dl-backtrace 0.0.16.dev4__py3-none-any.whl → 0.0.17__py3-none-any.whl

dl_backtrace/old_backtrace/__init__.py

@@ -0,0 +1 @@
+

dl_backtrace/old_backtrace/pytorch_backtrace/__init__.py

@@ -0,0 +1 @@
+from .backtrace import *

dl_backtrace/old_backtrace/pytorch_backtrace/backtrace/__init__.py

@@ -0,0 +1,4 @@
+from .backtrace import Backtrace
+
+from .utils import *
+from .config import *

dl_backtrace/old_backtrace/pytorch_backtrace/backtrace/backtrace.py

@@ -0,0 +1,639 @@
+import numpy as np
+import torch
+import torch.nn as nn
+from dl_backtrace.pytorch_backtrace.backtrace.utils import contrast as UC
+from dl_backtrace.pytorch_backtrace.backtrace.utils import prop as UP
+from dl_backtrace.pytorch_backtrace.backtrace.config import activation_master
+
+class Backtrace(object):
+    """
+    This is the constructor method for the Backtrace class. It initializes an instance of the class.
+    It takes two optional parameters: model (a neural network model) and activation_dict (a dictionary that maps layer names to activation functions).
+    """
+
+    def __init__(self, model=None, activation_dict={}):
+
+        # create a tree-like structure that represents the layers of the neural network model
+        self.create_tree(model)
+
+        # create a new model (an instance of tf.keras.Model) that produces the output of each layer in the neural network.
+        self.create_model_output(model)
+
+        # create a new model (an instance of tf.keras.Model) that produces the output of each layer in the neural network.
+        self.create_every_model_output(model)
+
+        # create a layer stack that defines the order in which layers should be processed during backpropagation.
+        self.create_layer_stack()
+
+        # checks if the model is sequential or not. If it's sequential, it adds the input layer to the layer stack.
+        # identity
+
+        inp_name = 'identity'
+        self.layer_stack.append(inp_name)
+        self.model_resource[1][inp_name] = {}
+        self.model_resource[1][inp_name]["name"] = inp_name
+        self.model_resource[1][inp_name]["type"] = "input"
+        self.model_resource[1][inp_name]["parent"] = []
+        self.model_resource[1][inp_name]["child"] = None
+        self.model_resource[3].append(inp_name)
+        self.sequential = True
+        try:
+
+            # calls the build_activation_dict method to build a dictionary that maps layer names to activation functions.
+            # If that fails, it creates a temporary dictionary with default activation functions.
+            if len(activation_dict) == 0:
+                self.build_activation_dict(model)
+            else:
+                self.activation_dict = activation_dict
+
+        except Exception as e:
+            print(e)
+            temp_dict = {}
+            for l in model.layers:
+                temp_dict[l.name] = activation_master["None"]
+            self.activation_dict = temp_dict
+
+    def build_activation_dict(self, model):
+        model_resource = self.model_resource
+        layer_list = list(model_resource[0].keys())
+        activation_dict = {}
+        activation_functions = ['relu', 'sigmoid', 'tanh', 'softmax']  # You can add more activation functions
+
+        for l in layer_list:
+            activation_found = False
+
+            try:  # could be activation for that layer
+                for activation in activation_functions:
+                    if activation in l.split('/')[1]:
+                        activation_dict[l.split('/')[0]] = activation
+                        activation_found = True
+            except:
+                activation_dict[l] = 'None'
+
+        # activation_master :
+        for key, value in activation_dict.items():
+            activation_dict[key] = activation_master.get(value)
+        self.activation_dict = activation_dict
+
+    def create_tree(self, model):
+        # create new layers same as tf version
+        layers = list(model.named_children())
+
+        activation_functions = ['relu', 'sigmoid', 'tanh', 'softmax']
+        layer_sequence = []
+
+        for i in range(len(layers) - 1):
+            current_layer, current_layer_obj = layers[i]
+            next_layer, next_layer_obj = layers[i + 1]
+            current_layer_name = current_layer
+            next_layer_name = next_layer
+
+            next_layer_type = next_layer_name.lower()
+            if any(af in next_layer_type for af in activation_functions):
+                layer_sequence.append((f"{current_layer_name}/{next_layer_name}", current_layer_obj))
+                i += 1
+            else:
+                if any(af in current_layer_name for af in activation_functions) is False:
+                    layer_sequence.append((current_layer_name, current_layer_obj))
