{"id":355973,"date":"2024-05-20T23:19:52","date_gmt":"2024-05-20T23:19:52","guid":{"rendered":"http:\/\/savepearlharbor.com\/?p=355973"},"modified":"-0001-11-30T00:00:00","modified_gmt":"2024-05-20T23:19:52","slug":"","status":"publish","type":"post","link":"https:\/\/savepearlharbor.com\/?p=355973","title":{"rendered":"\u041d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u044c, \u0433\u0435\u043d\u0435\u0440\u0438\u0440\u0443\u044e\u0449\u0430\u044f \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0438. \u0427\u0430\u0441\u0442\u044c 1. \u0413\u0435\u043d\u0435\u0440\u0430\u0446\u0438\u044f \u0441\u043b\u0443\u0447\u0430\u0439\u043d\u043e\u0439 \u0441\u0432\u0451\u0440\u0442\u043e\u0447\u043d\u043e\u0439 \u043d\u0435\u0439\u0440\u043e\u043d\u043d\u043e\u0439 \u0441\u0435\u0442\u0438<\/span>"},"content":{"rendered":"
<\/div>\n
\n
\n
\n
\n

\u041c\u043d\u043e\u0433\u0438\u0435 \u043b\u044e\u0434\u0438, \u043a\u043e\u0433\u0434\u0430 \u043b\u0438\u0431\u043e \u0438\u043c\u0435\u0432\u0448\u0438\u0435 \u0434\u0435\u043b\u043e \u0441 \u043d\u0435\u0439\u0440\u043e\u043d\u043d\u044b\u043c\u0438 \u0441\u0435\u0442\u044f\u043c\u0438, \u043d\u0430\u0432\u0435\u0440\u043d\u044f\u043a\u0430 \u0437\u0430\u0434\u0443\u043c\u044b\u0432\u0430\u043b\u0438\u0441\u044c, \u043c\u043e\u0436\u043d\u043e \u043b\u0438 \u043d\u0430\u043f\u0438\u0441\u0430\u0442\u044c \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u044c, \u043a\u043e\u0442\u043e\u0440\u0430\u044f \u0441\u0430\u043c\u0430 \u0431\u0443\u0434\u0435\u0442 \u0441\u043e\u0437\u0434\u0430\u0432\u0430\u0442\u044c \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0438 \u0434\u043b\u044f \u0440\u0435\u0448\u0435\u043d\u0438\u044f \u043a\u0430\u043a\u0438\u0445-\u043b\u0438\u0431\u043e \u0437\u0430\u0434\u0430\u0447. \u0422\u0430\u043a \u0432\u043e\u0442 \u0432 \u044d\u0442\u043e\u043c \u0446\u0438\u043a\u043b\u0435 \u0441\u0442\u0430\u0442\u0435\u0439 \u044f \u0440\u0435\u0448\u0438\u043b \u0440\u0435\u0430\u043b\u0438\u0437\u043e\u0432\u0430\u0442\u044c \u044d\u0442\u043e. \u041e\u0434\u043d\u0438\u043c \u0438\u0437 \u044d\u0442\u0430\u043f\u043e\u0432 \u0430\u043b\u0433\u043e\u0440\u0438\u0442\u043c\u0430 \u0431\u0443\u0434\u0435\u0442 \u0433\u0435\u043d\u0435\u0440\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u0435 \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0438 \u0438\u0437 \u0441\u043f\u0438\u0441\u043a\u0430 \u0441\u043b\u043e\u0451\u0432. \u0412 \u0441\u0432\u044f\u0437\u0438 \u0441 \u043d\u0435\u043a\u043e\u0442\u043e\u0440\u044b\u043c\u0438 \u043e\u0433\u0440\u0430\u043d\u0438\u0447\u0435\u043d\u0438\u044f\u043c\u0438, \u043d\u0430\u043a\u043b\u0430\u0434\u044b\u0432\u0430\u0435\u043c\u044b\u043c\u0438 \u043c\u0435\u0442\u043e\u0434\u0430\u043c\u0438 \u0440\u0435\u0430\u043b\u0438\u0437\u0430\u0446\u0438\u0438 (\u043e \u043a\u043e\u0442\u043e\u0440\u044b\u0445 \u0431\u0443\u0434\u0435\u0442 \u0441\u043a\u0430\u0437\u0430\u043d\u043e \u0432 \u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0438\u0445 \u0447\u0430\u0441\u0442\u044f\u0445, \u043a\u043e\u0433\u0434\u0430 \u043c\u044b \u043d\u0430\u0447\u043d\u0451\u043c \u043e\u0431\u044a\u0435\u0434\u0438\u043d\u044f\u0442\u044c \u043a\u043e\u0434 \u0438\u0437 \u044d\u0442\u043e\u0439 \u0441\u0442\u0430\u0442\u044c\u0438 \u0441 RL \u0295\u2299\u1d25\u2299\u0294 ), \u0432\u0445\u043e\u0434\u043d\u044b\u0435 \u0434\u0430\u043d\u043d\u044b\u0435 \u0434\u043b\u044f \u0433\u0435\u043d\u0435\u0440\u0430\u0442\u043e\u0440\u0430 \u0431\u0443\u0434\u0443\u0442 \u043f\u0440\u0435\u0434\u0441\u0442\u0430\u0432\u043b\u0435\u043d\u044b \u0432 \u0432\u0438\u0434\u0435 \u0441\u0442\u0440\u043e\u043a\u0438 \u0441\u043b\u0443\u0447\u0430\u0439\u043d\u043e\u0439 \u0434\u043b\u0438\u043d\u044b, \u0441\u043e\u0434\u0435\u0440\u0436\u0430\u0449\u0435\u0439 \u0443\u043f\u043e\u0440\u044f\u0434\u043e\u0447\u0435\u043d\u043d\u044b\u0439 \u043d\u0430\u0431\u043e\u0440 \u0441\u043b\u043e\u0451\u0432 \u0441 \u0438\u0445 \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u0430\u043c\u0438. \u0413\u0435\u043d\u0435\u0440\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u0441\u0435\u0442\u044c \u0431\u0443\u0434\u0435\u043c \u0434\u043b\u044f \u0437\u0430\u0434\u0430\u0447\u0438 \u043a\u043b\u0430\u0441\u0441\u0438\u0444\u0438\u043a\u0430\u0446\u0438\u0438 \u043a\u0430\u0440\u0442\u0438\u043d\u043e\u043a (\u0440\u0430\u0437\u043e\u0431\u044c\u0451\u043c \u044d\u0442\u043e \u043f\u0443\u0433\u0430\u043b\u043e \u043f\u0435\u0440\u0432\u044b\u043c).<\/p>\n

\u0424\u0443\u043d\u043a\u0446\u0438\u044f \u0441\u043e\u0437\u0434\u0430\u043d\u0438\u044f \u043e\u043f\u0438\u0441\u0430\u043d\u0438\u044f \u043f\u043e\u0441\u043b\u0435\u0434\u043e\u0432\u0430\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u0438 \u0441\u043b\u043e\u0451\u0432<\/h2>\n

\u041f\u0435\u0440\u0435\u0434 \u0433\u0435\u043d\u0435\u0440\u0430\u0446\u0438\u0435\u0439 \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0438 \u0441\u043e\u0437\u0434\u0430\u0434\u0438\u043c \u0441\u0442\u0440\u043e\u043a\u0443 \u0441\u043b\u0443\u0447\u0430\u0439\u043d\u043e \u0432\u044b\u0431\u0440\u0430\u043d\u043d\u044b\u0445 \u0441\u043b\u043e\u0451\u0432. \u041e\u0434\u043d\u0438\u043c \u0438\u0437 \u0431\u043e\u043b\u044c\u0448\u0438\u0445 \u043c\u0438\u043d\u0443\u0441\u043e\u0432, \u0441\u043e\u0437\u0434\u0430\u0432\u0430\u0435\u043c\u044b\u0445 \u043e\u0433\u0440\u0430\u043d\u0438\u0447\u0435\u043d\u0438\u044f\u043c\u0438, \u044f\u0432\u043b\u044f\u0435\u0442\u0441\u044f \u0442\u043e, \u0447\u0442\u043e \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u044b \u0441\u043b\u043e\u0451\u0432 \u043d\u0435\u043e\u0431\u0445\u043e\u0434\u0438\u043c\u043e \u0433\u0435\u043d\u0435\u0440\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u0432\u043c\u0435\u0441\u0442\u0435 \u0441\u043e \u0441\u043b\u043e\u044f\u043c\u0438, \u0442.\u0435. \u0433\u0435\u043d\u0435\u0440\u0438\u0440\u0443\u0435\u0442\u0441\u044f \u0441\u043b\u043e\u0439 \u0441\u0440\u0430\u0437\u0443 \u0441 \u0437\u0430\u0434\u0430\u043d\u043d\u044b\u043c\u0438 \u0444\u0438\u043a\u0441\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u044b\u043c\u0438 \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u0430\u043c\u0438. \u0420\u0430\u0441\u0448\u0438\u0440\u0435\u043d\u0438\u0435 \u043c\u043d\u043e\u0436\u0435\u0441\u0442\u0432\u0430 \u0432\u043e\u0437\u043c\u043e\u0436\u043d\u043e \u043d\u0430 \u0443\u0441\u043c\u043e\u0442\u0440\u0435\u043d\u0438\u0435 \u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u0435\u043b\u044f.<\/p>\n

layers_set = {                 # increase channels with kernel 3                 'conv=channel_factor:2,kernel_size:3,stride:1,padding:0-',                 'conv=channel_factor:3,kernel_size:3,stride:1,padding:0-',                 'conv=channel_factor:4,kernel_size:3,stride:1,padding:0-',                  # decrease channels with kernel 3                 'conv=channel_factor:0.4,kernel_size:3,stride:1,padding:0-',                 'conv=channel_factor:0.6,kernel_size:3,stride:1,padding:0-',                 'conv=channel_factor:0.8,kernel_size:3,stride:1,padding:0-',                    'batchnorm=eps:0.00001-',                  'batchnorm=eps:0.0001-',                   'avgpool=kernel_size:2,stride:2,padding:0-',                  'avgpool=kernel_size:4,stride:4,padding:0-',                   'maxpool=kernel_size:2,stride:2,padding:0-',                  'maxpool=kernel_size:4,stride:4,padding:0-',                   'dropout=p:0.2-',                  'dropout=p:0.4-',   }<\/code><\/pre>\n
channel_factor — \u043c\u043d\u043e\u0436\u0438\u0442\u0435\u043b\u044c \u043a\u043e\u043b-\u0432\u0430 \u043a\u0430\u043d\u0430\u043b\u043e\u0432.
