o yi,@sddlmZddlmZmZmZmZmZddlZddlm Z ddl m Z ddl m Z ddlmZddlmZGd d d e ZdS) )deepcopy)AnyDictListTupleUnionN)Tensor) ModuleList)MetricCollection)Metric)rank_zero_warncs0eZdZdZd!deeefdeeeefddffdd Z e de fd d Z d"d d Z d"d dZd"ddZdefddZdeeeeefffddZd"ddZd"ddZ d#dededeeee fedeeeedffeeeeedffeeee dfffffddZdeddfdd ZZS)$ MetricTrackeraHA wrapper class that can help keeping track of a metric or metric collection over time and implement useful methods. The wrapper implements the standard ``.update()``, ``.compute()``, ``.reset()`` methods that just calls corresponding method of the currently tracked metric. However, the following additional methods are provided: -``MetricTracker.n_steps``: number of metrics being tracked -``MetricTracker.increment()``: initialize a new metric for being tracked -``MetricTracker.compute_all()``: get the metric value for all steps -``MetricTracker.best_metric()``: returns the best value Args: metric: instance of a ``torchmetrics.Metric`` or ``torchmetrics.MetricCollection`` to keep track of at each timestep. maximize: either single bool or list of bool indicating if higher metric values are better (``True``) or lower is better (``False``). Example (single metric): >>> from torchmetrics import MetricTracker >>> from torchmetrics.classification import MulticlassAccuracy >>> _ = torch.manual_seed(42) >>> tracker = MetricTracker(MulticlassAccuracy(num_classes=10, average='micro')) >>> for epoch in range(5): ... tracker.increment() ... for batch_idx in range(5): ... preds, target = torch.randint(10, (100,)), torch.randint(10, (100,)) ... tracker.update(preds, target) ... print(f"current acc={tracker.compute()}") current acc=0.1120000034570694 current acc=0.08799999952316284 current acc=0.12600000202655792 current acc=0.07999999821186066 current acc=0.10199999809265137 >>> best_acc, which_epoch = tracker.best_metric(return_step=True) >>> best_acc # doctest: +ELLIPSIS 0.1260... >>> which_epoch 2 >>> tracker.compute_all() tensor([0.1120, 0.0880, 0.1260, 0.0800, 0.1020]) Example (multiple metrics using MetricCollection): >>> from torchmetrics import MetricTracker, MetricCollection, MeanSquaredError, ExplainedVariance >>> _ = torch.manual_seed(42) >>> tracker = MetricTracker(MetricCollection([MeanSquaredError(), ExplainedVariance()]), maximize=[False, True]) >>> for epoch in range(5): ... tracker.increment() ... for batch_idx in range(5): ... preds, target = torch.randn(100), torch.randn(100) ... tracker.update(preds, target) ... print(f"current stats={tracker.compute()}") # doctest: +NORMALIZE_WHITESPACE current stats={'MeanSquaredError': tensor(1.8218), 'ExplainedVariance': tensor(-0.8969)} current stats={'MeanSquaredError': tensor(2.0268), 'ExplainedVariance': tensor(-1.0206)} current stats={'MeanSquaredError': tensor(1.9491), 'ExplainedVariance': tensor(-0.8298)} current stats={'MeanSquaredError': tensor(1.9800), 'ExplainedVariance': tensor(-0.9199)} current stats={'MeanSquaredError': tensor(2.2481), 'ExplainedVariance': tensor(-1.1622)} >>> from pprint import pprint >>> best_res, which_epoch = tracker.best_metric(return_step=True) >>> pprint(best_res) # doctest: +ELLIPSIS {'ExplainedVariance': -0.829..., 'MeanSquaredError': 1.821...} >>> which_epoch {'MeanSquaredError': 0, 'ExplainedVariance': 2} >>> pprint(tracker.compute_all()) {'ExplainedVariance': tensor([-0.8969, -1.0206, -0.8298, -0.9199, -1.1622]), 'MeanSquaredError': tensor([1.8218, 2.0268, 1.9491, 1.9800, 2.2481])} TmetricmaximizereturnNcstt|ttfstd|||_t|ttfs!t dt|tr7t|tr7t |t |kr7t dt|trEt|tsEt d||_ d|_ dS)Nz[Metric arg need to be an instance of a torchmetrics `Metric` or `MetricCollection` but got zBArgument `maximize` should either be a single bool or list of boolzOThe len of argument `maximize` should match the length of the metric collectionzLArgument `maximize` should be a single bool when `metric` is a single MetricF) super__init__ isinstancer r TypeError _base_metricboollist ValueErrorlenr_increment_called)selfrr __class__Q/home/ubuntu/.local/lib/python3.10/site-packages/torchmetrics/wrappers/tracker.pyr^s $ zMetricTracker.