Spaces:
Running
Running
Sadjad Alikhani
commited on
Update app.py
Browse files
app.py
CHANGED
@@ -16,7 +16,7 @@ RAW_PATH = os.path.join("images", "raw")
|
|
16 |
EMBEDDINGS_PATH = os.path.join("images", "embeddings")
|
17 |
|
18 |
# Specific values for percentage of data for training
|
19 |
-
percentage_values = [10, 30, 50, 70, 100]
|
20 |
|
21 |
# Custom class to capture print output
|
22 |
class PrintCapture(io.StringIO):
|
@@ -47,6 +47,15 @@ def create_random_image(size=(300, 300)):
|
|
47 |
random_image = np.random.rand(*size, 3) * 255
|
48 |
return Image.fromarray(random_image.astype('uint8'))
|
49 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
50 |
# Function to dynamically load a Python module from a given file path
|
51 |
def load_module_from_path(module_name, file_path):
|
52 |
spec = importlib.util.spec_from_file_location(module_name, file_path)
|
@@ -59,19 +68,42 @@ def split_dataset(channels, labels, percentage_idx):
|
|
59 |
percentage = percentage_values[percentage_idx] / 100
|
60 |
num_samples = channels.shape[0]
|
61 |
train_size = int(num_samples * percentage)
|
|
|
|
|
62 |
indices = np.arange(num_samples)
|
63 |
np.random.shuffle(indices)
|
64 |
|
65 |
train_idx, test_idx = indices[:train_size], indices[train_size:]
|
|
|
66 |
train_data, test_data = channels[train_idx], channels[test_idx]
|
67 |
train_labels, test_labels = labels[train_idx], labels[test_idx]
|
68 |
|
69 |
return train_data, test_data, train_labels, test_labels
|
70 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
71 |
# Function to generate confusion matrix plot
|
72 |
def plot_confusion_matrix(y_true, y_pred, title):
|
73 |
cm = confusion_matrix(y_true, y_pred)
|
74 |
-
plt.figure(figsize=(
|
75 |
plt.imshow(cm, cmap='Blues')
|
76 |
plt.title(title)
|
77 |
plt.xlabel('Predicted')
|
@@ -84,76 +116,101 @@ def plot_confusion_matrix(y_true, y_pred, title):
|
|
84 |
return Image.open(f"{title}.png")
|
85 |
|
86 |
def identical_train_test_split(output_emb, output_raw, labels, percentage):
|
87 |
-
N = output_emb.shape[0]
|
88 |
-
|
|
|
|
|
|
|
|
|
89 |
split_index = int(N * percentage)
|
90 |
-
|
91 |
-
|
|
|
|
|
|
|
|
|
92 |
train_emb = output_emb[train_indices]
|
93 |
test_emb = output_emb[test_indices]
|
|
|
94 |
train_raw = output_raw[train_indices]
|
95 |
test_raw = output_raw[test_indices]
|
|
|
96 |
train_labels = labels[train_indices]
|
97 |
test_labels = labels[test_indices]
|
98 |
|
99 |
return train_emb, test_emb, train_raw, test_raw, train_labels, test_labels
|
100 |
|
101 |
-
# Function to classify test data based on distance to class centroids
|
102 |
-
def classify_based_on_distance(train_data, train_labels, test_data):
|
103 |
-
centroid_0 = train_data[train_labels == 0].mean(dim=0)
|
104 |
-
centroid_1 = train_data[train_labels == 1].mean(dim=0)
|
105 |
-
|
106 |
-
predictions = []
|
107 |
-
for test_point in test_data:
|
108 |
-
dist_0 = torch.norm(test_point - centroid_0)
|
109 |
-
dist_1 = torch.norm(test_point - centroid_1)
|
110 |
-
predictions.append(0 if dist_0 < dist_1 else 1)
|
111 |
-
|
112 |
-
return torch.tensor(predictions)
|
113 |
-
|
114 |
# Store the original working directory when the app starts
|
115 |
original_dir = os.getcwd()
|
116 |
|
117 |
def process_hdf5_file(uploaded_file, percentage_idx):
|
118 |
capture = PrintCapture()
|
119 |
-
sys.stdout = capture
|
120 |
|
121 |
try:
|
122 |
model_repo_url = "https://huggingface.co/sadjadalikhani/LWM"
|
123 |
