[common] Add beginning of lib.

Author: Arthur Douillard

inclearn/lib/__init__.py

@@ -0,0 +1 @@
+from . import callbacks, metrics

inclearn/lib/callbacks.py

@@ -0,0 +1,38 @@
+import torch
+
+
+class GaussianNoiseAnnealing:
+    """Add gaussian noise to the gradients.
+
+    Add gaussian noise to the gradients with the given mean & std. The std will
+    decrease at each batch up to 0.
+
+    # References:
+    - Adding Gradient Noise Improves Learning for Very Deep Networks
+    - https://arxiv.org/abs/1511.06807
+
+    :param eta: TODO
+    :param gamma: Decay rate.
+    """
+    def __init__(self, parameters, eta=0.3, gamma=0.55):
+        self._parameters = parameters
+        self._eta = eta
+        self._gamma = gamma
+
+        self._iteration = 0
+
+    def add_noise(self):
+        variance = self._eta / ((1 + self._iteration) ** self._gamma)
+
+        for param in self._parameters:
+            # L2 regularization on gradients
+            param.grad.add_(0.0001, torch.norm(param.grad, p=2))
+
+            # Noise on gradients:
+            noise = torch.randn(param.grad.shape, device=param.grad.device) * variance
+            param.grad.add_(noise)
+
+            param.grad.clamp_(min=-5, max=5)
+
+    def step(self):
+        self._iteration += 1

inclearn/lib/metrics.py

@@ -0,0 +1,59 @@
+import numpy as np
+
+
+class MetricLogger:
+    def __init__(self):
+        self._accuracy_per_task = []
+        self._accuracy = []
+        self._incremental_accuracy = []
+
+    def add_task(ypred, ytrue, task_name):
+        pass
+
+
+
+def compute_metrics(ypred, ytrue, task_size=10):
+    metrics = {}
+
+    metrics["accuracy"] = accuracy(ypred, ytrue, task_size=task_size)
+    metrics["incremental_accuracy"] = incremental_accuracy(metrics["accuracy"])
+
+
+def accuracy(ypred, ytrue, task_size=10):
+    """Computes accuracy for the whole test & per task.
+
+    :param ypred: The predictions array.
+    :param ytrue: The ground-truth array.
+    :param task_size: The size of the task.
+    :return: A dictionnary.
+    """
+    all_acc = {}
+
+    all_acc["total"] = round((ypred == ytrue).sum() / len(ytrue), 3)
+
+    for class_id in range(0, np.max(ytrue), task_size):
+        idxes = np.where(
+                np.logical_and(ytrue >= class_id, ytrue < class_id + task_size)
+        )[0]
+
+        label = "{}-{}".format(
+                str(class_id).rjust(2, "0"),
+                str(class_id + task_size - 1).rjust(2, "0")
+        )
+        all_acc[label] = round((ypred[idxes] == ytrue[idxes]).sum() / len(idxes), 3)
+
+    return all_acc
+
+
+
+def incremental_accuracy(acc_dict):
+    """Computes the average incremental accuracy as described in iCaRL.
+
+    It is the average of the current task accuracy (tested on 0-X) with the
+    previous task accuracy.
+
+    :param acc_dict: A dict TODO
+    """
+    v, c = 0., 0
+
+    pass