This repository aims at providing efficient CNNs for Audio Tagging. We provide AudioSet pre-trained models ready for downstream training and extraction of audio embeddings.
License: MIT License
Hi there,
thanks for sharing the code.
I was in particular interested in the dcase20 dataset loader.
However, something in the MixupDataset classed confused me.
See this link: https://github.com/fschmid56/EfficientAT/blob/main/datasets/dcase20.py#L101
SimpleSelectionDataset is returning x, label, device, city, self.available_indices[index].
It is then interpreted as x1, f1, y1, d1, c1.
It turns out that during training, x1 and f1 are used.
However, I think the goal is to return a one-hot encoded and weighted version of the label (y1 * l + y2 * (1. - l)).
Maybe not relevant for this training but in case someone stumbles on this or wants to reuse it might find it useful!
Hey Florian,
I checked out your new fname_to_index file and trained some models, but performance is extremely bad.
Then just proceeded to print some scores from your provided fname_to_idx.pkl file and check out the ground truth.
I used this simple code to map your indexes to the fnames:
import torch
import numpy as np
import pandas as pd
clmaps = pd.read_csv('./class_labels_indices.csv').set_index('index')['display_name'].to_dict()
data = np.load('./passt_enemble_logits_mAP_495.npy', allow_pickle=True)
fnames_to_idx = np.load('./fname_to_index.pkl', allow_pickle=True)
idx_to_fnames = {v: k for k, v in fnames_to_idx.items()}
for idx, fname in idx_to_fnames.items():
values, idxs = torch.as_tensor(data[idx], dtype=torch.float32).sigmoid().topk(5)
print(f" ==== {fname} ==== ")
names = [clmaps[i] for i in idxs.numpy()]
for score, clname in zip(values, names):
print(f"{clname:<10} {score:<.3f}")Some of the outputs are:
==== 09c885WMtMw ====
Animal 0.914
Dog 0.887
Domestic animals, pets 0.847
Bark 0.643
Bow-wow 0.293
The ground truth for that file however, is "Music", you can check out the source at:
https://youtu.be/09c885WMtMw?t=80
Another sample is:
==== 09bFB0X-8QY ====
Speech 0.917
Female speech, woman speaking 0.598
Narration, monologue 0.406
Child speech, kid speaking 0.050
Inside, small room 0.028
Which can be viewed here: https://youtu.be/09bFB0X-8QY?t=16
I'm reasonably confident that your fname_to_index is somewhat wrong, could you maybe check it out if its the case?
EDIT:
In the above code snippet, the following will throw an error, which means there are some duplicate indexes:
assert len(fnames_to_idx) == len(idx_to_fnames)Kind Regards,
Heinrich
Warning: FMAX is None setting to 15000
Dataset from /data/xiaoshengchang/audioset/mp3/eval_segments_mp3.hdf with length 18887.
Running AudioSet evaluation for model 'mn10_as' on device 'cuda'
69%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▎ | 109/158 [00:29<00:07, 6.86it/s]Failed to read frame size: Could not seek to 1026.
82%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▉ | 129/158 [00:35<00:07, 3.66it/s]
Traceback (most recent call last):
File "/data/xiaoshengchang/EfficientAT-main/ex_audioset.py", line 348, in
evaluate(args)
File "/data/xiaoshengchang/EfficientAT-main/ex_audioset.py", line 266, in evaluate
for batch in tqdm(dl):
File "/opt/conda/lib/python3.9/site-packages/tqdm/std.py", line 1195, in iter
for obj in iterable:
File "/opt/conda/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 628, in next
data = self._next_data()
File "/opt/conda/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 1313, in _next_data
return self._process_data(data)
File "/opt/conda/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 1359, in _process_data
data.reraise()
File "/opt/conda/lib/python3.9/site-packages/torch/_utils.py", line 542, in reraise
raise RuntimeError(msg) from None
RuntimeError: Caught ValueError in DataLoader worker process 9.
