r"""Applies each `module` in :attr:`modules` in parallel on arguments contained in :attr:`inputs` (positional) and :attr:`kwargs_tup` (keyword) on each of :attr:`devices`. Args: modules (Module): modules to be parallelized inputs (tensor): inputs to the modules d
(modules, inputs, kwargs_tup=None, devices=None)
| 78 | |
| 79 | |
| 80 | def parallel_apply(modules, inputs, kwargs_tup=None, devices=None): # pragma: no cover |
| 81 | r"""Applies each `module` in :attr:`modules` in parallel on arguments |
| 82 | contained in :attr:`inputs` (positional) and :attr:`kwargs_tup` (keyword) |
| 83 | on each of :attr:`devices`. |
| 84 | |
| 85 | Args: |
| 86 | modules (Module): modules to be parallelized |
| 87 | inputs (tensor): inputs to the modules |
| 88 | devices (list of int or torch.device): CUDA devices |
| 89 | |
| 90 | :attr:`modules`, :attr:`inputs`, :attr:`kwargs_tup` (if given), and |
| 91 | :attr:`devices` (if given) should all have same length. Moreover, each |
| 92 | element of :attr:`inputs` can either be a single object as the only argument |
| 93 | to a module, or a collection of positional arguments. |
| 94 | """ |
| 95 | assert len(modules) == len(inputs) |
| 96 | if kwargs_tup is not None: |
| 97 | assert len(modules) == len(kwargs_tup) |
| 98 | else: |
| 99 | kwargs_tup = ({},) * len(modules) |
| 100 | if devices is not None: |
| 101 | assert len(modules) == len(devices) |
| 102 | else: |
| 103 | devices = [None] * len(modules) |
| 104 | devices = list(map(lambda x: _get_device_index(x, True), devices)) |
| 105 | lock = threading.Lock() |
| 106 | results = {} |
| 107 | grad_enabled = torch.is_grad_enabled() |
| 108 | |
| 109 | def _worker(i, module, input, kwargs, device=None): |
| 110 | torch.set_grad_enabled(grad_enabled) |
| 111 | if device is None: |
| 112 | device = get_a_var(input).get_device() |
| 113 | try: |
| 114 | with torch.cuda.device(device): |
| 115 | # this also avoids accidental slicing of `input` if it is a Tensor |
| 116 | if not isinstance(input, (list, tuple)): |
| 117 | input = (input,) |
| 118 | |
| 119 | # --------------- |
| 120 | # CHANGE |
| 121 | if module.training: |
| 122 | output = module.training_step(*input, **kwargs) |
| 123 | |
| 124 | elif module.testing: |
| 125 | output = module.test_step(*input, **kwargs) |
| 126 | |
| 127 | else: |
| 128 | output = module.validation_step(*input, **kwargs) |
| 129 | # --------------- |
| 130 | |
| 131 | with lock: |
| 132 | results[i] = output |
| 133 | except Exception as e: |
| 134 | with lock: |
| 135 | results[i] = e |
| 136 | |
| 137 | # make sure each module knows what training state it's in... |
no test coverage detected