model_repo_dir = "./LWM"
|
|
|
|
|
124 |
if not os.path.exists(model_repo_dir):
|
|
|
125 |
subprocess.run(["git", "clone", model_repo_url, model_repo_dir], check=True)
|
|
|
|
|
126 |
repo_work_dir = os.path.join(original_dir, model_repo_dir)
|
127 |
if os.path.exists(repo_work_dir):
|
128 |
-
os.chdir(repo_work_dir)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
129 |
lwm_model_path = os.path.join(os.getcwd(), 'lwm_model.py')
|
130 |
input_preprocess_path = os.path.join(os.getcwd(), 'input_preprocess.py')
|
131 |
inference_path = os.path.join(os.getcwd(), 'inference.py')
|
132 |
|
|
|
133 |
lwm_model = load_module_from_path("lwm_model", lwm_model_path)
|
|
|
|
|
134 |
input_preprocess = load_module_from_path("input_preprocess", input_preprocess_path)
|
|
|
|
|
135 |
inference = load_module_from_path("inference", inference_path)
|
136 |
|
|
|
137 |
device = 'cpu'
|
|
|
138 |
model = lwm_model.LWM.from_pretrained(device=device)
|
139 |
|
|
|
140 |
with h5py.File(uploaded_file.name, 'r') as f:
|
141 |
-
channels = np.array(f['channels'])
|
142 |
-
labels = np.array(f['labels'])
|
|
|
143 |
|
|
|
144 |
preprocessed_chs = input_preprocess.tokenizer(manual_data=channels)
|
|
|
|
|
145 |
output_emb = inference.lwm_inference(preprocessed_chs, 'channel_emb', model)
|
146 |
output_raw = inference.create_raw_dataset(preprocessed_chs, device)
|
147 |
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
|
|
|
|
153 |
|
|
|
154 |
pred_raw = classify_based_on_distance(train_data_raw, train_labels, test_data_raw)
|
155 |
pred_emb = classify_based_on_distance(train_data_emb, train_labels, test_data_emb)
|
156 |
|
|
|
157 |
raw_cm_image = plot_confusion_matrix(test_labels, pred_raw, title="Confusion Matrix (Raw Channels)")
|
158 |
emb_cm_image = plot_confusion_matrix(test_labels, pred_emb, title="Confusion Matrix (Embeddings)")
|
159 |
|
@@ -163,28 +220,29 @@ def process_hdf5_file(uploaded_file, percentage_idx):
|
|
163 |
return str(e), str(e), capture.get_output()
|
164 |
|
165 |
finally:
|
|
|
166 |
os.chdir(original_dir)
|
167 |
-
sys.stdout = sys.__stdout__
|
168 |
|
|
|
169 |
def los_nlos_classification(file, percentage_idx):
|
170 |
if file is not None:
|
171 |
return process_hdf5_file(file, percentage_idx)
|
172 |
else:
|
173 |
return display_predefined_images(percentage_idx), None
|
174 |
|
175 |
-
# Define the Gradio interface
|
176 |
with gr.Blocks(css="""
|
|
|
|
|
|
|
|
|
|
|
|
|
177 |
.slider-container {
|
|
|
|
|
178 |
text-align: center;
|
179 |
-
margin-bottom: 20px;
|
180 |
-
}
|
181 |
-
.image-row {
|
182 |
-
justify-content: center;
|
183 |
-
margin-top: 10px;
|
184 |
-
}
|
185 |
-
.output-box {
|
186 |
-
max-width: 600px;
|
187 |
-
margin: 0 auto;
|
188 |
}
|
189 |
""") as demo:
|
190 |
|
@@ -203,24 +261,33 @@ with gr.Blocks(css="""
|
|
203 |
# Tabs for Beam Prediction and LoS/NLoS Classification
|
204 |
with gr.Tab("Beam Prediction Task"):
|
205 |
gr.Markdown("### Beam Prediction Task")
|
|
|
206 |
with gr.Row():
|
207 |
with gr.Column(elem_id="slider-container"):
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
|
|
|
|
|
|
212 |
percentage_slider_bp.change(fn=display_predefined_images, inputs=[percentage_slider_bp], outputs=[raw_img_bp, embeddings_img_bp])
|
213 |
|
214 |
with gr.Tab("LoS/NLoS Classification Task"):
|
215 |
gr.Markdown("### LoS/NLoS Classification Task")
|
|
|
216 |
file_input = gr.File(label="Upload HDF5 Dataset", file_types=[".h5"])
|
|
|
217 |
with gr.Row():