Original Traceback (most recent call last):
File "/opt/conda/lib/python3.9/site-packages/torch/utils/data/_utils/worker.py", line 302, in _worker_loop
data = fetcher.fetch(index)
File "/opt/conda/lib/python3.9/site-packages/torch/utils/data/_utils/fetch.py", line 58, in fetch
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/opt/conda/lib/python3.9/site-packages/torch/utils/data/_utils/fetch.py", line 58, in
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/data/xiaoshengchang/EfficientAT-main/datasets/audioset.py", line 154, in getitem
waveform = decode_mp3(self.dataset_file['mp3'][index])
File "/data/xiaoshengchang/EfficientAT-main/datasets/audioset.py", line 37, in decode_mp3
container = av.open(io.BytesIO(mp3_arr.tobytes()))
File "av/container/core.pyx", line 401, in av.container.core.open
File "av/container/core.pyx", line 272, in av.container.core.Container.cinit
File "av/container/core.pyx", line 292, in av.container.core.Container.err_check
File "av/error.pyx", line 336, in av.error.err_check
av.error.ValueError: [Errno 22] Invalid argument: ''; last error log: [mp3] Failed to read frame size: Could not seek to 1026.
Thanks for the great work. I would like to ask the input(i.e.,x) shape of the mobilenet model, and is it (batch_size, 1, time_steps, mel_bins) or (batch_size, 1, mel_bins, time_steps)?
x = _mel_forward(x, mel)
y_hat, _ = model(x)
I'm so sorry to disturb you. May I apply for the .pt model of the Passt model?
Hi,
i try to run the training script, and some data was loss during downloading. So the teacher pred index is not right and can not be used.
I am wondering if your have time to update a Dict version of teacher pred with audio_name as key, because it would be more complex for new guy to get the ensemble logit
# here the teacher_preds is indexed by dataset index
if args.kd_lambda > 0:
y_soft_teacher = teacher_preds[i]
Thanks
how to convert this model to onnx type?
Hi,
Would you mind uploading the pre-trained baseline model without KD? I need it to do the model comparison experiments on our self-collected dataset. Thanks.
Hello, how to get the frame-level output results similar to clipwise_output in panns
Thanks
Hi,
how to train the script using multiple gpus? thanks
Hi，fschmid，I have a question: How to accurately identify the sound event offset? As is known, the sound events in our life are endless, and some events are not the sound event categories in the Audioset. How to accurately identify these?
In lucidrains/audiolm-pytorch#177 I describe an approach for using EfficientAT as a discriminator for GAN training.
The only code change needed is this:
def _forward_impl(self, x: Tensor, return_fmaps: bool = False) -> Union[Tuple[Tensor, Tensor], Tuple[Tensor, List[Tensor]]]:
fmaps = []
for i, layer in enumerate(self.features):
x = layer(x)
if return_fmaps:
fmaps.append(x)
features = F.adaptive_avg_pool2d(x, (1, 1)).squeeze()
x = self.classifier(x).squeeze()
if features.dim() == 1 and x.dim() == 1:
# squeezed batch dimension
features = features.unsqueeze(0)
x = x.unsqueeze(0)
if return_fmaps:
return x, fmaps
else:
return x, features
def forward(self, x: Tensor) -> Union[Tuple[Tensor, Tensor], Tuple[Tensor, List[Tensor]]]:
return self._forward_impl(x)
to optionally return feature maps for learning feature matching in the generator.
I'm curious if you could release a pretrained model with a much shorter receptive length?
This would be useful for fine-grained tasks, like music transcription and event transcription (with a smaller hop size)
Hi,
I'm trying to use mn10_as as a pre-trained model, and want to fine-tune it to fit my collected dataset (3 classes, 50 10-second clips each class, sampling rate: 16KHz).
I prepraed my dataset and made a format like DCASE20, but only filename and scene_label are used. Then I modified ex_dcase20.py to ex_my_dataset.py. Follows point out where I modified, but errors occured, and hope you could give me some hits. Many thanks.
