LayerFactory.cs

using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Threading.Tasks;
using MyCaffe.basecode;
using MyCaffe.common;
using MyCaffe.db.image;
using MyCaffe.param;
 
namespace MyCaffe.layers.gpt
{
    public class LayerFactory : ILayerCreator
    {
        public Layer<double> CreateDouble(CudaDnn<double> cuda, Log log, LayerParameter p, CancelEvent evtCancel, IXDatabaseBase db)
        {
            switch (p.type)
            {
                case LayerParameter.LayerType.CAUSAL_SELF_ATTENTION:
                    return new CausalSelfAttentionLayer<double>(cuda, log, p);
 
                case LayerParameter.LayerType.MULTIHEAD_ATTENTION:
                    return new MultiheadAttentionLayer<double>(cuda, log, p);
 
                case LayerParameter.LayerType.POSITIONAL_ENCODER:
                    return new PositionalEncodingLayer<double>(cuda, log, p);
 
                case LayerParameter.LayerType.GELU:
                    return new GeluLayer<double>(cuda, log, p);
 
                case LayerParameter.LayerType.LAYERNORM:
                    return new LayerNormLayer<double>(cuda, log, p);
 
                case LayerParameter.LayerType.TRANSFORMER_BLOCK:
                    return new TransformerBlockLayer<double>(cuda, log, p);
                    
                case LayerParameter.LayerType.TOKENIZED_DATA:
                    return new TokenizedDataLayer<double>(cuda, log, p, db, evtCancel);
 
                case LayerParameter.LayerType.TOKENIZED_DATA_PAIRS:
                    return new TokenizedDataPairsLayer<double>(cuda, log, p, db, evtCancel);
                    
                case LayerParameter.LayerType.NLL_LOSS:
                    return new NLLLossLayer<double>(cuda, log, p);
 
                default:
                    return null;
            }
        }
 
        public Layer<float> CreateSingle(CudaDnn<float> cuda, Log log, LayerParameter p, CancelEvent evtCancel, IXDatabaseBase db)
        {
            switch (p.type)
            {
                case LayerParameter.LayerType.CAUSAL_SELF_ATTENTION:
                    return new CausalSelfAttentionLayer<float>(cuda, log, p);
 
                case LayerParameter.LayerType.MULTIHEAD_ATTENTION:
                    return new MultiheadAttentionLayer<float>(cuda, log, p);
 
                case LayerParameter.LayerType.POSITIONAL_ENCODER:
                    return new PositionalEncodingLayer<float>(cuda, log, p);
 
                case LayerParameter.LayerType.GELU:
                    return new GeluLayer<float>(cuda, log, p);
 
                case LayerParameter.LayerType.LAYERNORM:
                    return new LayerNormLayer<float>(cuda, log, p);
                    
                case LayerParameter.LayerType.TRANSFORMER_BLOCK:
                    return new TransformerBlockLayer<float>(cuda, log, p);
 
                case LayerParameter.LayerType.TOKENIZED_DATA:
                    return new TokenizedDataLayer<float>(cuda, log, p, db, evtCancel);
 
                case LayerParameter.LayerType.TOKENIZED_DATA_PAIRS:
                    return new TokenizedDataPairsLayer<float>(cuda, log, p, db, evtCancel);
 
                case LayerParameter.LayerType.NLL_LOSS:
                    return new NLLLossLayer<float>(cuda, log, p);
 
                default:
                    return null;
            }
        }
    }
}