| #------------------------------------------------------------- |
| # |
| # 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. |
| # |
| #------------------------------------------------------------- |
| |
| source("scripts/nn/networks/resnet.dml") as resnet |
| source("src/test/scripts/applications/nn/util.dml") as test_util |
| source("scripts/nn/layers/conv2d_builtin.dml") as conv2d |
| |
| |
| basic_test_basic_block_forward = function(int C_in, int C_base, int strideh, int stridew, |
| int Hin, int Win, int N, Boolean downsample) { |
| /* |
| * Basic input tests for basic_block_forward() function |
| */ |
| X = matrix(0, rows=N, cols=C_in*Hin*Win) |
| W_conv1 = matrix(1, rows=C_base, cols=C_in*3*3) |
| gamma_bn1 = matrix(1, rows=C_base, cols=1) |
| beta_bn1 = matrix(0, rows=C_base, cols=1) |
| W_conv2 = matrix(0, rows=C_base, cols=C_base*3*3) |
| gamma_bn2 = matrix(1, rows=C_base, cols=1) |
| beta_bn2 = matrix(0, rows=C_base, cols=1) |
| weights = list(W_conv1, gamma_bn1, beta_bn1, W_conv2, gamma_bn2, beta_bn2) |
| if (downsample) { |
| W_conv3 = matrix(0, rows=C_base, cols=C_in*1*1) |
| gamma_bn3 = matrix(1, rows=C_base, cols=1) |
| beta_bn3 = matrix(0, rows=C_base, cols=1) |
| weights = append(weights, W_conv3) |
| weights = append(weights, gamma_bn3) |
| weights = append(weights, beta_bn3) |
| } |
| |
| |
| mode = "train" |
| ema_mean_bn1 = matrix(0, rows=C_base, cols=1) |
| ema_var_bn1 = matrix(1, rows=C_base, cols=1) |
| ema_mean_bn2 = matrix(0, rows=C_base, cols=1) |
| ema_var_bn2 = matrix(1, rows=C_base, cols=1) |
| ema_means_vars = list(ema_mean_bn1, ema_var_bn1, ema_mean_bn2, ema_var_bn2) |
| if (downsample) { |
| ema_mean_bn3 = matrix(0, rows=C_base, cols=1) |
| ema_var_bn3 = matrix(1, rows=C_base, cols=1) |
| ema_means_vars = append(ema_means_vars, ema_mean_bn3) |
| ema_means_vars = append(ema_means_vars, ema_var_bn3) |
| } |
| |
| [out, Hout, Wout, ema_means_vars_up, cached_out, cached_means_vars] = resnet::basic_block_forward(X, weights, C_in, |
| C_base, Hin, Win, strideh, stridew, mode, ema_means_vars) |
| Hout_expected = Hin / strideh; Wout_expected = Win / strideh |
| out_cols_expected = C_base * Hout_expected * Wout_expected |
| if (Hout_expected != Hout | Wout_expected != Wout | out_cols_expected != ncol(out)) { |
| test_util::fail("Output shapes of basic_block_forward() are not as expected!") |
| test_util::fail("Output shapes of basic_block_forward() are not as expected!") |
| } |
| } |
| |
| |
| values_test_basic_block_forward_1 = function() { |
| /* |
| * Testing of values for forward pass of basic block against PyTorch. |
| */ |
| strideh = 1; stridew = 1 |
| C_in = 2; C_base = 2 |
| Hin = 4; Win = 4 |
| N = 2 |
| X = matrix(1, rows=N, cols=C_in*Hin*Win) |
| W_conv1 = matrix("-0.13904892 0.12838013 0.08027278 -0.06143695 0.07755502 -0.16483936 0.06582125 0.00754158 0.10763083 0.04604699 0.03576668 -0.07599333 -0.06836346 0.19890614 0.01955454 -0.02767003 0.21198983 0.12785362 |
| 0.04019578 -0.14636862 -0.02285126 -0.00971214 0.12590824 -0.06414033 -0.1034085 -0.23452668 -0.0999288 0.12418596 -0.03290591 0.02420332 0.17950852 0.00047226 0.13068716 -0.00955899 0.03092374 0.05555834", |
| rows=C_base, cols=C_in*3*3) |
| gamma_bn1 = matrix(1, rows=C_base, cols=1) |
| beta_bn1 = matrix(0, rows=C_base, cols=1) |
