Neuron dropout #165

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open

leoscholl wants to merge 20 commits into master from leo-neuron-dropout

db/tracker/json_param.py

-Original file line number
+Diff line change
@@ Expand Up / @@ -83,6 +83,8 @@ def norm_trait(trait, value): @@
                 record = Model.objects.get(pk=value)
                 value = record.get()
             # Otherwise, let's hope it's already an instance
+        elif ttype == 'Array':
+            value = np.array(value)
         elif ttype == 'DataFile':
             # Similar to Instance traits, except we always know to use models.DataFile as the database table to look up the primary key
             from . import models
@@ Expand Down @@

features/__init__.py

-Original file line number
+Diff line change
@@ Expand Up / @@ -15,7 +15,7 @@ @@
     from .plexon_features import PlexonBMI, RelayPlexon, RelayPlexByte
     from .hdf_features import SaveHDF
     from .video_recording_features import SingleChannelVideo, E3Video
-    from .bmi_task_features import NormFiringRates
+    from .bmi_task_features import NormFiringRates, RandomUnitDropout
     from .arduino_features import PlexonSerialDIORowByte
     from .blackrock_features import BlackrockBMI
     from .blackrock_features import RelayBlackrockByte
@@ Expand Down Expand Up / @@ -79,6 +79,7 @@ @@
         eye_calibration=EyeCalibration,
         force_sensor=ForceControl,
         show_fixation_progress=Progressbar_fixation,
+        random_unit_dropout=RandomUnitDropout,
         clda_kfrml=CLDA_KFRML_IntendedVelocity
     )
@@ Expand Down @@

