Por qué GradientTape es la característica más subestimada en TensorFlow

Contenido Overview Diferenciación automática y gradientes Instalación Computadores Gradientes Gradientes de tapas Gradientes con respecto a un modelo Controlar lo que mira la cinta Resultados intermedios Notas sobre el rendimiento Gradientes de los objetivos no escalares Control de flujo Diferenciación automática y gradientes es útil para la implementación de algoritmos de aprendizaje automático como Formación en Redes Neurales. Diferenciación automática retropropaganda En esta guía, explorará maneras de calcular gradientes con TensorFlow, especialmente en la ejecución ansiosa. Instalación import numpy as np import matplotlib.pyplot as plt import tensorflow as tf 2024-08-15 01:30:07.003169: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:485] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered 2024-08-15 01:30:07.023862: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:8454] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered 2024-08-15 01:30:07.029954: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1452] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered Computadores Gradientes Para diferenciar automáticamente, TensorFlow necesita recordar qué operaciones ocurren en qué orden durante la Entonces, durante el , TensorFlow cruza esta lista de operaciones en orden inverso para calcular gradientes. adelante Paso trasero Gradientes de tapas TensorFlow proporciona la API para la diferenciación automática; es decir, calcular el gradiente de un cálculo con respecto a algunas entradas, generalmente s. TensorFlow "recoge" las operaciones relevantes ejecutadas dentro del contexto de una TensorFlow luego utiliza esa cinta para calcular los gradientes de una computación "recordada" utilizando . tf.GradientTape tf.Variable tf.GradientTape Modo de diferenciación inversa He aquí un ejemplo sencillo: x = tf.Variable(3.0) with tf.GradientTape() as tape: y = x**2 WARNING: All log messages before absl::InitializeLog() is called are written to STDERR I0000 00:00:1723685409.408818 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685409.412555 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685409.416343 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685409.420087 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685409.431667 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685409.435229 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685409.438777 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685409.442350 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685409.445712 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685409.449141 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685409.452491 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685409.456034 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.685265 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.687389 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.689411 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.691490 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.693542 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.695541 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.697441 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.699432 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.701351 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.703333 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.705229 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.707222 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.744994 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.747037 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.749507 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.751538 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.753500 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.755501 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.757421 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.759404 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.761363 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.763858 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.766199 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723685410.768560 20970 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 Una