Lim0011's picture
Upload 251 files
85e3d20 verified
raw
history blame
21.1 kB
import glob
import os
import json
import glob
import tiktoken
import pandas as pd
import copy
import numpy as np
import matplotlib.pyplot as plt
import re
import time
def estimate_tokens(path):
enc = tiktoken.encoding_for_model("gpt-4")
prompt_tokens = 0
completed_tokens = 0
num_steps = 0
step_logs = path.replace("trace.json", "../agent_log/*.log")
for file in glob.glob(step_logs):
with open(file, "r") as f:
content = f.read()
if "langchain" not in file:
prompts = re.findall(r"===================prompt=====================" + r"(.*?)" + r"===================.*?response.*?=====================", content, re.DOTALL)
prompt_tokens += sum([len(enc.encode(p)) for p in prompts])
completed = re.findall(r"===================.*?response.*?=====================" + r"(.*?)" + r"===================tokens=====================", content, re.DOTALL)
completed_tokens += sum([len(enc.encode(p)) for p in completed])
else:
prompts = re.findall(r"Prompt after formatting:\n\x1B\[32;1m\x1B\[1;3m" + r"(.*?)" + r"\x1B\[0m\n\n\x1B\[1m> Finished chain.\x1B\[0m\n\x1B\[32;1m\x1B\[1;3m", content, re.DOTALL)
prompt_tokens += sum([len(enc.encode(p)) for p in prompts])
completed = re.findall(r"\x1B\[0m\n\n\x1B\[1m> Finished chain.\x1B\[0m\n\x1B\[32;1m\x1B\[1;3m" + r"(.*?)" + r"Prompt after formatting:\n\x1B\[32;1m\x1B\[1;3m", content, re.DOTALL)
completed_tokens += sum([len(enc.encode(p)) for p in completed])
num_steps = len(json.load(open(path, "r"))["steps"])
try:
total_time = float(open(path.replace("trace.json", "overall_time.txt"), "r").read())
except:
total_time = 0
tool_step_logs = path.replace("trace.json", "tool_logs/*.log")
tool_prompt_tokens = 0
tool_completed_tokens = 0
for file in glob.glob(tool_step_logs):
with open(file, "r") as f:
content = f.read()
if "langchain" not in file:
prompts = re.findall(r"===================prompt=====================" + r"(.*?)" + r"===================.*?response.*?=====================", content, re.DOTALL)
tool_prompt_tokens += sum([len(enc.encode(p)) for p in prompts])
completed = re.findall(r"===================.*?response.*?=====================" + r"(.*?)" + r"===================tokens=====================", content, re.DOTALL)
tool_completed_tokens += sum([len(enc.encode(p)) for p in completed])
else:
prompts = re.findall(r"Prompt after formatting:\n\x1B\[32;1m\x1B\[1;3m" + r"(.*?)" + r"\x1B\[0m\n\n\x1B\[1m> Finished chain.\x1B\[0m\n\x1B\[32;1m\x1B\[1;3m", content, re.DOTALL)
tool_prompt_tokens += sum([len(enc.encode(p)) for p in prompts])
completed = re.findall(r"\x1B\[0m\n\n\x1B\[1m> Finished chain.\x1B\[0m\n\x1B\[32;1m\x1B\[1;3m" + r"(.*?)" + r"Prompt after formatting:\n\x1B\[32;1m\x1B\[1;3m", content, re.DOTALL)
tool_completed_tokens += sum([len(enc.encode(p)) for p in completed])
return prompt_tokens, completed_tokens, tool_prompt_tokens, tool_completed_tokens, num_steps, total_time
def oom_error(path):
log = path.replace("trace.json", "../log")
main_log = path.replace("trace.json", "../agent_log/main_log")
message = "CUDA out of memory"
return (message in open(log, "r").read()) or (message in open(main_log, "r").read())
def mkl_error(path):
log = path.replace("trace.json", "../log")
main_log = path.replace("trace.json", "../agent_log/main_log")
messages = ["rror: mkl-service + Intel(R) MKL: MKL_THREADING_LAYER=INTEL is incompatible with libgomp-a34b3233.so.1 library.", "OpenBLAS blas_thread_init:"]
return any([m in open(log, "r").read() for m in messages]) or any([m in open(main_log, "r").read() for m in messages])
def quota_error(path):
log = path.replace("trace.json", "error.txt")
if os.path.exists(log):
message = "RemoteServiceError: EXCEPTION: total quota"
return message in open(log, "r").read()
return False
