Working classification for grid cell memories

Author: C-Earl

scripts/Chris/DQN/Eval.ipynb

@@ -16,7 +16,7 @@ def __getitem__(self, idx):
     # Compress spike train into windows for dimension reduction
     return self.samples[idx].flatten(), self.labels[idx]
 
-def classify_recalls(out_dim, train_ratio, batch_size):
+def classify_recalls(out_dim, train_ratio, batch_size, epochs):
   print("Classifying recalled memories...")
 
   ## Load recalled memory samples ##
@@ -38,7 +38,7 @@ def classify_recalls(out_dim, train_ratio, batch_size):
   ## Training ##
   loss_log = []
   accuracy_log = []
-  for epoch in range(20):
+  for epoch in range(epochs):
     total_loss = 0
     correct = 0
     for memory_batch, positions in train_loader:
@@ -68,13 +68,29 @@ def classify_recalls(out_dim, train_ratio, batch_size):
   ## Testing ##
   total = 0
   correct = 0
+  confusion_matrix = torch.zeros(25, 25)
+  out_of_bounds = 0
   with torch.no_grad():
     for memories, labels in test_loader:
       outputs = model(memories)
       loss = criterion(outputs, labels)
       total += len(labels)
       correct += torch.all(outputs.round() == labels.round(),
                            dim=1).sum().item()  # Check if prediction for both x and y are correct
+      for t, p in zip(labels, outputs):
+        label_ind = int(t[0].round() * 5 + t[1].round())
+        pred_ind = int(p[0].round() * 5 + p[1].round())
+        if label_ind < 0 or label_ind >= 25 or pred_ind < 0 or pred_ind >= 25:
+          out_of_bounds += 1
+        else:
+          confusion_matrix[label_ind, pred_ind] += 1
+
+  plt.imshow(confusion_matrix)
+  plt.title('Confusion Matrix')
+  plt.xlabel('Predicted')
+  plt.ylabel('True Label')
+  plt.colorbar()
+  plt.show()
 
   print(f'Accuracy: {round(correct / total, 3)*100}%')
 

scripts/Chris/DQN/classify_recalls.py

@@ -11,7 +11,7 @@
   ## Constants ##
   WIDTH = 5
   HEIGHT = 5
-  SAMPLES_PER_POS = 10
+  SAMPLES_PER_POS = 1000
   NOISE = 0.1   # Noise in sampling
   WINDOW_FREQ = 10
   WINDOW_SIZE = 10
@@ -29,6 +29,7 @@
   OUT_DIM = 2
   TRAIN_RATIO = 0.8
   BATCH_SIZE = 10
+  TRAIN_EPOCHS = 15
   PLOT = True
   exc_hyper_params = {
     'thresh_exc': -55,
@@ -63,15 +64,15 @@
   #
   # # Spike Train Generation ##
   # spike_trains, labels, sorted_spike_trains = spike_train_generator(samples, labels, SIM_TIME, GC_MULTIPLES, MAX_SPIKE_FREQ)
-
-  # ## Association (Store) ##
-  store_reservoir(EXC_SIZE, INH_SIZE, STORE_SAMPLES, NUM_CELLS, GC_MULTIPLES, SIM_TIME, hyper_params, PLOT)
-
-  # ## Association (Recall) ##
-  recall_reservoir(EXC_SIZE, INH_SIZE, SIM_TIME, PLOT)
-
+  #
+  # # ## Association (Store) ##
+  # store_reservoir(EXC_SIZE, INH_SIZE, STORE_SAMPLES, NUM_CELLS, GC_MULTIPLES, SIM_TIME, hyper_params, PLOT)
+  #
+  # # ## Association (Recall) ##
+  # recall_reservoir(EXC_SIZE, INH_SIZE, SIM_TIME, PLOT)
+  #
   # # Preprocess Recalls ##
   # recalled_mem_preprocessing(WINDOW_FREQ, WINDOW_SIZE, PLOT)
 
-  ## Train ANN ##
-  # classify_recalls(OUT_DIM, TRAIN_RATIO, BATCH_SIZE)
+  # Train ANN ##
+  classify_recalls(OUT_DIM, TRAIN_RATIO, BATCH_SIZE, TRAIN_EPOCHS)

scripts/Chris/DQN/pipeline_executor.py

@@ -56,6 +56,7 @@ def recalled_mem_preprocessing(window_freq, window_size, plot):
     # plt.tight_layout()
     # plt.show()
 
+    positions = np.array([key for key in new_samples_sorted.keys()])
     fig = plt.figure(figsize=(10, 10))
     gs = fig.add_gridspec(nrows=5, ncols=5)
     for i, pos in enumerate(positions):

scripts/Chris/DQN/recalled_mem_preprocessing.py

@@ -16,12 +16,12 @@ def store_reservoir(exc_size, inh_size, num_samples, num_grid_cells, gc_multiple
   w_exc_exc = torch.rand(exc_size, exc_size)
   w_exc_inh = torch.rand(exc_size, inh_size)
   w_inh_exc = -torch.rand(inh_size, exc_size)
-  w_inh_inh = torch.rand(inh_size, inh_size)
+  w_inh_inh = -torch.rand(inh_size, inh_size)
   w_in_exc = sparsify(w_in_exc, 0.85)   # 0 x% of weights
   w_in_inh = sparsify(w_in_inh, 0.85)
-  w_exc_exc = sparsify(w_exc_exc, 0.85)
-  w_exc_inh = sparsify(w_exc_inh, 0.85)
-  w_inh_exc = sparsify(w_inh_exc, 0.85)
+  w_exc_exc = sparsify(w_exc_exc, 0.8)
+  w_exc_inh = sparsify(w_exc_inh, 0.5)
+  w_inh_exc = sparsify(w_inh_exc, 0.7)
   w_inh_inh = sparsify(w_inh_inh, 0.85)
   res = Reservoir(in_size, exc_size, inh_size, hyper_params,
                   w_in_exc, w_in_inh, w_exc_exc, w_exc_inh, w_inh_exc, w_inh_inh)