Update app.py
Browse files
app.py
CHANGED
@@ -16,6 +16,7 @@ import pandas as pd # Para formatear la salida en tabla
|
|
16 |
# --- Configuración ---
|
17 |
MODEL_REPO_ID = "google/cxr-foundation"
|
18 |
MODEL_DOWNLOAD_DIR = './hf_cxr_foundation_space' # Directorio dentro del contenedor del Space
|
|
|
19 |
SIMILARITY_DIFFERENCE_THRESHOLD = 0.1
|
20 |
POSITIVE_SIMILARITY_THRESHOLD = 0.1
|
21 |
print(f"Usando umbrales: Comp Δ={SIMILARITY_DIFFERENCE_THRESHOLD}, Simp τ={POSITIVE_SIMILARITY_THRESHOLD}")
|
@@ -30,151 +31,255 @@ criteria_list_negative = [
|
|
30 |
"cropped image", "scapulae overlying lungs", "blurred image", "obscuring artifact"
|
31 |
]
|
32 |
|
33 |
-
# --- Funciones Auxiliares ---
|
|
|
|
|
|
|
|
|
|
|
34 |
def bert_tokenize(text, preprocessor):
|
35 |
-
|
|
|
|
|
36 |
if not isinstance(text, str): text = str(text)
|
|
|
|
|
37 |
out = preprocessor(tf.constant([text.lower()]))
|
|
|
|
|
38 |
ids = out['input_word_ids'].numpy().astype(np.int32)
|
39 |
masks = out['input_mask'].numpy().astype(np.float32)
|
40 |
paddings = 1.0 - masks
|
|
|
|
|
41 |
end_token_idx = (ids == 102)
|
42 |
ids[end_token_idx] = 0
|
43 |
paddings[end_token_idx] = 1.0
|
|
|
|
|
|
|
44 |
if ids.ndim == 2: ids = np.expand_dims(ids, axis=1)
|
45 |
if paddings.ndim == 2: paddings = np.expand_dims(paddings, axis=1)
|
|
|
|
|
46 |
expected_shape = (1, 1, 128)
|
47 |
if ids.shape != expected_shape:
|
|
|
48 |
if ids.shape == (1,128): ids = np.expand_dims(ids, axis=1)
|
49 |
else: raise ValueError(f"Shape incorrecta para ids: {ids.shape}, esperado {expected_shape}")
|
50 |
if paddings.shape != expected_shape:
|
51 |
if paddings.shape == (1,128): paddings = np.expand_dims(paddings, axis=1)
|
52 |
else: raise ValueError(f"Shape incorrecta para paddings: {paddings.shape}, esperado {expected_shape}")
|
|
|
53 |
return ids, paddings
|
54 |
|
55 |
def png_to_tfexample(image_array: np.ndarray) -> tf.train.Example:
|
|
|
56 |
if image_array.ndim == 3 and image_array.shape[2] == 1:
|
57 |
-
image_array = np.squeeze(image_array, axis=2)
|
58 |
elif image_array.ndim != 2:
|
59 |
-
raise ValueError(f'Array debe ser 2-D. Dimensiones: {image_array.ndim}')
|
|
|
60 |
image = image_array.astype(np.float32)
|
61 |
-
min_val
|
|
|
|
|
|
|
62 |
if max_val <= min_val:
|
|
|
|
|
63 |
if image_array.dtype == np.uint8 or (min_val >= 0 and max_val <= 255):
|
64 |
-
pixel_array = image.astype(np.uint8)
|
65 |
-
|
66 |
-
|
|
|
|
|
67 |
else:
|
68 |
-
image -= min_val
|
69 |
current_max = max_val - min_val
|
|
|
70 |
if image_array.dtype != np.uint8:
|
71 |
image *= 65535 / current_max
|
72 |
-
pixel_array = image.astype(np.uint16)
|
|
|
73 |
else:
|
|
|
|
|
|
|
74 |
image *= 255 / current_max
|
75 |
-
pixel_array = image.astype(np.uint8)
|
|
|
|
|
|
|
76 |
output = io.BytesIO()
|
77 |
-
png.Writer(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
78 |
example = tf.train.Example()
|
79 |
features = example.features.feature
|
80 |
-
features['image/encoded'].bytes_list.value.append(
|
81 |
features['image/format'].bytes_list.value.append(b'png')
|
82 |
return example
|
83 |
|
84 |
def generate_image_embedding(img_np, elixrc_infer, qformer_infer):
|
85 |
-
|
|
|
|
|
|
|
86 |
try:
|
|
|
87 |
serialized_img_tf_example = png_to_tfexample(img_np).SerializeToString()
|
88 |
elixrc_output = elixrc_infer(input_example=tf.constant([serialized_img_tf_example]))
|
89 |
elixrc_embedding = elixrc_output['feature_maps_0'].numpy()
|
|
|
|
|
|
|
90 |
qformer_input_img = {
|
91 |
'image_feature': elixrc_embedding.tolist(),
|
92 |
-
'ids': np.zeros((1, 1, 128), dtype=np.int32).tolist(),
|
93 |
-
'paddings': np.ones((1, 1, 128), dtype=np.float32).tolist(),
|
94 |
}
|
95 |
qformer_output_img = qformer_infer(**qformer_input_img)
|
96 |
image_embedding = qformer_output_img['all_contrastive_img_emb'].numpy()
|
|
|
|
|
97 |
if image_embedding.ndim > 2:
|
98 |
-
|
99 |
-
|
100 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
101 |
return image_embedding
|
|
|
102 |
except Exception as e:
|
103 |
-
print(f"Error generando embedding imagen: {e}")
|
|
|
|
|
104 |
|
105 |
def calculate_similarities_and_classify(image_embedding, bert_preprocessor, qformer_infer):
|
106 |
-
|
|
|
107 |
if bert_preprocessor is None: raise ValueError("Preprocesador BERT es None.")