| W_conv2 = matrix(" 0.10092591 0.15790914 -0.17673795 0.10573213 0.13680543 -0.22161855 -0.10239416 0.10747905 0.03636803 -0.00693908 -0.19976966 -0.15770042 -0.23468268 0.1040463 0.08357517 0.0780759 0.21764557 -0.11318331 |
| 0.1958775 0.00366694 0.05713235 -0.0768708 -0.04275537 -0.23076743 -0.029018 -0.02308315 -0.05915356 -0.12383241 0.16292028 0.20669906 -0.19045494 0.10580237 0.21305619 0.19072767 -0.19292024 0.15425198", |
| rows=C_base, cols=C_base*3*3) |
| gamma_bn2 = matrix(1, rows=C_base, cols=1) |
| beta_bn2 = matrix(0, rows=C_base, cols=1) |
| weights = list(W_conv1, gamma_bn1, beta_bn1, W_conv2, gamma_bn2, beta_bn2) |
| mode = "train" |
| |
| ema_mean_bn1 = matrix(0, rows=C_base, cols=1) |
| ema_var_bn1 = matrix(0, rows=C_base, cols=1) |
| ema_mean_bn2 = matrix(0, rows=C_base, cols=1) |
| ema_var_bn2 = matrix(0, rows=C_base, cols=1) |
| ema_means_vars = list(ema_mean_bn1, ema_var_bn1, ema_mean_bn2, ema_var_bn2) |
| |
| [out, Hout, Wout, ema_means_vars_up, cached_out, cached_means_vars] = resnet::basic_block_forward(X, weights, C_in, |
| C_base, Hin, Win, strideh, stridew, mode, ema_means_vars) |
| |
| out_expected = matrix("1.1192019 0.13680267 0.3965159 0.09488004 1.5221035 0.60268176 0.870202 0.3568437 0.19053036 2.023053 2.810772 2.7142973 1.4163418 2.2421117 0.22204155 0. 0.31268513 0.14521259 0.57843214 0.65275586 0. 0.02211368 0.4826215 0.65296173 1.2726448 0.6964331 1.6637247 1.2155424 2.3015604 3.9708042 1.6967016 0.40145046 |
| 1.1192019 0.13680267 0.3965159 0.09488004 1.5221035 0.60268176 0.870202 0.3568437 0.19053036 2.023053 2.810772 2.7142973 1.4163418 2.2421117 0.22204155 0. 0.31268513 0.14521259 0.57843214 0.65275586 0. 0.02211368 0.4826215 0.65296173 1.2726448 0.6964331 1.6637247 1.2155424 2.3015604 3.9708042 1.6967016 0.40145046", |
| rows=N, cols=Hout*Wout*C_base) |
| |
| test_util::check_all_close(out, out_expected, 0.00001) |
| } |
| |
| |
| values_test_basic_block_forward_2 = function() { |
| /* |
| * Testing of values for forward pass of basic block against PyTorch. |
| */ |
| strideh = 2; stridew = 2 |
| C_in = 2; C_base = 2 |
| Hin = 4; Win = 4 |
| N = 2 |
| X = matrix(1, rows=N, cols=C_in*Hin*Win) |
| W_conv1 = matrix("-0.14026615 -0.06974511 0.21421503 0.00487083 -0.17600328 -0.05576494 0.08433063 -0.04809754 -0.0021321 -0.1935787 -0.04766957 0.15073563 0.14598249 -0.1946578 -0.01819092 -0.11103764 -0.01316494 -0.14351277 |
| -0.0036971 -0.18704589 -0.09860466 0.20417325 -0.20006022 0.00959031 0.13883735 -0.11765242 -0.17820978 -0.03428984 -0.02357996 0.11326601 -0.22515622 0.2001556 -0.0103206 -0.0384565 0.13819869 -0.03230184", |
| rows=C_base, cols=C_in*3*3) |
| gamma_bn1 = matrix(1, rows=C_base, cols=1) |
| beta_bn1 = matrix(0, rows=C_base, cols=1) |
| W_conv2 = matrix(" 0.1952378 -0.13218941 0.20359151 0.21437167 0.20657437 0.07917522 -0.20072569 -0.16550082 0.14789648 0.03155191 0.10938872 -0.18765432 0.2069266 -0.0324703 0.14553984 -0.15199026 -0.01177226 0.05884366 |
| -0.16591048 -0.11745082 0.11246873 0.21435808 0.000237 -0.02330555 0.03408287 -0.09445126 0.09905426 -0.022421 -0.01720028 -0.08738072 -0.13018131 0.2277623 -0.22259445 0.06712689 -0.08571149 0.17849205", |
| rows=C_base, cols=C_base*3*3) |
| gamma_bn2 = matrix(1, rows=C_base, cols=1) |
| beta_bn2 = matrix(0, rows=C_base, cols=1) |
