我可以在具有自定义损失的多输出模型上使用 Keras model.fit()，该模型使用 Tensorflow 2 中的所有输出目标和预测吗？

Question

I have tried and failed to make Keras model.fit() work on my multi-output model with a custom loss that uses all outputs' targets and predictions (specifically for 2 outputs) in TF 2.

When I tried to do this on a model made with the Keras functional API, I get the error: "SymbolicException: Inputs to eager execution function cannot be Keras symbolic tensors, but found ..." meaning I can't use my loss function because it returns an eager tensor to a Keras DAG that works with symbolic tensors (functional API model). To get around this, I used model.add_loss() instead of passing my loss function into model.compile(), but I believe this hogged GPU memory and caused OOM errors.

I've tried workarounds, where I put my functional API model inside a Keras subclassed model or make a completely new Keras subclassed model.

Workaround 1 is below in code, and runs yet gives me NaNs across the epochs on training on a variety of gradient clippings, and gives 0-valued outputs.

Workaround 2 gives me an error inside the override call() method because the inputs param is different shapes during model compile-time and run-time because my model (in a quirky way) has 3 inputs: 1 is the actual input to the DLNN, and the 2 others are the targets for the input sample. This is so that I can get the targets from each sample into the loss function.

from scipy.io import wavfile
import scipy.signal as sg
import matplotlib.pyplot as plt
import tensorflow as tf
from tensorflow.keras.layers import Input, SimpleRNN, Dense, Lambda, TimeDistributed, Layer, LSTM, Bidirectional, BatchNormalization, Concatenate
from tensorflow.keras.models import Model
from tensorflow.keras.activations import relu
from tensorflow.keras.callbacks import EarlyStopping
import numpy as np
import datetime
import numpy as np
import math
import random
import json
import os
import sys



# Loss function
def discriminative_loss(piano_true, noise_true, piano_pred, noise_pred, loss_const):
    last_dim = piano_pred.shape[1] * piano_pred.shape[2]
    return (
        tf.math.reduce_mean(tf.reshape(noise_pred - noise_true, shape=(-1, last_dim)) ** 2, axis=-1) - 
        (loss_const * tf.math.reduce_mean(tf.reshape(noise_pred - piano_true, shape=(-1, last_dim)) ** 2, axis=-1)) +
        tf.math.reduce_mean(tf.reshape(piano_pred - piano_true, shape=(-1, last_dim)) ** 2, axis=-1) -
        (loss_const * tf.math.reduce_mean(tf.reshape(piano_pred - noise_true, shape=(-1, last_dim)) ** 2, axis=-1))
    )



def make_model(features, sequences, name='Model'):

    input_layer = Input(shape=(sequences, features), dtype='float32', 
                        name='piano_noise_mixed')
    piano_true = Input(shape=(sequences, features), dtype='float32', 
                       name='piano_true')
    noise_true = Input(shape=(sequences, features), dtype='float32', 
                       name='noise_true')

    x = SimpleRNN(features // 2, 
                  activation='relu', 
                  return_sequences=True) (input_layer) 
    piano_pred = TimeDistributed(Dense(features), name='piano_hat') (x)  # source 1 branch
    noise_pred = TimeDistributed(Dense(features), name='noise_hat') (x)  # source 2 branch
  
    model = Model(inputs=[input_layer, piano_true, noise_true],
                  outputs=[piano_pred, noise_pred])

    return model



# Model "wrapper" for many-input loss function
class RestorationModel2(Model):
    def __init__(self, model, loss_const):
        super(RestorationModel2, self).__init__()
        self.model = model
        self.loss_const = loss_const
       
    def call(self, inputs):
        return self.model(inputs)

    def compile(self, optimizer, loss):
        super(RestorationModel2, self).compile()
        self.optimizer = optimizer
        self.loss = loss

    def train_step(self, data):
        # Unpack data - what generator yeilds
        x, piano_true, noise_true = data

        with tf.GradientTape() as tape:
            piano_pred, noise_pred = self.model((x, piano_true, noise_true), training=True)
            loss = self.loss(piano_true, noise_true, piano_pred, noise_pred, self.loss_const)

