Upload 5 files

inference.py

@@ -0,0 +1,171 @@
+# -*- coding: utf-8 -*-
+"""
+Created on Sun Sep 15 18:27:17 2024
+
+@author: salikha4
+"""
+
+import os
+import csv
+import json
+import shutil
+import random
+import argparse
+from datetime import datetime
+import pandas as pd
+import time
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.utils.data import Dataset, DataLoader, TensorDataset
+from torch.optim import Adam
+import numpy as np
+from lwm_model import LWM, load_model
+import warnings
+warnings.filterwarnings('ignore')
+from input_preprocess import *
+
+# Device configuration
+device_idx_ds = 3
+device = torch.device(f'cuda:{device_idx_ds}' if torch.cuda.is_available() else "cpu")
+if torch.cuda.is_available():
+    torch.cuda.empty_cache()
+
+# Folders
+# MODELS_FOLDER = 'models/'
+
+def dataset_gen(preprocessed_chs, input_type, scenario_idxs, lwm_model):
+    
+    if input_type in ['cls_emb', 'channel_emb']:
+        dataset = prepare_for_LWM(preprocessed_chs, device)
+    elif input_type == 'raw':
+        dataset = create_raw_dataset(preprocessed_chs, device)
+    
+    if input_type in ['cls_emb','channel_emb']:
+        # model = LWM().to(device)
+        # ckpt_name = 'model_weights.pth'
+        # ckpt_path = os.path.join(MODELS_FOLDER, ckpt_name)
+        # lwm_model = load_model(model, ckpt_path, device)
+        # print(f"Model loaded successfully on {device}")
+
+        # Process data through LWM
+        lwm_loss, embedding_data = evaluate(lwm_model, dataset)
+        
+        print(f'LWM loss: {lwm_loss:.4f}')
+        
+        if input_type == 'cls_emb':
+            embedding_data = embedding_data[:, 0]
+        elif input_type == 'channel_emb':  
+            embedding_data = embedding_data[:, 1:]
+        
+        dataset = embedding_data.float()
+        
+    return dataset
+
+
+def prepare_for_LWM(data, device, batch_size=64, shuffle=False):
+
+    input_ids, masked_tokens, masked_pos = zip(*data)
+    
+    input_ids_tensor = torch.tensor(input_ids, device=device).float()
+    masked_tokens_tensor = torch.tensor(masked_tokens, device=device).float()
+    masked_pos_tensor = torch.tensor(masked_pos, device=device).long()
+
+    dataset = TensorDataset(input_ids_tensor, masked_tokens_tensor, masked_pos_tensor)
+    
+    return DataLoader(dataset, batch_size=batch_size, shuffle=shuffle)
+
+
+def create_raw_dataset(data, device):
+    """Create a dataset for raw channel data."""
+    input_ids, _, _ = zip(*data)
+    input_data = torch.tensor(input_ids, device=device)[:, 1:]
+    return input_data.float()
+
+
+def label_gen(task, data, scenario, n_beams=64):
+    
+    idxs = np.where(data['user']['LoS'] != -1)[0]
+            
+    if task == 'LoS/NLoS Classification':
+        label = data['user']['LoS'][idxs]
+    elif task == 'Beam Prediction':
+        parameters, row_column_users, n_ant_bs, n_ant_ue, n_subcarriers = get_parameters(scenario)
+        n_users = len(data['user']['channel'])
+        n_subbands = 1
+        fov = 120
+
+        # Setup Beamformers
+        beam_angles = np.around(np.arange(-fov/2, fov/2+.1, fov/(n_beams-1)), 2)
+
+        F1 = np.array([steering_vec(parameters['bs_antenna']['shape'], 
+                                    phi=azi*np.pi/180, 
+                                    kd=2*np.pi*parameters['bs_antenna']['spacing']).squeeze()
+                       for azi in beam_angles])
+
+        full_dbm = np.zeros((n_beams, n_subbands, n_users), dtype=float)
+        for ue_idx in tqdm(range(n_users), desc='Computing the channel for each user'):