|
108 |
if qformer_infer is None: raise ValueError("QFormer es None.")
|
|
|
109 |
detailed_results = {}
|
110 |
-
print("\n--- Calculando similitudes ---")
|
|
|
111 |
for i in range(len(criteria_list_positive)):
|
112 |
-
positive_text
|
113 |
-
|
114 |
-
|
|
|
|
|
115 |
similarity_positive, similarity_negative, difference = None, None, None
|
116 |
classification_comp, classification_simp = "ERROR", "ERROR"
|
|
|
117 |
try:
|
|
|
118 |
tokens_pos, paddings_pos = bert_tokenize(positive_text, bert_preprocessor)
|
119 |
-
|
120 |
-
|
|
|
|
|
|
|
121 |
if text_embedding_pos.ndim == 1: text_embedding_pos = np.expand_dims(text_embedding_pos, axis=0)
|
122 |
|
|
|
123 |
tokens_neg, paddings_neg = bert_tokenize(negative_text, bert_preprocessor)
|
124 |
-
|
125 |
-
|
|
|
|
|
|
|
126 |
if text_embedding_neg.ndim == 1: text_embedding_neg = np.expand_dims(text_embedding_neg, axis=0)
|
127 |
|
128 |
-
|
129 |
-
if image_embedding.shape[1] !=
|
|
|
|
|
|
|
130 |
|
|
|
131 |
similarity_positive = cosine_similarity(image_embedding, text_embedding_pos)[0][0]
|
132 |
similarity_negative = cosine_similarity(image_embedding, text_embedding_neg)[0][0]
|
|
|
133 |
|
|
|
134 |
difference = similarity_positive - similarity_negative
|
135 |
classification_comp = "PASS" if difference > SIMILARITY_DIFFERENCE_THRESHOLD else "FAIL"
|
136 |
classification_simp = "PASS" if similarity_positive > POSITIVE_SIMILARITY_THRESHOLD else "FAIL"
|
137 |
-
print(f"
|
|
|
138 |
except Exception as e:
|
139 |
-
print(f" ERROR criterio '{criterion_name}': {e}")
|
|
|
|
|
|
|
|
|
140 |
detailed_results[criterion_name] = {
|
141 |
-
'positive_prompt': positive_text,
|
|
|
142 |
'similarity_positive': float(similarity_positive) if similarity_positive is not None else None,
|
143 |
'similarity_negative': float(similarity_negative) if similarity_negative is not None else None,
|
144 |
'difference': float(difference) if difference is not None else None,
|
145 |
-
'classification_comparative': classification_comp,
|
|
|
146 |
}
|
147 |
return detailed_results
|
148 |
|
149 |
# --- Carga Global de Modelos ---
|
|
|
150 |
print("--- Iniciando carga global de modelos ---")
|
151 |
start_time = time.time()
|
152 |
models_loaded = False
|
153 |
bert_preprocessor_global = None
|
154 |
elixrc_infer_global = None
|
155 |
qformer_infer_global = None
|
|
|
156 |
try:
|
157 |
-
|
158 |
-
if
|
|
|
|
|
|
|
159 |
|
|
|
160 |
os.makedirs(MODEL_DOWNLOAD_DIR, exist_ok=True)
|
161 |
print(f"Descargando/verificando modelos en: {MODEL_DOWNLOAD_DIR}")
|
162 |
snapshot_download(repo_id=MODEL_REPO_ID, local_dir=MODEL_DOWNLOAD_DIR,
|
163 |
allow_patterns=['elixr-c-v2-pooled/*', 'pax-elixr-b-text/*'],
|
164 |
-
local_dir_use_symlinks=False
|
165 |
print("Modelos descargados/verificados.")
|
166 |
|
|
|
167 |
print("Cargando Preprocesador BERT...")
|
|
|
168 |
bert_preprocess_handle = "https://tfhub.dev/tensorflow/bert_en_uncased_preprocess/3"
|
169 |
bert_preprocessor_global = tf_hub.KerasLayer(bert_preprocess_handle)
|
170 |
print("Preprocesador BERT cargado.")
|
171 |
|
|
|
172 |
print("Cargando ELIXR-C...")
|
173 |
elixrc_model_path = os.path.join(MODEL_DOWNLOAD_DIR, 'elixr-c-v2-pooled')
|
174 |
elixrc_model = tf.saved_model.load(elixrc_model_path)
|
175 |
elixrc_infer_global = elixrc_model.signatures['serving_default']
|
176 |
print("Modelo ELIXR-C cargado.")
|
177 |
|
|
|
178 |
print("Cargando QFormer (ELIXR-B Text)...")
|
179 |
qformer_model_path = os.path.join(MODEL_DOWNLOAD_DIR, 'pax-elixr-b-text')
|
180 |
qformer_model = tf.saved_model.load(qformer_model_path)
|
@@ -184,167 +289,155 @@ try:
|
|
184 |
models_loaded = True
|
185 |
end_time = time.time()
|
186 |
print(f"--- Modelos cargados globalmente con éxito en {end_time - start_time:.2f} segundos ---")
|
|
|
187 |
except Exception as e:
|
188 |
models_loaded = False
|
189 |
-
print(f"--- ERROR CRÍTICO DURANTE LA CARGA GLOBAL DE MODELOS ---")
|
|
|
|
|
|
|
190 |
|
191 |
# --- Función Principal de Procesamiento para Gradio ---
|
192 |
-
def
|
193 |
-
"""
|
194 |
if not models_loaded:
|
195 |
raise gr.Error("Error: Los modelos no se pudieron cargar. La aplicación no puede procesar imágenes.")
