Handle multi-dimension inputs

README.md

@@ -38,7 +38,7 @@ Neumann, Anas. (2023). Simple Python and TensorFlow implementation of the optimi
 
 ## Complete code
 ```python
-def MAML(model, alpha=0.005, beta=0.005, optimizer=keras.optimizers.SGD, c_loss=keras.losses.mse, f_loss=keras.losses.MeanSquaredError(), meta_epochs=100, meta_tasks_per_epoch=[10, 30], validation_split=0.2, k_folds=0, tasks=[], cumul=False):
+def MAML(model, alpha=0.005, beta=0.005, optimizer=keras.optimizers.SGD, c_loss=keras.losses.mse, f_loss=keras.losses.MeanSquaredError(), meta_epochs=100, meta_tasks_per_epoch=[10, 30], inputs_dimension=1, validation_split=0.2, k_folds=0, tasks=[], cumul=False):
  """
  Simple MAML algorithm implementation for supervised regression.
  :param model: A Keras model to be trained using MAML.
@@ -48,6 +48,7 @@ def MAML(model, alpha=0.005, beta=0.005, optimizer=keras.optimizers.SGD, c_loss=
  :param c_loss: Loss function for calculating training loss.
  :param meta_epochs: Number of meta-training epochs.
  :param meta_tasks_per_epoch: Range of tasks to sample per epoch.
+ :param inputs_dimension: the input dimension (for sequence-to-sequence models).
  :param validation_split: Ratio of data to use for validation in each task (could be fixed or random between two values).
  :param k_folds: cross-validation with k_folds each time a task is called for meta-learning.
  :param tasks: List of tasks for meta-training.
@@ -56,37 +57,44 @@ def MAML(model, alpha=0.005, beta=0.005, optimizer=keras.optimizers.SGD, c_loss=
  """
  if tf.config.list_physical_devices('GPU'):
  with tf.device('/GPU:0'):
- return _MAML_compute(model, alpha, beta, optimizer, c_loss, f_loss, meta_epochs, meta_tasks_per_epoch, validation_split, k_folds, tasks, cumul)
+ return _MAML_compute(model, alpha, beta, optimizer, c_loss, f_loss, meta_epochs, meta_tasks_per_epoch, inputs_dimension, validation_split, k_folds, tasks, cumul)
  else:
- return _MAML_compute(model, alpha, beta, optimizer, c_loss, f_loss, meta_epochs, meta_tasks_per_epoch, validation_split, k_folds, tasks, cumul)
+ return _MAML_compute(model, alpha, beta, optimizer, c_loss, f_loss, meta_epochs, meta_tasks_per_epoch, inputs_dimension, validation_split, k_folds, tasks, cumul)
 
-def _build_task(t, validation_split, k_folds):
+def _build_task(t, inputs_dimension, validation_split, k_folds):
  """
  Build task t by splitting train_input, test_input, train_target, test_target if it's not already done.
  This function is flexible and handle both randon validation_splits and k_folds.
  :param t: a task to learn during the meta-pre-training stage
+ :param inputs_dimension: the input dimension (for sequence-to-sequence models).
  :param validation_split: optional ratio of data to use for training in each task (could be fixed or random between two values).
  :param k_folds: optional cross-validation with k_folds each time a task is called for meta-learning.
  :return: train_input, test_input, train_target, test_target
  """
  if "train" in t and "test" in t:
- return t["train"]["inputs"], t["test"]["inputs"], t["train"]["target"], t["test"]["target"]
+ train_input = t["train"]["inputs"] if inputs_dimension<=1 else [t["train"]["inputs"] for d in range(inputs_dimension)]
+ test_input = t["test"]["inputs"] if inputs_dimension<=1 else [t["test"]["inputs"] for d in range(inputs_dimension)]
+ return t["train"]["inputs"], t["test"]["inputs"], t["train"]["target"], t["test"]["target"] 
  elif k_folds>0:
  fold = random.randint(0, k_folds-1)
  fold_size = (len(t["inputs"]) // k_folds)
  v_start = fold * fold_size
  v_end = (fold + 1) * fold_size if fold < k_folds - 1 else len(t["inputs"])
- train_input, train_target = np.concatenate((t["inputs"][:v_start], t["inputs"][v_end:]), axis=0), np.concatenate((t["target"][:v_start], t["target"][v_end:]), axis=0)
- train_target, test_target = t["inputs"][v_start:v_end], t["target"][v_start:v_end]