+            if data['user']['LoS'][ue_idx] == -1:
+                full_dbm[:,:,ue_idx] = np.nan
+            else:
+                chs = F1 @ data['user']['channel'][ue_idx]
+                full_linear = np.abs(np.mean(chs.squeeze().reshape((n_beams, n_subbands, -1)), axis=-1))
+                full_dbm[:,:,ue_idx] = np.around(20*np.log10(full_linear) + 30, 1)
+
+        best_beams = np.argmax(np.mean(full_dbm,axis=1), axis=0)
+        best_beams = best_beams.astype(float)
+        best_beams[np.isnan(full_dbm[0,0,:])] = np.nan
+        max_bf_pwr = np.max(np.mean(full_dbm,axis=1), axis=0) 
+    
+        label = best_beams[idxs]
+        
+    return label.astype(int)
+
+
+def steering_vec(array, phi=0, theta=0, kd=np.pi):
+    # phi = azimuth
+    # theta = elevation
+    idxs = DeepMIMOv3.ant_indices(array)
+    resp = DeepMIMOv3.array_response(idxs, phi, theta+np.pi/2, kd)
+    return resp / np.linalg.norm(resp)
+
+
+def evaluate(model, dataloader):
+
+    model.eval()
+    running_loss = 0.0
+    outputs = []
+    criterionMCM = nn.MSELoss()
+    
+    with torch.no_grad():
+        for batch in dataloader:
+            input_ids = batch[0]
+            masked_tokens = batch[1]
+            masked_pos = batch[2]
+            
+            logits_lm, output = model(input_ids, masked_pos)
+            
+            output_batch_preproc = output 
+            outputs.append(output_batch_preproc)
+
+            loss_lm = criterionMCM(logits_lm, masked_tokens)
+            loss = loss_lm/torch.var(masked_tokens)
+            running_loss += loss.item()
+            
+    average_loss = running_loss / len(dataloader)
+    output_total = torch.cat(outputs, dim=0)
+    
+    return average_loss, output_total
+
+
+def label_prepend(deepmimo_data, preprocessed_chs, task, scenario_idxs, n_beams=64):
+    labels = []
+    for scenario_idx in scenario_idxs:
+        scenario_name = scenarios_list()[scenario_idx]
+        # data = DeepMIMO_data_gen(scenario_name)
+        data = deepmimo_data[scenario_idx]
+        labels.extend(label_gen(task, data, scenario_name, n_beams=n_beams))
+    
+    preprocessed_chs = [preprocessed_chs[i] + [labels[i]] for i in range(len(preprocessed_chs))]
+    
+    return preprocessed_chs

input_preprocess.py

@@ -0,0 +1,296 @@
+# -*- coding: utf-8 -*-
+"""
+Created on Fri Sep 13 16:13:29 2024
+
+This script generates preprocessed data from wireless communication scenarios, 
+including token generation, patch creation, and data sampling for machine learning models.
+
+@author: salikha4
+"""
+
+import numpy as np
+import os
+from tqdm import tqdm
+import time
+import pickle
+import DeepMIMOv3
+
+vars_folder = 'variables/'
+os.makedirs(vars_folder, exist_ok=True)
+
+#%% Scenarios List
+def scenarios_list():
+    """Returns an array of available scenarios."""
+    return np.array([
+        'city_18_denver', 'city_15_indianapolis', 'city_19_oklahoma', 
+        'city_12_fortworth', 'city_11_santaclara', 'city_7_sandiego'
+    ])
+
+#%% Token Generation
+def tokenizer(deepmimo_data, gen_raw=True):
+    """
+    Generates tokens by preparing and preprocessing the dataset.
+
+    Args:
+        scenario_idxs (list): Indices of the scenarios.
+        patch_gen (bool): Whether to generate patches. Defaults to True.
+        patch_size (int): Size of each patch. Defaults to 16.
+        gen_deepMIMO_data (bool): Whether to generate DeepMIMO data. Defaults to False.
+        gen_raw (bool): Whether to generate raw data. Defaults to False.