|
218 |
with gr.Column(elem_id="slider-container"):
|
219 |
-
|
220 |
-
|
221 |
-
|
222 |
-
|
223 |
-
|
|
|
|
|
|
|
224 |
file_input.change(fn=los_nlos_classification, inputs=[file_input, percentage_slider_los], outputs=[raw_img_los, embeddings_img_los, output_textbox])
|
225 |
percentage_slider_los.change(fn=los_nlos_classification, inputs=[file_input, percentage_slider_los], outputs=[raw_img_los, embeddings_img_los, output_textbox])
|
226 |
|
|
|
16 |
EMBEDDINGS_PATH = os.path.join("images", "embeddings")
|
17 |
|
18 |
# Specific values for percentage of data for training
|
19 |
+
percentage_values = [10, 20, 30, 40, 50, 60, 70, 80, 90, 100]
|
20 |
|
21 |
# Custom class to capture print output
|
22 |
class PrintCapture(io.StringIO):
|
|
|
47 |
random_image = np.random.rand(*size, 3) * 255
|
48 |
return Image.fromarray(random_image.astype('uint8'))
|
49 |
|
50 |
+
# Function to load the pre-trained model from your cloned repository
|
51 |
+
def load_custom_model():
|
52 |
+
from lwm_model import LWM # Assuming the model is defined in lwm_model.py
|
53 |
+
model = LWM() # Modify this according to your model initialization
|
54 |
+
model.eval()
|
55 |
+
return model
|
56 |
+
|
57 |
+
import importlib.util
|
58 |
+
|
59 |
# Function to dynamically load a Python module from a given file path
|
60 |
def load_module_from_path(module_name, file_path):
|
61 |
spec = importlib.util.spec_from_file_location(module_name, file_path)
|
|
|
68 |
percentage = percentage_values[percentage_idx] / 100
|
69 |
num_samples = channels.shape[0]
|
70 |
train_size = int(num_samples * percentage)
|
71 |
+
print(f'Number of Training Samples: {train_size}')
|
72 |
+
|
73 |
indices = np.arange(num_samples)
|
74 |
np.random.shuffle(indices)
|
75 |
|
76 |
train_idx, test_idx = indices[:train_size], indices[train_size:]
|
77 |
+
|
78 |
train_data, test_data = channels[train_idx], channels[test_idx]
|
79 |
train_labels, test_labels = labels[train_idx], labels[test_idx]
|
80 |
|
81 |
return train_data, test_data, train_labels, test_labels
|
82 |
|
83 |
+
# Function to calculate Euclidean distance between a point and a centroid
|
84 |
+
def euclidean_distance(x, centroid):
|
85 |
+
return np.linalg.norm(x - centroid)
|
86 |
+
|
87 |
+
import torch
|
88 |
+
|
89 |
+
def classify_based_on_distance(train_data, train_labels, test_data):
|
90 |
+
# Compute the centroids for the two classes
|
91 |
+
centroid_0 = train_data[train_labels == 0].mean(dim=0) # Use torch.mean
|
92 |
+
centroid_1 = train_data[train_labels == 1].mean(dim=0) # Use torch.mean
|
93 |
+
|
94 |
+
predictions = []
|
95 |
+
for test_point in test_data:
|
96 |
+
# Compute Euclidean distance between the test point and each centroid
|
97 |
+
dist_0 = euclidean_distance(test_point, centroid_0)
|
98 |
+
dist_1 = euclidean_distance(test_point, centroid_1)
|
99 |
+
predictions.append(0 if dist_0 < dist_1 else 1)
|
100 |
+
|
101 |
+
return torch.tensor(predictions) # Return predictions as a PyTorch tensor
|
102 |
+
|
103 |
# Function to generate confusion matrix plot
|
104 |
def plot_confusion_matrix(y_true, y_pred, title):
|
105 |
cm = confusion_matrix(y_true, y_pred)
|
106 |
+
plt.figure(figsize=(5, 5))
|
107 |
plt.imshow(cm, cmap='Blues')
|
108 |
plt.title(title)
|
109 |
plt.xlabel('Predicted')
|
|
|
116 |
return Image.open(f"{title}.png")
|
117 |
|
118 |
def identical_train_test_split(output_emb, output_raw, labels, percentage):
|
119 |
+