«-» —  \u0440\u0430\u0437\u0434\u0435\u043b\u0438\u0442\u0435\u043b\u044c \u0441\u043b\u043e\u0451\u0432
«,» — \u0440\u0430\u0437\u0434\u0435\u043b\u0438\u0442\u0435\u043b\u044c \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u043e\u0432 \u0441\u043b\u043e\u044f
«:» — \u0440\u0430\u0437\u0434\u0435\u043b\u0438\u0442\u0435\u043b\u044c \u043d\u0430\u0437\u0432\u0430\u043d\u0438\u044f \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u0430 \u0438 \u0435\u0433\u043e \u0437\u043d\u0430\u0447\u0435\u043d\u0438\u044f
«=» — \u0440\u0430\u0437\u0434\u0435\u043b\u0438\u0442\u0435\u043b\u044c \u043d\u0430\u0437\u0432\u0430\u043d\u0438\u044f \u0441\u043b\u043e\u044f \u043e\u0442 \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u043e\u0432<\/p>\n
\u0420\u0435\u0430\u043b\u0438\u0437\u0443\u0435\u043c \u0444\u0443\u043d\u043a\u0446\u0438\u044e, \u0441\u043e\u0437\u0434\u0430\u044e\u0449\u0443\u044e \u0441\u0442\u0440\u043e\u043a\u0443 \u0441\u043b\u043e\u0451\u0432 \u0441 \u0434\u043b\u0438\u043d\u043e\u0439 min_len <= L <= max_len \u0441\u043b\u043e\u0451\u0432.<\/p>\n
def create_layers_string(min_len=2, max_len=10):     if min_len < 1 or max_len < min_len:       print('Parameters are incorrect!')       return None     length = random.randint(min_len, max_len)     text = random.sample(layers_set, length)     text = ''.join((layer for layer in text))     return text<\/code><\/pre>\n
\u041a\u043b\u0430\u0441\u0441 \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0438<\/h2>\n
\u0412 \u043a\u0430\u0447\u0435\u0441\u0442\u0432\u0435 \u043a\u043b\u0430\u0441\u0441\u0430 \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0438 \u0432\u043e\u0437\u044c\u043c\u0451\u043c \u043f\u0440\u043e\u0441\u0442\u043e\u0439 \u043f\u0443\u0441\u0442\u043e\u0439 \u0448\u0430\u0431\u043b\u043e\u043d, \u043f\u043e\u043b\u0435 «layers» \u043a\u043e\u0442\u043e\u0440\u043e\u0433\u043e \u0431\u0443\u0434\u0435\u0442 \u0441\u043e\u0434\u0435\u0440\u0436\u0430\u0442\u044c \u043d\u0430\u0448\u0443 \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0435\u0432\u0443\u044e \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0443, \u043a\u043e\u0433\u0434\u0430 \u043c\u044b \u0435\u0451 \u0441\u0433\u0435\u043d\u0435\u0440\u0438\u0440\u0443\u0435\u043c.<\/p>\n
class NN(nn.Module):      def __init__(self, in_channels):       super().__init__()        self.layers = nn.Sequential()      def forward(self, x):       x = self.layers(x)       return x      def __call__(self, x):       return self.forward(x)<\/code><\/pre>\n
\u0417\u0430\u043c\u0435\u0447\u0443, \u0447\u0442\u043e \u043a\u043b\u0430\u0441\u0441\u0438\u0444\u0438\u043a\u0430\u0442\u043e\u0440\u0430 \u0432 \u0432\u0438\u0434\u0435 \u043b\u0438\u043d\u0435\u0439\u043d\u043e\u0433\u043e \u0441\u043b\u043e\u044f \u0442\u0443\u0442 \u043d\u0435\u0442. \u041e\u043d \u0431\u0443\u0434\u0435\u0442 \u0434\u043e\u0431\u0430\u0432\u043b\u0435\u043d \u0432 \u043a\u043e\u043d\u0435\u0446 layers \u043f\u043e\u0441\u043b\u0435 \u0432\u044b\u0447\u0438\u0441\u043b\u0435\u043d\u0438\u044f \u0432\u0445\u043e\u0434\u043d\u044b\u0445 \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u043e\u0432 \u0434\u043b\u044f \u043d\u0435\u0433\u043e.<\/p>\n
\u041a\u043b\u0430\u0441\u0441 nnGenerator <\/h2>\n
\u0421\u0430\u043c\u043e\u0439 \u0431\u043e\u043b\u044c\u0448\u043e\u0439 \u0447\u0430\u0441\u0442\u044c\u044e \u043a\u043e\u0434\u0430 \u0431\u0443\u0434\u0435\u0442 \u043a\u043b\u0430\u0441\u0441 nnGenerator, \u0441\u043e\u0434\u0435\u0440\u0436\u0430\u0449\u0438\u0439 \u0432 \u0441\u0435\u0431\u0435 \u043f\u043e\u043c\u0438\u043c\u043e \u043c\u0435\u0442\u043e\u0434\u0430 \u0433\u0435\u043d\u0435\u0440\u0430\u0446\u0438\u0438 \u043c\u0435\u0442\u043e\u0434 \u043f\u0430\u0440\u0441\u0438\u043d\u0433\u0430 \u043d\u0430\u0448\u0435\u0433\u043e \u0432\u0445\u043e\u0434\u043d\u043e\u0433\u043e \u0442\u0435\u043a\u0441\u0442\u0430, \u0441\u043e\u0434\u0435\u0440\u0436\u0430\u0449\u0435\u0433\u043e \u0441\u043b\u043e\u0438. 
\u041f\u0430\u0440\u0441\u0438\u043d\u0433 \u0441\u0442\u0440\u043e\u043a\u0438 \u0431\u0443\u0434\u0435\u0442 \u043f\u0440\u043e\u0438\u0437\u0432\u0435\u0434\u0451\u043d \u043f\u043e \u043e\u0431\u043e\u0437\u043d\u0430\u0447\u0435\u043d\u043d\u044b\u043c \u0432\u044b\u0448\u0435 \u0440\u0430\u0437\u0434\u0435\u043b\u0438\u0442\u0435\u043b\u044f\u043c.<\/p>\n
class nnGenerator():     def __init__(self):       self.text_layers_dict = dict({})       self.nn_len = -1<\/code><\/pre>\n
\u041a\u043b\u0430\u0441\u0441 \u0441\u043e\u0434\u0435\u0440\u0436\u0438\u0442 \u0441\u043b\u043e\u0432\u0430\u0440\u044c, \u0445\u0440\u0430\u043d\u044f\u0449\u0438\u0439 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442 \u043f\u0430\u0440\u0441\u0438\u043d\u0433\u0430 \u0441\u0442\u0440\u043e\u043a\u0438 \u0441\u043b\u043e\u0451\u0432.
\u0421\u0442\u0440\u0443\u043a\u0442\u0443\u0440\u0430 \u0441\u043b\u043e\u0432\u0430\u0440\u044f \u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0430\u044f: 1 \u044d\u043b\u0435\u043c\u0435\u043d\u0442 \u0435\u0433\u043e \u0441\u043e\u0441\u0442\u043e\u0438\u0442 \u0438\u0437 \u043d\u0430\u0437\u0432\u0430\u043d\u0438\u0435 \u0441\u043b\u043e\u044f, \u0441\u043a\u043e\u043d\u043a\u0430\u0442\u0435\u043d\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u043e\u0435 \u0441 \u0435\u0433\u043e id + \u043f\u0430\u0440\u0430: id + \u0441\u043b\u043e\u0432\u0430\u0440\u044c \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u043e\u0432. \u041f\u0440\u0438\u043c\u0435\u0440 \u043e\u0434\u043d\u043e\u0433\u043e \u044d\u043b\u0435\u043c\u0435\u043d\u0442\u0430 \u0441\u043b\u043e\u0432\u0430\u0440\u044f text_layers_dict:
«dropout1»: (1, {«p»: 0.2} )<\/p>\n
    def parseTextNet(self, text_net):       self.text_layers_dict = dict({})       self.nn_len = -1       print(text_net)       if text_net[-1] == '-':         text_net = text_net[:-1]       text_layers = text_net.split('-')       id = 0       for text_layer in text_layers:         tmp = text_layer.split('=')         layer_name, layer_params = tmp[0], tmp[1].split(',')         layer_params_dict = dict({})         for param in layer_params:           param = param.split(':')           param_name, param_value = param[0], param[1]           layer_params_dict[param_name] = param_value         self.text_layers_dict[layer_name + str(id)] = (id, layer_params_dict)         id += 1       print(self.text_layers_dict) <\/code><\/pre>\n
\u041f\u0430\u0440\u0441\u0435\u0440 \u043f\u0440\u0438\u043d\u0438\u043c\u0430\u0435\u0442 \u043d\u0430 \u0432\u0445\u043e\u0434 \u0441\u0442\u0440\u043e\u043a\u0443 \u0441\u043b\u043e\u0451\u0432 \u0438 \u043f\u0440\u0435\u043e\u0431\u0440\u0430\u0437\u0443\u0435\u0442 \u0435\u0451, \u0437\u0430\u043f\u0438\u0441\u044b\u0432\u0430\u044f \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442 \u0432 \u0441\u043b\u043e\u0432\u0430\u0440\u044c text_layers_dict. <\/p>\n
    def conv_output_shape(self, h, w, kernel_size=1, stride=1, pad=0, dilation=1):       h = math.floor( ((h + (2 * pad) - ( dilation * (kernel_size - 1) ) - 1 )\/ stride) + 1)       w = math.floor( ((w + (2 * pad) - ( dilation * (kernel_size - 1) ) - 1 )\/ stride) + 1)       return h, w  <\/code><\/pre>\n
\u041c\u0435\u0442\u043e\u0434 conv_output_shape \u0440\u0435\u0430\u043b\u0438\u0437\u0443\u0435\u0442 \u0432\u044b\u0447\u0438\u0441\u043b\u0435\u043d\u0438\u0435 \u0432\u044b\u0445\u043e\u0434\u0430 \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0435\u0432\u043e\u0433\u043e \u0441\u043b\u043e\u044f. 