+        save_data (bool): Whether to save the preprocessed data. Defaults to False.
+    
+    Returns:
+        preprocessed_data, sequence_length, element_length: Preprocessed data and related dimensions.
+    """
+    
+    # Patch generation or loading
+    n_scenarios = len(deepmimo_data)
+    patches = [patch_maker(deepmimo_data[scenario_idx]) for scenario_idx in range(n_scenarios)]
+    patches = np.vstack(patches)
+    
+    # Define dimensions
+    patch_size = patches.shape[2]
+    n_patches = patches.shape[1]
+    n_masks_half = int(0.15 * n_patches / 2)
+    sequence_length = n_patches + 1
+    element_length = patch_size
+    
+    word2id = {'[CLS]': 0.2 * np.ones((patch_size)), '[MASK]': 0.1 * np.ones((patch_size))}
+
+    # Generate preprocessed channels
+    preprocessed_data = []
+    for user_idx in tqdm(range(len(patches)), desc="Processing items"):
+        sample = make_sample(user_idx, patches, word2id, n_patches, n_masks_half, patch_size, gen_raw=gen_raw)
+        preprocessed_data.append(sample)
+            
+    return preprocessed_data
+
+#%% Patch Creation
+def patch_maker(data, patch_size=16, norm_factor=1e6):
+    """
+    Creates patches from the dataset based on the scenario.
+
+    Args:-
+        patch_size (int): Size of each patch.
+        scenario (str): Selected scenario for data generation.
+        gen_deepMIMO_data (bool): Whether to generate DeepMIMO data.
+        norm_factor (int): Normalization factor for channels.
+
+    Returns:
+        patch (numpy array): Generated patches.
+    """
+    idxs = np.where(data['user']['LoS'] != -1)[0]
+    
+    # Reshaping and normalizing channels
+    original_ch = data['user']['channel'][idxs]
+    flat_channels = original_ch.reshape((original_ch.shape[0], -1)).astype(np.csingle)
+    flat_channels_complex = np.hstack((flat_channels.real, flat_channels.imag)) * norm_factor
+    
+    # Create patches
+    n_patches = flat_channels_complex.shape[1] // patch_size
+    patch = np.zeros((len(idxs), n_patches, patch_size))
+    for idx in range(n_patches):
+        patch[:, idx, :] = flat_channels_complex[:, idx * patch_size:(idx + 1) * patch_size]
+    
+    return patch
+
+
+#%% Data Generation for Scenario Areas
+def DeepMIMO_data_gen(scenario):
+    """
+    Generates or loads data for a given scenario.
+
+    Args:
+        scenario (str): Scenario name.
+        gen_deepMIMO_data (bool): Whether to generate DeepMIMO data.
+        save_data (bool): Whether to save generated data.
+    
+    Returns:
+        data (dict): Loaded or generated data.
+    """
+    
+    parameters, row_column_users, n_ant_bs, n_ant_ue, n_subcarriers = get_parameters(scenario)
+    
+    deepMIMO_dataset = DeepMIMOv3.generate_data(parameters)
+    uniform_idxs = uniform_sampling(deepMIMO_dataset, [1, 1], len(parameters['user_rows']), 
+                                    users_per_row=row_column_users[scenario]['n_per_row'])
+    data = select_by_idx(deepMIMO_dataset, uniform_idxs)[0]
+        
+    return data
+
+#%%%
+def get_parameters(scenario):
+    
+    n_ant_bs = 32 #32
+    n_ant_ue = 1
+    n_subcarriers = 32 #32
+    scs = 30e3
+        
+    row_column_users = {
+    'city_18_denver': {
+        'n_rows': 85,
+        'n_per_row': 82
+    },
+    'city_15_indianapolis': {
+        'n_rows': 80,
+        'n_per_row': 79
+    },
+    'city_19_oklahoma': {
+        'n_rows': 82,
+        'n_per_row': 75
+    },
+    'city_12_fortworth': {
+        'n_rows': 86,
+        'n_per_row': 72
+    },
+    'city_11_santaclara': {
+        'n_rows': 47,
+        'n_per_row': 114
+    },
+    'city_7_sandiego': {