| |
| # downsampling weights |
| W_conv3 = matrix("-0.44065592 -0.29570574 |
| -0.60239863 0.43495506", |
| rows=C_base, cols=C_in*1*1) |
| gamma_bn3 = matrix(1, rows=C_base, cols=1) |
| beta_bn3 = matrix(0, rows=C_base, cols=1) |
| |
| weights = list(W_conv1, gamma_bn1, beta_bn1, W_conv2, gamma_bn2, beta_bn2, W_conv3, gamma_bn3, beta_bn3) |
| mode = "train" |
| |
| ema_mean_bn1 = matrix(0, rows=C_base, cols=1) |
| ema_var_bn1 = matrix(0, rows=C_base, cols=1) |
| ema_mean_bn2 = matrix(0, rows=C_base, cols=1) |
| ema_var_bn2 = matrix(0, rows=C_base, cols=1) |
| ema_mean_bn3 = matrix(0, rows=C_base, cols=1) |
| ema_var_bn3 = matrix(0, rows=C_base, cols=1) |
| ema_means_vars = list(ema_mean_bn1, ema_var_bn1, ema_mean_bn2, ema_var_bn2, ema_mean_bn3, ema_var_bn3) |
| |
| [out, Hout, Wout, ema_means_vars_up, cached_out, cached_means_vars] = resnet::basic_block_forward(X, weights, C_in, |
| C_base, Hin, Win, strideh, stridew, mode, ema_means_vars) |
| |
| out_expected = matrix("0. 0. 0.33147347 1.4695541 0. 0.9726007 0. 0.9382379 |
| 0. 0. 0.33147347 1.4695541 0. 0.9726007 0. 0.9382379 ", |
| rows=N, cols=Hout*Wout*C_base) |
| |
| test_util::check_all_close(out, out_expected, 0.0001) |
| } |
| |
| |
| values_test_basic_block_forward_3 = function() { |
| /* |
| * Testing of values for forward pass of basic block against PyTorch. |
| */ |
| strideh = 2; stridew = 2 |
| C_in = 2; C_base = 4 |
| Hin = 4; Win = 4 |
| N = 2 |
| X = matrix(1, rows=N, cols=C_in*Hin*Win) |
| W_conv1 = matrix(" 0.23060845 0.01255383 0.10554366 -0.11032246 0.04110865 0.12300454 0.03407042 0.03677405 0.1022801 0.08667193 0.15104999 -0.08343385 -0.16137402 0.2004693 -0.15173802 -0.14732602 -0.16977917 -0.11108103 |
| -0.02234201 -0.13497168 -0.15396744 -0.11581142 0.19164546 -0.02277191 0.1283987 -0.06767575 -0.05453977 -0.13944787 -0.16732863 0.08283444 -0.20333174 0.15993242 0.09864204 0.02714513 -0.05416261 -0.16831529 |
| -0.02864511 0.17540906 0.10217993 0.16238032 -0.09703699 0.13528688 -0.12437965 0.22771217 0.13020585 0.06029092 0.03008728 0.08683048 0.18321039 -0.0570219 -0.04176761 -0.10389957 -0.21008791 0.21670572 |
| -0.17345327 0.05192728 -0.137008 -0.16411738 0.2257642 0.1580276 -0.17054762 -0.15276143 0.18090443 -0.00081959 -0.20020181 -0.15055852 -0.08635178 0.11620347 -0.17689864 -0.17850053 0.14149593 0.05391319", |
| rows=C_base, cols=C_in*3*3) |
| gamma_bn1 = matrix(1, rows=C_base, cols=1) |
| beta_bn1 = matrix(0, rows=C_base, cols=1) |
| W_conv2 = matrix("-0.12502055 0.04218115 -0.01587057 0.02996665 0.03902064 -0.10042268 0.14614715 0.04653394 -0.16000232 0.16580684 -0.00197132 -0.12345098 0.16189565 -0.08695424 -0.02277309 -0.02887423 -0.06450109 -0.02360182 -0.11778401 -0.0124789 -0.14320037 -0.1436774 -0.00914766 -0.01130253 0.13241099 0.03841829 0.14280184 0.07521738 0.11815999 -0.13276237 0.12992252 -0.04222127 -0.03073458 0.01725562 0.15965255 0.10338821 |
| -0.1398921 0.05605571 -0.13995157 -0.12097194 -0.1035642 0.03313832 0.13476823 0.02719207 0.02726786 -0.08288203 0.10799147 0.03092675 -0.01539116 -0.08172278 -0.05077231 0.03913508 0.02528121 0.08431648 0.16408543 -0.09090649 -0.09221806 0.00649713 0.08248532 0.05170746 -0.03424133 0.12494816 -0.03637959 0.01817816 0.10356762 0.03744942 -0.10864812 0.10180093 -0.04949838 -0.10033202 -0.10501622 -0.05735092 |