+
+        # creating model_resource variable
+        layer_sequence
+        ltree = {}
+        layer_tree = {}
+        inputs = []
+        outputs = []
+        intermediates = []
+
+        prev_layer_id = None
+
+        num_layers = len(layer_sequence)
+
+        for i, (layer_name, layer) in enumerate(layer_sequence):
+            layer_id = layer_name
+            ltree[layer_id] = {}
+            layer_tree[layer_id] = layer
+
+            layer_type = layer.__class__.__name__
+            ltree[layer_id]["name"] = layer_id.split("/")[0]
+            ltree[layer_id]["class"] = layer_type
+
+            if i < num_layers - 1:
+                ltree[layer_id]["type"] = "intermediate"
+                intermediates.append(layer_id)
+            else:
+                ltree[layer_id]["type"] = "output"
+                outputs.append(layer_id)
+
+            if prev_layer_id is not None:
+                ltree[layer_id]["child"] = [prev_layer_id]
+                ltree[prev_layer_id]["parent"] = [layer_id]
+
+            prev_layer_id = layer_id
+
+        # Set child of the last layer as an empty list
+        if prev_layer_id is not None:
+            ltree[prev_layer_id]["parent"] = []
+
+        layer_tree.pop('identity')
+        ltree.pop('identity')
+        self.model_resource = (layer_tree, ltree, outputs, inputs)
+
+    def create_layer_stack(self):
+        model_resource = self.model_resource
+        start_layer = model_resource[2][0]
+        layer_stack = [start_layer]
+        temp_stack = [start_layer]
+        while len(layer_stack) < len(model_resource[0]):
+            start_layer = temp_stack.pop(0)
+            if model_resource[1][start_layer]["child"]:
+                child_nodes = model_resource[1][start_layer]["child"]
+                for ch in child_nodes:
+                    node_check = True
+                    for pa in model_resource[1][ch]["parent"]:
+                        if pa not in layer_stack:
+                            node_check = False
+                            break
+                    if node_check:
+                        if ch not in layer_stack:
+                            layer_stack.append(ch)
+                    temp_stack.append(ch)
+        self.layer_stack = layer_stack
+
+    def create_every_model_output(self, model):
+        class ModelWithEveryOutputs(nn.Module):
+            def __init__(self, base_model):
+                super(ModelWithEveryOutputs, self).__init__()
+                self.base_model = base_model
+
+            def forward(self, x):
+                outputs = []
+                for layer_name, layer in self.base_model._modules.items():
+                    if isinstance(x, tuple):
+                        if isinstance(layer, nn.LSTM):
+                            # Assuming you want to take the last LSTM output
+                            x, _ = layer(x[0])  # Pass the first element of the tuple (assumes one LSTM layer)
+                        else:
+                            x = layer(x[0])  # Pass the first element of the tuple
+                    else:
+                        x = layer(x)
+                    outputs.append((layer_name, x))
+                return outputs
+
+        self.every_out_model = ModelWithEveryOutputs(model)
+
+    def create_model_output(self, model):
+        class ModelWithOutputs(nn.Module):
+            def __init__(self, base_model):
+                super(ModelWithOutputs, self).__init__()
+                self.base_model = base_model
+
+            def forward(self, x):
+                outputs = []
+                for layer_name, layer in self.base_model._modules.items():
+                    if isinstance(layer, nn.LSTM):
+                        lstm_output, _ = layer(x)
+                        if lstm_output.dim() == 3:
+                            x = lstm_output[:, -1, :]  # Take the output of the last time step
+                        else:
+                            x = lstm_output
+                    else:
+                        x = layer(x)
+                    outputs.append((layer_name, x))
+                return outputs
+
+        # all_out_model = ModelWithOutputs(model)
+        self.all_out_model = ModelWithOutputs(model)
+        model.eval()
+        model_resource = self.model_resource
+        self.layers = [[], []]
+        for l in model_resource[0]:
+            self.layers[0].append(l)
+            self.layers[1].append(model_resource[0][l])
+
+    def predict_every(self, inputs):
+        every_out = self.every_out_model(inputs)
+        activation_functions = ['relu', 'sigmoid', 'tanh', 'softmax']
+        every_temp_out = {}
+
+        for i in range(len(every_out)):
+
