.. testsetup:: * import torch from lightning.pytorch import LightningModule ################## Structure Overview ################## TorchMetrics is a Metrics API created for easy metric development and usage in PyTorch and PyTorch Lightning. It is rigorously tested for all edge cases and includes a growing list of common metric implementations. The metrics API provides ``update()``, ``compute()``, ``reset()`` functions to the user. The metric :ref:`base class ` inherits :class:`torch.nn.Module` which allows us to call ``metric(...)`` directly. The ``forward()`` method of the base ``Metric`` class serves the dual purpose of calling ``update()`` on its input and simultaneously returning the value of the metric over the provided input. These metrics work with DDP in PyTorch and PyTorch Lightning by default. When ``.compute()`` is called in distributed mode, the internal state of each metric is synced and reduced across each process, so that the logic present in ``.compute()`` is applied to state information from all processes. This metrics API is independent of PyTorch Lightning. Metrics can directly be used in PyTorch as shown in the example: .. code-block:: python from torchmetrics.classification import BinaryAccuracy train_accuracy = BinaryAccuracy() valid_accuracy = BinaryAccuracy() for epoch in range(epochs): for x, y in train_data: y_hat = model(x) # training step accuracy batch_acc = train_accuracy(y_hat, y) print(f"Accuracy of batch{i} is {batch_acc}") for x, y in valid_data: y_hat = model(x) valid_accuracy.update(y_hat, y) # total accuracy over all training batches total_train_accuracy = train_accuracy.compute() # total accuracy over all validation batches total_valid_accuracy = valid_accuracy.compute() print(f"Training acc for epoch {epoch}: {total_train_accuracy}") print(f"Validation acc for epoch {epoch}: {total_valid_accuracy}") # Reset metric states after each epoch train_accuracy.reset() valid_accuracy.reset() .. note:: Metrics contain internal states that keep track of the data seen so far. Do not mix metric states across training, validation and testing. It is highly recommended to re-initialize the metric per mode as shown in the examples above. .. note:: Metric states are **not** added to the models ``state_dict`` by default. To change this, after initializing the metric, the method ``.persistent(mode)`` can be used to enable (``mode=True``) or disable (``mode=False``) this behaviour. .. note:: Due to specialized logic around metric states, we in general do **not** recommend that metrics are initialized inside other metrics (nested metrics), as this can lead to weird behaviour. Instead consider subclassing a metric or use ``torchmetrics.MetricCollection``. ******************* Metrics and devices ******************* Metrics are simple subclasses of :class:`~torch.nn.Module` and their metric states behave similar to buffers and parameters of modules. This means that metrics states should be moved to the same device as the input of the metric: .. code-block:: python from torchmetrics.classification import BinaryAccuracy target = torch.tensor([1, 1, 0, 0], device=torch.device("cuda", 0)) preds = torch.tensor([0, 1, 0, 0], device=torch.device("cuda", 0)) # Metric states are always initialized on cpu, and needs to be moved to # the correct device confmat = BinaryAccuracy().to(torch.device("cuda", 0)) out = confmat(preds, target) print(out.device) # cuda:0 However, when **properly defined** inside a :class:`~torch.nn.Module` or `LightningModule `_ the metric will be automatically moved to the same device as the module when using ``.to(device)``. Being **properly defined** means that the metric is correctly identified as a child module of the model (check ``.children()`` attribute of the model). Therefore, metrics cannot be placed in native python ``list`` and ``dict``, as they will not be correctly identified as child modules. Instead of ``list`` use :class:`~torch.nn.ModuleList` and instead of ``dict`` use :class:`~torch.nn.ModuleDict`. Furthermore, when working with multiple metrics the native `MetricCollection`_ module can also be used to wrap multiple metrics. .. testcode:: from torchmetrics import MetricCollection from torchmetrics.classification import BinaryAccuracy class MyModule(torch.nn.Module): def __init__(self): ... # valid ways metrics will be identified as child modules self.metric1 = BinaryAccuracy() self.metric2 = nn.ModuleList(BinaryAccuracy()) self.metric3 = nn.ModuleDict({'accuracy': BinaryAccuracy()}) self.metric4 = MetricCollection([BinaryAccuracy()]) # torchmetrics built-in collection class def forward(self, batch): data, target = batch preds = self(data) ... val1 = self.metric1(preds, target) val2 = self.metric2[0](preds, target) val3 = self.metric3['accuracy'](preds, target) val4 = self.metric4(preds, target) You can always check which device the metric is located on using the `.device` property. ***************************** Metrics and memory management ***************************** As stated before, metrics have states and those states take up a certain amount of memory depending on the metric. In general metrics can be divided into two categories when we talk about memory management: * Metrics with tensor states: These metrics only have states that are instances of :class:`~torch.Tensor`. When these kind of metrics are updated the values of those tensors are updated. Importantly the size of the tensors is **constant** meaning that regardless of how much data is passed to the metric, its memory footprint will not change. * Metrics with list states: These metrics have at least one state that is a list, which gets tensors appended as the metric is updated. Importantly the size of the list is therefore **not constant** and will grow. The growth depends on the particular metric (some metrics only need to store a single value per sample, some much more). You can always check the current metric state by accessing the `.metric_state` property, and checking if any of the states are lists. .. testcode:: import torch from torchmetrics.regression import SpearmanCorrCoef gen = torch.manual_seed(42) metric = SpearmanCorrCoef() metric(torch.rand(2,), torch.rand(2,)) print(metric.metric_state) metric(torch.rand(2,), torch.rand(2,)) print(metric.metric_state) .. testoutput:: :options: +NORMALIZE_WHITESPACE {'preds': [tensor([0.8823, 0.9150])], 'target': [tensor([0.3829, 0.9593])]} {'preds': [tensor([0.8823, 0.9150]), tensor([0.3904, 0.6009])], 'target': [tensor([0.3829, 0.9593]), tensor([0.2566, 0.7936])]} In general we have a few recommendations for memory management: * When done with a metric, we always recommend calling the `reset` method. The reason for this being that the python garbage collector can struggle to totally clean the metric states if this is not done. In the worst case, this can lead to a memory leak if multiple instances of the same metric for different purposes are created in the same script. * Better to always try to reuse the same instance of a metric instead of initializing a new one. Calling the `reset` method returns the metric to its initial state, and can therefore be used to reuse the same instance. However, we still highly recommend to use **different** instances from training, validation and testing. * If only the results on a batch level are needed e.g no aggregation or alternatively if you have a small dataset that fits into iteration of evaluation, we can recommend using the functional API instead as it does not keep an internal state and memory is therefore freed after each call. See :ref:`Metric kwargs` for different advanced settings for controlling the memory footprint of metrics. ************************** Saving and loading metrics ************************** Because metrics are essentially just a subclass of :class:`torch.nn.Module`, saving and loading metrics works in the same as any other `nn.Module`, with a key difference. Similar to `nn.Module` it is also recommended to save the state dict instead of the actual metric e.g.: .. code-block:: python # Instead of this torch.save(metric, "metric.pt") # do this torch.save(metric.state_dict(), "metric.pt") The key difference is that metric states are not automatically a part of the state dict. This is to make sure that torchmetrics is backward compatible with models that did not use the specific metrics when they were created. This behavior can be overwritten by using the `metric.persistent` method, which will mark all metric states to also be saved when `.state_dict` is called. Alternatively, for custom metrics, you can set the `persistent` argument when initializing the state in the `self.add_state` method. Therefore a correct example for saving and loading a metric would be: .. code-block:: python import torch from torchmetrics.classification import MulticlassAccuracy metric = MulticlassAccuracy(num_classes=5).to("cuda") metric.persistent(True) metric.update(torch.randint(5, (100,)).cuda(), torch.randint(5, (100,)).cuda()) torch.save(metric.state_dict(), "metric.pth") metric2 = MulticlassAccuracy(num_classes=5).to("cpu") metric2.load_state_dict(torch.load("metric.pth", map_location="cpu")) # These will match, but be on different devices print(metric.metric_state) print(metric2.metric_state) In the example, we also account for the initial metric state that is being saved on a different device than the metric it is being loaded into by using the `map_location` argument. *********************************************** Metrics