| -0.05820473 -0.11734504 0.16419913 -0.05231454 -0.07497393 0.1414146 -0.07572757 0.12433673 0.11995722 -0.08965874 0.01813734 0.07857008 -0.01808423 -0.10376819 0.02973495 -0.06675623 0.12945338 0.11593701 -0.0270998 0.06052397 -0.09865837 0.05997723 -0.09147376 -0.13572678 0.04277535 0.06700458 0.10086431 0.0624107 0.01380444 0.02379382 -0.06924826 0.16204901 0.09410296 0.08837719 0.08246924 -0.02000479 |
| -0.05427806 -0.06499916 -0.07410125 -0.09293389 -0.05310518 -0.14569238 0.0565486 0.06677081 0.11938848 0.140608 0.08632036 0.06824701 0.08058478 0.1592765 0.05660491 0.000421 0.08094937 0.02867652 0.00291246 -0.10608425 0.03485543 -0.1534006 -0.05454665 -0.12841377 0.15747286 0.15352447 0.00212862 -0.06103357 -0.05826634 -0.12362739 0.05188899 0.0780222 -0.08359687 -0.07310607 -0.04413005 0.16476734", |
| rows=C_base, cols=C_base*3*3) |
| gamma_bn2 = matrix(1, rows=C_base, cols=1) |
| beta_bn2 = matrix(0, rows=C_base, cols=1) |
| |
| # downsampling weights |
| W_conv3 = matrix("-0.434205 -0.51763165 |
| -0.48000953 -0.4031318 |
| 0.38555235 0.09546977 |
| -0.35429037 -0.35176745", |
| rows=C_base, cols=C_in*1*1) |
| gamma_bn3 = matrix(1, rows=C_base, cols=1) |
| beta_bn3 = matrix(0, rows=C_base, cols=1) |
| |
| weights = list(W_conv1, gamma_bn1, beta_bn1, W_conv2, gamma_bn2, beta_bn2, W_conv3, gamma_bn3, beta_bn3) |
| mode = "train" |
| |
| ema_mean_bn1 = matrix(0, rows=C_base, cols=1) |
| ema_var_bn1 = matrix(0, rows=C_base, cols=1) |
| ema_mean_bn2 = matrix(0, rows=C_base, cols=1) |
| ema_var_bn2 = matrix(0, rows=C_base, cols=1) |
| ema_mean_bn3 = matrix(0, rows=C_base, cols=1) |
| ema_var_bn3 = matrix(0, rows=C_base, cols=1) |
| ema_means_vars = list(ema_mean_bn1, ema_var_bn1, ema_mean_bn2, ema_var_bn2, ema_mean_bn3, ema_var_bn3) |
| |
| [out, Hout, Wout, ema_means_vars_up, cached_out, cached_means_vars] = resnet::basic_block_forward(X, weights, C_in, |
| C_base, Hin, Win, strideh, stridew, mode, ema_means_vars) |
| |
| out_expected = matrix("0. 1.576729 0. 0.14989257 0. 0. 1.5978024 0. 1.6740767 0. 0. 0. 0.89626473 1.0869961 0. 0. |
| 0. 1.576729 0. 0.14989257 0. 0. 1.5978024 0. 1.6740767 0. 0. 0. 0.89626473 1.0869961 0. 0. ", |
| rows=N, cols=Hout*Wout*C_base) |
| |
| test_util::check_all_close(out, out_expected, 0.0001) |
| } |
| |
| values_test_residual_layer_forward = function() { |
| Hin = 4; Win = 4; |
| C_in = 2; C_base = 4; |
| N = 2 |
| blocks = 2 |
| strideh = 2; stridew = 2 |
| mode = "train" |
| |
| X = matrix(seq(1, N*Hin*Win*C_in), rows=N, cols=C_in*Hin*Win) |
| |
| # weights for block 1 |
| W_conv1 = matrix(" 0.18088163 -0.1094396 0.16322954 |
| -0.16315436 -0.21879527 -0.13885644 |
| -0.17897591 -0.03466086 0.00447898 |
| |
| -0.02060366 0.07391576 -0.13932472 |
| 0.19118743 -0.1424667 0.13759573 |
| 0.01909433 -0.09354833 -0.04176301 |
| |
| |
| 0.1272765 0.03356279 0.09614499 |
| -0.19478372 0.00336599 -0.11832798 |
| -0.167036 -0.21355109 0.09034546 |
| |
| 0.08218841 0.09808768 -0.13677552 |
| 0.09094055 -0.2075183 -0.11292712 |
| 0.01358929 0.06142236 0.15713598 |
| |
| |
| -0.07507452 0.17012559 -0.00778644 |
| -0.17894624 -0.16280204 -0.20480031 |
| 0.17800058 -0.08244179 0.21474986 |
| |
| 0.13779633 -0.09157459 -0.16890329 |
| -0.18981671 0.01013687 0.19278602 |
