¿ªÆô×ó²à

ʹÓÃPyTorchʵÏÖÄñÀàÒôƵ¼ì²â¾í»ýÍøÂçÄ£ÐÍ

4279057

2020-7-16 15:44:31

ÒôÊÓÆµ
ʹÓÃPyTorchʵÏÖÄñÀàÒôƵ¼ì²â¾í»ýÍøÂçÄ£ÐÍ

ÒÔ¼°ÎªÊ²Ã´ÄñÀàµÄÉùÒô¼ì²â¶ÔÎÒÃÇ»·¾³µÄδÀ´Èç´ËÖØÒª
t01370dd55124de4c56.webp.jpg
ÄãÌý˵¹ý×Ô¶¯ÓïÒôʶ±ð£¬ÄãÌý˵¹ýÒôÀÖ±êÇ©ºÍÉú³É£¬µ«ÊÇÄãÌý˵¹ýÄñµÄÉùÒô¼ì²âÂð?
´óÔ¼ÔÚÒ»Äêǰ£¬ÔÚÎҸ߶þµÄʱºò£¬ÎÒµÚÒ»´ÎÌýµ½ÕâÖÖÒôƵÉî¶ÈѧϰµÄÓÃÀý¡£ÊÂʵÉÏ,ÄñÒôƵ¼ì²âÊÇÎÒ×öÉî¶ÈѧϰºÍ¼ÆËã»ú¿ÆÑ§µÄµÚÒ»¸öÏîÄ¿¡£ÎÒ²ÎÓëÁËÒ»¸öÑо¿ÏîÄ¿£¬ÔÚ±±°¢À­Ë¹¼ÓµÄ½¼ÇøÓô¿´âµÄÉùÒôÀ´Ì½²âÄñÀàµÄ´æÔÚ¡£ÌøÈëÆäÖУ¬ÄñµÄÒôƵ¼ì²â³öÏÖÁËÕâÑùÒ»¸öÀû»ù£¨ÓÐÀû¿ÉͼµÄÐÎʽ£©£¬ÔÚ±¾ÎÄÖУ¬ÎÒ½«ÏòÄúչʾÈçºÎÔÚBirdVox-70kÊý¾Ý¼¯ÉÏʹÓÃÒ»¸ö¼òµ¥µÄ¾í»ýÉñ¾­ÍøÂç(CNN)À´ÊµÏÖÕâÒ»µã¡£ÎªÊ²Ã´ÄñÀàµÄÉùÒô¼ì²âºÜÖØÒªÄØ?
¶àÄêÀ´£¬ÄñÀàÒôƵ¼ì²âÓ¦ÓÃÖÐÉî¶ÈѧϰģÐ͵ÄʹÓÃÒ»Ö±ÔÚ²»¶Ï·¢Õ¹½ø²½£¬ÕâÒ²ÊÇһЩÈ˶Դ˷dz£¸ÐÐËȤµÄÔ­Òò¡£Ê×ÏÈ£¬ÄñÀàÒÆ¶¯Ëٶȿ죬ÌåÐÍС£¬ÒѾ­ºÜÄÑ×·×Ù¡£´ËÍ⣬²»Í¬ÖÖÀàµÄÄñ¶¼ÓÐ×Ô¼º¶ÀÌØµÄ·¢Éù·½Ê½£¬ÕâʹµÃÄñÀàͨ¹ýÉùÒôºÍÌý¾õÀ´¼ì²â±Èͨ¹ýÊÓ¾õÀ´¼ì²â¸ü¼Ó¿ÉÈ¡¡£×îºó£¬ÓÉÓÚÄñÀàÊÇÉú̬ϵͳµÄÒ»²¿·Ö£¬ËüÃǵĴæÔÚºÍǨáãģʽÍùÍùÊÇÈκÎÌØ¶¨µØÇø»·¾³½¡¿µµÄ¾¯Ê¾Ðźš£
ÄÇô£¬ÎªÊ²Ã´²»Â¼ÖÆÒ»¶ÎÒôƵ£¬È»ºó·¢Ë͸øÈËÀàÉÔºóÔÙÌýÄØ?
ÊÖ¶¯±ê¼ÇÒôƵÊǰº¹óµÄ£¬·¦Î¶µÄ£¬¶øÇÒ¿ÉÄܲ»½Ó½üʵʱ¡£
ËùÒÔ£¬Õâ¾ÍÊÇÉî¶ÈѧϰºÍcnn·¢»Ó×÷Óõĵط½¡£Èç¹ûÎÒÃǵÄÄ£ÐÍ×ã¹»¾«È·£¬ÎÒÃÇ¿ÉÒÔͨ¹ýÔÚÒ°ÍâÉèÖÃÂó¿Ë·çÀ´×Ô¶¯¼Ç¼ÄñÀàǨáãģʽºÍ×·×ÙÄñÀàÖÖÀà¡£ÓÐÁËÕâЩÊý¾Ý£¬ÎÒÃÇ¿ÉÒÔ¸üÉîÈëµØÁ˽âÊÀ½çµÄÉú̬ϵͳ£¬²¢³ä·ÖÁË½âÆøºò±ä»¯¶Ô»·¾³µÄÓ°Ïì¡£
ÄÇô£¬ÉÏÃæÕâÕÅͼÊÇÊ²Ã´ÄØ?
ÈÃÎÒÃÇÏÈ̸̸ÉùÒô¡£ÉùÒô»ù±¾ÉÏÊÇÓÉÆæÌصÄѹÁ¦²¨×é³ÉµÄ£¬ÕâЩѹÁ¦²¨Í¨¹ý¿ÕÆø½øÈëÎÒÃǵĶúĤ£¬¿ÉÒÔ±»Êý×ÖÉ豸×÷ΪÐźżÇ¼ÏÂÀ´¡£½á¹û±íÃ÷£¬ÎÒÃÇ¿ÉÒÔ½«Õâ¸öÐźŻæÖÆÈçÏÂͼ:

t010acfc7bce05f2043.webp.jpg
È»¶ø£¬ÕâÖÖ³¤Ïßͼ¿´ÆðÀ´²¢²»ÊÇÌØ±ðÊʺÏÓÃÓÚÈκÎÉî¶ÈѧϰģÐÍ£¬¸ü²»ÓÃ˵CNNÁË¡£ÊÂʵÉÏ£¬Èç¹ûÎÒÏÖÔÚ¸æËßÄ㣬ÉÏÃæµÄÒôƵÐźŴú±íÁËÒ»Ö»ÄñµÄÃù½Ð£¬Äã»áÏàÐÅÎÒÂð?Äã·´¶ø»áÈÏΪÕâÊÇÎÒËæÒâ´´×÷µÄÇé½Ú¡£Èç¹ûÄãÈÏΪÄñÃùÉùÊÇÔÚµÚ10000 ²¨¶¯µã ×óÓÒ£¬Èç¹ûÎÒ˵ÄñÃùÉùÊÇÔÚÕýÖмäÔÚµÚ6000²¨¶¯µã?
ÄÇô£¬Èç¹ûÎÒÃÇ×Ô¼º¶¼²»ÄÜ×ö³öÈκÎÇ¿ÓÐÁ¦µÄ¼ÙÉ裬ÄÇôÉî¶ÈѧϰģÐÍÓÖÈçºÎÄÜ×öµ½ÕâÒ»µãÄØ?
ÎÒÃDz¢Ã»ÓÐÍêÈ«ÏÝÈëÀ§¾³¡£Êµ¼ÊÉÏ»¹ÓÐÁíÒ»ÖÖ±íʾÉùÒôµÄ·½·¨:ÉùÆ×ͼ¡£ÆµÆ×ͼÊÇͨ¹ýÔÚÒ»¸öСʱ¼ä´°¿ÚÖмǼƵÂʵĴæÔں͸÷×ÔµÄÇ¿¶È£¬ÑØ×ÅxÖá(ÓÃʱ¼äµ¥Î»±íʾ)¶Ñµþ£¬Ö±µ½ÆµÆ×ͼ¸²¸ÇÒôƵÐźŵÄ×ܳÖÐøÊ±¼ä¡£ÔÚ´´½¨Æ×ͼµÄ¹ý³ÌÖУ¬Ê±¼ä´°±¾Éí»á·¢ÉúÖØµþ£¬Í¨³£ÆµÂÊÇ¿¶È(ÒôÁ¿»òÏì¶È)ÓÃÑÕÉ«±íʾ£¬»òÕßÓÃÊý×ÖÀ´±íʾ¸ß/µÍÖµ¡£
t01d337510406173c3c.webp.jpg
´ÓÉÏÃæËùʾµÄÍêÈ«ÏàͬµÄ²¨ÐÎÖжÍÔì³öµÄ¹âÆ×ͼ¡£xÖá¡¢yÖáºÍ±ê»æÑÕÉ«·Ö±ð±íʾʱ¼äµ¥Î»¡¢ÆµÂÊ(ºÕ×È)ºÍƵÂÊÇ¿¶È¡£ÏÖÔÚ£¬ÄãÄÜÇå³þµØ·Ö±æ³öÄñ¶ùÊÇ·ñÃù½ÐÒÔ¼°ºÎʱÃù½ÐÂð?
Èç¹ûÓÐÒ»¸öÀ´×ÔÉÏÃæµÄÐźŵĹâÆ×ͼ£¬ÄÇôÄñ¶ùÊÇ·ñÃù½ÐÒÔ¼°ºÎʱÃù½Ð¾Í»áÇå³þµÃ¶à¡£¾ßÌåÀ´Ëµ£¬¡°Ö»ÊÇÒòΪ¡±£¬Äã¿ÉÒÔ¿´µ½¶ÀÌØµÄ»î¶¯´óÔ¼6000ºÕ×ÈÔÚ250ºÁÃëµÄÂí¿Ë;ÄÇÊÇÄñµÄ½ÐÉù¡£Òò´Ë£¬CNNÒôƵ·ÖÀàÆ÷¾­³£ÒÔ¹âÆ×ͼ×÷ΪÊäÈ룬Äñ½ÐÉùµÄÒôƵ¼ì²âÄ£ÐÍÒ²²»ÀýÍâ¡£×¼±¸
