Skip to main content

How to clone/duplicate a TensorFlow Probability neural network model

I have a TensorFlow Probability model that is built similar to models described in this YouTube Video.

I'm using

  • python==3.8.11
  • tensorflow==2.10.0
  • tensorflow-probability==0.18.0

Here's the code to build the model:

def posterior_mean_field(kernel_size: int, bias_size: int, dtype: Any) -> tf.keras.Model:
    n = kernel_size + bias_size
    c = np.log(np.expm1(1.))

    return tf.keras.Sequential([
        tfp.layers.VariableLayer(2 * n, dtype=dtype),
        tfp.layers.DistributionLambda(lambda t: tfd.Independent(tfd.Normal(loc=t[..., :n],
                                                                           scale=1e-5 + tf.nn.softplus(c + t[..., n:])),
                                                                reinterpreted_batch_ndims=1)),
    ])


def prior_trainable(kernel_size: int, bias_size: int, dtype: Any) -> tf.keras.Model:
    n = kernel_size + bias_size
    return tf.keras.Sequential([
        tfp.layers.VariableLayer(n, dtype=dtype),
        tfp.layers.DistributionLambda(lambda t: tfd.Independent(
            tfd.Normal(loc=t, scale=1),
            reinterpreted_batch_ndims=1)),
    ])


def build_model():
    model = keras.Sequential([
        tfp.layers.DenseVariational(64, activation='relu', input_shape=[len(train_dataset.keys())],
                                    make_posterior_fn=posterior_mean_field,
                                    make_prior_fn=prior_trainable),
        layers.Dense(64, activation='relu'),
        layers.Dense(1),
    ])
    optimizer = tf.keras.optimizers.RMSprop(0.001)
    model.compile(loss='mse', optimizer=optimizer, metrics=['mae', 'mse'])
    return model


model = build_model()
model.build((3, 10))

When I remove the TensorFlow Probability layer (1st layer) in the model, I can clone the model and copy its weights like this:

import copy
from tensorflow.keras.models import clone_model
model_weights = copy.deepcopy(model.get_weights())
model_copy = clone_model(model)
model_copy.set_weights(model_weights)

However, when the TensorFlow Probability layer is present I get this error:

Traceback (most recent call last):
  File "/Users/jisom/opt/miniconda3/envs/ic-hours/lib/python3.8/site-packages/IPython/core/interactiveshell.py", line 3398, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "<ipython-input-6-349eb0e7c1e5>", line 1, in <cell line: 1>
    model_new = clone_model(model)
  File "/Users/jisom/opt/miniconda3/envs/ic-hours/lib/python3.8/site-packages/keras/models.py", line 448, in clone_model
    return _clone_sequential_model(
  File "/Users/jisom/opt/miniconda3/envs/ic-hours/lib/python3.8/site-packages/keras/models.py", line 326, in _clone_sequential_model
    if isinstance(layer, InputLayer) else layer_fn(layer))
  File "/Users/jisom/opt/miniconda3/envs/ic-hours/lib/python3.8/site-packages/keras/models.py", line 56, in _clone_layer
    return layer.__class__.from_config(layer.get_config())
  File "/Users/jisom/opt/miniconda3/envs/ic-hours/lib/python3.8/site-packages/keras/engine/base_layer.py", line 727, in get_config
    raise NotImplementedError('Layer %s has arguments in `__init__` and '
NotImplementedError: Layer DenseVariational has arguments in `__init__` and therefore must override `get_config`.

I can see some information about how to deal with this error in this StackOverflow question, but in that question there's a custom-built transformer class that can be modified. I'm trying to use the clone_model function in keras, which I don't directly control. And, the error seems to be coming from the TFP DenseVariational layer that doesn't override get_config. How can I clone/duplicate a model, including its weights, if the model includes TensorFlow Probability layers?

Or should I be creating an issue in the TensorFlow Probability Github repo to get this functionality added to the DenseVariational layer?



source https://stackoverflow.com/questions/73832532/how-to-clone-duplicate-a-tensorflow-probability-neural-network-model

Labels

Labels:

Comments

Post a Comment

Popular posts from this blog

ValueError: X has 10 features, but LinearRegression is expecting 1 features as input

So, I am trying to predict the model but its throwing error like it has 10 features but it expacts only 1. So I am confused can anyone help me with it? more importantly its not working for me when my friend runs it. It works perfectly fine dose anyone know the reason about it? cv = KFold(n_splits = 10) all_loss = [] for i in range(9): # 1st for loop over polynomial orders poly_order = i X_train = make_polynomial(x, poly_order) loss_at_order = [] # initiate a set to collect loss for CV for train_index, test_index in cv.split(X_train): print('TRAIN:', train_index, 'TEST:', test_index) X_train_cv, X_test_cv = X_train[train_index], X_test[test_index] t_train_cv, t_test_cv = t[train_index], t[test_index] reg.fit(X_train_cv, t_train_cv) loss_at_order.append(np.mean((t_test_cv - reg.predict(X_test_cv))**2)) # collect loss at fold all_loss.append(np.mean(loss_at_order)) # collect loss at order plt.plot(np.log(al...
Post a Comment
Read more

Sorting large arrays of big numeric stings

I was solving bigSorting() problem from hackerrank: Consider an array of numeric strings where each string is a positive number with anywhere from to digits. Sort the array's elements in non-decreasing, or ascending order of their integer values and return the sorted array. I know it works as follows: def bigSorting(unsorted): return sorted(unsorted, key=int) But I didnt guess this approach earlier. Initially I tried below: def bigSorting(unsorted): int_unsorted = [int(i) for i in unsorted] int_sorted = sorted(int_unsorted) return [str(i) for i in int_sorted] However, for some of the test cases, it was showing time limit exceeded. Why is it so? PS: I dont know exactly what those test cases were as hacker rank does not reveal all test cases. source https://stackoverflow.com/questions/73007397/sorting-large-arrays-of-big-numeric-stings
Post a Comment
Read more

How to load Javascript with imported modules?

I am trying to import modules from tensorflowjs, and below is my code. test.html <!DOCTYPE html> <html lang="en"> <head> <meta charset="UTF-8"> <title>Document</title </head> <body> <script src="https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@2.0.0/dist/tf.min.js"></script> <script type="module" src="./test.js"></script> </body> </html> test.js import * as tf from "./node_modules/@tensorflow/tfjs"; import {loadGraphModel} from "./node_modules/@tensorflow/tfjs-converter"; const MODEL_URL = './model.json'; const model = await loadGraphModel(MODEL_URL); const cat = document.getElementById('cat'); model.execute(tf.browser.fromPixels(cat)); Besides, I run the server using python -m http.server in my command prompt(Windows 10), and this is the error prompt in the console log of my browser: Failed to loa...
Post a Comment
Read more