vez que haya registrado algunas operaciones, utilice para calcular el gradiente de alguna meta (a menudo una pérdida) en relación con alguna fuente (a menudo las variables del modelo): GradientTape.gradient(target, sources) # dy = 2x * dx dy_dx = tape.gradient(y, x) dy_dx.numpy() 6.0 El ejemplo anterior utiliza escaleras, pero Funciona tan fácilmente en cualquier tensor: tf.GradientTape w = tf.Variable(tf.random.normal((3, 2)), name='w') b = tf.Variable(tf.zeros(2, dtype=tf.float32), name='b') x = [[1., 2., 3.]] with tf.GradientTape(persistent=True) as tape: y = x @ w + b loss = tf.reduce_mean(y**2) Para obtener el gradiente de con respecto a ambas variables, puede pasar ambas como fuentes a la La cinta es flexible sobre cómo se pasan las fuentes y aceptará cualquier combinación de listas o diccionarios y devolverá el gradiente estructurado de la misma manera (ver ) de loss gradient tf.nest [dl_dw, dl_db] = tape.gradient(loss, [w, b]) El gradiente con respecto a cada fuente tiene la forma de la fuente: print(w.shape) print(dl_dw.shape) (3, 2) (3, 2) Aquí está el cálculo de gradientes de nuevo, esta vez pasando un diccionario de variables: my_vars = { 'w': w, 'b': b } grad = tape.gradient(loss, my_vars) grad['b'] Gradientes con respecto a un modelo Es común recoger En la A o una de sus subcategorías ( , de ) por y . tf.Variables tf.Module layers.Layer keras.Model Verificación Exportaciones En la mayoría de los casos, desea calcular gradientes con respecto a las variables entrenables de un modelo. agregar sus variables en el Propiedad, puede calcular estos gradientes en unas pocas líneas de código: tf.Module Module.trainable_variables layer = tf.keras.layers.Dense(2, activation='relu') x = tf.constant([[1., 2., 3.]]) with tf.GradientTape() as tape: # Forward pass y = layer(x) loss = tf.reduce_mean(y**2) # Calculate gradients with respect to every trainable variable grad = tape.gradient(loss, layer.trainable_variables) for var, g in zip(layer.trainable_variables, grad): print(f'{var.name}, shape: {g.shape}') kernel, shape: (3, 2) bias, shape: (2,) Controlar lo que mira la cinta El comportamiento predeterminado es registrar todas las operaciones después de acceder a un Las razones para esto son: tf.Variable La cinta necesita saber qué operaciones registrar en el paso adelante para calcular los gradientes en el paso atrás. La cinta contiene referencias a las salidas intermedias, por lo que no desea grabar operaciones innecesarias. El caso de uso más común consiste en calcular el gradiente de una pérdida con respecto a todas las variables entrenables de un modelo. Por ejemplo, el siguiente no calcula un gradiente porque el no está "visto" por defecto, y el No es entrenable: tf.Tensor tf.Variable # A trainable variable x0 = tf.Variable(3.0, name='x0') # Not trainable x1 = tf.Variable(3.0, name='x1', trainable=False) # Not a Variable: A variable + tensor returns a tensor. x2 = tf.Variable(2.0, name='x2') + 1.0 # Not a variable x3 = tf.constant(3.0, name='x3') with tf.GradientTape() as tape: y = (x0**2) + (x1**2) + (x2**2) grad = tape.gradient(y, [x0, x1, x2, x3]) for g in grad: print(g) tf.Tensor(6.0, shape=(), dtype=float32) None None None Puede listar las variables que está viendo la cinta usando el El método: GradientTape.watched_variables [var.name for var in tape.watched_variables()] ['x0:0'] Ofrece hooks que dan al usuario el control sobre lo que está o no está viendo. tf.GradientTape Para registrar gradientes con respecto a a Es necesario llamar : tf.Tensor GradientTape.watch(x) x = tf.constant(3.0) with tf.GradientTape() as tape: tape.watch(x) y = x**2 # dy = 2x * dx dy_dx = tape.gradient(y, x) print(dy_dx.numpy()) 6.0 Por el contrario, para desactivar el comportamiento predeterminado de ver todos los y set Este cálculo utiliza dos variables, pero sólo conecta el gradiente para una de las variables: tf.Variables watch_accessed_variables=False x0 = tf.Variable(0.0) x1 = tf.Variable(10.0) with tf.GradientTape(watch_accessed_variables=False) as tape: tape.watch(x1) y0 = tf.math.sin(x0) y1 = tf.nn.softplus(x1) y = y0 + y1 ys = tf.reduce_sum(y) Desde No se llamó a , no se calcula un gradiente con respecto a él: GradientTape.watch x0 # dys/dx1 = exp(x1) / (1 + exp(x1)) = sigmoid(x1) grad = tape.gradient(ys, {'x0': x0, 'x1': x1}) print('dy/dx0:', grad['x0']) print('dy/dx1:', grad['x1'].numpy()) dy/dx0: None dy/dx1: 0.9999546 Resultados intermedios También puede solicitar gradientes de la salida con respecto a los valores intermedios calculados dentro de la El contexto. tf.GradientTape x = tf.constant(3.0) with tf.GradientTape() as tape: tape.watch(x) y = x * x z = y * y # Use the tape to compute the gradient of z with respect to the # intermediate value y. # dz_dy = 2 * y and y = x ** 2 = 9 print(tape.gradient(z, y).numpy()) 18.0 Por defecto, los recursos mantenidos por un serán liberados tan pronto como Para calcular múltiples gradientes sobre el mismo cálculo, crea una cinta de gradientes con Esto permite múltiples llamadas a la método como recursos se liberan cuando el objeto de cinta es basura recogida. Por ejemplo: GradientTape GradientTape.gradient persistent=True gradient x = tf.constant([1, 3.0]) with tf.GradientTape(persistent=True) as tape: tape.watch(x) y = x * x z = y * y print(tape.gradient(z, x).numpy()) # [4.0, 108.0] (4 * x**3 at x = [1.0, 3.0]) print(tape.gradient(y, x).numpy()) # [2.0, 6.0] (2 * x at x = [1.0, 3.0]) [ 4. 108.] [2. 6.] del tape # Drop the reference to the tape Notas sobre el rendimiento There is a tiny overhead associated with doing operations inside a gradient tape context. For most eager execution this will not be a noticeable cost, but you should still use tape context around the areas only where it is required. Gradient tapes use memory to store intermediate results, including inputs and outputs, for use during the backwards pass. For efficiency, some ops (like ) don't need to keep their intermediate results and they are pruned during the forward pass. However, if you use on your tape, and your peak memory usage will be higher. ReLU persistent=True nothing is discarded Gradientes de los objetivos no escalares Un gradiente es fundamentalmente una operación en un escalar. x = tf.Variable(2.0) with tf.GradientTape(persistent=True) as tape: y0 = x**2 y1 = 1 / x print(tape.gradient(y0, x).numpy()) print(tape.gradient(y1, x).numpy()) 4.0 -0.25 Por lo tanto, si se solicita el gradiente de múltiples objetivos, el resultado para cada fuente es: El gradiente de la suma de los objetivos, o equivalente La suma de los gradientes de cada objetivo. x = tf.Variable(2.0) with tf.GradientTape() as tape: y0 = x**2 y1 = 1 / x print(tape.gradient({'y0': y0, 'y1': y1}, x).numpy()) 3.75 Del mismo modo, si el (los) objetivo(s) no son escalares, se calcula el gradiente de la suma: x = tf.Variable(2.) with tf.GradientTape() as tape: y = x * [3., 4.] print(tape.gradient(y, x).numpy()) 7.0 Esto hace que sea fácil tomar el gradiente de la suma de una colección de pérdidas, o el gradiente de la suma de un cálculo de pérdidas elemental. Si necesita un gradiente separado para cada elemento, consulte . Jacobianos Para un cálculo elemental, el gradiente de la suma da la derivada de cada elemento con respecto a su elemento de entrada, ya que cada elemento es independiente: x = tf.linspace(-10.0, 10.0, 200+1) with tf.GradientTape() as tape: tape.watch(x) y = tf.nn.sigmoid(x) dy_dx = tape.gradient(y, x) plt.plot(x, y, label='y') plt.plot(x, dy_dx, label='dy/dx') plt.legend() _ = plt.xlabel('x') Control de flujo Debido a que una cinta de gradientes registra operaciones a medida que se ejecutan, el flujo de control de Python se maneja naturalmente (por ejemplo, y las declaraciones). if while Aquí se utiliza una variable diferente en cada rama de un El gradiente sólo se conecta a la variable que se usó: if x = tf.constant(1.0) v0 = tf.Variable(2.0) v1 = tf.Variable(2.0) with tf.GradientTape(persistent=True) as tape: tape.watch(x) if x > 0.0: result = v0 else: result = v1**2 dv0, dv1 = tape.gradient(result, [v0, v1]) print(dv0) print(dv1) tf.Tensor(1.0, shape=(), dtype=float32) None Recuerde que las declaraciones de control no son diferenciables, por lo que son invisibles para los optimizadores basados en gradientes. Dependiendo del valor de En el ejemplo anterior, la cinta o o El gradiente con respecto a Siempre es . x result = v0 result = v1**2 x None dx = tape.gradient(result, x) print(dx) None Originalmente publicado en el sitio web de TensorFlow, este artículo aparece aquí bajo un nuevo título y está licenciado bajo CC BY 4.0. Originalmente publicado en el sitio web de TensorFlow, este artículo aparece aquí bajo un nuevo título y está licenciado bajo CC BY 4.0. TensorFlow TensorFlow