fix(torch): csvts forecast mode needs sequence of length backcast during predict

Author: Bycob

src/backends/torch/torchinputconns.cc

@@ -848,20 +848,21 @@ namespace dd
       {
         vecindex++;
         long int tstart = 0;
-        if (static_cast<long int>(seq.size())
-            < _backcast_timesteps + _forecast_timesteps)
+        long int timesteps = _train ? _backcast_timesteps + _forecast_timesteps
+                                    : _backcast_timesteps;
+        if (static_cast<long int>(seq.size()) < timesteps)
           {
             discard_warn(vecindex, seq.size(), test);
             continue;
           }
-        for (; tstart + _backcast_timesteps + _forecast_timesteps
-               < static_cast<long int>(seq.size());
+        for (; tstart + timesteps < static_cast<long int>(seq.size());
              tstart += _offset)
-          add_data_instance_forecast(tstart, vecindex, dataset, seq);
+          {
+            add_data_instance_forecast(tstart, vecindex, dataset, seq);
+          }
         if (tstart < static_cast<long int>(seq.size()) - 1)
-          add_data_instance_forecast(seq.size() - _backcast_timesteps
-                                         - _forecast_timesteps,
-                                     vecindex, dataset, seq);
+          add_data_instance_forecast(seq.size() - timesteps, vecindex, dataset,
+                                     seq);
       }
   }