def DeepArchNet(self, tensor):
tensor1 = self.Conv2d(tensor, out_c=50, size=(
8, 1), stride=(1, 1), padding=0)
tensor2 = self.Conv2d(tensor1, out_c=50, size=(
8, 1), stride=(1, 1), padding=0)
tensor3 = self.Conv2d(tensor2, out_c=50, size=(
8, 1), stride=(1, 1), padding=0)
tensor1 = Tensor(tensor1.h, tensor1.c, tensor1.w)
tensor3 = Tensor(tensor3.h, tensor3.c, tensor3.w)
tensor = self.Concat([tensor1, tensor3])
tensor_after_concat = Tensor(1, tensor.h*tensor.w, tensor.c)
tensor = self.LSTM(tensor_after_concat, 11, 0)
return tensor
def calculate(self):
'''
This function calculates the FLOPs
'''
tensor = Tensor(1, 128, 2)
tensor = self.IntroNet(tensor)
print('params: {}, flops: {}'.format(self.params, self.flops))
def calculate(self):
tensor = Tensor(1, 128, 2)
tensor = self.DeepArchNet(tensor)
print('params: {}, flops: {}'.format(self.params, self.flops))
def calculate(self):
tensor = Tensor(1, 2000, 600)
tensor = self.DistLearnWscNet(tensor)
print('params: {}, flops: {}'.format(self.params, self.flops))
def calculate(self):
tensor = Tensor(1, 64, 64)
tensor = self.SpecMonitorForRadarNet(tensor)
print('params: {}, flops: {}'.format(self.params, self.flops))
out.w *= self.num_heads
return self.Linear(out, 768)
def OutputEmbedding(self, input):
intermediate = self.Linear(input, 3072)
act = self.ReLU(intermediate)
out = self.Linear(act, 768)
return out
def BertLayer(self, embeddings):
att = self.SelfAttention(embeddings)
residual_att = self.Add(att, embeddings)
norm_att = self.LayerNorm(residual_att)
out = self.OutputEmbedding(norm_att)
residual_out = self.Add(out, norm_att)
norm_out = self.LayerNorm(residual_out)
return norm_out
def pipeline(self, embeddings):
for i in range(12):
embeddings = self.BertLayer(embeddings)
return embeddings
if __name__ == '__main__':
cal = BERTCalc(only_mac=True)
res = cal.pipeline(Tensor(1, 320, 768))
res.peek()
print("flops: ", cal.flops / 1e6)
def calculate(self):
tensor = Tensor(1, 128, 2)
tensor = self.ConvRadioModRecogNet(tensor)
print('params: {}, flops: {}'.format(self.params, self.flops))