def build_model(input_layer=None): ################# # Regular model # ################# input_size = data_sizes["sliced:data:singleslice"] if input_layer: l0 = input_layer else: l0 = nn.layers.InputLayer(input_size) # Reshape to framemodel l0_slices = nn.layers.ReshapeLayer(l0, (-1, 1, [2], [3])) l1a = nn.layers.dnn.Conv2DDNNLayer(l0_slices, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=16, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l1b = nn.layers.dnn.Conv2DDNNLayer(l1a, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=16, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l1 = nn.layers.dnn.MaxPool2DDNNLayer(l1b, pool_size=(2, 2), stride=(2, 2)) l2a = nn.layers.dnn.Conv2DDNNLayer(l1, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l2b = nn.layers.dnn.Conv2DDNNLayer(l2a, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l2 = nn.layers.dnn.MaxPool2DDNNLayer(l2b, pool_size=(2, 2), stride=(2, 2)) l3a = nn.layers.dnn.Conv2DDNNLayer(l2, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=64, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l3b = nn.layers.dnn.Conv2DDNNLayer(l3a, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=64, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l3c = nn.layers.dnn.Conv2DDNNLayer(l3b, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=64, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l3 = nn.layers.dnn.MaxPool2DDNNLayer(l3c, pool_size=(2, 2), stride=(2, 2)) l4a = nn.layers.dnn.Conv2DDNNLayer(l3, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=128, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l4b = nn.layers.dnn.Conv2DDNNLayer(l4a, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=128, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l4c = nn.layers.dnn.Conv2DDNNLayer(l4b, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=128, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l4 = nn.layers.dnn.MaxPool2DDNNLayer(l4c, pool_size=(2, 2), stride=(2, 2)) l5a = nn.layers.dnn.Conv2DDNNLayer(l4, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=256, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l5b = nn.layers.dnn.Conv2DDNNLayer(l5a, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=256, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l5c = nn.layers.dnn.Conv2DDNNLayer(l5b, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=256, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l5 = nn.layers.dnn.MaxPool2DDNNLayer(l5c, pool_size=(2, 2), stride=(2, 2)) l5drop = nn.layers.dropout(l5, p=0.0) ld1 = nn.layers.DenseLayer(l5drop, num_units=512, W=nn.init.Orthogonal("relu"), b=nn.init.Constant(0.1), nonlinearity=nn.nonlinearities.rectify) ld1drop = nn.layers.dropout(ld1, p=0.0) ld2 = nn.layers.DenseLayer(ld1drop, num_units=512, W=nn.init.Orthogonal("relu"), b=nn.init.Constant(0.1), nonlinearity=nn.nonlinearities.rectify) ld2drop = nn.layers.dropout(ld2, p=0.0) ld3mu = nn.layers.DenseLayer(ld2drop, num_units=1, W=nn.init.Orthogonal("relu"), b=nn.init.Constant(200.0), nonlinearity=None) ld3sigma = nn.layers.DenseLayer(ld2drop, num_units=1, W=nn.init.Orthogonal("relu"), b=nn.init.Constant(50.0), nonlinearity=lb_softplus(3)) ld3musigma = nn.layers.ConcatLayer([ld3mu, ld3sigma], axis=1) # Reshape back to slicemodel ld3musigma_slices = nn.layers.ReshapeLayer(ld3musigma, (-1, NR_FRAMES, 2)) l_systole_musigma = layers.ArgmaxAndMaxLayer(ld3musigma_slices, 'min') l_systole = layers.MuSigmaErfLayer(l_systole_musigma) l_diastole_musigma = layers.ArgmaxAndMaxLayer(ld3musigma_slices, 'max') l_diastole = layers.MuSigmaErfLayer(l_diastole_musigma) return { "inputs": { "sliced:data:singleslice": l0 }, "outputs": { "systole": l_systole, "diastole": l_diastole, }, "regularizable": { ld1: l2_weight, ld2: l2_weight, ld3mu: l2_weight_out, ld3musigma: l2_weight_out, } }