| 0.12158521 0.14861913 0.16235258 |
| |
| |
| -0.06423138 -0.08338779 0.03925912 |
| 0.08258285 -0.10299681 -0.02878706 |
| 0.09201117 0.1826442 0.04269032 |
| |
| 0.23345782 -0.19535708 -0.04301685 |
| -0.20099604 0.13898961 -0.13627604 |
| -0.02552247 0.19012617 0.11112581", rows=C_base, cols=C_in*3*3) |
| gamma_bn1 = matrix(1, rows=C_base, cols=1) |
| beta_bn1 = matrix(0, rows=C_base, cols=1) |
| W_conv2 = matrix("-0.05845656 -0.10387342 -0.00347954 |
| -0.10443704 0.07762128 0.08962621 |
| -0.15254496 -0.11626967 -0.07093517 |
| |
| 0.04007399 0.08974029 -0.08184759 |
| -0.03026156 -0.12850255 0.11625351 |
| -0.06768056 0.04290299 -0.11035781 |
| |
| -0.08182923 -0.07303425 -0.14692405 |
| 0.07654093 0.01254661 -0.08789502 |
| 0.12506868 -0.03339644 0.09265955 |
| |
| -0.0010149 0.06057359 0.13991983 |
| 0.07688661 0.0267771 0.0457534 |
| -0.12409463 0.16171788 0.04298592 |
| |
| |
| 0.04832537 -0.03725861 -0.06689632 |
| -0.06220818 0.14289887 -0.15294018 |
| 0.00162084 0.01681285 -0.16070186 |
| |
| -0.06003825 0.04543029 -0.10140823 |
| 0.12588196 0.0608999 0.12295659 |
| -0.06011419 -0.08439511 0.14588566 |
| |
| -0.06961578 0.03729546 0.10786648 |
| -0.04095362 -0.02719462 0.01085408 |
| 0.03948607 0.08386 -0.03532495 |
| |
| 0.00769222 0.04593278 -0.09789048 |
| -0.03487056 0.05989157 -0.09724655 |
| -0.15379873 -0.07006194 0.14572401 |
| |
| |
| -0.15647712 0.08853446 -0.0191143 |
| -0.15235935 0.09655853 0.02425753 |
| -0.03920817 -0.13884489 -0.15873407 |
| |
| -0.16632351 -0.15446958 -0.02832182 |
| 0.15606628 -0.07041912 0.09881757 |
| -0.07226615 0.07085086 -0.03745939 |
| |
| -0.0293621 -0.14463028 -0.15513223 |
| 0.10102965 -0.11223143 0.16495053 |
| 0.00031853 0.158157 -0.13941693 |
| |
| 0.10985805 0.11699991 -0.06803058 |
| -0.08846518 0.13454668 -0.07047473 |
| 0.0816289 -0.03807278 -0.01125084 |
| |
| |
| -0.1425793 0.04159175 0.0873092 |
| -0.07806729 0.06501202 0.09965105 |
| 0.06275028 0.07400332 0.09444918 |
| |
| -0.08856728 -0.09136113 -0.07333919 |
| 0.04255192 -0.10251606 0.15050472 |
| 0.07791793 0.10539879 -0.06219628 |
| |
| 0.12434496 0.16624264 0.08779152 |
| 0.00117178 0.13169001 -0.04333049 |
| -0.07304269 0.14722325 -0.06679092 |
| |
| -0.08179037 0.15500171 -0.00718816 |
| -0.1278541 -0.08474605 -0.10129128 |
| 0.07862437 -0.06843086 -0.04310509", rows=C_base, cols=C_base*3*3) |
| gamma_bn2 = matrix(1, rows=C_base, cols=1) |
| beta_bn2 = matrix(0, rows=C_base, cols=1) |
| # downsample parameters |
| W_conv3 = matrix("-0.41760927 |
| |
| -0.0473721 |
| |
| |
| -0.14902914 |
| |
| -0.3806823 |
| |
| |
| 0.04111391 |
| |
| 0.6815012 |
| |
| |
| 0.16695487 |
| |
| 0.5603499 ", rows=C_base, cols=C_in*1*1) |
| gamma_bn3 = matrix(1, rows=C_base, cols=1) |
| beta_bn3 = matrix(0, rows=C_base, cols=1) |
| block1_weights = list(W_conv1, gamma_bn1, beta_bn1, W_conv2, gamma_bn2, beta_bn2, W_conv3, gamma_bn3, beta_bn3) |
| |
| # EMAS for block 1 |
| ema_mean_bn1 = matrix(1, rows=C_base, cols=1) |
| ema_var_bn1 = matrix(0, rows=C_base, cols=1) |
| ema_mean_bn2 = matrix(1, rows=C_base, cols=1) |
| ema_var_bn2 = matrix(0, rows=C_base, cols=1) |
| ema_mean_bn3 = matrix(1, rows=C_base, cols=1) |
| ema_var_bn3 = matrix(0, rows=C_base, cols=1) |
| block1_EMAs = list(ema_mean_bn1, ema_var_bn1, ema_mean_bn2, ema_var_bn2, ema_mean_bn3, ema_var_bn3) |
