gr-libs 0.1.6.post1__py3-none-any.whl → 0.1.8__py3-none-any.whl

gr_libs/ml/planner/mcts/utils/__init__.py

@@ -1,2 +1,2 @@
 from .node import Node
-from .tree import Tree
+from .tree import Tree

gr_libs/ml/planner/mcts/utils/node.py

@@ -1,8 +1,11 @@
 import random
 
+
 class Node:
 
-    def __init__(self, identifier, state, action, action_space, reward, terminal, pos, depth):
+    def __init__(
+        self, identifier, state, action, action_space, reward, terminal, pos, depth
+    ):
         self.identifier = identifier
         self.parent_identifier = None
         self.children_identifiers = []
@@ -21,13 +24,14 @@ class Node:
 
     def __str__(self):
         return "{}: (action={}, visits={}, reward={:d}, ratio={:0.4f})".format(
-                                                  self.state,
-                                                  self.action,
-                                                  self.num_visits,
-                                                  int(self.total_simulation_reward),
-                                                  self.performance)
+            self.state,
+            self.action,
+            self.num_visits,
+            int(self.total_simulation_reward),
+            self.performance,
+        )
 
     def untried_action(self):
         action = random.choice(self.untried_actions)
         self.untried_actions.remove(action)
-        return action
+        return action

gr_libs/ml/planner/mcts/utils/tree.py

@@ -1,22 +1,24 @@
 def vertical_lines(last_node_flags):
     vertical_lines = []
-    vertical_line = '\u2502'
+    vertical_line = "\u2502"
     for last_node_flag in last_node_flags[0:-1]:
         if last_node_flag == False:
-            vertical_lines.append(vertical_line + ' ' * 3)
+            vertical_lines.append(vertical_line + " " * 3)
         else:
             # space between vertical lines
-            vertical_lines.append(' ' * 4)
-    return ''.join(vertical_lines)
+            vertical_lines.append(" " * 4)
+    return "".join(vertical_lines)
+
 
 def horizontal_line(last_node_flags):
-    horizontal_line = '\u251c\u2500\u2500 '
-    horizontal_line_end = '\u2514\u2500\u2500 '
+    horizontal_line = "\u251c\u2500\u2500 "
+    horizontal_line_end = "\u2514\u2500\u2500 "
     if last_node_flags[-1]:
         return horizontal_line_end
     else:
         return horizontal_line
 
+
 class Tree:
 
     def __init__(self):
@@ -39,7 +41,9 @@ class Tree:
         if depth == 0:
             yield "", node
         else:
-            yield vertical_lines(last_node_flags) + horizontal_line(last_node_flags), node
+            yield vertical_lines(last_node_flags) + horizontal_line(
+                last_node_flags
+            ), node
 
         children = [self.nodes[identifier] for identifier in node.children_identifiers]
         last_index = len(children) - 1
@@ -60,7 +64,7 @@ class Tree:
             self.nodes[node.identifier].parent = None
         else:
             self.nodes[parent.identifier].children_identifiers.append(node.identifier)
-            self.nodes[node.identifier].parent_identifier=parent.identifier
+            self.nodes[node.identifier].parent_identifier = parent.identifier
 
     def update_id(self, old_id, new_id):
         assert new_id not in self.nodes.keys()
@@ -78,7 +82,7 @@ class Tree:
         # update the node's children (if there are any?...)
         for child_id in node.children_identifiers:
             self.nodes[child_id].parent_identifier = new_id
-        
+
         self.nodes.pop(old_id)
         self.nodes.update({node.identifier: node})
 
@@ -99,4 +103,4 @@ class Tree:
         lines = ""
         for edge, node in self.iter(identifier=None, depth=0, last_node_flags=[]):
             lines += "{}{}\n".format(edge, node)
-        print(lines)
+        print(lines)

gr_libs/ml/sequential/__init__.py

@@ -1 +1 @@
-from gr_libs.ml.sequential.lstm_model import LstmObservations
+from gr_libs.ml.sequential.lstm_model import LstmObservations

gr_libs/ml/sequential/lstm_model.py

@@ -10,183 +10,264 @@ from torch.nn.utils.rnn import pack_padded_sequence
 
 
 def accuracy_per_epoch(model, data_loader):
-	model.eval()
-	correct = total = 0.0
-	sum_loss = 0.0
-	with torch.no_grad():
-		for (first_traces, second_traces, is_same_goals, first_traces_lengths, second_traces_lengths) in data_loader:
-			y_pred = model.forward_tab(first_traces, second_traces, first_traces_lengths, second_traces_lengths)
-			loss = F.binary_cross_entropy(y_pred, is_same_goals)
-			sum_loss += loss.item()
-			y_pred = (y_pred >= 0.5)
-			correct += torch.sum(y_pred == is_same_goals)
-			total += len(is_same_goals)
-	return correct / total, sum_loss / 32
+    model.eval()
+    correct = total = 0.0
+    sum_loss = 0.0
+    with torch.no_grad():
+        for (
+            first_traces,
+            second_traces,
+            is_same_goals,
+            first_traces_lengths,
+            second_traces_lengths,
+        ) in data_loader:
+            y_pred = model.forward_tab(
+                first_traces, second_traces, first_traces_lengths, second_traces_lengths
+            )
+            loss = F.binary_cross_entropy(y_pred, is_same_goals)
+            sum_loss += loss.item()
+            y_pred = y_pred >= 0.5
+            correct += torch.sum(y_pred == is_same_goals)
+            total += len(is_same_goals)
+    return correct / total, sum_loss / 32
+
 
