Tensorflow Backend Questions

[wcqc]

Hi,

Two questions about using TC with the TensorFlow backend:

1. When using the Tensorflow backend and keras.Sequential, does TensorCircuit support putting additional Keras layers before the qml_layer, such as a dense layer to transform the input before being fed to a qml_layer defined in TC? If yes, would it be possible to let TC fine-tune the input embeddings? Something like this:

model = tf.keras.Sequential() model.add(tf.keras.layers.Embedding(100, 5)) model.add(tf.keras.layers.Dense(num_params)) model.add(qml_layer) model.add(tf.keras.layers.Dense(1, activation="sigmoid")) 

2. What's the best way to debug and print the variables in the debugger when using the TF backend? tf.print() doesn't seem to work?

Thanks!

[refraction-ray]

1. of course, qmllayer is just a keras layer, you can compose any computational graph with all kinds of classical and quantum layers in one model
2. you can always directly print(model.variables) outside a jitted function?

[wcqc]

Thanks! Is there a good recommended way to debug and print within a quantum layer such as (when used as a Keras layer -- set run_eagerly=True doesn't seem to help):

def qlayer(x): c = tc.Circuit(n, inputs=x) ### print here??? for j in range(max_sent_len): for i in range(n): c.rx(i, theta=x[i]) 

[refraction-ray]

K = tc.set_backend("tensorflow") def f(param, weights): c = tc.Circuit(2) print(param) return 1.0 qmllayer = tc.keras.HardwareLayer(f, [2]) qmllayer(K.ones([2]))

what about this, HardwareLayer is a non jit version wrapper for keras layer, after you debug, you can switch back to KerasLayer

[wcqc]

One quick additional question:

If the JAX backend is used, would it be possible to use jax.lax.scan or jax.lax.fori_loop in place of standard for loops to speed up circuit compile especially for deep circuits?

[refraction-ray]

yes, of course, pls see this nice example: https://github.com/tencent-quantum-lab/tensorcircuit/blob/master/examples/jax_scan_jit_acc.py