logical-reasoning / llm_toolkit /eval_logical_reasoning.py
inflaton's picture
08a_InternLM_eval_NV4080 complete
a5837c1
raw
history blame
2.75 kB
import os
import sys
import torch
from dotenv import find_dotenv, load_dotenv
found_dotenv = find_dotenv(".env")
if len(found_dotenv) == 0:
found_dotenv = find_dotenv(".env.example")
print(f"loading env vars from: {found_dotenv}")
load_dotenv(found_dotenv, override=False)
path = os.path.dirname(found_dotenv)
print(f"Adding {path} to sys.path")
sys.path.append(path)
from llm_toolkit.llm_utils import *
from llm_toolkit.logical_reasoning_utils import *
model_name = os.getenv("MODEL_NAME")
adapter_name_or_path = os.getenv("ADAPTER_NAME_OR_PATH")
load_in_4bit = os.getenv("LOAD_IN_4BIT") == "true"
data_path = os.getenv("LOGICAL_REASONING_DATA_PATH")
results_path = os.getenv("LOGICAL_REASONING_RESULTS_PATH")
use_english_datasets = os.getenv("USE_ENGLISH_DATASETS") == "true"
print(model_name, adapter_name_or_path, load_in_4bit, data_path, results_path)
gpu_stats = torch.cuda.get_device_properties(0)
start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
print(f"(1) GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
print(f"{start_gpu_memory} GB of memory reserved.")
model, tokenizer = load_model(
model_name, load_in_4bit=load_in_4bit, adapter_name_or_path=adapter_name_or_path
)
gpu_stats = torch.cuda.get_device_properties(0)
start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
print(f"(2) GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
print(f"{start_gpu_memory} GB of memory reserved.")
datasets = load_logical_reasoning_dataset(
data_path, tokenizer=tokenizer, chinese_prompt=not use_english_datasets
)
if len(sys.argv) > 1:
num = int(sys.argv[1])
if num > 0:
print(f"--- evaluating {num} entries")
# create new dataset exluding those idx
datasets["test"] = datasets["test"].select(range(num))
print_row_details(datasets["test"].to_pandas(), indices=[0, -1])
print("Evaluating model: " + model_name)
predictions = eval_model(model, tokenizer, datasets["test"])
gpu_stats = torch.cuda.get_device_properties(0)
start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
print(f"(3) GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
print(f"{start_gpu_memory} GB of memory reserved.")
if adapter_name_or_path is not None:
model_name += "_" + adapter_name_or_path.split("/")[-1]
save_results(
model_name,
results_path,
datasets["test"],
predictions,
debug=True,
)
metrics = calc_metrics(datasets["test"]["label"], predictions, debug=True)
print(metrics)