 def accuracy_per_epoch_cont(model, data_loader):
-	model.eval()
-	correct = total = 0.0
-	sum_loss = 0.0
-	with torch.no_grad():
-		for (first_traces_images, first_traces_texts, second_traces_images, second_traces_texts, is_same_goals, first_traces_lengths, second_traces_lengths) in data_loader:
-			y_pred = model.forward_cont(first_traces_images, first_traces_texts, second_traces_images, second_traces_texts, first_traces_lengths, second_traces_lengths)
-			loss = F.binary_cross_entropy(y_pred, is_same_goals)
-			sum_loss += loss.item()
-			y_pred = (y_pred >= 0.5)
-			correct += torch.sum(y_pred == is_same_goals)
-			total += len(is_same_goals)
-	return correct / total, sum_loss / 32
-
-# class CNNImageEmbeddor(nn.Module):
-# 	def __init__(self, obs_space, action_space, use_text=False):
-# 		super().__init__()
-# 		self.use_text = use_text
-# 		self.image_conv = nn.Sequential(
-# 			nn.Conv2d(3, 4, kernel_size=(3, 3), padding=1),  # Reduced filters, added padding
-#             nn.ReLU(),
-#             nn.MaxPool2d((2, 2)),
-#             nn.Conv2d(4, 4, (3, 3), padding=1),  # Reduced filters, added padding
-#             nn.ReLU(),
-#             nn.MaxPool2d((2, 2)),  # Added additional pooling to reduce size
-#             nn.Conv2d(4, 8, (3, 3), padding=1),  # Reduced filters, added padding
-#             nn.ReLU(),
-#             nn.BatchNorm2d(8)
-# 		)
-# 		n = obs_space["image"][0]
-# 		m = obs_space["image"][1]
-# 		self.image_embedding_size = ((n - 4) // 4 - 3) * ((m - 4) // 4 - 3) * 8
-# 		if self.use_text:
-# 			self.word_embedding_size = 32
-# 			self.word_embedding = nn.Embedding(obs_space["text"], self.word_embedding_size)
-# 			self.text_embedding_size = 128
-# 			self.text_rnn = nn.GRU(self.word_embedding_size, self.text_embedding_size, batch_first=True)
-
-	def forward(self, images, texts):
-	 	# images shape: batch_size X max_sequence_len X sample_size. same for text.
-		# need to reshape image to num_channels X height X width, like nn.Conv expects it to be.
-		x = images.transpose(2, 4).transpose(3, 4)
-		orig_shape = x.shape
-		# combine batch and sequence to 1 dimension so conv could handle it
-		x = x.view(orig_shape[0]*orig_shape[1], orig_shape[2], orig_shape[3], orig_shape[4]) # x shape: batch_size * max_sequence_len X sample_size
-		x = self.image_conv(x) # x shape:  batch_size * max_sequence_len X last_conv_size X 1 X 1
-		# reshape x back to divide batches from sequences
-		x = x.view(orig_shape[0], orig_shape[1], x.shape[1]) # x shape: batch_size X max_sequence_len X last_conv_size. last 2 dimensions (1,1) are collapsed to last conv.
-		embedding = x
-  
-		if self.use_text:
-			embed_text = self._get_embed_text(texts)
-			embedding = torch.cat((embedding, embed_text), dim=1)
-			
-		return embedding
-
-	def _get_embed_text(self, text):
-		_, hidden = self.text_rnn(self.word_embedding(text))
-		return hidden[-1]
+    model.eval()
+    correct = total = 0.0
+    sum_loss = 0.0
+    with torch.no_grad():
+        for (
+            first_traces_images,
+            first_traces_texts,
+            second_traces_images,
+            second_traces_texts,
+            is_same_goals,
