c3d_symbol.py

import find_mxnet
import mxnet as mx
'''
def get_symbol(num_classes):
    input_data = mx.symbol.Variable('data')
    label =mx.symbol.Variable('label')

    data = mx.symbol.Flatten(data=input_data)

    fc6 = mx.symbol.FullyConnected(data=data, num_hidden=20)
    relu6 = mx.symbol.Activation(data=fc6, act_type='relu')
    #Loss
    fc8 = mx.symbol.FullyConnected(data=relu6, num_hidden=num_classes)
    softmax = mx.symbol.SoftmaxOutput(data=fc8, label=label, name='softmax')
    return softmax
'''
def get_symbol(num_classes):
    input_data = mx.symbol.Variable('data')
    label = mx.symbol.Variable('label')
    
    #input_data = mx.symbol.Crop(data=input_data, num_args=2, h_w=(112,112))
    ## 1st group 
    conv1 = mx.symbol.Convolution(data=input_data, kernel=(3,3,3), stride=(1,1,1), pad=(1,1,1), num_filter=64, name='conv1', cudnn_tune='fastest', layout='NCDHW')
    relu1 = mx.symbol.Activation(data=conv1, act_type='relu')
    pool1 = mx.symbol.Pooling(data=relu1, pool_type='max', kernel=(1,2,2), stride=(1,2,2))
    
    ## 2nd group
    conv2 = mx.symbol.Convolution(data=pool1, kernel=(3,3,3), stride=(1,1,1), pad=(1,1,1), num_filter=128, name='conv2', cudnn_tune='fastest', layout='NCDHW')
    relu2 = mx.symbol.Activation(data=conv2, act_type='relu')
    pool2 = mx.symbol.Pooling(data=relu2, pool_type='max', kernel=(2,2,2), stride=(2,2,2))

    ## 3rd group
    conv3a = mx.symbol.Convolution(data=pool2, kernel=(3,3,3), stride=(1,1,1), pad=(1,1,1), num_filter=256, cudnn_tune='fastest', layout='NCDHW')
    relu3a = mx.symbol.Activation(data=conv3a, act_type='relu')
    #conv3b = mx.symbol.Convolution(data=relu3a, kernel=(3,3,3), stride=(1,1,1), num_filter=256)
    #relu3b = mx.symbol.Activation(data=conv3b, act_type='relu')
    pool3b = mx.symbol.Pooling(data=relu3a, pool_type='max', kernel=(2,2,2), stride=(2,2,2))

    ## 4th group
    conv4a = mx.symbol.Convolution(data=pool3b, kernel=(3,3,3), stride=(1,1,1), pad=(1,1,1), num_filter=256, cudnn_tune='fastest', layout='NCDHW')
    relu4a = mx.symbol.Activation(data=conv4a, act_type='relu')
    #conv4b = mx.symbol.Convolution(data=relu4a, kernel=(3,3,3), stride=(1,1,1), num_filter=512)
    #relu4b = mx.symbol.Activation(data=conv4b, act_type='relu')
    pool4b = mx.symbol.Pooling(data=relu4a, pool_type='max', kernel=(2,2,2), stride=(2,2,2))
    
    ## 5th group
    conv5a = mx.symbol.Convolution(data=pool4b, kernel=(3,3,3), stride=(1,1,1), pad=(1,1,1), num_filter=256, cudnn_tune='fastest', layout='NCDHW')
    relu5a = mx.symbol.Activation(data=conv5a, act_type='relu')
    #conv5b = mx.symbol.Convolution(data=relu5a, kernel=(3,3,3), stride=(1,1,1), num_filter=512)
    #relu5b = mx.symbol.Activation(data=conv5b, act_type='relu')
    pool5b = mx.symbol.Pooling(data=relu5a, pool_type='max', kernel=(2,2,2), stride=(2,2,2))
    
    ## 6th group
    fc6 = mx.symbol.FullyConnected(data=pool5b, num_hidden=2048)
    relu6 = mx.symbol.Activation(data=fc6, act_type='relu')
    drop6 = mx.symbol.Dropout(data=relu6, p=0.5)
    fc7 = mx.symbol.FullyConnected(data=drop6, num_hidden=2048)
    relu7 = mx.symbol.Activation(data=fc7, act_type='relu')
    drop7 = mx.symbol.Dropout(data=relu7, p=0.5)

    #Loss
    fc8 = mx.symbol.FullyConnected(data=drop7, num_hidden=num_classes)
    softmax = mx.symbol.SoftmaxOutput(data=fc8, label=label, name='softmax')
    return softmax