Speed up Keras profiling

[AdrianAlan]

Description

Changes to get_ymodel_keras speeding up (Q)Keras profiling. Instead of compiling a new model on every step the output can be an array as suggested in Keras FAQ.

Type of change

• Other: (non-breaking enhancement)

Tests

• I added new tests in test_trace.
• I tried it with ResNet and I didn't find any issues.
• I have tried it on a simple example of LeNet on T4 on lxplus705:

def get_model():
  model = keras.Sequential()
  model.add(layers.Conv2D(filters=6, kernel_size=(3, 3), activation='relu', input_shape=(32,32,1)))
  model.add(layers.AveragePooling2D())
  model.add(layers.Conv2D(filters=16, kernel_size=(3, 3), activation='relu'))
  model.add(layers.AveragePooling2D())
  model.add(layers.Flatten())
  model.add(layers.Dense(units=120))
  model.add(layers.Dense(units=84, activation='linear'))
  model.add(layers.Dense(units=10, activation='softmax'))
  model.compile()
  return model

warmup = np.random.random((1000, 32, 32, 1))
model = get_model()
for _ in range(10):
  _ = hls4ml.model.profiling.get_ymodel_keras_old(model, warmup)

new, old = [], []
for _ in range(10):
  X = np.random.random((1000, 32, 32, 1))

  start = time.time()
  _trace_new = hls4ml.model.profiling.get_ymodel_keras(model, X)
  end = time.time()
  new.append(end-start)

  start = time.time()
  _trace_old = hls4ml.model.profiling.get_ymodel_keras_old(model, X)
  end = time.time()
  old.append(end-start)

  assert _trace_old.keys() == _trace_new.keys()
  for l in _trace_old.keys():
    assert np.all(_trace_old[l] == _trace_new[l])

print("New implementation: {}".format(np.mean(new)))
print("Old implementation: {}".format(np.mean(old)))

and I got

New implementation: 0.7009074449539184
Old implementation: 1.334557008743286

Checklist

• I have read the guidelines for contributing.
• I have commented my code, particularly in hard-to-understand areas.
• I have made corresponding changes to the documentation.
• My changes generate no new warnings.
• I have installed and run pre-commit on the files I edited or added.
• I have added tests that prove my fix is effective or that my feature works.

[jmduarte]

Looks good thanks @AdrianAlan!