+            current_layer, current_layer_obj = every_out[i]
+            try:
+                next_layer, next_layer_obj = every_out[i + 1]
+
+                current_layer_name = current_layer
+                next_layer_name = next_layer
+
+                next_layer_type = next_layer_name.lower()
+                if any(af in next_layer_type for af in activation_functions):
+                    if isinstance(next_layer_obj, tuple):
+                        # Assuming you want the first tensor from the tuple
+                        next_layer_tensor = next_layer_obj[0]
+                    else:
+                        next_layer_tensor = next_layer_obj
+
+                    every_temp_out[
+                        f"{current_layer_name}/{next_layer_name}"] = next_layer_tensor.detach().numpy().astype(
+                        np.float32)
+                    i += 1
+
+                else:
+                    if any(af in current_layer_name for af in activation_functions) is False:
+                        if isinstance(current_layer_obj, tuple):
+                            # Assuming you want the first tensor from the tuple
+                            current_layer_tensor = current_layer_obj[0]
+                        else:
+                            current_layer_tensor = current_layer_obj
+
+                        every_temp_out[current_layer_name] = current_layer_tensor.detach().numpy().astype(np.float32)
+            except:
+                if any(af in next_layer_type for af in activation_functions):
+                    pass
+
+                else:
+                    if any(af in current_layer for af in activation_functions) is False:
+                        if isinstance(current_layer_obj, tuple):
+                            # Assuming you want the first tensor from the tuple
+                            current_layer_tensor = current_layer_obj[0]
+                        else:
+                            current_layer_tensor = current_layer_obj
+
+                        every_temp_out[current_layer] = current_layer_tensor.detach().cpu().numpy().astype(np.float32)
+        return every_temp_out
+
+    def predict(self, inputs):
+        all_out = self.all_out_model(inputs)
+        activation_functions = ['relu', 'sigmoid', 'tanh', 'softmax']
+        temp_out = {}
+
+        for i in range(len(all_out)):
+
+            current_layer, current_layer_obj = all_out[i]
+            try:
+                next_layer, next_layer_obj = all_out[i + 1]
+
+                current_layer_name = current_layer
+                next_layer_name = next_layer
+
+                next_layer_type = next_layer_name.lower()
+                if any(af in next_layer_type for af in activation_functions):
+                    if isinstance(next_layer_obj, tuple):
+                        # Assuming you want the first tensor from the tuple
+                        next_layer_tensor = next_layer_obj[0]
+                    else:
+                        next_layer_tensor = next_layer_obj
+
+                    temp_out[
+                        f"{current_layer_name}/{next_layer_name}"] = next_layer_tensor.detach().cpu().numpy().astype(
+                        np.float32)
+                    i += 1
+
+                else:
+                    if any(af in current_layer_name for af in activation_functions) is False:
+                        if isinstance(current_layer_obj, tuple):
+                            # Assuming you want the first tensor from the tuple
+                            current_layer_tensor = current_layer_obj[0]
+                        else:
+                            current_layer_tensor = current_layer_obj
+
+                        temp_out[current_layer_name] = current_layer_tensor.detach().numpy().astype(np.float32)
+            except:
+                if any(af in next_layer_type for af in activation_functions):
+                    pass
+
+                else:
+                    if any(af in current_layer for af in activation_functions) is False:
+                        if isinstance(current_layer_obj, tuple):
+                            # Assuming you want the first tensor from the tuple
+                            current_layer_tensor = current_layer_obj[0]
+                        else:
+                            current_layer_tensor = current_layer_obj
+
+                        temp_out[current_layer] = current_layer_tensor.detach().cpu().numpy().astype(np.float32)
+
+        return temp_out
+
+    def eval(
+            self,
+            all_out,
+            mode,
+            start_wt=[],
+            multiplier=100.0,
+            scaler=0,
+            max_unit=0,
+    ):
+        # This method is used for evaluating layer-wise relevance based on different modes.