+            first_traces_lengths,
+            second_traces_lengths,
+        ) in data_loader:
+            y_pred = model.forward_cont(
+                first_traces_images,
+                first_traces_texts,
+                second_traces_images,
+                second_traces_texts,
+                first_traces_lengths,
+                second_traces_lengths,
+            )
+            loss = F.binary_cross_entropy(y_pred, is_same_goals)
+            sum_loss += loss.item()
+            y_pred = y_pred >= 0.5
+            correct += torch.sum(y_pred == is_same_goals)
+            total += len(is_same_goals)
+    return correct / total, sum_loss / 32
+
+    # class CNNImageEmbeddor(nn.Module):
+    # 	def __init__(self, obs_space, action_space, use_text=False):
+    # 		super().__init__()
+    # 		self.use_text = use_text
+    # 		self.image_conv = nn.Sequential(
+    # 			nn.Conv2d(3, 4, kernel_size=(3, 3), padding=1),  # Reduced filters, added padding
+    #             nn.ReLU(),
+    #             nn.MaxPool2d((2, 2)),
+    #             nn.Conv2d(4, 4, (3, 3), padding=1),  # Reduced filters, added padding
+    #             nn.ReLU(),
+    #             nn.MaxPool2d((2, 2)),  # Added additional pooling to reduce size
+    #             nn.Conv2d(4, 8, (3, 3), padding=1),  # Reduced filters, added padding
+    #             nn.ReLU(),
+    #             nn.BatchNorm2d(8)
+    # 		)
+    # 		n = obs_space["image"][0]
+    # 		m = obs_space["image"][1]
+    # 		self.image_embedding_size = ((n - 4) // 4 - 3) * ((m - 4) // 4 - 3) * 8
+    # 		if self.use_text:
+    # 			self.word_embedding_size = 32
+    # 			self.word_embedding = nn.Embedding(obs_space["text"], self.word_embedding_size)
+    # 			self.text_embedding_size = 128
+    # 			self.text_rnn = nn.GRU(self.word_embedding_size, self.text_embedding_size, batch_first=True)
+
+    def forward(self, images, texts):
+        # images shape: batch_size X max_sequence_len X sample_size. same for text.
+        # need to reshape image to num_channels X height X width, like nn.Conv expects it to be.
+        x = images.transpose(2, 4).transpose(3, 4)
+        orig_shape = x.shape
+        # combine batch and sequence to 1 dimension so conv could handle it
+        x = x.view(
+            orig_shape[0] * orig_shape[1], orig_shape[2], orig_shape[3], orig_shape[4]
+        )  # x shape: batch_size * max_sequence_len X sample_size
+        x = self.image_conv(
+            x
+        )  # x shape:  batch_size * max_sequence_len X last_conv_size X 1 X 1
+        # reshape x back to divide batches from sequences
+        x = x.view(
+            orig_shape[0], orig_shape[1], x.shape[1]
+        )  # x shape: batch_size X max_sequence_len X last_conv_size. last 2 dimensions (1,1) are collapsed to last conv.
+        embedding = x
+
+        if self.use_text:
+            embed_text = self._get_embed_text(texts)
+            embedding = torch.cat((embedding, embed_text), dim=1)
+
+        return embedding
+
+    def _get_embed_text(self, text):
+        _, hidden = self.text_rnn(self.word_embedding(text))
+        return hidden[-1]
+
 
