expected_timestamps) assert torch.allclose(frame_indices, expected_frame_indices) assert torch.allclose(fps_video_frame_indices, expected_fps_video_frame_indices) def test_different_fps_get_timestamps_indices(self) -> None: video = FrameSoccerNetVideo(duration=2700, fps_video=25, fps=2, num_frames=67500, **self.default_args) (timestamps, frame_indices, fps_video_frame_indices) = video.get_timestamps_and_frame_indices(0.75, 15.02) expected_timestamps = torch.arange(0.5, 15, 0.5) expected_frame_indices = torch.arange(1, (15 * 2)) expected_fps_video_frame_indices = torch.floor(((expected_frame_indices / 2) * 25)).to(dtype=torch.long) assert torch.allclose(timestamps, expected_timestamps) assert torch.allclose(frame_indices, expected_frame_indices) assert torch.allclose(fps_video_frame_indices, expected_fps_video_frame_indices) video = FrameSoccerNetVideo(duration=2700, fps_video=25, fps=2, num_frames=67500, **self.default_args) (timestamps, frame_indices, fps_video_frame_indices) = video.get_timestamps_and_frame_indices(0.75, 15.52) expected_timestamps = torch.arange(0.5, 15.5, 0.5) expected_frame_indices = torch.arange(1, ((15 * 2) + 1)) expected_fps_video_frame_indices = torch.floor(((expected_frame_indices / 2) * 25)).to(dtype=torch.long) assert torch.allclose(timestamps, expected_timestamps) assert torch.allclose(frame_indices, expected_frame_indices) assert torch.allclose(fps_video_frame_indices, expected_fps_video_frame_indices) video = FrameSoccerNetVideo(duration=2699.96, fps_video=25, fps=2, num_frames=67499, **self.default_args) (timestamps, frame_indices, fps_video_frame_indices) = video.get_timestamps_and_frame_indices(0.0, 2699.96) expected_timestamps = torch.arange(0.0, 2699.5, 0.5) expected_frame_indices = torch.arange(0, ((2699 * 2) + 1)) expected_fps_video_frame_indices = torch.floor(((expected_frame_indices / 2) * 25)).to(dtype=torch.long) assert torch.allclose(timestamps, expected_timestamps) assert torch.allclose(frame_indices, expected_frame_indices) assert torch.allclose(fps_video_frame_indices, expected_fps_video_frame_indices) video = FrameSoccerNetVideo(duration=2700, fps_video=25, fps=4, num_frames=67500, **self.default_args) (timestamps, frame_indices, fps_video_frame_indices) = video.get_timestamps_and_frame_indices(0.76, 15.02) expected_timestamps = torch.arange(0.75, 15, 0.25) expected_frame_indices = torch.arange(3, (15 * 4)) expected_fps_video_frame_indices = torch.tensor([19, 25, 31, 37, 44, 50, 56, 62, 69, 75, 81, 87, 94, 100, 106, 112, 119, 125, 131, 137, 144, 150, 156, 162, 169, 175, 181, 187, 194, 200, 206, 212, 219, 225, 231, 237, 244, 250, 256, 262, 269, 275, 281, 287, 294, 300, 306, 312, 319, 325, 331, 337, 344, 350, 356, 362, 369]) assert torch.allclose(timestamps, expected_timestamps) assert torch.allclose(frame_indices, expected_frame_indices) assert torch.allclose(fps_video_frame_indices, expected_fps_video_frame_indices)
class TestSoccerNetDataset(unittest.TestCase): def test_soccernet_dataset(self): dataset = soccernet_dataset((Path(os.path.realpath(__file__)).parent / 'small_annotations.json'), None, (Path(os.path.realpath(__file__)).parent / 'images'), label_args={'radius_label': 2, 'cache_dir': (Path(os.path.realpath(__file__)).parent / 'labels')}, decoder_args={'fps': 2}) sample = dataset[(0, 1, 1860, 1870)] expected_sample = {'timestamps': tensor([1860.0, 1860.5, 1861.0, 1861.5, 1862.0, 1862.5, 1863.0, 1863.5, 1864.0, 1864.5, 1865.0, 1865.5, 1866.0, 1866.5, 1867.0, 1867.5, 1868.0, 1868.5, 1869.0, 1869.5]), 'video_num_timestamps': tensor(5400), 'labels': tensor([[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]]), 'ignore_class': tensor([[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]]).bool()} for (key, value) in expected_sample.items(): assert torch.allclose(sample[key], value, rtol=0.001, atol=0.001)
class TestSoccerNetPredictions(unittest.TestCase): def test_aggregate_predictions(self): predictions = torch.tensor([[0.5, 0.6], [0.4, 0.7], [0.7, 0.3], [0.3, 0.5], [0.8, 0.2], [0.9, 0.9]]) timestamps = torch.tensor([[0.0, 0.0], [1.0, 1.0], [1.0, 1.0], [2.0, 2.0], [8.0, 2.0], [8.0, 8.0]]) expected_predictions = torch.tensor([[0.3, 0.6], [0.5, 0.7], [0.7, 0.0], [0.0, 0.0], [0.8, 0.0], [0.0, 0.5], [0.0, 0.0], [0.9, 0.9]]) expected_timestamps = torch.tensor([0, 1, 2, 3, 4, 5, 6, 7]) (new_predictions, new_timestamps) = aggregate_predictions(predictions, timestamps, 8, 1.0) assert torch.allclose(expected_predictions, new_predictions) assert torch.allclose(expected_timestamps, new_timestamps) def test_aggregate_predictions_with_ignore(self): predictions = torch.tensor([[0.5, 0.6], [0.4, 0.7], [0.7, 0.3], [0.3, 0.5], [0.8, 0.2], [0.9, 0.9]]) timestamps = torch.tensor([[0.0, 0.0], [1.0, 1.0], [1.0, 1.0], [2.0, 2.0], [8.0, 2.0], [8.0, 8.0]]) expected_predictions = torch.tensor([[0.0, 0.6], [0.5, 0.7], [0.7, 0.0], [0.0, 0.0], [0.8, 0.0], [0.0, 0.5], [0.0, 0.0], [0.9, 0.0], [0.0, 0.0]]) expected_timestamps = torch.tensor([0, 1, 2, 3, 4, 5, 6, 7, 8]) (new_predictions, new_timestamps) = aggregate_predictions(predictions, timestamps, 8, 1.0, True) assert torch.allclose(expected_predictions, new_predictions) assert torch.allclose(expected_timestamps, new_timestamps) def test_postprocess_spotting_half_predictions(self): predictions = torch.tensor([[0.5, 0.3], [1.0, 0.6], [0.2, 0.8], [0.3, 0.2], [0.1, 0.1], [0.2, 0.9], [0.6, 0.8]]) NMS_args = {'window': 3, 'threshold': 0.49} half_id = 0 half_predictions = postprocess_spotting_half_predictions(predictions, half_id, 15, NMS_args) expected_half_predictions = [{'gameTime': f'{half_id} - {int((15 / 60)):02d}:{int((15 % 60))}', 'label': REVERSE_ACTION_SPOTTING_LABELS[0], 'position': f'{int(15000)}', 'half': str(half_id), 'confidence': float(1.0)}, {'gameTime': f'{half_id} - {int((90 / 60)):02d}:{int((90 % 60))}', 'label': REVERSE_ACTION_SPOTTING_LABELS[0], 'position': f'{int(90000)}', 'half': str(half_id), 'confidence': float(0.6)}, {'gameTime': f'{half_id} - {int((30 / 60)):02d}:{int((30 % 60))}', 'label': REVERSE_ACTION_SPOTTING_LABELS[1], 'position': f'{int(30000)}', 'half': str(half_id), 'confidence': float(0.8)}, {'gameTime': f'{half_id} - {int((75 / 60)):02d}:{int((75 % 60))}', 'label': REVERSE_ACTION_SPOTTING_LABELS[1], 'position': f'{int(75000)}', 'half': str(half_id), 'confidence': float(0.9)}] for (expected, actual) in zip(expected_half_predictions, half_predictions): for key in expected: if (type(expected[key]) is float): assert math.isclose(expected[key], actual[key], rel_tol=1e-06, abs_tol=0.0) else: assert (expected[key] == actual[key])
class TestSCETokenMasks(unittest.TestCase): def test_perform_hard_NMS(self): values = torch.tensor([0.5, 1.0, 0.2, 0.3, 0.1, 0.2, 0.6]) window = 3 threshold = 0.49 keep_indexes = perform_hard_NMS(values, window, threshold) expected_keep_indexes = torch.tensor([False, True, False, False, False, False, True]) assert torch.allclose(keep_indexes, expected_keep_indexes) def test_perform_hard_NMS_longer_window(self): values = torch.tensor([0.5, 1.0, 0.2, 0.3, 0.1, 0.2, 0.6]) window = 14 threshold = 0.49 keep_indexes = perform_hard_NMS(values, window, threshold) expected_keep_indexes = torch.tensor([False, True, False, False, False, False, False]) assert torch.allclose(keep_indexes, expected_keep_indexes) def test_perform_all_classes_hard_NMS(self): values = torch.tensor([[0.5, 0.3], [1.0, 0.6], [0.2, 0.8], [0.3, 0.2], [0.1, 0.1], [0.2, 0.9], [0.6, 0.8]]) window = 3 threshold = 0.49 step_timestamp = 1.0 (kept_values, kept_timestamps_per_class) = perform_all_classes_NMS(values, step_timestamp, window, threshold, nms_type='hard') expected_kept_values = [torch.tensor([1.0, 0.6]), torch.tensor([0.8, 0.9])] expected_kept_timestamps_per_class = [torch.tensor([1.0, 6.0]), torch.tensor([2.0, 5.0])] assert all([torch.allclose(kept_value, expected_kept_value) for (kept_value, expected_kept_value) in zip(kept_values, expected_kept_values)]) assert all([torch.allclose(timestamp_per_class, expected_timestamp_per_class) for (timestamp_per_class, expected_timestamp_per_class) in zip(kept_timestamps_per_class, expected_kept_timestamps_per_class)]) def test_perform_all_classes_hard_NMS_step_timestamp(self): values = torch.tensor([[0.5, 0.3], [1.0, 0.6], [0.2, 0.8], [0.3, 0.2], [0.1, 0.1], [0.2, 0.9], [0.6, 0.8]]) window = 3 threshold = 0.49 step_timestamp = 0.5 (kept_values, kept_timestamps_per_class) = perform_all_classes_NMS(values, step_timestamp, window, threshold, nms_type='hard') expected_kept_values = [torch.tensor([1.0, 0.6]), torch.tensor([0.8, 0.9])] expected_kept_timestamps_per_class = [torch.tensor([0.5, 3.0]), torch.tensor([1.0, 2.5])] assert all([torch.allclose(kept_value, expected_kept_value) for (kept_value, expected_kept_value) in zip(kept_values, expected_kept_values)]) assert all([torch.allclose(timestamp_per_class, expected_timestamp_per_class) for (timestamp_per_class, expected_timestamp_per_class) in zip(kept_timestamps_per_class, expected_kept_timestamps_per_class)]) def test_perform_soft_NMS(self): values = torch.tensor([0.0001, 1.0, 0.2, 0.3, 0.1, 0.2, 0.6]) window = 3 threshold = 0.001 decayed_values = perform_soft_NMS(values, window, threshold) print(decayed_values) def test_perform_all_classes_soft_NMS_step_timestamp(self): values = torch.tensor([[0.5, 0.3], [1.0, 0.6], [0.2, 0.8], [0.3, 0.2], [0.1, 0.1], [0.2, 0.9], [0.6, 0.8]]) window = 3 threshold = 0.2 step_timestamp = 0.5 (kept_values, kept_timestamps_per_class) = perform_all_classes_NMS(values, step_timestamp, window, threshold, nms_type='soft') print(kept_values, kept_timestamps_per_class)
class BoringDataModule(LightningDataModule): def __init__(self, data_dir: str='./', dataset=RandomDataset((32, (64 * 4))), val_dataset=RandomDataset((32, (64 * 4))), batch_size: int=1): super().__init__() self.data_dir = data_dir self.non_picklable = None self.checkpoint_state: Optional[str] = None self.dataset = dataset self.val_dataset = val_dataset self.batch_size = batch_size @property def train_num_samples(self): return len(self.dataset) @property def val_num_samples(self): return len(self.val_dataset) @property def train_global_batch_size(self) -> int: return self.batch_size @property def val_global_batch_size(self) -> int: return self.batch_size @property def train_local_batch_size(self) -> int: return self.batch_size @property def val_local_batch_size(self) -> int: return self.batch_size def train_dataloader(self): return DataLoader(self.dataset, batch_size=self.batch_size, drop_last=True) def val_dataloader(self): return DataLoader(self.val_dataset, batch_size=self.batch_size)
class RandomDataset(Dataset): def __init__(self, size: Iterable[int]): self.length = size[0] self.data = torch.randn(size) def __getitem__(self, index): return {'input': self.data[index], 'idx': index} def __len__(self): return self.length
class RandomLabeledDataset(Dataset): def __init__(self, size: Iterable[int], num_classes: int=10): self.length = size[0] self.data = torch.randn(size) self.labels = torch.randint(num_classes, size=(size[0], 1)) def __getitem__(self, index): return {'input': self.data[index], 'label': self.labels[index], 'idx': index} def __len__(self): return self.length
class RandomVisionLabeledDataset(VisionDataset): def __init__(self, size: Iterable[int], num_classes: int=10, transform: Optional[Module]=None): super().__init__('data/', transform=transform) self.length = size[0] self.data = torch.randn(size) self.labels = torch.randint(num_classes, size=(size[0], 1)) def __getitem__(self, index): data = self.data[index] if (self.transform is not None): data = self.transform(data) return {'input': data, 'label': self.labels[index]} def __len__(self): return self.length
class BoringModel(LightningModule): def __init__(self): 'Testing PL Module. Use as follows:\n\n - subclass\n - modify the behavior for what you want\n class TestModel(BaseTestModel):\n def training_step(...):\n # do your own thing\n or:\n model = BaseTestModel()\n model.on_train_epoch_end = None\n ' super().__init__() self.layer = torch.nn.Linear(32, 2) @property def num_layers(self) -> int: return 1 def get_param_layer_id(self, name: str) -> int: return 0 def forward(self, x): return self.layer(x) def loss(self, batch, prediction): return torch.nn.functional.mse_loss(prediction, torch.ones_like(prediction)) def step(self, x): x = self(x) out = torch.nn.functional.mse_loss(x, torch.ones_like(x)) return out def training_step(self, batch, batch_idx): output = self(batch) loss = self.loss(batch, output) return {'loss': loss} def on_train_batch_end(self, training_step_outputs): return training_step_outputs def on_train_epoch_end(self, outputs) -> None: torch.stack([x['loss'] for x in outputs]).mean() def validation_step(self, batch, batch_idx): output = self(batch) loss = self.loss(batch, output) return {'x': loss} def on_validation_epoch_end(self, outputs) -> None: torch.stack([x['x'] for x in outputs]).mean() def test_step(self, batch, batch_idx): output = self(batch) loss = self.loss(batch, output) return {'y': loss} def on_test_epoch_end(self, outputs) -> None: torch.stack([x['y'] for x in outputs]).mean() def configure_optimizers(self): optimizer = torch.optim.SGD(self.layer.parameters(), lr=0.1) lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=1) return ([optimizer], [lr_scheduler]) def train_dataloader(self): return DataLoader(RandomDataset(32, 64)) def val_dataloader(self): return DataLoader(RandomDataset(32, 64)) def test_dataloader(self): return DataLoader(RandomDataset(32, 64)) def predict_dataloader(self): return DataLoader(RandomDataset(32, 64))