+        'n_rows': 71,
+        'n_per_row': 83
+    }}
+    
+    parameters = DeepMIMOv3.default_params()
+    parameters['dataset_folder'] = './scenarios'
+    parameters['scenario'] = scenario
+    
+    if scenario == 'O1_3p5':
+        parameters['active_BS'] = np.array([4])
+    elif scenario in ['city_18_denver', 'city_15_indianapolis']:
+        parameters['active_BS'] = np.array([3])
+    else:
+        parameters['active_BS'] = np.array([1])
+        
+    if scenario == 'Boston5G_3p5':
+        parameters['user_rows'] = np.arange(row_column_users[scenario]['n_rows'][0],
+                                            row_column_users[scenario]['n_rows'][1])
+    else:
+        parameters['user_rows'] = np.arange(row_column_users[scenario]['n_rows'])
+    parameters['bs_antenna']['shape'] = np.array([n_ant_bs, 1]) # Horizontal, Vertical 
+    parameters['bs_antenna']['rotation'] = np.array([0,0,-135]) # (x,y,z)
+    parameters['ue_antenna']['shape'] = np.array([n_ant_ue, 1])
+    parameters['enable_BS2BS'] = False
+    parameters['OFDM']['subcarriers'] = n_subcarriers
+    parameters['OFDM']['selected_subcarriers'] = np.arange(n_subcarriers)
+    
+    parameters['OFDM']['bandwidth'] = scs * n_subcarriers / 1e9
+    parameters['num_paths'] = 20
+    
+    return parameters, row_column_users, n_ant_bs, n_ant_ue, n_subcarriers
+
+
+#%% Sample Generation
+def make_sample(user_idx, patch, word2id, n_patches, n_masks, patch_size, gen_raw=False):
+    """
+    Generates a sample for each user, including masking and tokenizing.
+
+    Args:
+        user_idx (int): Index of the user.
+        patch (numpy array): Patches data.
+        word2id (dict): Dictionary for special tokens.
+        n_patches (int): Number of patches.
+        n_masks (int): Number of masks.
+        patch_size (int): Size of each patch.
+        gen_raw (bool): Whether to generate raw tokens.
+
+    Returns:
+        sample (list): Generated sample for the user.
+    """
+    
+    tokens = patch[user_idx]
+    input_ids = np.vstack((word2id['[CLS]'], tokens))
+    
+    real_tokens_size = int(n_patches / 2)
+    masks_pos_real = np.random.choice(range(0, real_tokens_size), size=n_masks, replace=False)
+    masks_pos_imag = masks_pos_real + real_tokens_size
+    masked_pos = np.hstack((masks_pos_real, masks_pos_imag)) + 1
+    
+    masked_tokens = []
+    for pos in masked_pos:
+        original_masked_tokens = input_ids[pos].copy()
+        masked_tokens.append(original_masked_tokens)
+        if not gen_raw:
+            rnd_num = np.random.rand()
+            if rnd_num < 0.1:
+                input_ids[pos] = np.random.rand(patch_size)
+            elif rnd_num < 0.9:
+                input_ids[pos] = word2id['[MASK]']
+
+    return [input_ids, masked_tokens, masked_pos]
+
+
+#%% Sampling and Data Selection
+def uniform_sampling(dataset, sampling_div, n_rows, users_per_row):
+    """
+    Performs uniform sampling on the dataset.
+
+    Args:
+        dataset (dict): DeepMIMO dataset.
+        sampling_div (list): Step sizes along [x, y] dimensions.
+        n_rows (int): Number of rows for user selection.
+        users_per_row (int): Number of users per row.
+
+    Returns:
+        uniform_idxs (numpy array): Indices of the selected samples.
+    """
+    cols = np.arange(users_per_row, step=sampling_div[0])
+    rows = np.arange(n_rows, step=sampling_div[1])
+    uniform_idxs = np.array([j + i * users_per_row for i in rows for j in cols])
+    
+    return uniform_idxs
+
+def select_by_idx(dataset, idxs):
+    """
+    Selects a subset of the dataset based on the provided indices.