__init__cCs t|dS)z=Returns the number of times the tracker has been incremented.)rrrrrn_stepsps zMetricTracker.n_stepscCsd|_|t|jdS)zECreates a new instance of the input metric that will be updated next.TN)rappendrrr!rrr incrementuszMetricTracker.incrementcOs|d|d|i|S)z2Calls forward of the current metric being tracked.forward)_check_for_incrementrargskwargsrrrr%zs zMetricTracker.forwardcOs"|d|dj|i|dS)z)Updates the current metric being tracked.updater&N)r'r+r(rrrr+s zMetricTracker.updatecCs|d|dS)z1Call compute of the current metric being tracked.computer&)r'r,r!rrrr,s  zMetricTracker.computecsT|dddt|Dt|jtr#d}fdd|DStjddS)asCompute the metric value for all tracked metrics. Return: Either a single tensor if the tracked base object is a single metric, else if a metric collection is provide a dict of tensors will be returned Raises: ValueError: If `self.increment` have not been called before this method is called. compute_allcSs g|] \}}|dkr|qS)r)r,).0irrrr s z-MetricTracker.compute_all..rcs*i|]tjfddDddqS)csg|]}|qSrr)r.rkrrr0sz8MetricTracker.compute_all...rdim)torchstack)r.resr2r s*z-MetricTracker.compute_all..r4)r' enumeraterrr keysr6r7)rr<rr8rr-s   zMetricTracker.compute_allcCs|ddS)z(Resets the current metric being tracked.r&Nresetr!rrrr>szMetricTracker.resetcCs|D]}|qdS)z!Resets all metrics being tracked.Nr=)rrrrr reset_alls zMetricTracker.reset_allF return_stepNNc Cst|jtrN|jr tjntj}z||d\}}|r$||fWS|WSt yM}zt d|dt |rBWYd}~dSWYd}~dSd}~ww|}t|jt r[|jnt ||jg}ii}}t|D]N\}\} } z"||r|tjntj}|| d} | d| d|| <|| <Wqnt y}zt d| d|d t d\|| <|| <WYd}~qnd}~ww|r||fS|S) aiReturns the highest metric out of all tracked. Args: return_step: If ``True`` will also return the step with the highest metric value. Returns: Either a single value or a tuple, depends on the value of ``return_step`` and the object being tracked. - If a single metric is being tracked and ``return_step=False`` then a single tensor will be returned - If a single metric is being tracked and ``return_step=True`` then a 2-element tuple will be returned, where the first value is optimal value and second value is the corresponding optimal step - If a metric collection is being tracked and ``return_step=False`` then a single dict will be returned, where keys correspond to the different values of the collection and the values are the optimal metric value - If a metric collection is being bracked and ``return_step=True`` then a 2-element tuple will be returned where each is a dict, with keys corresponding to the different values of th collection and the values of the first dict being the optimal values and the values of the second dict being the optimal step In addtion the value in all cases may be ``None`` if the underlying metric does have a proper defined way of being optimal. rzDEncountered the following error when trying to get the best metric: z^this is probably due to the 'best' not being defined for this metric.Returning `None` instead.NrAr zNEncountered the following error when trying to get the best metric for metric :z_ this is probably due to the 'best' not being defined for this metric.Returning `None` instead.)rrr rr6maxminr-itemrr UserWarningrrr;items) rr@fnvalueidxerrorr9rr/r3voutrrr best_metricsL    "  & zMetricTracker.best_metricmethodcCs|js td|ddS)z`Method for checking that a metric that can be updated/used for computations has been intialized.`z8` cannot be called before `.increment()` has been calledN)rr)rrOrrrr'sz"MetricTracker._check_for_increment)T)rN)F)__name__ __module__ __qualname____doc__rr r rrrpropertyintr"r$r%r+rr,rrstrr-r>r?floatrrNr' __classcell__rrrrr s60C      * Hr )copyrtypingrrrrrr6rtorch.nnr torchmetrics.collectionsr torchmetrics.metricr torchmetrics.utilities.printsr r rrrrs