in Distributed Data Parallel (DDP) mode *********************************************** When using metrics in `Distributed Data Parallel (DDP) `_ mode, one should be aware that DDP will add additional samples to your dataset if the size of your dataset is not equally divisible by ``batch_size * num_processors``. The added samples will always be replicates of datapoints already in your dataset. This is done to secure an equal load for all processes. However, this has the consequence that the calculated metric value will be slightly biased towards those replicated samples, leading to a wrong result. During training and/or validation this may not be important, however it is highly recommended when evaluating the test dataset to only run on a single gpu or use a `join `_ context in conjunction with DDP to prevent this behaviour. **************************** Metrics and 16-bit precision **************************** Most metrics in our collection can be used with 16-bit precision (``torch.half``) tensors. However, we have found the following limitations: * In general ``pytorch`` had better support for 16-bit precision much earlier on GPU than CPU. Therefore, we recommend that anyone that want to use metrics with half precision on CPU, upgrade to at least pytorch v1.6 where support for operations such as addition, subtraction, multiplication etc. was added. * Some metrics does not work at all in half precision on CPU. We have explicitly stated this in their docstring, but they are also listed below: - :ref:`image/peak_signal_noise_ratio:Peak Signal-to-Noise Ratio (PSNR)` - :ref:`image/structural_similarity:Structural Similarity Index Measure (SSIM)` - :ref:`regression/kl_divergence:KL Divergence` You can always check the precision/dtype of the metric by checking the `.dtype` property. ***************** Metric Arithmetic ***************** Metrics support most of python built-in operators for arithmetic, logic and bitwise operations. For example for a metric that should return the sum of two different metrics, implementing a new metric is an overhead that is not necessary. It can now be done with: .. code-block:: python first_metric = MyFirstMetric() second_metric = MySecondMetric() new_metric = first_metric + second_metric ``new_metric.update(*args, **kwargs)`` now calls update of ``first_metric`` and ``second_metric``. It forwards all positional arguments but forwards only the keyword arguments that are available in respective metric's update declaration. Similarly ``new_metric.compute()`` now calls compute of ``first_metric`` and ``second_metric`` and adds the results up. It is important to note that all implemented operations always return a new metric object. This means that the line ``first_metric == second_metric`` will not return a bool indicating if ``first_metric`` and ``second_metric`` is the same metric, but will return a new metric that checks if the ``first_metric.compute() == second_metric.compute()``. This pattern is implemented for the following operators (with ``a`` being metrics and ``b`` being metrics, tensors, integer or floats): * Addition (``a + b``) * Bitwise AND (``a & b``) * Equality (``a == b``) * Floordivision (``a // b``) * Greater Equal (``a >= b``) * Greater (``a > b``) * Less Equal (``a <= b``) * Less (``a < b``) * Matrix Multiplication (``a @ b``) * Modulo (``a % b``) * Multiplication (``a * b``) * Inequality (``a != b``) * Bitwise OR (``a | b``) * Power (``a ** b``) * Subtraction (``a - b``) * True Division (``a / b``) * Bitwise XOR (``a ^ b``) * Absolute Value (``abs(a)``) * Inversion (``~a``) * Negative Value (``neg(a)``) * Positive Value (``pos(a)``) * Indexing (``a[0]``) .. note:: Some of these operations are only fully supported from Pytorch v1.4 and onwards, explicitly we found: ``add``, ``mul``, ``rmatmul``, ``rsub``, ``rmod`` **************** MetricCollection **************** In many cases it is beneficial to evaluate the model output by multiple metrics. In this case the ``MetricCollection`` class may come in handy. It accepts a sequence of metrics and wraps these into a single callable metric class, with the same interface as any other metric. Example: .. testcode:: from torchmetrics import MetricCollection from torchmetrics.classification import MulticlassAccuracy, MulticlassPrecision, MulticlassRecall target = torch.tensor([0, 2, 0, 2, 0, 1, 0, 2]) preds = torch.tensor([2, 1, 2, 0, 1, 2, 2, 2]) metric_collection = MetricCollection([ MulticlassAccuracy(num_classes=3, average="micro"), MulticlassPrecision(num_classes=3, average="macro"), MulticlassRecall(num_classes=3, average="macro") ]) print(metric_collection(preds, target)) .. testoutput:: :options: +NORMALIZE_WHITESPACE {'MulticlassAccuracy': tensor(0.1250), 'MulticlassPrecision': tensor(0.0667), 'MulticlassRecall': tensor(0.1111)} Similarly it can also reduce the amount of code required to log multiple metrics inside your LightningModule. In most cases we just have to replace ``self.log`` with ``self.log_dict``. .. testcode:: from torchmetrics import MetricCollection from torchmetrics.classification import MulticlassAccuracy, MulticlassPrecision, MulticlassRecall class MyModule(LightningModule): def __init__(self, num_classes: int): super().