Ϊ´Ë£¬ÎÒʹÓÃÁËBirdVoxµÄBirdVox-70kÊý¾Ý¼¯£¬¸ÃÊý¾Ý¼¯°üº¬°ëÃë(500ms)µÄ¼Òô/²¨ÐΣ¬ÆäÖаüº¬»ò²»°üº¬Äñ½ÐÉù¡£Êý¾Ý¼¯À´×ÔÓÚËüµÄ´ó¸çBirdVox-full-night£¬ÕâÁ½ÖÖÄñ¶¼ÊÇÔÚ2015ÄêÇïÌìÔÚÃÀ¹úŦԼÖݵÄIthaca±»¼Ç¼µÄ¡£Ã¿¸ö²¨ÐεıêÇ©°üº¬ÔÚËüµÄÿ¸öÎļþÃû×÷Ϊ×îºóÒ»¸öÊý×Ö(0»ò1)¡£¾¡¹Ü70 kÊý¾Ý¼¯ÓµÓдóÔ¼70000µÄÊÓÆµÊý¾Ý(Òò´ËµÃÃû),ÎÒÖ»»áÓÃÈý·ÖÖ®¶þµÄ(¸ü¾ßÌåµØËµ,50000×óÓÒ)¡£
Êý¾Ý±»À¦°óÔÚ¶ÀÁ¢µÄHDF5ÎļþÖУ¬ÕâÒâζ×ÅÎҼƻ®Ê¹ÓõÄÒѾ­¼õÉÙµÄ50,000¸öÑù±¾Êµ¼ÊÉϱ»·Ö¸î³É4¸ö²»Í¬µÄ¡£HDF5Îļþ¡£Ã¿¸ö²¨Ðζ¼±»´æ´¢³ÉÒ»×飬ÓÐËü×Ô¼ºµÄ¡°ÎļþÃû¡±±êÇ©£¬ËùÓÐÕâЩ¶¼±»´æ´¢µ½³ÆÎª¡°²¨ÐΡ±µÄÁíÒ»×é¡£
¼´Ê¹ÔÚÔĶÁÁËh5pyÎĵµÖ®ºó£¬ÎÒҲûÓп´µ½ÏñÉÏÃæÃèÊöµÄÄÇÑùµÄ¿ª·¢¡£ÕâЩ·¢ÏÖÒѾ­ÈÃÎÒ×öÁËÁ½¸ö×îÔã¸âµÄÊý¾Ý×¼±¸µÄجÃΡ£ÒòΪ´æ´¢¿Õ¼äµÄÔ­Òò£¬ÎÒÎÞ·¨¸Ä±äÒѾ­¸øÎÒµÄÊý¾ÝµÄ¸ñʽ¡£Õâ¸øÎÒÁôÏÂÁË×îºóÒ»¸öÑ¡Ïî:´´½¨Ò»¸ö·Ç³£×Ô¶¨ÒåµÄPyTorchÊý¾Ý¼¯ÀàDataSet£¬ÕâÑùÎҾͿÉÒÔ¼ÌÐø¹¤×÷¡£
µ«ÔÚÎÒ¼ÌÐøÖ®Ç°£¬ÈÃÎÒÃÇÌýһЩ¼ò¶ÌµÄÒôƵÑù±¾£¬ÒÔ¼°Ò»Ð©¼òµ¥µÄÊý¾Ý·ÖÎöÊý¾Ý¼¯(¾¯¸æ-½µµÍÄãµÄÒôÁ¿£¬ÒòΪÑù±¾¿ÉÄܱÈÔ¤ÆÚµÄ¸ü´ó):
ÔÚ°ÑËùÓеÄÎļþ×éºÏÔÚÒ»ÆðÖ®ºó£¬bird-positiveÒôƵÑù±¾µÄ±ÈÀýÕýºÃÊÇ50?¾ÍÏñËûÃÇÍøÕ¾ÉϳÐŵµÄÄÇÑù¡£Ã¿¸öÄñÕý²¨ÐεÄÎļþÃû»¹°üÀ¨Äñ½ÐÉùµÄƵÂÊ¡£°ÑËùÓÐÕâЩÕûÀíÔÚÒ»Æð£¬ÎÒ·¢ÏÖÈçÏÂ:
t01fee686fdf91f235b.webp.jpg
ºÜ¶àµç»°µÄÒôµ÷»áºÜ¸ß¡£´ó²¿·ÖµÄÑù±¾ÊÇÕâÑùµÄ£¬µ«ÊǾø´ó¶àÊýʵ¼ÊÉϸü½Ó½ü½ÏµÍµÄƵÂÊ(2000-3500ºÕ×È)¡£ºÜÆæ¹Ö!Dataset & DataloaderÀà
ÊÕ²Ø
ËÍÔÞ
·ÖÏí

