-
Notifications
You must be signed in to change notification settings - Fork 0
/
my_main_eval_AudioCaps.py
210 lines (149 loc) · 6.91 KB
/
my_main_eval_AudioCaps.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
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
'''The goal of this script is to run audio retrieval with caption queries from the Clotho dev-test set
with audio files from AudioCaps. Then we will augment the clotho dev-train set with those and train a new model.
'''
import pandas as pd
import wandb
import os
import time
import numpy
import torch
import torch.optim as optim
import yaml
from torch.utils.data import DataLoader
from utils import criterion_utils, data_utils, eval_utils, model_utils
from eval_script import retrieval_metrics
# from models import ema
from utils.data_utils import EvalQueryAudioDataset
import h5py
torch_seed=0
numpy.random.seed(42)
torch.manual_seed(torch_seed)
import socket
host_name = socket.gethostname()
print(host_name)
if host_name == 'choro':
YAML_FPATH = "choro_conf.yaml"
# data_dir = '../clotho-dataset/data'
# data_dir = '/homelocal/thomas/clotho-dataset/data'
elif host_name == 'legolas':
YAML_FPATH = "my_conf.yaml"
# data_dir = '../clotho-dataset/data'
# data_dir = '/homelocal/thomas/clotho-dataset/data'
else:
# data_dir = '/tmpdir/pellegri/corpus/clotho-dataset/data/'
YAML_FPATH = "osirim_conf.yaml"
print(YAML_FPATH)
with open(YAML_FPATH, "rb") as stream:
config = yaml.full_load(stream)
training_config = config["training"]
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# model ckpt directory
# checkpoint_dir='best_checkpoints/checkpoint_seed0_none_margin_0.4_000018_mAP_2292'
checkpoint_dir='best_checkpoints/checkpoint_TAGS_sigmoidAS_0.800_margin_0.4_000026_mAP_2342'
# checkpoint_dir='best_checkpoints/checkpoint_TAGS_sigmoidAS_0.800_margin_0.4_000018_mAP_2339'
# checkpoint_dir='best_checkpoints/checkpoint_TAGS_sigmoidAS_0.800_margin_0.4_000018_mAP_2307'
# checkpoint_dir='best_checkpoints/checkpoint_TAGS_sigmoidAS_0.800_margin_0.4_000021'
# checkpoint_dir='best_checkpoints/checkpoint_pretrainedAudioCaps_seed1978_seed0_TAGS_sigmoidAS_0.800_margin_0.4_000053_mAP_2396'
# checkpoint_dir='best_checkpoints/checkpoint_pretrainedAudioCaps_seed1978_seed42_TAGS_sigmoidAS_0.800_margin_0.4_noScheduler_000043_mAP_2402'
_, vocabulary = data_utils.load_data(config["train_data"])
model_config = config[training_config["model"]]
model = model_utils.get_model(model_config, vocabulary)
print(model)
# Restore model states
model = model_utils.restore(model, checkpoint_dir)
model.to(device=device)
model.eval()
print("Nb of learnable parameters :",
sum(p.numel() for p in model.parameters() if p.requires_grad))
use_sentence_embeddings = True
use_auto_caption_embeddings = False
print("Using Etienne's auto caption embeddings?", use_auto_caption_embeddings)
# I - AUDIOCAPS: stack audio embeddings passed through the model
# Load audio features from AudioCaps
feats_path = os.path.join(config["audiocaps_train_data"]["input_path"], config["audiocaps_train_data"]["audio_features"])
audio_feats = h5py.File(feats_path, "r")
print("Load", feats_path)
sent_emb_path = os.path.join(config["audiocaps_train_data"]["input_path"],
config["audiocaps_train_data"]["caption_embeddings"])
sentence_embeddings = h5py.File(sent_emb_path, "r")
print("Load", sent_emb_path)