__init__() metrics = MetricCollection([ MulticlassAccuracy(num_classes), MulticlassPrecision(num_classes), MulticlassRecall(num_classes) ]) self.train_metrics = metrics.clone(prefix='train_') self.valid_metrics = metrics.clone(prefix='val_') def training_step(self, batch, batch_idx): logits = self(x) # ... output = self.train_metrics(logits, y) # use log_dict instead of log # metrics are logged with keys: train_Accuracy, train_Precision and train_Recall self.log_dict(output) def validation_step(self, batch, batch_idx): logits = self(x) # ... self.valid_metrics.update(logits, y) def on_validation_epoch_end(self): # use log_dict instead of log # metrics are logged with keys: val_Accuracy, val_Precision and val_Recall output = self.valid_metrics.compute() self.log_dict(output) # remember to reset metrics at the end of the epoch self.valid_metrics.reset() .. note:: `MetricCollection` as default assumes that all the metrics in the collection have the same call signature. If this is not the case, input that should be given to different metrics can given as keyword arguments to the collection. An additional advantage of using the ``MetricCollection`` object is that it will automatically try to reduce the computations needed by finding groups of metrics that share the same underlying metric state. If such a group of metrics is found only one of them is actually updated and the updated state will be broadcasted to the rest of the metrics within the group. In the example above, this will lead to a 2-3x lower computational cost compared to disabling this feature in the case of the validation metrics where only ``update`` is called (this feature does not work in combination with ``forward``). However, this speedup comes with a fixed cost upfront, where the state-groups have to be determined after the first update. In case the groups are known beforehand, these can also be set manually to avoid this extra cost of the dynamic search. See the *compute_groups* argument in the class docs below for more information on this topic. .. autoclass:: torchmetrics.MetricCollection :exclude-members: update, compute, forward *************** Metric wrappers *************** In some cases it is beneficial to transform the output of one metric in some way or add additional logic. For this we have implemented a few *Wrapper* metrics. Wrapper metrics always take another :class:`~torchmetrics.Metric` or ( :class:`~torchmetrics.MetricCollection`) as input and wraps it in some way. A good example of this is the :class:`~torchmetrics.wrappers.ClasswiseWrapper` that allows for easy altering the output of certain classification metrics to also include label information. .. testcode:: from torchmetrics.classification import MulticlassAccuracy from torchmetrics.wrappers import ClasswiseWrapper # creating metrics base_metric = MulticlassAccuracy(num_classes=3, average=None) wrapped_metric = ClasswiseWrapper(base_metric, labels=["cat", "dog", "fish"]) # sample prediction and GT target = torch.tensor([0, 2, 0, 2, 0, 1, 0, 2]) preds = torch.tensor([2, 1, 2, 0, 1, 2, 2, 2]) # showing the metric results print(base_metric(preds, target)) # this returns a simple tensor without label info print(wrapped_metric(preds, target)) # this returns a dict with label info .. testoutput:: :options: +NORMALIZE_WHITESPACE tensor([0.0000, 0.0000, 0.3333]) {'multiclassaccuracy_cat': tensor(0.), 'multiclassaccuracy_dog': tensor(0.), 'multiclassaccuracy_fish': tensor(0.3333)} Another good example of wrappers is the :class:`~torchmetrics.wrappers.BootStrapper` that allows for easy bootstrapping of metrics e.g. computation of confidence intervals by resampling of input data. .. testcode:: from torchmetrics.classification import MulticlassAccuracy from torchmetrics.wrappers import BootStrapper # creating metrics wrapped_metric = BootStrapper(MulticlassAccuracy(num_classes=3)) # sample prediction and GT target = torch.tensor([0, 2, 0, 2, 0, 1, 0, 2]) preds = torch.tensor([2, 1, 2, 0, 1, 2, 2, 2]) # showing the metric results print(wrapped_metric(preds, target)) # this returns a dict with label info .. testoutput:: :options: +NORMALIZE_WHITESPACE {'mean': tensor(0.1476), 'std': tensor(0.0613)} You can see all implemented wrappers under the wrapper