·¢±í»Ø¸´

µÇ¼ºó²ÅÄÜÆÀÂÛ

ÆÀÂÛÁÐ±í£¨1£©

4279057 Â¥Ö÷
ÎÒÃÇÖªµÀÓÐһЩ¶ÀÁ¢µÄ¡£hdf5ÎļþÒª´¦Àí³ÉÒ»¸ö¡°Êý¾Ý¼¯¡±£¬¶øÇÒÿ¸öÎļþ¶¼ÓÐÒ»Ð©Ææ¹ÖµÄÊý¾Ý½á¹¹¡£ÕâЩÌõ¼þ¿Ï¶¨ÐèÒªÒ»¸ö×Ô¶¨ÒåPyTorch DatasetÀàÀ´ÕýÈ·¼ÓÔØÒôƵÊý¾Ý£¬ÒÔ±ãÒÔºó½øÐÐѵÁ·¡£DatasetÀàÖ÷ÒªÐèҪʣϵÄ__init__ºÍ__getitem__º¯Êý(»¹ÓÐʣϵÄ__len__£¬µ«ÕâºÜ¼òµ¥)¡£ÒÔÏÂÊÇÎҼƻ®Òª×öµÄÊÂÇé:
__init__
±éÀúËùÓÐËĸöÎļþÖеÄÿ¸ö²¨ÐεÄÿ¸ö×éÃû£¬²¢½«ÆäËùÊôµÄÎļþºÍHDF5×é×·¼Óµ½ÊôÓÚ¸ÃÀàµÄÁбíÖС£
»ñÈ¡²¨ÐεıêÇ©(Îļþ/×éÃûµÄ×îºóһλ)²¢½«Æä¸½¼Óµ½ÊôÓÚ¸ÃÀàµÄÁíÒ»¸öÁбíÖС£ÀýÈ磬ÎļþÃûunit0301271222800000_0±íʾûÓÐÄñ½ÐÉù¡£
×¼±¸Ò»¸ö±ä»»º¯Êý£¬Ó¦Óõ½²¨ÐΣ¬½«Æäת»»Îª¹âÆ×ͼ(¾ßÌåµØËµ£¬mel-scaledµÄ¹âÆ×ͼ)£¬²¢¿¼Âǵ½ÆäËûÔöÇ¿¼¼Êõ¡£
__getitem__
ΪÔÚ³õʼ»¯Ö®Ê±´´½¨µÄÁбíÌṩË÷Òý
Ò»µ©Í¨¹ýÁбí½ÓÊÕµ½²¨ÐεÄλÖ㬴ò¿ª¸Ã²¨ÐεÄHDF5Îļþ¡£ËùÓеÄHDF5 I/O¶¼½«Ê¹ÓÃpython¿âh5pyÀ´´¦Àí
°ÑËü±ä³ÉPyTorchÕÅÁ¿²¢Ó¦ÓÃÈκα任£¬°üÀ¨Æ×ͼ±ä»»¡£
½«ÏîÄ¿·µ»Ø¸øµ÷ÓÃÕß
ÏÂÃæ£¬DatasetÀàµÄ´úÂë:
class BirdVox70kDS(Dataset):
   def __init__(self, root_dir, fnames, transforms=None):
       # store transforms func
       self.transforms = transforms
       # initialize storage arrays
       self.wave_loc = []
       self.labels = []

       # for each hdf5 file...
       for fname in fnames:
           # open the file
           fhdf5 = os.path.join(root_dir, fname)
           with h5py.File(fhdf5, 'r') as f:
               # navigate to `waveforms` group
               waveforms = f['waveforms']
               # for each piece of data...
               for waveform in waveforms.keys():
                   # append waveform filename for later access
                   self.wave_loc.append([fhdf5, waveform])
                   # (label == last digit of filename)
                   self.labels.append(waveform[-1])

       # turn them into np.arrays
       self.wave_loc = np.array(self.wave_loc)
       self.labels = np.array(self.labels)

       # melspec transform (similar to `librosa.feature.melspectrogram()`)
       self.melspec = T.MelSpectrogram(sample_rate=24000,
                                       n_fft=2048,
                                       hop_length=512)

   def __len__(self):
       # size of dataset
       return len(self.labels)

   def __getitem__(self, idx):
       # fetch waveform from hdf5 file & label
       fhdf5, waveform = self.wave_loc[idx]
       with h5py.File(fhdf5, 'r') as f:
           wave = f['waveforms'][waveform]
           # convert to np array for faster i/o performance