|
196 |
if image_pil is None:
|
197 |
-
|
198 |
-
|
199 |
-
gr.update(visible=False), # Oculta resultados
|
200 |
-
None, # Borra imagen de salida
|
201 |
-
gr.update(value="N/A"), # Borra etiqueta
|
202 |
-
pd.DataFrame(), # Borra dataframe
|
203 |
-
None # Borra JSON
|
204 |
-
)
|
205 |
|
206 |
print("\n--- Iniciando evaluación para nueva imagen ---")
|
207 |
start_process_time = time.time()
|
|
|
208 |
try:
|
|
|
|
|
209 |
img_np = np.array(image_pil.convert('L'))
|
|
|
|
|
|
|
|
|
210 |
image_embedding = generate_image_embedding(img_np, elixrc_infer_global, qformer_infer_global)
|
|
|
|
|
|
|
|
|
211 |
detailed_results = calculate_similarities_and_classify(image_embedding, bert_preprocessor_global, qformer_infer_global)
|
212 |
-
|
|
|
|
|
|
|
|
|
|
|
213 |
for criterion, details in detailed_results.items():
|
214 |
total_count += 1
|
215 |
-
|
216 |
-
|
217 |
-
|
218 |
-
|
219 |
-
|
220 |
-
output_data.append([
|
221 |
-
|
222 |
-
|
223 |
-
|
224 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
225 |
if total_count > 0:
|
226 |
pass_rate = passed_count / total_count
|
227 |
if pass_rate >= 0.85: overall_quality = "Excellent"
|
228 |
elif pass_rate >= 0.70: overall_quality = "Good"
|
229 |
elif pass_rate >= 0.50: overall_quality = "Fair"
|
230 |
else: overall_quality = "Poor"
|
231 |
-
quality_label = f"{overall_quality} ({passed_count}/{total_count} passed)"
|
|
|
232 |
end_process_time = time.time()
|
233 |
-
print(f"--- Evaluación completada en {end_process_time - start_process_time:.2f}
|
234 |
-
|
235 |
-
|
236 |
-
|
237 |
-
|
238 |
-
gr.update(value=quality_label), # Actualiza etiqueta
|
239 |
-
df_results, # Actualiza dataframe
|
240 |
-
detailed_results # Actualiza JSON
|
241 |
-
)
|
242 |
except Exception as e:
|
243 |
-
print(f"Error durante procesamiento Gradio: {e}")
|
244 |
-
|
245 |
-
|
246 |
-
|
247 |
-
def reset_ui():
|
248 |
-
print("Reseteando UI...")
|
249 |
-
return (
|
250 |
-
gr.update(visible=True), # Muestra bienvenida
|
251 |
-
gr.update(visible=False), # Oculta resultados
|
252 |
-
None, # Borra imagen de entrada
|
253 |
-
None, # Borra imagen de salida
|
254 |
-
gr.update(value="N/A"), # Borra etiqueta
|
255 |
-
pd.DataFrame(), # Borra dataframe
|
256 |
-
None # Borra JSON
|
257 |
-
)
|
258 |
|
259 |
-
|
260 |
-
|
261 |
-
|
262 |
-
|
263 |
-
|
264 |
-
font
|
265 |
-
|
266 |
-
|
267 |
-
|
268 |
-
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
|
275 |
-
|
276 |
-
block_title_text_color="#ffffff",
|
277 |
-
# Bordes
|
278 |
-
border_color_accent="#374151",
|
279 |
-
border_color_primary="#4b5563",
|
280 |
-
# Botones y Elementos Interactivos
|
281 |
-
button_primary_background_fill="*primary_600",
|
282 |
-
button_primary_text_color="#ffffff",
|
283 |
-
button_secondary_background_fill="*neutral_700",
|
284 |
-
button_secondary_text_color="#ffffff",
|
285 |
-
input_background_fill="#374151",
|
286 |
-
input_border_color="#4b5563",
|
287 |
-
# Sombras y Radios
|
288 |
-
shadow_drop="rgba(0,0,0,0.2) 0px 2px 4px",
|
289 |
-
block_shadow="rgba(0,0,0,0.2) 0px 2px 5px",
|
290 |
-
)
|
291 |
-
|
292 |
-
# --- Definir la Interfaz Gradio con Bloques y Tema ---
|
293 |
-
with gr.Blocks(theme=dark_theme, title="CXR Quality Assessment") as demo:
|
294 |
-
# --- Cabecera ---
|
295 |
with gr.Row():
|
296 |
-
gr.