def build_model(): ################# # Regular model # ################# input_size = data_sizes["sliced:data:sax"] input_size_mask = data_sizes["sliced:data:sax:is_not_padded"] input_size_locations = data_sizes["sliced:data:sax:locations"] l0 = nn.layers.InputLayer(input_size) lin_slice_mask = nn.layers.InputLayer(input_size_mask) lin_slice_locations = nn.layers.InputLayer(input_size_locations) # PREPROCESS SLICES SEPERATELY l0_slices = nn.layers.ReshapeLayer(l0, (batch_size * nr_slices, 30, patch_px, patch_px)) # (bxs, t, i, j) subsample_factor = 2 l0_slices_subsampled = nn.layers.SliceLayer(l0_slices, axis=1, indices=slice(0, 30, subsample_factor)) nr_frames_subsampled = 30 / subsample_factor # PREPROCESS FRAMES SEPERATELY l0_frames = nn.layers.ReshapeLayer(l0_slices_subsampled, (batch_size * nr_slices * nr_frames_subsampled, 1, patch_px, patch_px)) # (bxsxt, 1, i, j) l1a = nn.layers.dnn.Conv2DDNNLayer(l0_frames, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=16, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l1b = nn.layers.dnn.Conv2DDNNLayer(l1a, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=16, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l1c = nn.layers.dnn.Conv2DDNNLayer(l1b, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=16, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l1 = nn.layers.dnn.MaxPool2DDNNLayer(l1c, pool_size=(2,2), stride=(2,2)) l2a = nn.layers.dnn.Conv2DDNNLayer(l1, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=32, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l2b = nn.layers.dnn.Conv2DDNNLayer(l2a, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=32, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l2c = nn.layers.dnn.Conv2DDNNLayer(l2b, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=32, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l2 = nn.layers.dnn.MaxPool2DDNNLayer(l2c, pool_size=(2,2), stride=(2,2)) l3a = nn.layers.dnn.Conv2DDNNLayer(l2, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=64, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l3b = nn.layers.dnn.Conv2DDNNLayer(l3a, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=64, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l3c = nn.layers.dnn.Conv2DDNNLayer(l3b, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=64, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l3d = nn.layers.dnn.Conv2DDNNLayer(l3c, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=64, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l3 = nn.layers.dnn.MaxPool2DDNNLayer(l3d, pool_size=(2,2), stride=(2,2)) l4a = nn.layers.dnn.Conv2DDNNLayer(l3, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=128, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l4b = nn.layers.dnn.Conv2DDNNLayer(l4a, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=128, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l4c = nn.layers.dnn.Conv2DDNNLayer(l4b, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=128, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l4d = nn.layers.dnn.Conv2DDNNLayer(l4c, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=128, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l4 = nn.layers.dnn.MaxPool2DDNNLayer(l4d, pool_size=(2,2), stride=(2,2)) l5a = nn.layers.dnn.Conv2DDNNLayer(l4, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=128, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l5b = nn.layers.dnn.Conv2DDNNLayer(l5a, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=128, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l5c = nn.layers.dnn.Conv2DDNNLayer(l5b, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=128, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l5d = nn.layers.dnn.Conv2DDNNLayer(l5c, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=128, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l5 = nn.layers.dnn.MaxPool2DDNNLayer(l5d, pool_size=(2,2), stride=(2,2)) l5drop = nn.layers.dropout(l5, p=0.5) ld1 = nn.layers.DenseLayer(l5drop, num_units=256, W=nn.init.Orthogonal("relu"), b=nn.init.Constant(0.1), nonlinearity=nn.nonlinearities.rectify) ld1drop = nn.layers.dropout(ld1, p=0.5) ld2 = nn.layers.DenseLayer(ld1drop, num_units=256, W=nn.init.Orthogonal("relu"),b=nn.init.Constant(0.1), nonlinearity=nn.nonlinearities.rectify) ld2drop = nn.layers.dropout(ld2, p=0.5) ld3mu = nn.layers.DenseLayer(ld2drop, num_units=1, W=nn.init.Orthogonal("relu"), b=nn.init.Constant(15.0), nonlinearity=None) ld3sigma = nn.layers.DenseLayer(ld2drop, num_units=1, W=nn.init.Orthogonal("relu"), b=nn.init.Constant(5.0), nonlinearity=lb_softplus(.3)) ld3musigma = nn.layers.ConcatLayer([ld3mu, ld3sigma], axis=1) # Go back to a per slice model l_slices_musigma = nn.layers.ReshapeLayer(ld3musigma, (batch_size * nr_slices, nr_frames_subsampled, 2)) # (bxs, t, 2) l_slices_musigma_sys = layers.ArgmaxAndMaxLayer(l_slices_musigma, mode='min') # (bxs, 2) l_slices_musigma_dia = layers.ArgmaxAndMaxLayer(l_slices_musigma, mode='max') # (bxs, 2) # AGGREGATE SLICES PER PATIENT l_scaled_slice_locations = layers.TrainableScaleLayer(lin_slice_locations, scale=nn.init.Constant(0.1), trainable=False) # Systole l_pat_sys_ss_musigma = nn.layers.ReshapeLayer(l_slices_musigma_sys, (batch_size, nr_slices, 2)) l_pat_sys_ss_mu = nn.layers.SliceLayer(l_pat_sys_ss_musigma, indices=0, axis=-1) l_pat_sys_ss_sigma = nn.layers.SliceLayer(l_pat_sys_ss_musigma, indices=1, axis=-1) l_pat_sys_aggr_mu_sigma = layers.JeroenLayer([l_pat_sys_ss_mu, l_pat_sys_ss_sigma, lin_slice_mask, l_scaled_slice_locations], rescale_input=1.) l_systole = layers.MuSigmaErfLayer(l_pat_sys_aggr_mu_sigma) # Diastole l_pat_dia_ss_musigma = nn.layers.ReshapeLayer(l_slices_musigma_dia, (batch_size, nr_slices, 2)) l_pat_dia_ss_mu = nn.layers.SliceLayer(l_pat_dia_ss_musigma, indices=0, axis=-1) l_pat_dia_ss_sigma = nn.layers.SliceLayer(l_pat_dia_ss_musigma, indices=1, axis=-1) l_pat_dia_aggr_mu_sigma = layers.JeroenLayer([l_pat_dia_ss_mu, l_pat_dia_ss_sigma, lin_slice_mask, l_scaled_slice_locations], rescale_input=1.) l_diastole = layers.MuSigmaErfLayer(l_pat_dia_aggr_mu_sigma) return { "inputs":{ "sliced:data:sax": l0, "sliced:data:sax:is_not_padded": lin_slice_mask, "sliced:data:sax:locations": lin_slice_locations, }, "outputs": { "systole": l_systole, "diastole": l_diastole, }, "regularizable": { }, }
