-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathevaluate.py
190 lines (160 loc) · 8.78 KB
/
evaluate.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
182
183
184
185
186
187
188
189
190
import argparse
import os
from collections import defaultdict
from time import time
from typing import Dict
import torch
import numpy as np
import pandas as pd
from tqdm import tqdm
from yacs.config import CfgNode as CN
from config import _C as config
from criterion import get_loss_values
from util import load_config, load_model, get_criterions, prepare_dataloaders, set_seed
from data_utils import RMS
@torch.no_grad
def evaluate(epoch:int = 500, ckpt_dir:str = '', config:CN=None, drop_correct_0:bool=True, average:str=None)\
-> Dict[str, float]:
if ckpt_dir == '':
raise ValueError("ckpt_dir is empty")
if config is None:
raise ValueError("config is None")
print(f"Evaluating epoch {epoch}...")
print(f'Setting seed: {config.train.seed}')
set_seed(config.train.seed)
# load ckpt and prepare model
print('Loading model...')
model = load_model(epoch, ckpt_dir, config)
# prepare loss functions
print('Preparing loss functions...')
criterions = get_criterions(config, config.log.loss.types, average=average)
# prepare dataloader
print('Preparing dataset...')
test_loader = prepare_dataloaders(config.data, batch_size=1, train=False, test=True)
# evaluate
print('Evaluating...')
model.eval()
reduced_losses = defaultdict(list) # {} # None
reduced_losses['params'].append(sum(p.numel() for p in model.parameters() if p.requires_grad))
with torch.no_grad():
for idx, batch in enumerate(tqdm(test_loader)): # batch size 1
model.parse_batch(batch)
# make inference, and measure inference time
start_time = time()
model.forward()
infer_time = time() - start_time
reduced_losses['infer_time'].append(infer_time)
if drop_correct_0:
# if pred_rms class and gt_rms class are both 0, skip that index
# get indices where pred_rms class 0
if not config.data.rms_discretize:
model_device = model.pred_rms.device
mu_bins = RMS.get_mu_bins(config.data.rms_mu, config.data.rms_num_bins, config.data.rms_min)
mu_bins = mu_bins.to(model_device)
# make as if the prediction & gt is discretized
pred_rms_discretized = torch.zeros((model.pred_rms.shape[0], mu_bins.shape[0], model.pred_rms.shape[1]),
device=model_device)
model.gt_rms_continuous = model.gt_rms.clone()
for i in range(model.pred_rms.shape[0]):
# make RMS.discretize_rms(model.pred_rms[i], mu_bins).long() (axis 1) values to 1
assert RMS.discretize_rms(model.pred_rms[i], mu_bins).long().shape[0] == model.pred_rms.shape[1]
indices = RMS.discretize_rms(model.pred_rms[i], mu_bins).long() - 1
pred_rms_discretized[i, indices, torch.arange(indices.shape[0])] = 1
model.gt_rms[i] = RMS.discretize_rms(model.gt_rms[i], mu_bins).long()
model.pred_rms = pred_rms_discretized
pred_0_indices = torch.where(model.pred_rms.argmax(dim=1) == 0)[1]
gt_0_indices = torch.where(model.gt_rms == 0)[1]
# find intersection
drop_indices = pred_0_indices[torch.where(torch.isin(pred_0_indices, gt_0_indices))]
drop_mask = torch.ones(model.pred_rms.shape[-1], dtype=torch.bool)
drop_mask[drop_indices] = False
if len(drop_indices) > 0:
# drop indices (axis 2 of model.pred_rms, axis 1 of model.gt_rms)
model.gt_rms = model.gt_rms[:, drop_mask]
model.pred_rms = model.pred_rms[:, :, drop_mask]
model.gt_rms_continuous = model.gt_rms_continuous[:, drop_mask]
if not config.data.rms_discretize and not drop_correct_0:
assert model.pred_rms.shape[1] == model.gt_rms.shape[1]
model_device = model.pred_rms.device
