import gradio as gr import os from PIL import Image import numpy as np import pickle import io import sys import torch import subprocess import h5py from sklearn.metrics import confusion_matrix import matplotlib.pyplot as plt # Paths to the predefined images folder RAW_PATH = os.path.join("images", "raw") EMBEDDINGS_PATH = os.path.join("images", "embeddings") # Specific values for percentage of data for training percentage_values = np.arange(10) + 1 # Custom class to capture print output class PrintCapture(io.StringIO): def __init__(self): super().__init__() self.output = [] def write(self, txt): self.output.append(txt) super().write(txt) def get_output(self): return ''.join(self.output) # Function to load and display predefined images based on user selection def display_predefined_images(percentage_idx): percentage = percentage_values[percentage_idx] raw_image_path = os.path.join(RAW_PATH, f"percentage_{percentage}_complexity_16.png") embeddings_image_path = os.path.join(EMBEDDINGS_PATH, f"percentage_{percentage}_complexity_16.png") # Check if the images exist if os.path.exists(raw_image_path): raw_image = Image.open(raw_image_path) else: raw_image = create_random_image() # Use a fallback random image if os.path.exists(embeddings_image_path): embeddings_image = Image.open(embeddings_image_path) else: embeddings_image = create_random_image() # Use a fallback random image return raw_image, embeddings_image # Function to create random images for LoS/NLoS classification results def create_random_image(size=(300, 300)): random_image = np.random.rand(*size, 3) * 255 return Image.fromarray(random_image.astype('uint8')) # Function to split dataset into training and test sets based on user selection def identical_train_test_split(output_emb, output_raw, labels, percentage_idx): N = output_emb.shape[0] # Get the total number of samples # Generate the indices for shuffling and splitting indices = torch.randperm(N) # Randomly shuffle the indices # Calculate the split index split_index = int(N * percentage_values[percentage_idx] / 10) # Convert percentage index to percentage value print(f'Training Size: {split_index}') # Split indices into train and test train_indices = indices[:split_index] test_indices = indices[split_index:] # Select the same indices from both output_emb and output_raw train_emb = output_emb[train_indices] test_emb = output_emb[test_indices] train_raw = output_raw[train_indices] test_raw = output_raw[test_indices] train_labels = labels[train_indices] test_labels = labels[test_indices] return train_emb, test_emb, train_raw, test_raw, train_labels, test_labels # Function to calculate Euclidean distance between a point and a centroid def classify_based_on_distance(train_data, train_labels, test_data): centroid_0 = train_data[train_labels == 0].mean(dim=0) centroid_1 = train_data[train_labels == 1].mean(dim=0) predictions = [] for test_point in test_data: dist_0 = torch.norm(test_point - centroid_0) dist_1 = torch.norm(test_point - centroid_1) predictions.append(0 if dist_0 < dist_1 else 1) return torch.tensor(predictions) # Function to generate confusion matrix plot def plot_confusion_matrix(y_true, y_pred, title): cm = confusion_matrix(y_true, y_pred) plt.figure(figsize=(5, 5)) plt.imshow(cm, cmap='Blues') plt.title(title) plt.xlabel('Predicted') plt.ylabel('Actual') plt.colorbar() plt.xticks([0, 1], labels=[0, 1]) plt.yticks([0, 1], labels=[0, 1]) plt.tight_layout() plt.savefig(f"{title}.png") return Image.open(f"{title}.png") # Function to handle inference and return the results (store the results to state) def run_inference(uploaded_file): capture = PrintCapture() sys.stdout = capture # Redirect print statements to capture try: # Load the HDF5 file and extract channels and labels with h5py.File(uploaded_file.name, 'r') as f: channels = np.array(f['channels']) # Assuming 'channels' dataset in the HDF5 file labels = np.array(f['labels']) # Assuming 