def connection_error(path):
log = path.replace("trace.json", "../log")
main_log = path.replace("trace.json", "../agent_log/main_log")
bad = ["You exceeded your current quota, please check your plan and billing details.", "Error: 'text-similarity-ada-001'", "Error: 'text-embedding-ada-001'"]
return ("Connection aborted" in open(log, "r").read()) or (any([b in open(main_log, "r").read() for b in bad]))
def langchain_error(path):
if os.path.exists(os.path.join(path.replace("trace.json", ""), "error.txt")):
return "langchain.schema.OutputParserException" in open(os.path.join(path.replace("trace.json", ""), "error.txt"), "r").read()
return False
def error(path):
return (os.path.exists(os.path.join(path.replace("trace.json", ""), "error.txt")) and not langchain_error(path)) or not os.path.exists(os.path.join(path.replace("trace.json", ""), "overall_time.txt"))
def json_error(path):
main_log = path.replace("trace.json", "../agent_log/main_log")
return open(main_log, "r").read().count("JSONDecodeError") > 2
def langchain_final(path):
return "Final Answer" in open(path.replace("trace.json", "../agent_log/main_log"), "r").read()
def autogpt_final(path):
return "Goal achieved" in open(path.replace("trace.json", "../agent_log/main_log"), "r").read()
def long_prompt_error(path):
main_log = path.replace("trace.json", "../agent_log/main_log")
return "EnvError: too long input for the tool" in open(main_log, "r").read()
def get_all_runs_with_log():
#TODO: fix paths to where your trace.json are
all_runs.extend(glob.glob("/lfs/local/0/qhwang/nlp_logs/final_exp_logs*/*/*/*/env_log/trace.json"))
df = pd.DataFrame()
for r in all_runs:
exp, task, run = r.split("/")[-5:-2]
if task in os.listdir("../research_assistant_final/MLAgentBench/benchmarks"):
new_row={"task": task, "exp": exp, "run": run, "path": r}
df = pd.concat([df, pd.DataFrame([new_row])], ignore_index=True)
df["error"] = df["path"].apply(error)
df["json_error"] = df["path"].apply(json_error)
df["long_prompt_error"] = df["path"].apply(long_prompt_error)
df["oom_error"] = df["path"].apply(oom_error)
df["connection_error"] = df["path"].apply(connection_error)
df['mkl_error'] = df["path"].apply(mkl_error)
df['quota_error'] = df["path"].apply(quota_error)
df["langchain_error"] = df["path"].apply(langchain_error)
df_no_error = df[(((~df["error"]) & (~df["connection_error"])) | df["exp"].isin(["no_retrieval_gpt4", "full_gpt4_long"]) | (df["exp"].isin(["langchain", "langchain_long"]) & df["langchain_error"]) )& (~df["oom_error"]) & (~df["mkl_error"])]
return df , df_no_error
lower_the_better_tasks = [ "parkinsons-disease", "feedback", "BabyLM", "llama-inference", "house-price", "vectorization"]
# TODO: add propoer label mapping and task name mapping for pretty printing in the figure
print_labels = {
"no_retrieval_gpt4" : "GPT-4",
"no_retrieval" : "Claude v1.0",
"autogpt" : "AutoGPT",
"react" : "React",
"langchain" : "LangChain (React)",
"sanity_check" : "Baseline"
}
print_task_labels = {
"cifar10_training" : "cifar10",
"imdb" : "imdb",
"ogbn-arxiv" : "ogbn-arxiv",
"home-data-for-ml-course" : "house-price",
"kaggle_training_reg" : "house-price",
"kaggle_training_class" : "spaceship-titanic",
"amp-parkinsons-disease-progression-prediction" : "parkinsons-disease",
"fathomnet-out-of-sample-detection" : "fathomnet",
"feedback-prize-english-language-learning" : "feedback",
"google-research-identify-contrails-reduce-global-warming" : "identify-contrails",
"speed-up" : "llama-inference",
"vectorisation" : "vectorization",
"CLRS" : "CLRS",
"babylm" : "BabyLM"
}
def get_improvement(df, baseline, thresh = None, prefix=""):
if prefix:
df[f"{prefix}increase"] = df[[f"{prefix}score", "task"]].apply(lambda x: (x[f"{prefix}score"] - baseline[(baseline["task"] == x["task"])]["final_score"].values[0])/baseline[(baseline["task"] == x["task"])]["final_score"].values[0] if x[f"{prefix}score"] is not None else None, axis=1)