\u0414\u043b\u044f \u0441\u0432\u0451\u0440\u0442\u043e\u043a \u0438 \u043f\u0443\u043b\u043b\u0438\u043d\u0433\u043e\u0432 \u044d\u0442\u0430 \u0444\u0443\u043d\u043a\u0446\u0438\u044f \u0443\u043d\u0438\u0432\u0435\u0440\u0441\u0430\u043b\u044c\u043d\u0430, dropout \u0438 batchnorm \u043d\u0435 \u043c\u0435\u043d\u044f\u044e\u0442 \u0440\u0430\u0437\u043c\u0435\u0440\u043d\u043e\u0441\u0442\u0435\u0439, \u0430 \u0437\u043d\u0430\u0447\u0438\u0442 \u0432 \u0432\u044b\u0447\u0438\u0441\u043b\u0435\u043d\u0438\u0438 \u0432\u044b\u0445\u043e\u0434\u0430 \u0441\u043b\u043e\u044f \u043d\u0435 \u043d\u0443\u0436\u0434\u0430\u044e\u0442\u0441\u044f. \u0424\u043e\u0440\u043c\u0443\u043b\u0430 \u0432\u044b\u0447\u0438\u0441\u043b\u0435\u043d\u0438\u044f \u044d\u0442\u0438\u0445 \u0437\u043d\u0430\u0447\u0435\u043d\u0438\u0439 \u043f\u0440\u0435\u0434\u0441\u0442\u0430\u0432\u043b\u0435\u043d\u0430 \u0432 \u0434\u043e\u043a\u0443\u043c\u0435\u043d\u0442\u0430\u0446\u0438\u0438 \u0431\u0438\u0431\u043b\u0438\u043e\u0442\u0435\u043a\u0438 PyTorch. <\/a><\/p>\n
\"\u043a\u0432\u0430\u0434\u0440\u0430\u0442\u043d\u044b\u0435<\/p>\n
\u043a\u0432\u0430\u0434\u0440\u0430\u0442\u043d\u044b\u0435 \u0441\u043a\u043e\u0431\u043a\u0438 — \u043e\u043a\u0440\u0443\u0433\u043b\u0435\u043d\u0438\u0435 \u0432\u043d\u0438\u0437 (math.floor)<\/figcaption><\/div>\n<\/figure>\n
\u0414\u0430\u043b\u0435\u0435 \u043f\u043e\u0439\u0434\u0451\u0442 \u0441\u0430\u043c\u043e\u0435 \u0438\u043d\u0442\u0435\u0440\u0435\u0441\u043d\u043e\u0435 — \u0430\u0432\u0442\u043e\u043f\u043e\u0441\u0442\u0440\u043e\u0435\u043d\u0438\u0435 \u043d\u0430\u0448\u0435\u0439 \u0441\u0435\u0442\u0438 \u043f\u043e \u0441\u043b\u043e\u0432\u0430\u0440\u044e \u0441\u043b\u043e\u0451\u0432.<\/p>\n
    def generateNN(self, n_classes, test_batch):       success_state = False        backbone = nn.Sequential()       classifier = nn.Sequential()       optimizer = None       try:         data_shape = np.array(test_batch).shape # [B, C, H, W]         last_shape = data_shape         for layer_name in self.text_layers_dict.keys():             layer = None             layer_params= self.text_layers_dict[layer_name][1]             if layer_name.find('conv') >= 0:               kernel_size = int(layer_params['kernel_size'])               channel_factor = float(layer_params['channel_factor'])               stride = int(layer_params['stride'])               padding = int(layer_params['padding'])               dilation = 1                activation = nn.ReLU(inplace=True)                in_chan = last_shape[1]               assert(in_chan <= last_shape[2] and in_chan <= last_shape[3])               out_chan = math.floor(in_chan * channel_factor)               assert(out_chan > 0)                backbone.append(nn.Conv2d(in_chan, out_chan, kernel_size, stride, padding, dilation))               backbone.append(activation)                h, w = self.conv_output_shape(last_shape[2], last_shape[3], kernel_size, stride, padding, dilation)               last_shape = (last_shape[0], out_chan, h, w)              elif layer_name.find('batchnorm') >= 0:               eps = float(layer_params['eps'])               in_chan = last_shape[1]               backbone.append(nn.BatchNorm2d(in_chan, eps))              elif layer_name.find('avgpool') >= 0:               kernel_size = int(layer_params['kernel_size'])               stride = int(layer_params['stride'])               padding = int(layer_params['padding'])               dilation = 1               out_chan = last_shape[1]               backbone.append(nn.AvgPool2d(kernel_size, stride, padding))                h, w = self.conv_output_shape(last_shape[2], last_shape[3], kernel_size, stride, padding, dilation)               last_shape = (last_shape[0], out_chan, h, w)               elif layer_name.find('maxpool') >= 0:               kernel_size = int(layer_params['kernel_size'])               stride = int(layer_params['stride'])               padding = int(layer_params['padding'])               dilation = 1                backbone.append(nn.MaxPool2d(kernel_size, stride, padding))                h, w = self.conv_output_shape(last_shape[2], last_shape[3], kernel_size, stride, padding, dilation)               last_shape = (last_shape[0], last_shape[1], h, w)              elif layer_name.find('dropout') >= 0:               p = float(layer_params['p'])               backbone.append(nn.Dropout2d(p))          linear_in_shape = last_shape[1] * last_shape[2] * last_shape[3]         classifier = nn.Linear(linear_in_shape, n_classes)         success_state = True         print('NN build successfull!')                except Exception as e:         print('NN build failed!')         print(str(e))        net = nn.Sequential()       net.append(backbone)       net.append(nn.Flatten(start_dim=1))       net.append(classifier)       self.text_layers_dict = dict({})       return success_state, net<\/code><\/pre>\n
\u041f\u0441\u0435\u0432\u0434\u043e\u043a\u043e\u0434\u043e\u043c \u043c\u043e\u0436\u043d\u043e \u043e\u043f\u0438\u0441\u0430\u0442\u044c \u0430\u043b\u0433\u043e\u0440\u0438\u0442\u043c \u0442\u0430\u043a:
\u041f\u0440\u043e\u0445\u043e\u0434\u0438\u043c \u043f\u043e \u0432\u0441\u0435\u043c \u0441\u043b\u043e\u044f\u043c \u0438\u0437 \u0441\u043b\u043e\u0432\u0430\u0440\u044f:
     1) \u0411\u0435\u0440\u0451\u043c \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u044b \u0441\u043b\u043e\u044f
     2) \u0412\u044b\u0447\u0438\u0441\u043b\u044f\u0435\u043c \u0432\u044b\u0445\u043e\u0434\u043d\u043e\u0439 shape
     3) \u0421\u043e\u0437\u0434\u0430\u0451\u043c \u0441\u043b\u043e\u0439, \u0434\u043e\u0431\u0430\u0432\u043b\u044f\u0435\u043c \u0432 \u043a\u043e\u043d\u0435\u0446 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u044b
     4) \u0421\u043e\u0445\u0440\u0430\u043d\u044f\u0435\u043c \u0437\u043d\u0430\u0447\u0435\u043d\u0438\u0435 \u0432\u044b\u0445\u043e\u0434\u043d\u043e\u0433\u043e \u0440\u0430\u0437\u043c\u0435\u0440\u0430 \u043f\u043e\u0441\u043b\u0435\u0434\u043d\u0435\u0433\u043e \u0441\u043b\u043e\u044f
\u0412 \u043a\u043e\u043d\u0446\u0435 \u0434\u043e\u0431\u0430\u0432\u043b\u044f\u0435\u043c \u043b\u0438\u043d\u0435\u0439\u043d\u044b\u0439 \u043a\u043b\u0430\u0441\u0441\u0438\u0444\u0438\u043a\u0430\u0442\u043e\u0440. 