 class LstmObservations(nn.Module):
-	
-	def __init__(self, input_size, hidden_size): # TODO make sure the right cuda is used!
-		super(LstmObservations,self).__init__()
-		#self.embeddor = CNNImageEmbeddor(obs_space, action_space)
-		# check if the traces are a bunch of images
-		self.lstm = nn.LSTM(input_size=input_size, hidden_size=hidden_size, batch_first=True)
-		self.dropout = nn.Dropout(0.5)  # Added dropout layer
-		# Initialize weights
-		for name, param in self.lstm.named_parameters():
-			if 'weight' in name:
-				nn.init.xavier_uniform_(param)
-			elif 'bias' in name:
-				nn.init.zeros_(param)
-		
-
-	# tabular
-	def forward_tab(self, traces1, traces2, lengths1, lengths2):
-		out1, (ht1, ct1) = self.lstm(pack_padded_sequence(traces1, lengths1, batch_first=True, enforce_sorted=False), None) # traces1 & traces 2 shapes: batch_size X max sequence_length X embedding_size
-		out2, (ht2, ct2) = self.lstm(pack_padded_sequence(traces2, lengths2, batch_first=True, enforce_sorted=False), None)
-		# out1, _ = pad_packed_sequence(out1, batch_first=True, total_length=max(lengths1))
-		# out2, _ = pad_packed_sequence(out2, batch_first=True, total_length=max(lengths2))
-		manhattan_dis = torch.exp(-torch.sum(torch.abs(ht1[-1]-ht2[-1]),dim=1,keepdim=True))
-		return manhattan_dis.squeeze()
-
-	# continuous
-	# def forward_cont(self, traces1_images, traces1_texts, traces2_images, traces2_texts, lengths1, lengths2):
-	#  	# we also embed '0' images, but we take them out of the equation in the lstm (it knows to not treat them when batching)
-	# 	traces1 = self.embeddor(traces1_images, traces1_texts)
-	# 	traces2 = self.embeddor(traces2_images, traces2_texts) # traces1 & traces 2 shapes: batch_size X max_sequence_length X embedding_size
-	# 	out1, (ht1, ct1) = self.lstm(pack_padded_sequence(traces1, lengths1, batch_first=True, enforce_sorted=False), None)
-	# 	out2, (ht2, ct2) = self.lstm(pack_padded_sequence(traces2, lengths2, batch_first=True, enforce_sorted=False), None)
-	# 	manhattan_dis = torch.exp(-torch.sum(torch.abs(ht1[-1]-ht2[-1]),dim=1,keepdim=True))
-	# 	return manhattan_dis.squeeze()
-
-	def embed_sequence(self, trace):
-		trace = torch.stack([torch.tensor(observation, dtype=torch.float32) for observation in trace]).to(device)
-		out, (ht, ct) = self.lstm(trace, None)
-		return ht[-1]
-
-	# def embed_sequence_cont(self, sequence, preprocess_obss):
-	# 	sequence = [preprocess_obss([obs])[0] for ((obs, (_, _)), _) in sequence]
-	# 	trace_images = torch.tensor(np.expand_dims(torch.stack([step.image for step in sequence]), axis=0)).to(device)
-	# 	trace_texts = torch.tensor(np.expand_dims(torch.stack([step.text for step in sequence]), axis=0)).to(device)
-	# 	embedded_trace = self.embeddor(trace_images, trace_texts)
-	# 	out, (ht, ct) = self.lstm(embedded_trace)
-	# 	return ht[-1]
-
-def train_metric_model(model, train_loader, dev_loader, nepochs=5, patience = 2):
-	devAccuracy = []
-	best_dev_accuracy = 0.0
-	no_improvement_count = 0
-	optimizer = torch.optim.Adadelta(model.parameters(), weight_decay=0.1)
-	scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', patience=2, factor=0.5)
-	for epoch in range(nepochs):
-		sum_loss, denominator = 0.0, 0.0
-		model.train()
-		for (first_traces, second_traces, is_same_goals, first_traces_lengths, second_traces_lengths) in train_loader:
-			model.zero_grad()
-			y_pred = model.forward_tab(first_traces, second_traces, first_traces_lengths, second_traces_lengths)
-			if len(is_same_goals) == 1: is_same_goals = torch.squeeze(is_same_goals) # for the case of batches in size 1...
-			loss = F.binary_cross_entropy(y_pred, is_same_goals)
-			sum_loss += loss.item()
-			denominator += 1
-			loss.backward()
-			optimizer.step()
-   
-		dev_accuracy, dev_loss = accuracy_per_epoch(model, dev_loader)
-		devAccuracy.append(dev_accuracy)
-		if dev_accuracy > best_dev_accuracy:
-			best_dev_accuracy = dev_accuracy
-			no_improvement_count = 0
-		else:
-			no_improvement_count = 1
-
-		print("epoch - {}/{}...".format(epoch + 1, nepochs),
-				"train loss - {:.6f}...".format(sum_loss / denominator),
-				"dev loss - {:.6f}...".format(dev_loss),
-				"dev accuracy - {:.6f}".format(dev_accuracy))
-
-		if no_improvement_count >= patience:
-			print(f"Early stopping after {epoch + 1} epochs with no improvement.")
-			break
-  
+
+    def __init__(