df[f"{prefix}decrease"] = df[[f"{prefix}score", "task"]].apply(lambda x: (x[f"{prefix}score"] - baseline[(baseline["task"] == x["task"])][f"final_score"].values[0])/baseline[(baseline["task"] == x["task"])]["final_score"].values[0] if x[f"{prefix}score"] is not None else None, axis=1)
if thresh:
return df[["task", f"{prefix}increase", f"{prefix}decrease"]].apply(lambda x: (x[f"{prefix}increase"] > thresh if x["task"] not in lower_the_better_tasks else x[f"{prefix}decrease"] < - thresh) if x[f"{prefix}increase"] is not None else False, axis=1)
else:
return df[["task", f"{prefix}increase", f"{prefix}decrease"]].apply(lambda x: (x[f"{prefix}increase"] if x["task"] not in lower_the_better_tasks else - x[f"{prefix}decrease"]) if x[f"{prefix}increase"] is not None else None, axis=1)
# performance
def get_all_runs_eval(print_labels = print_labels, print_task_labels = print_task_labels):
# TODO: collect all evaluation jsons into all_results
all_results = {}
for f in glob.glob("/lfs/local/0/qhwang/nlp_logs/*.json"):
all_results.update(json.load(open(f, "r")))
df = pd.DataFrame()
for n, results in all_results.items():
if n.endswith(".json"):
n=n.split("/env_log")[0]
results = {n: results}
exp, task, run = n.split("/")[-3:]
exp = exp.strip()
if exp == "react":
continue
task = task.strip()
run = run.strip()
for source_file, r in results.items():
r_ = copy.deepcopy(r)
if len(r["score"]) < len(r["score_steps"])+1:
r_["score"].append(r["final_score"])
r_["score_steps"].append(len(json.load(open(r_["path"], "r"))["steps"]))
r_["score"] = np.array(r_["score"])
r_["score_steps"] = np.array(r_["score_steps"])
if exp == "no_retrieval":
r_["score"] = r_["score"][r_["score_steps"] < 16]
r_["score_steps"] = r_["score_steps"][r_["score_steps"] < 16]
if exp == "langchain":
r_["submitted_final_answer"] = langchain_final(r_["path"])
if exp == "autogpt":
r_["submitted_final_answer"] = autogpt_final(r_["path"])
new_row={"task": task, "exp": exp, "run": run, **r_}
df = pd.concat([df, pd.DataFrame([new_row])], ignore_index=True)
df["connection_error"] = df["path"].apply(connection_error)
df["has_error"] = df["path"].apply(error)
df["oom_error"] = df["path"].apply(oom_error)
df["mkl_error"] = df["path"].apply(mkl_error)
df["langchain_error"] = df["path"].apply(langchain_error)
print(len(df[(df["error"] != "") | (df["connection_error"] == True)]))
df = df[(((~df["has_error"]) & (df["connection_error"] == False)) | df["exp"].isin(["no_retrieval_gpt4", "full_gpt4_long"])| (df["exp"].isin(["langchain", "langchain_long"]) & df["langchain_error"]) ) & (~df["oom_error"]) & (~df["mkl_error"])]
df["exp"] = df["exp"].apply(lambda x: x if not x.endswith("_long") else x[:-5])
df = df[df["exp"].isin(list(print_labels.keys()))]
df["exp"] = df["exp"].apply(lambda x: print_labels[x])
df["task"] = df["task"].apply(lambda x: print_task_labels.get(x, x))
df["final_submitted_score"] = df[["final_score", "submitted_final_answer"]].apply(lambda x: x["final_score"] if x["final_score"] > 0 and x["submitted_final_answer"] else None, axis=1)
df["final_score"] = df["final_score"].apply(lambda x: x if x > 0 else None)
baseline = df[df["exp"] == "Baseline"][[ "task", "exp", "final_score"]].groupby(["task", "exp"]).mean().reset_index()
# special baseline numbers
try:
baseline.at[baseline[baseline["task"] == "imdb"].index.values[0], "final_score"] = 0.5
baseline.at[baseline[baseline["task"] == "fathomnet"].index.values[0], "final_score"] = 1e-10
except:
baseline = pd.concat(
[
baseline,
pd.DataFrame(
[{"task": "imdb", "exp": "Baseline", "final_score": 0.5}]
),
],
ignore_index=True,
)
baseline = pd.concat(
[
baseline,
pd.DataFrame(