| |
| # Weights for block 2 |
| W_conv1_2 = matrix("-7.28789419e-02 7.32977241e-02 -1.16737187e-01 |
| -1.09122857e-01 1.51534528e-02 -1.55087024e-01 |
| 1.54908732e-01 -1.46781862e-01 5.23662418e-02 |
| |
| -1.45408154e-01 9.34596211e-02 -1.57577261e-01 |
| -1.62250042e-01 -2.06526369e-02 1.58289358e-01 |
| -1.38804317e-04 -1.37716874e-01 -1.25059336e-01 |
| |
| 5.44795990e-02 -1.56691819e-02 4.58848923e-02 |
| -1.66475177e-01 -8.56291652e-02 9.89388674e-02 |
| -3.30540538e-02 -4.98571396e-02 -1.48154795e-03 |
| |
| -1.63741872e-01 7.52360672e-02 -7.26198778e-02 |
| 7.97830820e-02 -1.53044373e-01 1.05956629e-01 |
| 1.42732725e-01 -4.19619307e-02 3.91238928e-03 |
| |
| |
| -3.11080813e-02 6.24701530e-02 -8.18871707e-02 |
| -2.80226916e-02 5.65987229e-02 -5.63029051e-02 |
| 9.20275897e-02 -1.50979385e-01 3.14275920e-03 |
| |
| 4.63068485e-02 -7.43940473e-02 -1.23582803e-01 |
| -4.84378934e-02 -1.62422940e-01 2.17949301e-02 |
| 1.48192182e-01 -9.01084542e-02 9.67378765e-02 |
| |
| 1.82208419e-02 1.48985460e-01 -1.47735506e-01 |
| 8.09304416e-03 1.12461001e-02 2.95447111e-02 |
| -1.24866471e-01 3.81960124e-02 -6.68919683e-02 |
| |
| 5.13156503e-02 6.83855265e-02 -4.90674824e-02 |
| -9.68660563e-02 3.40797305e-02 -1.13457203e-01 |
| -2.67352313e-02 -8.27323049e-02 -7.86665529e-02 |
| |
| |
| -1.18089557e-01 1.66068569e-01 4.50132638e-02 |
| 4.65527624e-02 1.10370263e-01 1.24886349e-01 |
| -9.42516923e-02 -1.62573516e-01 7.66497254e-02 |
| |
| 1.08407423e-01 -4.26756591e-02 -1.11639105e-01 |
| -8.21658969e-02 2.47098356e-02 -7.98595399e-02 |
| -1.54958516e-02 2.23536491e-02 -1.03785992e-02 |
| |
| -1.53956562e-02 1.13292173e-01 -2.27067471e-02 |
| -2.12994069e-02 8.41291696e-02 1.61149070e-01 |
| -1.32289156e-01 -7.05852732e-02 7.90221095e-02 |
| |
| -1.43424913e-01 -1.21421874e-01 -1.27822340e-01 |
| -2.88078189e-02 -5.81898317e-02 -1.99964195e-02 |
| 1.16435274e-01 -1.30379200e-03 -4.03594971e-02 |
| |
| |
| -1.00988328e-01 6.64077997e-02 -1.31890640e-01 |
| -1.35123342e-01 -1.37298390e-01 8.09081197e-02 |
| 3.08579355e-02 7.35761523e-02 1.45316467e-01 |
| |
| 9.22436267e-02 3.85234505e-02 3.37007642e-02 |
| 4.96874899e-02 2.56118029e-02 -2.25261599e-02 |
| 6.15442246e-02 7.31151104e-02 1.47016749e-01 |
| |
| -7.32975453e-02 -6.77637309e-02 2.41905302e-02 |
| -1.47778392e-01 -1.27045453e-01 -3.37420404e-03 |
| -1.08250901e-01 -8.57535824e-02 -7.87658989e-03 |
| |
| -3.25922221e-02 -8.58756527e-02 7.44262338e-02 |
| -1.37389064e-01 7.38748461e-02 -7.68387318e-02 |
| 1.25040159e-01 -6.90028891e-02 -5.00665307e-02", rows=C_base, cols=C_base*3*3) |
| gamma_bn1_2 = matrix(1, rows=C_base, cols=1) |
| beta_bn1_2 = matrix(0, rows=C_base, cols=1) |
| W_conv2_2 = matrix("-0.02658759 0.02687277 -0.11249679 |
| 0.07110029 -0.1287723 0.14960568 |
| 0.00347126 0.082368 0.1592095 |
| |
| 0.01121612 -0.01054858 -0.11533582 |
| 0.00191922 -0.09345891 0.0201468 |
| 0.16406216 0.1631505 -0.12568823 |
| |
| 0.10923527 0.0047278 0.08264925 |
| 0.03556542 -0.11967081 -0.03904144 |
| 0.06578751 -0.11364846 -0.06719196 |
| |
| -0.02243698 -0.15126523 -0.16504839 |
| 0.03257662 0.03516276 -0.00604182 |
| 0.00565775 0.15348013 -0.08239889 |
| |
| |
| -0.10838002 0.03725785 -0.01738496 |
| 0.01061849 0.00593401 -0.14623034 |
