test: [RLOS2023] add test for slate (#4629)

* test: add test for slate

* test: test cleanup and slate test update

* test: minor cleanup and change assert_loss function to equal instead of lower

python/tests/test_framework/assert_job.py

@@ -68,11 +68,17 @@ def assert_loss(job, **kwargs):
     assert job.status == ExecutionStatus.Success, "job should be successful"
     assert type(job[0].loss) == float, "loss should be an float"
     decimal = kwargs.get("decimal", 2)
-    if job[0].loss < kwargs["expected_loss"]:
-        return
     assert_almost_equal(job[0].loss, kwargs["expected_loss"], decimal=decimal)
 
 
+def assert_loss_below(job, **kwargs):
+    assert job.status == ExecutionStatus.Success, "job should be successful"
+    assert type(job[0].loss) == float, "loss should be an float"
+    assert (
+        job[0].loss <= kwargs["expected_loss"]
+    ), f"loss should be below {kwargs['expected_loss']}"
+
+
 def assert_prediction_with_generated_data(job, **kwargs):
     assert job.status == ExecutionStatus.Success, "job should be successful"
     expected_class = []

python/tests/test_framework/cb/data_generation.py

@@ -17,8 +17,7 @@ def generate_cb_data(
     num_actions,
     reward_function,
     logging_policy,
-    no_context=1,
-    context_name=None,
+    context_name=["1"],
 ):
 
     dataFile = f"cb_test_{num_examples}_{num_actions}_{num_features}.txt"
@@ -32,16 +31,20 @@ def generate_cb_data(
     features = [f"feature{index}" for index in range(1, num_features + 1)]
     with open(os.path.join(script_directory, dataFile), "w") as f:
         for _ in range(num_examples):
+            no_context = len(context_name)
             if no_context > 1:
                 context = random.randint(1, no_context)
-                if not context_name:
-                    context_name = [f"{index}" for index in range(1, no_context + 1)]
+            else:
+                context = 1
 
             def return_cost_probability(chosen_action, context=1):
-                cost = reward_function_obj(
+                cost = -reward_function_obj(
                     chosen_action, context, **reward_function["params"]
                 )
+                if "params" not in logging_policy:
+                    logging_policy["params"] = {}
                 logging_policy["params"]["chosen_action"] = chosen_action
+                logging_policy["params"]["num_actions"] = num_actions
                 probability = logging_policy_obj(**logging_policy["params"])
                 return cost, probability
 

python/tests/test_framework/classification/data_generation.py

@@ -7,20 +7,20 @@
 
 
 def generate_classification_data(
-    num_sample,
+    num_example,
     num_classes,
     num_features,
     classify_func,
     bounds=None,
 ):
-    dataFile = f"classification_{num_classes}_{num_features}_{num_sample}.txt"
+    dataFile = f"classification_{num_classes}_{num_features}_{num_example}.txt"
     classify_func_obj = get_function_object(
         "classification.classification_functions", classify_func["name"]
     )
     if not bounds:
         bounds = [[0, 1] for _ in range(num_features)]
     with open(os.path.join(script_directory, dataFile), "w") as f:
-        for _ in range(num_sample):
+        for _ in range(num_example):
             x = [
                 random.uniform(bounds[index][0], bounds[index][1])
                 for index in range(num_features)

python/tests/test_framework/slate/action_space.py

@@ -0,0 +1,7 @@
+def new_action_after_threshold(**kwargs):
+    iteration = kwargs.get("iteration", 0)
+    threshold = kwargs.get("threshold", 0)
+    # before iteration 500, it is sunny and after it is raining
+    if iteration > threshold:
+        return kwargs["after"]
+    return kwargs["before"]