def train(args):
# Train Models for Acoustic Scene Classification
# logging is done using wandb
wandb.init(
project="my_dataset",
notes="Fine-tune Models for Acoustic Scene Classification.",
tags=[ "Acoustic Scene Classification", "Fine-Tuning"],
config=args,
name=args.experiment_name
)
device = torch.device('cuda') if args.cuda and torch.cuda.is_available() else torch.device('cpu')
# model to preprocess waveform into mel spectrograms
mel = AugmentMelSTFT(n_mels=args.n_mels,
sr=args.resample_rate,
win_length=args.window_size,
hopsize=args.hop_size,
n_fft=args.n_fft,
freqm=args.freqm,
timem=args.timem,
fmin=args.fmin,
fmax=args.fmax,
fmin_aug_range=args.fmin_aug_range,
fmax_aug_range=args.fmax_aug_range
)
mel.to(device)
# load prediction model
pretrained_name = args.pretrained_name
if pretrained_name:
model = get_mobilenet(width_mult=NAME_TO_WIDTH(pretrained_name), pretrained_name=pretrained_name,
head_type=args.head_type, se_dims=args.se_dims, num_classes=3)
else:
model = get_mobilenet(width_mult=args.model_width, head_type=args.head_type, se_dims=args.se_dims,
num_classes=3)
parser.add_argument('--pretrained_name', type=str, default='mn10_as')
sr=16000
resample_rate=sr
dataset_config = {
"dataset_name": "my_dataset",
"meta_csv": os.path.join(dataset_dir, "meta.csv"),
"train_files_csv": os.path.join(dataset_dir, "evaluation_setup", "fold1_train.csv"),
"test_files_csv": os.path.join(dataset_dir, "evaluation_setup", "fold1_evaluate.csv")
#
class BasicDCASE22Dataset(TorchDataset):
def __init__(self, meta_csv, sr=sr, cache_path=None):
"""
@param meta_csv: meta csv file for the dataset
@param sr: specify sampling rate
@param sr: specify cache path to store resampled waveforms
return: waveform, name of the file, label, device and cities
"""
df = pd.read_csv(meta_csv, sep="\t")
le = preprocessing.LabelEncoder()
self.labels = torch.from_numpy(le.fit_transform(df[['scene_label']].values.reshape(-1)))
**#self.devices = le.fit_transform(df[['source_label']].values.reshape(-1))
#self.cities = le.fit_transform(df['identifier'].apply(lambda loc: loc.split("-")[0]).values.reshape(-1))**
self.files = df[['filename']].values.reshape(-1)
self.sr = sr
if cache_path is not None:
self.cache_path = os.path.join(cache_path, dataset_config["dataset_name"] + f"_r{self.sr}", "files_cache")
os.makedirs(self.cache_path, exist_ok=True)
else:
self.cache_path = None
def __getitem__(self, index):
if self.cache_path:
cpath = os.path.join(self.cache_path, str(index) + ".pt")
try:
sig = torch.load(cpath)
except FileNotFoundError:
sig, _ = librosa.load(os.path.join(dataset_dir, self.files[index]), sr=self.sr, mono=True)
sig = torch.from_numpy(sig[np.newaxis])
torch.save(sig, cpath)
else:
sig, _ = librosa.load(os.path.join(dataset_dir, self.files[index]), sr=self.sr, mono=True)
sig = torch.from_numpy(sig[np.newaxis])
**#return sig, self.labels[index], self.devices[index], self.cities[index]**
return sig, self.labels[index]
def __len__(self):
return len(self.files)
class SimpleSelectionDataset(TorchDataset):
"""A dataset that selects a subsample from a dataset based on a set of sample ids.
Supporting integer indexing in range from 0 to len(self) exclusive.
"""
def __init__(self, dataset, available_indices):
"""
@param dataset: dataset to load data from
@param available_indices: available indices of samples for 'training', 'testing'
return: x, label, device, city, index
"""
self.available_indices = available_indices
self.dataset = dataset
def __getitem__(self, index):
#x, label, device, city = self.dataset[self.available_indices[index]]
x, label = self.dataset[self.available_indices[index]]
#return x, label, device, city, self.available_indices[index]
return x, label, self.available_indices[index]
def __len__(self):
return len(self.available_indices)
File C:\Anaconda3\envs\EfficientAT\lib\site-packages\spyder_kernels\py3compat.py:356 in compat_exec
exec(code, globals, locals)
File d:\users\2023-efficientat-main\ex_my_dataset.py:241
train(args)
File d:\users\2023-efficientat-main\ex_my_dataset.py:100 in train
for batch in pbar:
File C:\Anaconda3\envs\EfficientAT\lib\site-packages\tqdm\std.py:1195 in __iter__
for obj in iterable:
File C:\Anaconda3\envs\EfficientAT\lib\site-packages\torch\utils\data\dataloader.py:435 in __iter__
return self._get_iterator()
File C:\Anaconda3\envs\EfficientAT\lib\site-packages\torch\utils\data\dataloader.py:381 in _get_iterator
return _MultiProcessingDataLoaderIter(self)
File C:\Anaconda3\envs\EfficientAT\lib\site-packages\torch\utils\data\dataloader.py:1034 in __init__
w.start()
File C:\Anaconda3\envs\EfficientAT\lib\multiprocessing\process.py:121 in start
self._popen = self._Popen(self)
File C:\Anaconda3\envs\EfficientAT\lib\multiprocessing\context.py:224 in _Popen
return _default_context.get_context().Process._Popen(process_obj)
File C:\Anaconda3\envs\EfficientAT\lib\multiprocessing\context.py:336 in _Popen
return Popen(process_obj)
File C:\Anaconda3\envs\EfficientAT\lib\multiprocessing\popen_spawn_win32.py:93 in __init__
reduction.dump(process_obj, to_child)
File C:\Anaconda3\envs\EfficientAT\lib\multiprocessing\reduction.py:60 in dump
ForkingPickler(file, protocol).dump(obj)
AttributeError: Can't pickle local object 'get_roll_func.<locals>.roll_func'
Firstly I would like to state that this repo is great, so many models, all in pytorch and getting them to work on my machine was very easy.