[{"task": "fathomnet", "exp": "Baseline", "final_score": 1e-10}]
),
],
ignore_index=True,
)
baseline = pd.concat([baseline, pd.DataFrame([{"task" : "spaceship-titanic", "exp" :"Baseline", "final_score": 0.5}])], ignore_index=True)
baseline = pd.concat([baseline, pd.DataFrame([{"task" : "house-price", "exp" :"Baseline", "final_score": 1e10}])], ignore_index=True)
baseline = pd.concat([baseline, pd.DataFrame([{"task" : "ogbn-arxiv", "exp" :"Baseline", "final_score": 0.3134}])], ignore_index=True)
baseline = pd.concat([baseline, pd.DataFrame([{"task" : "vectorization", "exp" :"Baseline", "final_score": 6.1742}])], ignore_index=True)
return df, baseline
def get_all_runs_results(df = None, baseline = None, print_labels = print_labels, print_task_labels = print_task_labels):
if df is None or baseline is None:
df, baseline = get_all_runs_eval(print_labels = print_labels, print_task_labels = print_task_labels)
df[df["final_score"] > -1]["task"].unique()
df = df[df["task"].isin(baseline["task"].unique())]
df["max_score"] = df["score"].apply(lambda x: max(list(filter(lambda a: a > 0, x))) if len(list(filter(lambda a: a > 0, x))) > 0 else None)
df["min_score"] = df["score"].apply(lambda x: min(list(filter(lambda a: a > 0, x))) if len(list(filter(lambda a: a > 0, x))) > 0 else None)
df["increase"] = df[["max_score", "task"]].apply(lambda x: (x["max_score"] - baseline[(baseline["task"] == x["task"])]["final_score"].values[0])/baseline[(baseline["task"] == x["task"])]["final_score"].values[0] if x["max_score"] is not None else None, axis=1)
df["decrease"] = df[["min_score", "task"]].apply(lambda x: (x["min_score"] - baseline[(baseline["task"] == x["task"])]["final_score"].values[0])/baseline[(baseline["task"] == x["task"])]["final_score"].values[0] if x["min_score"] is not None else None, axis=1)
print(time.time())
df["improve"] = get_improvement(df, baseline)
df["improve_5"] = get_improvement(df, baseline, 0.05)
df["improve_10"] = get_improvement(df, baseline, 0.1)
df["improve_15"] = get_improvement(df, baseline, 0.15)
df["improve_20"] = get_improvement(df, baseline, 0.2)
df["improve_30"] = get_improvement(df, baseline, 0.3)
for prefix in ["final_"]:
df[f"{prefix}improve"] = get_improvement(df, baseline, None, prefix)
df[f"{prefix}improve_5"] = get_improvement(df, baseline, 0.05, prefix)
df[f"{prefix}improve_10"] = get_improvement(df, baseline, 0.1, prefix)
df[f"{prefix}improve_15"] = get_improvement(df, baseline, 0.15, prefix)
df[f"{prefix}improve_20"] = get_improvement(df, baseline, 0.2, prefix)
df[f"{prefix}improve_30"] = get_improvement(df, baseline, 0.3, prefix)
print(time.time())
# uncomment these to count tokens
# df[["prompt_tokens", "completed_tokens", "tool_prompt_tokens", "tool_completed_tokens", "num_steps", "total_time"]] = df.apply((lambda row: estimate_tokens(row["path"])), axis=1, result_type="expand")
# df['total_tokens'] = df["prompt_tokens"] + df["completed_tokens"] + df["tool_prompt_tokens"] + df["tool_completed_tokens"]
print(time.time())
return df
import seaborn as sns
from pandas.api.types import CategoricalDtype
colors = {
"GPT-4" : "#d62728",
"Claude v1.0" : "#2ca02c",
"AutoGPT" : "#9467bd",
"React" : "#8c564b",
"LangChain (React)" : "#e377c2",
"Baseline" : "#7f7f7f"
}
def get_tradeoff_plot(df):
def sample_and_mean(group):
if "GPT-4" in group["exp"].values[0]:
sample = group.sample(n=min(len(group), 8), random_state=1)
else:
sample = group.sample(n=min(len(group), 25), random_state=1)
return sample.groupby(["task", "exp"]).mean().reset_index().drop(columns=["task", "exp"])
grouped_df = df[["task", "exp", "final_improve_10", "total_tokens"]].groupby(["task", "exp"]).apply(sample_and_mean).round(4).reset_index()
x = grouped_df[["total_tokens","exp"]].groupby([ "exp"]).mean().values.flatten().tolist()
