Keras Compatibility with Present Version (2.2.4)

[zafarmah92]

Hi ,

I am trying to re-run the code but due to keras change of version I am unable to do such .

My problem here is with Model.py file in the function " build_icm_model ( ... ) "

Below the merge operation is performed but mode = that we can create the custom lambda operations, but in the updated keras this functionality is not longer available

can you kindly take a look how i can reproduce the merge and lambda operation below for l_i , loss0 and loss
Also what exactly the operation is performed here, would be great

Many thanks

def build_icm_model(state_shape, action_shape, lmd=1.0, beta=0.01):
    s_t0 = Input(shape=state_shape, name="state0")
    s_t1 = Input(shape=state_shape, name="state1")
    a_t = Input(shape=action_shape, name="action")
    reshape = Reshape(target_shape=(1,) + state_shape)
    fmap = build_feature_map((1,) + state_shape)
    f_t0 = fmap(reshape(s_t0))
    f_t1 = fmap(reshape(s_t1))
    act_hat = inverse_model()(f_t0, f_t1)
    f_t1_hat = forward_model()(f_t0, a_t)
    r_in = merge([f_t1, f_t1_hat], mode=lambda x: 0.5 * K.sum(K.square(x[0] - x[1]), axis=-1),
                 output_shape=(1,), name="reward_intrinsic")
    l_i = merge([a_t, act_hat], mode=lambda x: -K.sum(x[0] * K.log(x[1] + K.epsilon()), axis=-1),
                output_shape=(1,))
    loss0 = merge([r_in, l_i], mode=lambda x: beta * x[0] + (1.0 - beta) * x[1], output_shape=(1,))
    rwd = Input(shape=(1,))
    loss = merge([rwd, loss0], mode=lambda x: (-lmd * x[0].T + x[1]).T, output_shape=(1,))
    return Model([s_t0, s_t1, a_t, rwd], loss)