        trainable_vars = self.model.trainable_variables
        gradients = tape.gradient(loss, trainable_vars)
        self.optimizer.apply_gradients(zip(gradients, trainable_vars))
        
        return {'loss': loss}

    def test_step(self, data):
        x, piano_true, noise_true = data

        piano_pred, noise_pred = self.model((x, piano_true, noise_true), training=False)
        loss = self.loss(piano_true, noise_true, piano_pred, noise_pred, self.loss_const)
        
        return {'loss': loss}



def make_imp_model(features, sequences, loss_const=0.05, 
                   optimizer=tf.keras.optimizers.RMSprop(clipvalue=0.7),
                   name='Restoration Model', epsilon=10 ** (-10)):
    
    # NEW Semi-imperative model
    model = RestorationModel2(make_model(features, sequences, name='Training Model'),
                              loss_const=loss_const)

    model.compile(optimizer=optimizer, loss=discriminative_loss)

    return model



# MODEL TRAIN & EVAL FUNCTION
def evaluate_source_sep(train_generator, validation_generator,
                        num_train, num_val, n_feat, n_seq, batch_size, 
                        loss_const, epochs=20, 
                        optimizer=tf.keras.optimizers.RMSprop(clipvalue=0.75),
                        patience=10, epsilon=10 ** (-10)):
   
    print('Making model...')    # IMPERATIVE MODEL - Customize Fit
    model = make_imp_model(n_feat, n_seq, loss_const=loss_const, optimizer=optimizer, epsilon=epsilon)
    
    print('Going into training now...')
    hist = model.fit(train_generator,
                     steps_per_epoch=math.ceil(num_train / batch_size),
                     epochs=epochs,
                     validation_data=validation_generator,
                     validation_steps=math.ceil(num_val / batch_size),
                     callbacks=[EarlyStopping('val_loss', patience=patience, mode='min')])
    print(model.summary())



# NEURAL NETWORK DATA GENERATOR
def my_dummy_generator(num_samples, batch_size, train_seq, train_feat):

    while True:
        for offset in range(0, num_samples, batch_size):

            # Initialise x, y1 and y2 arrays for this batch
            x, y1, y2 = (np.empty((batch_size, train_seq, train_feat)),
                            np.empty((batch_size, train_seq, train_feat)),
                            np.empty((batch_size, train_seq, train_feat)))

            yield (x, y1, y2)



def main():
    epsilon = 10 ** (-10)
    train_batch_size = 5
    loss_const, epochs, val_split = 0.05, 10, 0.25
    optimizer = tf.keras.optimizers.RMSprop(clipvalue=0.9)

    TRAIN_SEQ_LEN, TRAIN_FEAT_LEN = 1847, 2049
    TOTAL_SMPLS = 60 

    # Validation & Training Split
    indices = list(range(TOTAL_SMPLS))
    val_indices = indices[:math.ceil(TOTAL_SMPLS * val_split)]
    num_val = len(val_indices)
    num_train = TOTAL_SMPLS - num_val
   
    train_seq, train_feat = TRAIN_SEQ_LEN, TRAIN_FEAT_LEN
    print('Train Input Stats:')
    print('N Feat:', train_feat, 'Seq Len:', train_seq, 'Batch Size:', train_batch_size)

    # Create data generators and evaluate model with them
    train_generator = my_dummy_generator(num_train,
                        batch_size=train_batch_size, train_seq=train_seq,
                        train_feat=train_feat)
    validation_generator = my_dummy_generator(num_val,
                        batch_size=train_batch_size, train_seq=train_seq,
                        train_feat=train_feat)

    evaluate_source_sep(train_generator, validation_generator, num_train, num_val,
                            n_feat=train_feat, n_seq=train_seq, 
                            batch_size=train_batch_size, 
                            loss_const=loss_const, epochs=epochs,
                            optimizer=optimizer, epsilon=epsilon)

if __name__ == '__main__':
    main()

Thanks for the help!