class LargeBoringModel(LightningModule): def __init__(self): 'Testing PL Module. Use as follows:\n\n - subclass\n - modify the behavior for what you want\n class TestModel(BaseTestModel):\n def training_step(...):\n # do your own thing\n or:\n model = BaseTestModel()\n model.on_train_epoch_end = None\n ' super().__init__() self.layer1 = torch.nn.Linear(32, 32, bias=False) self.bn1 = torch.nn.BatchNorm1d(32) self.layer2 = torch.nn.Linear(32, 32, bias=False) self.bn2 = torch.nn.BatchNorm1d(32) self.layer3 = torch.nn.Linear(32, 32, bias=False) self.bn3 = torch.nn.BatchNorm1d(32) self.layer4 = torch.nn.Linear(32, 2, bias=True) self.useless_layer = torch.nn.Linear(32, 2) @property def learnable_params(self) -> List[Parameter]: params = [param for layer in [self.layer1, self.bn1, self.layer2, self.bn2, self.layer3, self.bn3, self.layer4] for param in layer.parameters()] return params @property def num_layers(self) -> int: 'Number of layers of the model.' return 4 def get_param_layer_id(self, name: str) -> int: 'Get the layer id of the named parameter.\n\n Args:\n name: The name of the parameter.\n ' if (name.startswith('layer1') or name.startswith('bn1')): return 0 elif (name.startswith('layer2') or name.startswith('bn2')): return 1 elif (name.startswith('layer3') or name.startswith('bn3')): return 2 elif name.startswith('layer4'): return 3 elif name.startswith('useless_layer'): return 3 def forward(self, x): x = self.layer1(x) x = self.bn1(x) x = self.layer2(x) x = self.bn2(x) x = self.layer3(x) x = self.bn3(x) return x def loss(self, batch, prediction): return torch.nn.functional.mse_loss(prediction, torch.ones_like(prediction)) def step(self, x): x = self(x) out = torch.nn.functional.mse_loss(x, torch.ones_like(x)) return out def training_step(self, batch, batch_idx): output = self(batch) loss = self.loss(batch, output) return {'loss': loss} def on_train_batch_end(self, training_step_outputs): return training_step_outputs def on_train_epoch_end(self, outputs) -> None: torch.stack([x['loss'] for x in outputs]).mean() def validation_step(self, batch, batch_idx): output = self(batch) loss = self.loss(batch, output) return {'x': loss} def on_validation_epoch_end(self, outputs) -> None: torch.stack([x['x'] for x in outputs]).mean() def test_step(self, batch, batch_idx): output = self(batch) loss = self.loss(batch, output) return {'y': loss} def on_test_epoch_end(self, outputs) -> None: torch.stack([x['y'] for x in outputs]).mean() def configure_optimizers(self): optimizer = torch.optim.SGD(self.layer.parameters(), lr=0.1) lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=1) return ([optimizer], [lr_scheduler]) def train_dataloader(self): return DataLoader(RandomDataset(32, 64)) def val_dataloader(self): return DataLoader(RandomDataset(32, 64)) def test_dataloader(self): return DataLoader(RandomDataset(32, 64)) def predict_dataloader(self): return DataLoader(RandomDataset(32, 64))
class ManualOptimBoringModel(BoringModel): def __init__(self): super().__init__() self.automatic_optimization = False def training_step(self, batch, batch_idx): opt = self.optimizers() output = self(batch) loss = self.loss(batch, output) opt.zero_grad() self.manual_backward(loss) opt.step() return loss
class TestSCELoss(unittest.TestCase): def setUp(self) -> None: self.coeff = 0.5 self.temp = 0.1 self.temp_m = 0.07 def test_sce_loss_without_key(self): q = torch.arange(1.0, 9.0, 1.0).view((4, 2)) k = torch.tensor([[0, 0], [1, 1], [0, 0], [1.0, 1.0]]) queue = torch.tensor([[0.0, 0, 0, 2], [0.0, 0, 0, 2.0]]) sim_pos = torch.tensor([0.0, 7, 0, 15]).unsqueeze((- 1)) sim_qqueue = torch.Tensor([[0.0, 0, 0, 6], [0, 0, 0, 14], [0, 0, 0, 22], [0, 0, 0, 30]]) sim_kqueue = torch.Tensor([[0.0, 0, 0, 0], [0, 0, 0, 4], [0, 0, 0, 0], [0, 0, 0, 4]]) expected_mask = torch.Tensor([[1.0, 0, 0, 0, 0], [1.0, 0, 0, 0, 0], [1.0, 0, 0, 0, 0], [1.0, 0, 0, 0, 0]]) sim_q = torch.cat((sim_pos, sim_qqueue), 1) sim_k = torch.cat((torch.tensor([0.0, 0, 0, 0]).unsqueeze((- 1)), sim_kqueue), 1) logits_q = (sim_q / self.temp) logits_k = (sim_k / self.temp_m) prob_k = nn.functional.softmax(logits_k, dim=1) prob_q = nn.functional.normalize(((self.coeff * expected_mask) + ((1 - self.coeff) * prob_k)), p=1, dim=1) expected_loss = (- torch.sum((prob_q * nn.functional.log_softmax(logits_q, dim=1)), dim=1).mean(dim=0)) mask = compute_sce_mask(4, 4, False, 0, 1, 'cuda') loss = compute_sce_loss(q, k, k, False, queue, mask, self.coeff, self.temp, self.temp_m) assert torch.equal(expected_mask, mask) assert torch.equal(expected_loss, loss) def test_sce_loss_with_key(self): q = torch.arange(1.0, 9.0, 1.0).view((4, 2)) k = torch.tensor([[0, 0], [1, 1], [0, 0], [1.0, 1.0]]) queue = torch.tensor([[0.0, 0, 0, 2], [0.0, 0, 0, 2.0]]) sim_qk = torch.tensor([[0, 3, 0, 3], [0, 7, 0, 7], [0, 11, 0, 11], [0, 15.0, 0, 15]]) sim_kk = torch.tensor([[0, 0, 0, 0], [0, 2, 0, 2], [0, 0, 0, 0], [0, 2.0, 0, 2]]) sim_qqueue = torch.Tensor([[0.0, 0, 0, 6], [0, 0, 0, 14], [0, 0, 0, 22], [0, 0, 0, 30]]) sim_kqueue = torch.Tensor([[0.0, 0, 0, 0], [0, 0, 0, 4], [0, 0, 0, 0], [0, 0, 0, 4]]) expected_mask = torch.tensor([[1.0, 0, 0, 0, 0, 0, 0, 0], [0, 1.0, 0, 0, 0, 0, 0, 0], [0, 0, 1.0, 0, 0, 0, 0, 0], [0, 0, 0, 1.0, 0, 0, 0, 0]]) sim_q = torch.cat([sim_qk, sim_qqueue], dim=1) sim_k = torch.cat([sim_kk, sim_kqueue], dim=1) sim_k -= (1000000000.0 * expected_mask) logits_q = (sim_q / self.temp) logits_k = (sim_k / self.temp_m) prob_k = nn.functional.softmax(logits_k, dim=1) prob_q = nn.functional.normalize(((self.coeff * expected_mask) + ((1 - self.coeff) * prob_k)), p=1, dim=1) expected_loss = (- torch.sum((prob_q * nn.functional.log_softmax(logits_q, dim=1)), dim=1).mean(dim=0)) mask = compute_sce_mask(4, 4, True, 0, 1, 'cuda') loss = compute_sce_loss(q, k, k, True, queue, mask, self.coeff, self.temp, self.temp_m) assert torch.equal(expected_mask, mask) assert torch.equal(expected_loss, loss) def test_sce_loss_with_key_without_queue(self): q = torch.arange(1.0, 9.0, 1.0).view((4, 2)) k = torch.tensor([[0, 0], [1, 1], [0, 0], [1.0, 1.0]]) sim_qk = torch.tensor([[0, 3, 0, 3], [0, 7, 0, 7], [0, 11, 0, 11], [0, 15.0, 0, 15]]) sim_kk = torch.tensor([[0, 0, 0, 0], [0, 2, 0, 2], [0, 0, 0, 0], [0, 2.0, 0, 2]]) expected_mask = torch.tensor([[1.0, 0, 0, 0], [0, 1.0, 0, 0], [0, 0, 1.0, 0], [0, 0, 0, 1.0]]) sim_q = sim_qk sim_k = sim_kk logits_q = (sim_q / self.temp) logits_k = (sim_k / self.temp_m) logits_k -= (1000000000.0 * expected_mask) prob_k = nn.functional.softmax(logits_k, dim=1) prob_q = nn.functional.normalize(((self.coeff * expected_mask) + ((1 - self.coeff) * prob_k)), p=1, dim=1) expected_loss = (- torch.sum((prob_q * nn.functional.log_softmax(logits_q, dim=1)), dim=1).mean(dim=0)) mask = compute_sce_mask(4, 0, True, 0, 1, 'cuda') loss = compute_sce_loss(q, k, k, True, None, mask, self.coeff, self.temp, self.temp_m) assert torch.equal(expected_mask, mask) assert torch.equal(expected_loss, loss)
class TestSCETokenMasks(unittest.TestCase): def setUp(self) -> None: self.batch_size = 2 self.num_tokens = 8 self.num_negatives = 2 def test_one_device_zero_pos_radius_no_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=0, keep_aligned_positive=True, use_keys=False, use_all_keys=False, rank=0, world_size=1) expected_mask_prob_q = torch.tensor([[1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0]], device=mask_prob_q.device, dtype=mask_prob_q.dtype) expected_mask_sim_q = torch.tensor([[0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0]], device=mask_sim_q.device, dtype=mask_sim_q.dtype) expected_mask_sim_k = torch.tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], device=mask_sim_k.device, dtype=mask_sim_k.dtype) expected_num_positives = torch.tensor([[1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1]], device=num_positives_per_token.device, dtype=num_positives_per_token.dtype) expected_mask_log_q = torch.cat(((1 - mask_sim_q), torch.ones((mask_sim_q.shape[0], self.num_negatives), device=mask_sim_q.device, dtype=mask_sim_q.dtype)), 1).to(dtype=torch.bool) assert torch.allclose(mask_prob_q, expected_mask_prob_q) assert torch.allclose(mask_sim_k, expected_mask_sim_k) assert torch.allclose(mask_sim_q, expected_mask_sim_q) assert torch.allclose(mask_log_q, expected_mask_log_q) assert torch.allclose(num_positives_per_token, expected_num_positives) def test_one_device_two_pos_radius_no_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=True, use_keys=False, use_all_keys=False, rank=0, world_size=1) expected_mask_prob_q = torch.tensor([[1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0]], device=mask_prob_q.device, dtype=mask_prob_q.dtype) expected_mask_sim_q = torch.tensor([[0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0]], device=mask_sim_q.device, dtype=mask_sim_q.dtype) expected_mask_sim_k = torch.tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], device=mask_sim_k.device, dtype=mask_sim_k.dtype) expected_num_positives = torch.tensor([[3], [4], [5], [5], [5], [5], [4], [3], [3], [4], [5], [5], [5], [5], [4], [3]], device=num_positives_per_token.device, dtype=num_positives_per_token.dtype) expected_mask_log_q = torch.cat(((1 - mask_sim_q), torch.ones((mask_sim_q.shape[0], self.num_negatives), device=mask_sim_q.device, dtype=mask_sim_q.dtype)), 1).to(dtype=torch.bool) assert torch.allclose(mask_prob_q, (expected_mask_prob_q / expected_num_positives)) assert torch.allclose(mask_sim_k, expected_mask_sim_k) assert torch.allclose(mask_sim_q, expected_mask_sim_q) assert torch.allclose(mask_log_q, expected_mask_log_q) assert torch.allclose(num_positives_per_token, expected_num_positives) def test_one_device_two_pos_radius_no_keys_not_aligned_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=False, use_keys=False, use_all_keys=False, rank=0, world_size=1) expected_mask_prob_q = torch.tensor([[0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0]], device=mask_prob_q.device, dtype=mask_prob_q.dtype) expected_mask_sim_q = torch.tensor([[1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 0, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 0, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 0, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 0, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1]], device=mask_sim_q.device, dtype=mask_sim_q.dtype) expected_mask_sim_k = torch.tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], device=mask_sim_k.device, dtype=mask_sim_k.dtype) expected_num_positives = torch.tensor([[2], [3], [4], [4], [4], [4], [3], [2], [2], [3], [4], [4], [4], [4], [3], [2]], device=num_positives_per_token.device, dtype=num_positives_per_token.dtype) expected_mask_log_q = torch.cat(((1 - mask_sim_q), torch.ones((mask_sim_q.shape[0], self.num_negatives), device=mask_sim_q.device, dtype=mask_sim_q.dtype)), 1).to(dtype=torch.bool) assert torch.allclose(mask_prob_q, (expected_mask_prob_q / expected_num_positives)) assert torch.allclose(mask_sim_k, expected_mask_sim_k) assert torch.allclose(mask_sim_q, expected_mask_sim_q) assert torch.allclose(mask_log_q, expected_mask_log_q) assert torch.allclose(num_positives_per_token, expected_num_positives) def test_several_devices_zero_pos_radius_no_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=0, keep_aligned_positive=True, use_keys=False, use_all_keys=False, rank=1, world_size=3) expected_mask_prob_q = torch.tensor([[1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0]], device=mask_prob_q.device, dtype=mask_prob_q.dtype) expected_mask_sim_q = torch.tensor([[0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0]], device=mask_sim_q.device, dtype=mask_sim_q.dtype) expected_mask_sim_k = torch.tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], device=mask_sim_k.device, dtype=mask_sim_k.dtype) expected_num_positives = torch.tensor([[1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1]], device=num_positives_per_token.device, dtype=num_positives_per_token.dtype) expected_mask_log_q = torch.cat(((1 - mask_sim_q), torch.ones((mask_sim_q.shape[0], self.num_negatives), device=mask_sim_q.device, dtype=mask_sim_q.dtype)), 1).to(dtype=torch.bool) assert torch.allclose(mask_prob_q, expected_mask_prob_q) assert torch.allclose(mask_sim_k, expected_mask_sim_k) assert torch.allclose(mask_sim_q, expected_mask_sim_q) assert torch.allclose(mask_log_q, expected_mask_log_q) assert torch.allclose(num_positives_per_token, expected_num_positives) def test_several_devices_two_pos_radius_no_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=True, use_keys=False, use_all_keys=False, rank=1, world_size=3) expected_mask_prob_q = torch.tensor([[1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0]], device=mask_prob_q.device, dtype=mask_prob_q.dtype) expected_mask_sim_q = torch.tensor([[0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0]], device=mask_sim_q.device, dtype=mask_sim_q.dtype) expected_mask_sim_k = torch.tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], device=mask_sim_k.device, dtype=mask_sim_k.dtype) expected_num_positives = torch.tensor([[3], [4], [5], [5], [5], [5], [4], [3], [3], [4], [5], [5], [5], [5], [4], [3]], device=num_positives_per_token.device, dtype=num_positives_per_token.dtype) expected_mask_log_q = torch.cat(((1 - mask_sim_q), torch.ones((mask_sim_q.shape[0], self.num_negatives), device=mask_sim_q.device, dtype=mask_sim_q.dtype)), 1).to(dtype=torch.bool) assert torch.allclose(mask_prob_q, (expected_mask_prob_q / expected_num_positives)) assert torch.allclose(mask_sim_k, expected_mask_sim_k) assert torch.allclose(mask_sim_q, expected_mask_sim_q) assert torch.allclose(mask_log_q, expected_mask_log_q) assert torch.allclose(num_positives_per_token, expected_num_positives) def test_several_devices_two_pos_radius_no_keys_not_aligned_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=False, use_keys=False, use_all_keys=False, rank=1, world_size=3) expected_mask_prob_q = torch.tensor([[0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0]], device=mask_prob_q.device, dtype=mask_prob_q.dtype) expected_mask_sim_q = torch.tensor([[1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 0, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 0, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 0, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 0, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1]], device=mask_sim_q.device, dtype=mask_sim_q.dtype) expected_mask_sim_k = torch.tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], device=mask_sim_k.device, dtype=mask_sim_k.dtype) expected_num_positives = torch.tensor([[2], [3], [4], [4], [4], [4], [3], [2], [2], [3], [4], [4], [4], [4], [3], [2]], device=num_positives_per_token.device, dtype=num_positives_per_token.dtype) expected_mask_log_q = torch.cat(((1 - mask_sim_q), torch.ones((mask_sim_q.shape[0], self.num_negatives), device=mask_sim_q.device, dtype=mask_sim_q.dtype)), 1).to(dtype=torch.bool) assert torch.allclose(mask_prob_q, (expected_mask_prob_q / expected_num_positives)) assert torch.allclose(mask_sim_k, expected_mask_sim_k) assert torch.allclose(mask_sim_q, expected_mask_sim_q) assert torch.allclose(mask_log_q, expected_mask_log_q) assert torch.allclose(num_positives_per_token, expected_num_positives) def test_one_device_zero_pos_radius_with_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=0, keep_aligned_positive=True, use_keys=True, use_all_keys=False, rank=0, world_size=1) expected_mask_prob_q = torch.tensor([[1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0]], device=mask_prob_q.device, dtype=mask_prob_q.dtype) expected_mask_sim_q = torch.tensor([[0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]], device=mask_sim_q.device, dtype=mask_sim_q.dtype) expected_mask_sim_k = torch.tensor([[1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 1, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 1, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 1]], device=mask_sim_k.device, dtype=mask_sim_k.dtype) expected_num_positives = torch.tensor([[1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1]], device=num_positives_per_token.device, dtype=num_positives_per_token.dtype) expected_mask_log_q = torch.cat(((1 - mask_sim_q), torch.ones((mask_sim_q.shape[0], self.num_negatives), device=mask_sim_q.device, dtype=mask_sim_q.dtype)), 1).to(dtype=torch.bool) assert torch.allclose(mask_prob_q, expected_mask_prob_q) assert torch.allclose(mask_sim_k, expected_mask_sim_k) assert torch.allclose(mask_sim_q, expected_mask_sim_q) assert torch.allclose(mask_log_q, expected_mask_log_q) assert torch.allclose(num_positives_per_token, expected_num_positives) def test_one_device_two_pos_radius_with_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=True, use_keys=True, use_all_keys=False, rank=0, world_size=1) expected_mask_prob_q = torch.tensor([[1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0]], device=mask_prob_q.device, dtype=mask_prob_q.dtype) expected_mask_sim_q = torch.tensor([[0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]], device=mask_sim_q.device, dtype=mask_sim_q.dtype) expected_mask_sim_k = torch.tensor([[1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1]], device=mask_sim_k.device, dtype=mask_sim_k.dtype) expected_num_positives = torch.tensor([[3], [4], [5], [5], [5], [5], [4], [3], [3], [4], [5], [5], [5], [5], [4], [3]], device=num_positives_per_token.device, dtype=num_positives_per_token.dtype) expected_mask_log_q = torch.cat(((1 - mask_sim_q), torch.ones((mask_sim_q.shape[0], self.num_negatives), device=mask_sim_q.device, dtype=mask_sim_q.dtype)), 1).to(dtype=torch.bool) assert torch.allclose(mask_prob_q, (expected_mask_prob_q / expected_num_positives)) assert torch.allclose(mask_sim_k, expected_mask_sim_k) assert torch.allclose(mask_sim_q, expected_mask_sim_q) assert torch.allclose(mask_log_q, expected_mask_log_q) assert torch.allclose(num_positives_per_token, expected_num_positives) def test_one_device_two_pos_radius_with_keys_not_aligned_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=False, use_keys=True, use_all_keys=False, rank=0, world_size=1) expected_mask_prob_q = torch.tensor([[0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0]], device=mask_prob_q.device, dtype=mask_prob_q.dtype) expected_mask_sim_q = torch.tensor([[1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 1, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 1, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 1]], device=mask_sim_q.device, dtype=mask_sim_q.dtype) expected_mask_sim_k = torch.tensor([[1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1]], device=mask_sim_k.device, dtype=mask_sim_k.dtype) expected_num_positives = torch.tensor([[2], [3], [4], [4], [4], [4], [3], [2], [2], [3], [4], [4], [4], [4], [3], [2]], device=num_positives_per_token.device, dtype=num_positives_per_token.dtype) expected_mask_log_q = torch.cat(((1 - mask_sim_q), torch.ones((mask_sim_q.shape[0], self.num_negatives), device=mask_sim_q.device, dtype=mask_sim_q.dtype)), 1).to(dtype=torch.bool) assert torch.allclose(mask_prob_q, (expected_mask_prob_q / expected_num_positives)) assert torch.allclose(mask_sim_k, expected_mask_sim_k) assert torch.allclose(mask_sim_q, expected_mask_sim_q) assert torch.allclose(mask_log_q, expected_mask_log_q) assert torch.allclose(num_positives_per_token, expected_num_positives) def test_several_devices_zero_pos_radius_with_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=0, keep_aligned_positive=True, use_keys=True, use_all_keys=False, rank=1, world_size=3) expected_mask_prob_q = torch.tensor([[1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0]], device=mask_prob_q.device, dtype=mask_prob_q.dtype) expected_mask_sim_q = torch.tensor([[0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]], device=mask_sim_q.device, dtype=mask_sim_q.dtype) expected_mask_sim_k = torch.tensor([[1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 1, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 1, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 1]], device=mask_sim_k.device, dtype=mask_sim_k.dtype) expected_num_positives = torch.tensor([[1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1]], device=num_positives_per_token.device, dtype=num_positives_per_token.dtype) expected_mask_log_q = torch.cat(((1 - mask_sim_q), torch.ones((mask_sim_q.shape[0], self.num_negatives), device=mask_sim_q.device, dtype=mask_sim_q.dtype)), 1).to(dtype=torch.bool) assert torch.allclose(mask_prob_q, expected_mask_prob_q) assert torch.allclose(mask_sim_k, expected_mask_sim_k) assert torch.allclose(mask_sim_q, expected_mask_sim_q) assert torch.allclose(mask_log_q, expected_mask_log_q) assert torch.allclose(num_positives_per_token, expected_num_positives) def test_several_devices_two_pos_radius_with_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=True, use_keys=True, use_all_keys=False, rank=1, world_size=3) expected_mask_prob_q = torch.tensor([[1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0]], device=mask_prob_q.device, dtype=mask_prob_q.dtype) expected_mask_sim_q = torch.tensor([[0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]], device=mask_sim_q.device, dtype=mask_sim_q.dtype) expected_mask_sim_k = torch.tensor([[1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1]], device=mask_sim_k.device, dtype=mask_sim_k.dtype) expected_num_positives = torch.tensor([[3], [4], [5], [5], [5], [5], [4], [3], [3], [4], [5], [5], [5], [5], [4], [3]], device=num_positives_per_token.device, dtype=num_positives_per_token.dtype) expected_mask_log_q = torch.cat(((1 - mask_sim_q), torch.ones((mask_sim_q.shape[0], self.num_negatives), device=mask_sim_q.device, dtype=mask_sim_q.dtype)), 1).to(dtype=torch.bool) assert torch.allclose(mask_prob_q, (expected_mask_prob_q / expected_num_positives)) assert torch.allclose(mask_sim_k, expected_mask_sim_k) assert torch.allclose(mask_sim_q, expected_mask_sim_q) assert torch.allclose(mask_log_q, expected_mask_log_q) assert torch.allclose(num_positives_per_token, expected_num_positives) def test_several_devices_two_pos_radius_with_keys_not_aligned_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=False, use_keys=True, use_all_keys=False, rank=1, world_size=3) expected_mask_prob_q = torch.tensor([[0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0]], device=mask_prob_q.device, dtype=mask_prob_q.dtype) expected_mask_sim_q = torch.tensor([[1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 1, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 1, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 1]], device=mask_sim_q.device, dtype=mask_sim_q.dtype) expected_mask_sim_k = torch.tensor([[1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 0], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1]], device=mask_sim_k.device, dtype=mask_sim_k.dtype) expected_num_positives = torch.tensor([[2], [3], [4], [4], [4], [4], [3], [2], [2], [3], [4], [4], [4], [4], [3], [2]], device=num_positives_per_token.device, dtype=num_positives_per_token.dtype) expected_mask_log_q = torch.cat(((1 - mask_sim_q), torch.ones((mask_sim_q.shape[0], self.num_negatives), device=mask_sim_q.device, dtype=mask_sim_q.dtype)), 1).to(dtype=torch.bool) assert torch.allclose(mask_prob_q, (expected_mask_prob_q / expected_num_positives)) assert torch.allclose(mask_sim_k, expected_mask_sim_k) assert torch.allclose(mask_sim_q, expected_mask_sim_q) assert torch.allclose(mask_log_q, expected_mask_log_q) assert torch.allclose(num_positives_per_token, expected_num_positives) def test_one_device_zero_pos_radius_with_all_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=0, keep_aligned_positive=True, use_keys=False, use_all_keys=True, rank=0, world_size=1) expected_mask_prob_q = torch.tensor([[1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0]], device=mask_prob_q.device, dtype=mask_prob_q.dtype) expected_mask_sim_k = torch.tensor([[1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1]], device=mask_sim_k.device, dtype=mask_sim_k.dtype) expected_num_positives = torch.tensor([[1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1]], device=num_positives_per_token.device, dtype=num_positives_per_token.dtype) assert torch.allclose(mask_prob_q, expected_mask_prob_q) assert torch.allclose(mask_sim_k, expected_mask_sim_k) assert (mask_sim_q is None) assert (mask_log_q is None) assert torch.allclose(num_positives_per_token, expected_num_positives) def test_one_device_two_pos_radius_with_all_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=True, use_keys=False, use_all_keys=True, rank=0, world_size=1) expected_mask_prob_q = torch.tensor([[1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0]], device=mask_prob_q.device, dtype=mask_prob_q.dtype) expected_mask_sim_k = torch.tensor([[1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1]], device=mask_sim_k.device, dtype=mask_sim_k.dtype) expected_num_positives = torch.tensor([[3], [4], [5], [5], [5], [5], [4], [3], [3], [4], [5], [5], [5], [5], [4], [3]], device=num_positives_per_token.device, dtype=num_positives_per_token.dtype) assert torch.allclose(mask_prob_q, (expected_mask_prob_q / expected_num_positives)) assert torch.allclose(mask_sim_k, expected_mask_sim_k) assert (mask_sim_q is None) assert (mask_log_q is None) assert torch.allclose(num_positives_per_token, expected_num_positives) def test_one_device_two_pos_radius_with_all_keys_not_aligned_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=False, use_keys=False, use_all_keys=True, rank=0, world_size=1) expected_mask_prob_q = torch.tensor([[0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0]], device=mask_prob_q.device, dtype=mask_prob_q.dtype) expected_mask_sim_q = torch.tensor([[1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1]], device=mask_sim_q.device, dtype=mask_sim_q.dtype) expected_mask_sim_k = torch.tensor([[1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1]], device=mask_sim_k.device, dtype=mask_sim_k.dtype) expected_num_positives = torch.tensor([[2], [3], [4], [4], [4], [4], [3], [2], [2], [3], [4], [4], [4], [4], [3], [2]], device=num_positives_per_token.device, dtype=num_positives_per_token.dtype) expected_mask_log_q = torch.cat(((1 - mask_sim_q), torch.ones((mask_sim_q.shape[0], self.num_negatives), device=mask_sim_q.device, dtype=mask_sim_q.dtype)), 1).to(dtype=torch.bool) assert torch.allclose(mask_prob_q, (expected_mask_prob_q / expected_num_positives)) assert torch.allclose(mask_sim_k, expected_mask_sim_k) assert torch.allclose(mask_sim_q, expected_mask_sim_q) assert torch.allclose(mask_log_q, expected_mask_log_q) assert torch.allclose(num_positives_per_token, expected_num_positives) def test_several_devices_zero_pos_radius_with_all_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=0, keep_aligned_positive=True, use_keys=False, use_all_keys=True, rank=1, world_size=3) zeros_other_device = torch.zeros(((self.batch_size * self.num_tokens), (self.batch_size * self.num_tokens))) expected_mask_prob_q = torch.tensor([[1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0]], device=mask_prob_q.device, dtype=mask_prob_q.dtype) expected_mask_prob_q = torch.cat((zeros_other_device, expected_mask_prob_q, zeros_other_device), dim=1) expected_mask_sim_k = torch.tensor([[1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1]], device=mask_sim_k.device, dtype=mask_sim_k.dtype) expected_mask_sim_k = torch.cat((zeros_other_device, expected_mask_sim_k, zeros_other_device), dim=1) expected_num_positives = torch.tensor([[1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1], [1]], device=num_positives_per_token.device, dtype=num_positives_per_token.dtype) assert torch.allclose(mask_prob_q, expected_mask_prob_q) assert torch.allclose(mask_sim_k, expected_mask_sim_k) assert (mask_sim_q is None) assert (mask_log_q is None) assert torch.allclose(num_positives_per_token, expected_num_positives) def test_several_devices_two_pos_radius_with_all_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=True, use_keys=False, use_all_keys=True, rank=1, world_size=3) zeros_other_device = torch.zeros(((self.batch_size * self.num_tokens), (self.batch_size * self.num_tokens))) expected_mask_prob_q = torch.tensor([[1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0]], device=mask_prob_q.device, dtype=mask_prob_q.dtype) expected_mask_prob_q = torch.cat((zeros_other_device, expected_mask_prob_q, zeros_other_device), dim=1) expected_mask_sim_k = torch.tensor([[1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1]], device=mask_sim_k.device, dtype=mask_sim_k.dtype) expected_mask_sim_k = torch.cat((zeros_other_device, expected_mask_sim_k, zeros_other_device), dim=1) expected_num_positives = torch.tensor([[3], [4], [5], [5], [5], [5], [4], [3], [3], [4], [5], [5], [5], [5], [4], [3]], device=num_positives_per_token.device, dtype=num_positives_per_token.dtype) assert torch.allclose(mask_prob_q, (expected_mask_prob_q / expected_num_positives)) assert torch.allclose(mask_sim_k, expected_mask_sim_k) assert (mask_sim_q is None) assert (mask_log_q is None) assert torch.allclose(num_positives_per_token, expected_num_positives) def test_with_keys_and_all_keys(self): try: (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=False, use_keys=True, use_all_keys=True, rank=1, world_size=3) except NotImplementedError: return else: assert False