def build_model(input_layer=None): ################# # Regular model # ################# l_4ch = nn.layers.InputLayer(data_sizes["sliced:data:chanzoom:4ch"]) l_2ch = nn.layers.InputLayer(data_sizes["sliced:data:chanzoom:2ch"]) # Add an axis to concatenate over later l_4chr = nn.layers.ReshapeLayer(l_4ch, ( batch_size, 1, ) + l_4ch.output_shape[1:]) l_2chr = nn.layers.ReshapeLayer(l_2ch, ( batch_size, 1, ) + l_2ch.output_shape[1:]) # Cut the images in half, flip the left ones l_4ch_left = nn.layers.SliceLayer(l_4chr, indices=slice(image_size // 2 - 1, None, -1), axis=-1) l_4ch_right = nn.layers.SliceLayer(l_4chr, indices=slice(image_size // 2, None, 1), axis=-1) l_2ch_left = nn.layers.SliceLayer(l_2chr, indices=slice(image_size // 2 - 1, None, -1), axis=-1) l_2ch_right = nn.layers.SliceLayer(l_2chr, indices=slice(image_size // 2, None, 1), axis=-1) # Concatenate over second axis l_24lr = nn.layers.ConcatLayer( [l_4ch_left, l_4ch_right, l_2ch_left, l_2ch_right], axis=1) # b, 4, t, h, w # Subsample frames SUBSAMPLING_FACTOR = 2 nr_subsampled_frames = nr_frames // SUBSAMPLING_FACTOR l_24lr_ss = nn.layers.SliceLayer(l_24lr, indices=slice(None, None, SUBSAMPLING_FACTOR), axis=2) # Move frames and halves to batch, process them all in the same way, add channel axis l_halves = nn.layers.ReshapeLayer(l_24lr_ss, (batch_size * 4 * nr_subsampled_frames, 1, image_size, image_size // 2)) # First, do some convolutions in all directions l1a = nn.layers.dnn.Conv2DDNNLayer(l_halves, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=16, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l1b = nn.layers.dnn.Conv2DDNNLayer(l1a, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=16, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l1 = nn.layers.dnn.MaxPool2DDNNLayer(l1b, pool_size=(1, 2), stride=(1, 2)) # Then, only use the last axis l2a = nn.layers.dnn.Conv2DDNNLayer(l1, W=nn.init.Orthogonal("relu"), filter_size=(1, 3), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l2b = nn.layers.dnn.Conv2DDNNLayer(l2a, W=nn.init.Orthogonal("relu"), filter_size=(1, 3), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l2 = nn.layers.dnn.MaxPool2DDNNLayer(l2b, pool_size=(1, 2), stride=(1, 2)) l3a = nn.layers.dnn.Conv2DDNNLayer(l2, W=nn.init.Orthogonal("relu"), filter_size=(1, 3), num_filters=64, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l3b = nn.layers.dnn.Conv2DDNNLayer(l3a, W=nn.init.Orthogonal("relu"), filter_size=(1, 3), num_filters=64, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l3c = nn.layers.dnn.Conv2DDNNLayer(l3b, W=nn.init.Orthogonal("relu"), filter_size=(1, 3), num_filters=64, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l3 = nn.layers.dnn.MaxPool2DDNNLayer(l3c, pool_size=(1, 2), stride=(1, 2)) l4a = nn.layers.dnn.Conv2DDNNLayer(l3, W=nn.init.Orthogonal("relu"), filter_size=(1, 3), num_filters=128, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l4b = nn.layers.dnn.Conv2DDNNLayer(l4a, W=nn.init.Orthogonal("relu"), filter_size=(1, 3), num_filters=128, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l4c = nn.layers.dnn.Conv2DDNNLayer(l4b, W=nn.init.Orthogonal("relu"), filter_size=(1, 3), num_filters=128, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) l4 = nn.layers.dnn.MaxPool2DDNNLayer(l4c, pool_size=(1, 2), stride=(1, 2)) # Now, process each row