+        if mode == "default":
+            output = self.proportional_eval(
+                all_out=all_out,
+                start_wt=start_wt,
+                multiplier=multiplier,
+                scaler=0,
+                max_unit=0,
+            )
+            return output
+        elif mode == "contrast":
+            temp_output = self.contrast_eval(all_out=all_out, multiplier=multiplier)
+            output = {}
+            for k in temp_output[0].keys():
+                output[k] = {}
+                output[k]["Positive"] = temp_output[0][k]
+                output[k]["Negative"] = temp_output[1][k]
+            return output
+
+    def proportional_eval(
+            self, all_out, start_wt=[], multiplier=100.0, scaler=0, max_unit=0
+    ):
+        model_resource = self.model_resource
+        activation_dict = self.activation_dict
+        inputcheck = False
+        out_layer = model_resource[2][0]
+        all_wt = {}
+        if len(start_wt) == 0:
+            start_wt = UP.calculate_start_wt(all_out[out_layer])
+        all_wt[out_layer] = start_wt * multiplier
+        layer_stack = self.layer_stack
+
+        for start_layer in layer_stack:
+            if model_resource[1][start_layer]["child"]:
+                child_nodes = model_resource[1][start_layer]["child"]
+                for ch in child_nodes:
+                    if ch not in all_wt:
+                        if model_resource[1][start_layer]["class"] == 'LSTM':
+                            all_wt[ch] = np.zeros_like(every_temp_out[ch][0])
+                        else:
+                            all_wt[ch] = np.zeros_like(all_out[ch][0])
+
+                if model_resource[1][start_layer]["class"] == "Linear":
+                    l1 = model_resource[0][start_layer]
+                    w1 = l1.state_dict()['weight']
+                    b1 = l1.state_dict()['bias']
+                    temp_wt = UP.calculate_wt_fc(
+                        all_wt[start_layer],
+                        all_out[child_nodes[0]][0],
+                        w1,
+                        b1,
+                        activation_dict[model_resource[1][start_layer]["name"]],
+                    )
+                    all_wt[child_nodes[0]] += temp_wt
+                elif model_resource[1][start_layer]["class"] == "Conv2d":
+                    l1 = model_resource[0][start_layer]
+                    w1 = l1.state_dict()['weight']
+                    b1 = l1.state_dict()['bias']
+                    temp_wt = UP.calculate_wt_conv(
+                        all_wt[start_layer],
+                        all_out[child_nodes[0]][0],
+                        w1,
+                        b1,
+                        activation_dict[model_resource[1][start_layer]["name"]],
+                    )
+                    all_wt[child_nodes[0]] += temp_wt.T
+                elif model_resource[1][start_layer]["class"] == "Reshape":
+                    temp_wt = UP.calculate_wt_rshp(
+                        all_wt[start_layer], all_out[child_nodes[0]][0]
+                    )
+                    all_wt[child_nodes[0]] += temp_wt
+                elif model_resource[1][start_layer]["class"] == "Flatten":
+                    temp_wt = UP.calculate_wt_rshp(
+                        all_wt[start_layer], all_out[child_nodes[0]][0]
+                    )
+                    all_wt[child_nodes[0]] += temp_wt
+                elif (
+                        model_resource[1][start_layer]["class"] == "AdaptiveAvgPool2d"
+                ):
+                    temp_wt = UP.calculate_wt_gavgpool(
+                        all_wt[start_layer], all_out[child_nodes[0]][0]
+                    )
+                    all_wt[child_nodes[0]] += temp_wt.T
+                elif model_resource[1][start_layer]["class"] == "MaxPool2d":
+                    l1 = model_resource[0][start_layer]
+                    temp_wt = UP.calculate_wt_maxpool(
+                        all_wt[start_layer], all_out[child_nodes[0]][0], (l1.kernel_size, l1.kernel_size)
+                    )
+                    all_wt[child_nodes[0]] += temp_wt.T
+                elif model_resource[1][start_layer]["class"] == "AvgPool2d":