class TestSCETokenLoss(unittest.TestCase): def setUp(self) -> None: self.batch_size = 2 self.num_tokens = 8 self.num_negatives = 2 self.dim = 4 self.query = torch.randn(((self.batch_size * self.num_tokens), self.dim)) self.key = torch.randn(((self.batch_size * self.num_tokens), self.dim)) self.global_key = torch.randn((((3 * self.batch_size) * self.num_tokens), self.dim)) self.queue = torch.randn((self.dim, self.num_negatives)) def test_one_device_zero_pos_radius_no_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=0, keep_aligned_positive=True, use_keys=False, use_all_keys=False, rank=0, world_size=1) compute_sce_token_loss(self.query, self.key, self.key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5)) def test_one_device_two_pos_radius_no_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=True, use_keys=False, use_all_keys=False, rank=0, world_size=1) compute_sce_token_loss(self.query, self.key, self.key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5)) def test_one_device_two_pos_radius_no_keys_not_aligned_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=False, use_keys=False, use_all_keys=False, rank=0, world_size=1) compute_sce_token_loss(self.query, self.key, self.key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5)) def test_several_devices_zero_pos_radius_no_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=0, keep_aligned_positive=True, use_keys=False, use_all_keys=False, rank=1, world_size=3) compute_sce_token_loss(self.query, self.key, self.key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5)) def test_several_devices_two_pos_radius_no_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=True, use_keys=False, use_all_keys=False, rank=1, world_size=3) compute_sce_token_loss(self.query, self.key, self.key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5)) def test_several_devices_two_pos_radius_no_keys_not_aligned_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=False, use_keys=False, use_all_keys=False, rank=1, world_size=3) compute_sce_token_loss(self.query, self.key, self.key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5)) def test_one_device_zero_pos_radius_with_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=0, keep_aligned_positive=True, use_keys=True, use_all_keys=False, rank=0, world_size=1) compute_sce_token_loss(self.query, self.key, self.key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5)) def test_one_device_two_pos_radius_with_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=True, use_keys=True, use_all_keys=False, rank=0, world_size=1) compute_sce_token_loss(self.query, self.key, self.key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5)) def test_one_device_two_pos_radius_with_keys_not_aligned_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=False, use_keys=True, use_all_keys=False, rank=0, world_size=1) compute_sce_token_loss(self.query, self.key, self.key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5)) def test_several_devices_zero_pos_radius_with_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=0, keep_aligned_positive=True, use_keys=True, use_all_keys=False, rank=1, world_size=3) compute_sce_token_loss(self.query, self.key, self.key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5)) def test_several_devices_two_pos_radius_with_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=True, use_keys=True, use_all_keys=False, rank=1, world_size=3) compute_sce_token_loss(self.query, self.key, self.key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5)) def test_several_devices_two_pos_radius_with_all_keys_not_aligned_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=False, use_keys=True, use_all_keys=False, rank=1, world_size=3) compute_sce_token_loss(self.query, self.key, self.key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5)) def test_one_device_zero_pos_radius_with_all_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=0, keep_aligned_positive=True, use_keys=False, use_all_keys=True, rank=0, world_size=1) compute_sce_token_loss(self.query, self.key, self.key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5)) def test_one_device_two_pos_radius_with_all_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=True, use_keys=False, use_all_keys=True, rank=0, world_size=1) compute_sce_token_loss(self.query, self.key, self.key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5)) def test_one_device_two_pos_radius_with_all_keys_not_aligned_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=False, use_keys=False, use_all_keys=True, rank=0, world_size=1) compute_sce_token_loss(self.query, self.key, self.key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5)) def test_several_devices_zero_pos_radius_with_all_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=0, keep_aligned_positive=True, use_keys=False, use_all_keys=True, rank=1, world_size=3) compute_sce_token_loss(self.query, self.key, self.global_key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5)) def test_several_devices_two_pos_radius_with_all_keys_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=True, use_keys=False, use_all_keys=True, rank=1, world_size=3) compute_sce_token_loss(self.query, self.key, self.global_key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5)) def test_several_devices_two_pos_radius_with_all_keys_not_aligned_init(self): (mask_sim_q, mask_sim_k, mask_prob_q, mask_log_q, num_positives_per_token) = compute_sce_token_masks(self.batch_size, self.num_tokens, self.num_negatives, positive_radius=2, keep_aligned_positive=False, use_keys=False, use_all_keys=True, rank=1, world_size=3) compute_sce_token_loss(self.query, self.key, self.global_key, self.queue, mask_sim_q=mask_sim_q, mask_sim_k=mask_sim_k, mask_prob_q=mask_prob_q, mask_log_q=mask_log_q, coeff=torch.tensor(0.5))
class TestMoCoModel(unittest.TestCase): def setUp(self) -> None: self.trunk_cfg = DictConfig({'_target_': 'eztorch.models.trunks.create_resnet', 'name': 'resnet18', 'num_classes': 0, 'small_input': True}) self.projector_cfg = DictConfig({'_target_': 'eztorch.models.heads.MLPHead', 'input_dim': 512, 'output_dim': 2}) self.predictor_cfg = DictConfig({'_target_': 'eztorch.models.heads.MLPHead', 'input_dim': 512, 'output_dim': 2}) self.queue_cfg = DictConfig({'_target_': 'eztorch.models.queues.FIFOQueue', 'size': 8, 'feature_dim': 2}) self.temp = 0.2 def test_moco_init(self): MoCoModel(trunk=self.trunk_cfg, projector=None, predictor=None, optimizer={}, queue=None, num_devices=1, simulate_n_devices=1, temp=self.temp) MoCoModel(trunk=self.trunk_cfg, projector=None, predictor=None, optimizer={}, queue=None, num_devices=1, simulate_n_devices=8, temp=self.temp) MoCoModel(trunk=self.trunk_cfg, projector=None, predictor=None, optimizer={}, queue=None, num_devices=2, simulate_n_devices=8, temp=self.temp) MoCoModel(trunk=self.trunk_cfg, projector=self.projector_cfg, predictor=None, optimizer={}, queue=None, num_devices=2, temp=self.temp) MoCoModel(trunk=self.trunk_cfg, projector=self.projector_cfg, predictor=self.predictor_cfg, optimizer={}, queue=None, num_devices=2, temp=self.temp) MoCoModel(trunk=self.trunk_cfg, projector=self.projector_cfg, predictor=self.predictor_cfg, optimizer={}, queue=self.queue_cfg, num_devices=2, temp=self.temp) def test_moco_loss_without_key(self): q = torch.arange(1.0, 9.0, 1.0).view((4, 2)) k = torch.arange(9.0, 17.0, 1.0).view((4, 2)) queue = torch.tensor([[0.0, 0, 0, 2], [0.0, 0, 0, 2.0]]) model = MoCoModel(trunk=self.trunk_cfg, projector=self.projector_cfg, predictor=self.predictor_cfg, optimizer={}, queue=self.queue_cfg, num_devices=2, temp=self.temp) model.queue = queue labels = torch.tensor([0, 0, 0, 0]) sim = torch.tensor([[29, 0, 0, 0, 2.5], [81, 0, 0, 0, 5.5], [149, 0, 0, 0, 8.5], [233, 0, 0, 0, 11.5]]) logits = (sim / self.temp) loss = nn.functional.cross_entropy(logits, labels) output_loss = compute_moco_loss(q, k, k, False, queue, self.temp, 0) assert torch.equal(loss, output_loss) def test_moco_loss_with_key(self): q = torch.arange(1.0, 9.0, 1.0).view((4, 2)) k = torch.tensor([[0, 0], [1, 1], [0, 0], [0.0, 0.0]]) queue = torch.tensor([[0.0, 0, 0, 2], [0.0, 0, 0, 2.0]]) model = MoCoModel(trunk=self.trunk_cfg, projector=self.projector_cfg, predictor=self.predictor_cfg, optimizer={}, use_keys=True, queue=self.queue_cfg, num_devices=2, temp=self.temp) model.queue = queue labels = torch.tensor([0, 1, 2, 3]) sim = torch.tensor([[0.0, 3, 0, 0, 0, 0, 0, 6], [0, 7, 0, 0, 0, 0, 0, 14], [0, 11, 0, 0, 0, 0, 0, 22], [0, 15, 0, 0, 0, 0, 0, 30]]) logits = (sim / self.temp) loss = nn.functional.cross_entropy(logits, labels) output_loss = compute_moco_loss(q, k, k, True, queue, self.temp, 0) assert torch.equal(loss, output_loss) def test_moco_loss_with_key_without_queue(self): q = torch.arange(1.0, 9.0, 1.0).view((4, 2)) k = torch.tensor([[0, 0], [1, 1], [0, 0], [0.0, 0.0]]) model = MoCoModel(trunk=self.trunk_cfg, projector=self.projector_cfg, predictor=self.predictor_cfg, optimizer={}, use_keys=True, queue=None, num_devices=2, temp=self.temp) labels = torch.tensor([0, 1, 2, 3]) sim = torch.tensor([[0.0, 3, 0, 0], [0, 7, 0, 0], [0, 11, 0, 0], [0, 15, 0, 0]]) logits = (sim / self.temp) loss = nn.functional.cross_entropy(logits, labels) output_loss = compute_moco_loss(q, k, k, True, None, self.temp, 0) assert torch.equal(loss, output_loss) def test_moco_sym_fit(self): optimizer_cfg = DictConfig({'_target_': 'eztorch.optimizers.optimizer_factory', 'name': 'adam', 'scheduler': None, 'initial_lr': 0.06}) transform_cfg = [{'num_views': 2, 'transform': nn.Identity()}] model = MoCoModel(trunk=self.trunk_cfg, optimizer=optimizer_cfg, use_keys=True, queue=None, num_devices=1, simulate_n_devices=1, sym=True, temp=self.temp) datamodule = BoringDataModule(dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=MultiCropTransform(transform_cfg)), val_dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=None), batch_size=64) trainer = Trainer(fast_dev_run=1, devices=1) trainer.fit(model, datamodule) def test_moco_sym_mutual_fit(self): optimizer_cfg = DictConfig({'_target_': 'eztorch.optimizers.optimizer_factory', 'name': 'adam', 'scheduler': None, 'initial_lr': 0.06}) transform_cfg = [{'num_views': 2, 'transform': nn.Identity()}] model = MoCoModel(trunk=self.trunk_cfg, optimizer=optimizer_cfg, use_keys=True, queue=None, num_devices=1, simulate_n_devices=1, sym=True, temp=self.temp, mutual_pass=True) datamodule = BoringDataModule(dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=MultiCropTransform(transform_cfg)), val_dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=None), batch_size=64) trainer = Trainer(fast_dev_run=1, devices=1) trainer.fit(model, datamodule) def test_moco_fit(self): optimizer_cfg = DictConfig({'_target_': 'eztorch.optimizers.optimizer_factory', 'name': 'adam', 'scheduler': None, 'initial_lr': 0.06}) transform_cfg = [{'num_views': 2, 'transform': nn.Identity()}] model = MoCoModel(trunk=self.trunk_cfg, optimizer=optimizer_cfg, projector=self.projector_cfg, use_keys=True, queue=None, num_devices=1, simulate_n_devices=1, initial_momentum=0.98, scheduler_momentum='cosine', temp=self.temp) datamodule = BoringDataModule(dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=MultiCropTransform(transform_cfg)), val_dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=None), batch_size=64) assert (model.current_momentum == 0.98) trainer = Trainer(fast_dev_run=2, devices=1) trainer.fit(model, datamodule) assert (model.current_momentum == 0.98) def test_moco_mutual_fit(self): optimizer_cfg = DictConfig({'_target_': 'eztorch.optimizers.optimizer_factory', 'name': 'adam', 'scheduler': None, 'initial_lr': 0.06}) transform_cfg = [{'num_views': 2, 'transform': nn.Identity()}] model = MoCoModel(trunk=self.trunk_cfg, optimizer=optimizer_cfg, projector=self.projector_cfg, use_keys=True, queue=None, num_devices=1, simulate_n_devices=1, initial_momentum=0.98, scheduler_momentum='cosine', temp=self.temp, mutual_pass=True) datamodule = BoringDataModule(dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=MultiCropTransform(transform_cfg)), val_dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=None), batch_size=64) assert (model.current_momentum == 0.98) trainer = Trainer(fast_dev_run=2, devices=1) trainer.fit(model, datamodule) assert (model.current_momentum == 0.98)