section of the API docs. **************************** Module vs Functional Metrics **************************** The functional metrics follow the simple paradigm input in, output out. This means they don't provide any advanced mechanisms for syncing across DDP nodes or aggregation over batches. They simply compute the metric value based on the given inputs. Also, the integration within other parts of PyTorch Lightning will never be as tight as with the Module-based interface. If you look for just computing the values, the functional metrics are the way to go. However, if you are looking for the best integration and user experience, please consider also using the Module interface. ***************************** Metrics and differentiability ***************************** Metrics support backpropagation, if all computations involved in the metric calculation are differentiable. All modular metric classes have the property ``is_differentiable`` that determines if a metric is differentiable or not. However, note that the cached state is detached from the computational graph and cannot be back-propagated. Not doing this would mean storing the computational graph for each update call, which can lead to out-of-memory errors. In practice this means that: .. code-block:: python MyMetric.is_differentiable # returns True if metric is differentiable metric = MyMetric() val = metric(pred, target) # this value can be back-propagated val = metric.compute() # this value cannot be back-propagated A functional metric is differentiable if its corresponding modular metric is differentiable. *************************************** Metrics and hyperparameter optimization *************************************** If you want to directly optimize a metric it needs to support backpropagation (see section above). However, if you are just interested in using a metric for hyperparameter tuning and are not sure if the metric should be maximized or minimized, all modular metric classes have the ``higher_is_better`` property that can be used to determine this: .. code-block:: python # returns True because accuracy is optimal when it is maximized torchmetrics.classification.Accuracy.higher_is_better # returns False because the mean squared error is optimal when it is minimized torchmetrics.MeanSquaredError.higher_is_better .. _Metric kwargs: ************************ Advanced metric settings ************************ The following is a list of additional arguments that can be given to any metric class (in the ``**kwargs`` argument) that will alter how metric states are stored and synced. If you are running metrics on GPU and are encountering that you are running out of GPU VRAM then the following argument can help: - ``compute_on_cpu``: will automatically move the metric states to cpu after calling ``update``, making sure that GPU memory is not filling up. The consequence will be that the ``compute`` method will be called on CPU instead of GPU. Only applies to metric states that are lists. - ``compute_with_cache``: This argument indicates if the result after calling the ``compute`` method should be cached. By default this is ``True`` meaning that repeated calls to ``compute`` (with no change to the metric state in between) does not recompute the metric but just returns the cache. By setting it to ``False`` the metric will be recomputed every time ``compute`` is called, but it can also help clean up a bit of memory. If you are running in a distributed environment, TorchMetrics will automatically take care of the distributed synchronization for you. However, the following three keyword arguments can be given to any metric class for further control over the distributed aggregation: - ``sync_on_compute``: This argument is an ``bool`` that indicates if the metrics should automatically sync between devices whenever the ``compute`` method is called. By default this is ``True``, but by setting this to ``False`` you can manually control when the synchronization happens. - ``dist_sync_on_step``: This argument is ``bool`` that indicates if the metric should synchronize between different devices every time ``forward`` is called. Setting this to ``True`` is in general not recommended as synchronization is an expensive operation to do after each batch. - ``process_group``: By default we synchronize across the *world* i.e. all processes being computed on. You can provide an ``torch._C._distributed_c10d.ProcessGroup`` in this argument to specify exactly what devices should be synchronized over. - ``dist_sync_fn``: By default we use :func:`torch.distributed.all_gather` to perform the synchronization between devices. Provide another callable function for this argument to perform custom distributed synchronization.