Keras: only display NN-architecture without model memory allocation

Question

My goal is to use Keras to visualise the architecture of the model. No training, no inference.

For example if I just want to visualise the graph of the network of the classic VGG16-model with model = create_vgg_like_model() then Keras starts to pre-allocate the memory needed and the program crash (see image).

enter image description here

PC specs: CPU with 4 cores, 8GB RAM. Due to the "huge" amount of parameters of the model it uses even the swap partition.

---

I tried to play around with

cpu_devices = tf.config.list_physical_devices('CPU')
if cpu_devices:
 try:
 # avoid allocating all memory on the device
 tf.config.set_visible_devices(cpu_devices, 'CPU')
 tf.config.experimental.set_memory_growth(cpu_devices[0], True)
 model = create_vgg_like_model()
 dot = tf.keras.utils.model_to_dot(model) # example of display
 except ValueError as e:
 print(e)

but without success, it raises

ValueError: Cannot set memory growth on non-GPU and non-Pluggable devices.

Is there a way

• to post-pone the memory allocation on CPU and/or on GPU
• to implement a Model-like object containing only "shallow" information of the network such as layer name, input, output, number of parameters, ...?

• cards: it happens both with keras and tensorflow.keras
• mcvincekova: imho initialising the model with no weights should do the trick; so -> from keras.applications import VGG16 -> model = VGG16(weights=None) . You can get the model config with layer definitions from there like so: model.get_config().
• cards: weights=None is accepted only by the pre-build models from keras.applications and what it does is a "random initialization" of the weights.
• mcvincekova: Correct, it didn't hit me you will want to go beyond predefined models. In that case - as far as I know, there is no way in keras to get precise architecture stats (i.e. input/output & num. of params details per layer) without building/compiling the model. There is a lazy module in pytorch that might be helpful here, but even there all params/weights are still initialised as soon as you do the first forward pass.
• cards: @mcvincekova but maybe it is possible to disable that behaviour by overwrite a method of the class Model / Sequential

Answer 1

The problem is it didn't detect your gpu that's why it is throwing :ValueError: Cannot set memory growth on non-GPU and non-Pluggable devices. To avoid this issue try removing the line which is causing the issue :

tf.config.experimental.set_memory_growth(cpu_devices[0], True)

Coment it out and it might do the job just fine or if you want to use this try using a proper gpu.

• cards: I mentioned the specs of the machine and no gpu is mentioned. The question is, indeed, how to avoid/postpone that memory allocation