Have you tried fine-tuning the models on the temporally-strong labeled subset of the AudioSet dataset?
I'm not sure why this was needed, but I had to add this hack to get num_classes=1 to work:
< #num_classes = state_dict['classifier.1.bias'].size(0)
< num_classes = state_dict['classifier.2.bias'].size(0)
---
> num_classes = state_dict['classifier.1.bias'].size(0)
313,315d299
< if "classifier.2.weight" in state_dict:
< del state_dict['classifier.2.weight']
< del state_dict['classifier.2.bias']
I won't push a fix because I don't understand the impact of this on other users. Perhaps it should only be used when num_classes is 1?
The documentation for fine-tuning DyMN on OpenMic includes a reported example run, accessible via the following link: https://api.wandb.ai/links/florians/qo32vrgl.
In this report, the mean Average Precision (mAP) is stated as 91.6, whereas the paper reports a mAP of 84.4. Interestingly, the mAP value I obtained from fine-tuning on my personal computer aligns closely with the paper's reported value.
Upon closer inspection, I noticed that all mAP values correspond exactly to the values of the ROC statistics. This raises the question: could incorrect statistics be logged to W&B?
when i run ex_audioset.py, there is a problem:
in line 111 : x,f,y,i=batch
problem: not enough values to unpack(expected4, got3)
I think the value i dose not have a value and I want to know how to use the teacher preds.
Thanks!
Thank you for your great work and sharing it!
Do you have any recommendation to use your models to label audio at a higher resolution, say 1 sec or lower? Or even mel frame level?
I've tried applying your models on short windows but below 5 seconds, the results deteriorate a lot (for 1sec it seems to fail completely). I guess it's because the training AudioSet samples are ~10 seconds long.
I've also tried to modify the model to obtain frame-level predictions but it seems that they all use the "mlp" head and getting rid of the adaptative pooling would require a full retrain?
Thank you in advance!
Hello!
Congratulations on the model, very impressive results. This is more of a question than it is an issue.
I was wondering if there is a configured way to extract the scene embeddings w/o performing the classification step on this repo.
Much appreciated,
Caio
model = get_dymn(pretrained_name="dymn20_as")
gives
In [8]: get_dymn(pretrained_name="dymn20_as")
/opt/miniconda3/lib/python3.9/site-packages/torchvision/ops/misc.py:120: UserWarning: Don't use ConvNormActivation directly, please use Conv2dNormActivation and Conv3dNormActivation instead.
