-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathget_top_k.py
144 lines (121 loc) · 6.68 KB
/
get_top_k.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
import math
import torch
import pickle
import os
import argparse
import numpy as np
import itertools
from tqdm import tqdm
from utils import load_model, move_to
from utils.data_utils import save_dataset
from torch.utils.data import DataLoader
import time
from datetime import timedelta
from utils.functions import parse_softmax_temperature
mp = torch.multiprocessing.get_context('spawn')
def eval_dataset_mp(args):
(dataset_path, width, softmax_temp, opts, i, num_processes) = args
model, _ = load_model(opts.model)
val_size = opts.val_size // num_processes
dataset = model.problem.make_dataset(filename=dataset_path, num_samples=val_size, offset=opts.offset + val_size * i)
device = torch.device("cuda:{}".format(i))
return _eval_dataset(model, dataset, width, softmax_temp, opts, device)
def eval_dataset(dataset_path, width, softmax_temp, opts):
# Even with multiprocessing, we load the model here since it contains the name where to write results
model, _ = load_model(opts.model)
use_cuda = torch.cuda.is_available() and not opts.no_cuda
if opts.multiprocessing:
assert use_cuda, "Can only do multiprocessing with cuda"
num_processes = torch.cuda.device_count()
assert opts.val_size % num_processes == 0
with mp.Pool(num_processes) as pool:
results = list(itertools.chain.from_iterable(pool.map(
eval_dataset_mp,
[(dataset_path, width, softmax_temp, opts, i, num_processes) for i in range(num_processes)]
)))
else:
device = torch.device("cuda:0" if use_cuda else "cpu")
dataset = model.problem.make_dataset(filename=dataset_path, num_samples=opts.val_size, offset=opts.offset)
results = _eval_dataset(model, dataset, width, softmax_temp, opts, device)
def _eval_dataset(model, dataset, width, softmax_temp, opts, device):
model.to(device)
model.eval()
model.set_decode_type(
"greedy" if opts.decode_strategy in ('bs', 'greedy') else "sampling",
temp=softmax_temp)
dataloader = DataLoader(dataset, batch_size=opts.eval_batch_size)
results = []
# print(width, opts.eval_batch_size , opts.max_calc_batch_size)
for batch in tqdm(dataloader, disable=opts.no_progress_bar):
batch = move_to(batch, device)
start = time.time()
with torch.no_grad():
if opts.decode_strategy in ('sample', 'greedy'):
if opts.decode_strategy == 'greedy':
assert width == 0, "Do not set width when using greedy"
assert opts.eval_batch_size <= opts.max_calc_batch_size, \
"eval_batch_size should be smaller than calc batch size"
batch_rep = 1
iter_rep = 1
elif width * opts.eval_batch_size > opts.max_calc_batch_size:
assert opts.eval_batch_size == 1
assert width % opts.max_calc_batch_size == 0
batch_rep = opts.max_calc_batch_size
iter_rep = width // opts.max_calc_batch_size
else:
batch_rep = width
iter_rep = 1
assert batch_rep > 0
# This returns (batch_size, iter_rep shape)
# print(width, opts.eval_batch_size , opts.max_calc_batch_size, batch_rep)
top_k_sequences, top_k_costs = model.sample_many_top_k(batch, opts.k, batch_rep=batch_rep, iter_rep=iter_rep)
batch_size = len(top_k_costs)
ids = torch.arange(batch_size, dtype=torch.int64, device=top_k_costs.device)
else:
assert opts.decode_strategy == 'bs'
cum_log_p, sequences, costs, ids, batch_size = model.beam_search(
batch, beam_size=width,
compress_mask=opts.compress_mask,
max_calc_batch_size=opts.max_calc_batch_size
)
duration = time.time() - start
for seq, cost in zip(top_k_sequences, top_k_costs):
results.append((cost, seq, duration))
pickle.dump( results, open( opts.output_filename, "wb" ) )
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("datasets", nargs='+', help="Filename of the dataset(s) to evaluate")
parser.add_argument("-f", action='store_true', help="Set true to overwrite")
parser.add_argument("-o", default=None, help="Name of the results file to write")
parser.add_argument('--k', type=int, default=10,
help='top-k solutions from sampling')
parser.add_argument('--val_size', type=int, default=10000,
help='Number of instances used for reporting validation performance')
parser.add_argument('--offset', type=int, default=0,
help='Offset where to start in dataset (default 0)')
parser.add_argument('--eval_batch_size', type=int, default=1024,
help="Batch size to use during (baseline) evaluation")
# parser.add_argument('--decode_type', type=str, default='greedy',
# help='Decode type, greedy or sampling')
parser.add_argument('--width', type=int, nargs='+',
help='Sizes of beam to use for beam search (or number of samples for sampling), '
'0 to disable (default), -1 for infinite')
parser.add_argument('--decode_strategy', type=str,
help='Beam search (bs), Sampling (sample) or Greedy (greedy)')
parser.add_argument('--softmax_temperature', type=parse_softmax_temperature, default=1,
help="Softmax temperature (sampling or bs)")
parser.add_argument('--model', type=str)
parser.add_argument('--no_cuda', action='store_true', help='Disable CUDA')
parser.add_argument('--no_progress_bar', action='store_true', help='Disable progress bar')
parser.add_argument('--compress_mask', action='store_true', help='Compress mask into long')
parser.add_argument('--max_calc_batch_size', type=int, default=10000, help='Size for subbatches')
parser.add_argument('--output_filename', default='save.pkl', help="Name of output file")
parser.add_argument('--multiprocessing', action='store_true',
help='Use multiprocessing to parallelize over multiple GPUs')
opts = parser.parse_args()
assert opts.o is None or (len(opts.datasets) == 1 and len(opts.width) <= 1), \
"Cannot specify result filename with more than one dataset or more than one width"
widths = opts.width if opts.width is not None else [0]
for width in widths:
for dataset_path in opts.datasets:
eval_dataset(dataset_path, width, opts.softmax_temperature, opts)