|
297 |
-
"""
|
298 |
-
|
299 |
-
|
300 |
-
|
301 |
-
|
302 |
-
|
303 |
-
|
304 |
-
|
305 |
-
with gr.Row(equal_height=False):
|
306 |
-
|
307 |
-
# --- Columna Izquierda (Carga) ---
|
308 |
-
with gr.Column(scale=1, min_width=350):
|
309 |
-
gr.Markdown("### 1. Upload Image", elem_id="upload-title")
|
310 |
-
input_image = gr.Image(type="pil", label="Upload Chest X-ray", height=300)
|
311 |
-
with gr.Row():
|
312 |
-
analyze_btn = gr.Button("Analyze Image", variant="primary", scale=2)
|
313 |
-
reset_btn = gr.Button("Reset", variant="secondary", scale=1)
|
314 |
-
# gr.Examples( examples=[os.path.join("examples", "sample_cxr.png")], inputs=input_image, label="Example CXR" )
|
315 |
-
gr.Markdown( "<p style='color:#9ca3af; font-size:0.9em;'>Model loading on startup takes ~1 min. Analysis takes ~15-40 sec.</p>" )
|
316 |
-
|
317 |
-
# --- Columna Derecha (Bienvenida / Resultados) ---
|
318 |
with gr.Column(scale=2):
|
319 |
-
|
320 |
-
|
321 |
-
|
322 |
-
|
323 |
-
|
324 |
-
|
325 |
-
|
326 |
-
|
327 |
-
|
328 |
-
|
329 |
-
|
330 |
-
|
331 |
-
|
332 |
-
|
333 |
-
|
334 |
-
|
335 |
-
|
336 |
-
|
337 |
-
|
338 |
-
|
339 |
-
inputs=[input_image],
|
340 |
-
outputs=[ welcome_block, results_block, output_image, output_label, output_dataframe, output_json ]
|
341 |
)
|
342 |
-
|
343 |
-
|
344 |
-
|
345 |
-
|
|
|
|
|
346 |
)
|
347 |
|
348 |
# --- Iniciar la Aplicación Gradio ---
|
|
|
|
|
|
|
|
|
349 |
if __name__ == "__main__":
|
|
|
|
|
|
|
350 |
demo.launch(server_name="0.0.0.0", server_port=7860)
|
|
|
16 |
# --- Configuración ---
|
17 |
MODEL_REPO_ID = "google/cxr-foundation"
|
18 |
MODEL_DOWNLOAD_DIR = './hf_cxr_foundation_space' # Directorio dentro del contenedor del Space
|
19 |
+
# Umbrales
|
20 |
SIMILARITY_DIFFERENCE_THRESHOLD = 0.1
|
21 |
POSITIVE_SIMILARITY_THRESHOLD = 0.1
|
22 |
print(f"Usando umbrales: Comp Δ={SIMILARITY_DIFFERENCE_THRESHOLD}, Simp τ={POSITIVE_SIMILARITY_THRESHOLD}")
|
|
|
31 |
"cropped image", "scapulae overlying lungs", "blurred image", "obscuring artifact"
|
32 |
]
|
33 |
|
34 |
+
# --- Funciones Auxiliares (Integradas o adaptadas) ---
|
35 |
+
# @tf.function(input_signature=[tf.TensorSpec(shape=[None], dtype=tf.string)]) # Puede ayudar rendimiento
|
36 |
+
def preprocess_text(text):
|
37 |
+
"""Función interna del preprocesador BERT."""
|
38 |
+
return bert_preprocessor_global(text)
|
39 |
+
|
40 |
def bert_tokenize(text, preprocessor):
|
41 |
+
"""Tokeniza texto usando el preprocesador BERT cargado globalmente."""
|
42 |
+
if preprocessor is None:
|
43 |
+
raise ValueError("BERT preprocessor no está cargado.")
|
44 |
if not isinstance(text, str): text = str(text)
|
45 |
+
|
46 |
+
# Ejecutar el preprocesador
|
47 |
out = preprocessor(tf.constant([text.lower()]))
|
48 |
+
|
49 |
+
# Extraer y procesar IDs y máscaras
|
50 |
ids = out['input_word_ids'].numpy().astype(np.int32)
|
51 |
masks = out['input_mask'].numpy().astype(np.float32)
|
52 |
paddings = 1.0 - masks
|
53 |
+
|
54 |
+
# Reemplazar token [SEP] (102) por 0 y marcar como padding
|
55 |
end_token_idx = (ids == 102)
|
56 |
ids[end_token_idx] = 0
|
57 |
paddings[end_token_idx] = 1.0
|
58 |
+
|
59 |
+
# Asegurar las dimensiones (B, T, S) -> (1, 1, 128)
|
60 |
+
# El preprocesador puede devolver (1, 128), necesitamos (1, 1, 128)
|
61 |
if ids.ndim == 2: ids = np.expand_dims(ids, axis=1)
|
62 |
if paddings.ndim == 2: paddings = np.expand_dims(paddings, axis=1)
|
63 |
+
|
64 |
+
# Verificar formas finales
|
65 |
expected_shape = (1, 1, 128)
|
66 |
if ids.shape != expected_shape:
|
67 |
+
# Intentar reajustar si es necesario (puede pasar con algunas versiones)
|
68 |
if ids.shape == (1,128): ids = np.expand_dims(ids, axis=1)
|
69 |
else: raise ValueError(f"Shape incorrecta para ids: {ids.shape}, esperado {expected_shape}")
|
70 |
if paddings.shape != expected_shape:
|
71 |
if paddings.shape == (1,128): paddings = np.expand_dims(paddings, axis=1)
|
72 |
else: raise ValueError(f"Shape incorrecta para paddings: {paddings.shape}, esperado {expected_shape}")
|
73 |
+
|
74 |
return ids, paddings
|
75 |
|
76 |
def png_to_tfexample(image_array: np.ndarray) -> tf.train.Example:
|
77 |
+
"""Crea tf.train.Example desde NumPy array (escala de grises)."""