warnings.warn(
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
Cell In[8], line 1
----> 1 get_dymn(pretrained_name="dymn20_as")
File ~/dev/goodvibes/vocaltechnique/EfficientAT/models/dymn/model.py:321, in get_model(num_classes, pretrained_name, width_mult, strides, context_ratio, max_context_size, min_context_size, dyrelu_k, no_dyrelu, dyconv_k, no_dyconv, T_max, T0_slope, T1_slope, T_min, pretrain_final_temp, no_ca, use_dy_blocks)
317 T_max = pretrain_final_temp
319 temp_schedule = (T_max, T_min, T0_slope, T1_slope)
--> 321 m = dymn(num_classes=num_classes,
322 pretrained_name=pretrained_name,
323 block=block,
324 width_mult=width_mult,
325 strides=strides,
326 context_ratio=context_ratio,
327 max_context_size=max_context_size,
328 min_context_size=min_context_size,
329 dyrelu_k=dyrelu_k,
330 dyconv_k=dyconv_k,
331 no_dyrelu=no_dyrelu,
332 no_dyconv=no_dyconv,
333 no_ca=no_ca,
334 temp_schedule=temp_schedule,
335 use_dy_blocks=use_dy_blocks
336 )
337 print(m)
338 return m
File ~/dev/goodvibes/vocaltechnique/EfficientAT/models/dymn/model.py:263, in dymn(pretrained_name, **kwargs)
261 def dymn(pretrained_name: str = None, **kwargs: Any):
262 inverted_residual_setting, last_channel = _dymn_conf(**kwargs)
--> 263 return _dymn(inverted_residual_setting, last_channel, pretrained_name, **kwargs)
File ~/dev/goodvibes/vocaltechnique/EfficientAT/models/dymn/model.py:257, in _dymn(inverted_residual_setting, last_channel, pretrained_name, **kwargs)
255 model.load_state_dict(state_dict, strict=False)
256 else:
--> 257 model.load_state_dict(state_dict)
258 return model
File /opt/miniconda3/lib/python3.9/site-packages/torch/nn/modules/module.py:2152, in Module.load_state_dict(self, state_dict, strict, assign)
2147 error_msgs.insert(
2148 0, 'Missing key(s) in state_dict: {}. '.format(
2149 ', '.join(f'"{k}"' for k in missing_keys)))
2151 if len(error_msgs) > 0:
-> 2152 raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
2153 self.__class__.__name__, "\n\t".join(error_msgs)))
2154 return _IncompatibleKeys(missing_keys, unexpected_keys)
RuntimeError: Error(s) in loading state_dict for DyMN:
size mismatch for layers.0.depth_conv.weight: copying a param with shape torch.Size([1, 1, 4, 288]) from checkpoint, the shape in current model is torch.Size([1, 1, 4, 144]).
size mismatch for layers.0.depth_conv.residuals.0.weight: copying a param with shape torch.Size([4, 64]) from checkpoint, the shape in current model is torch.Size([4, 32]).
size mismatch for layers.0.depth_norm.weight: copying a param with shape torch.Size([32]) from checkpoint, the shape in current model is torch.Size([16]).
size mismatch for layers.0.depth_norm.bias: copying a param with shape torch.Size([32]) from checkpoint, the shape in current model is torch.Size([16]).
size mismatch for layers.0.depth_norm.running_mean: copying a param with shape torch.Size([32]) from checkpoint, the shape in current model is torch.Size([16]).
size mismatch for layers.0.depth_norm.running_var: copying a param with shape torch.Size([32]) from checkpoint, the shape in current model is torch.Size([16]).
size mismatch for layers.0.depth_act.coef_net.0.weight: copying a param with shape torch.Size([128, 64]) from checkpoint, the shape in current model is torch.Size([64, 32]).
size mismatch for layers.0.depth_act.coef_net.0.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([64]).
size mismatch for layers.0.proj_conv.weight: copying a param with shape torch.Size([1, 1, 4, 1024]) from checkpoint, the shape in current model is torch.Size([1, 1, 4, 256]).
size mismatch for layers.0.proj_conv.residuals.0.weight: copying a param with shape torch.Size([4, 64]) from checkpoint, the shape in current model is torch.Size([4, 32]).
size mismatch for layers.0.proj_norm.weight: copying a param with shape torch.Size([32]) from checkpoint, the shape in current model is torch.Size([16]).
size mismatch for layers.0.proj_norm.bias: copying a param with shape torch.Size([32]) from checkpoint, the shape in current model is torch.Size([16]).
size mismatch for layers.0.proj_norm.running_mean: copying a param with shape torch.Size([32]) from checkpoint, the shape in current model is torch.Size([16]).
size mismatch for layers.0.proj_norm.running_var: copying a param with shape torch.Size([32]) from checkpoint, the shape in current model is torch.Size([16]).
size mismatch for layers.0.context_gen.joint_conv.weight: copying a param with shape torch.Size([64, 32, 1, 1]) from checkpoint, the shape in current model is torch.Size([32, 16, 1, 1]).
size mismatch for layers.0.context_gen.joint_norm.weight: copying a param with shape torch.Size([64]) from checkpoint, the shape in current model is torch.Size([32]).
size mismatch for layers.0.context_gen.joint_norm.bias: copying a param with shape torch.Size([64]) from checkpoint, the shape in current model is torch.Size([32]).
size mismatch for layers.0.context_gen.joint_norm.running_mean: copying a param with shape torch.Size([64]) from checkpoint, the shape in current model is torch.Size([32]).
size mismatch for layers.0.context_gen.joint_norm.running_var: copying a param with shape torch.Size([64]) from checkpoint, the shape in current model is torch.Size([32]).
