-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathscore_eval_webshop.py
181 lines (160 loc) · 7.1 KB
/
score_eval_webshop.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
import argparse, json, os, re
from tqdm import tqdm
from typing import Dict, List
import numpy as np
import torch
import transformers
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
import adaptive_utils
parser = argparse.ArgumentParser()
parser.add_argument("--model", type=str, default="codellama/CodeLlama-7b-Instruct-hf", help="Model name")
parser.add_argument("--cache_dir", type=str, default="", help="Cache directory for the model")
parser.add_argument("--load_in_4bit", type=bool, default=True, help="Load in 4-bit")
parser.add_argument("--seed", type=int, default=42, help="Random seed")
parser.add_argument("--data_path", type=str, help="Path to the data")
parser.add_argument("--output_dir", type=str, help="Output directory")
parser.add_argument("--aggregation", type=str, default="mean", help="Aggregation method (mean, min, prod)")
parser.add_argument("--method", type=str, default="SC", help="Method (SC, SoftSC, AdaptiveSC, AdaptiveSoftSC)")
parser.add_argument("--threshold", type=float, default=0.8, help="Threshold for AdaptiveSC and AdaptiveSoftSC")
args = parser.parse_args()
args.device = "cuda" if torch.cuda.is_available() else "cpu"
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=True,
)
transformers.set_seed(args.seed)
tokenizer = AutoTokenizer.from_pretrained(args.model, cache_dir=args.cache_dir)
if args.load_in_4bit:
model = AutoModelForCausalLM.from_pretrained(args.model, cache_dir=args.cache_dir, device_map="auto", quantization_config=bnb_config)
else:
model = AutoModelForCausalLM.from_pretrained(args.model, cache_dir=args.cache_dir, device_map="auto",
torch_dtype=torch.float16 if args.device == "cuda" else torch.bfloat16)
tokenizer.pad_token_id = tokenizer.eos_token_id
def find_start(tokens):
match_token = ['<0x0A>', 'Action', ':']
for i in range(len(tokens) - 3, 0, -1):
if tokens[i : i + 3] == match_token:
break
return i + 3
softmax = torch.nn.Softmax(dim=1)
@torch.no_grad()
def get_probs(input):
input = tokenizer(input, return_tensors="pt").to(args["device"])
input_tokens = tokenizer.convert_ids_to_tokens(input["input_ids"][0])
output = model(**input, return_dict=True)
logits = output.logits[0].to("cpu")
probs = softmax(logits)
start = find_start(input_tokens)
probs_list = [probs[i-1][input["input_ids"][0][i]].item() for i in range(start, len(input_tokens)-1)]
probs_list_np = np.array(probs_list)
return probs_list_np
with open(args.data_path, "r") as f:
data = json.load(f)
results_path = os.path.join(args.output_dir, "results.json")
prompts = data["prompts"]
for key in data:
if key == "prompts":
continue
problem = data[key]["problem"]
products = data[key]["products"]
scores = data[key]["scores"]
rewards = data[key]["reward"]
run_keys = []
for subkey in data[key]:
if subkey.startswith("run"):
run_keys.append(subkey)
init_prompt_select = prompts["select"].format(problem)
for i in range(len(run_keys)):
act_history = data[key][run_keys[i]]["act_history"]
trace = data[key][run_keys[i]]["trace"]
prompt = init_prompt_select
j = 0
probs_list = []
avg_probs_list = []
lowest_probs_list = []
prod_logprobs_list = []
for act in act_history["select"]:
if act == "click[Back to Search]":
break
while act not in trace["select"][j]:
prompt += "\n" + trace["select"][j]
j += 1
input = prompt + "\n\nAction: " + act + "\n"
logprobs = get_probs(input)
probs_list.append(logprobs.tolist())
avg_probs_list.append(np.mean(logprobs))
lowest_probs_list.append(np.min(logprobs))
prod_logprobs_list.append(np.mean(np.log(logprobs)))
prompt = prompts["buy"].format(products[i], problem)
j = 0
for act in act_history["buy"]:
while act not in trace["buy"][j]:
prompt += "\n" + trace["buy"][j]
j += 1
input = prompt + "\n\nAction: " + act + "\n"
logprobs = get_probs(input)
probs_list.append(logprobs.tolist())
avg_probs_list.append(np.mean(logprobs))
lowest_probs_list.append(np.min(logprobs))
prod_logprobs_list.append(np.mean(np.log(logprobs)))
data[key][run_keys[i]]["probs_list"] = probs_list
data[key][run_keys[i]]["avg_probs_list"] = avg_probs_list
data[key][run_keys[i]]["lowest_probs_list"] = lowest_probs_list
data[key][run_keys[i]]["prod_logprobs_list"] = prod_logprobs_list
with open(results_path, "w") as f:
json.dump(data, f, indent=2)
data = {}
with open(results_path, "r") as f:
fdata = json.load(f)
for key in fdata:
if key.startswith("fixed_"):
data[key] = fdata[key]
success_num = 0
total_score = 0
for key in data:
scores = data[key]["scores"]
rewards = data[key]["reward"]
products = data[key]["products"]
run_keys = []
for subkey in data[key]:
if subkey.startswith("run"):
run_keys.append(subkey)
if args.method == "SC":
major_product = max(set(products), key=products.count)
idx = products.index(major_product)
total_score += scores[idx]
if rewards[idx]:
success_num += 1
elif args.method == "SoftSC":
if args.aggregation == "mean":
probs_list = [np.mean(np.array(data[key][run_key]["avg_probs_list"])) for run_key in run_keys]
elif args.aggregation == "min":
probs_list = [np.min(np.array(data[key][run_key]["lowest_probs_list"])) for run_key in run_keys]
elif args.aggregation == "prod":
probs_list = [np.mean(np.array(data[key][run_key]["prod_logprobs_list"])) for run_key in run_keys]
total_score += scores[np.argmax(probs_list)]
if rewards[np.argmax(probs_list)]:
success_num += 1
elif args.method == "AdaptiveSC":
adaptive_products = adaptive_utils.adptive_sampling(products, threshold=args.threshold)
major_product = max(set(adaptive_products), key=adaptive_products.count)
idx = products.index(major_product)
total_score += scores[idx]
if rewards[idx]:
success_num += 1
elif args.method == "AdaptiveSoftSC":
lowest_probs_list = [np.mean(np.array(data[key][run_key]["lowest_probs_list"])) for run_key in run_keys]
generations = []
probs_sum = 0.0
j = 0
while probs_sum < args.threshold and len(generations) < len(products):
probs_sum += lowest_probs_list[j]
generations.append((j, products[j], lowest_probs_list[j]))
j += 1
idx, best_generation, best_prob = sorted(generations, key=lambda x: x[2])[-1]
total_score += scores[idx]
if rewards[idx]:
success_num += 1
print(f"Success rate: {success_num / len(data)}, Average score: {total_score / len(data)}")