class TestReSSLModel(unittest.TestCase): def setUp(self) -> None: self.trunk_cfg = DictConfig({'_target_': 'eztorch.models.trunks.create_resnet', 'name': 'resnet18', 'num_classes': 0, 'small_input': True}) self.projector_cfg = DictConfig({'_target_': 'eztorch.models.heads.MLPHead', 'input_dim': 512, 'output_dim': 2}) self.predictor_cfg = DictConfig({'_target_': 'eztorch.models.heads.MLPHead', 'input_dim': 512, 'output_dim': 2}) self.queue_cfg = DictConfig({'_target_': 'eztorch.models.queues.FIFOQueue', 'size': 8, 'feature_dim': 2}) self.temp = 0.1 self.temp_m = 0.04 def test_ressl_init(self): ReSSLModel(trunk=self.trunk_cfg, projector=None, predictor=None, optimizer={}, queue=None, num_devices=1, simulate_n_devices=1, temp=self.temp, temp_m=self.temp_m) ReSSLModel(trunk=self.trunk_cfg, projector=None, predictor=None, optimizer={}, queue=None, num_devices=1, simulate_n_devices=8, temp=self.temp, temp_m=self.temp_m) ReSSLModel(trunk=self.trunk_cfg, projector=None, predictor=None, optimizer={}, queue=None, num_devices=2, simulate_n_devices=8, temp=self.temp, temp_m=self.temp_m) ReSSLModel(trunk=self.trunk_cfg, projector=self.projector_cfg, predictor=None, optimizer={}, queue=None, num_devices=2, temp=self.temp, temp_m=self.temp_m) ReSSLModel(trunk=self.trunk_cfg, projector=self.projector_cfg, predictor=self.predictor_cfg, optimizer={}, queue=None, num_devices=2, temp=self.temp, temp_m=self.temp_m) ReSSLModel(trunk=self.trunk_cfg, projector=self.projector_cfg, predictor=self.predictor_cfg, optimizer={}, queue=self.queue_cfg, num_devices=2, temp=self.temp, temp_m=self.temp_m) def test_ressl_loss_without_key(self): q = torch.arange(1.0, 9.0, 1.0).view((4, 2)) k = torch.tensor([[0, 0], [1, 1], [0, 0], [1.0, 1.0]]) queue = torch.tensor([[0.0, 0, 0, 2], [0.0, 0, 0, 2.0]]) queue model = ReSSLModel(trunk=self.trunk_cfg, projector=self.projector_cfg, predictor=self.predictor_cfg, optimizer={}, use_keys=False, queue=self.queue_cfg, num_devices=2, temp=self.temp, temp_m=self.temp_m) model.queue = queue sim_qqueue = torch.Tensor([[0.0, 0, 0, 6], [0, 0, 0, 14], [0, 0, 0, 22], [0, 0, 0, 30]]) sim_kqueue = torch.Tensor([[0.0, 0, 0, 0], [0, 0, 0, 4], [0, 0, 0, 0], [0, 0, 0, 4]]) sim_q = sim_qqueue sim_k = sim_kqueue logits_q = (sim_q / self.temp) logits_k = (sim_k / self.temp_m) loss = (- torch.sum((nn.functional.softmax(logits_k.detach(), dim=1) * nn.functional.log_softmax(logits_q, dim=1)), dim=1).mean(dim=0)) output_loss = compute_ressl_loss(q, k, k, False, queue, None, self.temp, self.temp_m) assert torch.equal(loss, output_loss) def test_ressl_loss_with_key(self): q = torch.arange(1.0, 9.0, 1.0).view((4, 2)) k = torch.tensor([[0, 0], [1, 1], [0, 0], [1.0, 1.0]]) queue = torch.tensor([[0.0, 0, 0, 2], [0.0, 0, 0, 2.0]]) model = ReSSLModel(trunk=self.trunk_cfg, projector=self.projector_cfg, predictor=self.predictor_cfg, optimizer={}, use_keys=True, queue=self.queue_cfg, num_devices=2, temp=self.temp, temp_m=self.temp_m) model.queue = queue sim_qk = torch.tensor([[0, 3, 0, 3], [0, 7, 0, 7], [0, 11, 0, 11], [0, 15.0, 0, 15]]) sim_kk = torch.tensor([[0, 0, 0, 0], [0, 2, 0, 2], [0, 0, 0, 0], [0, 2.0, 0, 2]]) sim_qqueue = torch.Tensor([[0.0, 0, 0, 6], [0, 0, 0, 14], [0, 0, 0, 22], [0, 0, 0, 30]]) sim_kqueue = torch.Tensor([[0.0, 0, 0, 0], [0, 0, 0, 4], [0, 0, 0, 0], [0, 0, 0, 4]]) mask = torch.tensor([[1.0, 0, 0, 0], [0, 1.0, 0, 0], [0, 0, 1.0, 0], [0, 0, 0, 1.0]]) sim_kk -= (1000000000.0 * mask) sim_qk -= (1000000000.0 * mask) sim_q = torch.cat([sim_qk, sim_qqueue], dim=1) sim_k = torch.cat([sim_kk, sim_kqueue], dim=1) logits_q = (sim_q / self.temp) logits_k = (sim_k / self.temp_m) loss = (- torch.sum((nn.functional.softmax(logits_k.detach(), dim=1) * nn.functional.log_softmax(logits_q, dim=1)), dim=1).mean(dim=0)) mask = compute_ressl_mask(q.shape[0], queue.shape[1], True, 0, 1, 'cuda') output_loss = compute_ressl_loss(q, k, k, True, queue, mask, self.temp, self.temp_m) assert torch.equal(loss, output_loss) def test_ressl_loss_with_key_without_queue(self): q = torch.arange(1.0, 9.0, 1.0).view((4, 2)) k = torch.tensor([[0, 0], [1, 1], [0, 0], [1.0, 1.0]]) model = ReSSLModel(trunk=self.trunk_cfg, projector=self.projector_cfg, predictor=self.predictor_cfg, optimizer={}, use_keys=True, queue=None, num_devices=2, temp=self.temp, temp_m=self.temp_m) sim_qk = torch.tensor([[0, 3, 0, 3], [0, 7, 0, 7], [0, 11, 0, 11], [0, 15.0, 0, 15]]) sim_kk = torch.tensor([[0, 0, 0, 0], [0, 2, 0, 2], [0, 0, 0, 0], [0, 2.0, 0, 2]]) mask = torch.tensor([[1.0, 0, 0, 0], [0, 1.0, 0, 0], [0, 0, 1.0, 0], [0, 0, 0, 1.0]]) sim_kk -= (1000000000.0 * mask) sim_qk -= (1000000000.0 * mask) sim_q = sim_qk sim_k = sim_kk logits_q = (sim_q / self.temp) logits_k = (sim_k / self.temp_m) loss = (- torch.sum((nn.functional.softmax(logits_k.detach(), dim=1) * nn.functional.log_softmax(logits_q, dim=1)), dim=1).mean(dim=0)) mask = compute_ressl_mask(q.shape[0], 0, True, 0, 1, 'cuda') output_loss = compute_ressl_loss(q, k, k, True, None, mask, self.temp, self.temp_m) assert torch.equal(loss, output_loss) def test_ressl_sym_fit(self): optimizer_cfg = DictConfig({'_target_': 'eztorch.optimizers.optimizer_factory', 'name': 'adam', 'scheduler': None, 'initial_lr': 0.06}) transform_cfg = [{'num_views': 2, 'transform': nn.Identity()}] model = ReSSLModel(trunk=self.trunk_cfg, optimizer=optimizer_cfg, use_keys=True, queue=None, num_devices=1, simulate_n_devices=1, sym=True, temp=self.temp, temp_m=self.temp_m) datamodule = BoringDataModule(dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=MultiCropTransform(transform_cfg)), val_dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=None), batch_size=64) trainer = Trainer(fast_dev_run=1, devices=1) trainer.fit(model, datamodule) def test_ressl_sym_mutual_fit(self): optimizer_cfg = DictConfig({'_target_': 'eztorch.optimizers.optimizer_factory', 'name': 'adam', 'scheduler': None, 'initial_lr': 0.06}) transform_cfg = [{'num_views': 2, 'transform': nn.Identity()}] model = ReSSLModel(trunk=self.trunk_cfg, optimizer=optimizer_cfg, use_keys=True, queue=None, num_devices=1, simulate_n_devices=1, sym=True, temp=self.temp, temp_m=self.temp_m, mutual_pass=True) datamodule = BoringDataModule(dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=MultiCropTransform(transform_cfg)), val_dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=None), batch_size=64) trainer = Trainer(fast_dev_run=1, devices=1) trainer.fit(model, datamodule) def test_ressl_fit(self): optimizer_cfg = DictConfig({'_target_': 'eztorch.optimizers.optimizer_factory', 'name': 'adam', 'scheduler': None, 'initial_lr': 0.06}) transform_cfg = [{'num_views': 2, 'transform': nn.Identity()}] model = ReSSLModel(trunk=self.trunk_cfg, optimizer=optimizer_cfg, projector=self.projector_cfg, use_keys=True, queue=None, num_devices=1, simulate_n_devices=1, initial_momentum=0.98, scheduler_momentum='cosine', temp=self.temp, temp_m=self.temp_m) datamodule = BoringDataModule(dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=MultiCropTransform(transform_cfg)), val_dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=None), batch_size=64) assert (model.current_momentum == 0.98) trainer = Trainer(fast_dev_run=2, devices=1) trainer.fit(model, datamodule) assert (model.current_momentum == 0.98) def test_ressl_mutual_fit(self): optimizer_cfg = DictConfig({'_target_': 'eztorch.optimizers.optimizer_factory', 'name': 'adam', 'scheduler': None, 'initial_lr': 0.06}) transform_cfg = [{'num_views': 2, 'transform': nn.Identity()}] model = ReSSLModel(trunk=self.trunk_cfg, optimizer=optimizer_cfg, projector=self.projector_cfg, use_keys=True, queue=None, num_devices=1, simulate_n_devices=1, initial_momentum=0.98, scheduler_momentum='cosine', temp=self.temp, temp_m=self.temp_m, mutual_pass=True) datamodule = BoringDataModule(dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=MultiCropTransform(transform_cfg)), val_dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=None), batch_size=64) assert (model.current_momentum == 0.98) trainer = Trainer(fast_dev_run=2, devices=1) trainer.fit(model, datamodule) assert (model.current_momentum == 0.98)
class TestSCEModel(unittest.TestCase): def setUp(self) -> None: self.trunk_cfg = DictConfig({'_target_': 'eztorch.models.trunks.create_resnet', 'name': 'resnet18', 'num_classes': 0, 'small_input': True}) self.projector_cfg = DictConfig({'_target_': 'eztorch.models.heads.MLPHead', 'input_dim': 512, 'output_dim': 2}) self.predictor_cfg = DictConfig({'_target_': 'eztorch.models.heads.MLPHead', 'input_dim': 512, 'output_dim': 2}) self.queue_cfg = DictConfig({'_target_': 'eztorch.models.queues.FIFOQueue', 'size': 8, 'feature_dim': 2}) self.temp = 0.1 self.temp_m = 0.05 self.coeff = 0.5 def test_sce_init(self): SCEModel(trunk=self.trunk_cfg, projector=None, predictor=None, optimizer={}, queue=None, num_devices=1, simulate_n_devices=1, temp=self.temp, temp_m=self.temp_m, coeff=self.coeff) SCEModel(trunk=self.trunk_cfg, projector=None, predictor=None, optimizer={}, queue=None, num_devices=1, simulate_n_devices=8, temp=self.temp, temp_m=self.temp_m, coeff=self.coeff) SCEModel(trunk=self.trunk_cfg, projector=None, predictor=None, optimizer={}, queue=None, num_devices=2, simulate_n_devices=8, temp=self.temp, temp_m=self.temp_m, coeff=self.coeff) SCEModel(trunk=self.trunk_cfg, projector=self.projector_cfg, predictor=None, optimizer={}, queue=None, num_devices=2, temp=self.temp, temp_m=self.temp_m, coeff=self.coeff) SCEModel(trunk=self.trunk_cfg, projector=self.projector_cfg, predictor=self.predictor_cfg, optimizer={}, queue=None, num_devices=2, temp=self.temp, temp_m=self.temp_m, coeff=self.coeff) SCEModel(trunk=self.trunk_cfg, projector=self.projector_cfg, predictor=self.predictor_cfg, optimizer={}, queue=self.queue_cfg, num_devices=2, temp=self.temp, temp_m=self.temp_m, coeff=self.coeff) def test_sce_sym_fit(self): optimizer_cfg = DictConfig({'_target_': 'eztorch.optimizers.optimizer_factory', 'name': 'adam', 'scheduler': None, 'initial_lr': 0.06}) transform_cfg = [{'num_views': 2, 'transform': nn.Identity()}] model = SCEModel(trunk=self.trunk_cfg, optimizer=optimizer_cfg, use_keys=True, queue=None, num_devices=1, simulate_n_devices=1, sym=True, temp=self.temp, temp_m=self.temp_m, coeff=self.coeff) datamodule = BoringDataModule(dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=MultiCropTransform(transform_cfg)), val_dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=None), batch_size=64) trainer = Trainer(fast_dev_run=1, devices=1) trainer.fit(model, datamodule) def test_sce_split_fit(self): optimizer_cfg = DictConfig({'_target_': 'eztorch.optimizers.optimizer_factory', 'name': 'adam', 'scheduler': None, 'initial_lr': 0.06}) transform_cfg = [{'num_views': 2, 'transform': nn.Identity()}] model = SCEModel(trunk=self.trunk_cfg, optimizer=optimizer_cfg, use_keys=True, queue=None, num_devices=1, simulate_n_devices=1, sym=False, num_splits=2, temp=self.temp, temp_m=self.temp_m, coeff=self.coeff) datamodule = BoringDataModule(dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=MultiCropTransform(transform_cfg)), val_dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=None), batch_size=64) trainer = Trainer(fast_dev_run=1, devices=1) trainer.fit(model, datamodule) def test_sce_split_sym_fit(self): optimizer_cfg = DictConfig({'_target_': 'eztorch.optimizers.optimizer_factory', 'name': 'adam', 'scheduler': None, 'initial_lr': 0.06}) transform_cfg = [{'num_views': 2, 'transform': nn.Identity()}] model = SCEModel(trunk=self.trunk_cfg, optimizer=optimizer_cfg, use_keys=True, queue=None, num_devices=1, simulate_n_devices=1, sym=True, num_splits=2, temp=self.temp, temp_m=self.temp_m, coeff=self.coeff) datamodule = BoringDataModule(dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=MultiCropTransform(transform_cfg)), val_dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=None), batch_size=64) trainer = Trainer(fast_dev_run=1, devices=1) trainer.fit(model, datamodule) def test_sce_split_sym_with_queue_fit(self): optimizer_cfg = DictConfig({'_target_': 'eztorch.optimizers.optimizer_factory', 'name': 'adam', 'scheduler': None, 'initial_lr': 0.06}) transform_cfg = [{'num_views': 2, 'transform': nn.Identity()}] model = SCEModel(trunk=self.trunk_cfg, optimizer=optimizer_cfg, use_keys=True, queue=DictConfig({'size': 128, 'feature_dim': 512}), num_devices=1, simulate_n_devices=1, sym=True, num_splits=2, temp=self.temp, temp_m=self.temp_m, coeff=self.coeff) datamodule = BoringDataModule(dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=MultiCropTransform(transform_cfg)), val_dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=None), batch_size=64) trainer = Trainer(fast_dev_run=2, devices=1) trainer.fit(model, datamodule) def test_sce_sym_mutual_fit(self): optimizer_cfg = DictConfig({'_target_': 'eztorch.optimizers.optimizer_factory', 'name': 'adam', 'scheduler': None, 'initial_lr': 0.06}) transform_cfg = [{'num_views': 2, 'transform': nn.Identity()}] model = SCEModel(trunk=self.trunk_cfg, optimizer=optimizer_cfg, use_keys=True, queue=None, num_devices=1, simulate_n_devices=1, sym=True, temp=self.temp, temp_m=self.temp_m, coeff=self.coeff, mutual_pass=True) datamodule = BoringDataModule(dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=MultiCropTransform(transform_cfg)), val_dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=None), batch_size=64) trainer = Trainer(fast_dev_run=1, devices=1) trainer.fit(model, datamodule) def test_sce_fit(self): optimizer_cfg = DictConfig({'_target_': 'eztorch.optimizers.optimizer_factory', 'name': 'adam', 'scheduler': None, 'initial_lr': 0.06}) transform_cfg = [{'num_views': 2, 'transform': nn.Identity()}] model = SCEModel(trunk=self.trunk_cfg, optimizer=optimizer_cfg, projector=self.projector_cfg, use_keys=True, queue=None, num_devices=1, simulate_n_devices=1, initial_momentum=0.98, scheduler_momentum='cosine', temp=self.temp, temp_m=self.temp_m, coeff=self.coeff) datamodule = BoringDataModule(dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=MultiCropTransform(transform_cfg)), val_dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=None), batch_size=64) assert (model.current_momentum == 0.98) trainer = Trainer(fast_dev_run=2, devices=1) trainer.fit(model, datamodule) assert (model.current_momentum == 0.98) def test_sce_mutual_fit(self): optimizer_cfg = DictConfig({'_target_': 'eztorch.optimizers.optimizer_factory', 'name': 'adam', 'scheduler': None, 'initial_lr': 0.06}) transform_cfg = [{'num_views': 2, 'transform': nn.Identity()}] model = SCEModel(trunk=self.trunk_cfg, optimizer=optimizer_cfg, projector=self.projector_cfg, use_keys=True, queue=None, num_devices=1, simulate_n_devices=1, initial_momentum=0.98, scheduler_momentum='cosine', temp=self.temp, temp_m=self.temp_m, coeff=self.coeff, mutual_pass=True) datamodule = BoringDataModule(dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=MultiCropTransform(transform_cfg)), val_dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=None), batch_size=64) assert (model.current_momentum == 0.98) trainer = Trainer(fast_dev_run=2, devices=1) trainer.fit(model, datamodule) assert (model.current_momentum == 0.98)