N = output_emb.shape[0] # Get the total number of samples
|
120 |
+
|
121 |
+
# Generate the indices for shuffling and splitting
|
122 |
+
indices = torch.randperm(N) # Randomly shuffle the indices
|
123 |
+
|
124 |
+
# Calculate the split index
|
125 |
split_index = int(N * percentage)
|
126 |
+
|
127 |
+
# Split indices into train and test
|
128 |
+
train_indices = indices[:split_index] # First 80% for training
|
129 |
+
test_indices = indices[split_index:] # Remaining 20% for testing
|
130 |
+
|
131 |
+
# Select the same indices from both output_emb and output_raw
|
132 |
train_emb = output_emb[train_indices]
|
133 |
test_emb = output_emb[test_indices]
|
134 |
+
|
135 |
train_raw = output_raw[train_indices]
|
136 |
test_raw = output_raw[test_indices]
|
137 |
+
|
138 |
train_labels = labels[train_indices]
|
139 |
test_labels = labels[test_indices]
|
140 |
|
141 |
return train_emb, test_emb, train_raw, test_raw, train_labels, test_labels
|
142 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
143 |
# Store the original working directory when the app starts
|
144 |
original_dir = os.getcwd()
|
145 |
|
146 |
def process_hdf5_file(uploaded_file, percentage_idx):
|
147 |
capture = PrintCapture()
|
148 |
+
sys.stdout = capture # Redirect print statements to capture
|
149 |
|
150 |
try:
|
151 |
model_repo_url = "https://huggingface.co/sadjadalikhani/LWM"
|
152 |
model_repo_dir = "./LWM"
|
153 |
+
|
154 |
+
# Step 1: Clone the repository if not already done
|
155 |
if not os.path.exists(model_repo_dir):
|
156 |
+
print(f"Cloning model repository from {model_repo_url}...")
|
157 |
subprocess.run(["git", "clone", model_repo_url, model_repo_dir], check=True)
|
158 |
+
|
159 |
+
# Step 2: Verify the repository was cloned and change the working directory
|
160 |
repo_work_dir = os.path.join(original_dir, model_repo_dir)
|
161 |
if os.path.exists(repo_work_dir):
|
162 |
+
os.chdir(repo_work_dir) # Change the working directory only once
|
163 |
+
print(f"Changed working directory to {os.getcwd()}")
|
164 |
+
print(f"Directory content: {os.listdir(os.getcwd())}") # Debugging: Check repo content
|
165 |
+
else:
|
166 |
+
print(f"Directory {repo_work_dir} does not exist.")
|
167 |
+
return
|
168 |
+
|
169 |
+
# Step 3: Dynamically load lwm_model.py, input_preprocess.py, and inference.py
|
170 |
lwm_model_path = os.path.join(os.getcwd(), 'lwm_model.py')
|
171 |
input_preprocess_path = os.path.join(os.getcwd(), 'input_preprocess.py')
|
172 |
inference_path = os.path.join(os.getcwd(), 'inference.py')
|
173 |
|
174 |
+
# Load lwm_model
|
175 |
lwm_model = load_module_from_path("lwm_model", lwm_model_path)
|
176 |
+
|
177 |
+
# Load input_preprocess
|
178 |
input_preprocess = load_module_from_path("input_preprocess", input_preprocess_path)
|
179 |
+
|
180 |
+
# Load inference
|
181 |
inference = load_module_from_path("inference", inference_path)
|
182 |
|
183 |
+
# Step 4: Load the model from lwm_model module
|
184 |
device = 'cpu'
|
185 |
+
print(f"Loading the LWM model on {device}...")
|
186 |
model = lwm_model.LWM.from_pretrained(device=device)
|
187 |
|
188 |
+
# Step 5: Load the HDF5 file and extract the channels and labels
|
189 |
with h5py.File(uploaded_file.name, 'r') as f:
|
190 |
+
channels = np.array(f['channels']) # Assuming 'channels' dataset in the HDF5 file
|
191 |
+
labels = np.array(f['labels']) # Assuming 'labels' dataset in the HDF5 file
|
192 |
+
print(f"Loaded dataset with {channels.shape[0]} samples.")