features/bmi_task_features.py

-Original file line number
+Diff line change
@@ -1,16 +1,110 @@
     '''
     BMI task features
     '''
+    import ast
     import time
     import numpy as np
     from riglib.experiment import traits, experiment
+    import copy
     ###### CONSTANTS
     sec_per_min = 60
     ########################################################################################################
     # Decoder/BMISystem add-ons
     ########################################################################################################
+    class RandomUnitDropout(traits.HasTraits):
+        '''
+        Randomly removes units from the decoder. Units are removed at the end of the delay period on each
+        trial and replaced when the trial ends (either in reward or penalty).The same units will be dropped
+        on repeated trials. The units to drop are specified in the `unit_drop_groups` attribute by a list of
+        lists of unit indices. The `unit_drop_targets` attribute specifies the target indices on which to
+        drop each group of units. Does not work with CLDA turned on.
+        '''
+        unit_drop_prob = traits.Float(0, desc="Probability of dropping a group of units from the decoder")
+        unit_drop_group1_channels = traits.List(value=[], desc="Channels to drop from the decoder one at a time")
+        unit_drop_group1_targets = traits.List(value=[], desc="Target indices on which to drop groups of units from the decoder")
+        unit_drop_group2_channels = traits.List(value=[], desc="Channels to drop from the decoder one at a time")
+        unit_drop_group2_targets = traits.List(value=[], desc="Target indices on which to drop groups of units from the decoder")
+        unit_drop_group3_channels = traits.List(value=[], desc="Channels to drop from the decoder one at a time")
+        unit_drop_group3_targets = traits.List(value=[], desc="Target indices on which to drop groups of units from the decoder")
+        unit_drop_group4_channels = traits.List(value=[], desc="Channels to drop from the decoder one at a time")
+        unit_drop_group4_targets = traits.List(value=[], desc="Target indices on which to drop groups of units from the decoder")
+        def init(self):
+            self.decoder_units_dropped = np.ones((len(self.decoder.units),), dtype='bool')
+            self.add_dtype('decoder_units_dropped', '?', self.decoder_units_dropped.shape)
+            self.unit_drop_groups = [self.unit_drop_group1_channels, self.unit_drop_group2_channels,
+                                     self.unit_drop_group3_channels, self.unit_drop_group4_channels]
+            self.unit_drop_groups = np.array([g for g in self.unit_drop_groups if not g == ['']])
+            self.unit_drop_targets = [self.unit_drop_group1_targets, self.unit_drop_group2_targets,
+                                     self.unit_drop_group3_targets, self.unit_drop_group4_targets]
+            self.unit_drop_targets = np.array([t for t in self.unit_drop_targets if not t == ['']])
+            self.unit_drop_group_idx = 0
+            super().init()
+            # Save a copy of the decoder
+            self.decoder_orig = copy.deepcopy(self.decoder)
+            self.reportstats['Units dropped'] = '' # Runtime stat displayed on the UI
+            print('--------------------\nUnit dropping settings:')
+            if len(self.unit_drop_groups) == len(self.unit_drop_targets):
+                for i in range(len(self.unit_drop_groups)):
+                    print(f'group {i} channels: {self.unit_drop_groups[i]}, targets: {self.unit_drop_targets[i]}')
+            else:
+                print('Unit dropping settings invalid! Please reformat into groups [[ch1, ch2], ...] and targets [[t1, t2, ...], ...]')
+                print('Current groups:', self.unit_drop_groups)
+                print('Current targets:', self.unit_drop_targets)
+            print('--------------------')
+        def _start_wait(self):
+            super()._start_wait()
+            # Decide which units to drop in this trial but don't actually drop them yet
+            if (self.gen_indices[self.target_index] in np.array(self.unit_drop_targets[self.unit_drop_group_idx]) and
+                np.random.rand() < self.unit_drop_prob):
+                self.decoder_units_dropped = np.isin(self.decoder.channels, self.unit_drop_groups[self.unit_drop_group_idx])
+                # Update the group for next trial
+                self.unit_drop_group_idx = (self.unit_drop_group_idx + 1) % len(self.unit_drop_groups)
+            else:
+                self.decoder_units_dropped = np.zeros((len(self.decoder.units),), dtype='bool')
+        def _start_targ_transition(self):
+            '''
+            Override the decoder to drop random units. Keep a record of what's going on in the `trial` data.
+            '''
+            super()._start_targ_transition()
+            if self.target_index == -1:
+                # Came from a penalty state
+                pass
+            elif self.target_index + 1 < self.chain_length and np.any(self.decoder_units_dropped):
+                if hasattr(self.decoder.filt, 'C'):
+                    self.decoder.filt.C[self.decoder_units_dropped, :] = 0
+                elif hasattr(self.decoder.filt, 'unit_to_state'):
+                    self.decoder.filt.unit_to_state[:, self.decoder_units_dropped] = 0
+                self.task_data['decoder_units_dropped'] = self.decoder_units_dropped
+                self.reportstats['Units dropped'] = str(self.decoder.channels[self.decoder_units_dropped])
+        def _reset_decoder_units(self):
+            if hasattr(self.decoder.filt, 'C'):
+                self.decoder.filt.C = self.decoder_orig.filt.C
+            elif hasattr(self.decoder.filt, 'unit_to_state'):
+                self.decoder.filt.unit_to_state = self.decoder_orig.filt.unit_to_state
+            self.task_data['decoder_units_dropped'] = np.zeros((len(self.decoder.units),), dtype='bool')
+            self.reportstats['Units dropped'] = '[]'
+        def _increment_tries(self):
+            super()._increment_tries()
+            self._reset_decoder_units()
+        def _start_reward(self):
+            super()._start_reward()
+            self._reset_decoder_units()
     class NormFiringRates(traits.HasTraits):
         ''' Docstring '''
@@ Expand Down @@

features/clda_features.py

-Original file line number
+Diff line change
@@ Expand Up / @@ -41,4 +41,4 @@ def create_updater(self): @@
             alter the decoder parameters to better match the intention estimates.
             '''
             self.updater = clda.KFRML(self.clda_batch_time, self.clda_update_half_life)
-            self.updater.init(self.decoder)
+            self.updater.init(self.decoder)

setup.py

-Original file line number
+Diff line change
@@ Expand Up / @@ -2,7 +2,7 @@ @@
     setuptools.setup(
         name="aolab-bmi3d",
-        version="1.0.5",
+        version="1.0.6",
         author="Lots of people",
         description="electrophysiology experimental rig library",
         packages=setuptools.find_packages(),
@@ Expand Down @@

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Neuron dropout #165

Uh oh!

Diff view

Diff view

There are no files selected for viewing

Neuron dropout #165

Are you sure you want to change the base?

Uh oh!

Neuron dropout #165

Uh oh!

Uh oh!

Diff view

Diff view

There are no files selected for viewing