seperately, by flipping the channel and height axis, and then putting height in the batch l4shuffle = nn.layers.DimshuffleLayer(l4, pattern=(0, 2, 1, 3)) l4rows = nn.layers.ReshapeLayer( l4shuffle, (batch_size * 4 * nr_subsampled_frames * image_size, l4shuffle.output_shape[-2], l4shuffle.output_shape[-1])) # Systole ld1 = nn.layers.DenseLayer(l4rows, num_units=256, W=nn.init.Orthogonal("relu"), b=nn.init.Constant(0.1), nonlinearity=nn.nonlinearities.rectify) ld1drop = nn.layers.dropout(ld1, p=0.5) ld2 = nn.layers.DenseLayer(ld1drop, num_units=256, W=nn.init.Orthogonal("relu"), b=nn.init.Constant(0.1), nonlinearity=nn.nonlinearities.rectify) ld2drop = nn.layers.dropout(ld2, p=0.5) ld3mu = nn.layers.DenseLayer(ld2drop, num_units=1, W=nn.init.Orthogonal("relu"), b=nn.init.Constant(16.0), nonlinearity=None) ld3sigma = nn.layers.DenseLayer(ld2drop, num_units=1, W=nn.init.Orthogonal("relu"), b=nn.init.Constant(4.0), nonlinearity=lb_softplus(.01)) ld3musigma = nn.layers.ConcatLayer([ld3mu, ld3sigma], axis=1) # Get the four halves back l_24lr_musigma = nn.layers.ReshapeLayer( ld3musigma, (batch_size, 4, nr_subsampled_frames, image_size, 2)) l_24lr_musigma_shuffle = nn.layers.DimshuffleLayer(l_24lr_musigma, pattern=(0, 2, 1, 3, 4)) l_24lr_musigma_re = nn.layers.ReshapeLayer( l_24lr_musigma_shuffle, (batch_size * nr_subsampled_frames, 4, image_size, 2)) l_4ch_left_musigma = nn.layers.SliceLayer(l_24lr_musigma_re, indices=0, axis=1) l_4ch_right_musigma = nn.layers.SliceLayer(l_24lr_musigma_re, indices=1, axis=1) l_2ch_left_musigma = nn.layers.SliceLayer(l_24lr_musigma_re, indices=2, axis=1) l_2ch_right_musigma = nn.layers.SliceLayer(l_24lr_musigma_re, indices=3, axis=1) l_4ch_musigma = layers.SumGaussLayer( [l_4ch_left_musigma, l_4ch_right_musigma]) l_2ch_musigma = layers.SumGaussLayer( [l_2ch_left_musigma, l_2ch_right_musigma]) l_musigma_frames = layers.IraLayerNoTime(l_4ch_musigma, l_2ch_musigma) # Minmax over time print l_musigma_frames.output_shape l_musigmas = nn.layers.ReshapeLayer(l_musigma_frames, (-1, nr_subsampled_frames, 2)) l_musigma_sys = layers.ArgmaxAndMaxLayer(l_musigmas, mode='min') l_musigma_dia = layers.ArgmaxAndMaxLayer(l_musigmas, mode='max') l_systole = layers.MuSigmaErfLayer(l_musigma_sys) l_diastole = layers.MuSigmaErfLayer(l_musigma_dia) return { "inputs": { "sliced:data:chanzoom:4ch": l_4ch, "sliced:data:chanzoom:2ch": l_2ch, }, "outputs": { "systole": l_systole, "diastole": l_diastole, }, "regularizable": {}, "meta_outputs": {} }
def build_model(): ################# # Regular model # ################# input_size = data_sizes["sliced:data:sax"] input_size_mask = data_sizes["sliced:data:sax:is_not_padded"] input_size_locations = data_sizes["sliced:data:sax:locations"] l0 = nn.layers.InputLayer(input_size) lin_slice_mask = nn.layers.InputLayer(input_size_mask) lin_slice_locations = nn.layers.InputLayer(input_size_locations) # PREPROCESS SLICES SEPERATELY l0_slices = nn.layers.ReshapeLayer( l0, (batch_size * nr_slices, 30, patch_px, patch_px)) # (bxs, t, i, j) subsample_factor = 2 l0_slices_subsampled = nn.layers.SliceLayer(l0_slices, axis=1, indices=slice( 0, 30, subsample_factor)) nr_frames_subsampled = 30 / subsample_factor # PREPROCESS FRAMES SEPERATELY l0_frames = nn.layers.ReshapeLayer( l0_slices_subsampled, (batch_size * nr_slices * nr_frames_subsampled, 1, patch_px, patch_px)) # (bxsxt, 1, i, j) # downsample