+
+    Args:
+        dataset (dict): Dataset to trim.
+        idxs (numpy array): Indices of users to select.
+
+    Returns:
+        dataset_t (list): Trimmed dataset based on selected indices.
+    """
+    dataset_t = []  # Trimmed dataset
+    for bs_idx in range(len(dataset)):
+        dataset_t.append({})
+        for key in dataset[bs_idx].keys():
+            dataset_t[bs_idx]['location'] = dataset[bs_idx]['location']
+            dataset_t[bs_idx]['user'] = {k: dataset[bs_idx]['user'][k][idxs] for k in dataset[bs_idx]['user']}
+    
+    return dataset_t
+
+#%% Save and Load Utilities
+def save_var(var, path):
+    """
+    Saves a variable to a pickle file.
+
+    Args:
+        var (object): Variable to be saved.
+        path (str): Path to save the file.
+
+    Returns:
+        None
+    """
+    path_full = path if path.endswith('.p') else (path + '.pickle')    
+    with open(path_full, 'wb') as handle:
+        pickle.dump(var, handle)
+
+def load_var(path):
+    """
+    Loads a variable from a pickle file.
+
+    Args:
+        path (str): Path of the file to load.
+
+    Returns:
+        var (object): Loaded variable.
+    """
+    path_full = path if path.endswith('.p') else (path + '.pickle')
+    with open(path_full, 'rb') as handle:
+        var = pickle.load(handle)
+    
+    return var
+
+#%%

load_data.py

@@ -0,0 +1,39 @@
+# -*- coding: utf-8 -*-
+"""
+Created on Sun Sep 15 17:19:21 2024
+
+@author: salikha4
+"""
+
+import requests
+import zipfile
+import os
+import pickle
+
+def load_DeepMIMO_data(zip_file_name='dataset.zip', p_file_name='deepmimo_data.p', extract_path='./data'):
+    url = "https://huggingface.co/datasets/sadjadalikhani/lwm/resolve/main/dataset.zip"
+    # Step 1: Download the ZIP file from Hugging Face
+    print(f"Downloading ZIP file from {url}...")
+    response = requests.get(url)
+    with open(zip_file_name, 'wb') as f:
+        f.write(response.content)
+    print(f"Downloaded ZIP file to {zip_file_name}.")
+
+    # Step 2: Unzip the file
+    print(f"Extracting ZIP file to {extract_path}...")
+    with zipfile.ZipFile(zip_file_name, 'r') as zip_ref:
+        zip_ref.extractall(extract_path)
+    print(f"Extracted ZIP file contents to {extract_path}.")
+
+    # Step 3: Load the .p file
+    p_file_path = os.path.join(extract_path, p_file_name)
+    print(f"Loading .p file from {p_file_path}...")
+    with open(p_file_path, 'rb') as f:
+        data = pickle.load(f)
+    
+    print("Data successfully loaded from the .p file.")
+    return data
+
+
+
+

lwm_model.py

@@ -0,0 +1,173 @@
+"""
+LWM (Large Wireless Model) Implementation and Loading
+
+@author: salikha4
+
+This module defines a Large Wireless Model (LWM) using PyTorch, including custom layers
+for embedding, self-attention, and feed-forward networks. It also provides functionality
+to load a pre-trained model from a checkpoint.