| -0.04515064 0.11624385 0.07984154 |
| |
| 0.02820455 0.14632984 0.11314054 |
| 0.1556999 0.06233911 0.04464059 |
| 0.16220112 0.12387766 -0.01516332 |
| |
| 0.11520021 0.03099509 0.11585347 |
| 0.0215022 0.09577711 0.12590684 |
| 0.12393482 -0.00796187 -0.1233204 |
| |
| -0.01217443 -0.09484772 -0.13615403 |
| -0.06195782 -0.10316825 -0.13738668 |
| 0.04165971 -0.16430686 0.1249779 |
| |
| |
| -0.06067413 0.09290363 -0.10419172 |
| 0.04424816 0.16639026 -0.01638204 |
| -0.01993459 0.16510008 -0.03844319 |
| |
| -0.06738343 -0.15954465 0.14164312 |
| -0.09711097 0.04057109 -0.06419432 |
| -0.15190187 0.02492356 0.14873762 |
| |
| 0.05357671 -0.02110486 -0.07781315 |
| -0.12230659 0.13541014 0.04158884 |
| 0.13525964 0.07432733 0.04886186 |
| |
| -0.04131328 0.05893086 0.08948417 |
| 0.15411825 0.05368501 -0.13857502 |
| 0.15523924 -0.1510986 0.01617928 |
| |
| |
| 0.14878072 0.15607525 -0.12842798 |
| 0.00907773 0.09931238 0.03955895 |
| 0.04165536 -0.0382842 0.06571688 |
| |
| -0.0926128 -0.15800306 -0.0235941 |
| 0.03582941 -0.13953064 0.03686035 |
| -0.15508795 0.06028162 -0.15286762 |
| |
| 0.00642897 -0.01605938 -0.10140433 |
| 0.09824272 0.13854371 0.13406266 |
| 0.13023908 -0.10159403 -0.08961493 |
| |
| -0.0350284 0.08208416 0.11221837 |
| -0.07019123 0.00895458 0.10123546 |
| -0.04459848 0.15377314 0.04990514", rows=C_base, cols=C_base*3*3) |
| gamma_bn2_2 = matrix(1, rows=C_base, cols=1) |
| beta_bn2_2 = matrix(0, rows=C_base, cols=1) |
| block2_weights = list(W_conv1_2, gamma_bn1_2, beta_bn1_2, W_conv2_2, gamma_bn2_2, beta_bn2_2) |
| |
| # EMAS for block 1 |
| ema_mean_bn1_2 = matrix(1, rows=C_base, cols=1) |
| ema_var_bn1_2 = matrix(0, rows=C_base, cols=1) |
| ema_mean_bn2_2 = matrix(1, rows=C_base, cols=1) |
| ema_var_bn2_2 = matrix(0, rows=C_base, cols=1) |
| block2_EMAs = list(ema_mean_bn1_2, ema_var_bn1_2, ema_mean_bn2_2, ema_var_bn2_2) |
| |
| expected_Hout = 2 |
| expected_Wout = 2 |
| expected_out = matrix("1.9154322 1.3386528 |
| 0. 0.17239538 |
| |
| 1.1703718 1.765771 |
| 0. 0.48162544 |
| |
| 0. 0. |
| 0.5448834 0. |
| |
| 0.5596865 0. |
| 1.7399819 0. |
| |
| |
| 0. 2.4118602 |
| 0. 0.32017422 |
| |
| 0.45649305 1.0522902 |
| 1.0418018 1.2802167 |
| |
| 3.914177 0.7292799 |
| 0. 1.287521 |
| |
| 0. 0.71220785 |
| 2.7913036 0.6481694 ", rows=2, cols=C_base*expected_Hout*expected_Wout) |
| |
| blocks_weights = list(block1_weights, block2_weights) |
| ema_means_vars = list(block1_EMAs, block2_EMAs) |
| block_type = "basic" |
| [out, Hout, Wout, ema_means_vars_upd, cached_out, cached_means_vars] = resnet::reslayer_forward(X, Hin, Win, |
| block_type, blocks, strideh, stridew, C_in, C_base, blocks_weights, mode, ema_means_vars) |
| |
| test_util::check_all_close(out, expected_out, 0.0001) |
| } |
| |
| values_test_basic_block_backward = function() { |
| /* |
| * Testing of values for forward pass of basic block against PyTorch. |
| */ |
| strideh = 2; stridew = 2 |
| C_in = 2; C_base = 4 |
| Hin = 4; Win = 4 |
| N = 2 |
| X = matrix(seq(1, N*C_in*Hin*Win), rows=N, cols=C_in*Hin*Win) |
| W_conv1 = matrix("0.22795902 0.17547427 0.21921514 |
| 0.13373677 0.1449876 -0.20703636 |
| 0.1049294 -0.07758546 0.01650636 |
| |
| 0.09977336 0.14042483 -0.10443689 |
| 0.18606941 0.07984604 0.1862065 |
| -0.0245543 -0.00999866 -0.06876408 |