downsample = lambda incoming: nn.layers.dnn.Pool2DDNNLayer( incoming, pool_size=(2, 2), stride=(2, 2), mode='average_inc_pad') upsample = lambda incoming: nn.layers.Upscale2DLayer(incoming, scale_factor=2) l0_frames_d0 = l0_frames l0_frames_d1 = downsample(l0_frames_d0) l0_frames_d2 = downsample(l0_frames_d1) l0_frames_d3 = downsample(l0_frames_d2) ld3a = nn.layers.dnn.Conv2DDNNLayer(l0_frames_d3, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=16, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld3b = nn.layers.dnn.Conv2DDNNLayer(ld3a, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=16, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld3c = nn.layers.dnn.Conv2DDNNLayer(ld3b, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=16, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld3o = nn.layers.dnn.Conv2DDNNLayer(ld3c, W=nn.init.Orthogonal("relu"), filter_size=(3, 3), num_filters=16, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld2i = nn.layers.ConcatLayer([l0_frames_d2, upsample(ld3o)], axis=1) ld2a = nn.layers.dnn.Conv2DDNNLayer(ld2i, W=nn.init.Orthogonal("relu"), filter_size=(5, 5), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld2b = nn.layers.dnn.Conv2DDNNLayer(ld2a, W=nn.init.Orthogonal("relu"), filter_size=(5, 5), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld2c = nn.layers.dnn.Conv2DDNNLayer(ld2b, W=nn.init.Orthogonal("relu"), filter_size=(5, 5), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld2d = nn.layers.dnn.Conv2DDNNLayer(ld2c, W=nn.init.Orthogonal("relu"), filter_size=(5, 5), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld2o = nn.layers.dnn.Conv2DDNNLayer(ld2d, W=nn.init.Orthogonal("relu"), filter_size=(5, 5), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld1i = nn.layers.ConcatLayer([l0_frames_d1, upsample(ld2o)], axis=1) ld1a = nn.layers.dnn.Conv2DDNNLayer(ld1i, W=nn.init.Orthogonal("relu"), filter_size=(5, 5), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld1b = nn.layers.dnn.Conv2DDNNLayer(ld1a, W=nn.init.Orthogonal("relu"), filter_size=(5, 5), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld1c = nn.layers.dnn.Conv2DDNNLayer(ld1b, W=nn.init.Orthogonal("relu"), filter_size=(5, 5), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld1d = nn.layers.dnn.Conv2DDNNLayer(ld1c, W=nn.init.Orthogonal("relu"), filter_size=(5, 5), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld1o = nn.layers.dnn.Conv2DDNNLayer(ld1d, W=nn.init.Orthogonal("relu"), filter_size=(5, 5), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld0i = nn.layers.ConcatLayer([l0_frames_d0, upsample(ld1o)], axis=1) ld0a = nn.layers.dnn.Conv2DDNNLayer(ld0i, W=nn.init.Orthogonal("relu"), filter_size=(5, 5), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld0b = nn.layers.dnn.Conv2DDNNLayer(ld0a, W=nn.init.Orthogonal("relu"), filter_size=(5, 5), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld0c = nn.layers.dnn.Conv2DDNNLayer(ld0b, W=nn.init.Orthogonal("relu"), filter_size=(5, 5), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld0d = nn.layers.dnn.Conv2DDNNLayer(ld0c, W=nn.init.Orthogonal("relu"), filter_size=(5, 5), num_filters=32, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.rectify) ld0o = nn.layers.dnn.Conv2DDNNLayer(ld0d, W=nn.init.Orthogonal("relu"), filter_size=(5, 5), num_filters=1, stride=(1, 