\u0414\u043b\u044f \u0432\u044b\u0447\u0438\u0441\u043b\u0435\u043d\u0438\u044f \u0432\u0445\u043e\u0434\u043d\u043e\u0433\u043e \u0440\u0430\u0437\u043c\u0435\u0440\u0430 \u0434\u0430\u043d\u043d\u044b\u0445 \u043f\u043e\u0434\u0430\u0451\u0442\u0441\u044f \u0442\u0435\u0441\u0442\u043e\u0432\u044b\u0439 \u0431\u0430\u0442\u0447, \u043d\u0430 \u0432\u0430\u0448\u0435 \u0443\u0441\u043c\u043e\u0442\u0440\u0435\u043d\u0438\u0435 \u043c\u043e\u0436\u043d\u043e \u043f\u0435\u0440\u0435\u0434\u0435\u043b\u0430\u0442\u044c \u0432 \u0431\u043e\u043b\u0435\u0435 \u0443\u0434\u043e\u0431\u043d\u044b\u0439 \u0432\u0438\u0434.<\/p>\n
\u041e\u0441\u043e\u0431\u044b\u0435 \u043c\u043e\u043c\u0435\u043d\u0442\u044b, \u043a\u043e\u0442\u043e\u0440\u044b\u0435 \u043d\u0435\u043e\u0431\u0445\u043e\u0434\u0438\u043c\u043e \u043e\u0442\u043b\u0430\u0432\u043b\u0438\u0432\u0430\u0442\u044c, \u044d\u0442\u043e \u0441\u043b\u0443\u0447\u0430\u0439, \u043f\u0440\u0438 \u043a\u043e\u0442\u043e\u0440\u043e\u043c \u0432 \u0445\u043e\u0434\u0435 \u043f\u043e\u0441\u0442\u0440\u043e\u0435\u043d\u0438\u044f \u0441\u0435\u0442\u0438 \u044f\u0434\u0440\u043e \u0441\u0442\u0430\u043d\u043e\u0432\u0438\u0442\u0441\u044f \u0431\u043e\u043b\u044c\u0448\u0435 \u0440\u0430\u0437\u043c\u0435\u0440\u0430 \u0434\u0430\u043d\u043d\u044b\u0445, \u0438 \u043f\u0440\u0438 \u043a\u043e\u0442\u043e\u0440\u043e\u043c \u043a\u043e\u043b-\u0432\u043e \u043a\u0430\u043d\u0430\u043b\u043e\u0432 \u0443\u043c\u0435\u043d\u044c\u0448\u0430\u0435\u0442\u0441\u044f \u0434\u043e \u0437\u043d\u0430\u0447\u0435\u043d\u0438\u044f < 1. <\/p>\n
\u0414\u0430\u043b\u0435\u0435 \u043a\u043e\u0434 \u0441\u043e\u0437\u0434\u0430\u043d\u0438\u044f dataloader’\u043e\u0432 \u043f\u043e\u0434 MNIST, \u0430 \u0442\u0430\u043a\u0436\u0435 \u043e\u0431\u0443\u0447\u0430\u044e\u0449\u0435\u0433\u043e \u0446\u0438\u043a\u043b\u0430, \u0434\u0443\u043c\u0430\u044e, \u0432 \u043a\u043e\u043c\u043c\u0435\u043d\u0442\u0430\u0440\u0438\u044f\u0445 \u044d\u0442\u043e \u043d\u0435 \u043d\u0443\u0436\u0434\u0430\u0435\u0442\u0441\u044f, \u043e\u0442\u043c\u0435\u0447\u0443 \u0442\u043e\u043b\u044c\u043a\u043e, \u0447\u0442\u043e \u0432\u0445\u043e\u0434\u043d\u044b\u0435 \u0434\u0430\u043d\u043d\u044b\u0435 \u0434\u043e\u043b\u0436\u043d\u044b \u0431\u044b\u0442\u044c (B, C, H, W) = (Batch_Size, Channels, height, width), \u0437\u043d\u0430\u0447\u0438\u0442 \u0434\u0430\u043d\u043d\u044b\u0435 \u0438\u0437 \u041c\u041d\u0418\u0421\u0422\u0430 \u0441 shape’\u043e\u043c (N, 28,28) \u043d\u0443\u0436\u043d\u043e \u043f\u0440\u0435\u0434\u0441\u0442\u0430\u0432\u0438\u0442\u044c \u0432 \u0432\u0438\u0434\u0435 (N, 1, 28, 28)<\/p>\n
 class ds(Dataset):     def __init__(self, X, y):        self.X = X       self.y = y      def __len__(self):         return len(self.y)      def __getitem__(self, idx):         x_ = self.X[idx]         y_ = self.y[idx]         return x_, y_  train_data = dsets.MNIST(root = '.\/data', train = True,                         transform = transforms.ToTensor(), download = True)  test_data = dsets.MNIST(root = '.\/data', train = False,                        transform = transforms.ToTensor())   train_samples = np.expand_dims(np.array(train_data.data), axis=1)[:5000] train_labels = np.array(train_data.targets)[:5000]   dataset = ds(X=train_samples, y=train_labels)  train_set, valid_set = random_split(dataset, [0.8, 0.2], generator=torch.Generator().manual_seed(42))  train_dataloader = torch.utils.data.DataLoader(   train_set,   batch_size=bs,   shuffle=True,   drop_last=True)  valid_dataloader = torch.utils.data.DataLoader(   valid_set,   batch_size=bs,   drop_last=True,   shuffle=True)  test_batch = None for b, _ in train_dataloader:   test_batch = b   break <\/code><\/pre>\n\n
\u0422\u0435\u0441\u0442\u044b<\/h2>\n
\u0413\u0435\u043d\u0435\u0440\u0438\u0440\u0443\u0435\u043c \u043f\u043e\u0441\u043b\u0435\u0434\u043e\u0432\u0430\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c \u0441\u043b\u043e\u0451\u0432<\/p>\n
test_batch = None for b, _ in train_dataloader:   test_batch = b   break  generator = nnGenerator()  success_state = False while not success_state:   text_layers = create_layers_string()   generator.parseTextNet(text_layers)   success_state, seq = generator.generateNN(n_classes=10, test_batch=test_batch) <\/code><\/pre>\n
\u0420\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u043e\u043c \u0431\u0443\u0434\u0435\u0442 \u0442\u0435\u043a\u0441\u0442 \u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0435\u0433\u043e \u0441\u043e\u0434\u0435\u0440\u0436\u0430\u043d\u0438\u044f:<\/p>\n
conv=channel_factor:0.4,kernel_size:3,stride:1,padding:0-batchnorm=eps:0.00001-conv=channel_factor:4,kernel_size:3,stride:1,padding:0-conv=channel_factor:0.8,kernel_size:3,stride:1,padding:0-maxpool=kernel_size:4,stride:4,padding:0- {'conv0': (0, {'channel_factor': '0.4', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'batchnorm1': (1, {'eps': '0.00001'}), 'conv2': (2, {'channel_factor': '4', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'conv3': (3, {'channel_factor': '0.8', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'maxpool4': (4, {'kernel_size': '4', 'stride': '4', 'padding': '0'})} NN build failed!  batchnorm=eps:0.00001-conv=channel_factor:3,kernel_size:3,stride:1,padding:0-avgpool=kernel_size:4,stride:4,padding:0-conv=channel_factor:0.6,kernel_size:3,stride:1,padding:0-maxpool=kernel_size:4,stride:4,padding:0-dropout=p:0.2-conv=channel_factor:0.8,kernel_size:3,stride:1,padding:0-dropout=p:0.4- {'batchnorm0': (0, {'eps': '0.00001'}), 'conv1': (1, {'channel_factor': '3', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'avgpool2': (2, {'kernel_size': '4', 'stride': '4', 'padding': '0'}), 'conv3': (3, {'channel_factor': '0.6', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'maxpool4': (4, {'kernel_size': '4', 'stride': '4', 'padding': '0'}), 'dropout5': (5, {'p': '0.2'}), 'conv6': (6, {'channel_factor': '0.8', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'dropout7': (7, {'p': '0.4'})} NN build failed!  avgpool=kernel_size:4,stride:4,padding:0-conv=channel_factor:0.6,kernel_size:3,stride:1,padding:0-dropout=p:0.4-conv=channel_factor:0.8,kernel_size:3,stride:1,padding:0- {'avgpool0': (0, {'kernel_size': '4', 'stride': '4', 'padding': '0'}), 'conv1': (1, {'channel_factor': '0.6', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'dropout2': (2, {'p': '0.4'}), 'conv3': (3, {'channel_factor': '0.8', 'kernel_size': '3', 'stride': '1', 'padding': '0'})} NN build failed!  conv=channel_factor:0.6,kernel_size:3,stride:1,padding:0-avgpool=kernel_size:2,stride:2,padding:0-conv=channel_factor:0.8,kernel_size:3,stride:1,padding:0-conv=channel_factor:0.4,kernel_size:3,stride:1,padding:0-conv=channel_factor:4,kernel_size:3,stride:1,padding:0-conv=channel_factor:3,kernel_size:3,stride:1,padding:0-maxpool=kernel_size:4,stride:4,padding:0-batchnorm=eps:0.00001-maxpool=kernel_size:2,stride:2,padding:0-batchnorm=eps:0.0001- {'conv0': (0, {'channel_factor': '0.6', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'avgpool1': (1, {'kernel_size': '2', 'stride': '2', 'padding': '0'}), 'conv2': (2, {'channel_factor': '0.8', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'conv3': (3, {'channel_factor': '0.4', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'conv4': (4, {'channel_factor': '4', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'conv5': (5, {'channel_factor': '3', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'maxpool6': (6, {'kernel_size': '4', 'stride': '4', 'padding': '0'}), 'batchnorm7': (7, {'eps': '0.00001'}), 'maxpool8': (8, {'kernel_size': '2', 'stride': '2', 'padding': '0'}), 'batchnorm9': (9, {'eps': '0.0001'})} NN build failed!  