+                    l1 = model_resource[0][start_layer]
+                    temp_wt = UP.calculate_wt_avgpool(
+                        all_wt[start_layer], all_out[child_nodes[0]][0], (l1.kernel_size, l1.kernel_size)
+                    )
+                    all_wt[child_nodes[0]] += temp_wt.T
+                elif model_resource[1][start_layer]["class"] == "Concatenate":
+                    temp_wt = UP.calculate_wt_concat(
+                        all_wt[start_layer],
+                        [all_out[ch] for ch in child_nodes],
+                        model_resource[0][start_layer].axis,
+                    )
+                    for ind, ch in enumerate(child_nodes):
+                        all_wt[ch] += temp_wt[ind]
+                elif model_resource[1][start_layer]["class"] == "Add":
+                    temp_wt = UP.calculate_wt_add(
+                        all_wt[start_layer], [all_out[ch] for ch in child_nodes]
+                    )
+                    for ind, ch in enumerate(child_nodes):
+                        all_wt[ch] += temp_wt[ind]
+                elif model_resource[1][start_layer]["class"] == "LSTM":
+                    l1 = model_resource[0][start_layer]
+                    return_sequence = l1.return_sequences
+                    units = l1.units
+                    num_of_cells = l1.input_shape[1]
+                    lstm_obj_f = UP.LSTM_forward(
+                        num_of_cells, units, l1.weights, return_sequence, False
+                    )
+                    lstm_obj_b = UP.LSTM_backtrace(
+                        num_of_cells,
+                        units,
+                        [i.numpy() for i in l1.weights],
+                        return_sequence,
+                        False,
+                    )
+                    temp_out_f = lstm_obj_f.calculate_lstm_wt(
+                        every_temp_out[child_nodes[0]][0]
+                    )
+                    temp_wt = lstm_obj_b.calculate_lstm_wt(
+                        all_wt[start_layer], lstm_obj_f.compute_log
+                    )
+                    all_wt[child_nodes[0]] += temp_wt
+                else:
+                    temp_wt = all_wt[start_layer]
+                    all_wt[child_nodes[0]] += temp_wt
+        if max_unit > 0 and scaler == 0:
+            temp_dict = {}
+            for k in all_wt.keys():
+                temp_dict[k] = UC.weight_normalize(all_wt[k], max_val=max_unit)
+            all_wt = temp_dict
+        elif scaler > 0:
+            temp_dict = {}
+            for k in all_wt.keys():
+                temp_dict[k] = UC.weight_scaler(all_wt[k], scaler=scaler)
+            all_wt = temp_dict
+
+        return all_wt
+
+    def contrast_eval(self, all_out, multiplier=100.0):
+        model_resource = self.model_resource
+        activation_dict = self.activation_dict
+        inputcheck = False
+        out_layer = model_resource[2][0]
+        all_wt_pos = {}
+        all_wt_neg = {}
+        start_wt_pos, start_wt_neg = UC.calculate_start_wt(all_out[out_layer])
+        all_wt_pos[out_layer] = start_wt_pos * multiplier
+        all_wt_neg[out_layer] = start_wt_neg * multiplier
+        layer_stack = [out_layer]
+
+        while len(layer_stack) > 0:
+            start_layer = layer_stack.pop(0)
+            if model_resource[1][start_layer]["child"]:
+                child_nodes = model_resource[1][start_layer]["child"]
+                for ch in child_nodes:
+                    if ch not in all_wt_pos:
+                        all_wt_pos[ch] = np.zeros_like(all_out[ch][0])
+                        all_wt_neg[ch] = np.zeros_like(all_out[ch][0])
+                if model_resource[1][start_layer]["class"] == "Linear":
+                    l1 = model_resource[0][start_layer]
+                    w1 = l1.state_dict()['weight']
+                    b1 = l1.state_dict()['bias']
+                    temp_wt_pos, temp_wt_neg = UC.calculate_wt_fc(
+                        all_wt_pos[start_layer],
+                        all_wt_neg[start_layer],
+                        all_out[child_nodes[0]][0],
+                        w1,
+                        b1,