1), pad="same", nonlinearity=nn.nonlinearities.sigmoid) ld0r = nn.layers.ReshapeLayer( ld0o, (batch_size * nr_slices * nr_frames_subsampled, patch_px, patch_px)) l_frames_musigma = layers.IntegrateAreaLayer(ld0r, sigma_mode='scale', sigma_scale=.1) area_per_pixel_cm = (float(patch_mm) / float(patch_px))**2 / 100.0 l_frames_musigma_cm = layers.TrainableScaleLayer( l_frames_musigma, scale=nn.init.Constant(area_per_pixel_cm), trainable=False) # Go back to a per slice model l_slices_musigma_cm = nn.layers.ReshapeLayer( l_frames_musigma_cm, (batch_size * nr_slices, nr_frames_subsampled, 2)) # (bxs, t, 2) l_slices_musigma_cm_sys = layers.ArgmaxAndMaxLayer(l_slices_musigma_cm, mode='min') # (bxs, 2) l_slices_musigma_cm_dia = layers.ArgmaxAndMaxLayer(l_slices_musigma_cm, mode='max') # (bxs, 2) l_slices_musigma_cm_avg = layers.ArgmaxAndMaxLayer(l_slices_musigma_cm, mode='mean') # AGGREGATE SLICES PER PATIENT l_scaled_slice_locations = layers.TrainableScaleLayer( lin_slice_locations, scale=nn.init.Constant(0.1), trainable=False) # Systole l_pat_sys_ss_musigma_cm = nn.layers.ReshapeLayer( l_slices_musigma_cm_sys, (batch_size, nr_slices, 2)) l_pat_sys_ss_mu_cm = nn.layers.SliceLayer(l_pat_sys_ss_musigma_cm, indices=0, axis=-1) l_pat_sys_ss_sigma_cm = nn.layers.SliceLayer(l_pat_sys_ss_musigma_cm, indices=1, axis=-1) l_pat_sys_ss_sigma_cm = nn.layers.TrainableScaleLayer( l_pat_sys_ss_sigma_cm) l_pat_sys_aggr_mu_sigma = layers.JeroenLayer([ l_pat_sys_ss_mu_cm, l_pat_sys_ss_sigma_cm, lin_slice_mask, l_scaled_slice_locations ], rescale_input=1.) l_systole = layers.MuSigmaErfLayer(l_pat_sys_aggr_mu_sigma) # Diastole l_pat_dia_ss_musigma_cm = nn.layers.ReshapeLayer( l_slices_musigma_cm_dia, (batch_size, nr_slices, 2)) l_pat_dia_ss_mu_cm = nn.layers.SliceLayer(l_pat_dia_ss_musigma_cm, indices=0, axis=-1) l_pat_dia_ss_sigma_cm = nn.layers.SliceLayer(l_pat_dia_ss_musigma_cm, indices=1, axis=-1) l_pat_dia_ss_sigma_cm = nn.layers.TrainableScaleLayer( l_pat_dia_ss_sigma_cm) l_pat_dia_aggr_mu_sigma = layers.JeroenLayer([ l_pat_dia_ss_mu_cm, l_pat_dia_ss_sigma_cm, lin_slice_mask, l_scaled_slice_locations ], rescale_input=1.) l_diastole = layers.MuSigmaErfLayer(l_pat_dia_aggr_mu_sigma) # Average l_pat_avg_ss_musigma_cm = nn.layers.ReshapeLayer( l_slices_musigma_cm_avg, (batch_size, nr_slices, 2)) l_pat_avg_ss_mu_cm = nn.layers.SliceLayer(l_pat_avg_ss_musigma_cm, indices=0, axis=-1) l_pat_avg_ss_sigma_cm = nn.layers.SliceLayer(l_pat_avg_ss_musigma_cm, indices=1, axis=-1) l_pat_avg_aggr_mu_sigma = layers.JeroenLayer([ l_pat_avg_ss_mu_cm, l_pat_avg_ss_sigma_cm, lin_slice_mask, l_scaled_slice_locations ], rescale_input=1.) l_mean = layers.MuSigmaErfLayer(l_pat_avg_aggr_mu_sigma) return { "inputs": { "sliced:data:sax": l0, "sliced:data:sax:is_not_padded": lin_slice_mask, "sliced:data:sax:locations": lin_slice_locations, }, "outputs": { "systole": l_systole, "diastole": l_diastole, "average": l_mean, }, "regularizable": {}, }