+        self, input_size, hidden_size
+    ):  # TODO make sure the right cuda is used!
+        super(LstmObservations, self).__init__()
+        # self.embeddor = CNNImageEmbeddor(obs_space, action_space)
+        # check if the traces are a bunch of images
+        self.lstm = nn.LSTM(
+            input_size=input_size, hidden_size=hidden_size, batch_first=True
+        )
+        self.dropout = nn.Dropout(0.5)  # Added dropout layer
+        # Initialize weights
+        for name, param in self.lstm.named_parameters():
+            if "weight" in name:
+                nn.init.xavier_uniform_(param)
+            elif "bias" in name:
+                nn.init.zeros_(param)
+
+    # tabular
+    def forward_tab(self, traces1, traces2, lengths1, lengths2):
+        out1, (ht1, ct1) = self.lstm(
+            pack_padded_sequence(
+                traces1, lengths1, batch_first=True, enforce_sorted=False
+            ),
+            None,
+        )  # traces1 & traces 2 shapes: batch_size X max sequence_length X embedding_size
+        out2, (ht2, ct2) = self.lstm(
+            pack_padded_sequence(
+                traces2, lengths2, batch_first=True, enforce_sorted=False
+            ),
+            None,
+        )
+        # out1, _ = pad_packed_sequence(out1, batch_first=True, total_length=max(lengths1))
+        # out2, _ = pad_packed_sequence(out2, batch_first=True, total_length=max(lengths2))
+        manhattan_dis = torch.exp(
+            -torch.sum(torch.abs(ht1[-1] - ht2[-1]), dim=1, keepdim=True)
+        )
+        return manhattan_dis.squeeze()
+
+    # continuous
+    # def forward_cont(self, traces1_images, traces1_texts, traces2_images, traces2_texts, lengths1, lengths2):
+    #  	# we also embed '0' images, but we take them out of the equation in the lstm (it knows to not treat them when batching)
+    # 	traces1 = self.embeddor(traces1_images, traces1_texts)
+    # 	traces2 = self.embeddor(traces2_images, traces2_texts) # traces1 & traces 2 shapes: batch_size X max_sequence_length X embedding_size
+    # 	out1, (ht1, ct1) = self.lstm(pack_padded_sequence(traces1, lengths1, batch_first=True, enforce_sorted=False), None)
+    # 	out2, (ht2, ct2) = self.lstm(pack_padded_sequence(traces2, lengths2, batch_first=True, enforce_sorted=False), None)
+    # 	manhattan_dis = torch.exp(-torch.sum(torch.abs(ht1[-1]-ht2[-1]),dim=1,keepdim=True))
+    # 	return manhattan_dis.squeeze()
+
+    def embed_sequence(self, trace):
+        trace = torch.stack(
+            [torch.tensor(observation, dtype=torch.float32) for observation in trace]
+        ).to(device)
+        out, (ht, ct) = self.lstm(trace, None)
+        return ht[-1]
+
+    # def embed_sequence_cont(self, sequence, preprocess_obss):
+    # 	sequence = [preprocess_obss([obs])[0] for ((obs, (_, _)), _) in sequence]
+    # 	trace_images = torch.tensor(np.expand_dims(torch.stack([step.image for step in sequence]), axis=0)).to(device)
+    # 	trace_texts = torch.tensor(np.expand_dims(torch.stack([step.text for step in sequence]), axis=0)).to(device)
+    # 	embedded_trace = self.embeddor(trace_images, trace_texts)
+    # 	out, (ht, ct) = self.lstm(embedded_trace)
+    # 	return ht[-1]
+
+
+def train_metric_model(model, train_loader, dev_loader, nepochs=5, patience=2):
+    devAccuracy = []
+    best_dev_accuracy = 0.0
+    no_improvement_count = 0
+    optimizer = torch.optim.Adadelta(model.parameters(), weight_decay=0.1)
+    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
+        optimizer, mode="min", patience=2, factor=0.5
+    )
+    for epoch in range(nepochs):
+        sum_loss, denominator = 0.0, 0.0
+        model.train()
+        for (
+            first_traces,
+            second_traces,
+            is_same_goals,
+            first_traces_lengths,
+            second_traces_lengths,
+        ) in train_loader:
+            model.zero_grad()
+            y_pred = model.forward_tab(
+                first_traces, second_traces, first_traces_lengths, second_traces_lengths
+            )
+            if len(is_same_goals) == 1:
+                is_same_goals = torch.squeeze(
+                    is_same_goals
+                )  # for the case of batches in size 1...
+            loss = F.binary_cross_entropy(y_pred, is_same_goals)
+            sum_loss += loss.item()
+            denominator += 1
