opfromgraph_doesnt_work_yet

opfromgraph_test.py

import theano
from theano import tensor as T
from theano import OpFromGraph
import lasagne
from lasagne.layers import *
from lasagne.init import *
from lasagne.nonlinearities import *
from lasagne.objectives import *
from lasagne.updates import *
from lasagne.regularization import *
import sys
import os


class OpFromGraphLayer(Layer):
    def __init__(self, incoming, block_fn, block_fn_args, inline=False,**kwargs):
        super(OpFromGraphLayer, self).__init__(incoming, **kwargs)    
        if isinstance(incoming, InputLayer):
            l_in = incoming
        else:
            # we must create a new input layer (and therefore a new input var)
            # which has the incoming layer's output shape
            l_in = InputLayer(incoming.output_shape)
        final = block_fn(l_in, **block_fn_args)
        final_out = get_output(final, l_in.input_var)
        self.final = final
        self.op = OpFromGraph([l_in.input_var], [final_out], inline=inline)
        for layer in get_all_layers(final):
            self.params.update( layer.params )
    def get_output_shape_for(self, input_shape):
        #return self.out_shp
        return get_output_shape(self.final, input_shape)
    def get_output_for(self, input, **kwargs):
        return self.op(input)


def residual_block(l_inp, n_out_channels, stride=1):
    conv1 = Conv2DLayer(l_inp, num_filters=n_out_channels, filter_size=3, stride=stride)
    #conv1 = BatchNormLayer(conv1)
    conv2 = Conv2DLayer(conv1, num_filters=n_out_channels, filter_size=3, stride=stride)
    return conv2

l_in = InputLayer((None,3,224,224))
l_resblock = OpFromGraphLayer(l_in, block_fn=residual_block, block_fn_args={'n_out_channels':32}, inline=False)
l_resblock2 = OpFromGraphLayer(l_resblock, block_fn=residual_block, block_fn_args={'n_out_channels':64, 'stride':2}, inline=False)

for layer in get_all_layers(l_resblock2):
    print layer, layer.output_shape

X = T.tensor4('X')
tmp = get_output(l_resblock, X)
theano.printing.pydotprint(tmp, outfile='test.png', var_with_name_simple=True)
tmp.eval({X:np.ones((1,3,224,224))})


theano.printing.pydotprint(tmp, outfile='test.png', var_with_name_simple=True)

# doesn't work for l_resblock2, but does for l_resblock
tmp.eval({X:np.ones((1,3,224,224))})