           wave = np.array(wave)
           # apply other specified transforms
           if self.transforms:
               wave = self.transforms()(wave)
           # convert into tensor & apply melspec
           wave = self.melspec(torch.Tensor(wave))
           # unsqueeze adds dimension needed for pytorch's `Conv2d`
           wave = wave.unsqueeze(0)
       # parse label (still a string)
       label = self.labels[idx]
       return wave, int(label)
ÎÒÃǶ¨ÒåÊý¾Ý¼¯ÎļþµÄĿ¼ºÍÎļþÃû±¾Éí¡£ÎÒ¾ö¶¨Ê¹ÓÃ4¸öÎļþÖеÄ3¸ö×÷Ϊ²âÊÔÊý¾Ý£¬×îºóÒ»¸ö×÷ΪÑéÖ¤/²âÊÔ¼¯À´¶ÈÁ¿Ä£Ð͵ÄÐÔÄÜ£¬ÎªºóÕßÁôÏÂ×îСµÄÎļþ¡£ÓÉÓÚÿ¸öÎļþµÄËùÓмǼÉ豸¶¼ÉèÖõıȽϽü(ÔÚÒÁÈø¿¨µÄͬһ¸ö³ÇÊÐ)£¬ÎÒÈÏÎªÈÆ¹ýËæ»ú·Ö¸î²»»áÒýÈë´óÁ¿µÄÆ«²î¡£
root_dir = '/notebooks/storage/'
fnames = ['BirdVox-70k_unit03.hdf5',
         'BirdVox-70k_unit07.hdf5',
         'BirdVox-70k_unit10.hdf5']
train_ds = BirdVox70kDS(root_dir, fnames)
val_ds = BirdVox70kDS(root_dir, ['BirdVox-70k_unit01.hdf5'])
ÔÚÎÒÃǵÄÊý¾Ý¼¯Ö®ºó£¬ÏÂÃæ¾ÍÊÇʹÓÃdataloadersÁË:
batch_size = 128
train_dl = DataLoader(train_ds, batch_size, shuffle=True, pin_memory=True)
val_dl = DataLoader(val_ds, batch_size * 2, pin_memory=True)
# having `num_workers > 1` will crash dataloader when working w/ h5 files

¶¨Òådataloaders£¬Ëü½«·ÖÅú·µ»ØÊý¾Ý¡£ÔÚʹÓÃPyTorchºÍHDF5Îļþʱ£¬ÎÒ³¢ÊÔ¹ýÉèÖöà¸ö¡°num_workers¡±£¬µ«·¢ÏÖ´æÔÚÒ»¸öbugÄ£ÐÍ
ÎÒΪÎÒµÄÄ£ÐÍÉèÖÃÁ˱ØÒªµÄ¸¨Öúº¯Êý£¬ÒÔ±ãÒÔºó½øÐÐѵÁ·:
class ModelBase(nn.Module):

   # defines mechanism when training each batch in dl
   def train_step(self, batch):
       xb, labels = batch
       outs = self(xb)
       loss = F.cross_entropy(outs, labels)
       return loss

   # similar to `train_step`, but includes acc calculation & detach
   def val_step(self, batch):
       xb, labels = batch
       outs = self(xb)
       loss = F.cross_entropy(outs, labels)
       acc = accuracy(outs, labels)
       return {'loss': loss.detach(), 'acc': acc.detach()}

   # average out losses & accuracies from validation epoch
   def val_epoch_end(self, outputs):
       batch_loss = [x['loss'] for x in outputs]
       batch_acc = [x['acc'] for x in outputs]
       avg_loss = torch.stack(batch_loss).mean()
       avg_acc = torch.stack(batch_acc).mean()
       return {'avg_loss': avg_loss, 'avg_acc': avg_acc}

   # print all data once done
   def epoch_end(self, epoch, avgs, test=False):
       s = 'test' if test else 'val'
       print(f'Epoch #{epoch + 1}, {s}_loss:{avgs["avg_loss"]}, {s}_acc:{avgs["avg_acc"]}')
¶¨Òå¶à¸öº¯Êý£¬ÒÔºó¿ÉÒÔʹÓÃÕâЩº¯ÊýѵÁ·¼Ì³ÐÕâ¸öÀàµÄPyTorchÄ£ÐÍ¡£
def accuracy(outs, labels):
   _, preds = torch.max(outs, dim=1)
   return torch.tensor(torch.sum(preds == labels).item() / len(preds))
Õâ¸öº¯ÊýÔÚÉÏÊöÀàµÄval_stepº¯ÊýÖб»ÓÃÀ´È·¶¨ÑéÖ¤dataloaderÉÏÄ£Ð͵Ä?È·ÐÔ¡£
²¢¶¨ÒåÓÃÓÚÄâºÏ/ѵÁ·Ä£ÐͺÍÔÚÑéÖ¤Êý¾Ý¼¯ÉϲâÊÔÄ£Ð͵ÄÖ÷Òª¹¦ÄÜ
@torch.no_grad()
def evaluate(model, val_dl):
   # eval mode
   model.eval()
   outputs = [model.val_step(batch) for batch in val_dl]
   return model.val_epoch_end(outputs)