|
78 |
if image_array.ndim == 3 and image_array.shape[2] == 1:
|
79 |
+
image_array = np.squeeze(image_array, axis=2) # Asegurar 2D
|
80 |
elif image_array.ndim != 2:
|
81 |
+
raise ValueError(f'Array debe ser 2-D (escala de grises). Dimensiones actuales: {image_array.ndim}')
|
82 |
+
|
83 |
image = image_array.astype(np.float32)
|
84 |
+
min_val = image.min()
|
85 |
+
max_val = image.max()
|
86 |
+
|
87 |
+
# Evitar división por cero si la imagen es constante
|
88 |
if max_val <= min_val:
|
89 |
+
# Si es constante, tratar como uint8 si el rango original lo permitía,
|
90 |
+
# o simplemente ponerla a 0 si es float.
|
91 |
if image_array.dtype == np.uint8 or (min_val >= 0 and max_val <= 255):
|
92 |
+
pixel_array = image.astype(np.uint8)
|
93 |
+
bitdepth = 8
|
94 |
+
else: # Caso flotante constante o fuera de rango uint8
|
95 |
+
pixel_array = np.zeros_like(image, dtype=np.uint16)
|
96 |
+
bitdepth = 16
|
97 |
else:
|
98 |
+
image -= min_val # Mover mínimo a cero
|
99 |
current_max = max_val - min_val
|
100 |
+
# Escalar a 16-bit para mayor precisión si no era uint8 originalmente
|
101 |
if image_array.dtype != np.uint8:
|
102 |
image *= 65535 / current_max
|
103 |
+
pixel_array = image.astype(np.uint16)
|
104 |
+
bitdepth = 16
|
105 |
else:
|
106 |
+
# Si era uint8, mantener el rango y tipo
|
107 |
+
# La resta del min ya la dejó en [0, current_max]
|
108 |
+
# Escalar a 255 si es necesario
|
109 |
image *= 255 / current_max
|
110 |
+
pixel_array = image.astype(np.uint8)
|
111 |
+
bitdepth = 8
|
112 |
+
|
113 |
+
# Codificar como PNG
|
114 |
output = io.BytesIO()
|
115 |
+
png.Writer(
|
116 |
+
width=pixel_array.shape[1],
|
117 |
+
height=pixel_array.shape[0],
|
118 |
+
greyscale=True,
|
119 |
+
bitdepth=bitdepth
|
120 |
+
).write(output, pixel_array.tolist())
|
121 |
+
png_bytes = output.getvalue()
|
122 |
+
|
123 |
+
# Crear tf.train.Example
|
124 |
example = tf.train.Example()
|
125 |
features = example.features.feature
|
126 |
+
features['image/encoded'].bytes_list.value.append(png_bytes)
|
127 |
features['image/format'].bytes_list.value.append(b'png')
|
128 |
return example
|
129 |
|
130 |
def generate_image_embedding(img_np, elixrc_infer, qformer_infer):
|
131 |
+
"""Genera embedding final de imagen."""
|
132 |
+
if elixrc_infer is None or qformer_infer is None:
|
133 |
+
raise ValueError("Modelos ELIXR-C o QFormer no cargados.")
|
134 |
+
|
135 |
try:
|
136 |
+
# 1. ELIXR-C
|
137 |
serialized_img_tf_example = png_to_tfexample(img_np).SerializeToString()
|
138 |
elixrc_output = elixrc_infer(input_example=tf.constant([serialized_img_tf_example]))
|
139 |
elixrc_embedding = elixrc_output['feature_maps_0'].numpy()
|
140 |
+
print(f" Embedding ELIXR-C shape: {elixrc_embedding.shape}")
|
141 |
+
|
142 |
+
# 2. QFormer (Imagen)
|
143 |
qformer_input_img = {
|
144 |
'image_feature': elixrc_embedding.tolist(),
|
145 |
+
'ids': np.zeros((1, 1, 128), dtype=np.int32).tolist(), # Texto vacío
|
146 |
+
'paddings': np.ones((1, 1, 128), dtype=np.float32).tolist(), # Todo padding
|
147 |
}
|
148 |
qformer_output_img = qformer_infer(**qformer_input_img)
|
149 |
image_embedding = qformer_output_img['all_contrastive_img_emb'].numpy()
|
150 |
+
|
151 |
+
# Ajustar dimensiones si es necesario
|
152 |
if image_embedding.ndim > 2:
|
153 |
+
print(f" Ajustando dimensiones embedding imagen (original: {image_embedding.shape})")
|
154 |
+
image_embedding = np.mean(
|
155 |
+
image_embedding,
|
156 |
+
axis=tuple(range(1, image_embedding.ndim - 1))
|
157 |
+
)
|
158 |
+
if image_embedding.ndim == 1:
|
159 |
+
image_embedding = np.expand_dims(image_embedding, axis=0)
|
160 |
+
elif image_embedding.ndim == 1:
|
161 |
+
image_embedding = np.expand_dims(image_embedding, axis=0) # Asegurar 2D
|
162 |
+
|
163 |
+
print(f" Embedding final imagen shape: {image_embedding.shape}")
|
164 |
+
if image_embedding.ndim != 2:
|
165 |
+
raise ValueError(f"Embedding final de imagen no tiene 2 dimensiones: {image_embedding.shape}")
|
166 |
return image_embedding
|
167 |
+
|
168 |
except Exception as e:
|
169 |
+
print(f"Error generando embedding de imagen: {e}")
|
170 |
+
traceback.print_exc()
|
171 |
+
raise # Re-lanzar la excepción para que Gradio la maneje
|
172 |
|
173 |
def calculate_similarities_and_classify(image_embedding, bert_preprocessor, qformer_infer):
|
174 |
+
"""Calcula similitudes y clasifica."""
|
175 |
+
if image_embedding is None: raise ValueError("Embedding de imagen es None.")
|
176 |
if bert_preprocessor is None: raise ValueError("Preprocesador BERT es None.")
|
177 |
if qformer_infer is None: raise ValueError("QFormer es None.")
