TrainerNoisyDqn.cs

using MyCaffe;
using MyCaffe.basecode;
using MyCaffe.common;
using MyCaffe.data;
using MyCaffe.layers;
using MyCaffe.param;
using MyCaffe.solvers;
using MyCaffe.trainers;
using MyCaffe.trainers.common;
using System;
using System.Collections;
using System.Collections.Generic;
using System.Diagnostics;
using System.Drawing;
using System.IO;
using System.Linq;
using System.Text;
using System.Threading.Tasks;
 
namespace MyCaffe.trainers.dqn.noisy.simple
{
    public class TrainerNoisyDqn<T> : IxTrainerRL, IDisposable
    {
        IxTrainerCallback m_icallback;
        CryptoRandom m_random = new CryptoRandom();
        MyCaffeControl<T> m_mycaffe;
        PropertySet m_properties;
 
        public TrainerNoisyDqn(MyCaffeControl<T> mycaffe, PropertySet properties, CryptoRandom random, IxTrainerCallback icallback)
        {
            m_icallback = icallback;
            m_mycaffe = mycaffe;
            m_properties = properties;
            m_random = random;
        }
 
        public void Dispose()
        {
        }
 
        public bool Initialize()
        {
            m_mycaffe.CancelEvent.Reset();
            m_icallback.OnInitialize(new InitializeArgs(m_mycaffe));
            return true;
        }
 
        public bool Shutdown(int nWait)
        {
            if (m_mycaffe != null)
            {
                m_mycaffe.CancelEvent.Set();
                wait(nWait);
            }
 
            m_icallback.OnShutdown();
 
            return true;
        }
 
        private void wait(int nWait)
        {
            int nWaitInc = 250;
            int nTotalWait = 0;
 
            while (nTotalWait < nWait)
            {
                m_icallback.OnWait(new WaitArgs(nWaitInc));
                nTotalWait += nWaitInc;
            }
        }
 
        public ResultCollection RunOne(int nDelay = 1000)
        {
            m_mycaffe.CancelEvent.Reset();
            DqnAgent<T> agent = new DqnAgent<T>(m_icallback, m_mycaffe, m_properties, m_random, Phase.TRAIN);
            agent.Run(Phase.TEST, 1, ITERATOR_TYPE.ITERATION, TRAIN_STEP.NONE);
            agent.Dispose();
            return null;
        }
 
        public byte[] Run(int nN, PropertySet runProp, out string type)
        {
            m_mycaffe.CancelEvent.Reset();
            DqnAgent<T> agent = new DqnAgent<T>(m_icallback, m_mycaffe, m_properties, m_random, Phase.RUN);
            byte[] rgResults = agent.Run(nN, out type);
            agent.Dispose();
 
            return rgResults;
        }
 
        public bool Test(int nN, ITERATOR_TYPE type)
        {
            int nDelay = 1000;
            string strProp = m_properties.ToString();
 
            // Turn off the num-skip to run at normal speed.
            strProp += "EnableNumSkip=False;";
            PropertySet properties = new PropertySet(strProp);
 
            m_mycaffe.CancelEvent.Reset();
            DqnAgent<T> agent = new DqnAgent<T>(m_icallback, m_mycaffe, properties, m_random, Phase.TRAIN);
            agent.Run(Phase.TEST, nN, type, TRAIN_STEP.NONE);
 
            agent.Dispose();
            Shutdown(nDelay);
 
            return true;
        }
 
        public bool Train(int nN, ITERATOR_TYPE type, TRAIN_STEP step)
        {
            m_mycaffe.CancelEvent.Reset();
            DqnAgent<T> agent = new DqnAgent<T>(m_icallback, m_mycaffe, m_properties, m_random, Phase.TRAIN);
            agent.Run(Phase.TRAIN, nN, type, step);
            agent.Dispose();
 
            return false;
        }
    }
 
 
    class DqnAgent<T> : IDisposable 
    {
        IxTrainerCallback m_icallback;
        Brain<T> m_brain;
        PropertySet m_properties;
        CryptoRandom m_random;
        float m_fGamma = 0.99f;
        bool m_bUseRawInput = true;
        double m_dfBetaStart = 0.4;
        int m_nBetaFrames = 1000;
        int m_nMemorySize = 10000;
        float m_fPriorityAlpha = 0.6f;
        int m_nUpdateTargetFreq = 1000;
        MEMTYPE m_memType = MEMTYPE.PRIORITY;
 
 
        public DqnAgent(IxTrainerCallback icallback, MyCaffeControl<T> mycaffe, PropertySet properties, CryptoRandom random, Phase phase)
        {
            m_icallback = icallback;
            m_brain = new Brain<T>(mycaffe, properties, random, phase);
            m_properties = properties;
            m_random = random;
 
            m_fGamma = (float)properties.GetPropertyAsDouble("Gamma", m_fGamma);
            m_bUseRawInput = properties.GetPropertyAsBool("UseRawInput", m_bUseRawInput);
        }
 