|
193 |
|
194 |
+
# Step 7: Tokenize the data using the tokenizer from input_preprocess
|
195 |
preprocessed_chs = input_preprocess.tokenizer(manual_data=channels)
|
196 |
+
|
197 |
+
# Step 7: Perform inference using the functions from inference.py
|
198 |
output_emb = inference.lwm_inference(preprocessed_chs, 'channel_emb', model)
|
199 |
output_raw = inference.create_raw_dataset(preprocessed_chs, device)
|
200 |
|
201 |
+
print(f"Output Embeddings Shape: {output_emb.shape}")
|
202 |
+
print(f"Output Raw Shape: {output_raw.shape}")
|
203 |
+
|
204 |
+
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),
|
205 |
+
output_raw.view(len(output_raw),-1),
|
206 |
+
labels,
|
207 |
+
percentage_idx)
|
208 |
|
209 |
+
# Step 8: Perform classification using the Euclidean distance for both raw and embeddings
|
210 |
pred_raw = classify_based_on_distance(train_data_raw, train_labels, test_data_raw)
|
211 |
pred_emb = classify_based_on_distance(train_data_emb, train_labels, test_data_emb)
|
212 |
|
213 |
+
# Step 9: Generate confusion matrices for both raw and embeddings
|
214 |
raw_cm_image = plot_confusion_matrix(test_labels, pred_raw, title="Confusion Matrix (Raw Channels)")
|
215 |
emb_cm_image = plot_confusion_matrix(test_labels, pred_emb, title="Confusion Matrix (Embeddings)")
|
216 |
|
|
|
220 |
return str(e), str(e), capture.get_output()
|
221 |
|
222 |
finally:
|
223 |
+
# Always return to the original working directory after processing
|
224 |
os.chdir(original_dir)
|
225 |
+
sys.stdout = sys.__stdout__ # Reset print statements
|
226 |
|
227 |
+
# Function to handle logic based on whether a file is uploaded or not
|
228 |
def los_nlos_classification(file, percentage_idx):
|
229 |
if file is not None:
|
230 |
return process_hdf5_file(file, percentage_idx)
|
231 |
else:
|
232 |
return display_predefined_images(percentage_idx), None
|
233 |
|
234 |
+
# Define the Gradio interface
|
235 |
with gr.Blocks(css="""
|
236 |
+
.vertical-slider input[type=range] {
|
237 |
+
writing-mode: bt-lr; /* IE */
|
238 |
+
-webkit-appearance: slider-vertical; /* WebKit */
|
239 |
+
width: 8px;
|
240 |
+
height: 200px;
|
241 |
+
}
|
242 |
.slider-container {
|
243 |
+
display: inline-block;
|
244 |
+
margin-right: 50px;
|
245 |
text-align: center;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
246 |
}
|
247 |
""") as demo:
|
248 |
|
|
|
261 |
# Tabs for Beam Prediction and LoS/NLoS Classification
|
262 |
with gr.Tab("Beam Prediction Task"):
|
263 |
gr.Markdown("### Beam Prediction Task")
|
264 |
+
|
265 |
with gr.Row():
|
266 |
with gr.Column(elem_id="slider-container"):
|
267 |
+
gr.Markdown("Percentage of Data for Training")
|
268 |
+
percentage_slider_bp = gr.Slider(minimum=0, maximum=4, step=1, value=0, interactive=True, elem_id="vertical-slider")
|
269 |
+
|
270 |
+
with gr.Row():
|
271 |
+
raw_img_bp = gr.Image(label="Raw Channels", type="pil", width=300, height=300, interactive=False)
|
272 |
+
embeddings_img_bp = gr.Image(label="Embeddings", type="pil", width=300, height=300, interactive=False)
|
273 |
+
|
274 |
percentage_slider_bp.change(fn=display_predefined_images, inputs=[percentage_slider_bp], outputs=[raw_img_bp, embeddings_img_bp])
|
275 |
|
276 |
with gr.Tab("LoS/NLoS Classification Task"):
|
277 |
gr.Markdown("### LoS/NLoS Classification Task")
|
278 |
+
|
279 |
file_input = gr.File(label="Upload HDF5 Dataset", file_types=[".h5"])
|
280 |
+
|
281 |
with gr.Row():
|
282 |
with gr.Column(elem_id="slider-container"):
|
283 |
+
gr.Markdown("Percentage of Data for Training")
|
284 |
+
percentage_slider_los = gr.Slider(minimum=0, maximum=4, step=1, value=0, interactive=True, elem_id="vertical-slider")
|
285 |
+
|
286 |
+
with gr.Row():
|
287 |
+
raw_img_los = gr.Image(label="Raw Channels", type="pil", width=300, height=300, interactive=False)
|
288 |
+
embeddings_img_los = gr.Image(label="Embeddings", type="pil", width=300, height=300, interactive=False)
|
289 |
+
output_textbox = gr.Textbox(label="Console Output", lines=10)
|
290 |
+
|
291 |
file_input.change(fn=los_nlos_classification, inputs=[file_input, percentage_slider_los], outputs=[raw_img_los, embeddings_img_los, output_textbox])
|
292 |
percentage_slider_los.change(fn=los_nlos_classification, inputs=[file_input, percentage_slider_los], outputs=[raw_img_los, embeddings_img_los, output_textbox])
|
293 |
|