+ t_i = np.concatenate((t["inputs"][:v_start], t["inputs"][v_end:]), axis=0)
+ train_input = t_i if inputs_dimension<=1 else [t_i for d in range(inputs_dimension)]
+ test_input = t["inputs"][v_start:v_end] if inputs_dimension<=1 else [t["inputs"][v_start:v_end] for d in range(inputs_dimension)]
+ train_target = np.concatenate((t["target"][:v_start], t["target"][v_end:]), axis=0)
+ test_target = t["target"][v_start:v_end]
  return train_input, test_input, train_target, test_target
  else:
  v = random.uniform(validation_split[0], validation_split[1]) if isinstance(validation_split,list) else validation_split
- split_idx = int(len(t["inputs"]) * v) 
- train_input, test_input = t["inputs"][:split_idx], t["inputs"][split_idx:]
+ split_idx = int(len(t["inputs"]) * v)
+ train_input = t["inputs"][:split_idx] if inputs_dimension<=1 else [t["inputs"][:split_idx] for d in range(inputs_dimension)]
+ test_input = t["inputs"][split_idx:] if inputs_dimension<=1 else [t["inputs"][split_idx:] for d in range(inputs_dimension)]
  train_target, test_target = t["target"][:split_idx], t["target"][split_idx:]
  return train_input, test_input, train_target, test_target
 
-def _MAML_compute(model, alpha, beta, optimizer, c_loss, f_loss, meta_epochs, meta_tasks_per_epoch, validation_split, k_folds, tasks, cumul):
+def _MAML_compute(model, alpha, beta, optimizer, c_loss, f_loss, meta_epochs, meta_tasks_per_epoch, inputs_dimension, validation_split, k_folds, tasks, cumul):
  log_step = meta_epochs // 10 if meta_epochs > 10 else 1
  optim_test=optimizer(learning_rate=alpha)
  optim_test.build(model.trainable_variables)
@@ -104,7 +112,7 @@ def _MAML_compute(model, alpha, beta, optimizer, c_loss, f_loss, meta_epochs, me
  model_copy.compile(loss=f_loss, optimizer=optim_train)
  for _ in range(num_tasks_sampled):
  t = random.choice(tasks)
- train_input, test_input, train_target, test_target = _build_task(t, validation_split, k_folds)
+ train_input, test_input, train_target, test_target = _build_task(t, input_dimenstion, validation_split, k_folds)
 
  # 1. Inner loop: Update the model copy on the current task
  with tf.GradientTape(watch_accessed_variables=False) as train_tape:

setup.py

@@ -5,7 +5,7 @@
 
 setup(
  name='simplemaml',
- version='1.2.6',
+ version='1.2.7',
  description='A generic Python and TensorFlow function that implements a simple version of the "Model-Agnostic Meta-Learning (MAML) Algorithm for Fast Adaptation of Deep Networks" as designed by Chelsea Finn et al. 2017',
  long_description=long_description,
  long_description_content_type='text/markdown',

simplemaml.py

@@ -3,7 +3,7 @@
 import numpy as np
 import random
 
-def MAML(model, alpha=0.005, beta=0.005, optimizer=keras.optimizers.SGD, c_loss=keras.losses.mse, f_loss=keras.losses.MeanSquaredError(), meta_epochs=100, meta_tasks_per_epoch=[10, 30], validation_split=0.2, k_folds=0, tasks=[], cumul=False):
+def MAML(model, alpha=0.005, beta=0.005, optimizer=keras.optimizers.SGD, c_loss=keras.losses.mse, f_loss=keras.losses.MeanSquaredError(), meta_epochs=100, meta_tasks_per_epoch=[10, 30], inputs_dimension=1, validation_split=0.2, k_folds=0, tasks=[], cumul=False):
  """
  Simple MAML algorithm implementation for supervised regression.
  :param model: A Keras model to be trained using MAML.
@@ -13,6 +13,7 @@ def MAML(model, alpha=0.005, beta=0.005, optimizer=keras.optimizers.SGD, c_loss=
  :param c_loss: Loss function for calculating training loss.
  :param meta_epochs: Number of meta-training epochs.
  :param meta_tasks_per_epoch: Range of tasks to sample per epoch.
+ :param inputs_dimension: the input dimension (for sequence-to-sequence models).
  :param validation_split: Ratio of data to use for validation in each task (could be fixed or random between two values).
  :param k_folds: cross-validation with k_folds each time a task is called for meta-learning.
  :param tasks: List of tasks for meta-training.
