| #------------------------------------------------------------- |
| # |
| # 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. |
| # |
| #------------------------------------------------------------- |
| args <- commandArgs(TRUE) |
| library("Matrix") |
| imgSize=as.integer(args[1]) |
| numImg=as.integer(args[2]) |
| numChannels=as.integer(args[3]) |
| numFilters=as.integer(args[4]) |
| filterSize=as.integer(args[5]) |
| stride=as.integer(args[6]) |
| pad=as.integer(args[7]) |
| P=as.integer(args[8]) |
| Q=as.integer(args[9]) |
| |
| # Assumption: NCHW image format |
| x=matrix(seq(1, numImg*numChannels*imgSize*imgSize), numImg, numChannels*imgSize*imgSize, byrow=TRUE) |
| dout=matrix(seq(1, numImg*numFilters*P*Q), numImg, numFilters*P*Q, byrow=TRUE) |
| |
| if(as.logical(args[11])) { |
| zero_mask = (x - mean(x)*1.5) > 0 |
| x = x * zero_mask |
| } else { |
| x = x - mean(x) |
| } |
| if(as.logical(args[12])) { |
| zero_mask = (dout - mean(dout)*1.5) > 0 |
| dout = dout * zero_mask |
| } else { |
| dout = dout - mean(dout) |
| } |
| pad_image <- function(img, Hin, Win, padh, padw){ |
| C = nrow(img) |
| img_padded = matrix(0, C, (Hin+2*padh)*(Win+2*padw)) # zeros |
| for (c in 1:C) { |
| img_slice = matrix(img[c,], Hin, Win, byrow=TRUE) # depth slice C reshaped |
| img_padded_slice = matrix(0, Hin+2*padh, Win+2*padw) |
| img_padded_slice[(padh+1):(padh+Hin), (padw+1):(padw+Win)] = img_slice |
| img_padded[c,] = matrix(t(img_padded_slice), 1, (Hin+2*padh)*(Win+2*padw)) # reshape |
| } |
| img_padded |
| } |
| |
| im2col <- function(img, Hin, Win, Hf, Wf, strideh, stridew) { |
| C = nrow(img) |
| Hout = as.integer((Hin - Hf) / strideh + 1) |
| Wout = as.integer((Win - Wf) / stridew + 1) |
| |
| img_cols = matrix(0, C*Hf*Wf, Hout*Wout) # zeros |
| for (hout in 1:Hout) { # all output rows |
| hin = (hout-1) * strideh + 1 |
| for (wout in 1:Wout) { # all output columns |
| win = (wout-1) * stridew + 1 |
| # Extract a local patch of the input image corresponding spatially to the filter sizes. |
| img_patch = matrix(0, C, Hf*Wf) # zeros |
| for (c in 1:C) { # all channels |
| img_slice = matrix(img[c,], Hin, Win, byrow=TRUE) # reshape |
| img_patch[c,] = matrix(t(img_slice[hin:(hin+Hf-1), win:(win+Wf-1)]), 1, Hf*Wf) |
| } |
| img_cols[,(hout-1)*Wout + wout] = matrix(t(img_patch), C*Hf*Wf, 1) # reshape |
| } |
| } |
| img_cols |
| } |
| |
| conv2d_backward_filter <- function(dout, Hout, Wout, |
| X, N, K, C, Hin, Win, Hf, Wf, |
| strideh, stridew, padh, padw) { |
| |
| F = K |
| |
| # Create gradient volumes |
| dW = matrix(0, F, C*Hf*Wf, byrow=TRUE) |
| |
| # Create convenience gradient volumes for dW and db that will allow |
| # for one gradient to be stored per example, allowing for parallel |
| # computation at the expense of memory. We will reduce at the end. |
| dWN = matrix(0, N, F*C*Hf*Wf, byrow=TRUE) |
| |
| # Partial derivatives for convolution - im2col implementation |
| for (n in 1:N) { # all examples |
| doutn = matrix(dout[n,], F, Hout*Wout, byrow=TRUE) |
| |
| # Compute dW |
| Xn = matrix(X[n,], C, Hin*Win, byrow=TRUE) # reshape |
| Xn_padded = pad_image(Xn, Hin, Win, padh, padw) # shape (C, (Hin+2*padh)*(Win+2*padw)) |
| Xn_padded_cols = im2col(Xn_padded, Hin+2*padh, Win+2*padw, Hf, Wf, strideh, stridew) |
| #dW = dW + doutn %*% t(Xn_padded_cols) |
| dWN[n,] = matrix(t(doutn %*% t(Xn_padded_cols)), 1, F*C*Hf*Wf) |
| } |
| |
| # Reduce convenience gradient volumes with one gradient per example |
| # into single gradients for W and b. |
| dW = matrix(colSums(dWN), F, C*Hf*Wf, byrow=TRUE) |
| dW |
| } |
| |
| dw = conv2d_backward_filter(dout, P, Q, x, numImg, numFilters, numChannels, imgSize, imgSize, filterSize, filterSize, stride, stride, pad, pad) |
| |
| |
| writeMM(as(dw,"CsparseMatrix"), paste(args[10], "B", sep="")) |