conv=channel_factor:0.6,kernel_size:3,stride:1,padding:0-dropout=p:0.2-batchnorm=eps:0.00001-maxpool=kernel_size:4,stride:4,padding:0-avgpool=kernel_size:2,stride:2,padding:0-conv=channel_factor:0.4,kernel_size:3,stride:1,padding:0-conv=channel_factor:4,kernel_size:3,stride:1,padding:0-avgpool=kernel_size:4,stride:4,padding:0-conv=channel_factor:0.8,kernel_size:3,stride:1,padding:0-batchnorm=eps:0.0001- {'conv0': (0, {'channel_factor': '0.6', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'dropout1': (1, {'p': '0.2'}), 'batchnorm2': (2, {'eps': '0.00001'}), 'maxpool3': (3, {'kernel_size': '4', 'stride': '4', 'padding': '0'}), 'avgpool4': (4, {'kernel_size': '2', 'stride': '2', 'padding': '0'}), 'conv5': (5, {'channel_factor': '0.4', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'conv6': (6, {'channel_factor': '4', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'avgpool7': (7, {'kernel_size': '4', 'stride': '4', 'padding': '0'}), 'conv8': (8, {'channel_factor': '0.8', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'batchnorm9': (9, {'eps': '0.0001'})} NN build failed!  batchnorm=eps:0.00001-dropout=p:0.4-batchnorm=eps:0.0001-conv=channel_factor:0.8,kernel_size:3,stride:1,padding:0-maxpool=kernel_size:4,stride:4,padding:0-conv=channel_factor:3,kernel_size:3,stride:1,padding:0- {'batchnorm0': (0, {'eps': '0.00001'}), 'dropout1': (1, {'p': '0.4'}), 'batchnorm2': (2, {'eps': '0.0001'}), 'conv3': (3, {'channel_factor': '0.8', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'maxpool4': (4, {'kernel_size': '4', 'stride': '4', 'padding': '0'}), 'conv5': (5, {'channel_factor': '3', 'kernel_size': '3', 'stride': '1', 'padding': '0'})} NN build failed!  dropout=p:0.2-conv=channel_factor:3,kernel_size:3,stride:1,padding:0-dropout=p:0.4- {'dropout0': (0, {'p': '0.2'}), 'conv1': (1, {'channel_factor': '3', 'kernel_size': '3', 'stride': '1', 'padding': '0'}), 'dropout2': (2, {'p': '0.4'})} NN build successfull! <ipython-input-20-4e57f99010fc>:6: DeprecationWarning: Sampling from a set deprecated since Python 3.9 and will be removed in a subsequent version.   text = random.sample(layers_set, length)<\/code><\/pre>\n
\u0421 \u0441\u0435\u0434\u044c\u043c\u043e\u0439 \u043f\u043e\u043f\u044b\u0442\u043a\u0438 \u043f\u043e\u043b\u0443\u0447\u0438\u043b\u043e\u0441\u044c (\uff1b\u4e00_\u4e00)  <\/p>\n
\u0421\u043e\u0437\u0434\u0430\u0451\u043c \u043e\u0431\u044a\u0435\u043a\u0442 \u043f\u0443\u0441\u0442\u043e\u0439 \u0441\u0435\u0442\u0438, \u0433\u0435\u043d\u0435\u0440\u0438\u0440\u0443\u0435\u043c \u0441\u0435\u0442\u044c<\/p>\n
model = NN(in_channels=1) model.layers = seq display(model)<\/code><\/pre>\n
\u041f\u043e\u043b\u0443\u0447\u0430\u0435\u043c \u043a\u0430\u043a\u0443\u044e-\u0442\u043e \u043d\u0430\u0448\u0443 \u0441\u043b\u0443\u0447\u0430\u0439\u043d\u0443\u044e \u0441\u0435\u0442\u044c<\/p>\n
NN(   (layers): Sequential(     (0): Sequential(       (0): Dropout2d(p=0.2, inplace=False)       (1): Conv2d(1, 3, kernel_size=(3, 3), stride=(1, 1))       (2): ReLU(inplace=True)       (3): Dropout2d(p=0.4, inplace=False)     )     (1): Flatten(start_dim=1, end_dim=-1)     (2): Linear(in_features=2028, out_features=10, bias=True)   ) )<\/code><\/pre>\n
\u0414\u0430\u043b\u0435\u0435 \u043e\u043f\u0442\u0438\u043c\u0430\u0439\u0437\u0435\u0440 + \u0444-\u044f \u043e\u0448\u0438\u0431\u043a\u0438<\/p>\n
criterion = torch.nn.CrossEntropyLoss() optimizer = torch.optim.SGD(model.parameters(), lr=lr)<\/code><\/pre>\n
\u0422\u0440\u0435\u043d\u0438\u0440\u043e\u0432\u043e\u0447\u043d\u044b\u0439 \u0446\u0438\u043a\u043b \u0441\u0430\u043c\u044b\u0439 \u043f\u0440\u043e\u0441\u0442\u043e\u0439<\/p>\n
def train():     train_losses = []     valid_losses = []     # TODO calculate metrics and return them after train     def CalcValLoss():         with torch.no_grad():             losses = []             for X, Y in valid_dataloader:                 X = X.float().to(device)                 Y = Y.float().to(device)                 preds = model(X)                 preds, _ = torch.max(preds,1)                 loss = criterion(preds,Y)                 losses.append(loss.item())             print(\"Valid Loss : {:.6f}\".format(torch.tensor(losses).mean()))             valid_losses.append(torch.tensor(losses).mean())      for i in range(1, epochs):         losses = []         for X, Y in tqdm(train_dataloader):             X = X.float().to(device)             Y = Y.float().to(device)             preds = model(X)             preds, _ = torch.max(preds,1)             loss = criterion(preds, Y)             losses.append(loss.item())              optimizer.zero_grad()             loss.backward()             optimizer.step()          print(\"Train Loss : {:.6f}\".format(torch.tensor(losses).mean()))         train_losses.append(torch.tensor(losses).mean())     return train_losses, valid_losses<\/code><\/pre>\n
\u0417\u0430\u043f\u0443\u0441\u043a\u0430\u0435\u043c, \u043e\u0431\u0443\u0447\u0430\u0435\u043c, \u0444\u0438\u043a\u0441\u0438\u0440\u0443\u0435\u043c \u043f\u0440\u0438\u0431\u044b\u043b\u044c<\/p>\n
train_losses, valid_losses = train()  100%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 125\/125 [00:00<00:00, 213.39it\/s] Train Loss : 8579.939453 100%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 125\/125 [00:00<00:00, 219.21it\/s] Train Loss : 493.860870 100%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 125\/125 [00:00<00:00, 224.44it\/s] Train Loss : 493.228210 100%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 125\/125 [00:00<00:00, 347.04it\/s] Train Loss : 493.066498 100%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 125\/125 [00:00<00:00, 359.57it\/s] Train Loss : 492.938690 100%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 125\/125 [00:00<00:00, 340.28it\/s] Train Loss : 492.947876 100%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 125\/125 [00:00<00:00, 339.37it\/s] Train Loss : 492.833740 100%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 125\/125 [00:00<00:00, 335.57it\/s] Train Loss : 492.833679 100%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 125\/125 [00:00<00:00, 350.72it\/s] Train Loss : 492.908295<\/code><\/pre>\n
\u0418\u0442\u043e\u0433<\/h2>\n
\u041c\u044b \u043f\u043e\u043b\u0443\u0447\u0438\u043b\u0438 \u0433\u0435\u043d\u0435\u0440\u0430\u0442\u043e\u0440 \u0441\u0432\u0451\u0440\u0442\u043e\u0447\u043d\u044b\u0445 \u043d\u0435\u0439\u0440\u043e\u043d\u043d\u044b\u0445 \u0441\u0435\u0442\u0435\u0439 \u0441\u043b\u0443\u0447\u0430\u0439\u043d\u043e\u0439 \u0434\u043b\u0438\u043d\u044b \u0434\u043b\u044f \u043a\u043b\u0430\u0441\u0441\u0438\u0444\u0438\u043a\u0430\u0446\u0438\u0438 \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0439.<\/p>\n
\u041f\u043b\u0430\u043d\u044b<\/h2>\n
\u0412 \u043a\u0430\u0447\u0435\u0441\u0442\u0432\u0435 \u0443\u043b\u0443\u0447\u0448\u0435\u043d\u0438\u044f \u043f\u0440\u0435\u0434\u0441\u0442\u0430\u0432\u043b\u0435\u043d\u043d\u043e\u0433\u043e \u0430\u043b\u0433\u043e\u0440\u0438\u0442\u043c\u0430 \u043c\u043e\u0436\u043d\u043e \u0440\u0430\u0441\u0448\u0438\u0440\u0438\u0442\u044c \u043d\u0430\u0431\u043e\u0440 \u0441\u043b\u043e\u0451\u0432, \u0434\u043e\u0431\u0430\u0432\u0438\u0442\u044c skip connections, \u0440\u0435\u0430\u043b\u0438\u0437\u043e\u0432\u0430\u0442\u044c \u0432\u043e\u0437\u043c\u043e\u0436\u043d\u043e\u0441\u0442\u044c \u0441\u043e\u0437\u0434\u0430\u043d\u0438\u044f \u0431\u043e\u043b\u0435\u0435 \u0441\u043b\u043e\u0436\u043d\u043e\u0439 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u044b, \u0430 \u0442\u0430\u043a\u0436\u0435 \u0441\u0434\u0435\u043b\u0430\u0442\u044c \u0431\u043e\u043b\u0435\u0435 \u043c\u044f\u0433\u043a\u0438\u0435 \u0442\u0440\u0435\u0431\u043e\u0432\u0430\u043d\u0438\u044f \u043a\u043e \u0432\u0445\u043e\u0434\u043d\u044b\u043c \u0434\u0430\u043d\u043d\u044b\u043c \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0438.<\/p>\n