|
178 |
+
|
179 |
detailed_results = {}
|
180 |
+
print("\n--- Calculando similitudes y clasificando ---")
|
181 |
+
|
182 |
for i in range(len(criteria_list_positive)):
|
183 |
+
positive_text = criteria_list_positive[i]
|
184 |
+
negative_text = criteria_list_negative[i]
|
185 |
+
criterion_name = positive_text # Usar prompt positivo como clave
|
186 |
+
|
187 |
+
print(f"Procesando criterio: \"{criterion_name}\"")
|
188 |
similarity_positive, similarity_negative, difference = None, None, None
|
189 |
classification_comp, classification_simp = "ERROR", "ERROR"
|
190 |
+
|
191 |
try:
|
192 |
+
# 1. Embedding Texto Positivo
|
193 |
tokens_pos, paddings_pos = bert_tokenize(positive_text, bert_preprocessor)
|
194 |
+
qformer_input_text_pos = {
|
195 |
+
'image_feature': np.zeros([1, 8, 8, 1376], dtype=np.float32).tolist(), # Dummy
|
196 |
+
'ids': tokens_pos.tolist(), 'paddings': paddings_pos.tolist(),
|
197 |
+
}
|
198 |
+
text_embedding_pos = qformer_infer(**qformer_input_text_pos)['contrastive_txt_emb'].numpy()
|
199 |
if text_embedding_pos.ndim == 1: text_embedding_pos = np.expand_dims(text_embedding_pos, axis=0)
|
200 |
|
201 |
+
# 2. Embedding Texto Negativo
|
202 |
tokens_neg, paddings_neg = bert_tokenize(negative_text, bert_preprocessor)
|
203 |
+
qformer_input_text_neg = {
|
204 |
+
'image_feature': np.zeros([1, 8, 8, 1376], dtype=np.float32).tolist(), # Dummy
|
205 |
+
'ids': tokens_neg.tolist(), 'paddings': paddings_neg.tolist(),
|
206 |
+
}
|
207 |
+
text_embedding_neg = qformer_infer(**qformer_input_text_neg)['contrastive_txt_emb'].numpy()
|
208 |
if text_embedding_neg.ndim == 1: text_embedding_neg = np.expand_dims(text_embedding_neg, axis=0)
|
209 |
|
210 |
+
# Verificar compatibilidad de dimensiones para similitud
|
211 |
+
if image_embedding.shape[1] != text_embedding_pos.shape[1]:
|
212 |
+
raise ValueError(f"Dimensión incompatible: Imagen ({image_embedding.shape[1]}) vs Texto Pos ({text_embedding_pos.shape[1]})")
|
213 |
+
if image_embedding.shape[1] != text_embedding_neg.shape[1]:
|
214 |
+
raise ValueError(f"Dimensión incompatible: Imagen ({image_embedding.shape[1]}) vs Texto Neg ({text_embedding_neg.shape[1]})")
|
215 |
|
216 |
+
# 3. Calcular Similitudes
|
217 |
similarity_positive = cosine_similarity(image_embedding, text_embedding_pos)[0][0]
|
218 |
similarity_negative = cosine_similarity(image_embedding, text_embedding_neg)[0][0]
|
219 |
+
print(f" Sim (+)={similarity_positive:.4f}, Sim (-)={similarity_negative:.4f}")
|
220 |
|
221 |
+
# 4. Clasificar
|
222 |
difference = similarity_positive - similarity_negative
|
223 |
classification_comp = "PASS" if difference > SIMILARITY_DIFFERENCE_THRESHOLD else "FAIL"
|
224 |
classification_simp = "PASS" if similarity_positive > POSITIVE_SIMILARITY_THRESHOLD else "FAIL"
|
225 |
+
print(f" Diff={difference:.4f} -> Comp: {classification_comp}, Simp: {classification_simp}")
|
226 |
+
|
227 |
except Exception as e:
|
228 |
+
print(f" ERROR procesando criterio '{criterion_name}': {e}")
|
229 |
+
traceback.print_exc()
|
230 |
+
# Mantener clasificaciones como "ERROR"
|
231 |
+
|
232 |
+
# Guardar resultados
|
233 |
detailed_results[criterion_name] = {
|
234 |
+
'positive_prompt': positive_text,
|
235 |
+
'negative_prompt': negative_text,
|
236 |
'similarity_positive': float(similarity_positive) if similarity_positive is not None else None,
|
237 |
'similarity_negative': float(similarity_negative) if similarity_negative is not None else None,
|
238 |
'difference': float(difference) if difference is not None else None,
|
239 |
+
'classification_comparative': classification_comp,
|
240 |
+
'classification_simplified': classification_simp
|
241 |
}
|
242 |
return detailed_results
|
243 |
|
244 |
# --- Carga Global de Modelos ---
|
245 |
+
# Se ejecuta UNA VEZ al iniciar la aplicación Gradio/Space
|
246 |
print("--- Iniciando carga global de modelos ---")
|
247 |
start_time = time.time()
|
248 |
models_loaded = False
|
249 |
bert_preprocessor_global = None
|
250 |
elixrc_infer_global = None
|
251 |
qformer_infer_global = None
|
252 |
+
|
253 |
try:
|
254 |
+
# Verificar autenticación HF (útil si se usan modelos privados, aunque no es el caso aquí)
|
255 |
+
# if HfFolder.get_token() is None:
|
256 |
+
# print("Advertencia: No se encontró token de Hugging Face.")
|
257 |
+
# else:
|
258 |
+
# print("Token de Hugging Face encontrado.")
|
259 |
|
260 |
+
# Crear directorio si no existe
|
261 |
os.makedirs(MODEL_DOWNLOAD_DIR, exist_ok=True)
|
262 |
print(f"Descargando/verificando modelos en: {MODEL_DOWNLOAD_DIR}")
|
263 |
snapshot_download(repo_id=MODEL_REPO_ID, local_dir=MODEL_DOWNLOAD_DIR,
|
264 |
allow_patterns=['elixr-c-v2-pooled/*', 'pax-elixr-b-text/*'],
|
265 |
+
local_dir_use_symlinks=False) # Evitar symlinks
|
266 |
print("Modelos descargados/verificados.")
