| #------------------------------------------------------------- |
| # |
| # Licensed to the Apache Software Foundation (ASF) under one |
| # or more contributor license agreements. See the NOTICE file |
| # distributed with this work for additional information |
| # regarding copyright ownership. The ASF licenses this file |
| # to you under the Apache License, Version 2.0 (the |
| # "License"); you may not use this file except in compliance |
| # with the License. You may obtain a copy of the License at |
| # |
| # http://www.apache.org/licenses/LICENSE-2.0 |
| # |
| # Unless required by applicable law or agreed to in writing, |
| # software distributed under the License is distributed on an |
| # "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY |
| # KIND, either express or implied. See the License for the |
| # specific language governing permissions and limitations |
| # under the License. |
| # |
| #------------------------------------------------------------- |
| |
| /* |
| * 2D Transpose Convolutional layer. |
| * |
| * Utilizes built-in convolution operators for higher performance. |
| */ |
| source("scripts/nn/util.dml") as util |
| |
| forward = function(matrix[double] X, matrix[double] W, matrix[double] b, |
| int C, int Hin, int Win, int Hf, int Wf, |
| int strideh, int stridew, int padh, int padw, |
| int out_padh, int out_padw) |
| return (matrix[double] out, int Hout, int Wout){ |
| /* |
| * Computes the forward pass for a 2D spatial transpose convolutional |
| * layer with F filters. The input data has N examples, each |
| * represented as a 3D tensor flattened into a single vector. |
| * |
| * Inputs: |
| * - X: Inputs, of shape (N, C*Hin*Win). |
| * - W: Weights, of shape (C, F*Hf*Wf). |
| * - b: Biases, of shape (F, 1). |
| * - C: Number of input channels (dimensionality of depth). |
| * - Hin: Input height. |
| * - Win: Input width. |
| * - Hf: Filter height. |
| * - Wf: Filter width. |
| * - strideh: Stride over height. |
| * - stridew: Stride over width. |
| * - padh: Padding for top and bottom sides. |
| * - padw: Padding for left and right sides. |
| * - out_padh: extra padding for top side. This should |
| * lie in [0, strideh-1]. |
| * - out_padw: extra padding for right side. This should |
| * lie in [0, stridew-1]. |
| * |
| * Outputs: |
| * - out: Outputs, of shape (N, F*Hout*Wout). |
| * - Hout: Output height. |
| * - Wout: Output width. |
| */ |
| N = nrow(X) |
| F = nrow(b) |
| Hout = strideh*(Hin-1) - 2*padh + Hf + out_padh |
| Wout = stridew*(Win-1) - 2*padw + Wf + out_padw |
| |
| # Transpose convolution aims to go in the other direction of |
| # (direct) convolution, i.e., given input X, produce output O such |
| # that running convolution on O recovers X. This is achieved by |
| # conv2d_backward_data (since the derivative wrt data must produce |
| # output of same size as the input to conv2d). By reusing a built-in |
| # operator we achieve efficiency and restrict the number of built-in |
| # operators to manageable levels. Plus, most other deep-learning |
| # packages make use of the same strategy which means this |
| # implementation of transpose convolution is 'in-sync' with them. |
| # |
| # One potential downside of reusing conv2d_backward_data is the fact |
| # that it rotates the filter by 180 degrees before applying it. This |
| # needs to be kept in mind when interpreting the output of transpose |
| # convolution. |
| out = conv2d_backward_data(W, X, stride=[strideh,stridew], padding=[padh,padw], |
| input_shape=[N,F,Hout,Wout], filter_shape=[C,F,Hf,Wf]) |
| |
| # Add bias term to each output filter |
| out = bias_add(out, b) |
| } |
| |
| backward = function(matrix[double] dout, int Hout, int Wout, |
| matrix[double] X, matrix[double] W, matrix[double] b, |
| int C, int Hin, int Win, int Hf, int Wf, |
| int strideh, int stridew, int padh, int padw) |
| return (matrix[double] dX, matrix[double] dW, matrix[double] db){ |
| /* |
| * Computes the backward pass for a 2D spatial transpose |
| * convolutional layer with F filters. |
| * |
| * Inputs: |
| * - dout: Gradient wrt `out` from upstream, of |
| * shape (N, F*Hout*Wout). |
| * - Hout: Output height. |
| * - Wout: Output width. |
| * - X: Inputs, of shape (N, C*Hin*Win). |
| * - W: Weights, of shape (C, F*Hf*Wf). |
| * - b: Biases, of shape (F, 1). |
| * - C: Number of input channels (dimensionality of depth). |
| * - Hin: Input height. |
| * - Win: Input width. |
| * - Hf: Filter height. |
| * - Wf: Filter width. |
| * - strideh: Stride over height. |
| * - stridew: Stride over width. |
| * - padh: Padding for top and bottom sides. |
| * - padw: Padding for left and right sides. |
| * |
| * Outputs: |
| * - dX: Gradient wrt `X`, of shape (N, C*Hin*Win). |
| * - dW: Gradient wrt `W`, of shape (C, F*Hf*Wf). |
| * - db: Gradient wrt `b`, of shape (F, 1). |
| */ |
| N = nrow(X) |
| F = nrow(b) |
| |
| # conv2d_backward_filter takes the input and delta map as first and |
| # second args, respectively. Given that we need to compute the |