def build_model(input_layer = None): ################# # Regular model # ################# l_4ch = nn.layers.InputLayer(data_sizes["sliced:data:chanzoom:4ch"]) l_2ch = nn.layers.InputLayer(data_sizes["sliced:data:chanzoom:2ch"]) # Add an axis to concatenate over later l_4chr = nn.layers.ReshapeLayer(l_4ch, (batch_size, 1, ) + l_4ch.output_shape[1:]) l_2chr = nn.layers.ReshapeLayer(l_2ch, (batch_size, 1, ) + l_2ch.output_shape[1:]) # Cut the images in half, flip the left ones l_4ch_left = nn.layers.SliceLayer(l_4chr, indices=slice(image_size//2-1, None, -1), axis=-1) l_4ch_right = nn.layers.SliceLayer(l_4chr, indices=slice(image_size//2, None, 1), axis=-1) l_2ch_left = nn.layers.SliceLayer(l_2chr, indices=slice(image_size//2-1, None, -1), axis=-1) l_2ch_right = nn.layers.SliceLayer(l_2chr, indices=slice(image_size//2, None, 1), axis=-1) # Concatenate over second axis l_24lr = nn.layers.ConcatLayer([l_4ch_left, l_4ch_right, l_2ch_left, l_2ch_right], axis=1) # b, 4, t, h, w # Subsample frames SUBSAMPLING_FACTOR = 2 nr_subsampled_frames = nr_frames // SUBSAMPLING_FACTOR l_24lr_ss = nn.layers.SliceLayer(l_24lr, indices=slice(None, None, SUBSAMPLING_FACTOR), axis=2) # Move frames and halves to batch, process them all in the same way, add channel axis l_halves = nn.layers.ReshapeLayer(l_24lr_ss, (batch_size * 4 * nr_subsampled_frames, 1, image_size, image_size//2)) # First, do some convolutions in all directions num_channels = 64 l1a = nn.layers.dnn.Conv2DDNNLayer(l_halves, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=num_channels, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) l1b = nn.layers.dnn.Conv2DDNNLayer(l1a, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=num_channels, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify) # Then, put an rnn over the last axis l1_shuffle = nn.layers.DimshuffleLayer(l1b, pattern=(0, 2, 3, 1)) l1_r = nn.layers.ReshapeLayer(l1_shuffle, (batch_size * 4 * nr_subsampled_frames * image_size, image_size//2, num_channels)) l_rnn = rnn_layer(l1_r, 1024) ld3mu = nn.layers.DenseLayer(l_rnn, num_units=1, W=nn.init.Orthogonal("relu"), b=nn.init.Constant(16.0), nonlinearity=None) ld3sigma = nn.layers.DenseLayer(l_rnn, num_units=1, W=nn.init.Orthogonal("relu"), b=nn.init.Constant(4.0), nonlinearity=lb_softplus(.01)) ld3musigma = nn.layers.ConcatLayer([ld3mu, ld3sigma], axis=1) # Get the four halves back l_24lr_musigma = nn.layers.ReshapeLayer(ld3musigma, (batch_size, 4, nr_subsampled_frames, image_size, 2)) l_24lr_musigma_shuffle = nn.layers.DimshuffleLayer(l_24lr_musigma, pattern=(0, 2, 1, 3, 4)) l_24lr_musigma_re = nn.layers.ReshapeLayer(l_24lr_musigma_shuffle, (batch_size * nr_subsampled_frames, 4, image_size, 2)) l_4ch_left_musigma = nn.layers.SliceLayer(l_24lr_musigma_re, indices=0, axis=1) l_4ch_right_musigma = nn.layers.SliceLayer(l_24lr_musigma_re, indices=1, axis=1) l_2ch_left_musigma = nn.layers.SliceLayer(l_24lr_musigma_re, indices=2, axis=1) l_2ch_right_musigma = nn.layers.SliceLayer(l_24lr_musigma_re, indices=3, axis=1) l_4ch_musigma = layers.SumGaussLayer([l_4ch_left_musigma, l_4ch_right_musigma]) l_2ch_musigma = layers.SumGaussLayer([l_2ch_left_musigma, l_2ch_right_musigma]) l_musigma_frames = layers.IraLayerNoTime(l_4ch_musigma, l_2ch_musigma) # Minmax over time print(l_musigma_frames.output_shape) l_musigmas = nn.layers.ReshapeLayer(l_musigma_frames, (-1, nr_subsampled_frames, 2)) l_musigma_sys = layers.ArgmaxAndMaxLayer(l_musigmas, mode='min') l_musigma_dia = layers.ArgmaxAndMaxLayer(l_musigmas, mode='max') l_systole = layers.MuSigmaErfLayer(l_musigma_sys) l_diastole = layers.MuSigmaErfLayer(l_musigma_dia) return { "inputs":{ "sliced:data:chanzoom:4ch": l_4ch, "sliced:data:chanzoom:2ch": l_2ch, }, "outputs": { "systole": l_systole, "diastole": l_diastole, }, "regularizable": { }, "meta_outputs": { } }