        public void Dispose()
        {
            if (m_brain != null)
            {
                m_brain.Dispose();
                m_brain = null;
            }
        }
 
        private double beta_by_frame(int nFrameIdx)
        {
            return Math.Min(1.0, m_dfBetaStart + nFrameIdx * (1.0 - m_dfBetaStart) / m_nBetaFrames);
        }
 
        private StateBase getData(Phase phase, int nAction, int nIdx)
        {
            GetDataArgs args = m_brain.getDataArgs(phase, nAction);
            m_icallback.OnGetData(args);
            args.State.Data.Index = nIdx;
            return args.State;
        }
 
        private void updateStatus(int nIteration, int nEpisodeCount, double dfRewardSum, double dfRunningReward, double dfLoss, double dfLearningRate, bool bModelUpdated)
        {
            GetStatusArgs args = new GetStatusArgs(0, nIteration, nEpisodeCount, 1000000, dfRunningReward, dfRewardSum, 0, 0, dfLoss, dfLearningRate, bModelUpdated);
            m_icallback.OnUpdateStatus(args);
        }
 
        public byte[] Run(int nIterations, out string type)
        {
            IxTrainerCallbackRNN icallback = m_icallback as IxTrainerCallbackRNN;
            if (icallback == null)
                throw new Exception("The Run method requires an IxTrainerCallbackRNN interface to convert the results into the native format!");
 
            StateBase s = getData(Phase.RUN, -1, 0);
            int nIteration = 0;
            List<float> rgResults = new List<float>();
            bool bDifferent;
 
            while (!m_brain.Cancel.WaitOne(0) && (nIterations == -1 || nIteration < nIterations))
            {
                // Preprocess the observation.
                SimpleDatum x = m_brain.Preprocess(s, m_bUseRawInput, out bDifferent);
 
                // Forward the policy network and sample an action.
                int action = m_brain.act(x, s.Clip, s.ActionCount);
 
                rgResults.Add(s.Data.TimeStamp.ToFileTime());
                rgResults.Add(s.Data.GetDataAtF(0));
                rgResults.Add(action);
 
                nIteration++;
 
                // Take the next step using the action
                s = getData(Phase.RUN, action, nIteration);
            }
 
            ConvertOutputArgs args = new ConvertOutputArgs(nIterations, rgResults.ToArray());
            icallback.OnConvertOutput(args);
 
            type = args.RawType;
            return args.RawOutput;
        }
 
        private bool isAtIteration(int nN, ITERATOR_TYPE type, int nIteration, int nEpisode)
        {
            if (nN == -1)
                return false;
 
            if (type == ITERATOR_TYPE.EPISODE)
            {
                if (nEpisode < nN)
                    return false;
 
                return true;
            }
            else
            {
                if (nIteration < nN)
                    return false;
 
                return true;
            }
        }
 
        public void Run(Phase phase, int nN, ITERATOR_TYPE type, TRAIN_STEP step)
        {
            IMemoryCollection iMemory = MemoryCollectionFactory.CreateMemory(m_memType, m_nMemorySize, m_fPriorityAlpha);
            int nIteration = 1;
            double dfRunningReward = 0;
            double dfEpisodeReward = 0;
            int nEpisode = 0;
            bool bDifferent = false;
 
            StateBase state = getData(phase, -1, -1);
            // Preprocess the observation.
            SimpleDatum x = m_brain.Preprocess(state, m_bUseRawInput, out bDifferent, true);
 
            // Set the initial target model to the current model.
            m_brain.UpdateTargetModel();
 
            while (!m_brain.Cancel.WaitOne(0) && !isAtIteration(nN, type, nIteration, nEpisode))
            {
                // Forward the policy network and sample an action.
                int action = m_brain.act(x, state.Clip, state.ActionCount);
 
                // Take the next step using the action
                StateBase state_next = getData(phase, action, nIteration);
 
                // Preprocess the next observation.
                SimpleDatum x_next = m_brain.Preprocess(state_next, m_bUseRawInput, out bDifferent);
                if (!bDifferent)
                    m_brain.Log.WriteLine("WARNING: The current state is the same as the previous state!");
 
                // Build up episode memory, using reward for taking the action.
                iMemory.Add(new MemoryItem(state, x, action, state_next, x_next, state_next.Reward, state_next.Done, nIteration, nEpisode));
                dfEpisodeReward += state_next.Reward;
 