def fit(epochs, lr, model, train_dl, val_dl, opt_func=torch.optim.Adam):
   torch.cuda.empty_cache()
   history = []
   # define optimizer
   optimizer = opt_func(model.parameters(), lr)
   # for each epoch...
   for epoch in range(epochs):
       # training mode
       model.train()
       # (training) for each batch in train_dl...
       for batch in tqdm(train_dl):
           # pass thru model
           loss = model.train_step(batch)
           # perform gradient descent
           loss.backward()
           optimizer.step()
           optimizer.zero_grad()
       # validation
       res = evaluate(model, val_dl)
       # print everything useful
       model.epoch_end(epoch, res, test=False)
       # append to history
       history.append(res)
   return history
×îºó£¬ÕâÊÇÎÒÃǵȴýÒѾõļòµ¥CNNÄ£ÐÍ:
class Classifier(ModelBase):
   def __init__(self):
       super().__init__()                                      # 1 x 128 x 24
       self.conv1 = nn.Conv2d(1, 4, kernel_size=3, padding=1)  # 4 x 128 x 24
       self.conv2 = nn.Conv2d(4, 8, kernel_size=3, padding=1)  # 8 x 128 x 24
       self.bm1 = nn.MaxPool2d(2)                              # 8 x 64 x 12
       self.conv3 = nn.Conv2d(8, 8, kernel_size=3, padding=1)  # 8 x 64 x 12
       self.bm2 = nn.MaxPool2d(2)                              # 8 x 32 x 6
       self.fc1 = nn.Linear(8*32*6, 64)
       self.fc2 = nn.Linear(64, 2)

   def forward(self, xb):
       out = F.relu(self.conv1(xb))
       out = F.relu(self.conv2(out))
       out = self.bm1(out)
       out = F.relu(self.conv3(out))
       out = self.bm2(out)
       out = torch.flatten(out, 1)
       out = F.relu(self.fc1(out))
       out = self.fc2(out)
       return out
ÎÒʹÓÃÁ˶à¸ö¾í»ý²ã£¬ÕýÈçÎÒÃÇ֮ǰµÄÀíÂÛÍÆ¶ÏËù½¨ÒéµÄÄÇÑù£¬ÎÒÃǵÄÄ£ÐÍʹÓÃÁËһЩ×î´ó³Ø»¯²ã£¬È»ºóʹÓÃÒ»¸ö·Ç³£¼òµ¥µÄÈ«Á¬½ÓÍøÂçÀ´½øÐÐʵ¼ÊµÄ·ÖÀà¡£ÁîÈ˾ªÑȵÄÊÇ£¬Õâ¸ö¼Ü¹¹ºóÀ´±íÏÖµÃÏ൱ºÃ£¬ÉõÖÁ³¬¹ýÁËÎÒ×Ô¼ºµÄÔ¤ÆÚ¡£ÀûÓÃGPU
¼¸ºõÿ¸öÈ˶¼ÐèÒªGPUÀ´ÑµÁ·±ÈÒ»°ãµÄǰÀ¡Éñ¾­ÍøÂç¸ü¸´ÔӵĶ«Î÷¡£ÐÒÔ˵ÄÊÇ£¬PyTorchÈÃÎÒÃÇ¿ÉÒÔºÜÈÝÒ×µØÀûÓÃÏÖÓÐGPUµÄÄÜÁ¦¡£Ê×ÏÈ£¬ÎÒÃǽ«ÎÒÃǵÄcudaÉ豸¶¨ÒåΪ¹Ø¼ü´ÊÉ豸£¬ÒÔ±ã¸üÈÝÒ×·ÃÎÊ:
device = 'cuda' if torch.cuda.is_available() else 'cpu'
ÎÒÃÇ»¹È·±£Èç¹ûûÓÐGPU, CPU»á±»Ê¹Óá£
ÕâÀﻹÓÐÁíÒ»¸ö¼¼ÇÉ:
torch.backends.cudnn.benchmark = True
Õâ¿ÉÒÔ°ïÖúÌá¸ßÄãµÄѵÁ·ËÙ¶È(Èç¹ûÄãµÄÊäÈëÔÚ´óС/ÐÎ×´ÉÏûÓб仯)
ÏÔÈ»£¬Äã¿ÉÒÔ¡°¸æËß¡±PyTorchÔÚÒ»´ÎÓÖÒ»´ÎµÄѵÁ·ÖÐÓÅ»¯×Ô¼º£¬Ö»ÒªÑµÁ·ÊäÈëÔÚ´óСºÍÐÎ×´Éϱ£³Ö²»±ä¡£Ëü»áÖªµÀΪÄãµÄÌØ¶¨Ó²¼þ(GPU)ʹÓÃ×î¿ìµÄËã·¨¡£!
È»ºóÎÒÃǶ¨Òå°ïÖúº¯ÊýÀ´Òƶ¯dataloaders¡¢tensorsºÍÎÒÃǵÄÄ£Ð͵½ÎÒÃǵÄGPUÉ豸¡£
def to_device(data, device):
   """Move tensor(s) to chosen device"""
   if isinstance(data, (list, tuple)):
       return [to_device(x, device) for x in data]
   return data.to(device, non_blocking=True)


class DeviceDataLoader():
   """Wrap a dataloader to move data to a device"""

   def __init__(self, dl, device):
       self.dl = dl
       self.device = device

   def __iter__(self):
       """Yield a batch of data after moving it to device"""
       for b in self.dl:
           yield to_device(b, self.device)