size mismatch for layers.0.context_gen.conv_f.weight: copying a param with shape torch.Size([32, 64, 1, 1]) from checkpoint, the shape in current model is torch.Size([16, 32, 1, 1]).
size mismatch for layers.0.context_gen.conv_f.bias: copying a param with shape torch.Size([32]) from checkpoint, the shape in current model is torch.Size([16]).
size mismatch for layers.0.context_gen.conv_t.weight: copying a param with shape torch.Size([32, 64, 1, 1]) from checkpoint, the shape in current model is torch.Size([16, 32, 1, 1]).
size mismatch for layers.0.context_gen.conv_t.bias: copying a param with shape torch.Size([32]) from checkpoint, the shape in current model is torch.Size([16]).
size mismatch for layers.1.exp_conv.weight: copying a param with shape torch.Size([1, 1, 4, 4096]) from checkpoint, the shape in current model is torch.Size([1, 1, 4, 1024]).
size mismatch for layers.1.exp_conv.residuals.0.weight: copying a param with shape torch.Size([4, 64]) from checkpoint, the shape in current model is torch.Size([4, 32]).
size mismatch for layers.1.exp_norm.weight: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([64]).
size mismatch for layers.1.exp_norm.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([64]).
size mismatch for layers.1.exp_norm.running_mean: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([64]).
etc
could u provide the all config detail for reproducing the paper results? For example, the model with pretrained and without pretrained version. By the way, must i use the audio with resample rate 32k?
Hi,
I converted dymn10_as to pytorch mobile, But mobile output is so different with python.
I checked both torch version and the model file. What may be the problem? Thanks :)
Expected Behavior:
Same or similar outputs
Actual Behavior:
So different output
Reproduce:
1.convert dymn10_as to pytorch mobile(android)
2.compare python and mobile(android) output
if __name__ == '__main__':
model_name = 'dymn10_as'
model_input = torch.rand(1, 1, 128, 210)
ptmobile_name = 'eat_' + model_name + '_ptmobile.ptl'
if model_name.startswith('dymn'):
model = get_dymn(width_mult=NAME_TO_WIDTH(model_name), pretrained_name=model_name, strides=[2, 2, 2, 2])
else:
model = get_mn(width_mult=NAME_TO_WIDTH(model_name), pretrained_name=model_name, strides=[2, 2, 2, 2])
model.to(torch.device('cpu'))
model.eval()
model = torch.jit.trace(model, model_input)
print(model.code)
# https://github.com/pytorch/pytorch/issues/96639
# model = mobile_optimizer.optimize_for_mobile(model,
# {
# MobileOptimizerType.CONV_BN_FUSION,
# # I'm only disabling CONV_BN_FUSION
# # MobileOptimizerType.FUSE_ADD_RELU,
# # MobileOptimizerType.HOIST_CONV_PACKED_PARAMS,
# # MobileOptimizerType.INSERT_FOLD_PREPACK_OPS,
# # MobileOptimizerType.REMOVE_DROPOUT
# })
model._save_for_lite_interpreter(ptmobile_name)
print('android model save success!')
Why do you use nn.BCEWithLogitsLoss for distillation loss while other KD works using F.kl_div? Is it special setting for audio classification task?
i run this python ex_dcase20.py --cuda --pretrained --model_name=dymn04_as --cache_path=cache for custom dataset like urban dataset csv,but sitll fail,
File "/root/miniconda3/envs/EfficientAT/lib/python3.10/site-packages/torch/utils/data/_utils/worker.py", line 302, in _worker_loop
data = fetcher.fetch(index)
File "/root/miniconda3/envs/EfficientAT/lib/python3.10/site-packages/torch/utils/data/_utils/fetch.py", line 58, in fetch
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/root/miniconda3/envs/EfficientAT/lib/python3.10/site-packages/torch/utils/data/_utils/fetch.py", line 58, in
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/opt/EfficientAT/datasets/dcase20.py", line 129, in getitem
x = (x1 * l + x2 * (1. - l))
RuntimeError: The size of tensor a (17248) must match the size of tensor b (64000) at non-singleton dimension 1
can you help me
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
Django
The Web framework for perfectionists with deadlines.
Laravel
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
Microsoft
Open source projects and samples from Microsoft.
Google
Google ❤️ Open Source for everyone.
Alibaba
Alibaba Open Source for everyone
D3
Data-Driven Documents codes.
Tencent
China tencent open source team.