\u0421\u0441\u044b\u043b\u043a\u0430 \u043d\u0430 \u0433\u0438\u0442\u0445\u0430\u0431 \u0441 ipynb \u0444\u0430\u0439\u043b\u043e\u043c \u043a\u043e\u0434\u0430.<\/a><\/p>\n
\u0422\u044d\u0433\u0438: pytorch, RL, reinforcement learning, \u0433\u0435\u043d\u0435\u0440\u0430\u0446\u0438\u044f \u043d\u0435\u0439\u0440\u043e\u043d\u043d\u044b\u0445 \u0441\u0435\u0442\u0435\u0439, \u043d\u0435\u0439\u0440\u043e\u043d\u043d\u0430\u044f \u0441\u0435\u0442\u044c, neural network.  <\/p>\n<\/p>\n<\/div>\n<\/div>\n<\/div>\n
<\/div>\n

 \u0441\u0441\u044b\u043b\u043a\u0430 \u043d\u0430 \u043e\u0440\u0438\u0433\u0438\u043d\u0430\u043b \u0441\u0442\u0430\u0442\u044c\u0438  https:\/\/habr.com\/ru\/articles\/763706\/<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"
<\/div>\n
\n
\n
\n
\n
\u041c\u043d\u043e\u0433\u0438\u0435 \u043b\u044e\u0434\u0438, \u043a\u043e\u0433\u0434\u0430 \u043b\u0438\u0431\u043e \u0438\u043c\u0435\u0432\u0448\u0438\u0435 \u0434\u0435\u043b\u043e \u0441 \u043d\u0435\u0439\u0440\u043e\u043d\u043d\u044b\u043c\u0438 \u0441\u0435\u0442\u044f\u043c\u0438, \u043d\u0430\u0432\u0435\u0440\u043d\u044f\u043a\u0430 \u0437\u0430\u0434\u0443\u043c\u044b\u0432\u0430\u043b\u0438\u0441\u044c, \u043c\u043e\u0436\u043d\u043e \u043b\u0438 \u043d\u0430\u043f\u0438\u0441\u0430\u0442\u044c \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u044c, \u043a\u043e\u0442\u043e\u0440\u0430\u044f \u0441\u0430\u043c\u0430 \u0431\u0443\u0434\u0435\u0442 \u0441\u043e\u0437\u0434\u0430\u0432\u0430\u0442\u044c \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0438 \u0434\u043b\u044f \u0440\u0435\u0448\u0435\u043d\u0438\u044f \u043a\u0430\u043a\u0438\u0445-\u043b\u0438\u0431\u043e \u0437\u0430\u0434\u0430\u0447. \u0422\u0430\u043a \u0432\u043e\u0442 \u0432 \u044d\u0442\u043e\u043c \u0446\u0438\u043a\u043b\u0435 \u0441\u0442\u0430\u0442\u0435\u0439 \u044f \u0440\u0435\u0448\u0438\u043b \u0440\u0435\u0430\u043b\u0438\u0437\u043e\u0432\u0430\u0442\u044c \u044d\u0442\u043e. \u041e\u0434\u043d\u0438\u043c \u0438\u0437 \u044d\u0442\u0430\u043f\u043e\u0432 \u0430\u043b\u0433\u043e\u0440\u0438\u0442\u043c\u0430 \u0431\u0443\u0434\u0435\u0442 \u0433\u0435\u043d\u0435\u0440\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u0435 \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0438 \u0438\u0437 \u0441\u043f\u0438\u0441\u043a\u0430 \u0441\u043b\u043e\u0451\u0432. \u0412 \u0441\u0432\u044f\u0437\u0438 \u0441 \u043d\u0435\u043a\u043e\u0442\u043e\u0440\u044b\u043c\u0438 \u043e\u0433\u0440\u0430\u043d\u0438\u0447\u0435\u043d\u0438\u044f\u043c\u0438, \u043d\u0430\u043a\u043b\u0430\u0434\u044b\u0432\u0430\u0435\u043c\u044b\u043c\u0438 \u043c\u0435\u0442\u043e\u0434\u0430\u043c\u0438 \u0440\u0435\u0430\u043b\u0438\u0437\u0430\u0446\u0438\u0438 (\u043e \u043a\u043e\u0442\u043e\u0440\u044b\u0445 \u0431\u0443\u0434\u0435\u0442 \u0441\u043a\u0430\u0437\u0430\u043d\u043e \u0432 \u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0438\u0445 \u0447\u0430\u0441\u0442\u044f\u0445, \u043a\u043e\u0433\u0434\u0430 \u043c\u044b \u043d\u0430\u0447\u043d\u0451\u043c \u043e\u0431\u044a\u0435\u0434\u0438\u043d\u044f\u0442\u044c \u043a\u043e\u0434 \u0438\u0437 \u044d\u0442\u043e\u0439 \u0441\u0442\u0430\u0442\u044c\u0438 \u0441 RL \u0295\u2299\u1d25\u2299\u0294 ), \u0432\u0445\u043e\u0434\u043d\u044b\u0435 \u0434\u0430\u043d\u043d\u044b\u0435 \u0434\u043b\u044f \u0433\u0435\u043d\u0435\u0440\u0430\u0442\u043e\u0440\u0430 \u0431\u0443\u0434\u0443\u0442 \u043f\u0440\u0435\u0434\u0441\u0442\u0430\u0432\u043b\u0435\u043d\u044b \u0432 \u0432\u0438\u0434\u0435 \u0441\u0442\u0440\u043e\u043a\u0438 \u0441\u043b\u0443\u0447\u0430\u0439\u043d\u043e\u0439 \u0434\u043b\u0438\u043d\u044b, \u0441\u043e\u0434\u0435\u0440\u0436\u0430\u0449\u0435\u0439 \u0443\u043f\u043e\u0440\u044f\u0434\u043e\u0447\u0435\u043d\u043d\u044b\u0439 \u043d\u0430\u0431\u043e\u0440 \u0441\u043b\u043e\u0451\u0432 \u0441 \u0438\u0445 \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u0430\u043c\u0438. \u0413\u0435\u043d\u0435\u0440\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u0441\u0435\u0442\u044c \u0431\u0443\u0434\u0435\u043c \u0434\u043b\u044f \u0437\u0430\u0434\u0430\u0447\u0438 \u043a\u043b\u0430\u0441\u0441\u0438\u0444\u0438\u043a\u0430\u0446\u0438\u0438 \u043a\u0430\u0440\u0442\u0438\u043d\u043e\u043a (\u0440\u0430\u0437\u043e\u0431\u044c\u0451\u043c \u044d\u0442\u043e \u043f\u0443\u0433\u0430\u043b\u043e \u043f\u0435\u0440\u0432\u044b\u043c).<\/p>\n
\u0424\u0443\u043d\u043a\u0446\u0438\u044f \u0441\u043e\u0437\u0434\u0430\u043d\u0438\u044f \u043e\u043f\u0438\u0441\u0430\u043d\u0438\u044f \u043f\u043e\u0441\u043b\u0435\u0434\u043e\u0432\u0430\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u0438 \u0441\u043b\u043e\u0451\u0432<\/h2>\n
\u041f\u0435\u0440\u0435\u0434 \u0433\u0435\u043d\u0435\u0440\u0430\u0446\u0438\u0435\u0439 \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0438 \u0441\u043e\u0437\u0434\u0430\u0434\u0438\u043c \u0441\u0442\u0440\u043e\u043a\u0443 \u0441\u043b\u0443\u0447\u0430\u0439\u043d\u043e \u0432\u044b\u0431\u0440\u0430\u043d\u043d\u044b\u0445 \u0441\u043b\u043e\u0451\u0432. \u041e\u0434\u043d\u0438\u043c \u0438\u0437 \u0431\u043e\u043b\u044c\u0448\u0438\u0445 \u043c\u0438\u043d\u0443\u0441\u043e\u0432, \u0441\u043e\u0437\u0434\u0430\u0432\u0430\u0435\u043c\u044b\u0445 \u043e\u0433\u0440\u0430\u043d\u0438\u0447\u0435\u043d\u0438\u044f\u043c\u0438, \u044f\u0432\u043b\u044f\u0435\u0442\u0441\u044f \u0442\u043e, \u0447\u0442\u043e \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u044b \u0441\u043b\u043e\u0451\u0432 \u043d\u0435\u043e\u0431\u0445\u043e\u0434\u0438\u043c\u043e \u0433\u0435\u043d\u0435\u0440\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u0432\u043c\u0435\u0441\u0442\u0435 \u0441\u043e \u0441\u043b\u043e\u044f\u043c\u0438, \u0442.\u0435. \u0433\u0435\u043d\u0435\u0440\u0438\u0440\u0443\u0435\u0442\u0441\u044f \u0441\u043b\u043e\u0439 \u0441\u0440\u0430\u0437\u0443 \u0441 \u0437\u0430\u0434\u0430\u043d\u043d\u044b\u043c\u0438 \u0444\u0438\u043a\u0441\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u044b\u043c\u0438 \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u0430\u043c\u0438. \u0420\u0430\u0441\u0448\u0438\u0440\u0435\u043d\u0438\u0435 \u043c\u043d\u043e\u0436\u0435\u0441\u0442\u0432\u0430 \u0432\u043e\u0437\u043c\u043e\u0436\u043d\u043e \u043d\u0430 \u0443\u0441\u043c\u043e\u0442\u0440\u0435\u043d\u0438\u0435 \u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u0435\u043b\u044f.<\/p>\n
layers_set = {                 # increase channels with kernel 3                 'conv=channel_factor:2,kernel_size:3,stride:1,padding:0-',                 'conv=channel_factor:3,kernel_size:3,stride:1,padding:0-',                 'conv=channel_factor:4,kernel_size:3,stride:1,padding:0-',                  # decrease channels with kernel 3                 'conv=channel_factor:0.4,kernel_size:3,stride:1,padding:0-',                 'conv=channel_factor:0.6,kernel_size:3,stride:1,padding:0-',                 'conv=channel_factor:0.8,kernel_size:3,stride:1,padding:0-',                    'batchnorm=eps:0.00001-',                  'batchnorm=eps:0.0001-',                   'avgpool=kernel_size:2,stride:2,padding:0-',                  'avgpool=kernel_size:4,stride:4,padding:0-',                   'maxpool=kernel_size:2,stride:2,padding:0-',                  'maxpool=kernel_size:4,stride:4,padding:0-',                   'dropout=p:0.2-',                  'dropout=p:0.4-',   }<\/code><\/pre>\n
channel_factor — \u043c\u043d\u043e\u0436\u0438\u0442\u0435\u043b\u044c \u043a\u043e\u043b-\u0432\u0430 \u043a\u0430\u043d\u0430\u043b\u043e\u0432.