'labels' dataset in the HDF5 file print(f"Loaded dataset with {channels.shape[0]} samples.") # Run the tokenization and model inference model_repo_url = "https://huggingface.co/sadjadalikhani/LWM" model_repo_dir = "./LWM" if not os.path.exists(model_repo_dir): print(f"Cloning model repository from {model_repo_url}...") subprocess.run(["git", "clone", model_repo_url, model_repo_dir], check=True) # Load the model lwm_model_path = os.path.join(model_repo_dir, 'lwm_model.py') input_preprocess_path = os.path.join(model_repo_dir, 'input_preprocess.py') inference_path = os.path.join(model_repo_dir, 'inference.py') # Load dynamically lwm_model = load_module_from_path("lwm_model", lwm_model_path) input_preprocess = load_module_from_path("input_preprocess", input_preprocess_path) inference = load_module_from_path("inference", inference_path) device = 'cuda' if torch.cuda.is_available() else 'cpu' print(f"Loading LWM model on {device}...") model = lwm_model.LWM.from_pretrained(device=device).to(torch.float32) # Preprocess and inference preprocessed_chs = input_preprocess.tokenizer(manual_data=channels) output_emb = inference.lwm_inference(preprocessed_chs, 'channel_emb', model) output_raw = inference.create_raw_dataset(preprocessed_chs, device) print(f"Output Embeddings Shape: {output_emb.shape}") print(f"Output Raw Shape: {output_raw.shape}") return output_emb, output_raw, labels, capture.get_output() except Exception as e: return None, None, None, str(e) finally: sys.stdout = sys.__stdout__ # Reset print statements # Function to handle classification after inference (using Gradio state) def los_nlos_classification(output_emb, output_raw, labels, percentage_idx): if output_emb is not None and output_raw is not None: train_data_emb, test_data_emb, train_data_raw, test_data_raw, train_labels, test_labels = identical_train_test_split( output_emb.view(len(output_emb), -1), output_raw.view(len(output_raw), -1), labels, percentage_idx ) pred_raw = classify_based_on_distance(train_data_raw, train_labels, test_data_raw) pred_emb = classify_based_on_distance(train_data_emb, train_labels, test_data_emb) raw_cm_image = plot_confusion_matrix(test_labels, pred_raw, title="Confusion Matrix (Raw Channels)") emb_cm_image = plot_confusion_matrix(test_labels, pred_emb, title="Confusion Matrix (Embeddings)") return raw_cm_image, emb_cm_image, "Classification successful" return create_random_image(), create_random_image(), "No valid inference outputs" # Define the Gradio interface with gr.Blocks(css=""" .vertical-slider input[type=range] { writing-mode: bt-lr; /* IE */ -webkit-appearance: slider-vertical; /* WebKit */ width: 8px; height: 200px; } .slider-container { display: inline-block; margin-right: 50px; text-align: center; } """) as demo: # Tabs for Beam Prediction and LoS/NLoS Classification with gr.Tab("LoS/NLoS Classification Task"): gr.Markdown("### LoS/NLoS Classification Task") file_input = gr.File(label="Upload HDF5 Dataset", file_types=[".h5"]) with gr.Row(): percentage_dropdown_los = gr.Dropdown( choices=[str(v) for v in percentage_values * 10], value=10, label="Percentage of Data for Training", interactive=True ) with gr.Row(): raw_img_los = gr.Image(label="Raw Channels", type="pil", width=300, height=300, interactive=False) embeddings_img_los = gr.Image(label="Embeddings", type="pil", width=300, height=300, interactive=False) output_textbox = gr.Textbox(label="Console Output", lines=10) # Process file upload to run inference inference_output = gr.State() file_input.upload(run_inference, inputs=file_input, outputs=inference_output) # Handle dropdown change for classification percentage_dropdown_los.change( fn=los_nlos_classification, inputs=[inference_output['output_emb'], inference_output['output_raw'], inference_output['labels'], percentage_dropdown_los], outputs=[raw_img_los, embeddings_img_los, output_textbox] ) # Launch the app if __name__ == "__main__": demo.launch()