| # grad (wrt to filter) for transpose convolution where the roles of |
| # the input and output are reversed, we reverse the order of the |
| # args (along with setting input_shape to the delta map shape). |
| # Effectively, we are running a direct convolution with X as the |
| # filter and the dout as the input. To convince oneself that the |
| # interconnections between the cells of the filter, input and delta |
| # map are preserved please keep in mind that the forward of |
| # convolution transpose rotates the filter by 180 degrees before |
| # applying it. |
| dW = conv2d_backward_filter(dout, X, stride=[strideh,stridew], padding=[padh,padw], |
| input_shape=[N,F,Hout,Wout], filter_shape=[C,F,Hf,Wf]) |
| |
| # Since the forward for transpose convolution makes a call to |
| # conv2d_backward_data, to compute its derivative wrt to data |
| # we can run conv2d by applying the filter on the delta |
| # map (this makes sense because convolution transpose is the |
| # 'reverse' of convolution). Its easy to see that this will produce |
| # output of the required size. To convince oneself that conv2d will |
| # respect the interconnections between the cells in the delta map |
| # and the filter, keep in mind that the forward function rotates the |
| # filter by 180 degrees before applying it. |
| dX = conv2d(dout, W, input_shape=[N,F,Hout,Wout], filter_shape=[C,F,Hf,Wf], |
| stride=[strideh,stridew], padding=[padh,padw]) |
| |
| # Partial derivatives for bias vector |
| db = util::channel_sums(dout, F, Hout, Wout) |
| } |
| |
| init = function(int F, int C, int Hf, int Wf) |
| return (matrix[double] W, matrix[double] b){ |
| /* |
| * Utility function to initialize the parameters of this layer. |
| * |
| * We use the heuristic by He et al., which limits the magnification |
| * of inputs/gradients during forward/backward passes by scaling |
| * unit-Gaussian weights by a factor of sqrt(2/n), under the |
| * assumption of relu neurons. |
| * - http://arxiv.org/abs/1502.01852 |
| * |
| * Inputs: |
| * - F: Number of filters. |
| * - C: Number of input channels (dimensionality of depth). |
| * - Hf: Filter height. |
| * - Wf: Filter width. |
| * |
| * Outputs: |
| * - W: Weights, of shape (C, F*Hf*Wf). |
| * - b: Biases, of shape (F, 1). |
| */ |
| W = rand(rows=C, cols=F*Hf*Wf, pdf="normal") * sqrt(2/(C*Hf*Wf)) |
| b = matrix(0, rows=F, cols=1) |
| } |
| |
| init_bilinear = function(int C, int K) |
| return (matrix[double] W, matrix[double] b){ |
| /* |
| * Utility function to upsample using this layer. |
| * |
| * Upsampling the input by factor f (each side) requires |
| * channel-wise independent kernels of size K = 2f - f%2, |
| * stride = f and pad = ceil((f-1)/2). The weights are set |
| * via bilinear interpolation, bias is set to 0. |
| * |
| * Inputs: |
| * - C: Number of input channels (dimensionality of depth). |
| * - K: Kernel size (upsampling requires a square filter |
| * of size K X K). |
| * |
| * Outputs: |
| * - W: Weights, of shape (C, C*K*K). |
| * - b: Biases, of shape (C, 1). |
| */ |
| factor_up = ceil(K / 2) |
| center = (2 * factor_up - factor_up %% 2 - 1) / 2 / factor_up |
| vect = 1 - abs(seq(0, K-1) / factor_up - center) |
| weights = matrix(vect %*% t(vect), rows=1, cols=K*K) |
| |
| # To create a multi-channel channel-independent upsampling filter, |
| # we need to intersperse the filter weights with 0s. For instance, |
| # consider the case of 2X upsampling. In this case, K=4 and we have |
| # K^2=16 weights to include into the 3D tensor representing the |
| # filter which should look like the following (assuming 3 channels): |
| # |
| # <-16 weights-> <---------32 0s---------> |
| # X X ...... X X 0 0 0 ............. 0 0 0 |
| # 0 .......... 0 X X .... X X 0 ...... 0 0 |
| # 0 0 0 ............... 0 0 0 X X .... X X |
| # |
| # To be clear, the second row should have 16 0s followed by 16 |
| # weights followed by 16 0s. |
| # |
| # To create the above filter, we take advantage of the fact that |
| # between two sets of non-zero weights, there is always a sequence |
| # of C*K*K 0s. In the above example, C*K^2 = 48 (e.g., 32 trailing |
| # 0s in the first row and 16 leading 0s in the second row). |
| # |
| # Note that, in the special case of C=1 we do not need to |
| # intersperse with 0s (no question of being channel-wise independent |
| # since we have only 1 channel). |
| |
| # Append C*K*K trailing 0s to the K*K kernel and replicate the |
| # resulting row C times |
| repl_weights = matrix(1, rows=C, cols=1) %*% cbind(weights, matrix(0, rows=1, cols=C*K*K)) |
| |
| # The above operation added extra C*K*K trailing 0s in the last row |
| # that we do not need. Thus, we need to: |
| # 1) reshape the resulting matrix into a row |
| # 2) 'Clip off' the last few 0s using indexing and reshape the |
| # result into the expected filter shape ([C, C, K, K]) |
| repl_weights_row = matrix(repl_weights, rows=1, cols=C*(C+1)*K^2) |
| W = matrix(repl_weights_row[1,1:(C*K)^2], rows=C, cols=C*K^2) |
| |
| b = matrix(0, rows=C, cols=1) |
| } |