   def __len__(self):
       """Number of batches"""
       return len(self.dl)ѵÁ·
ʹÓÃÎÒÃǵÄÉ豸¸¨Öú¹¦ÄÜ£¬ÎÒÃǽ«Ò»ÇÐÒÆ¶¯µ½ÎÒÃǵÄGPUÈçÏÂ:
train_dl = DeviceDataLoader(train_dl, device)
val_dl = DeviceDataLoader(val_dl, device)
model = to_device(Classifier(), device)
ÎÒÃÇ»¹Ö¸¶¨ÁËÎÒÃǵÄѧϰËٶȺÍÎÒÃǵÄѵÁ·´ÎÊý£¬ÎÒ»¨Á˺ܳ¤Ê±¼äÀ´ÕÒµ½Ò»¸öºÃµÄÖµ¡£
lr = 1e-5
epochs = 8
ÔÚ½øÐÐÈκÎѵÁ·Ö®Ç°£¬ÎÒÃǻᷢÏÖÄ£Ð͵ıíÏÖ:
history = [evaluate(model, val_dl)]
[{'avg_loss': tensor(0.7133, device='cuda:0'), 'avg_acc': tensor(0.6042)}]
ºÜÏÔÈ»£¬ÎÒÃÇÔÚ¿ªÊ¼Ê±µÄ׼ȷÐÔÉÏÓеãÐÒÔË£¬µ«¾­¹ý¶à´ÎÊÔÑ飬×îÖÕ½á¹û¼´Ê¹²»Ïàͬ£¬Ò²»á·Ç³£ÏàËÆ¡£
½ÓÏÂÀ´£¬ÎÒÃÇʹÓÃ֮ǰ¶¨ÒåµÄfitº¯ÊýÀ´ÑµÁ·ÎÒÃǵļòµ¥·ÖÀàÆ÷Ä£ÐÍʵÀý:
history += fit(epochs, lr, model, train_dl, val_dl)
HBox(children=(FloatProgress(value=0.0, max=352.0), HTML(value='')))
Epoch #1, val_loss:0.6354132294654846, val_acc:0.7176136374473572
HBox(children=(FloatProgress(value=0.0, max=352.0), HTML(value='')))
Epoch #2, val_loss:0.6065077781677246, val_acc:0.7439352869987488
HBox(children=(FloatProgress(value=0.0, max=352.0), HTML(value='')))
Epoch #3, val_loss:0.56722491979599, val_acc:0.77376788854599
HBox(children=(FloatProgress(value=0.0, max=352.0), HTML(value='')))
Epoch #4, val_loss:0.5528884530067444, val_acc:0.7751822471618652
HBox(children=(FloatProgress(value=0.0, max=352.0), HTML(value='')))
Epoch #5, val_loss:0.5130119323730469, val_acc:0.8004600405693054
HBox(children=(FloatProgress(value=0.0, max=352.0), HTML(value='')))
Epoch #6, val_loss:0.4849482774734497, val_acc:0.8157732486724854
HBox(children=(FloatProgress(value=0.0, max=352.0), HTML(value='')))
Epoch #7, val_loss:0.4655478596687317, val_acc:0.8293880224227905
HBox(children=(FloatProgress(value=0.0, max=352.0), HTML(value='')))
Epoch #8, val_loss:0.4765000343322754, val_acc:0.8155447244644165
ÔÚ·¢ÏÖ¾«È·¶ÈÓÐÒ»µãϽµºó£¬ÎÒ¾ö¶¨ÔÙѵÁ·Ò»µãµã£¬ÈÃÄ£ÐÍÖØ¶¨Ïò»Øµ½ÕýÈ·µÄ·½Ïò:
history += fit(3, 1e-6, model, train_dl, val_dl)