                // Do the training
                if (iMemory.Count > m_brain.BatchSize)
                {
                    double dfBeta = beta_by_frame(nIteration);
                    MemoryCollection rgSamples = iMemory.GetSamples(m_random, m_brain.BatchSize, dfBeta);
                    m_brain.Train(nIteration, rgSamples, state.ActionCount);
                    iMemory.Update(rgSamples);
 
                    if (nIteration % m_nUpdateTargetFreq == 0)
                        m_brain.UpdateTargetModel();
                }
           
                if (state_next.Done)
                {
                    // Update reward running
                    dfRunningReward = dfRunningReward * 0.99 + dfEpisodeReward * 0.01;
 
                    nEpisode++;
                    updateStatus(nIteration, nEpisode, dfEpisodeReward, dfRunningReward, 0, 0, m_brain.GetModelUpdated());
 
                    state = getData(phase, -1, -1);
                    x = m_brain.Preprocess(state, m_bUseRawInput, out bDifferent, true);
                    dfEpisodeReward = 0;
                }
                else
                {
                    state = state_next;
                    x = x_next;
                }
               
                nIteration++;
            }
 
            iMemory.CleanUp();
        }
    }
 
    class Brain<T> : IDisposable, IxTrainerGetDataCallback
    {
        MyCaffeControl<T> m_mycaffe;
        Solver<T> m_solver;
        Net<T> m_netOutput;
        Net<T> m_netTarget;
        PropertySet m_properties;
        CryptoRandom m_random;
        SimpleDatum m_sdLast = null;
        DataTransformer<T> m_transformer;
        MemoryLossLayer<T> m_memLoss;
        Blob<T> m_blobActions = null;
        Blob<T> m_blobQValue = null;
        Blob<T> m_blobNextQValue = null;
        Blob<T> m_blobExpectedQValue = null;
        Blob<T> m_blobDone = null;
        Blob<T> m_blobLoss = null;
        Blob<T> m_blobWeights = null;
        BlobCollection<T> m_colAccumulatedGradients = new BlobCollection<T>();
        bool m_bUseAcceleratedTraining = false;
        double m_dfLearningRate;
        int m_nMiniBatch = 1;  
        float m_fGamma = 0.99f;
        int m_nBatchSize = 32;
        MemoryCollection m_rgSamples;
        int m_nActionCount = 2;
        bool m_bModelUpdated = false;
        Font m_font = null;
        Dictionary<Color, Tuple<Brush, Brush, Pen, Brush>> m_rgStyle = new Dictionary<Color, Tuple<Brush, Brush, Pen, Brush>>();
        List<SimpleDatum> m_rgX = new List<SimpleDatum>();
        float[] m_rgOverlay = null;
 
 
        public Brain(MyCaffeControl<T> mycaffe, PropertySet properties, CryptoRandom random, Phase phase)
        {
            m_mycaffe = mycaffe;
            m_solver = mycaffe.GetInternalSolver();
            m_netOutput = mycaffe.GetInternalNet(phase);
            m_netTarget = new Net<T>(m_mycaffe.Cuda, m_mycaffe.Log, m_netOutput.net_param, m_mycaffe.CancelEvent, null, phase);
            m_properties = properties;
            m_random = random;
 
            Blob<T> data = m_netOutput.blob_by_name("data");
            if (data == null)
                m_mycaffe.Log.FAIL("Missing the expected input 'data' blob!");
 
            m_nBatchSize = data.num;
 
            Blob<T> logits = m_netOutput.blob_by_name("logits");
            if (logits == null)
                m_mycaffe.Log.FAIL("Missing the expected input 'logits' blob!");
 
            m_nActionCount = logits.channels;
 
            m_transformer = m_mycaffe.DataTransformer;
            if (m_transformer == null)
            {
                TransformationParameter trans_param = new TransformationParameter();
                int nC = m_mycaffe.CurrentProject.Dataset.TrainingSource.Channels;
                int nH = m_mycaffe.CurrentProject.Dataset.TrainingSource.Height;
                int nW = m_mycaffe.CurrentProject.Dataset.TrainingSource.Width;
                m_transformer = new DataTransformer<T>(m_mycaffe.Cuda, m_mycaffe.Log, trans_param, phase, nC, nH, nW);
            }
            m_blobActions = new Blob<T>(m_mycaffe.Cuda, m_mycaffe.Log, false);
            m_blobQValue = new Blob<T>(m_mycaffe.Cuda, m_mycaffe.Log);
            m_blobNextQValue = new Blob<T>(m_mycaffe.Cuda, m_mycaffe.Log);
            m_blobExpectedQValue = new Blob<T>(m_mycaffe.Cuda, m_mycaffe.Log);
            m_blobDone = new Blob<T>(m_mycaffe.Cuda, m_mycaffe.Log, false);
            m_blobLoss = new Blob<T>(m_mycaffe.Cuda, m_mycaffe.Log);
            m_blobWeights = new Blob<T>(m_mycaffe.Cuda, m_mycaffe.Log, false);
 
            m_fGamma = (float)properties.GetPropertyAsDouble("Gamma", m_fGamma);
 
            m_memLoss = m_netOutput.FindLastLayer(LayerParameter.LayerType.MEMORY_LOSS) as MemoryLossLayer<T>;
            if (m_memLoss == null)
                m_mycaffe.Log.FAIL("Missing the expected MEMORY_LOSS layer!");
 
            double? dfRate = mycaffe.CurrentProject.GetSolverSettingAsNumeric("base_lr");
            if (dfRate.HasValue)
                m_dfLearningRate = dfRate.Value;
 