mu_bins = RMS.get_mu_bins(config.data.rms_mu, config.data.rms_num_bins, config.data.rms_min)
mu_bins = mu_bins.to(model_device)
for i in range(model.pred_rms.shape[0]):
model.pred_rms[i] = RMS.undiscretize_rms(
RMS.discretize_rms(torch.tensor(model.pred_rms[i]), mu_bins).long(), mu_bins
)
model_gt = model.gt_rms
else:
assert model.pred_rms.shape[2] == model.gt_rms.shape[1] == model.gt_rms_continuous.shape[1]
model_gt = (model.gt_rms, model.gt_rms_continuous)
if model.pred_rms.shape[-1] == 0:
continue
_reduced_losses = get_loss_values(model.pred_rms, model_gt, criterions,
average=average if average != 'macro' else None,
model_output_onset=model.pred_onset if config.train.onset_supervision else None,
targets_onset=model.gt_onset if config.train.onset_supervision else None)
if idx == 0:
print(_reduced_losses)
nan_indices = []
for loss_type, loss in _reduced_losses.items():
if loss_type in ['CE', 'CE_GLS', 'MSE', 'MAE', 'PRAUC', 'ROCAUC']:
reduced_losses[loss_type].append(loss)
else:
if average in ['macro', None]:
if loss_type == 'ACC':
nan_indices = [i for i, l in enumerate(loss) if np.isnan(l)]
reduced_losses[loss_type].append(np.mean([l for idx, l in enumerate(loss) if idx not in nan_indices])) # macro
if average is None:
for i in range(len(loss)):
if i not in nan_indices:
reduced_losses[f'{loss_type}_{i}'].append(loss[i])
elif average == 'micro':
reduced_losses[loss_type].append(loss)
reduced_losses = {loss_type: np.mean(losses) for loss_type, losses in reduced_losses.items()}
return reduced_losses
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-c', '--ckpt_dir', type=str, default='', required=True,
help='directory for model checkpoint')
parser.add_argument('-e', '--epoch', type=int, default=500, required=True,
help='number of epochs to evaluate')
parser.add_argument('-o', '--output_file', type=str, default='./evaluate.csv', required=False,
help='path to csv file to save scores')
parser.add_argument('-a', '--average', default='micro', required=False, type=str,
choices=['micro', 'macro', None],
help='average type for ACC, PREC, RECALL, F1')
parser.add_argument('-0', '--preserve_correct_0', action='store_true', default=False, required=False,
help='preserve correct predictions of class 0')
args = parser.parse_args()
config = load_config(os.path.join(args.ckpt_dir, 'opts.yml'))
config.freeze()
scores_dict = evaluate(epoch=args.epoch, ckpt_dir=args.ckpt_dir, config=config,
drop_correct_0=(not args.preserve_correct_0), average=args.average)
scores_dict = {key: float(f'{value:.5f}') for key, value in scores_dict.items()}
### append and save score values to existing csv file
# get values
save_dir = config.log.save_dir
loss_type = config.train.loss.type
rms_num_bins = config.data.rms_num_bins
gls_blur_range = config.log.loss.gls_blur_range if 'CE_GLS' in config.log.loss.types else None
epoch = args.epoch
# create dataframe
df = pd.DataFrame({
'save_dir': [save_dir],
'loss_type': [loss_type],
'rms_num_bins': [rms_num_bins],
'gls_blur_range': [gls_blur_range],
'epoch': [epoch],
**{key: [value] for key, value in scores_dict.items()}
})
# if file does not exist write header
if not os.path.isfile(args.output_file):
df.to_csv(args.output_file, index=False)
else:
# read existing file
existing_df = pd.read_csv(args.output_file)
# concatenate new data with existing data
df = pd.concat([existing_df, df], ignore_index=True)
# save the concatenated dataframe
df.to_csv(args.output_file, index=False)
print(scores_dict)
print(f"Scores saved to {args.output_file}")