HBox(children=(FloatProgress(value=0.0, max=352.0), HTML(value='')))
Epoch #1, val_loss:0.4524107873439789, val_acc:0.8329823613166809
HBox(children=(FloatProgress(value=0.0, max=352.0), HTML(value='')))
Epoch #2, val_loss:0.44666698575019836, val_acc:0.8373703360557556
HBox(children=(FloatProgress(value=0.0, max=352.0), HTML(value='')))
Epoch #3, val_loss:0.4442901611328125, val_acc:0.8412765860557556
ºöÂÔ¡°HBox¡±¹¤¼þ£¬ÕâÊÇtqdmÌṩµÄ£¬Çë¶à¹Ø×¢×¼È·ÐÔ!
×ܽáһϣ¬ÒÔÏÂÊÇÎÒÃǹýÈ¥11¸öʱÆÚµÄѵÁ·Í³¼ÆÊý¾Ý:
plt.plot([x['avg_loss'] for x in history])
plt.title('Losses over epochs')
plt.xlabel('epochs')
plt.ylabel('loss')
plt.show()
t01f7ec4b075035d9cb.webp.jpg
plt.plot([x['avg_acc'] for x in history])
plt.title('Accuracy over epochs')
plt.xlabel('epochs')
plt.ylabel('acc')
plt.show()
t019c5078e95a32f8e4.webp.jpg
×ܵÄÀ´Ëµ£¬ÎÒÃǵÄÄ£ÐÍѵÁ·µÃÏ൱ºÃ£¬´ÓËüµÄÍâ¹ÛÀ´¿´£¬ÎÒÃÇ¿ÉÄÜÒѾ­ÎªÎÒÃǵÄÄ£Ð͵ÄËðʧÕÒµ½ÁËÒ»¸öÏà¶Ô×îСµÄÖµ¡£
µÈµÈ£¬Ò»¸ö¸ü¸´ÔÓµÄÄ£ÐÍ»òÕßʹÓò»Í¬µÄת»»ÔõôÑù?
ÏàÐÅÎÒ£¬ÔÚÎÒµÄsimple Classifier()µÚÒ»´Î³¢ÊԳɹ¦Ö®ºó²»¾Ã£¬ÎÒÒ²³¢ÊÔ¹ýÕâÁ½ÖÖ·½·¨¡£ÎÒ¾ö¶¨²»°üÀ¨ÕâÁ½¸öϸ½Ú£¬ÒòΪÎÒ·¢ÏÖËûÃǵĽá¹ûʵ¼ÊÉϱÈÎÒÃÇÒѾ­È¡µÃµÄ½á¹û¸üÔã¸â£¬ÕâºÜÆæ¹Ö¡£
¶ÔÓÚ¶îÍâµÄÆ×ͼת»»£¬ÎÒ³¢ÊÔÁËËæ»úÊ±ÒÆºÍÔëÉù×¢Èë¡£³¤»°¶Ì˵£¬ËüËÆºõ¸ù±¾Ã»ÓÐÌá¸ßÑéÖ¤µÄ׼ȷÐÔ¡£´Ëºó,ÎÒÈÏΪÊÇÕâÑù,ÒòΪÊý¾Ý¼¯¹æ·¶Ã÷È·±íʾ,ËùÓÐà±à±½«Î»ÓÚÖмäµÄ¼Òô,Òò´ËËæ»ú±ä»¯µÄ¹âÆ×ͼµÄÄ¿µÄÔÊÐí¸üºÃµÄÄ£ÐÍ·º»¯Êµ¼ÊÉÏ¿ÉÄÜÒѾ­×÷ΪË𺦵ıíÏÖ¡£È»¶ø£¬ÎÒ»¹Ã»Óг¢ÊÔ¹ýËæ»úÔëÉù×¢Èë¡£
ÎÒ»¹³¢ÊÔѵÁ·Ò»¸öResNet50Ä£ÐÍ£¬Ï£Íû½øÒ»²½Ìá¸ßÑéÖ¤µÄ׼ȷÐÔ¡£ÕâÊÇ×îÁîÈËÄÑÒÔÖÃÐŵIJ¿·Ö:ÎÒµÄÄ£ÐÍ´ÓÀ´Ã»Óг¬¹ý50?׼ȷÂÊ!Ö±µ½½ñÌìÎÒдÕâÆªÎÄÕµÄʱºò£¬ÎÒ»¹²»È·¶¨ÎÒ×ö´íÁËʲô£¬ËùÒÔÈç¹ûÆäËûÈËÄÜ¿´¿´±Ê¼Ç±¾²¢°ïÖúÎÒ£¬ÎҺܸßÐËÊÕµ½Èκν¨Òé!½áÂÛ
×ܶøÑÔÖ®£¬ÕâÊÇÒ»¸öÕæÕýÓÐȤµÄŬÁ¦£¬»¨Ò»Ð©Ê±¼ä½øÐÐÑо¿¡£Ê×ÏÈ£¬ÎÒµÃÖØÐÂÉóÊÓÎÒÈ¥ÄêÏÄÌìµ÷²é¹ýµÄ¶«Î÷£¬ÎÞ¿É·ñÈÏ£¬ÕâÓÐÒ»ÖÖ»³¾ÉµÄ¸Ð¾õ¡£¸üÖØÒªµÄÊÇ£¬ÎÒÃÇѧϰÁËÈçºÎʵÏÖÒ»¸öºÜ¿ÉÄÜÓÃÓÚÕæÊµ³¡¾°µÄPyTorchÊý¾Ý¼¯À࣬ÔÚÕæÊµ³¡¾°ÖУ¬Êý¾Ý²»Ò»¶¨ÏñÄúÔ¤ÆÚµÄÄÇÑùÉèÖá£×îºó£¬×îÖÕµÄÑéÖ¤·ÖÊýΪ84?¶ÔÓÚÎÒ¼´ÐË´´½¨µÄÈç´Ë¼òµ¥µÄÍøÂç¼Ü¹¹À´Ëµ£¬ÕâÊÇÏ൱Õû½àµÄ!
2020-7-16 15:48:42 »Ø¸´

ÒôƵ·ÖÀà´óÈ«

+ ²é¿´È«²¿·ÖÀà