            m_nMiniBatch = m_properties.GetPropertyAsInt("MiniBatch", m_nMiniBatch);
            m_bUseAcceleratedTraining = properties.GetPropertyAsBool("UseAcceleratedTraining", false);
 
            if (m_nMiniBatch > 1)
            {
                m_colAccumulatedGradients = m_netOutput.learnable_parameters.Clone();
                m_colAccumulatedGradients.SetDiff(0);
            }
        }
 
        private void dispose(ref Blob<T> b)
        {
            if (b != null)
            {
                b.Dispose();
                b = null;
            }
        }
 
        public void Dispose()
        {
            dispose(ref m_blobActions);
            dispose(ref m_blobQValue);
            dispose(ref m_blobNextQValue);
            dispose(ref m_blobExpectedQValue);
            dispose(ref m_blobDone);
            dispose(ref m_blobLoss);
            dispose(ref m_blobWeights);
 
            if (m_colAccumulatedGradients != null)
            {
                m_colAccumulatedGradients.Dispose();
                m_colAccumulatedGradients = null;
            }
 
            if (m_netTarget != null)
            {
                m_netTarget.Dispose();
                m_netTarget = null;
            }
 
            if (m_font != null)
            {
                m_font.Dispose();
                m_font = null;
            }
 
            foreach (KeyValuePair<Color, Tuple<Brush, Brush, Pen, Brush>> kv in m_rgStyle)
            {
                kv.Value.Item1.Dispose();
                kv.Value.Item2.Dispose();
                kv.Value.Item3.Dispose();
                kv.Value.Item4.Dispose();
            }
 
            m_rgStyle.Clear();
        }
 
        public GetDataArgs getDataArgs(Phase phase, int nAction)
        {
            bool bReset = (nAction == -1) ? true : false;
            return new GetDataArgs(phase, 0, m_mycaffe, m_mycaffe.Log, m_mycaffe.CancelEvent, bReset, nAction, true, false, false, this);
        }
 
        public int BatchSize
        {
            get { return m_nBatchSize; }
        }
 
        public Log Log
        {
            get { return m_mycaffe.Log; }
        }
 
        public CancelEvent Cancel
        {
            get { return m_mycaffe.CancelEvent; }
        }
 
        public SimpleDatum Preprocess(StateBase s, bool bUseRawInput, out bool bDifferent, bool bReset = false)
        {
            bDifferent = false;
 
            SimpleDatum sd = new SimpleDatum(s.Data, true);
 
            if (!bUseRawInput)
            {
                if (bReset)
                    m_sdLast = null;
 
                if (m_sdLast == null)
                    sd.Zero();
                else
                    bDifferent = sd.Sub(m_sdLast);
 
                m_sdLast = new SimpleDatum(s.Data, true);
            }
            else
            {
                bDifferent = true;
            }
 
            sd.Tag = bReset;
 
            return sd;
        }
 
        public int act(SimpleDatum sd, SimpleDatum sdClip, int nActionCount)
        {
            setData(m_netOutput, sd, sdClip);
            m_netOutput.ForwardFromTo(0, m_netOutput.layers.Count - 2);
 
            Blob<T> output = m_netOutput.blob_by_name("logits");
            if (output == null)
                throw new Exception("Missing expected 'logits' blob!");
 
            // Choose greedy action
            return argmax(Utility.ConvertVecF<T>(output.mutable_cpu_data));
        }
 
        public bool GetModelUpdated()
        {
            bool bModelUpdated = m_bModelUpdated;
            m_bModelUpdated = false;
            return bModelUpdated;
        }
 
        public void UpdateTargetModel()
        {
            m_mycaffe.Log.Enable = false;
            m_netOutput.CopyTrainedLayersTo(m_netTarget);
            m_netOutput.CopyInternalBlobsTo(m_netTarget);
            m_mycaffe.Log.Enable = true;
            m_bModelUpdated = true;
        }
 
        public void Train(int nIteration, MemoryCollection rgSamples, int nActionCount)
        {
            m_rgSamples = rgSamples;
 
            if (m_nActionCount != nActionCount)
                throw new Exception("The logit output of '" + m_nActionCount.ToString() + "' does not match the action count of '" + nActionCount.ToString() + "'!");
 
            // Get next_q_values
            m_mycaffe.Log.Enable = false;
            setNextStateData(m_netTarget, rgSamples);
            m_netTarget.ForwardFromTo(0, m_netTarget.layers.Count - 2);
 
            setCurrentStateData(m_netOutput, rgSamples);
            m_memLoss.OnGetLoss += m_memLoss_ComputeTdLoss;
 
            if (m_nMiniBatch == 1)
            {
                m_solver.Step(1);
            }
            else
            {
                m_solver.Step(1, TRAIN_STEP.NONE, true, m_bUseAcceleratedTraining, true, true);
                m_colAccumulatedGradients.Accumulate(m_mycaffe.Cuda, m_netOutput.learnable_parameters, true);
 
                if (nIteration % m_nMiniBatch == 0)
                {
                    m_netOutput.learnable_parameters.CopyFrom(m_colAccumulatedGradients, true);
                    m_colAccumulatedGradients.SetDiff(0);
                    m_dfLearningRate = m_solver.ApplyUpdate(nIteration);
                    m_netOutput.ClearParamDiffs();
                }
            }
 
            m_memLoss.OnGetLoss -= m_memLoss_ComputeTdLoss;
            m_mycaffe.Log.Enable = true;
 
            resetNoise(m_netOutput);
            resetNoise(m_netTarget);
        }
 
        private void m_memLoss_ComputeTdLoss(object sender, MemoryLossLayerGetLossArgs<T> e)
        {
            MemoryCollection rgMem = m_rgSamples;
 
            Blob<T> q_values = m_netOutput.blob_by_name("logits");
            Blob<T> next_q_values = m_netTarget.blob_by_name("logits");
 
            float[] rgActions = rgMem.GetActionsAsOneHotVector(m_nActionCount);
            m_blobActions.ReshapeLike(q_values);
            m_blobActions.mutable_cpu_data = Utility.ConvertVec<T>(rgActions);
            m_blobQValue.ReshapeLike(q_values);
 
            // q_value = q_values.gather(1, action.unsqueeze(1)).squeeze(1)
            m_mycaffe.Cuda.mul(m_blobActions.count(), m_blobActions.gpu_data, q_values.gpu_data, m_blobQValue.mutable_gpu_data);
            reduce_sum_axis1(m_blobQValue);
 
            // next_q_value = next_q_values.max(1)[0]
            m_blobNextQValue.CopyFrom(next_q_values, false, true);
            reduce_argmax_axis1(m_blobNextQValue);
 
            // expected_q_values
            float[] rgRewards = rgMem.GetRewards();
            m_blobExpectedQValue.ReshapeLike(m_blobQValue);
            m_blobExpectedQValue.mutable_cpu_data = Utility.ConvertVec<T>(rgRewards);
 
            float[] rgDone = rgMem.GetInvertedDoneAsOneHotVector();
            m_blobDone.ReshapeLike(m_blobQValue);
            m_blobDone.mutable_cpu_data = Utility.ConvertVec<T>(rgDone);
 
            m_mycaffe.Cuda.mul(m_blobNextQValue.count(), m_blobNextQValue.gpu_data, m_blobDone.gpu_data, m_blobExpectedQValue.mutable_gpu_diff);           // next_q_val * (1- done)
            m_mycaffe.Cuda.mul_scalar(m_blobExpectedQValue.count(), m_fGamma, m_blobExpectedQValue.mutable_gpu_diff);                                      // gamma *  ^
            m_mycaffe.Cuda.add(m_blobExpectedQValue.count(), m_blobExpectedQValue.gpu_diff, m_blobExpectedQValue.gpu_data, m_blobExpectedQValue.gpu_data); // reward + ^
 
            // loss = (q_value - expected_q_value.detach()).pow(2) 
            m_blobLoss.ReshapeLike(m_blobQValue);
            m_mycaffe.Cuda.sub(m_blobQValue.count(), m_blobQValue.gpu_data, m_blobExpectedQValue.gpu_data, m_blobQValue.mutable_gpu_diff); // q_value - expected_q_value
            m_mycaffe.Cuda.powx(m_blobLoss.count(), m_blobQValue.gpu_diff, 2.0, m_blobLoss.mutable_gpu_data);                              // (q_value - expected_q_value)^2
 
            // loss = (q_value - expected_q_value.detach()).pow(2) * weights
            m_blobWeights.ReshapeLike(m_blobQValue);
            m_blobWeights.mutable_cpu_data = Utility.ConvertVec<T>(m_rgSamples.Priorities); // weights
            m_mycaffe.Cuda.mul(m_blobLoss.count(), m_blobLoss.gpu_data, m_blobWeights.gpu_data, m_blobLoss.mutable_gpu_data);               //    ^ * weights
 
            // prios = loss + 1e-5
            m_mycaffe.Cuda.copy(m_blobLoss.count(), m_blobLoss.gpu_data, m_blobLoss.mutable_gpu_diff);
            m_mycaffe.Cuda.add_scalar(m_blobLoss.count(), 1e-5, m_blobLoss.mutable_gpu_diff);
            double[] rgPrios = Utility.ConvertVec<T>(m_blobLoss.mutable_cpu_diff);
 
            for (int i = 0; i < rgPrios.Length; i++)
            {
                m_rgSamples.Priorities[i] = rgPrios[i];
            }
 
 
            //-------------------------------------------------------
            //  Calculate the gradient - unroll the operations
            //  (autograd - psha! how about manualgrad :-D)
            //-------------------------------------------------------
 
            // initial gradient
            double dfGradient = 1.0;
            if (m_memLoss.layer_param.loss_weight.Count > 0)
                dfGradient *= m_memLoss.layer_param.loss_weight[0];
 
            // mean gradient - expand and divide by batch count
            dfGradient /= m_blobLoss.count();
            m_blobLoss.SetDiff(dfGradient);
 
            // multiplication gradient - multiply by the other side.
            m_mycaffe.Cuda.mul(m_blobLoss.count(), m_blobLoss.gpu_diff, m_blobWeights.gpu_data, m_blobLoss.mutable_gpu_diff);
 
            // power gradient - multiply by the exponent.
            m_mycaffe.Cuda.mul_scalar(m_blobLoss.count(), 2.0, m_blobLoss.mutable_gpu_diff);
 
            // q_value - expected_q_value gradient
            m_mycaffe.Cuda.mul(m_blobLoss.count(), m_blobLoss.gpu_diff, m_blobQValue.gpu_diff, m_blobLoss.mutable_gpu_diff);
 
            // squeeze/gather gradient
            mul(m_blobLoss, m_blobActions, e.Bottom[0]);
 
            e.Loss = reduce_mean(m_blobLoss, false);
            e.EnableLossUpdate = false;
        }
 
        private void resetNoise(Net<T> net)
        {
            foreach (Layer<T> layer in net.layers)
            {
                if (layer.type == LayerParameter.LayerType.INNERPRODUCT)
                {
                    if (layer.layer_param.inner_product_param.enable_noise)
                        ((InnerProductLayer<T>)layer).ResetNoise();
                }
            }
        }
 
        private void mul(Blob<T> val, Blob<T> actions, Blob<T> result)
        {
            float[] rgVal = Utility.ConvertVecF<T>(val.mutable_cpu_diff);
            float[] rgActions = Utility.ConvertVecF<T>(actions.mutable_cpu_data);
            float[] rgResult = new float[rgActions.Length];
 
            for (int i = 0; i < actions.num; i++)
            {
                float fPred = rgVal[i];
 
                for (int j = 0; j < actions.channels; j++)
                {
                    int nIdx = (i * actions.channels) + j;
                    rgResult[nIdx] = rgActions[nIdx] * fPred; 
                }
            }
 
            result.mutable_cpu_diff = Utility.ConvertVec<T>(rgResult);
        }
 
        private float reduce_mean(Blob<T> b, bool bDiff)
        {
            float[] rg = Utility.ConvertVecF<T>((bDiff) ? b.mutable_cpu_diff : b.mutable_cpu_data);
            float fSum = rg.Sum(p => p);
            return fSum / rg.Length;
        }
 
        private void reduce_sum_axis1(Blob<T> b)
        {
            int nNum = b.shape(0);
            int nActions = b.shape(1);
            int nInnerCount = b.count(2);
            float[] rg = Utility.ConvertVecF<T>(b.mutable_cpu_data);
            float[] rgSum = new float[nNum * nInnerCount];
 
            for (int i = 0; i < nNum; i++)
            {
                for (int j = 0; j < nInnerCount; j++)
                {
                    float fSum = 0;
 
                    for (int k = 0; k < nActions; k++)
                    {
                        int nIdx = (i * nActions * nInnerCount) + (k * nInnerCount);
                        fSum += rg[nIdx + j];
                    }
 
                    int nIdxR = i * nInnerCount;
                    rgSum[nIdxR + j] = fSum;
                }
            }
 
            b.Reshape(nNum, nInnerCount, 1, 1);
            b.mutable_cpu_data = Utility.ConvertVec<T>(rgSum);
        }
 
        private void reduce_argmax_axis1(Blob<T> b)
        {
            int nNum = b.shape(0);
            int nActions = b.shape(1);
            int nInnerCount = b.count(2);
            float[] rg = Utility.ConvertVecF<T>(b.mutable_cpu_data);
            float[] rgMax = new float[nNum * nInnerCount];
 
            for (int i = 0; i < nNum; i++)
            {
                for (int j = 0; j < nInnerCount; j++)
                {
                    float fMax = -float.MaxValue;
 
                    for (int k = 0; k < nActions; k++)
                    {
                        int nIdx = (i * nActions * nInnerCount) + (k * nInnerCount);
                        fMax = Math.Max(fMax, rg[nIdx + j]);
                    }
 
                    int nIdxR = i * nInnerCount;
                    rgMax[nIdxR + j] = fMax;
                }
            }
 
            b.Reshape(nNum, nInnerCount, 1, 1);
            b.mutable_cpu_data = Utility.ConvertVec<T>(rgMax);
        }
 
        private int argmax(float[] rgProb, int nActionCount, int nSampleIdx)
        {
            float[] rgfProb = new float[nActionCount];
 
            for (int j = 0; j < nActionCount; j++)
            {
                int nIdx = (nSampleIdx * nActionCount) + j;
                rgfProb[j] = rgProb[nIdx];
            }
 
            return argmax(rgfProb);
        }
 
        private int argmax(float[] rgfAprob)
        {
            double fMax = -float.MaxValue;
            int nIdx = 0;
 
