added crafter example

examples/crafter/README.md

@@ -0,0 +1,6 @@
+# Using AgentKit for Crafter
+
+Run
+```
+python main.py
+```

examples/crafter/baseline.py

@@ -0,0 +1,124 @@
+import os
+os.environ["MINEDOJO_HEADLESS"]="1"
+import argparse
+import numpy as np
+from tqdm import tqdm
+import gym
+import crafter
+from crafter_description import describe_frame, action_list, match_act
+from functools import partial
+from utils import get_ctxt, describe_achievements
+MANUAL = get_ctxt()
+
+parser = argparse.ArgumentParser()
+parser.add_argument('--llm_name', type=str, default='yintat-all-gpt-4', help='Name of the LLM')
+
+args = parser.parse_args()
+
+LLM_name = args.llm_name
+
+env = crafter.Env(area=(256, 256))
+action_space = env.action_space
+
+# Replace with your own LLM API.
+# Note: query_model takes two arguments: 1) message in openai chat completion form (list of dictionaries), 
+#                                        2) an index to indicate where the message should be truncated if the length exceeds LLM context length.
+from llm_api import get_query
+query_model = partial(get_query(LLM_name), max_gen=2048)
+
+def compose_ingame_prompt(info, question, past_qa=[]):
+    messages = [
+        {"role": "system", "content" : "You’re a player trying to play the game."}
+    ]
+
+    if len(info['manual'])>0:
+        messages.append({"role": "system", "content": info['manual']})
+
+    messages.append({"role": "system", "content": "{}".format(info['obs'])})
+
+    if len(past_qa)>0:
+        for q,a in past_qa:
+            messages.append({"role": "user", "content": q})
+            messages.append({"role": "assistant", "content": a})
+
+    messages.append({"role": "user", "content": question})
+
+    return messages, 1 # This is the index of the history, we will truncate the history if it is too long for LLM
+
+questions=[
+        "What is the best action to take? Let's think step by step, ",
+        "Choose the best executable action from the list of all actions. Write the exact chosen action."
+    ]
+
+def run():
+    env = crafter.Env(area=(256, 256))
+    env_steps = 1000000
+    num_iter = 2
+
+    rewards = []
+    progresses = []
+    for eps in tqdm(range(num_iter), desc="Evaluating LLM {}".format(LLM_name)):
+        import wandb
+        wandb.init(project="Crafter_baseline", config={"LLM": LLM_name, "eps": eps, "num_iter": num_iter, "env_steps": env_steps})
+        step = 0
+        trajectories = []
+        qa_history = []
+        progress = [0]
+        reward = 0
+        rewards = []
+        done=False
+
+        columns=["Context", "Step", "OBS", "Score", "Reward", "Total Reward"] + questions + ["Action"]
+        wandb_table = wandb.Table(columns=columns)
+
+        env.reset()
+        a = action_list.index("noop")
+        obs, reward, done, info = env.step(a)
+
+        while step < env_steps:
+            last_act_desc, desc = describe_frame(info, 1)
+            if len(trajectories)>0:
+                trajectories[-1][1] = last_act_desc
+            trajectories.append([step, None, desc])
+            text_obs = "\n\n".join(["== Gamestep {}{} ==\n\n".format(i, "" if i!=trajectories[-1][0] else " (current)",) + "{}{}".format(d, "\n\nAction:\n{}".format(a) if a is not None else "") for i, a, d in trajectories[-2:]])
+            info['obs'] = text_obs
+            info['manual'] = describe_achievements(info, MANUAL)
+            info['reward'] = reward
+            info['score'] = sum(rewards)
+            new_row = [info['manual'], step, info['obs'], info['score'], reward, sum(rewards)]
+            wandb.log({"metric/total_reward".format(eps): sum(rewards), 
+                       "metric/score".format(eps): info['score'],
+                       "metric/reward".format(eps): reward,
+                       })
+
+            if done:
+                break
+
+            qa_history = []
+            for question in questions:
+                prompt = compose_ingame_prompt(info, question, qa_history)
+                answer, _ = query_model(*prompt)
+                qa_history.append((question, answer))
+                new_row.append(answer)
+                answer_act = answer
+
+            a, _, _ = match_act(answer_act)
+            if a is None:
+                a = action_list.index("noop")
+            new_row.append(action_list[a])
+            obs, reward, done, info = env.step(a)
+            rewards.append(reward)
+
+            step += 1
+            wandb_table.add_data(*new_row)
+
+        progresses.append(np.max(progress))
+        wandb.log({"rollout/rollout-{}".format(eps): wandb_table, 
+                "final/total_reward":sum(rewards),
+                "final/episodic_step":step,
+                "final/eps":eps,
+                })
+        del wandb_table
+        wandb.finish()
+
+run()