|
267 |
|
268 |
+
# Cargar Preprocesador BERT desde TF Hub
|
269 |
print("Cargando Preprocesador BERT...")
|
270 |
+
# Usar handle explícito puede ser más robusto en algunos entornos
|
271 |
bert_preprocess_handle = "https://tfhub.dev/tensorflow/bert_en_uncased_preprocess/3"
|
272 |
bert_preprocessor_global = tf_hub.KerasLayer(bert_preprocess_handle)
|
273 |
print("Preprocesador BERT cargado.")
|
274 |
|
275 |
+
# Cargar ELIXR-C
|
276 |
print("Cargando ELIXR-C...")
|
277 |
elixrc_model_path = os.path.join(MODEL_DOWNLOAD_DIR, 'elixr-c-v2-pooled')
|
278 |
elixrc_model = tf.saved_model.load(elixrc_model_path)
|
279 |
elixrc_infer_global = elixrc_model.signatures['serving_default']
|
280 |
print("Modelo ELIXR-C cargado.")
|
281 |
|
282 |
+
# Cargar QFormer (ELIXR-B Text)
|
283 |
print("Cargando QFormer (ELIXR-B Text)...")
|
284 |
qformer_model_path = os.path.join(MODEL_DOWNLOAD_DIR, 'pax-elixr-b-text')
|
285 |
qformer_model = tf.saved_model.load(qformer_model_path)
|
|
|
289 |
models_loaded = True
|
290 |
end_time = time.time()
|
291 |
print(f"--- Modelos cargados globalmente con éxito en {end_time - start_time:.2f} segundos ---")
|
292 |
+
|
293 |
except Exception as e:
|
294 |
models_loaded = False
|
295 |
+
print(f"--- ERROR CRÍTICO DURANTE LA CARGA GLOBAL DE MODELOS ---")
|
296 |
+
print(e)
|
297 |
+
traceback.print_exc()
|
298 |
+
# Gradio se iniciará, pero la función de análisis fallará.
|
299 |
|
300 |
# --- Función Principal de Procesamiento para Gradio ---
|
301 |
+
def assess_quality(image_pil):
|
302 |
+
"""Función que Gradio llamará con la imagen de entrada."""
|
303 |
if not models_loaded:
|
304 |
raise gr.Error("Error: Los modelos no se pudieron cargar. La aplicación no puede procesar imágenes.")
|
305 |
if image_pil is None:
|
306 |
+
# Devolver resultados vacíos o un mensaje de error si no hay imagen
|
307 |
+
return pd.DataFrame(), "N/A", None # Dataframe vacío, Label vacío, JSON vacío
|
|
|
|
|
|
|
|
|
|
|
|
|
308 |
|
309 |
print("\n--- Iniciando evaluación para nueva imagen ---")
|
310 |
start_process_time = time.time()
|
311 |
+
|
312 |
try:
|
313 |
+
# 1. Convertir PIL Image a NumPy array (escala de grises)
|
314 |
+
# Gradio con type="pil" ya la entrega como objeto PIL
|
315 |
img_np = np.array(image_pil.convert('L'))
|
316 |
+
print(f"Imagen convertida a NumPy. Shape: {img_np.shape}, Tipo: {img_np.dtype}")
|
317 |
+
|
318 |
+
# 2. Generar Embedding de Imagen
|
319 |
+
print("Generando embedding de imagen...")
|
320 |
image_embedding = generate_image_embedding(img_np, elixrc_infer_global, qformer_infer_global)
|
321 |
+
print("Embedding de imagen generado.")
|
322 |
+
|
323 |
+
# 3. Calcular Similitudes y Clasificar
|
324 |
+
print("Calculando similitudes y clasificando criterios...")
|
325 |
detailed_results = calculate_similarities_and_classify(image_embedding, bert_preprocessor_global, qformer_infer_global)
|
326 |
+
print("Clasificación completada.")