+
+Dependencies:
+    - torch
+    - numpy
+"""
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+
+ELEMENT_LENGTH = 16
+D_MODEL = 64
+MAX_LEN = 129
+N_LAYERS = 12
+N_HEADS = 12
+D_FF = D_MODEL * 4
+D_K = D_MODEL // N_HEADS
+D_V = D_MODEL // N_HEADS
+DROPOUT = 0.1
+
+class LayerNormalization(nn.Module):
+    def __init__(self, d_model: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.eps = eps
+        self.alpha = nn.Parameter(torch.ones(d_model))
+        self.bias = nn.Parameter(torch.zeros(d_model))
+
+    def forward(self, x):
+        mean = x.mean(dim=-1, keepdim=True)
+        std = x.std(dim=-1, keepdim=True)
+        return self.alpha * (x - mean) / (std + self.eps) + self.bias
+
+class Embedding(nn.Module):
+    def __init__(self, element_length, d_model, max_len):
+        super().__init__()
+        self.element_length = element_length
+        self.d_model = d_model
+        self.proj = nn.Linear(element_length, d_model)
+        self.pos_embed = nn.Embedding(max_len, d_model)
+        self.norm = LayerNormalization(d_model)
+
+    def forward(self, x):
+        seq_len = x.size(1)
+        pos = torch.arange(seq_len, dtype=torch.long, device=x.device)
+        pos = pos.unsqueeze(0).expand_as(x[:, :, 0])
+        tok_emb = self.proj(x.float())
+        embedding = tok_emb + self.pos_embed(pos)
+        return self.norm(embedding)
+
+class ScaledDotProductAttention(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, Q, K, V):
+        scores = torch.matmul(Q, K.transpose(-1, -2)) / np.sqrt(D_K)
+        attn = F.softmax(scores, dim=-1)
+        context = torch.matmul(attn, V)
+        return context, attn
+
+class MultiHeadAttention(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.W_Q = nn.Linear(D_MODEL, D_K * N_HEADS)
+        self.W_K = nn.Linear(D_MODEL, D_K * N_HEADS)
+        self.W_V = nn.Linear(D_MODEL, D_V * N_HEADS)
+        self.linear = nn.Linear(N_HEADS * D_V, D_MODEL)
+        self.norm = LayerNormalization(D_MODEL)
+        self.dropout = nn.Dropout(DROPOUT)
+        
+    def forward(self, Q, K, V):
+        residual, batch_size = Q, Q.size(0)
+        q_s = self.W_Q(Q).view(batch_size, -1, N_HEADS, D_K).transpose(1, 2)
+        k_s = self.W_K(K).view(batch_size, -1, N_HEADS, D_K).transpose(1, 2)
+        v_s = self.W_V(V).view(batch_size, -1, N_HEADS, D_V).transpose(1, 2)
+
+        context, attn = ScaledDotProductAttention()(q_s, k_s, v_s)
+        output = context.transpose(1, 2).contiguous().view(batch_size, -1, N_HEADS * D_V)
+        output = self.linear(output)
+        return residual + self.dropout(output), attn #residual + self.dropout(output), attn
+
+class PoswiseFeedForwardNet(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.fc1 = nn.Linear(D_MODEL, D_FF)
+        self.fc2 = nn.Linear(D_FF, D_MODEL)
+        self.dropout = nn.Dropout(DROPOUT)
+        self.norm = LayerNormalization(D_MODEL)
+
+    def forward(self, x):
+        output = self.fc2(self.dropout(F.relu(self.fc1(x))))
+        return x + self.dropout(output) #x + self.dropout(output)
+
+class EncoderLayer(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.enc_self_attn = MultiHeadAttention()
+        self.pos_ffn = PoswiseFeedForwardNet()
+        self.norm = LayerNormalization(D_MODEL)
+
+    def forward(self, enc_inputs):
+        attn_outputs, attn = self.enc_self_attn(enc_inputs, enc_inputs, enc_inputs)
+        attn_outputs = self.norm(attn_outputs)
+        enc_outputs = self.pos_ffn(attn_outputs)
+        return enc_outputs, attn
+
+class LWM(nn.Module):
+    def __init__(self, element_length=16, d_model=64, max_len=129, n_layers=12):
+        super().__init__()
+
+        self.embedding = Embedding(element_length, d_model, max_len)
+        self.layers = nn.ModuleList([EncoderLayer() for _ in range(n_layers)])
+        self.linear = nn.Linear(d_model, d_model)
+        self.norm = LayerNormalization(d_model)
+
+        embed_weight = self.embedding.proj.weight
+        d_model, n_dim = embed_weight.size()
+        self.decoder = nn.Linear(d_model, n_dim, bias=False)
+        self.decoder.weight = nn.Parameter(embed_weight.transpose(0, 1))
+        self.decoder_bias = nn.Parameter(torch.zeros(n_dim))
+
+    def forward(self, input_ids, masked_pos):
+        output = self.embedding(input_ids)
+        
+        for layer in self.layers:
+            output, _ = layer(output)
+
+        masked_pos = masked_pos.long()[:, :, None].expand(-1, -1, output.size(-1))
+        h_masked = torch.gather(output, 1, masked_pos)
+        h_masked = self.norm(F.relu(self.linear(h_masked)))
+        logits_lm = self.decoder(h_masked) + self.decoder_bias
+
+        return logits_lm, output
+
+def load_model(model, model_path, device=None):
+    """
+    Load a pre-trained LWM model from a checkpoint.