+                        activation_dict[model_resource[1][start_layer]["name"]],
+                    )
+                    all_wt_pos[child_nodes[0]] += temp_wt_pos
+                    all_wt_neg[child_nodes[0]] += temp_wt_neg
+                elif model_resource[1][start_layer]["class"] == "Conv2d":
+                    l1 = model_resource[0][start_layer]
+                    w1 = l1.state_dict()['weight']
+                    b1 = l1.state_dict()['bias']
+                    temp_wt_pos, temp_wt_neg = UC.calculate_wt_conv(
+                        all_wt_pos[start_layer],
+                        all_wt_neg[start_layer],
+                        all_out[child_nodes[0]][0],
+                        w1,
+                        b1,
+                        activation_dict[model_resource[1][start_layer]["name"]],
+                    )
+                    all_wt_pos[child_nodes[0]] += temp_wt_pos.T
+                    all_wt_neg[child_nodes[0]] += temp_wt_neg.T
+                elif model_resource[1][start_layer]["class"] == "Reshape":
+                    temp_wt_pos = UC.calculate_wt_rshp(
+                        all_wt_pos[start_layer], all_out[child_nodes[0]][0]
+                    )
+                    temp_wt_neg = UC.calculate_wt_rshp(
+                        all_wt_neg[start_layer], all_out[child_nodes[0]][0]
+                    )
+                    all_wt_pos[child_nodes[0]] += temp_wt_pos
+                    all_wt_neg[child_nodes[0]] += temp_wt_neg
+                elif (
+                        model_resource[1][start_layer]["class"] == "AdaptiveAvgPool2d"
+                ):
+                    temp_wt_pos, temp_wt_neg = UC.calculate_wt_gavgpool(
+                        all_wt_pos[start_layer],
+                        all_wt_neg[start_layer],
+                        all_out[child_nodes[0]][0],
+                    )
+                    all_wt_pos[child_nodes[0]] += temp_wt_pos.T
+                    all_wt_neg[child_nodes[0]] += temp_wt_neg.T
+                elif model_resource[1][start_layer]["class"] == "Flatten":
+                    temp_wt = UC.calculate_wt_rshp(
+                        all_wt_pos[start_layer], all_out[child_nodes[0]][0]
+                    )
+                    all_wt_pos[child_nodes[0]] += temp_wt
+                    temp_wt = UC.calculate_wt_rshp(
+                        all_wt_neg[start_layer], all_out[child_nodes[0]][0]
+                    )
+                    all_wt_neg[child_nodes[0]] += temp_wt
+                elif (
+                        model_resource[1][start_layer]["class"] == "AdaptiveAvgPool2d"
+                ):
+                    temp_wt_pos, temp_wt_neg = UC.calculate_wt_gavgpool(
+                        all_wt_pos[start_layer],
+                        all_wt_neg[start_layer],
+                        all_out[child_nodes[0]][0],
+                    )
+                    all_wt_pos[child_nodes[0]] += temp_wt_pos.T
+                    all_wt_neg[child_nodes[0]] += temp_wt_neg.T
+                elif model_resource[1][start_layer]["class"] == "MaxPool2d":
+                    l1 = model_resource[0][start_layer]
+                    temp_wt = UC.calculate_wt_maxpool(
+                        all_wt_pos[start_layer],
+                        all_out[child_nodes[0]][0],
+                        (l1.kernel_size, l1.kernel_size),
+                    )
+                    all_wt_pos[child_nodes[0]] += temp_wt.T
+                    temp_wt = UC.calculate_wt_maxpool(
+                        all_wt_neg[start_layer],
+                        all_out[child_nodes[0]][0],
+                        (l1.kernel_size, l1.kernel_size),
+                    )
+                    all_wt_neg[child_nodes[0]] += temp_wt.T
+                elif model_resource[1][start_layer]["class"] == "AvgPool2d":
+                    l1 = model_resource[0][start_layer]
+                    temp_wt_pos, temp_wt_neg = UC.calculate_wt_avgpool(
+                        all_wt_pos[start_layer],
+                        all_wt_neg[start_layer],
+                        all_out[child_nodes[0]][0],
+                        (l1.kernel_size, l1.kernel_size),
+                    )