|
327 |
+
|
328 |
+
# 4. Formatear Resultados para Gradio
|
329 |
+
output_data = []
|
330 |
+
passed_count = 0
|
331 |
+
total_count = 0
|
332 |
for criterion, details in detailed_results.items():
|
333 |
total_count += 1
|
334 |
+
sim_pos_str = f"{details['similarity_positive']:.4f}" if details['similarity_positive'] is not None else "N/A"
|
335 |
+
sim_neg_str = f"{details['similarity_negative']:.4f}" if details['similarity_negative'] is not None else "N/A"
|
336 |
+
diff_str = f"{details['difference']:.4f}" if details['difference'] is not None else "N/A"
|
337 |
+
assessment_comp = details['classification_comparative']
|
338 |
+
assessment_simp = details['classification_simplified']
|
339 |
+
output_data.append([
|
340 |
+
criterion,
|
341 |
+
sim_pos_str,
|
342 |
+
sim_neg_str,
|
343 |
+
diff_str,
|
344 |
+
assessment_comp,
|
345 |
+
assessment_simp
|
346 |
+
])
|
347 |
+
if assessment_comp == "PASS":
|
348 |
+
passed_count += 1
|
349 |
+
|
350 |
+
# Crear DataFrame
|
351 |
+
df_results = pd.DataFrame(output_data, columns=[
|
352 |
+
"Criterion", "Sim (+)", "Sim (-)", "Difference", "Assessment (Comp)", "Assessment (Simp)"
|
353 |
+
])
|
354 |
+
|
355 |
+
# Calcular etiqueta de calidad general
|
356 |
+
overall_quality = "Error"
|
357 |
if total_count > 0:
|
358 |
pass_rate = passed_count / total_count
|
359 |
if pass_rate >= 0.85: overall_quality = "Excellent"
|
360 |
elif pass_rate >= 0.70: overall_quality = "Good"
|
361 |
elif pass_rate >= 0.50: overall_quality = "Fair"
|
362 |
else: overall_quality = "Poor"
|
363 |
+
quality_label = f"{overall_quality} ({passed_count}/{total_count} criteria passed)"
|
364 |
+
|
365 |
end_process_time = time.time()
|
366 |
+
print(f"--- Evaluación completada en {end_process_time - start_process_time:.2f} segundos ---")
|
367 |
+
|
368 |
+
# Devolver DataFrame, Etiqueta y JSON
|
369 |
+
return df_results, quality_label, detailed_results
|
370 |
+
|
|
|
|
|
|
|
|
|
371 |
except Exception as e:
|
372 |
+
print(f"Error durante el procesamiento de la imagen en Gradio: {e}")
|
373 |
+
traceback.print_exc()
|
374 |
+
# Lanzar un gr.Error para mostrarlo en la UI de Gradio
|
375 |
+
raise gr.Error(f"Error procesando la imagen: {str(e)}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
376 |
|
377 |
+
|
378 |
+
# --- Definir la Interfaz Gradio ---
|
379 |
+
css = """
|
380 |
+
#quality-label label {
|
381 |
+
font-size: 1.1em;
|
382 |
+
font-weight: bold;
|
383 |
+
}
|
384 |
+
"""
|
385 |
+
with gr.Blocks(css=css) as demo:
|
386 |
+
gr.Markdown(
|
387 |
+
"""
|
388 |
+
# Chest X-ray Technical Quality Assessment
|
389 |
+
Upload a chest X-ray image (PNG, JPG, etc.) to evaluate its technical quality based on 7 standard criteria
|
390 |
+
using the ELIXR model family (comparative strategy: Positive vs Negative prompts).
|
391 |
+
**Note:** Model loading on startup might take a minute. Processing an image can take 10-30 seconds depending on server load.
|
392 |
+
"""
|
393 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
394 |
with gr.Row():
|
395 |
+
with gr.Column(scale=1):
|
396 |
+
input_image = gr.Image(type="pil", label="Upload Chest X-ray")
|
397 |
+
submit_button = gr.Button("Assess Quality", variant="primary")
|
398 |
+
# Añadir ejemplos si tienes imágenes de ejemplo
|
399 |
+
# Asegúrate de que la carpeta 'examples' exista y contenga las imágenes
|
400 |
+
# gr.Examples(
|
401 |
+
# examples=[os.path.join("examples", "sample_cxr.png")], # Lista de rutas a ejemplos
|
402 |
+
# inputs=input_image
|
403 |
+
# )
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
404 |
with gr.Column(scale=2):
|
405 |
+
output_label = gr.Label(label="Overall Quality Estimate", elem_id="quality-label")
|
406 |
+
output_dataframe = gr.DataFrame(
|
407 |
+
headers=["Criterion", "Sim (+)", "Sim (-)", "Difference", "Assessment (Comp)", "Assessment (Simp)"],
|
408 |
+
label="Detailed Quality Assessment",
|
409 |
+
wrap=True,
|
410 |
+
height=350
|
411 |
+
)
|
412 |
+
output_json = gr.JSON(label="Raw Results (for debugging)")
|
413 |
+
|
414 |
+
|
415 |
+
gr.Markdown(
|
416 |
+
f"""
|
417 |
+
**Explanation:**
|
418 |
+
* **Criterion:** The quality aspect being evaluated (using the positive prompt text).
|
419 |
+
* **Sim (+):** Cosine similarity between the image and the *positive* text prompt (e.g., "optimal centering"). Higher is better.
|
420 |
+
* **Sim (-):** Cosine similarity between the image and the *negative* text prompt (e.g., "poorly centered"). Lower is better.
|
421 |
+
* **Difference:** Sim (+) - Sim (-). A large positive difference indicates the image is much closer to the positive description.
|
422 |
+
* **Assessment (Comp):** PASS if Difference > {SIMILARITY_DIFFERENCE_THRESHOLD}, otherwise FAIL. This is the main comparative assessment.
|
423 |
+
* **Assessment (Simp):** PASS if Sim (+) > {POSITIVE_SIMILARITY_THRESHOLD}, otherwise FAIL. A simpler check based only on positive similarity.
|
424 |
+
"""
|
|
|
|
|
425 |
)
|
426 |
+
|
427 |
+
# Conectar el botón a la función de procesamiento
|
428 |
+
submit_button.click(
|
429 |
+
fn=assess_quality,
|
430 |
+
inputs=input_image,
|
431 |
+
outputs=[output_dataframe, output_label, output_json]
|
432 |
)
|
433 |
|
434 |
# --- Iniciar la Aplicación Gradio ---
|
435 |
+
# Al desplegar en Spaces, Gradio se encarga de esto automáticamente.
|
436 |
+
# Para ejecutar localmente: demo.launch()
|
437 |
+
# Para Spaces, es mejor dejar que HF maneje el launch.
|
438 |
+
# demo.launch(share=True) # Para obtener un link público temporal si corres localmente
|
439 |
if __name__ == "__main__":
|
440 |
+
# share=True solo si quieres un enlace público temporal desde local
|
441 |
+
# server_name="0.0.0.0" para permitir conexiones de red local
|
442 |
+
# server_port=7860 es el puerto estándar de HF Spaces
|
443 |
demo.launch(server_name="0.0.0.0", server_port=7860)
|