将线性代数应用于 tensorflow 中的参差不齐的张量

Question

I'm using tensorflow v2.7.0 and trying to create a ML model using ragged tensor.

The issue is that tf.linalg.diag, tf.matmul and tf.linalg.det are not working with ragged tensor. I've found a workaround by converting the ragged tensor in numpy and converting it back to a ragged tensor but it's not working when applying the layer in a global model.

The following code is working

 import tensorflow as tf class LRDet(tf.keras.layers.Layer): def __init__(self,numItems,rank=10): super(LRDet,self).__init__() self.numItems = numItems self.rank = rank def build(self,input_shape): V_init = tf.random_normal_initializer(mean=0.0,stddev=0.01) D_init = tf.random_normal_initializer(mean=1.0,stddev=0.01) self.V = tf.Variable(name='V',initial_value=V_init(shape=(self.numItems, self.rank)),trainable=True) self.D = tf.Variable(name='D',initial_value=D_init(shape=(self.numItems,)),trainable=True) def call(self,inputs): batch_size = inputs.nrows() subV = tf.gather(self.V,inputs) subD = tf.square(tf.gather(self.D,inputs,batch_dims=0))#tf.linalg.diag(tf.square(tf.gather(D,Xrag,batch_dims=0))) subD = tf.ragged.constant([tf.linalg.diag(subD[i]).numpy() for i in tf.range(batch_size)]) K = tf.ragged.constant([tf.matmul(subV[i],subV[i],transpose_b=True).numpy() for i in tf.range(batch_size)]) K = tf.add(K,subD) res = tf.ragged.constant([tf.linalg.det(K[i].to_tensor()).numpy() for i in tf.range(batch_size)]) return res numItems = 10 rank = 3 detX = LRDet(numItems,rank) X = [[1,2],[3],[4,5,6]] Xrag = tf.ragged.constant(X) _ = detX(Xrag)

But once I used this layer in a more global model I have the following error

OperatorNotAllowedInGraphError: Exception encountered when calling layer "lr_det_10" (type LRDet).

 in user code: File "<ipython-input-57-6b073a14386e>", line 18, in call * subD = tf.ragged.constant([tf.linalg.diag(subD[i]).numpy() for i in tf.range(batch_size)]) OperatorNotAllowedInGraphError: iterating over `tf.Tensor` is not allowed: AutoGraph did convert this function. This might indicate you are trying to use an unsupported feature.

I tried to used tf.map_fn instead of list comprehension with.numpy() but without success.

Any help would be very much appreciated.

Thanks !

Answer 1

Here is an option running with tf.map_fn ; however, it currently only runs on the CPU due to a very recent bug regarding tf.map_fn , ragged tensors and the GPU:

import os
os.environ["CUDA_VISIBLE_DEVICES"] = "-1" # do not access GPU
import tensorflow as tf

class LRDet(tf.keras.layers.Layer):
    
    def __init__(self,numItems,rank=10):
        super(LRDet,self).__init__()
        self.numItems = numItems
        self.rank = rank
    
    def build(self,input_shape):
        V_init = tf.random_normal_initializer(mean=0.0,stddev=0.01)
        D_init = tf.random_normal_initializer(mean=1.0,stddev=0.01)
        self.V = tf.Variable(name='V',initial_value=V_init(shape=(self.numItems, self.rank)),trainable=True)
        self.D = tf.Variable(name='D',initial_value=D_init(shape=(self.numItems,)),trainable=True)

    def call(self,inputs):
        batch_size = inputs.nrows()
        subV = tf.gather(self.V,inputs)
    
        subD = tf.square(tf.gather(self.D, inputs, batch_dims=0))
        subD = tf.map_fn(self.diag, subD, fn_output_signature=tf.RaggedTensorSpec(shape=[1, None, None],
                                                                    dtype=tf.type_spec_from_value(subD).dtype,
                                                                    ragged_rank=2,
                                                                    row_splits_dtype=tf.type_spec_from_value(subD).row_splits_dtype))
        subD = tf.squeeze(subD, 1)


        K = tf.map_fn(self.matmul, subV, fn_output_signature=tf.RaggedTensorSpec(shape=[1, None, None],
                                                                    dtype=tf.type_spec_from_value(subV).dtype,
                                                                    ragged_rank=2,
                                                                    row_splits_dtype=tf.type_spec_from_value(subV).row_splits_dtype))
        K = tf.squeeze(K, 1)
        K = tf.add(K,subD)

        res = tf.map_fn(self.det, K, tf.TensorSpec(shape=(), dtype=tf.float32, name=None))

        return res
        

    def diag(self, x):
      return tf.ragged.stack(tf.linalg.diag(x))

    def matmul(self, x):
      return tf.ragged.stack(tf.matmul(x, x,transpose_b=True))

    def det(self, x):
      return tf.linalg.det(x.to_tensor())


numItems = 10
rank = 3
input = tf.keras.layers.Input(shape=(None,), ragged=True, dtype=tf.int32)
detX = LRDet(numItems,rank)
output = detX(input)
model = tf.keras.Model(input, output)

X = [[1,2],[3],[4,5,6]]
Xrag = tf.ragged.constant(X)
y = tf.random.normal((3, 1))
model.compile(loss='mse', optimizer='adam')
model.fit(Xrag, y, batch_size=1, epochs=1)