csv_path = os.path.join(config["audiocaps_train_data"]["input_path"],
config["audiocaps_train_data"]["data_splits"]['train'])
df = pd.read_csv(csv_path)
print("Load", csv_path)
from utils.data_utils import AudiocapsQueryAudioDataset
audiocaps_dataset = AudiocapsQueryAudioDataset(audio_feats, df, sentence_embeddings)
fid_embs, fid_fnames = {}, {}
with torch.no_grad():
for index in range(len(audiocaps_dataset)):
audio, word_query, info, query = audiocaps_dataset[index]
audio = torch.unsqueeze(audio, dim=0).to(device=device)
audio_emb, _, _ = model(audio, query)
audio_emb = torch.squeeze(audio_emb, dim=0).to(device=device)
fid_embs[info["fid"]] = audio_emb
fid_fnames[info["fid"]] = info["fname"]
# Stack audio embeddings
audiocaps_embs, audiocaps_fnames = [], []
for fid in fid_embs:
audiocaps_embs.append(fid_embs[fid])
audiocaps_fnames.append(fid_fnames[fid])
audiocaps_audio_embs = torch.vstack(audiocaps_embs) # dim [N, E]
# II - Load queries from Clotho dev-test, and do retrieval on the AUDIOCAPS audio embeddings
#
# nb of audio files to retrieve
K = 10
# Load audio features
feats_path = os.path.join(config["eval_data"]["input_path"], config["eval_data"]["audio_features"])
audio_feats_clotho = h5py.File(feats_path, "r")
print("Load", feats_path)
sent_emb_path = os.path.join(config["eval_data"]["input_path"], config["eval_data"]["caption_embeddings"])
sentence_embeddings_clotho = h5py.File(sent_emb_path, "r")
print("Load", sent_emb_path)
# Load json from Clotho dev-test
json_path = os.path.join(config["train_data"]["input_path"], config["train_data"]["data_splits"]["test"])
df_clotho = pd.read_json(json_path)
print("Load", json_path)
clothotest_dataset = EvalQueryAudioDataset(audio_feats_clotho, df_clotho, sentence_embeddings_clotho)
# Retrieve audio files for eval captions
split_name = 'audiocaps_with_clotho_dev_eval_queries'
scores_output_fpath = os.path.join(checkpoint_dir, "{}_scores.h5".format(split_name))
# scores_h5_output_file = None
with h5py.File(scores_output_fpath, "w") as score_store:
with torch.no_grad():
cid_embs, cid_infos = {}, {}
# Encode audio signals and captions
for index in range(len(clothotest_dataset)):
audio, word_query, info, query = clothotest_dataset[index]
audio = torch.unsqueeze(audio, dim=0).to(device=device)
query = torch.unsqueeze(query, dim=0).to(device=device)
_, query_emb, _ = model(audio, query)
query_emb = torch.squeeze(query_emb, dim=0).to(device=device)
cid_embs[info["cid"]] = query_emb
cid_infos[info["cid"]] = info
# Compute similarities
output_rows = []
for cid in cid_embs:
# print(cid)
sims = torch.mm(torch.vstack([cid_embs[cid]]), audiocaps_audio_embs.T).flatten().to(device=device)
# print(cid, sims.size(), sims[:5])
sorted_idx = torch.argsort(sims, dim=-1, descending=True)
csv_row = [cid_infos[cid]["caption"]] # caption
for idx in sorted_idx[:K]: # top-K retrieved fnames
csv_row.append(audiocaps_fnames[idx])
score_store[str(cid)] = sims.clone().detach().cpu().numpy()
output_rows.append(csv_row)
csv_fields = ["caption",
"file_name_1",
"file_name_2",
"file_name_3",
"file_name_4",
"file_name_5",
"file_name_6",
"file_name_7",
"file_name_8",
"file_name_9",
"file_name_10"]
output = pd.DataFrame(data=output_rows, columns=csv_fields)
output.to_csv(os.path.join(checkpoint_dir, "{}.output.csv".format(split_name)),
index=False)
print("Saved", "{}.output.csv".format(split_name))