+                    all_wt_pos[child_nodes[0]] += temp_wt_pos.T
+                    all_wt_neg[child_nodes[0]] += temp_wt_neg.T
+                elif model_resource[1][start_layer]["class"] == "Concatenate":
+                    temp_wt = UC.calculate_wt_concat(
+                        all_wt_pos[start_layer],
+                        [all_out[ch] for ch in child_nodes],
+                        model_resource[0][start_layer].axis,
+                    )
+                    for ind, ch in enumerate(child_nodes):
+                        all_wt_pos[ch] += temp_wt[ind]
+                    temp_wt = UC.calculate_wt_concat(
+                        all_wt_neg[start_layer],
+                        [all_out[ch] for ch in child_nodes],
+                        model_resource[0][start_layer].axis,
+                    )
+                    for ind, ch in enumerate(child_nodes):
+                        all_wt_neg[ch] += temp_wt[ind]
+                elif model_resource[1][start_layer]["class"] == "Add":
+                    temp_wt = UC.calculate_wt_add(
+                        all_wt_pos[start_layer],
+                        all_wt_neg[start_layer],
+                        [all_out[ch] for ch in child_nodes],
+                    )
+                    for ind, ch in enumerate(child_nodes):
+                        all_wt_pos[ch] += temp_wt[ind][0]
+                        all_wt_neg[ch] += temp_wt[ind][1]
+                elif model_resource[1][start_layer]["class"] == "LSTM":
+                    l1 = model_resource[0][start_layer]
+                    return_sequence = l1.return_sequences
+                    units = l1.units
+                    num_of_cells = l1.input_shape[1]
+                    lstm_obj_f = UC.LSTM_forward(
+                        num_of_cells, units, l1.weights, return_sequence, False
+                    )
+                    lstm_obj_b = UC.LSTM_backtrace(
+                        num_of_cells,
+                        units,
+                        [i.numpy() for i in l1.weights],
+                        return_sequence,
+                        False,
+                    )
+                    temp_out_f = lstm_obj_f.calculate_lstm_wt(
+                        all_out[child_nodes[0]][0]
+                    )
+                    temp_wt_pos, temp_wt_neg = lstm_obj_b.calculate_lstm_wt(
+                        all_wt_pos[start_layer],
+                        all_wt_neg[start_layer],
+                        lstm_obj_f.compute_log,
+                    )
+                    all_wt_pos[child_nodes[0]] = temp_wt_pos
+                    all_wt_neg[child_nodes[0]] = temp_wt_neg
+                else:
+                    temp_wt_pos = all_wt_pos[start_layer]
+                    temp_wt_neg = all_wt_neg[start_layer]
+                    all_wt_pos[child_nodes[0]] += temp_wt_pos
+                    all_wt_neg[child_nodes[0]] += temp_wt_neg
+                for ch in child_nodes:
+                    if not (ch in layer_stack):
+                        layer_stack.append(ch)
+        return all_wt_pos, all_wt_neg

dl_backtrace/old_backtrace/pytorch_backtrace/backtrace/config.py

@@ -0,0 +1,41 @@
+from dl_backtrace.pytorch_backtrace.backtrace.utils.prop import np_swish
+
+activation_master = {
+    "None": {
+        "name": None,
+        "range": {"l": None, "u": None},
+        "type": "null",
+        "func": None,
+    },
+    "linear": {
+        "name": None,
+        "range": {"l": None, "u": None},
+        "type": "mono",
+        "func": None,
+    },
+    "tanh": {"name": "tanh", "range": {"l": -2, "u": 2}, "type": "mono", "func": None},
+    "sigmoid": {
+        "name": "sigmoid",
+        "range": {"l": -4, "u": 4},
+        "type": "mono",
+        "func": None,
+    },
+    "relu": {
+        "name": "relu",
+        "range": {"l": 0, "u": None},
+        "type": "mono",
+        "func": None,
+    },
+    "swish": {
+        "name": "swish",
+        "range": {"l": -6, "u": None},
+        "type": "non_mono",
+        "func": np_swish,
+    },
+    "softmax": {
+        "name": "softmax",
+        "range": {"l": -1, "u": 2},
+        "type": "mono",
+        "func": None,
+    },
+}