private IEnumerable <EvalData> OfflineEvaluateInternal(TrainerResult trainerResult)
{
this.performanceCounters.Stage4_Evaluation_PerSec.Increment();
this.performanceCounters.Stage4_Evaluation_Total.Increment();
if (trainerResult == null)
{
this.telemetry.TrackTrace($"Received invalid data: trainerResult is null");
yield break;
}
if (trainerResult.Label == null)
{
this.telemetry.TrackTrace($"Received invalid data: trainerResult.Label is null");
yield break;
}
if (trainerResult.ProgressivePrediction == null)
{
this.telemetry.TrackTrace($"Received invalid data: trainerResult.ProgressivePrediction is null");
yield break;
}
var pi_a_x = trainerResult.Probabilities[trainerResult.Label.Action - 1];
var p_a_x = trainerResult.Label.Probability * (1 - trainerResult.ProbabilityOfDrop);
yield return(new EvalData
{
PolicyName = "Latest Policy",
JSON = JsonConvert.SerializeObject(
new
{
name = "Latest Policy",
// calcuate expectation under current randomized policy (using current exploration strategy)
// VW action is 0-based, label Action is 1 based
cost = (trainerResult.Label.Cost * pi_a_x) / p_a_x,
prob = pi_a_x / p_a_x
})
});
// the one currently running
yield return(new EvalData
{
PolicyName = "Deployed Policy",
JSON = JsonConvert.SerializeObject(
new
{
name = "Deployed Policy",
cost = trainerResult.Label.Cost,
prob = trainerResult.Label.Probability
})
});
for (int action = 1; action <= trainerResult.ProgressivePrediction.Length; action++)
{
string tag;
if (!trainerResult.ActionsTags.TryGetValue(action, out tag))
{
tag = action.ToString(CultureInfo.InvariantCulture);
}
var name = $"Constant Policy {tag}";
yield return(new EvalData
{
PolicyName = name,
JSON = JsonConvert.SerializeObject(
new
{
name = name,
cost = VowpalWabbitContextualBanditUtil.GetUnbiasedCost(trainerResult.Label.Action, (uint)action, trainerResult.Label.Cost, trainerResult.Label.Probability),
prob = trainerResult.Label.Action == action ? 1 / (trainerResult.Probabilities[action - 1] * (1 - trainerResult.ProbabilityOfDrop)) : 0
})
});
}
}
private IEnumerable <EvalEventData> OfflineEvaluateInternal(TrainerResult trainerResult)
{
this.performanceCounters.Stage4_Evaluation_PerSec.Increment();
this.performanceCounters.Stage4_Evaluation_Total.Increment();
if (trainerResult == null)
{
yield break;
}
if (trainerResult.Label == null)
{
this.telemetry.TrackTrace($"Received invalid data: trainerResult.Label is null");
yield break;
}
if (trainerResult.ProgressiveProbabilities == null)
{
this.telemetry.TrackTrace($"Received invalid data: trainerResult.Probabilities is null");
yield break;
}
var pi_a_x = trainerResult.ProgressiveProbabilities[trainerResult.Label.Action - 1];
var p_a_x = trainerResult.Label.Probability * (1 - trainerResult.ProbabilityOfDrop);
// the latest one we're currently training
yield return(new EvalEventData
{
Name = "Latest Policy",
// calcuate expectation under current randomized policy (using current exploration strategy)
// VW action is 0-based, label Action is 1 based
WeightedCost = (trainerResult.Label.Cost * pi_a_x) / p_a_x,
ImportanceWeight = pi_a_x / p_a_x
});
// the one currently running
yield return(new EvalEventData
{
Name = "Deployed Policy",
WeightedCost = trainerResult.Label.Cost,
ImportanceWeight = 1 // for deployed policy just use the observed cost
});
// Default = choosing the action that's supplied by caller
yield return(new EvalEventData
{
Name = "Default Policy",
WeightedCost = VowpalWabbitContextualBanditUtil.GetUnbiasedCost(trainerResult.Label.Action, (uint)1, trainerResult.Label.Cost, trainerResult.Label.Probability),
ImportanceWeight = trainerResult.Label.Action == 1 ? 1 / (trainerResult.ObservedProbabilities[0] * (1 - trainerResult.ProbabilityOfDrop)) : 0
});
// per action tag policies
for (int action = 1; action <= trainerResult.ProgressiveRanking.Length; action++)
{
string tag;
if (!trainerResult.ActionsTags.TryGetValue(action, out tag))
{
tag = action.ToString(CultureInfo.InvariantCulture);
}
var name = $"Constant Policy {tag}";
yield return(new EvalEventData
{
Name = name,
WeightedCost = VowpalWabbitContextualBanditUtil.GetUnbiasedCost(trainerResult.Label.Action, (uint)action, trainerResult.Label.Cost, trainerResult.Label.Probability),
ImportanceWeight = trainerResult.Label.Action == action ? 1 / (trainerResult.ObservedProbabilities[action - 1] * (1 - trainerResult.ProbabilityOfDrop)) : 0
});
}
}
private IEnumerable <EvalData> OfflineEvaluateInternal(TrainerResult trainerResult)
{
this.performanceCounters.Stage4_Evaluation_PerSec.Increment();
this.performanceCounters.Stage4_Evaluation_Total.Increment();
if (trainerResult == null)
{
this.telemetry.TrackTrace($"Received invalid data: trainerResult is null");
yield break;
}
if (trainerResult.Label == null)
{
this.telemetry.TrackTrace($"Received invalid data: trainerResult.Label is null");
yield break;
}
if (trainerResult.ProgressivePrediction == null)
{
this.telemetry.TrackTrace($"Received invalid data: trainerResult.ProgressivePrediction is null");
yield break;
}
yield return(new EvalData
{
PolicyName = "Latest Policy",
JSON = JsonConvert.SerializeObject(
new
{
name = "Latest Policy",
// calcuate expectation under current randomized policy (using current exploration strategy)
// VW action is 0-based, label Action is 1 based
cost = trainerResult.ProgressivePrediction
.Sum(ap => ap.Score * VowpalWabbitContextualBanditUtil.GetUnbiasedCost(trainerResult.Label.Action, ap.Action + 1, trainerResult.Label.Cost, trainerResult.Label.Probability)),
prob = trainerResult.ProgressivePrediction
.Sum(ap => ap.Score / (trainerResult.Probabilities[ap.Action] * (1 - trainerResult.ProbabilityOfDrop)))
})
});
// the one currently running
yield return(new EvalData
{
PolicyName = "Deployed Policy",
JSON = JsonConvert.SerializeObject(
new
{
name = "Deployed Policy",
cost = trainerResult.Label.Cost,
prob = trainerResult.Label.Probability
})
});
for (int action = 1; action <= trainerResult.ProgressivePrediction.Length; action++)
{
yield return new EvalData
{
PolicyName = $"Constant Policy {action}",
JSON = JsonConvert.SerializeObject(
new
{
name = $"Constant Policy {action}",
cost = VowpalWabbitContextualBanditUtil.GetUnbiasedCost(trainerResult.Label.Action, (uint)action, trainerResult.Label.Cost, trainerResult.Label.Probability),
prob = trainerResult.Probabilities[action - 1] * (1 - trainerResult.ProbabilityOfDrop)
})
}
}
;
}
