Could you share which version of tf and keras you use? As this no longer works with latest versions.

Hello,

Thanks for a great project. I learnt a lot from this!

I was wondering one thing: Why does value of t has to be between [0 - 1]? I notice when I change it to 10, it is significantly lower to train the model.

Also, I notice t is a variable. Do we optimize t together with network's parameters, or t fixed? If t is not fixed, what is the value of t after training? (Maybe I don't understand the paper is well enough and this can be a naive question).

Thanks a lot!

Best,

It would be useful to know training times of both solutions in order to compare them and see what solution is better.

Hi Jason,

Thanks for sharing. When I try training the ode-block, seeing no response in terminal, it keeps waiting for long time and I am killing the process at the end. When trying for Resnet, it is working and getting results as normal. Do you have any suggestion about this?
I am actually using gtx 1080 Ti with 12 gb and I think it should work.

Using TensorFlow backend.
Traceback (most recent call last):
File "predict.py", line 63, in
model = load_model('./saved_models/OdeNetF95.h5')
File "/anaconda3/envs/experiment/lib/python3.5/site-packages/keras/engine/saving.py", line 419, in load_model
model = _deserialize_model(f, custom_objects, compile)
File "/anaconda3/envs/experiment/lib/python3.5/site-packages/keras/engine/saving.py", line 225, in _deserialize_model
model = model_from_config(model_config, custom_objects=custom_objects)
File "/anaconda3/envs/experiment/lib/python3.5/site-packages/keras/engine/saving.py", line 458, in model_from_config
return deserialize(config, custom_objects=custom_objects)
File "/anaconda3/envs/experiment/lib/python3.5/site-packages/keras/layers/init.py", line 55, in deserialize
printable_module_name='layer')
File "/anaconda3/envs/experiment/lib/python3.5/site-packages/keras/utils/generic_utils.py", line 145, in deserialize_keras_object
list(custom_objects.items())))
File "/anaconda3/envs/experiment/lib/python3.5/site-packages/keras/engine/network.py", line 1022, in from_config
process_layer(layer_data)
File "/anaconda3/envs/experiment/lib/python3.5/site-packages/keras/engine/network.py", line 1008, in process_layer
custom_objects=custom_objects)
File "/anaconda3/envs/experiment/lib/python3.5/site-packages/keras/layers/init.py", line 55, in deserialize
printable_module_name='layer')
File "/anaconda3/envs/experiment/lib/python3.5/site-packages/keras/utils/generic_utils.py", line 138, in deserialize_keras_object
': ' + class_name)
ValueError: Unknown layer: ODEBlock

FOR THE CODE:

from keras.models import load_model
import cv2
from keras.layers import Conv2D, Dense, Flatten, Input, MaxPooling2D, Layer
import numpy as np

class ODEBlock(Layer):

def __init__(self, filters, kernel_size, **kwargs):
    self.filters = filters
    self.kernel_size = kernel_size
    super(ODEBlock, self).__init__(**kwargs)

def build(self, input_shape):
    self.conv2d_w1 = self.add_weight("conv2d_w1", self.kernel_size + (self.filters + 1, self.filters), initializer='glorot_uniform')
    self.conv2d_w2 = self.add_weight("conv2d_w2", self.kernel_size + (self.filters + 1, self.filters), initializer='glorot_uniform')
    self.conv2d_b1 = self.add_weight("conv2d_b1", (self.filters,), initializer='zero')
    self.conv2d_b2 = self.add_weight("conv2d_b2", (self.filters,), initializer='zero')
    super(ODEBlock, self).build(input_shape)

def call(self, x):
    t = K.constant([0, 1], dtype="float32")
    return tf.contrib.integrate.odeint(self.ode_func, x, t, rtol=1e-3, atol=1e-3)[1]

def compute_output_shape(self, input_shape):
    return input_shape

def ode_func(self, x, t):
    y = self.concat_t(x, t)
    y = K.conv2d(y, self.conv2d_w1, padding="same")
    y = K.bias_add(y, self.conv2d_b1)
    y = K.relu(y)

    y = self.concat_t(y, t)
    y = K.conv2d(y, self.conv2d_w2, padding="same")
    y = K.bias_add(y, self.conv2d_b2)
    y = K.relu(y)

    return y

def concat_t(self, x, t):
    new_shape = tf.concat(
        [
            tf.shape(x)[:-1],
            tf.constant([1],dtype="int32",shape=(1,))
        ], axis=0)

    t = tf.ones(shape=new_shape) * tf.reshape(t, (1, 1, 1, 1))
    return tf.concat([x, t], axis=-1)

def build_model(input_shape, num_classes):
x = Input(input_shape)
y = Conv2D(32, (3, 3), activation='relu')(x)
y = MaxPooling2D((2,2))(y)
y = Conv2D(64, (3, 3), activation='relu')(y)
y = MaxPooling2D((2,2))(y)
y = ODEBlock(64, (3, 3))(y)
y = Flatten()(y)
y = Dense(num_classes, activation='softmax')(y)
return Model(x,y)

model = load_model('./saved_models/OdeNetF95.h5')

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

model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=keras.optimizers.Adam(),
metrics=['accuracy'])
img = cv2.imread('test.png')
img = np.reshape(img,[-1,71, 64,1])

classes = model.predict_classes(img)

print(classes)

Are we using the adjoint method as described in the paper?