class TestSimCLRModel(unittest.TestCase): def setUp(self) -> None: self.trunk_cfg = DictConfig({'_target_': 'eztorch.models.trunks.create_resnet', 'name': 'resnet18', 'num_classes': 0, 'small_input': True}) self.projector_cfg = DictConfig({'_target_': 'eztorch.models.heads.MLPHead', 'input_dim': 512, 'output_dim': 2}) self.queue_cfg = DictConfig({'_target_': 'eztorch.models.queues.FIFOQueue', 'size': 8, 'feature_dim': 2}) self.temp = 10.0 def test_simclr_init(self): SimCLRModel(trunk=self.trunk_cfg, projector=None, optimizer={}, temp=self.temp) SimCLRModel(trunk=self.trunk_cfg, projector=None, optimizer={}, temp=self.temp) SimCLRModel(trunk=self.trunk_cfg, projector=None, optimizer={}, temp=self.temp) SimCLRModel(trunk=self.trunk_cfg, projector=self.projector_cfg, optimizer={}, temp=self.temp) def test_simclr_loss(self): z = torch.tensor([[1.0, 2], [3, 4], [5, 6], [7, 8], [0, 0], [1, 1], [0, 0], [0, 0]]) pos_mask = torch.tensor([[0.0, 0, 0, 0, 1, 0, 0, 0], [0.0, 0, 0, 0, 0, 1, 0, 0], [0.0, 0, 0, 0, 0, 0, 1, 0], [0.0, 0, 0, 0, 0, 0, 0, 1], [1.0, 0, 0, 0, 0, 0, 0, 0], [0.0, 1, 0, 0, 0, 0, 0, 0], [0.0, 0, 1, 0, 0, 0, 0, 0], [0.0, 0, 0, 1, 0, 0, 0, 0]]) neg_mask = torch.tensor([[0.0, 1, 1, 1, 0, 1, 1, 1], [1.0, 0, 1, 1, 1, 0, 1, 1], [1.0, 1, 0, 1, 1, 1, 0, 1], [1.0, 1, 1, 0, 1, 1, 1, 0], [0.0, 1, 1, 1, 0, 1, 1, 1], [1.0, 0, 1, 1, 1, 0, 1, 1], [1.0, 1, 0, 1, 1, 1, 0, 1], [1.0, 1, 1, 0, 1, 1, 1, 0]]) sim = torch.tensor([[5.0, 11, 17, 23, 0, 3, 0, 0], [11, 25, 39, 53, 0, 7, 0, 0], [17, 39, 61, 83, 0, 11, 0, 0], [23, 53, 83, 113, 0, 15, 0, 0], [0.0, 0, 0, 0, 0, 0, 0, 0], [3, 7, 11, 15, 0, 2, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0]]) logits = torch.exp((sim / self.temp)) pos = torch.sum((logits * pos_mask), 1) neg = torch.sum((logits * neg_mask), 1) loss = (- torch.mean(torch.log((pos / (neg + pos))))) output_loss = compute_simclr_loss(z, z, pos_mask, neg_mask, self.temp) assert torch.equal(loss, output_loss) def test_simclr_fit(self): optimizer_cfg = DictConfig({'_target_': 'eztorch.optimizers.optimizer_factory', 'name': 'adam', 'scheduler': None, 'initial_lr': 0.06}) transform_cfg = [{'num_views': 2, 'transform': nn.Identity()}] model = SimCLRModel(trunk=self.trunk_cfg, optimizer=optimizer_cfg, temp=self.temp) datamodule = BoringDataModule(dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=MultiCropTransform(transform_cfg)), val_dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=None), batch_size=64) trainer = Trainer(fast_dev_run=1, devices=1) trainer.fit(model, datamodule) def test_simclr_mutual_fit(self): optimizer_cfg = DictConfig({'_target_': 'eztorch.optimizers.optimizer_factory', 'name': 'adam', 'scheduler': None, 'initial_lr': 0.06}) transform_cfg = [{'num_views': 2, 'transform': nn.Identity()}] model = SimCLRModel(trunk=self.trunk_cfg, optimizer=optimizer_cfg, temp=self.temp, mutual_pass=True) datamodule = BoringDataModule(dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=MultiCropTransform(transform_cfg)), val_dataset=RandomVisionLabeledDataset((128, 3, 32, 32), transform=None), batch_size=64) trainer = Trainer(fast_dev_run=1, devices=1) trainer.fit(model, datamodule)
class TestResnet(unittest.TestCase): def test_all_resnets(self): for resnet in _ResNets: create_resnet(resnet) def test_resnet_small_input_with_fc(self): resnet = create_resnet('resnet18', small_input=True) assert isinstance(resnet.fc, nn.Linear) assert (resnet.conv1.kernel_size == (3, 3)) assert (resnet.conv1.stride == (1, 1)) assert (resnet.conv1.padding == (1, 1)) assert isinstance(resnet.maxpool, nn.Identity) def test_resnet_small_input_without_fc(self): resnet = create_resnet('resnet18', small_input=True, num_classes=0) assert isinstance(resnet.fc, nn.Identity) assert (resnet.conv1.kernel_size == (3, 3)) assert (resnet.conv1.stride == (1, 1)) assert (resnet.conv1.padding == (1, 1)) assert isinstance(resnet.maxpool, nn.Identity) def test_resnet_large_input_with_fc(self): resnet = create_resnet('resnet18', small_input=False) assert isinstance(resnet.fc, nn.Linear) assert (resnet.conv1.kernel_size == (7, 7)) assert (resnet.conv1.stride == (2, 2)) assert (resnet.conv1.padding == (3, 3)) assert isinstance(resnet.maxpool, nn.MaxPool2d) def test_resnet_large_input_without_fc(self): resnet = create_resnet('resnet18', small_input=False, num_classes=0) assert isinstance(resnet.fc, nn.Identity) assert (resnet.conv1.kernel_size == (7, 7)) assert (resnet.conv1.stride == (2, 2)) assert (resnet.conv1.padding == (3, 3)) assert isinstance(resnet.maxpool, nn.MaxPool2d) def test_resnet_forward(self): resnet = create_resnet('resnet18') x = torch.rand((1, 3, 224, 224)) resnet(x)
class TestOptimizerFactory(unittest.TestCase): def test_init_lars(self): model = BoringModel() LARS(model.parameters(), lr=0.1)
class TestOptimizerFactory(unittest.TestCase): def setUp(self): self.base_model = BoringModel() self.large_model = LargeBoringModel() self.sgd_config = DictConfig({'name': 'sgd', 'initial_lr': 2.0, 'batch_size': None, 'num_steps_per_epoch': None, 'exclude_wd_norm': False, 'exclude_wd_bias': False, 'scaler': None, 'params': {}, 'scheduler': None}) def test_no_exclude_optimizer_factory(self): (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=[], model=self.base_model) assert (len(optimizer.param_groups) == 1) assert (len(optimizer.param_groups[0]['params']) == 2) (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=[], model=self.large_model) assert (len(optimizer.param_groups) == 1) assert (len(optimizer.param_groups[0]['params']) == 11) def test_no_exclude_optimizer_factory_layer_decay(self): (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=[], model=self.base_model, layer_decay_lr=0.7) print(optimizer.param_groups) assert (len(optimizer.param_groups) == 1) assert (len(optimizer.param_groups[0]['params']) == 2) (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=[], model=self.large_model, layer_decay_lr=0.7) assert (len(optimizer.param_groups) == 4) assert (len(optimizer.param_groups[0]['params']) == 3) assert (len(optimizer.param_groups[1]['params']) == 3) assert (len(optimizer.param_groups[2]['params']) == 3) assert (len(optimizer.param_groups[3]['params']) == 2) def test_exclude_bias_optimizer_factory(self): self.sgd_config.exclude_wd_bias = True (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=[], model=self.base_model) assert (len(optimizer.param_groups) == 2) assert (len(optimizer.param_groups[0]['params']) == 1) assert (len(optimizer.param_groups[1]['params']) == 1) (optimizer, _) = optimizer_factory(**self.sgd_config, model=self.large_model) assert (len(optimizer.param_groups) == 2) assert (len(optimizer.param_groups[0]['params']) == 7) assert (len(optimizer.param_groups[1]['params']) == 4) def test_exclude_bias_optimizer_factory_layer_decay(self): self.sgd_config.exclude_wd_bias = True (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=[], model=self.base_model, layer_decay_lr=0.7) assert (len(optimizer.param_groups) == 2) assert (len(optimizer.param_groups[0]['params']) == 1) assert (len(optimizer.param_groups[1]['params']) == 1) (optimizer, _) = optimizer_factory(**self.sgd_config, model=self.large_model, layer_decay_lr=0.7) assert (len(optimizer.param_groups) == 8) assert (len(optimizer.param_groups[0]['params']) == 2) assert (len(optimizer.param_groups[1]['params']) == 2) assert (len(optimizer.param_groups[2]['params']) == 2) assert (len(optimizer.param_groups[3]['params']) == 1) assert (len(optimizer.param_groups[4]['params']) == 1) assert (len(optimizer.param_groups[5]['params']) == 1) assert (len(optimizer.param_groups[6]['params']) == 1) assert (len(optimizer.param_groups[7]['params']) == 1) def test_exclude_norm_optimizer_factory(self): self.sgd_config.exclude_wd_norm = True (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=[], model=self.base_model) assert (len(optimizer.param_groups) == 1) assert (len(optimizer.param_groups[0]['params']) == 2) (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=[], model=self.large_model) assert (len(optimizer.param_groups) == 2) assert (len(optimizer.param_groups[0]['params']) == 5) assert (len(optimizer.param_groups[1]['params']) == 6) def test_exclude_norm_optimizer_factory_layer_decay(self): self.sgd_config.exclude_wd_norm = True (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=[], model=self.base_model, layer_decay_lr=0.7) assert (len(optimizer.param_groups) == 1) assert (len(optimizer.param_groups[0]['params']) == 2) (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=[], model=self.large_model, layer_decay_lr=0.7) assert (len(optimizer.param_groups) == 7) assert (len(optimizer.param_groups[0]['params']) == 1) assert (len(optimizer.param_groups[1]['params']) == 1) assert (len(optimizer.param_groups[2]['params']) == 1) assert (len(optimizer.param_groups[3]['params']) == 2) assert (len(optimizer.param_groups[4]['params']) == 2) assert (len(optimizer.param_groups[5]['params']) == 2) assert (len(optimizer.param_groups[6]['params']) == 2) def test_exclude_bias_and_norm_optimizer_factory(self): self.sgd_config.exclude_wd_norm = True self.sgd_config.exclude_wd_bias = True (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=[], model=self.base_model) assert (len(optimizer.param_groups) == 2) assert (len(optimizer.param_groups[0]['params']) == 1) assert (len(optimizer.param_groups[1]['params']) == 1) (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=[], model=self.large_model) assert (len(optimizer.param_groups) == 2) assert (len(optimizer.param_groups[0]['params']) == 4) assert (len(optimizer.param_groups[1]['params']) == 7) def test_exclude_bias_and_norm_optimizer_factory_layer_decay(self): self.sgd_config.exclude_wd_norm = True self.sgd_config.exclude_wd_bias = True (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=[], model=self.base_model, layer_decay_lr=0.7) assert (len(optimizer.param_groups) == 2) assert (len(optimizer.param_groups[0]['params']) == 1) assert (len(optimizer.param_groups[1]['params']) == 1) (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=[], model=self.large_model, layer_decay_lr=0.7) assert (len(optimizer.param_groups) == 8) assert (len(optimizer.param_groups[0]['params']) == 1) assert (len(optimizer.param_groups[1]['params']) == 1) assert (len(optimizer.param_groups[2]['params']) == 1) assert (len(optimizer.param_groups[3]['params']) == 1) assert (len(optimizer.param_groups[4]['params']) == 2) assert (len(optimizer.param_groups[5]['params']) == 2) assert (len(optimizer.param_groups[6]['params']) == 2) assert (len(optimizer.param_groups[7]['params']) == 1) def test_exclude_bias_via_key_and_norm_optimizer_factory(self): self.sgd_config.exclude_wd_norm = True self.sgd_config.exclude_wd_bias = False (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=['bias'], model=self.base_model) assert (len(optimizer.param_groups) == 2) assert (len(optimizer.param_groups[0]['params']) == 1) assert (len(optimizer.param_groups[1]['params']) == 1) (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=['bias'], model=self.large_model) assert (len(optimizer.param_groups) == 2) assert (len(optimizer.param_groups[0]['params']) == 4) assert (len(optimizer.param_groups[1]['params']) == 7) def test_exclude_bias_via_key_and_norm_optimizer_factory_layer_decay(self): self.sgd_config.exclude_wd_norm = True self.sgd_config.exclude_wd_bias = False (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=['bias'], model=self.base_model, layer_decay_lr=0.7) assert (len(optimizer.param_groups) == 2) assert (len(optimizer.param_groups[0]['params']) == 1) assert (len(optimizer.param_groups[1]['params']) == 1) (optimizer, _) = optimizer_factory(**self.sgd_config, keys_without_decay=['bias'], model=self.large_model, layer_decay_lr=0.7) assert (len(optimizer.param_groups) == 8) assert (len(optimizer.param_groups[0]['params']) == 1) assert (len(optimizer.param_groups[1]['params']) == 1) assert (len(optimizer.param_groups[2]['params']) == 1) assert (len(optimizer.param_groups[3]['params']) == 1) assert (len(optimizer.param_groups[4]['params']) == 2) assert (len(optimizer.param_groups[5]['params']) == 2) assert (len(optimizer.param_groups[6]['params']) == 2) assert (len(optimizer.param_groups[7]['params']) == 1)
class TestFilterLearnableParmams(unittest.TestCase): def test_filter_learnable_params(self) -> None: boring_model = BoringModel() large_boring_model = LargeBoringModel() boring_model_params = list(boring_model.parameters()) filtered_boring_model_params = filter_learnable_params(boring_model_params, boring_model) assert all([any([(param is filtered_param) for filtered_param in filtered_boring_model_params]) for param in boring_model_params]) assert (len(boring_model_params) == len(filtered_boring_model_params)) large_boring_model_params = list(large_boring_model.parameters()) filtered_large_boring_model_params = filter_learnable_params(large_boring_model_params, large_boring_model) assert any([(not any([(param is filtered_param) for filtered_param in filtered_large_boring_model_params])) for param in boring_model_params]) assert (len(large_boring_model_params) == (len(filtered_large_boring_model_params) + 2))
class TestLrScaler(unittest.TestCase): def setUp(self) -> None: self.initial_lr = 2.0 self.batch_size = 16 def test_none_scaler(self) -> None: lr = scale_learning_rate(self.initial_lr, None, self.batch_size) assert (lr == self.initial_lr) lr = scale_learning_rate(self.initial_lr, 'none', self.batch_size) assert (lr == self.initial_lr) def test_linear_scaler(self) -> None: lr = scale_learning_rate(self.initial_lr, 'linear', self.batch_size) assert (lr == ((self.initial_lr * self.batch_size) / 256)) def test_sqrt_scaler(self) -> None: lr = scale_learning_rate(self.initial_lr, 'sqrt', self.batch_size) assert (lr == (self.initial_lr * math.sqrt(self.batch_size)))