«-» —  \u0440\u0430\u0437\u0434\u0435\u043b\u0438\u0442\u0435\u043b\u044c \u0441\u043b\u043e\u0451\u0432
«,» — \u0440\u0430\u0437\u0434\u0435\u043b\u0438\u0442\u0435\u043b\u044c \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u043e\u0432 \u0441\u043b\u043e\u044f
«:» — \u0440\u0430\u0437\u0434\u0435\u043b\u0438\u0442\u0435\u043b\u044c \u043d\u0430\u0437\u0432\u0430\u043d\u0438\u044f \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u0430 \u0438 \u0435\u0433\u043e \u0437\u043d\u0430\u0447\u0435\u043d\u0438\u044f
«=» — \u0440\u0430\u0437\u0434\u0435\u043b\u0438\u0442\u0435\u043b\u044c \u043d\u0430\u0437\u0432\u0430\u043d\u0438\u044f \u0441\u043b\u043e\u044f \u043e\u0442 \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u043e\u0432<\/p>\n
\u0420\u0435\u0430\u043b\u0438\u0437\u0443\u0435\u043c \u0444\u0443\u043d\u043a\u0446\u0438\u044e, \u0441\u043e\u0437\u0434\u0430\u044e\u0449\u0443\u044e \u0441\u0442\u0440\u043e\u043a\u0443 \u0441\u043b\u043e\u0451\u0432 \u0441 \u0434\u043b\u0438\u043d\u043e\u0439 min_len <= L <= max_len \u0441\u043b\u043e\u0451\u0432.<\/p>\n
def create_layers_string(min_len=2, max_len=10):     if min_len < 1 or max_len < min_len:       print('Parameters are incorrect!')       return None     length = random.randint(min_len, max_len)     text = random.sample(layers_set, length)     text = ''.join((layer for layer in text))     return text<\/code><\/pre>\n
\u041a\u043b\u0430\u0441\u0441 \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0438<\/h2>\n
\u0412 \u043a\u0430\u0447\u0435\u0441\u0442\u0432\u0435 \u043a\u043b\u0430\u0441\u0441\u0430 \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0438 \u0432\u043e\u0437\u044c\u043c\u0451\u043c \u043f\u0440\u043e\u0441\u0442\u043e\u0439 \u043f\u0443\u0441\u0442\u043e\u0439 \u0448\u0430\u0431\u043b\u043e\u043d, \u043f\u043e\u043b\u0435 «layers» \u043a\u043e\u0442\u043e\u0440\u043e\u0433\u043e \u0431\u0443\u0434\u0435\u0442 \u0441\u043e\u0434\u0435\u0440\u0436\u0430\u0442\u044c \u043d\u0430\u0448\u0443 \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0435\u0432\u0443\u044e \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0443, \u043a\u043e\u0433\u0434\u0430 \u043c\u044b \u0435\u0451 \u0441\u0433\u0435\u043d\u0435\u0440\u0438\u0440\u0443\u0435\u043c.<\/p>\n
class NN(nn.Module):      def __init__(self, in_channels):       super().__init__()        self.layers = nn.Sequential()      def forward(self, x):       x = self.layers(x)       return x      def __call__(self, x):       return self.forward(x)<\/code><\/pre>\n
\u0417\u0430\u043c\u0435\u0447\u0443, \u0447\u0442\u043e \u043a\u043b\u0430\u0441\u0441\u0438\u0444\u0438\u043a\u0430\u0442\u043e\u0440\u0430 \u0432 \u0432\u0438\u0434\u0435 \u043b\u0438\u043d\u0435\u0439\u043d\u043e\u0433\u043e \u0441\u043b\u043e\u044f \u0442\u0443\u0442 \u043d\u0435\u0442. \u041e\u043d \u0431\u0443\u0434\u0435\u0442 \u0434\u043e\u0431\u0430\u0432\u043b\u0435\u043d \u0432 \u043a\u043e\u043d\u0435\u0446 layers \u043f\u043e\u0441\u043b\u0435 \u0432\u044b\u0447\u0438\u0441\u043b\u0435\u043d\u0438\u044f \u0432\u0445\u043e\u0434\u043d\u044b\u0445 \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u043e\u0432 \u0434\u043b\u044f \u043d\u0435\u0433\u043e.<\/p>\n
\u041a\u043b\u0430\u0441\u0441 nnGenerator <\/h2>\n
\u0421\u0430\u043c\u043e\u0439 \u0431\u043e\u043b\u044c\u0448\u043e\u0439 \u0447\u0430\u0441\u0442\u044c\u044e \u043a\u043e\u0434\u0430 \u0431\u0443\u0434\u0435\u0442 \u043a\u043b\u0430\u0441\u0441 nnGenerator, \u0441\u043e\u0434\u0435\u0440\u0436\u0430\u0449\u0438\u0439 \u0432 \u0441\u0435\u0431\u0435 \u043f\u043e\u043c\u0438\u043c\u043e \u043c\u0435\u0442\u043e\u0434\u0430 \u0433\u0435\u043d\u0435\u0440\u0430\u0446\u0438\u0438 \u043c\u0435\u0442\u043e\u0434 \u043f\u0430\u0440\u0441\u0438\u043d\u0433\u0430 \u043d\u0430\u0448\u0435\u0433\u043e \u0432\u0445\u043e\u0434\u043d\u043e\u0433\u043e \u0442\u0435\u043a\u0441\u0442\u0430, \u0441\u043e\u0434\u0435\u0440\u0436\u0430\u0449\u0435\u0433\u043e \u0441\u043b\u043e\u0438. 
\u041f\u0430\u0440\u0441\u0438\u043d\u0433 \u0441\u0442\u0440\u043e\u043a\u0438 \u0431\u0443\u0434\u0435\u0442 \u043f\u0440\u043e\u0438\u0437\u0432\u0435\u0434\u0451\u043d \u043f\u043e \u043e\u0431\u043e\u0437\u043d\u0430\u0447\u0435\u043d\u043d\u044b\u043c \u0432\u044b\u0448\u0435 \u0440\u0430\u0437\u0434\u0435\u043b\u0438\u0442\u0435\u043b\u044f\u043c.<\/p>\n
class nnGenerator():     def __init__(self):       self.text_layers_dict = dict({})       self.nn_len = -1<\/code><\/pre>\n
\u041a\u043b\u0430\u0441\u0441 \u0441\u043e\u0434\u0435\u0440\u0436\u0438\u0442 \u0441\u043b\u043e\u0432\u0430\u0440\u044c, \u0445\u0440\u0430\u043d\u044f\u0449\u0438\u0439 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442 \u043f\u0430\u0440\u0441\u0438\u043d\u0433\u0430 \u0441\u0442\u0440\u043e\u043a\u0438 \u0441\u043b\u043e\u0451\u0432.
\u0421\u0442\u0440\u0443\u043a\u0442\u0443\u0440\u0430 \u0441\u043b\u043e\u0432\u0430\u0440\u044f \u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0430\u044f: 1 \u044d\u043b\u0435\u043c\u0435\u043d\u0442 \u0435\u0433\u043e \u0441\u043e\u0441\u0442\u043e\u0438\u0442 \u0438\u0437 \u043d\u0430\u0437\u0432\u0430\u043d\u0438\u0435 \u0441\u043b\u043e\u044f, \u0441\u043a\u043e\u043d\u043a\u0430\u0442\u0435\u043d\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u043e\u0435 \u0441 \u0435\u0433\u043e id + \u043f\u0430\u0440\u0430: id + \u0441\u043b\u043e\u0432\u0430\u0440\u044c \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u043e\u0432. \u041f\u0440\u0438\u043c\u0435\u0440 \u043e\u0434\u043d\u043e\u0433\u043e \u044d\u043b\u0435\u043c\u0435\u043d\u0442\u0430 \u0441\u043b\u043e\u0432\u0430\u0440\u044f text_layers_dict:
«dropout1»: (1, {«p»: 0.2} )<\/p>\n
    def parseTextNet(self, text_net):       self.text_layers_dict = dict({})       self.nn_len = -1       print(text_net)       if text_net[-1] == '-':         text_net = text_net[:-1]       text_layers = text_net.split('-')       id = 0       for text_layer in text_layers:         tmp = text_layer.split('=')         layer_name, layer_params = tmp[0], tmp[1].split(',')         layer_params_dict = dict({})         for param in layer_params:           param = param.split(':')           param_name, param_value = param[0], param[1]           layer_params_dict[param_name] = param_value         self.text_layers_dict[layer_name + str(id)] = (id, layer_params_dict)         id += 1       print(self.text_layers_dict) <\/code><\/pre>\n
\u041f\u0430\u0440\u0441\u0435\u0440 \u043f\u0440\u0438\u043d\u0438\u043c\u0430\u0435\u0442 \u043d\u0430 \u0432\u0445\u043e\u0434 \u0441\u0442\u0440\u043e\u043a\u0443 \u0441\u043b\u043e\u0451\u0432 \u0438 \u043f\u0440\u0435\u043e\u0431\u0440\u0430\u0437\u0443\u0435\u0442 \u0435\u0451, \u0437\u0430\u043f\u0438\u0441\u044b\u0432\u0430\u044f \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442 \u0432 \u0441\u043b\u043e\u0432\u0430\u0440\u044c text_layers_dict. <\/p>\n
    def conv_output_shape(self, h, w, kernel_size=1, stride=1, pad=0, dilation=1):       h = math.floor( ((h + (2 * pad) - ( dilation * (kernel_size - 1) ) - 1 )\/ stride) + 1)       w = math.floor( ((w + (2 * pad) - ( dilation * (kernel_size - 1) ) - 1 )\/ stride) + 1)       return h, w  <\/code><\/pre>\n
\u041c\u0435\u0442\u043e\u0434 conv_output_shape \u0440\u0435\u0430\u043b\u0438\u0437\u0443\u0435\u0442 \u0432\u044b\u0447\u0438\u0441\u043b\u0435\u043d\u0438\u0435 \u0432\u044b\u0445\u043e\u0434\u0430 \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0435\u0432\u043e\u0433\u043e \u0441\u043b\u043e\u044f. 
