Merge pull request #230 from pollytur/encoder3d_shifter_fix

Added shifter and updated 3d-encoder nonlinearity

neuralpredictors/layers/encoders/encoder3d.py

@@ -12,14 +12,30 @@ def __init__(self, core, readout, readout_nonlinearity, elu_xshift, elu_yshift):
         else:
             self.nonlinearity = core.nonlinearities[readout_nonlinearity]()
 
-    def forward(self, x, data_key=None):
+    def forward(self, x, data_key=None, pupil_center=None, trial_idx=None, shift=None, detach_core=False, **kwargs):
         out_core = self.core(x)
+        if detach_core:
+            out_core = out_core.detach()
+
+        if self.shifter:
+            if pupil_center is None:
+                raise ValueError("pupil_center is not given")
+            if shift is None:
+                time_points = x.shape[1]
+                pupil_center = pupil_center[:, :, -time_points:]
+                pupil_center = torch.transpose(pupil_center, 1, 2)
+                pupil_center = pupil_center.reshape(((-1,) + pupil_center.size()[2:]))
+                shift = self.shifter[data_key](pupil_center, trial_idx)
+
         out_core = torch.transpose(out_core, 1, 2)
         # the expected readout is 2d whereas the core can output 3d matrices
         # therefore, the first two dimensions (representing depth and batch size) are flattened and then passed
         # through the readout
         out_core = out_core.reshape(((-1,) + out_core.size()[2:]))
+        readout_out = self.readout(out_core, data_key=data_key, shift=shift, **kwargs)
 
-        readout_out = self.readout(out_core)
-        out = self.nonlinearity(readout_out)
+        if self.nonlinearity_type == "elu":
+            out = self.nonlinearity_fn(readout_out + self.offset) + 1
+        else:
+            out = self.nonlinearity_fn(readout_out)
         return out