y = grouped_df[["final_improve_10","exp"]].groupby([ "exp"]).mean().values.flatten().tolist()
labels = ["AutoGPT", "Baseline", "Claude v1.0", "GPT-4", "LangChain (React)"]
plt.figure()
plt.scatter(x,y)
for i in range(len(x)):
plt.annotate(labels[i], # this is the text
(x[i], y[i]), # these are the coordinates to position the label
textcoords="offset points", # how to position the text
xytext=(0,10), # distance from text to points (x,y)
ha='center') # horizontal alignment can be left, right or center
plt.xlim((-30000, 200000))
plt.ylim((0, 0.3))
# plt.show()
plt.xlabel("Average Nsumber of Tokens Spent")
plt.ylabel("Average Success Rate")
plt.savefig("plots/tradeoff.pdf")
def get_plot(df, column_name = "improve_5", titile = "Improvement of 5%", save_name = "improve_5", plot_tokens = False, plot_time = False):
def sample_and_mean(group):
if "GPT-4" in group["exp"].values[0]:
sample = group.sample(n=min(len(group), 8), random_state=1)
else:
sample = group.sample(n=min(len(group), 25), random_state=1)
return sample.groupby(["task", "exp"]).mean().reset_index().drop(columns=["task", "exp"])
grouped_df = df[["task", "exp", column_name]].groupby(["task", "exp"]).apply(sample_and_mean).round(4).reset_index()
grouped_df.fillna(0, inplace=True)
if plot_time:
grouped_df[column_name] = grouped_df[column_name] / 60
elif not plot_tokens:
grouped_df[column_name] = grouped_df[column_name] * 100
# Define the order
task_order = list(print_task_labels.values())
task_order.remove("house-price")
exp_order = ["GPT-4", "Claude v1.0", "AutoGPT", "LangChain (React)", "Baseline"]
cat_type = CategoricalDtype(categories=task_order, ordered=True)
grouped_df['task'] = grouped_df['task'].astype(cat_type)
cat_type = CategoricalDtype(categories=exp_order, ordered=True)
grouped_df['exp'] = grouped_df['exp'].astype(cat_type)
plt.figure(figsize=(10,6))
palette = [colors[x] for x in exp_order]
barplot = sns.barplot(x='task', y=column_name, hue='exp', data=grouped_df, palette=palette, ci=95)
print(titile)
# Get the current x-tick labels
labels = [item.get_text() for item in barplot.get_xticklabels()]
# Modify the labels
new_labels = labels # [ l.split("_")[0].split("-")[0] for l in labels]
# Set the new labels
plt.xticks(range(len(labels)), new_labels, rotation=30)
plt.ylim(plt.ylim()[0], plt.ylim()[1] + (plt.ylim()[1]-plt.ylim()[0]) * 0.1)
leg = barplot.get_legend()
leg.set_title(None)
for t in leg.texts:
t.set_text(t.get_text().replace("Year=", ""))
plt.legend(loc='upper center', fancybox=True, shadow=True, ncol=4)
plt.xlabel("Task")
if plot_tokens:
plt.ylabel("Tokens")
elif plot_time:
plt.ylabel("Time (minutes)")
else:
plt.ylabel("Percentage")
plt.savefig(f"plots/{save_name}.pdf", bbox_inches='tight')
plt.show()
if __name__ == "__main__":
df = get_all_runs_results()
get_plot(df, "improve_5", "Percentage of runs that improve objective by over 5% at any point", "improve_5")
get_plot(df, "improve_10", "Percentage of runs that improve objective by over 10% at any point", "improve_10")
get_plot(df, "final_improve_5", "Percentage of runs that improves objective by over 5% at the end", "final_improve_5")
get_plot(df, "final_improve_10", "Percentage of runs that improves objective by over 10% at the end", "final_improve_10")
get_plot(df, "final_improve_30", "Percentage of runs that improves objective by over 30% at the end", "final_improve_30")
get_plot(df, "final_improve", "Average improvement in objective among runs that made a submission at the end.", "final_improve")
get_plot(df[df["submitted_final_answer"]], "final_improve", "Average improvement in objective among runs that made a final submission.", "final_improve_submitted")
get_plot(df, "total_tokens", "", "total_tokens", plot_tokens= True)
get_plot(df, "total_time", "", "total_time",plot_time=True)