\u0414\u043b\u044f \u0441\u0432\u0451\u0440\u0442\u043e\u043a \u0438 \u043f\u0443\u043b\u043b\u0438\u043d\u0433\u043e\u0432 \u044d\u0442\u0430 \u0444\u0443\u043d\u043a\u0446\u0438\u044f \u0443\u043d\u0438\u0432\u0435\u0440\u0441\u0430\u043b\u044c\u043d\u0430, dropout \u0438 batchnorm \u043d\u0435 \u043c\u0435\u043d\u044f\u044e\u0442 \u0440\u0430\u0437\u043c\u0435\u0440\u043d\u043e\u0441\u0442\u0435\u0439, \u0430 \u0437\u043d\u0430\u0447\u0438\u0442 \u0432 \u0432\u044b\u0447\u0438\u0441\u043b\u0435\u043d\u0438\u0438 \u0432\u044b\u0445\u043e\u0434\u0430 \u0441\u043b\u043e\u044f \u043d\u0435 \u043d\u0443\u0436\u0434\u0430\u044e\u0442\u0441\u044f. \u0424\u043e\u0440\u043c\u0443\u043b\u0430 \u0432\u044b\u0447\u0438\u0441\u043b\u0435\u043d\u0438\u044f \u044d\u0442\u0438\u0445 \u0437\u043d\u0430\u0447\u0435\u043d\u0438\u0439 \u043f\u0440\u0435\u0434\u0441\u0442\u0430\u0432\u043b\u0435\u043d\u0430 \u0432 \u0434\u043e\u043a\u0443\u043c\u0435\u043d\u0442\u0430\u0446\u0438\u0438 \u0431\u0438\u0431\u043b\u0438\u043e\u0442\u0435\u043a\u0438 PyTorch. <\/a><\/p>\n
\n
\u043a\u0432\u0430\u0434\u0440\u0430\u0442\u043d\u044b\u0435 \u0441\u043a\u043e\u0431\u043a\u0438 — \u043e\u043a\u0440\u0443\u0433\u043b\u0435\u043d\u0438\u0435 \u0432\u043d\u0438\u0437 (math.floor)<\/figcaption><\/div>\n<\/figure>\n
\u0414\u0430\u043b\u0435\u0435 \u043f\u043e\u0439\u0434\u0451\u0442 \u0441\u0430\u043c\u043e\u0435 \u0438\u043d\u0442\u0435\u0440\u0435\u0441\u043d\u043e\u0435 — \u0430\u0432\u0442\u043e\u043f\u043e\u0441\u0442\u0440\u043e\u0435\u043d\u0438\u0435 \u043d\u0430\u0448\u0435\u0439 \u0441\u0435\u0442\u0438 \u043f\u043e \u0441\u043b\u043e\u0432\u0430\u0440\u044e \u0441\u043b\u043e\u0451\u0432.<\/p>\n
    def generateNN(self, n_classes, test_batch):       success_state = False        backbone = nn.Sequential()       classifier = nn.Sequential()       optimizer = None       try:         data_shape = np.array(test_batch).shape # [B, C, H, W]         last_shape = data_shape         for layer_name in self.text_layers_dict.keys():             layer = None             layer_params= self.text_layers_dict[layer_name][1]             if layer_name.find('conv') >= 0:               kernel_size = int(layer_params['kernel_size'])               channel_factor = float(layer_params['channel_factor'])               stride = int(layer_params['stride'])               padding = int(layer_params['padding'])               dilation = 1                activation = nn.ReLU(inplace=True)                in_chan = last_shape[1]               assert(in_chan <= last_shape[2] and in_chan <= last_shape[3])               out_chan = math.floor(in_chan * channel_factor)               assert(out_chan > 0)                backbone.append(nn.Conv2d(in_chan, out_chan, kernel_size, stride, padding, dilation))               backbone.append(activation)                h, w = self.conv_output_shape(last_shape[2], last_shape[3], kernel_size, stride, padding, dilation)               last_shape = (last_shape[0], out_chan, h, w)              elif layer_name.find('batchnorm') >= 0:               eps = float(layer_params['eps'])               in_chan = last_shape[1]               backbone.append(nn.BatchNorm2d(in_chan, eps))              elif layer_name.find('avgpool') >= 0:               kernel_size = int(layer_params['kernel_size'])               stride = int(layer_params['stride'])               padding = int(layer_params['padding'])               dilation = 1               out_chan = last_shape[1]               backbone.append(nn.AvgPool2d(kernel_size, stride, padding))                h, w = self.conv_output_shape(last_shape[2], last_shape[3], kernel_size, stride, padding, dilation)               last_shape = (last_shape[0], out_chan, h, w)               elif layer_name.find('maxpool') >= 0:               kernel_size = int(layer_params['kernel_size'])               stride = int(layer_params['stride'])               padding = int(layer_params['padding'])               dilation = 1                backbone.append(nn.MaxPool2d(kernel_size, stride, padding))                h, w = self.conv_output_shape(last_shape[2], last_shape[3], kernel_size, stride, padding, dilation)               last_shape = (last_shape[0], last_shape[1], h, w)              elif layer_name.find('dropout') >= 0:               p = float(layer_params['p'])               backbone.append(nn.Dropout2d(p))          linear_in_shape = last_shape[1] * last_shape[2] * last_shape[3]         classifier = nn.Linear(linear_in_shape, n_classes)         success_state = True         print('NN build successfull!')                except Exception as e:         print('NN build failed!')         print(str(e))        net = nn.Sequential()       net.append(backbone)       net.append(nn.Flatten(start_dim=1))       net.append(classifier)       self.text_layers_dict = dict({})       return success_state, net<\/code><\/pre>\n
\u041f\u0441\u0435\u0432\u0434\u043e\u043a\u043e\u0434\u043e\u043c \u043c\u043e\u0436\u043d\u043e \u043e\u043f\u0438\u0441\u0430\u0442\u044c \u0430\u043b\u0433\u043e\u0440\u0438\u0442\u043c \u0442\u0430\u043a:
\u041f\u0440\u043e\u0445\u043e\u0434\u0438\u043c \u043f\u043e \u0432\u0441\u0435\u043c \u0441\u043b\u043e\u044f\u043c \u0438\u0437 \u0441\u043b\u043e\u0432\u0430\u0440\u044f:
     1) \u0411\u0435\u0440\u0451\u043c \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u044b \u0441\u043b\u043e\u044f
     2) \u0412\u044b\u0447\u0438\u0441\u043b\u044f\u0435\u043c \u0432\u044b\u0445\u043e\u0434\u043d\u043e\u0439 shape
     3) \u0421\u043e\u0437\u0434\u0430\u0451\u043c \u0441\u043b\u043e\u0439, \u0434\u043e\u0431\u0430\u0432\u043b\u044f\u0435\u043c \u0432 \u043a\u043e\u043d\u0435\u0446 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u044b
     4) \u0421\u043e\u0445\u0440\u0430\u043d\u044f\u0435\u043c \u0437\u043d\u0430\u0447\u0435\u043d\u0438\u0435 \u0432\u044b\u0445\u043e\u0434\u043d\u043e\u0433\u043e \u0440\u0430\u0437\u043c\u0435\u0440\u0430 \u043f\u043e\u0441\u043b\u0435\u0434\u043d\u0435\u0433\u043e \u0441\u043b\u043e\u044f
\u0412 \u043a\u043e\u043d\u0446\u0435 \u0434\u043e\u0431\u0430\u0432\u043b\u044f\u0435\u043c \u043b\u0438\u043d\u0435\u0439\u043d\u044b\u0439 \u043a\u043b\u0430\u0441\u0441\u0438\u0444\u0438\u043a\u0430\u0442\u043e\u0440. 
\u0414\u043b\u044f \u0432\u044b\u0447\u0438\u0441\u043b\u0435\u043d\u0438\u044f \u0432\u0445\u043e\u0434\u043d\u043e\u0433\u043e \u0440\u0430\u0437\u043c\u0435\u0440\u0430 \u0434\u0430\u043d\u043d\u044b\u0445 \u043f\u043e\u0434\u0430\u0451\u0442\u0441\u044f \u0442\u0435\u0441\u0442\u043e\u0432\u044b\u0439 \u0431\u0430\u0442\u0447, \u043d\u0430 \u0432\u0430\u0448\u0435 \u0443\u0441\u043c\u043e\u0442\u0440\u0435\u043d\u0438\u0435 \u043c\u043e\u0436\u043d\u043e \u043f\u0435\u0440\u0435\u0434\u0435\u043b\u0430\u0442\u044c \u0432 \u0431\u043e\u043b\u0435\u0435 \u0443\u0434\u043e\u0431\u043d\u044b\u0439 \u0432\u0438\u0434.<\/p>\n
\u041e\u0441\u043e\u0431\u044b\u0435 \u043c\u043e\u043c\u0435\u043d\u0442\u044b, \u043a\u043e\u0442\u043e\u0440\u044b\u0435 \u043d\u0435\u043e\u0431\u0445\u043e\u0434\u0438\u043c\u043e \u043e\u0442\u043b\u0430\u0432\u043b\u0438\u0432\u0430\u0442\u044c, \u044d\u0442\u043e \u0441\u043b\u0443\u0447\u0430\u0439, \u043f\u0440\u0438 \u043a\u043e\u0442\u043e\u0440\u043e\u043c \u0432 \u0445\u043e\u0434\u0435 \u043f\u043e\u0441\u0442\u0440\u043e\u0435\u043d\u0438\u044f \u0441\u0435\u0442\u0438 \u044f\u0434\u0440\u043e \u0441\u0442\u0430\u043d\u043e\u0432\u0438\u0442\u0441\u044f \u0431\u043e\u043b\u044c\u0448\u0435 \u0440\u0430\u0437\u043c\u0435\u0440\u0430 \u0434\u0430\u043d\u043d\u044b\u0445, \u0438 \u043f\u0440\u0438 \u043a\u043e\u0442\u043e\u0440\u043e\u043c \u043a\u043e\u043b-\u0432\u043e \u043a\u0430\u043d\u0430\u043b\u043e\u0432 \u0443\u043c\u0435\u043d\u044c\u0448\u0430\u0435\u0442\u0441\u044f \u0434\u043e \u0437\u043d\u0430\u0447\u0435\u043d\u0438\u044f < 1. <\/p>\n
\u0414\u0430\u043b\u0435\u0435 \u043a\u043e\u0434 \u0441\u043e\u0437\u0434\u0430\u043d\u0438\u044f dataloader’\u043e\u0432 \u043f\u043e\u0434 MNIST, \u0430 \u0442\u0430\u043a\u0436\u0435 \u043e\u0431\u0443\u0447\u0430\u044e\u0449\u0435\u0433\u043e \u0446\u0438\u043a\u043b\u0430, \u0434\u0443\u043c\u0430\u044e, \u0432 \u043a\u043e\u043c\u043c\u0435\u043d\u0442\u0430\u0440\u0438\u044f\u0445 \u044d\u0442\u043e \u043d\u0435 \u043d\u0443\u0436\u0434\u0430\u0435\u0442\u0441\u044f, \u043e\u0442\u043c\u0435\u0447\u0443 \u0442\u043e\u043b\u044c\u043a\u043e, \u0447\u0442\u043e \u0432\u0445\u043e\u0434\u043d\u044b\u0435 \u0434\u0430\u043d\u043d\u044b\u0435 \u0434\u043e\u043b\u0436\u043d\u044b \u0431\u044b\u0442\u044c (B, C, H, W) = (Batch_Size, Channels, height, width), \u0437\u043d\u0430\u0447\u0438\u0442 \u0434\u0430\u043d\u043d\u044b\u0435 \u0438\u0437 \u041c\u041d\u0418\u0421\u0422\u0430 \u0441 shape’\u043e\u043c (N, 28,28) \u043d\u0443\u0436\u043d\u043e \u043f\u0440\u0435\u0434\u0441\u0442\u0430\u0432\u0438\u0442\u044c \u0432 \u0432\u0438\u0434\u0435 (N, 1, 28, 28)<\/p>\n
 class ds(Dataset):     def __init__(self, X, y):        self.X = X       self.y = y      def __len__(self):         return len(self.y)      def __getitem__(self, idx):         x_ = self.X[idx]         y_ = self.y[idx]         return x_, y_  train_data = dsets.MNIST(root = '.\/data', train = True,                         transform = transforms.ToTensor(), download = True)  test_data = dsets.MNIST(root = '.\/data', train = False,                        transform = transforms.ToTensor())   train_samples = np.expand_dims(np.array(train_data.data), axis=1)[:5000] train_labels =<\/code><\/pre>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[],"tags":[],"class_list":["post-355973","post","type-post","status-publish","format-standard","hentry"],"_links":{"self":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/posts\/355973","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=355973"}],"version-history":[{"count":0,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/posts\/355973\/revisions"}],"wp:attachment":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=355973"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=355973"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=355973"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}