python/tests/test_framework/slate/assert_job.py

@@ -0,0 +1,43 @@
+from numpy.testing import assert_allclose, assert_almost_equal
+from vw_executor.vw import ExecutionStatus
+import numpy as np
+
+
+def majority_close(arr1, arr2, rtol, atol, threshold):
+    # Check if the majority of elements are close
+    close_count = np.count_nonzero(np.isclose(arr1, arr2, rtol=rtol, atol=atol))
+    return close_count >= len(arr1) * threshold
+
+
+def assert_prediction(job, **kwargs):
+    assert job.status == ExecutionStatus.Success, "job should be successful"
+    atol = kwargs.get("atol", 10e-8)
+    rtol = kwargs.get("rtol", 10e-5)
+    threshold = kwargs.get("threshold", 0.9)
+    expected_value = kwargs["expected_value"]
+    predictions = job.outputs["-p"]
+    res = []
+    with open(predictions[0], "r") as f:
+        exampleRes = []
+        while True:
+            line = f.readline()
+            if not line:
+                break
+            if line.count(":") == 0:
+                res.append(exampleRes)
+                exampleRes = []
+                continue
+            slotRes = [0] * line.count(":")
+            slot = line.split(",")
+            for i in range(len(slot)):
+                actionInd = int(slot[i].split(":")[0])
+                slotRes[i] = float(slot[actionInd].split(":")[1])
+            exampleRes.append(slotRes)
+
+        assert majority_close(
+            res,
+            [expected_value] * len(res),
+            rtol=rtol,
+            atol=atol,
+            threshold=threshold,
+        ), f"predicted value should be {expected_value}, \n actual values are {res}"

python/tests/test_framework/slate/data_generation.py

@@ -0,0 +1,70 @@
+import random
+import os
+from test_helper import get_function_object
+
+script_directory = os.path.dirname(os.path.realpath(__file__))
+random.seed(10)
+
+
+def generate_slate_data(
+    num_examples,
+    reward_function,
+    logging_policy,
+    action_space,
+    context_name=["1"],
+):
+
+    action_space_obj = get_function_object("slate.action_space", action_space["name"])
+
+    reward_function_obj = get_function_object(
+        "slate.reward_functions", reward_function["name"]
+    )
+    logging_policy_obj = get_function_object(
+        "slate.logging_policies", logging_policy["name"]
+    )
+
+    def return_cost_probability(chosen_action, chosen_slot, context):
+        cost = -reward_function_obj(
+            chosen_action, context, chosen_slot, **reward_function["params"]
+        )
+        logging_policy["params"]["num_action"] = num_actions[chosen_slot - 1]
+        logging_policy["params"]["chosen_action"] = chosen_action
+        probability = logging_policy_obj(**logging_policy["params"])
+        return cost, probability
+
+    dataFile = f"slate_test_{num_examples}_{generate_slate_data.__name__}.txt"
+    with open(os.path.join(script_directory, dataFile), "w") as f:
+        for i in range(num_examples):
+            action_space["params"]["iteration"] = i
+            action_spaces = action_space_obj(**action_space["params"])
+            num_slots = len(action_spaces)
+            num_actions = [len(slot) for slot in action_spaces]
+            slot_name = [f"slot_{index}" for index in range(1, num_slots + 1)]
+            chosen_actions = []
+            num_context = len(context_name)
+            if num_context > 1:
+                context = random.randint(1, num_context)
+            else:
+                context = 1
+            for s in range(num_slots):
+                chosen_actions.append(random.randint(1, num_actions[s]))
+            chosen_actions_cost_prob = [
+                return_cost_probability(action, slot + 1, context)
+                for slot, action in enumerate(chosen_actions)
+            ]
+            total_cost = sum([cost for cost, _ in chosen_actions_cost_prob])
+
+            f.write(f"slates shared {total_cost} |User {context_name[context-1]}\n")
+            # write actions
+            for ind, slot in enumerate(action_spaces):
+                for a in slot:
+                    f.write(
+                        f"slates action {ind} |Action {a}\n",
+                    )
+
+            for s in range(num_slots):
+                f.write(
+                    f"slates slot {chosen_actions[s]}:{chosen_actions_cost_prob[s][1]} |Slot {slot_name[s]}\n"
+                )
+            f.write("\n")
+    return os.path.join(script_directory, dataFile)

python/tests/test_framework/slate/logging_policies.py

@@ -0,0 +1,3 @@
+def even_probability(chosen_action, **kwargs):
+    num_actions = kwargs["num_action"]
+    return round(1 / num_actions, 2)

python/tests/test_framework/slate/reward_functions.py

@@ -0,0 +1,12 @@
+def fixed_reward(chosen_action, context, slot, **kwargs):
+    reward = kwargs["reward"]
+    return reward[slot - 1][chosen_action - 1]
+
+
+def reverse_reward_after_threshold(chosen_action, context, slot, **kwargs):
+    reward = kwargs["reward"]
+    iteration = kwargs.get("iteration", 0)
+    threshold = kwargs.get("threshold", 0)
+    if iteration > threshold:
+        reward = [i[::-1] for i in reward]
+    return reward[slot - 1][chosen_action - 1]