class TestRetrieveModelParams(unittest.TestCase): def setUp(self) -> None: self.linear1 = nn.Linear(5, 5) self.bn1 = nn.BatchNorm1d(5) self.linear2 = nn.Linear(5, 5, bias=True) self.model = nn.Sequential(self.linear1, self.bn1, self.linear2) self.module_list = list(self.model.modules()) def test_retrieve_model_params_no_filter(self) -> None: modules_to_filter = [] keys_to_filter = [] (filtered_parameters, other_parameters) = retrieve_model_params(self.model, modules_to_filter, keys_to_filter) for param in self.model.parameters(): self.assertTrue(any([(other_parameter is not param) for other_parameter in other_parameters])) self.assertTrue((len(filtered_parameters) == 0)) self.assertTrue((len(list(self.model.parameters())) == len(other_parameters))) self.assertTrue((len(list(self.model.parameters())) == (len(filtered_parameters) + len(other_parameters)))) def test_retrieve_model_params_filter_module(self) -> None: modules_to_filter = [nn.BatchNorm1d] keys_to_filter = [] (filtered_parameters, other_parameters) = retrieve_model_params(self.model, modules_to_filter, keys_to_filter) for param in self.linear1.parameters(): self.assertTrue(all([(param is not filtered_parameter) for filtered_parameter in filtered_parameters])) self.assertTrue(any([(param is other_parameter) for other_parameter in other_parameters])) for param in self.bn1.parameters(): self.assertTrue(any([(param is filtered_parameter) for filtered_parameter in filtered_parameters])) self.assertTrue(all([(param is not other_parameter) for other_parameter in other_parameters])) for param in self.linear2.parameters(): self.assertTrue(all([(param is not filtered_parameter) for filtered_parameter in filtered_parameters])) self.assertTrue(any([(param is other_parameter) for other_parameter in other_parameters])) self.assertTrue((len(filtered_parameters) == 2)) self.assertTrue((len(other_parameters) == 4)) self.assertTrue((len(list(self.model.parameters())) == (len(filtered_parameters) + len(other_parameters)))) def test_retrieve_model_params_filter_key(self) -> None: modules_to_filter = [] keys_to_filter = ['bias'] (filtered_parameters, other_parameters) = retrieve_model_params(self.model, modules_to_filter, keys_to_filter) for (name_param, param) in self.linear1.named_parameters(): if (name_param == 'bias'): self.assertTrue(any([(param is filtered_parameter) for filtered_parameter in filtered_parameters])) self.assertTrue(all([(param is not other_parameter) for other_parameter in other_parameters])) else: self.assertTrue(all([(param is not filtered_parameter) for filtered_parameter in filtered_parameters])) self.assertTrue(any([(param is other_parameter) for other_parameter in other_parameters])) for (name_param, param) in self.bn1.named_parameters(): if (name_param == 'bias'): self.assertTrue(any([(param is filtered_parameter) for filtered_parameter in filtered_parameters])) self.assertTrue(all([(param is not other_parameter) for other_parameter in other_parameters])) else: self.assertTrue(all([(param is not filtered_parameter) for filtered_parameter in filtered_parameters])) self.assertTrue(any([(param is other_parameter) for other_parameter in other_parameters])) for (name_param, param) in self.linear2.named_parameters(): if (name_param == 'bias'): self.assertTrue((not all([(param is not filtered_parameter) for filtered_parameter in filtered_parameters]))) self.assertTrue(all([(param is not other_parameter) for other_parameter in other_parameters])) else: self.assertTrue(all([(param is not filtered_parameter) for filtered_parameter in filtered_parameters])) self.assertTrue(any([(param is other_parameter) for other_parameter in other_parameters])) self.assertTrue((len(filtered_parameters) == 3)) self.assertTrue((len(other_parameters) == 3)) self.assertTrue((len(list(self.model.parameters())) == (len(filtered_parameters) + len(other_parameters)))) def test_retrieve_model_params_filter_key_and_module(self) -> None: modules_to_filter = [nn.BatchNorm1d] keys_to_filter = ['bias'] (filtered_parameters, other_parameters) = retrieve_model_params(self.model, modules_to_filter, keys_to_filter) for (name_param, param) in self.linear1.named_parameters(): if (name_param == 'bias'): self.assertTrue(any([(param is filtered_parameter) for filtered_parameter in filtered_parameters])) self.assertTrue(all([(param is not other_parameter) for other_parameter in other_parameters])) else: self.assertTrue(all([(param is not filtered_parameter) for filtered_parameter in filtered_parameters])) self.assertTrue(any([(param is other_parameter) for other_parameter in other_parameters])) for (name_param, param) in self.bn1.named_parameters(): if (name_param == 'bias'): self.assertTrue(any([(param is filtered_parameter) for filtered_parameter in filtered_parameters])) self.assertTrue(all([(param is not other_parameter) for other_parameter in other_parameters])) else: self.assertTrue(any([(param is filtered_parameter) for filtered_parameter in filtered_parameters])) self.assertTrue(all([(param is not other_parameter) for other_parameter in other_parameters])) for (name_param, param) in self.linear2.named_parameters(): if (name_param == 'bias'): self.assertTrue(any([(param is filtered_parameter) for filtered_parameter in filtered_parameters])) self.assertTrue(all([(param is not other_parameter) for other_parameter in other_parameters])) else: self.assertTrue(all([(param is not filtered_parameter) for filtered_parameter in filtered_parameters])) self.assertTrue(any([(param is other_parameter) for other_parameter in other_parameters])) self.assertTrue((len(filtered_parameters) == 4)) self.assertTrue((len(other_parameters) == 2)) self.assertTrue((len(list(self.model.parameters())) == (len(filtered_parameters) + len(other_parameters))))
class TestReducedTimestamps(unittest.TestCase): def test_batch_reduced_timestamps(self): x = torch.tensor([[[0.5, 0.8], [0.4, 0.6]], [[0.5, 0.8], [0.4, 0.6]], [[0.2, 0.8], [0.4, 0.6]], [[0.2, 0.8], [0.4, 0.6]], [[0.2, 0.8], [0.4, 0.6]], [[0.2, 0.8], [0.4, 0.6]]]) labels = torch.tensor([[[1.0, 0.0], [0.0, 0.0], [1.0, 1.0], [0.0, 1.0]], [[1.0, 0.0], [0.0, 0.0], [1.0, 1.0], [0.0, 1.0]], [[0.0, 1.0], [0.0, 0.0], [1.0, 1.0], [0.0, 1.0]], [[0.0, 1.0], [0.0, 0.0], [1.0, 1.0], [1.0, 1.0]], [[0.0, 0.0], [0.0, 1.0], [0.0, 1.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 1.0], [0.0, 1.0], [0.0, 0.0]]]) has_label = labels.bool() ignore_class = torch.tensor([[[0.0, 0.0], [0.0, 0.0], [0.0, 0.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 0.0], [0.0, 0.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 0.0], [0.0, 0.0], [0.0, 0.0]], [[0.0, 1.0], [0.0, 0.0], [0.0, 0.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 0.0], [0.0, 0.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 0.0], [0.0, 0.0], [0.0, 0.0]]]) timestamps = torch.tensor([[0.0, 0.48, 1.0, 1.48], [0.0, 0.5, 1.0, 1.5], [2.0, 2.5, 3.0, 3.5], [4.0, 4.5, 5.0, 5.5], [6.0, 6.48, 7.0, 7.5], [6.0, 6.5, 7.0, 7.5]]) reduced_batch = BatchReduceTimestamps()({'input': x, 'labels': labels, 'has_label': has_label, 'ignore_class': ignore_class, 'timestamps': timestamps}) expected_reduced_batch = {'input': x, 'labels': torch.tensor([[[1.0, 0.0], [1.0, 1.0]], [[1.0, 0.0], [1.0, 1.0]], [[0.0, 1.0], [1.0, 1.0]], [[0.0, 1.0], [1.0, 1.0]], [[0.0, 1.0], [0.0, 1.0]], [[0.0, 1.0], [0.0, 1.0]]]), 'has_label': torch.tensor([[[1.0, 0.0], [1.0, 1.0]], [[1.0, 0.0], [1.0, 1.0]], [[0.0, 1.0], [1.0, 1.0]], [[0.0, 1.0], [1.0, 1.0]], [[0.0, 1.0], [0.0, 1.0]], [[0.0, 1.0], [0.0, 1.0]]]).bool(), 'ignore_class': torch.tensor([[[0, 0], [0, 0]], [[0, 0], [0, 0]], [[0, 0], [0, 0]], [[0, 1], [0, 0]], [[0, 0], [0, 0]], [[0, 0], [0, 0]]]).to(torch.bool), 'timestamps': torch.tensor([[0.24, 1.24], [0.25, 1.25], [2.25, 3.25], [4.25, 5.25], [6.24, 7.25], [6.25, 7.25]])} for (key, value) in expected_reduced_batch.items(): assert torch.allclose(reduced_batch[key], value, rtol=0.001, atol=0.001) def test_reduced_timestamps(self): x = torch.tensor([[[0.5, 0.8], [0.4, 0.6]], [[0.2, 0.8], [0.4, 0.6]], [[0.2, 0.8], [0.4, 0.6]], [[0.2, 0.8], [0.4, 0.6]]]) labels = torch.tensor([[[1.0, 0.0], [0.0, 0.0], [1.0, 1.0], [0.0, 1.0]], [[0.0, 1.0], [0.0, 0.0], [1.0, 1.0], [0.0, 1.0]], [[0.0, 1.0], [0.0, 0.0], [1.0, 1.0], [1.0, 1.0]], [[0.0, 0.0], [0.0, 1.0], [0.0, 1.0], [0.0, 0.0]]]) has_label = labels.bool() ignore_class = torch.tensor([[[0.0, 0.0], [0.0, 0.0], [0.0, 0.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 0.0], [0.0, 0.0], [0.0, 0.0]], [[0.0, 1.0], [0.0, 0.0], [0.0, 0.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 0.0], [0.0, 0.0], [0.0, 0.0]]]) timestamps = torch.tensor([[0.0, 0.5, 1.0, 1.5], [2.0, 2.5, 3.0, 3.5], [4.0, 4.5, 5.0, 5.5], [6.0, 6.5, 7.0, 7.5]]) expected_reduced_batch = {'input': x, 'labels': torch.tensor([[[1.0, 0.0], [1.0, 1.0]], [[0.0, 1.0], [1.0, 1.0]], [[0.0, 1.0], [1.0, 1.0]], [[0.0, 1.0], [0.0, 1.0]]]), 'has_label': torch.tensor([[[1.0, 0.0], [1.0, 1.0]], [[0.0, 1.0], [1.0, 1.0]], [[0.0, 1.0], [1.0, 1.0]], [[0.0, 1.0], [0.0, 1.0]]]).bool(), 'timestamps': torch.tensor([[0.25, 1.25], [2.25, 3.25], [4.25, 5.25], [6.25, 7.25]])} for i in range(len(x)): reduced = ReduceTimestamps()({'input': x[i], 'labels': labels[i], 'has_label': has_label[i], 'ignore_class': ignore_class[i], 'timestamps': timestamps[i]}) for (key, value) in expected_reduced_batch.items(): assert torch.allclose(reduced[key], value[i], rtol=0.001, atol=0.001)
class TestActionSpottingMixup(unittest.TestCase): def test_mix_action_spotting(self): x = torch.tensor([[[0.5, 0.8], [0.4, 0.6]], [[0.2, 0.8], [0.4, 0.6]], [[0.5, 0.8], [0.4, 0.5]], [[0.4, 0.2], [0.2, 0.5]]]) mix_value = torch.tensor([[[0.5]], [[0.3]], [[0.7]], [[0.2]]]) permutation = torch.tensor([1, 0, 2, 3]) labels = torch.tensor([[[1.0, 0.0], [0.0, 0.0]], [[0.0, 1.0], [0.0, 0.0]], [[1.0, 1.0], [1.0, 0.0]], [[1.0, 1.0], [0.0, 1.0]]]) ignore_class = torch.tensor([[[1.0, 0.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 1.0]]]) (x_mixed, labels_mixed, has_label_mixed, ignore_class_mixed, mixed_weights) = mix_spotting(x, mix_value, permutation, labels, labels.bool(), ignore_class) expected_x_mixed = ((mix_value * x) + ((1 - mix_value) * x[permutation])) expected_labels_mixed = torch.tensor([[[1.0, 0.0], [0.0, 0.0]], [[0.0, 1.0], [0.0, 0.0]], [[1.0, 1.0], [1.0, 0.0]], [[1.0, 1.0], [0.0, 1.0]], [[0.0, 1.0], [0.0, 0.0]], [[1.0, 0.0], [0.0, 0.0]], [[1.0, 1.0], [1.0, 0.0]], [[1.0, 1.0], [0.0, 1.0]]]) expected_has_label_mixed = torch.tensor([[[1.0, 0.0], [0.0, 0.0]], [[0.0, 1.0], [0.0, 0.0]], [[1.0, 1.0], [1.0, 0.0]], [[1.0, 1.0], [0.0, 1.0]], [[0.0, 1.0], [0.0, 0.0]], [[1.0, 0.0], [0.0, 0.0]], [[1.0, 1.0], [1.0, 0.0]], [[1.0, 1.0], [0.0, 1.0]]]).bool() expected_ignore_class_mixed = torch.tensor([[[1.0, 0.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 1.0]], [[0.0, 0.0], [0.0, 0.0]], [[1.0, 0.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 1.0]]]) expected_mixed_weights = torch.tensor([[[0.5]], [[0.3]], [[0.7]], [[0.2]], [[0.5]], [[0.7]], [[0.3]], [[0.8]]]) assert torch.allclose(x_mixed, expected_x_mixed) assert torch.allclose(labels_mixed, expected_labels_mixed) assert torch.allclose(has_label_mixed, expected_has_label_mixed) assert torch.allclose(ignore_class_mixed, expected_ignore_class_mixed) assert torch.allclose(mixed_weights, expected_mixed_weights) def test_action_spotting_mixup(self): spotting_mixup = SpottingMixup(alpha=0.5) x = torch.tensor([[[0.5, 0.8], [0.4, 0.6]], [[0.2, 0.8], [0.4, 0.6]], [[0.5, 0.8], [0.4, 0.5]], [[0.4, 0.2], [0.2, 0.5]]]) labels = torch.tensor([[[1.0, 0.0], [0.0, 0.0]], [[0.0, 1.0], [0.0, 0.0]], [[1.0, 1.0], [1.0, 0.0]], [[1.0, 1.0], [0.0, 1.0]]]) ignore_class = torch.tensor([[[1.0, 0.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, 1.0]]]) spotting_mixup(batch={'input': x, 'labels': labels, 'has_label': labels.bool(), 'ignore_class': ignore_class})
class RSSMPrior(nn.Module): c: Config @nn.compact def __call__(self, prev_state, context): inputs = jnp.concatenate([prev_state['sample'], context], (- 1)) hl = nn.relu(nn.Dense(self.c.cell_embed_size)(inputs)) (det_state, det_out) = GRUCell()(prev_state['det_state'], hl) hl = nn.relu(nn.Dense(self.c.cell_embed_size)(det_out)) mean = nn.Dense(self.c.cell_stoch_size)(hl) stddev = (nn.softplus(nn.Dense(self.c.cell_stoch_size)((hl + 0.54))) + self.c.cell_min_stddev) dist = tfd.MultivariateNormalDiag(mean, stddev) sample = dist.sample(seed=self.make_rng('sample')) return dict(mean=mean, stddev=stddev, sample=sample, det_out=det_out, det_state=det_state, output=jnp.concatenate([sample, det_out], (- 1)))
class RSSMPosterior(nn.Module): c: Config @nn.compact def __call__(self, prior, obs_inputs): inputs = jnp.concatenate([prior['det_out'], obs_inputs], (- 1)) hl = nn.relu(nn.Dense(self.c.cell_embed_size)(inputs)) hl = nn.relu(nn.Dense(self.c.cell_embed_size)(hl)) mean = nn.Dense(self.c.cell_stoch_size)(hl) stddev = (nn.softplus(nn.Dense(self.c.cell_stoch_size)((hl + 0.54))) + self.c.cell_min_stddev) dist = tfd.MultivariateNormalDiag(mean, stddev) sample = dist.sample(seed=self.make_rng('sample')) return dict(mean=mean, stddev=stddev, sample=sample, det_out=prior['det_out'], det_state=prior['det_state'], output=jnp.concatenate([sample, prior['det_out']], (- 1)))
class RSSMCell(nn.Module): c: Config @property def state_size(self): return dict(mean=self.c.cell_stoch_size, stddev=self.c.cell_stoch_size, sample=self.c.cell_stoch_size, det_out=self.c.cell_deter_size, det_state=self.c.cell_deter_size, output=(self.c.cell_stoch_size + self.c.cell_deter_size)) def zero_state(self, batch_size, dtype=jnp.float32): return {k: jnp.zeros((batch_size, v), dtype=dtype) for (k, v) in self.state_size.items()} @nn.compact def __call__(self, state, inputs, use_obs): (obs_input, context) = inputs prior = RSSMPrior(self.c)(state, context) posterior = (RSSMPosterior(self.c)(prior, obs_input) if use_obs else prior) return (posterior, (prior, posterior))
class Encoder(nn.Module): '\n Multi-level Video Encoder.\n 1. Extracts hierarchical features from a sequence of observations.\n 2. Encodes observations using Conv layers, uses them directly for the bottom-most level.\n 3. Uses dense features for each level of the hierarchy above the bottom-most level.\n ' c: Config @nn.compact def __call__(self, obs): '\n Arguments:\n obs : Tensor\n Un-flattened observations (videos) of shape (batch size, timesteps, height, width, channels)\n ' x = obs.reshape((((- 1),) + obs.shape[2:])) Conv = partial(nn.Conv, kernel_size=(4, 4), strides=(2, 2), padding='VALID') x = leaky_relu(Conv(self.c.total_filters)(x)) x = leaky_relu(Conv((self.c.total_filters * 2))(x)) x = leaky_relu(Conv((self.c.total_filters * 4))(x)) x = leaky_relu(Conv((self.c.total_filters * 8))(x)) x = x.reshape((obs.shape[:2] + ((- 1),))) layers = [x] print(f'Input shape at level 0: {x.shape}') feat_size = x.shape[(- 1)] for level in range(1, self.c.levels): for _ in range((self.c.enc_dense_layers - 1)): x = nn.relu(nn.Dense(self.c.enc_dense_embed_size)(x)) if (self.c.enc_dense_layers > 0): x = nn.Dense(feat_size)(x) layer = x timesteps_to_merge = (self.c.tmp_abs_factor ** level) timesteps_to_pad = ((- layer.shape[1]) % timesteps_to_merge) layer = jnp.pad(layer, ((0, 0), (0, timesteps_to_pad), (0, 0))) layer = layer.reshape((layer.shape[0], (- 1), timesteps_to_merge, layer.shape[2])) layer = jnp.sum(layer, axis=2) layers.append(layer) print(f'Input shape at level {level}: {layer.shape}') return layers
