Model learns when using keras but doesn't with tf.keras

Question

The model that I am using is this:

from keras.layers import (Input, MaxPooling1D, Dropout,
                          BatchNormalization, Activation, Add,
                          Flatten, Conv1D, Dense)
from keras.models import Model
import numpy as np


class ResidualUnit(object):
    """References
    ----------
    .. [1] K. He, X. Zhang, S. Ren, and J. Sun, "Identity Mappings in Deep Residual Networks,"
           arXiv:1603.05027 [cs], Mar. 2016. https://arxiv.org/pdf/1603.05027.pdf.
    .. [2] K. He, X. Zhang, S. Ren, and J. Sun, "Deep Residual Learning for Image Recognition," in 2016 IEEE Conference
           on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770-778. https://arxiv.org/pdf/1512.03385.pdf
    """

    def __init__(self, n_samples_out, n_filters_out, kernel_initializer='he_normal',
                 dropout_rate=0.8, kernel_size=17, preactivation=True,
                 postactivation_bn=False, activation_function='relu'):
        self.n_samples_out = n_samples_out
        self.n_filters_out = n_filters_out
        self.kernel_initializer = kernel_initializer
        self.dropout_rate = dropout_rate
        self.kernel_size = kernel_size
        self.preactivation = preactivation
        self.postactivation_bn = postactivation_bn
        self.activation_function = activation_function

    def _skip_connection(self, y, downsample, n_filters_in):
        """Implement skip connection."""
        # Deal with downsampling
        if downsample > 1:
            y = MaxPooling1D(downsample, strides=downsample, padding='same')(y)
        elif downsample == 1:
            y = y
        else:
            raise ValueError("Number of samples should always decrease.")
        # Deal with n_filters dimension increase
        if n_filters_in != self.n_filters_out:
            # This is one of the two alternatives presented in ResNet paper
            # Other option is to just fill the matrix with zeros.
            y = Conv1D(self.n_filters_out, 1, padding='same',
                       use_bias=False,
                       kernel_initializer=self.kernel_initializer
                       )(y)
            return y

    def _batch_norm_plus_activation(self, x):
        if self.postactivation_bn:
            x = Activation(self.activation_function)(x)
            x = BatchNormalization(center=False, scale=False)(x)
        else:
            x = BatchNormalization()(x)
            x = Activation(self.activation_function)(x)
        return x

    def __call__(self, inputs):
        """Residual unit."""
        x, y = inputs
        n_samples_in = y.shape[1]
        downsample = n_samples_in // self.n_samples_out
        n_filters_in = y.shape[2]
        y = self._skip_connection(y, downsample, n_filters_in)
        # 1st layer
        x = Conv1D(self.n_filters_out, self.kernel_size, padding='same',
                   use_bias=False,
                   kernel_initializer=self.kernel_initializer
                   )(x)
        x = self._batch_norm_plus_activation(x)
        if self.dropout_rate > 0:
            x = Dropout(self.dropout_rate)(x)

        # 2nd layer
        x = Conv1D(self.n_filters_out, self.kernel_size, strides=downsample,
                   padding='same', use_bias=False,
                   kernel_initializer=self.kernel_initializer
                   )(x)
        if self.preactivation:
            x = Add()([x, y])  # Sum skip connection and main connection
            y = x
            x = self._batch_norm_plus_activation(x)
            if self.dropout_rate > 0:
                x = Dropout(self.dropout_rate)(x)
        else:
            x = BatchNormalization()(x)
            x = Add()([x, y])  # Sum skip connection and main connection
            x = Activation(self.activation_function)(x)
            if self.dropout_rate > 0:
                x = Dropout(self.dropout_rate)(x)
            y = x
        return [x, y]


# ----- Model ----- #
kernel_size = 16
kernel_initializer = 'he_normal'
signal = Input(shape=(1000, 12), dtype=np.float32, name='signal')
age_range = Input(shape=(6,), dtype=np.float32, name='age_range')
is_male = Input(shape=(1,), dtype=np.float32, name='is_male')
x = signal
x = Conv1D(64, kernel_size, padding='same', use_bias=False,
           kernel_initializer=kernel_initializer
           )(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x, y = ResidualUnit(512, 128, kernel_size=kernel_size,
                    kernel_initializer=kernel_initializer
                    )([x, x])
x, y = ResidualUnit(256, 196, kernel_size=kernel_size,
                    kernel_initializer=kernel_initializer
                    )([x, y])
x, y = ResidualUnit(64, 256, kernel_size=kernel_size,
                    kernel_initializer=kernel_initializer
                    )([x, y])
x, _ = ResidualUnit(16, 320, kernel_size=kernel_size, kernel_initializer=kernel_initializer
                    )([x, y])
x = Flatten()(x)
diagn = Dense(2, activation='sigmoid', kernel_initializer=kernel_initializer)(x)
model = Model(signal, diagn)

model.summary()


# ----- Train ----- #

from keras.optimizers import Adam
from keras.callbacks import ModelCheckpoint, ReduceLROnPlateau

loss = 'binary_crossentropy'
lr = 0.001
batch_size = 64
opt = Adam(learning_rate=0.001)
callbacks = [ReduceLROnPlateau(monitor='val_loss',
                                factor=0.1,
                                patience=7,
                                min_lr=lr / 100)]

model.compile(optimizer=opt, loss=loss, metrics=['accuracy'])

history = model.fit(x_train, y_train,
                    batch_size=batch_size,
                    epochs=70,
                    initial_epoch=0, 
                    validation_split=0.1,
                    shuffle='batch', 
                    callbacks=callbacks,
                    verbose=1)

# Save final result
model.save("./final_model_middle_one.hdf5")

When I substitute the use of Keras with tf.keras, which I need to use the qkeras library, the model doesn't learn and gets stuck at a much lower accuracy at every iteration. What could be causing this?

When I use keras the accuracy start high at 83% and slightly increases during training.

Train on 17340 samples, validate on 1927 samples
Epoch 1/70
17340/17340 [==============================] - 33s 2ms/step - loss: 0.3908 - accuracy: 0.8314 - val_loss: 0.3283 - val_accuracy: 0.8710
Epoch 2/70
17340/17340 [==============================] - 31s 2ms/step - loss: 0.3641 - accuracy: 0.8416 - val_loss: 0.3340 - val_accuracy: 0.8612
Epoch 3/70
17340/17340 [==============================] - 31s 2ms/step - loss: 0.3525 - accuracy: 0.8483 - val_loss: 0.3847 - val_accuracy: 0.8550
Epoch 4/70
17340/17340 [==============================] - 31s 2ms/step - loss: 0.3354 - accuracy: 0.8563 - val_loss: 0.4641 - val_accuracy: 0.8215
Epoch 5/70
17340/17340 [==============================] - 31s 2ms/step - loss: 0.3269 - accuracy: 0.8590 - val_loss: 0.7172 - val_accuracy: 0.7870
Epoch 6/70
17340/17340 [==============================] - 31s 2ms/step - loss: 0.3202 - accuracy: 0.8630 - val_loss: 0.3599 - val_accuracy: 0.8617
Epoch 7/70
17340/17340 [==============================] - 31s 2ms/step - loss: 0.3101 - accuracy: 0.8678 - val_loss: 0.2659 - val_accuracy: 0.8934
Epoch 8/70
17340/17340 [==============================] - 31s 2ms/step - loss: 0.3058 - accuracy: 0.8688 - val_loss: 0.5683 - val_accuracy: 0.8293
Epoch 9/70
17340/17340 [==============================] - 31s 2ms/step - loss: 0.2980 - accuracy: 0.8739 - val_loss: 0.3442 - val_accuracy: 0.8643
Epoch 10/70
 7424/17340 [===========>..................] - ETA: 17s - loss: 0.2966 - accuracy: 0.8707

When I use tf.keras the accuracy starts at 50% and does not increase considerably during training:

Epoch 1/70
271/271 [==============================] - 30s 110ms/step - loss: 0.9325 - accuracy: 0.5093 - val_loss: 0.6973 - val_accuracy: 0.5470 - lr: 0.0010
Epoch 2/70
271/271 [==============================] - 29s 108ms/step - loss: 0.8424 - accuracy: 0.5157 - val_loss: 0.6660 - val_accuracy: 0.6528 - lr: 0.0010
Epoch 3/70
271/271 [==============================] - 29s 108ms/step - loss: 0.8066 - accuracy: 0.5213 - val_loss: 0.6441 - val_accuracy: 0.6539 - lr: 0.0010
Epoch 4/70
271/271 [==============================] - 29s 108ms/step - loss: 0.7884 - accuracy: 0.5272 - val_loss: 0.6649 - val_accuracy: 0.6559 - lr: 0.0010
Epoch 5/70
271/271 [==============================] - 29s 108ms/step - loss: 0.7888 - accuracy: 0.5368 - val_loss: 0.6899 - val_accuracy: 0.5760 - lr: 0.0010
Epoch 6/70
271/271 [==============================] - 29s 108ms/step - loss: 0.7617 - accuracy: 0.5304 - val_loss: 0.6641 - val_accuracy: 0.6533 - lr: 0.0010
Epoch 7/70
271/271 [==============================] - 29s 108ms/step - loss: 0.7485 - accuracy: 0.5333 - val_loss: 0.6450 - val_accuracy: 0.6544 - lr: 0.0010
Epoch 8/70
271/271 [==============================] - 29s 108ms/step - loss: 0.7431 - accuracy: 0.5382 - val_loss: 0.6599 - val_accuracy: 0.6539 - lr: 0.0010
Epoch 9/70
271/271 [==============================] - 29s 108ms/step - loss: 0.7336 - accuracy: 0.5421 - val_loss: 0.6532 - val_accuracy: 0.6554 - lr: 0.0010
Epoch 10/70
271/271 [==============================] - 29s 108ms/step - loss: 0.7274 - accuracy: 0.5379 - val_loss: 0.6753 - val_accuracy: 0.6492 - lr: 0.0010

The lines that have been changed between the two trials are the lines where I import keras modules by adding 'tensorflow.' in front of them. I don't know why the results would be so different, possibly due to different default values of certain parameters?

Answer 1

It might be related to how the accuracy metric is computed in keras vs tf.keras . As far as I can tell the accuracy function is usually used when you have one-hot-encoded output. However, it seems that you are outputting two values [A, B] with a sigmoid function applied to each value. Since I don't know the labels you're using, there might be two cases:

a) You want to predict A or B . If sos I would change the activation function to softmax

b) You want to predict between A or not A and B or not B . In this case I would modify the output tensor shape to have two heads, each with two values: head_A = [A, not_A] and head_B = [B, not_B] . I would then hot-encode the labels respectively and then I would assume you could use the accuracy metric.

Alternatively, you can create a custom metric that is appropriate to your output shape.

Answer 2

I have a similar (same?) problem, I was manipulating some examples from Kaggle, and was unable to save the model using keras. After much Googling I realised that I needed to use tensorflow.keras. This solved my problem, but the 60000 data items I have and was using for training dropped to a reported 1875. Although the error was still 10%.

1875 * 32 = 60000.

This is my fit.

model.fit(X_train, y_train, validation_data=(X_test, y_test), epochs=epochs, verbose=True, callbacks=[early_stopping_monitor])

1539/1875 [=======================>......] - ETA: 3s - loss: 0.4445 - accuracy: 0.8418

It turns out that fit defaults to a batch size of 32. If I increase the batch size to 64 I get half the reported data sets, which makes sense:

model.fit(X_train, y_train, batch_size=64, validation_data=(X_test, y_test), epochs=epochs, verbose=True, callbacks=[early_stopping_monitor])

938/938 [==============================] - 16s 17ms/step - loss: 0.4568 - accuracy: 0.8388

I noticed from your code that you've set batch_size to 64, and your reported data items reduce from 17340 to 271 which is about a 64th, this must also affect your accuracy due to the data you are using.

From the docs here: https://www.tensorflow.org/api_docs/python/tf/keras/Sequential

batch_size
Integer or None. Number of samples per gradient update. If unspecified, batch_size will default to 32. Do not specify the batch_size if your data is in the form of a dataset, generators, or keras.utils.Sequence instances (since they generate batches).

From the Keras docs: https://keras.rstudio.com/reference/fit.html , it also says that the batch size defaults to 32, it must just be reported differently when training the model.

Hope this helps.