+
+    Args:
+        model_path (str): Path to the checkpoint file.
+        device (torch.device, optional): Device to load the model onto.
+
+    Returns:
+        LWM: Loaded model instance.
+    """
+    if device is None:
+        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    
+    #model = LWM(ELEMENT_LENGTH, D_MODEL, MAX_LEN, N_LAYERS)
+    state_dict = torch.load(model_path, map_location=device)
+    model.load_state_dict(state_dict)
+    model.to(device)
+    return model
+
+# Usage example
+if __name__ == "__main__":
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    model_name = 'model_weights.pth'
+    model_path = f'models/{model_name}'
+
+    model = LWM()
+    
+    model = load_model(model, model_path, device)
+    print(f"Model loaded successfully on {device}")
+    print(f"Model parameters: {sum(p.numel() for p in model.parameters())}")

model.py

@@ -0,0 +1,160 @@
+# -*- coding: utf-8 -*-
+"""
+Created on Sun Sep 15 19:55:23 2024
+
+@author: salikha4
+"""
+
+import os
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+
+from inference import *
+from load_data import load_DeepMIMO_data
+from input_preprocess import *
+from lwm_model import LWM, load_model
+
+
+ELEMENT_LENGTH = 16
+D_MODEL = 64
+MAX_LEN = 129
+N_LAYERS = 12
+N_HEADS = 12
+D_FF = D_MODEL * 4
+D_K = D_MODEL // N_HEADS
+D_V = D_MODEL // N_HEADS
+DROPOUT = 0.1
+
+class LayerNormalization(nn.Module):
+    def __init__(self, d_model: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.eps = eps
+        self.alpha = nn.Parameter(torch.ones(d_model))
+        self.bias = nn.Parameter(torch.zeros(d_model))
+
+    def forward(self, x):
+        mean = x.mean(dim=-1, keepdim=True)
+        std = x.std(dim=-1, keepdim=True)
+        return self.alpha * (x - mean) / (std + self.eps) + self.bias
+
+class Embedding(nn.Module):
+    def __init__(self, element_length, d_model, max_len):
+        super().__init__()
+        self.element_length = element_length
+        self.d_model = d_model
+        self.proj = nn.Linear(element_length, d_model)
+        self.pos_embed = nn.Embedding(max_len, d_model)
+        self.norm = LayerNormalization(d_model)
+
+    def forward(self, x):
+        seq_len = x.size(1)
+        pos = torch.arange(seq_len, dtype=torch.long, device=x.device)
+        pos = pos.unsqueeze(0).expand_as(x[:, :, 0])
+        tok_emb = self.proj(x.float())
+        embedding = tok_emb + self.pos_embed(pos)
+        return self.norm(embedding)
+
+class ScaledDotProductAttention(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, Q, K, V):
+        scores = torch.matmul(Q, K.transpose(-1, -2)) / np.sqrt(D_K)
+        attn = F.softmax(scores, dim=-1)
+        context = torch.matmul(attn, V)
+        return context, attn
+
+class MultiHeadAttention(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.W_Q = nn.Linear(D_MODEL, D_K * N_HEADS)
+        self.W_K = nn.Linear(D_MODEL, D_K * N_HEADS)
+        self.W_V = nn.Linear(D_MODEL, D_V * N_HEADS)
+        self.linear = nn.Linear(N_HEADS * D_V, D_MODEL)
+        self.norm = LayerNormalization(D_MODEL)
+        self.dropout = nn.Dropout(DROPOUT)
+        
+    def forward(self, Q, K, V):
+        residual, batch_size = Q, Q.size(0)