            for (int i = 0; i < rgfAprob.Length; i++)
            {
                if (rgfAprob[i] == fMax)
                {
                    if (m_random.NextDouble() > 0.5)
                        nIdx = i;
                }
                else if (fMax < rgfAprob[i])
                {
                    fMax = rgfAprob[i];
                    nIdx = i;
                }
            }
 
            return nIdx;
        }
 
        private void setData(Net<T> net, SimpleDatum sdData, SimpleDatum sdClip)
        {
            SimpleDatum[] rgData = new SimpleDatum[] { sdData };
            SimpleDatum[] rgClip = null;
 
            if (sdClip != null)
                rgClip = new SimpleDatum[] { sdClip };
 
            setData(net, rgData, rgClip);
        }
 
        private void setCurrentStateData(Net<T> net, MemoryCollection rgSamples)
        {
            List<SimpleDatum> rgData0 = rgSamples.GetCurrentStateData();
            List<SimpleDatum> rgClip0 = rgSamples.GetCurrentStateClip();
 
            SimpleDatum[] rgData = rgData0.ToArray();
            SimpleDatum[] rgClip = (rgClip0 != null) ? rgClip0.ToArray() : null;
 
            setData(net, rgData, rgClip);
        }
 
        private void setNextStateData(Net<T> net, MemoryCollection rgSamples)
        {
            List<SimpleDatum> rgData1 = rgSamples.GetNextStateData();
            List<SimpleDatum> rgClip1 = rgSamples.GetNextStateClip();
 
            SimpleDatum[] rgData = rgData1.ToArray();
            SimpleDatum[] rgClip = (rgClip1 != null) ? rgClip1.ToArray() : null;
 
            setData(net, rgData, rgClip);
        }
 
        private void setData(Net<T> net, SimpleDatum[] rgData, SimpleDatum[] rgClip)
        {
            Blob<T> data = net.blob_by_name("data");
 
            data.Reshape(rgData.Length, data.channels, data.height, data.width);
            m_transformer.Transform(rgData, data, m_mycaffe.Cuda, m_mycaffe.Log);
 
            if (rgClip != null)
            {
                Blob<T> clip = net.blob_by_name("clip");
 
                if (clip != null)
                {
                    clip.Reshape(rgClip.Length, rgClip[0].Channels, rgClip[0].Height, rgClip[0].Width);
                    m_transformer.Transform(rgClip, clip, m_mycaffe.Cuda, m_mycaffe.Log, true);
                }
            }
        }
 
        public void OnOverlay(OverlayArgs e)
        {
            Blob<T> logits = m_netOutput.blob_by_name("logits");
            if (logits == null)
                return;
 
            if (logits.num == 1)
                m_rgOverlay = Utility.ConvertVecF<T>(logits.mutable_cpu_data);
 
            if (m_rgOverlay == null)
                return;
 
            using (Graphics g = Graphics.FromImage(e.DisplayImage))
            {
                int nBorder = 30;
                int nWid = e.DisplayImage.Width - (nBorder * 2);
                int nWid1 = nWid / m_rgOverlay.Length;
                int nHt1 = (int)(e.DisplayImage.Height * 0.3);
                int nX = nBorder;
                int nY = e.DisplayImage.Height - nHt1;
                ColorMapper clrMap = new ColorMapper(0, m_rgOverlay.Length + 1, Color.Black, Color.Red);            
                float fMax = -float.MaxValue;
                int nMaxIdx = 0;
                float fMin1 = m_rgOverlay.Min(p => p);
                float fMax1 = m_rgOverlay.Max(p => p);
 
                for (int i=0; i<m_rgOverlay.Length; i++)
                {
                    if (fMin1 < 0 || fMax1 > 1)
                       m_rgOverlay[i] = (m_rgOverlay[i] - fMin1) / (fMax1 - fMin1);
 
                    if (m_rgOverlay[i] > fMax)
                    {
                        fMax = m_rgOverlay[i];
                        nMaxIdx = i;
                    }
                }
 
                for (int i = 0; i < m_rgOverlay.Length; i++)
                {
                    drawProbabilities(g, nX, nY, nWid1, nHt1, i, m_rgOverlay[i], fMin1, fMax1, clrMap.GetColor(i + 1), (i == nMaxIdx) ? true : false);
                    nX += nWid1;
                }
            }
        }
 
        private void drawProbabilities(Graphics g, int nX, int nY, int nWid, int nHt, int nAction, float fProb, float fMin, float fMax, Color clr, bool bMax)
        {
            string str = "";
 
            if (m_font == null)
                m_font = new Font("Century Gothic", 9.0f);
 
            if (!m_rgStyle.ContainsKey(clr))
            {
                Color clr1 = Color.FromArgb(128, clr);
                Brush br1 = new SolidBrush(clr1);
                Color clr2 = Color.FromArgb(64, clr);
                Pen pen = new Pen(clr2, 1.0f);
                Brush br2 = new SolidBrush(clr2);
                Brush brBright = new SolidBrush(clr);
                m_rgStyle.Add(clr, new Tuple<Brush, Brush, Pen, Brush>(br1, br2, pen, brBright));
            }
 
