我想计算tf.extract_volume_patches的逆。使用梯度(如#6743和stackoverflow中所建议的(来计算tf.extract_volume_patches的倒数适用于静态已知的输入形状，但当与可变输入形状一起使用时，它确实会返回错误。
合成3D MNIST数据集的一个最小示例，(当input=(28,28,28,1)时有效，但我需要对可变大小的数据输入(大型生物医学图像(进行提取(：
import tensorflow as tf
import numpy as np
def gaussian(x, amp=1, mu=None, sig=None):
""" Gaussian function over d dimensions of x
"""
if mu is None:
mu = np.zeros_like(x)
if sig is None:
sig = np.ones_like(x)
return amp * np.exp(-np.sum(np.square(x - mu) / (2 * np.square(sig))))
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()
x_train = x_train.astype('float32') / 255.
x_test = x_test.astype('float32') / 255.
multiplier = np.zeros((28,), dtype=np.float32)
for i in range(4):
multiplier[13-i] = round(gaussian(i),2)
multiplier[14+i] = round(gaussian(i),2)
print(multiplier)
x_train = np.einsum('bhw,d->bdhw', x_train, multiplier)[...,np.newaxis]
class ExtractPatches(tf.keras.layers.Layer):
def __init__(self, ksizes, strides, shape):
super(ExtractPatches, self).__init__()
self.ksizes = ksizes
self.strides = strides
self.shape = shape
def call(self, inputs):
patches = tf.extract_volume_patches(inputs,
ksizes=self.ksizes,
strides=self.strides,
padding="VALID")
return tf.reshape(patches, self.shape), tf.shape(inputs)
class CombinePatches(tf.keras.layers.Layer):
def __init__(self, ksizes, strides):
super(CombinePatches, self).__init__()
self.ksizes = ksizes
self.strides = strides
def call(self, patches, inputs):
target_volume = tf.zeros_like(inputs)
target_patches = tf.extract_volume_patches(
target_volume,
ksizes=self.ksizes,
strides=self.strides,
padding="VALID"
)
# Creates list of gradient mappings from patches to target shape
# Patches without overlap get 1, elements that overlap receive 1
# times the number of overlaps.
target_grad_mapping = tf.gradients(target_patches, target_volume)[0]
# Computes gradients again and dividing by grad, otherwise its just summed.
return tf.gradients(target_patches, target_volume, patches)[0] / target_grad_mapping
def create_model():
inputs = tf.keras.layers.Input(shape=(None,None,None,1))
patches, shape = ExtractPatches(ksizes=[1,14,14,14,1], strides=[1,14,14,14,1], shape=(-1,14,14,14,1))(inputs)
encoded = tf.keras.layers.Conv3D(filters=28, kernel_size=(14,14,14), strides=(14,14,14))(patches)
decoded = tf.keras.layers.Conv3DTranspose(filters=1, kernel_size=(14,14,14), strides=(14,14,14))(encoded)
merged = CombinePatches(ksizes=[1,14,14,14,1], strides=[1,14,14,14,1])(decoded, inputs)
return tf.keras.models.Model(inputs=inputs, outputs=merged)
ae = create_model()
ae.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=1e-3),
loss=tf.keras.losses.MeanSquaredError(),
metrics=['accuracy'])
test_history = ae.fit(x_train,
x_train,
batch_size=1,
epochs=1,
callbacks=None)
Stacktrace
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
<ipython-input-5-733ae79232be> in <module>()
55
56
---> 57 ae = create_model()
58
59 ae.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=1e-3),
5 frames
/usr/local/lib/python3.7/dist-packages/tensorflow/python/autograph/impl/api.py in wrapper(*args, **kwargs)
668 except Exception as e: # pylint:disable=broad-except
669 if hasattr(e, 'ag_error_metadata'):
--> 670 raise e.ag_error_metadata.to_exception(e)
671 else:
672 raise
TypeError: in user code:
<ipython-input-5-733ae79232be>:41 call *
target_grad_mapping = tf.gradients(target_patches, target_volume)[0]
/usr/local/lib/python3.7/dist-packages/tensorflow/python/ops/gradients_impl.py:318 gradients_v2 **
unconnected_gradients)
/usr/local/lib/python3.7/dist-packages/tensorflow/python/ops/gradients_util.py:684 _GradientsHelper
lambda: grad_fn(op, *out_grads))
/usr/local/lib/python3.7/dist-packages/tensorflow/python/ops/gradients_util.py:340 _MaybeCompile
return grad_fn() # Exit early
/usr/local/lib/python3.7/dist-packages/tensorflow/python/ops/gradients_util.py:684 <lambda>
lambda: grad_fn(op, *out_grads))
/usr/local/lib/python3.7/dist-packages/tensorflow/python/ops/array_grad.py:1067 _ExtractVolumePatchesGrad
input_indices_num = 1 + planes_in * rows_in * cols_in
TypeError: unsupported operand type(s) for *: 'NoneType' and 'NoneType'
明显改变类CombinePatches调用方法
target_grad_mapping = tf.gradients(target_patches, target_volume)[0]
# Computes gradients again and dividing by grad, otherwise its just summed.
return tf.gradients(target_patches, target_volume, patches)[0] / target_grad_mapping
至
if target_patches.shape[0] is not None:
target_grad_mapping = tf.gradients(target_patches, target_volume)[0]
# Computes gradients again and dividing by grad, otherwise its just summed.
return tf.gradients(target_patches, target_volume, patches)[0] / target_grad_mapping
return target_volume
成功了。