+        q_s = self.W_Q(Q).view(batch_size, -1, N_HEADS, D_K).transpose(1, 2)
+        k_s = self.W_K(K).view(batch_size, -1, N_HEADS, D_K).transpose(1, 2)
+        v_s = self.W_V(V).view(batch_size, -1, N_HEADS, D_V).transpose(1, 2)
+
+        context, attn = ScaledDotProductAttention()(q_s, k_s, v_s)
+        output = context.transpose(1, 2).contiguous().view(batch_size, -1, N_HEADS * D_V)
+        output = self.linear(output)
+        return residual + self.dropout(output), attn #residual + self.dropout(output), attn
+
+class PoswiseFeedForwardNet(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.fc1 = nn.Linear(D_MODEL, D_FF)
+        self.fc2 = nn.Linear(D_FF, D_MODEL)
+        self.dropout = nn.Dropout(DROPOUT)
+        self.norm = LayerNormalization(D_MODEL)
+
+    def forward(self, x):
+        output = self.fc2(self.dropout(F.relu(self.fc1(x))))
+        return x + self.dropout(output) #x + self.dropout(output)
+
+class EncoderLayer(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.enc_self_attn = MultiHeadAttention()
+        self.pos_ffn = PoswiseFeedForwardNet()
+        self.norm = LayerNormalization(D_MODEL)
+
+    def forward(self, enc_inputs):
+        attn_outputs, attn = self.enc_self_attn(enc_inputs, enc_inputs, enc_inputs)
+        attn_outputs = self.norm(attn_outputs)
+        enc_outputs = self.pos_ffn(attn_outputs)
+        return enc_outputs, attn
+    
+class LWM(torch.nn.Module):
+    def __init__(self, element_length=16, d_model=64, max_len=129, n_layers=12):
+        super().__init__()
+
+        self.embedding = Embedding(element_length, d_model, max_len)
+        self.layers = nn.ModuleList([EncoderLayer() for _ in range(n_layers)])
+        self.linear = nn.Linear(d_model, d_model)
+        self.norm = LayerNormalization(d_model)
+
+        embed_weight = self.embedding.proj.weight
+        d_model, n_dim = embed_weight.size()
+        self.decoder = nn.Linear(d_model, n_dim, bias=False)
+        self.decoder.weight = nn.Parameter(embed_weight.transpose(0, 1))
+        self.decoder_bias = nn.Parameter(torch.zeros(n_dim))
+
+    def forward(self, input_ids, masked_pos):
+        output = self.embedding(input_ids)
+        
+        for layer in self.layers:
+            output, _ = layer(output)
+
+        masked_pos = masked_pos.long()[:, :, None].expand(-1, -1, output.size(-1))
+        h_masked = torch.gather(output, 1, masked_pos)
+        h_masked = self.norm(F.relu(self.linear(h_masked)))
+        logits_lm = self.decoder(h_masked) + self.decoder_bias
+
+        return logits_lm, output
+
+    @classmethod
+    def from_pretrained(cls, ckpt_name='model_weights.pth', device='cuda'):
+        # Define model
+        model = cls().to(device)
+        
+        # Download the model weights (from a remote or local repository)
+        ckpt_path = f'https://huggingface.co/sadjadalikhani/lwm/resolve/main/{ckpt_name}'
+        
+        # Load the model weights
+        model.load_state_dict(torch.hub.load_state_dict_from_url(ckpt_path, map_location=device))
+        print(f"Model loaded successfully from {ckpt_path} to {device}")
+        
+        return model
+
+    def forward(self, x):
+        # Define the forward pass for the model
+        pass
+