@@ -21,37 +22,44 @@ def MAML(model, alpha=0.005, beta=0.005, optimizer=keras.optimizers.SGD, c_loss=
  """
  if tf.config.list_physical_devices('GPU'):
  with tf.device('/GPU:0'):
- return _MAML_compute(model, alpha, beta, optimizer, c_loss, f_loss, meta_epochs, meta_tasks_per_epoch, validation_split, k_folds, tasks, cumul)
+ return _MAML_compute(model, alpha, beta, optimizer, c_loss, f_loss, meta_epochs, meta_tasks_per_epoch, inputs_dimension, validation_split, k_folds, tasks, cumul)
  else:
- return _MAML_compute(model, alpha, beta, optimizer, c_loss, f_loss, meta_epochs, meta_tasks_per_epoch, validation_split, k_folds, tasks, cumul)
+ return _MAML_compute(model, alpha, beta, optimizer, c_loss, f_loss, meta_epochs, meta_tasks_per_epoch, inputs_dimension, validation_split, k_folds, tasks, cumul)
 
-def _build_task(t, validation_split, k_folds):
+def _build_task(t, inputs_dimension, validation_split, k_folds):
  """
  Build task t by splitting train_input, test_input, train_target, test_target if it's not already done.
  This function is flexible and handle both randon validation_splits and k_folds.
  :param t: a task to learn during the meta-pre-training stage
+ :param inputs_dimension: the input dimension (for sequence-to-sequence models).
  :param validation_split: optional ratio of data to use for training in each task (could be fixed or random between two values).
  :param k_folds: optional cross-validation with k_folds each time a task is called for meta-learning.
  :return: train_input, test_input, train_target, test_target
  """
  if "train" in t and "test" in t:
+ train_input = t["train"]["inputs"] if inputs_dimension<=1 else [t["train"]["inputs"] for d in range(inputs_dimension)]
+ test_input = t["test"]["inputs"] if inputs_dimension<=1 else [t["test"]["inputs"] for d in range(inputs_dimension)]
  return t["train"]["inputs"], t["test"]["inputs"], t["train"]["target"], t["test"]["target"] 
  elif k_folds>0:
  fold = random.randint(0, k_folds-1)
  fold_size = (len(t["inputs"]) // k_folds)
  v_start = fold * fold_size
  v_end = (fold + 1) * fold_size if fold < k_folds - 1 else len(t["inputs"])
- train_input, train_target = np.concatenate((t["inputs"][:v_start], t["inputs"][v_end:]), axis=0), np.concatenate((t["target"][:v_start], t["target"][v_end:]), axis=0)
- train_target, test_target = t["inputs"][v_start:v_end], t["target"][v_start:v_end]
+ t_i = np.concatenate((t["inputs"][:v_start], t["inputs"][v_end:]), axis=0)
+ train_input = t_i if inputs_dimension<=1 else [t_i for d in range(inputs_dimension)]
+ test_input = t["inputs"][v_start:v_end] if inputs_dimension<=1 else [t["inputs"][v_start:v_end] for d in range(inputs_dimension)]
+ train_target = np.concatenate((t["target"][:v_start], t["target"][v_end:]), axis=0)
+ test_target = t["target"][v_start:v_end]
  return train_input, test_input, train_target, test_target
  else:
  v = random.uniform(validation_split[0], validation_split[1]) if isinstance(validation_split,list) else validation_split
- split_idx = int(len(t["inputs"]) * v) 
- train_input, test_input = t["inputs"][:split_idx], t["inputs"][split_idx:]
+ split_idx = int(len(t["inputs"]) * v)
+ train_input = t["inputs"][:split_idx] if inputs_dimension<=1 else [t["inputs"][:split_idx] for d in range(inputs_dimension)]
+ test_input = t["inputs"][split_idx:] if inputs_dimension<=1 else [t["inputs"][split_idx:] for d in range(inputs_dimension)]
  train_target, test_target = t["target"][:split_idx], t["target"][split_idx:]
  return train_input, test_input, train_target, test_target
 
-def _MAML_compute(model, alpha, beta, optimizer, c_loss, f_loss, meta_epochs, meta_tasks_per_epoch, validation_split, k_folds, tasks, cumul):
+def _MAML_compute(model, alpha, beta, optimizer, c_loss, f_loss, meta_epochs, meta_tasks_per_epoch, inputs_dimension, validation_split, k_folds, tasks, cumul):
  log_step = meta_epochs // 10 if meta_epochs > 10 else 1
  optim_test=optimizer(learning_rate=alpha)
  optim_test.build(model.trainable_variables)
@@ -69,7 +77,7 @@ def _MAML_compute(model, alpha, beta, optimizer, c_loss, f_loss, meta_epochs, me
  model_copy.compile(loss=f_loss, optimizer=optim_train)
  for _ in range(num_tasks_sampled):
  t = random.choice(tasks)
- train_input, test_input, train_target, test_target = _build_task(t, validation_split, k_folds)
+ train_input, test_input, train_target, test_target = _build_task(t, input_dimenstion, validation_split, k_folds)
 
  # 1. Inner loop: Update the model copy on the current task
  with tf.GradientTape(watch_accessed_variables=False) as train_tape: