我尝试实施MAML。因此,我需要一步一步地训练模型的副本(model_copy), 那么我就需要对我的meta_model进行培训,同时丢失了我的model_copy。
我想训练一个函数中的model_copy。 如果我将代码复制到函数中,则不会得到正确的gradients_meta(它们全都不是)。
似乎图形未连接-如何连接图形?
任何关于我做错事情的想法吗?我看到了很多变量,但这似乎没有什么区别。
以下是重现此问题的代码:
import tensorflow as tf
import tensorflow.keras as keras
import tensorflow.keras.backend as keras_backend
def copy_model(model):
copied_model = keras.Sequential()
copied_model.add(keras.layers.Dense(5, input_shape=(1,)))
copied_model.add(keras.layers.Dense(1))
copied_model.set_weights(model.get_weights())
return copied_model
def compute_loss(model, x, y):
logits = model(x) # prediction of my model
mse = keras_backend.mean(keras.losses.mean_squared_error(y, logits)) # compute loss between prediciton and label/truth
return mse, logits
# meta_model to learn in outer gradient tape
meta_model = keras.Sequential()
meta_model.add(keras.layers.Dense(5, input_shape=(1,)))
meta_model.add(keras.layers.Dense(1))
# optimizer for training
optimizer = keras.optimizers.Adam()
# function to calculate model_copys params
def do_calc(x, y, meta_model):
with tf.GradientTape() as gg:
model_copy = copy_model(meta_model)
gg.watch(x)
gg.watch(meta_model.trainable_variables)
gg.watch(model_copy.trainable_variables)
loss, _ = compute_loss(model_copy, x, y)
gradient = gg.gradient(loss, model_copy.trainable_variables)
optimizer.apply_gradients(zip(gradient, model_copy.trainable_variables))
return model_copy
# inputs for training
x = tf.constant(3.0, shape=(1, 1, 1))
y = tf.constant(3.0, shape=(1, 1, 1))
with tf.GradientTape() as g:
g.watch(x)
g.watch(y)
model_copy = do_calc(x, y, meta_model)
g.watch(model_copy.trainable_variables)
# calculate loss of model_copy
test_loss, _ = compute_loss(model_copy, x, y)
# build gradients for meta_model update
gradients_meta = g.gradient(test_loss, meta_model.trainable_variables)
# gradients always None !?!!11 elf
optimizer.apply_gradients(zip(gradients_meta, meta_model.trainable_variables))
在此先感谢您的帮助。
答案 0 :(得分:1)
我找到了解决方案: 我需要以某种方式“连接”元模型和模型复制。
任何人都可以解释为什么这样做有效,以及如何使用“适当的”优化程序来实现这一目标吗?
import tensorflow as tf
import tensorflow.keras as keras
import tensorflow.keras.backend as keras_backend
def copy_model(model):
copied_model = keras.Sequential()
copied_model.add(keras.layers.Dense(5, input_shape=(1,)))
copied_model.add(keras.layers.Dense(1))
copied_model.set_weights(model.get_weights())
return copied_model
def compute_loss(model, x, y):
logits = model(x) # prediction of my model
mse = keras_backend.mean(keras.losses.mean_squared_error(y, logits)) # compute loss between prediciton and label/truth
return mse, logits
# meta_model to learn in outer gradient tape
meta_model = keras.Sequential()
meta_model.add(keras.layers.Dense(5, input_shape=(1,)))
meta_model.add(keras.layers.Dense(1))
# optimizer for training
optimizer = keras.optimizers.Adam()
# function to calculate model_copys params
def do_calc(meta_model, x, y, gg, alpha=0.01):
model_copy = copy_model(meta_model)
loss, _ = compute_loss(model_copy, x, y)
gradients = gg.gradient(loss, model_copy.trainable_variables)
k = 0
for layer in range(len(model_copy.layers)):
# calculate adapted parameters w/ gradient descent
# \theta_i' = \theta - \alpha * gradients
model_copy.layers[layer].kernel = tf.subtract(meta_model.layers[layer].kernel,
tf.multiply(alpha, gradients[k]))
model_copy.layers[layer].bias = tf.subtract(meta_model.layers[layer].bias,
tf.multiply(alpha, gradients[k + 1]))
k += 2
return model_copy
with tf.GradientTape() as g:
# inputs for training
x = tf.constant(3.0, shape=(1, 1, 1))
y = tf.constant(3.0, shape=(1, 1, 1))
adapted_models = []
# model_copy = meta_model
with tf.GradientTape() as gg:
model_copy = do_calc(meta_model, x, y, gg)
# calculate loss of model_copy
test_loss, _ = compute_loss(model_copy, x, y)
# build gradients for meta_model update
gradients_meta = g.gradient(test_loss, meta_model.trainable_variables)
# gradients work. Why???
optimizer.apply_gradients(zip(gradients_meta, meta_model.trainable_variables))