Answer 1

Solution, don't pass your loss into model.add_loss(). Instead concatenate your outputs together which lets you pass your custom loss into model.compile(). Then deal with the outputs in the custom loss function.

class TimeFreqMasking(Layer):
    # Init is for input-independent variables
    def __init__(self, epsilon, **kwargs):
        super(TimeFreqMasking, self).__init__(**kwargs)
        self.epsilon = epsilon

    # No build method, b/c passing in multiple inputs to layer (no single shape)

    def call(self, inputs):
        y_hat_self, y_hat_other, x_mixed = inputs
        mask = tf.abs(y_hat_self) / (tf.abs(y_hat_self) + tf.abs(y_hat_other) + self.epsilon)
        y_tilde_self = mask * x_mixed
        return y_tilde_self


def discrim_loss(y_true, y_pred):
    piano_true, noise_true = tf.split(y_true, num_or_size_splits=2, axis=-1)
    loss_const = y_pred[-1, :, :][0][0]
    piano_pred, noise_pred = tf.split(y_pred[:-1, :, :], num_or_size_splits=2, axis=0)

    last_dim = piano_pred.shape[1] * piano_pred.shape[2]
    return (
        tf.math.reduce_mean(tf.reshape(noise_pred - noise_true, shape=(-1, last_dim)) ** 2) - 
        (loss_const * tf.math.reduce_mean(tf.reshape(noise_pred - piano_true, shape=(-1, last_dim)) ** 2)) +
        tf.math.reduce_mean(tf.reshape(piano_pred - piano_true, shape=(-1, last_dim)) ** 2) -
        (loss_const * tf.math.reduce_mean(tf.reshape(piano_pred - noise_true, shape=(-1, last_dim)) ** 2))
    )


def make_model(features, sequences, epsilon, loss_const):
    input_layer = Input(shape=(sequences, features), name='piano_noise_mixed')
    x = SimpleRNN(features // 2, 
                activation='relu', 
                return_sequences=True) (input_layer) 
    x = SimpleRNN(features // 2, 
            activation='relu',
            return_sequences=True) (x)
    piano_hat = TimeDistributed(Dense(features), name='piano_hat') (x)  # source 1 branch
    noise_hat = TimeDistributed(Dense(features), name='noise_hat') (x)  # source 2 branch
    piano_pred = TimeFreqMasking(epsilon=epsilon, 
                                name='piano_pred') ((piano_hat, noise_hat, input_layer))
    noise_pred = TimeFreqMasking(epsilon=epsilon, 
                                name='noise_pred') ((noise_hat, piano_hat, input_layer))

    preds_and_gamma = Concatenate(axis=0) ([piano_pred, 
                                        noise_pred, 
                                        #  loss_const_tensor
                                        tf.broadcast_to(tf.constant(loss_const), [1, sequences, features])
                                        ])
    model = Model(inputs=input_layer, outputs=preds_and_gamma)
    model.compile(optimizer=optimizer, loss=discrim_loss)
    return model


def dummy_generator(num_samples, batch_size, num_seq, num_feat):
    while True:
        for _ in range(0, num_samples, batch_size):
            x, y1, y2 = (np.random.rand(batch_size, num_seq, num_feat),
                        np.random.rand(batch_size, num_seq, num_feat),
                        np.random.rand(batch_size, num_seq, num_feat))

            yield ([x, np.concatenate((y1, y2), axis=-1)])


total_samples = 6
batch_size = 2
time_steps = 3
features = 4
loss_const = 2
epochs = 10
val_split = 0.25
epsilon = 10 ** (-10)

model = make_model(features, time_steps, epsilon, loss_const)
print(model.summary())

num_val = math.ceil(actual_samples * val_split)
num_train = total_samples - val_samples
train_dataset = dummy_generator(num_train, batch_size, time_steps, features)
val_dataset = dummy_generator(num_val, batch_size, time_steps, features)

model.fit(train_dataset,
                steps_per_epoch=math.ceil(num_train / batch_size),
                epochs=epochs,
                validation_data=val_dataset,
                validation_steps=math.ceil(num_val / batch_size)