            Brush brBack = m_rgStyle[clr].Item1;
            Brush brFront = m_rgStyle[clr].Item2;
            Brush brTop = m_rgStyle[clr].Item4;
            Pen penLine = m_rgStyle[clr].Item3;
 
            if (fMin != 0 || fMax != 0)
            {
                str = "Action " + nAction.ToString() + " (" + fProb.ToString("N7") + ")";
            }
            else
            {
                str = "Action " + nAction.ToString() + " - No Probabilities";
            }
 
            SizeF sz = g.MeasureString(str, m_font);
 
            int nY1 = (int)(nY + (nHt - sz.Height));
            int nX1 = (int)(nX + (nWid / 2) - (sz.Width / 2));
            g.DrawString(str, m_font, (bMax) ? brTop : brFront, new Point(nX1, nY1));
 
            if (fMin != 0 || fMax != 0)
            {
                float fX = nX;
                float fWid = nWid ;
                nHt -= (int)sz.Height;                
 
                float fHt = nHt * fProb;
                float fHt1 = nHt - fHt;
                RectangleF rc1 = new RectangleF(fX, nY + fHt1, fWid, fHt);
                g.FillRectangle(brBack, rc1);
                g.DrawRectangle(penLine, rc1.X, rc1.Y, rc1.Width, rc1.Height);
            }
        }
 
        public void SaveWeights(string strFile)
        {
            if (File.Exists(strFile))
                File.Delete(strFile);
 
            using (StreamWriter sw = new StreamWriter(strFile))
            {
                save(sw, m_netOutput.layer_by_name("linear") as InnerProductLayer<T>);
                save(sw, m_netOutput.layer_by_name("noisy1") as InnerProductLayer<T>);
                save(sw, m_netOutput.layer_by_name("noisy2") as InnerProductLayer<T>);
                save(sw, m_netTarget.layer_by_name("linear") as InnerProductLayer<T>);
                save(sw, m_netTarget.layer_by_name("noisy1") as InnerProductLayer<T>);
                save(sw, m_netTarget.layer_by_name("noisy2") as InnerProductLayer<T>);
            }
        }
 
        private void save(StreamWriter sw, InnerProductLayer<T> layer)
        {
            for (int i = 0; i < layer.blobs.Count; i++)
            {
                float[] rgf = Utility.ConvertVecF<T>(layer.blobs[i].mutable_cpu_data);
                string strLine = "";
 
                for (int j = 0; j < rgf.Length; j++)
                {
                    strLine += rgf[j].ToString() + ",";
                }
 
                sw.WriteLine(strLine.TrimEnd(','));
            }
 
            for (int i = 0; i < layer.internal_blobs.Count; i++)
            {
                float[] rgf = Utility.ConvertVecF<T>(layer.internal_blobs[i].mutable_cpu_data);
                string strLine = "";
 
                for (int j = 0; j < rgf.Length; j++)
                {
                    strLine += rgf[j].ToString() + ",";
                }
 
                sw.WriteLine(strLine.TrimEnd(','));
            }
        }
 
        public void LoadWeights(string strFile)
        {
            if (!File.Exists(strFile))
                return;
 
            using (StreamReader sr = new StreamReader(strFile))
            {
                load(sr, m_netOutput.layer_by_name("linear") as InnerProductLayer<T>);
                load(sr, m_netOutput.layer_by_name("noisy1") as InnerProductLayer<T>);
                load(sr, m_netOutput.layer_by_name("noisy2") as InnerProductLayer<T>);
                load(sr, m_netTarget.layer_by_name("linear") as InnerProductLayer<T>);
                load(sr, m_netTarget.layer_by_name("noisy1") as InnerProductLayer<T>);
                load(sr, m_netTarget.layer_by_name("noisy2") as InnerProductLayer<T>);
            }
        }
 
        private void load(StreamReader sr, InnerProductLayer<T> layer)
        {
            for (int i = 0; i < layer.blobs.Count; i++)
            {
                List<float> rgf = new List<float>();
                string strLine = sr.ReadLine();
                string[] rgstr = strLine.Split(',');
 
                for (int j = 0; j < rgstr.Length; j++)
                {
                    rgf.Add(BaseParameter.ParseFloat(rgstr[j]));
                }
 
                layer.blobs[i].mutable_cpu_data = Utility.ConvertVec<T>(rgf.ToArray());
            }
 
            for (int i = 0; i < layer.internal_blobs.Count; i++)
            {
                List<float> rgf = new List<float>();
                string strLine = sr.ReadLine();
                string[] rgstr = strLine.Split(',');
 
                for (int j = 0; j < rgstr.Length; j++)
                {
                    rgf.Add(BaseParameter.ParseFloat(rgstr[j]));
                }
 
                layer.internal_blobs[i].mutable_cpu_data = Utility.ConvertVec<T>(rgf.ToArray());
            }
        }
    }
}