+            loss.backward()
+            optimizer.step()
+
+        dev_accuracy, dev_loss = accuracy_per_epoch(model, dev_loader)
+        devAccuracy.append(dev_accuracy)
+        if dev_accuracy > best_dev_accuracy:
+            best_dev_accuracy = dev_accuracy
+            no_improvement_count = 0
+        else:
+            no_improvement_count = 1
+
+        print(
+            "epoch - {}/{}...".format(epoch + 1, nepochs),
+            "train loss - {:.6f}...".format(sum_loss / denominator),
+            "dev loss - {:.6f}...".format(dev_loss),
+            "dev accuracy - {:.6f}".format(dev_accuracy),
+        )
+
+        if no_improvement_count >= patience:
+            print(f"Early stopping after {epoch + 1} epochs with no improvement.")
+            break
+
+
 def train_metric_model_cont(model, train_loader, dev_loader, nepochs=5):
-	devAccuracy = []
-	optimizer = torch.optim.Adadelta(model.parameters(),weight_decay=1.25)
-	for epoch in range(nepochs):
-		sum_loss, denominator = 0.0, 0.0
-		model.train()
-		for (first_traces_images, first_traces_texts, second_traces_images, second_traces_texts, is_same_goals, first_traces_lengths, second_traces_lengths) in train_loader:
-			model.zero_grad()
-			y_pred = model.forward_cont(first_traces_images, first_traces_texts, second_traces_images, second_traces_texts, first_traces_lengths, second_traces_lengths)
-			loss = F.binary_cross_entropy(y_pred, is_same_goals)
-			sum_loss += loss.item()
-			denominator += 1
-			loss.backward()
-			optimizer.step()
-   
-		dev_accuracy, dev_loss = accuracy_per_epoch_cont(model, dev_loader)
-		devAccuracy.append(dev_accuracy)
-
-		print("epoch - {}/{}...".format(epoch + 1, nepochs),
-				"train loss - {:.6f}...".format(sum_loss / denominator),
-				"dev loss - {:.6f}...".format(dev_loss),
-				"dev accuracy - {:.6f}".format(dev_accuracy))
-  
+    devAccuracy = []
+    optimizer = torch.optim.Adadelta(model.parameters(), weight_decay=1.25)
+    for epoch in range(nepochs):
+        sum_loss, denominator = 0.0, 0.0
+        model.train()
+        for (
+            first_traces_images,
+            first_traces_texts,
+            second_traces_images,
+            second_traces_texts,
+            is_same_goals,
+            first_traces_lengths,
+            second_traces_lengths,
+        ) in train_loader:
+            model.zero_grad()
+            y_pred = model.forward_cont(
+                first_traces_images,
+                first_traces_texts,
+                second_traces_images,
+                second_traces_texts,
+                first_traces_lengths,
+                second_traces_lengths,
+            )
+            loss = F.binary_cross_entropy(y_pred, is_same_goals)
+            sum_loss += loss.item()
+            denominator += 1
+            loss.backward()
+            optimizer.step()
+
+        dev_accuracy, dev_loss = accuracy_per_epoch_cont(model, dev_loader)
+        devAccuracy.append(dev_accuracy)
+
+        print(
+            "epoch - {}/{}...".format(epoch + 1, nepochs),
+            "train loss - {:.6f}...".format(sum_loss / denominator),
+            "dev loss - {:.6f}...".format(dev_loss),
+            "dev accuracy - {:.6f}".format(dev_accuracy),
+        )

gr_libs/ml/tabular/state.py

@@ -2,11 +2,9 @@ from abc import ABC
 
 
 class TabularState(ABC):
-    def __init__(self,
-                 agent_x_position: int,
-                 agent_y_position: int,
-                 agent_direction: int
-                 ):
+    def __init__(
+        self, agent_x_position: int, agent_y_position: int, agent_direction: int
+    ):
         self._agent_x_position = agent_x_position
         self._agent_y_position = agent_y_position
         self._agent_direction = agent_direction
@@ -14,8 +12,10 @@ class TabularState(ABC):
     @staticmethod
     def gen_tabular_state(environment, observation):
         x, y = environment.unwrapped.agent_pos
-        direction = observation['direction']
-        return TabularState(agent_x_position=x, agent_y_position=y, agent_direction=direction)
+        direction = observation["direction"]
+        return TabularState(
+            agent_x_position=x, agent_y_position=y, agent_direction=direction
+        )
 
     def __str__(self):
         return f"({self._agent_x_position},{self._agent_y_position}):{self._agent_direction}"