class Decoder(nn.Module): ' States to Images Decoder.' c: Config @nn.compact def __call__(self, bottom_layer_output): '\n Arguments:\n bottom_layer_output : Tensor\n State tensor of shape (batch_size, timesteps, feature_dim)\n\n Returns:\n Output video of shape (batch_size, timesteps, 64, 64, out_channels)\n ' x = nn.Dense((self.c.channels_mult * 1024))(bottom_layer_output) x = jnp.reshape(x, ((- 1), 1, 1, x.shape[(- 1)])) ConvT = partial(nn.ConvTranspose, strides=(2, 2), padding='VALID') x = leaky_relu(ConvT((self.c.total_filters * 4), (5, 5))(x)) x = leaky_relu(ConvT((self.c.total_filters * 2), (5, 5))(x)) x = leaky_relu(ConvT(self.c.total_filters, (6, 6))(x)) x = nn.tanh(ConvT(self.c.channels, (6, 6))(x)) return x.reshape((bottom_layer_output.shape[:2] + x.shape[1:]))
def must_be(value): return field(default=value, metadata=dict(choices=[value]))
@dataclass class Config(): config: str datadir: str logdir: str levels: int = 3 tmp_abs_factor: int = 6 dec_stddev: float = 1.0 enc_dense_layers: int = 3 enc_dense_embed_size: int = 1000 cell_stoch_size: int = 20 cell_deter_size: int = 200 cell_embed_size: int = 200 cell_min_stddev: float = 0.0001 use_obs: Optional[str] = None channels_mult: int = 1 filters: int = 32 dataset: str = field(default='mmnist', metadata=dict(choices=['mmnist', 'minerl', 'mazes'])) seq_len: int = 100 eval_seq_len: int = 1000 channels: int = 1 lr: float = 0.0003 batch_size: int = 50 num_epochs: int = 300 clip_grad_norm_by: float = 10000 seed: int = np.random.randint(np.iinfo(np.int32).max) open_loop_ctx: int = 36 save_gifs: bool = True save_scalars_every: int = 1000 save_gifs_every: int = 1000 save_model_every: int = 1000 save_named_model_every: int = 5000 num_examples: int = 100 num_samples: int = 1 no_save_grid: bool = False cell_type: str = must_be('RSSMCell') cell_mean_only: str = must_be('false') cell_reset_state: str = must_be('false') beta: Optional[float] = must_be(None) free_nats: Optional[float] = must_be(None) kl_grad_post_perc: Optional[float] = must_be(None) num_val_batches: int = must_be(1) def config_file(self, eval): return ((Path(self.logdir).parent / 'config.yml') if eval else Path(self.config)) @property def _run_name(self): return f'{self.dataset}_cwvae_{self.cell_type.lower()}_{self.levels}l_f{self.tmp_abs_factor}_decsd{self.dec_stddev}_enchl{self.enc_dense_layers}_ences{self.enc_dense_embed_size}_edchnlmult{self.channels_mult}_ss{self.cell_stoch_size}_ds{self.cell_deter_size}_es{self.cell_embed_size}_seq{self.seq_len}_lr{self.lr}_bs{self.batch_size}' @property def exp_rootdir(self): return ((Path(self.logdir) / self.dataset) / self._run_name) def save(self): self.exp_rootdir.mkdir(parents=True, exist_ok=True) with (self.exp_rootdir / 'config.yml').open('w') as f: yaml.dump(asdict(self), f, default_flow_style=False) @property def total_filters(self): return (self.filters * self.channels_mult) @property def use_observations(self) -> List[bool]: if (self.use_obs is None): return ([True] * self.levels) assert (len(self.use_obs) == self.levels) return [dict(T=True, F=False)[c] for c in self.use_obs.upper()] @property def _dataset_name(self): return dict(minerl='minerl_navigate', mmnist='moving_mnist_2digit', mazes='gqn_mazes')[self.dataset] def load_dataset(self, eval=False): import tensorflow as tf import tensorflow_datasets as tfds np.random.seed(self.seed) tf.random.set_seed(self.seed) if (self.dataset == 'minerl'): import minerl_navigate elif (self.dataset == 'mmnist'): import datasets.moving_mnist elif (self.dataset == 'mazes'): import datasets.gqn_mazes d = tfds.load(self._dataset_name, data_dir=self.datadir, shuffle_files=(not eval)) d = d[('test' if eval else 'train')] d = d.map((lambda vid: (tf.cast(vid['video'], tf.float32) / 255.0))) seq_len = (self.eval_seq_len if eval else self.seq_len) if seq_len: def split_to_seq_len(seq): usable_len = (tf.shape(seq)[0] - (tf.shape(seq)[0] % seq_len)) seq = tf.reshape(seq[:usable_len], tf.concat([[(usable_len // seq_len), seq_len], tf.shape(seq)[1:]], (- 1))) return tf.data.Dataset.from_tensor_slices(seq) d = d.flat_map(split_to_seq_len) d = d.prefetch(tf.data.experimental.AUTOTUNE) if (not eval): d = d.repeat(self.num_epochs).shuffle((10 * self.batch_size)) d = d.batch(self.batch_size).prefetch(tf.data.experimental.AUTOTUNE) return tfds.as_numpy(d)
def parse_config(eval=False): p = ArgumentParser() for f in fields(Config): kwargs = (dict(action='store_true') if ((f.type is bool) and (not f.default)) else dict(default=f.default, type=f.type)) p.add_argument(f'--{f.name}', **kwargs, **f.metadata) c = Config(**vars(p.parse_args())) p.set_defaults(**yaml.full_load(c.config_file(eval).read_text())) return replace(c, **vars(p.parse_args()))
class GqnMazes(tfds.core.GeneratorBasedBuilder): 'DatasetBuilder for GQN Mazes dataset.' VERSION = tfds.core.Version('1.0.0') RELEASE_NOTES = {'1.0.0': 'Initial release.'} def _info(self) -> tfds.core.DatasetInfo: 'Returns the dataset metadata.' return tfds.core.DatasetInfo(builder=self, description=_DESCRIPTION, features=tfds.features.FeaturesDict(dict(video=tfds.features.Video(shape=(None, 64, 64, 3)))), supervised_keys=None, homepage='https://archive.org/details/gqn_mazes', citation=_CITATION) def _split_generators(self, dl_manager: tfds.download.DownloadManager): 'Returns SplitGenerators.' path = dl_manager.download_and_extract(_DOWNLOAD_URL) return dict(train=self._generate_examples((path / 'train')), test=self._generate_examples((path / 'test'))) def _generate_examples(self, path): 'Yields examples.' for f in path.glob('*.mp4'): (yield (str(f), dict(video=str(f.resolve()))))
class MovingMnist_2digit(tfds.core.GeneratorBasedBuilder): 'DatasetBuilder for Moving MNIST dataset.' VERSION = tfds.core.Version('1.0.0') RELEASE_NOTES = {'1.0.0': 'Initial release.'} def _info(self) -> tfds.core.DatasetInfo: 'Returns the dataset metadata.' return tfds.core.DatasetInfo(builder=self, description=_DESCRIPTION, features=tfds.features.FeaturesDict(dict(video=tfds.features.Video(shape=(None, 64, 64, 1)))), supervised_keys=None, homepage='https://archive.org/details/moving_mnist', citation=_CITATION) def _split_generators(self, dl_manager: tfds.download.DownloadManager): 'Returns SplitGenerators.' path = dl_manager.download_and_extract(_DOWNLOAD_URL) return dict(train=self._generate_examples((path / 'train-seq100')), test=self._generate_examples((path / 'test-seq1000'))) def _generate_examples(self, path): 'Yields examples.' for f in path.glob('*.mp4'): (yield (str(f), dict(video=str(f.resolve()))))
class SSFetcher(threading.Thread): def __init__(self, parent, init_offset=0, init_reshuffle_count=1, eos_sym=(- 1), skip_utterance=False, skip_utterance_predict_both=False): threading.Thread.__init__(self) self.parent = parent self.rng = numpy.random.RandomState(self.parent.seed) self.indexes = numpy.arange(parent.data_len) self.init_offset = init_offset self.init_reshuffle_count = init_reshuffle_count self.offset = 0 self.reshuffle_count = 0 self.eos_sym = eos_sym self.skip_utterance = skip_utterance self.skip_utterance_predict_both = skip_utterance_predict_both def apply_reshuffle(self): self.rng.shuffle(self.indexes) self.offset = 0 self.reshuffle_count += 1 def run(self): diter = self.parent while (self.reshuffle_count < self.init_reshuffle_count): self.apply_reshuffle() self.offset = self.init_offset while (not diter.exit_flag): last_batch = False dialogues = [] while (len(dialogues) < diter.batch_size): if (self.offset == diter.data_len): if (not diter.use_infinite_loop): last_batch = True break else: self.apply_reshuffle() index = self.indexes[self.offset] s = diter.data[index] if (len(s) > 0): if isinstance(s[0], list): s = [item for sublist in s for item in sublist] if (not self.skip_utterance): if ((diter.max_len == (- 1)) or (len(s) <= diter.max_len)): dialogues.append([s, self.offset, self.reshuffle_count]) else: s = copy.deepcopy(s) eos_indices = numpy.where((numpy.asarray(s) == self.eos_sym))[0] if (not (s[0] == self.eos_sym)): eos_indices = numpy.insert(eos_indices, 0, [self.eos_sym]) if (not (s[(- 1)] == self.eos_sym)): eos_indices = numpy.append(eos_indices, [self.eos_sym]) if (len(eos_indices) > 2): first_utterance_index = self.rng.randint(0, (len(eos_indices) - 2)) s_forward = s[eos_indices[first_utterance_index]:(eos_indices[(first_utterance_index + 2)] + 1)] s_backward_a = s[eos_indices[(first_utterance_index + 1)]:eos_indices[(first_utterance_index + 2)]] s_backward_b = s[eos_indices[first_utterance_index]:(eos_indices[(first_utterance_index + 1)] + 1)] if ((s_backward_a[(- 1)] == self.eos_sym) or (s_backward_b[0] == self.eos_sym)): s_backward = (s_backward_a + s_backward_b) else: s_backward = ((s_backward_a + [self.eos_sym]) + s_backward_b) else: s_forward = [self.eos_sym] s_backward = [self.eos_sym] if self.skip_utterance_predict_both: if ((diter.max_len == (- 1)) or (len(s_forward) <= diter.max_len)): dialogues.append([s_forward, self.offset, self.reshuffle_count]) if ((diter.max_len == (- 1)) or (len(s_backward) <= diter.max_len)): dialogues.append([s_backward, self.offset, self.reshuffle_count]) elif (self.rng.randint(0, 2) == 0): if ((diter.max_len == (- 1)) or (len(s_forward) <= diter.max_len)): dialogues.append([s_forward, self.offset, self.reshuffle_count]) elif ((diter.max_len == (- 1)) or (len(s_backward) <= diter.max_len)): dialogues.append([s_backward, self.offset, self.reshuffle_count]) self.offset += 1 if len(dialogues): diter.queue.put(dialogues) if last_batch: diter.queue.put(None) return
class SSIterator(object): def __init__(self, dialogue_file, batch_size, seed, max_len=(- 1), use_infinite_loop=True, init_offset=0, init_reshuffle_count=1, eos_sym=(- 1), skip_utterance=False, skip_utterance_predict_both=False): self.dialogue_file = dialogue_file self.batch_size = batch_size self.init_offset = init_offset self.init_reshuffle_count = init_reshuffle_count self.eos_sym = eos_sym self.skip_utterance = skip_utterance self.skip_utterance_predict_both = skip_utterance_predict_both args = locals() args.pop('self') self.__dict__.update(args) self.load_files() self.exit_flag = False def load_files(self): self.data = cPickle.load(open(self.dialogue_file, 'r')) self.data_len = len(self.data) logger.debug(('Data len is %d' % self.data_len)) def start(self): self.exit_flag = False self.queue = Queue.Queue(maxsize=1000) self.gather = SSFetcher(self, self.init_offset, self.init_reshuffle_count, self.eos_sym, self.skip_utterance, self.skip_utterance_predict_both) self.gather.daemon = True self.gather.start() def __del__(self): if hasattr(self, 'gather'): self.gather.exitFlag = True self.gather.join() def __iter__(self): return self def next(self): if self.exit_flag: return None batch = self.queue.get() if (not batch): self.exit_flag = True return batch
def sharedX(value, name=None, borrow=False, dtype=None): if (dtype is None): dtype = theano.config.floatX return theano.shared(theano._asarray(value, dtype=dtype), name=name, borrow=borrow)
def Adam(grads, lr=0.0002, b1=0.1, b2=0.001, e=1e-08): updates = [] varlist = [] i = sharedX(0.0) i_t = (i + 1.0) fix1 = (1.0 - ((1.0 - b1) ** i_t)) fix2 = (1.0 - ((1.0 - b2) ** i_t)) lr_t = (lr * (T.sqrt(fix2) / fix1)) for (p, g) in grads.items(): m = sharedX((p.get_value() * 0.0), name=(p.name + '_adam_optimizer_m')) v = sharedX((p.get_value() * 0.0), name=(p.name + '_adam_optimizer_v')) m_t = ((b1 * g) + ((1.0 - b1) * m)) v_t = ((b2 * T.sqr(g)) + ((1.0 - b2) * v)) g_t = (m_t / (T.sqrt(v_t) + e)) p_t = (p - (lr_t * g_t)) updates.append((m, m_t)) updates.append((v, v_t)) updates.append((p, p_t)) varlist.append(m) varlist.append(v) updates.append((i, i_t)) return (updates, varlist)
def safe_pickle(obj, filename): if os.path.isfile(filename): logger.info(('Overwriting %s.' % filename)) else: logger.info(('Saving to %s.' % filename)) with open(filename, 'wb') as f: cPickle.dump(obj, f, protocol=cPickle.HIGHEST_PROTOCOL)
class Model(object): def __init__(self): self.floatX = theano.config.floatX self.params = [] def save(self, filename): '\n Save the model to file `filename`\n ' vals = dict([(x.name, x.get_value()) for x in self.params]) numpy.savez(filename, **vals) def load(self, filename, parameter_strings_to_ignore=[]): '\n Load the model.\n\n Any parameter which has one of the strings inside parameter_strings_to_ignore as a substring,\n will not be loaded from the file (but instead initialized as a new model, which usually means random).\n ' vals = numpy.load(filename) for p in self.params: load_parameter = True for string_to_ignore in parameter_strings_to_ignore: if (string_to_ignore in p.name): logger.debug('Initializing parameter {} as in new model'.format(p.name)) load_parameter = False if load_parameter: if (p.name in vals): logger.debug('Loading {} of {}'.format(p.name, p.get_value(borrow=True).shape)) if (p.get_value().shape != vals[p.name].shape): raise Exception('Shape mismatch: {} != {} for {}'.format(p.get_value().shape, vals[p.name].shape, p.name)) p.set_value(vals[p.name]) else: logger.error('No parameter {} given: default initialization used'.format(p.name)) unknown = (set(vals.keys()) - {p.name for p in self.params}) if len(unknown): logger.error('Unknown parameters {} given'.format(unknown))
class Timer(object): def __init__(self): self.total = 0 def start(self): self.start_time = time.time() def finish(self): self.total += (time.time() - self.start_time)