| |
| |
| -0.21368428 0.11649923 0.1329887 |
| -0.03319308 -0.18250495 0.16087545 |
| -0.0010442 -0.06631051 0.11644398 |
| |
| -0.19472131 -0.0089076 -0.00361156 |
| 0.11273773 -0.04975031 -0.22203343 |
| 0.12707166 0.05724771 -0.19370714 |
| |
| |
| -0.03382689 -0.2220618 0.09851243 |
| -0.04797092 -0.15873244 0.10475938 |
| -0.14053479 0.14056166 0.16246034 |
| |
| -0.22719866 -0.18615595 0.13337635 |
| -0.18140101 0.05603398 0.11490373 |
| 0.0046075 -0.11261812 0.01054354 |
| |
| |
| 0.06122501 0.18338178 0.21308573 |
| 0.00820886 0.01341644 -0.08846006 |
| -0.04187146 -0.05795708 0.07099648 |
| |
| 0.0377063 -0.19591397 0.08679475 |
| 0.05578776 -0.01468489 -0.22213465 |
| 0.21319999 -0.19887432 0.00093569 |
| ", rows=C_base, cols=C_in*3*3) |
| gamma_bn1 = matrix(1, rows=C_base, cols=1) |
| beta_bn1 = matrix(0, rows=C_base, cols=1) |
| W_conv2 = matrix("-0.15316814 0.01649837 0.02066648 |
| 0.15951438 0.14952205 0.05908944 |
| 0.12360035 0.16258992 -0.05275005 |
| |
| -0.09005231 0.10635151 -0.06287201 |
| 0.0140198 0.00389086 -0.13122693 |
| -0.01728405 0.06940794 0.1438715 |
| |
| -0.06239128 -0.04787713 0.12102707 |
| -0.09782796 -0.12092899 0.15389048 |
| 0.13240601 -0.07352956 -0.04799255 |
| |
| -0.04824746 0.15890409 -0.09615827 |
| 0.00646619 0.07214041 0.05696104 |
| 0.09033443 -0.15197693 -0.1104077 |
| |
| |
| 0.12790756 -0.14982054 -0.08584315 |
| 0.1374556 0.06672513 0.05649196 |
| -0.15129058 0.07952888 0.07835837 |
| |
| 0.05553712 -0.02423525 0.12602611 |
| -0.03570884 0.07482989 -0.07258789 |
| -0.15282819 0.09200995 0.00603025 |
| |
| -0.05926096 0.0686207 0.11548664 |
| -0.08733666 -0.14680144 -0.11286473 |
| 0.01421845 -0.04764497 0.15172698 |
| |
| -0.13044584 0.04871863 -0.14850675 |
| -0.13645135 -0.11802016 -0.01737686 |
| 0.10504811 -0.13870925 -0.07346682 |
| |
| |
| -0.0705896 -0.09258682 0.1341825 |
| 0.05714887 -0.06405389 -0.14810124 |
| 0.06753898 0.00227241 0.06100388 |
| |
| 0.09614421 0.12490447 -0.09217326 |
| -0.04312347 0.08090664 0.12942137 |
| 0.04428639 0.0577697 0.12196468 |
| |
| 0.10778381 -0.10680336 0.11534978 |
| -0.0510295 -0.01859938 0.01497144 |
| -0.03191304 0.06286548 0.11793233 |
| |
| -0.12459338 0.09700234 -0.11604425 |
| 0.09004898 -0.02339827 0.15153421 |
| 0.06829721 -0.01891816 -0.0645818 |
| |
| |
| 0.03223832 0.10086165 -0.13695073 |
| -0.01531534 0.07804878 -0.00895005 |
| -0.03381938 -0.10908765 0.03976715 |
| |
| -0.01888643 0.08189081 -0.01315074 |
| -0.05945349 -0.03236644 -0.07914945 |
| -0.06206207 -0.01699001 0.0371331 |
| |
| 0.04449199 0.05730323 0.05018614 |
| -0.16624898 -0.00204407 0.05084966 |
| 0.1588928 -0.00098683 -0.08426523 |
| |
| -0.04337545 0.06106292 -0.00310864 |
| -0.05939015 -0.02901964 -0.12184718 |
| -0.09405826 0.11087091 -0.10842602 |
| ", rows=C_base, cols=C_base*3*3) |
| gamma_bn2 = matrix(1, rows=C_base, cols=1) |
| beta_bn2 = matrix(0, rows=C_base, cols=1) |
| W_conv_down = matrix("0.31054777 |
| |
| 0.3262263 |
| |
| |
| 0.46358436 |
| |
| -0.51710904 |
| |
| |
| 0.18096805 |
| |
| 0.32488567 |
| |
| |
| -0.29951102 |
| |
| -0.40844375 |
| ", rows=C_base, cols=C_in) |
| gamma_bn3 = matrix(1, rows=C_base, cols=1) |
| beta_bn3 = matrix(0, rows=C_base, cols=1) |
| |
| weights = list(W_conv1, gamma_bn1, beta_bn1, W_conv2, gamma_bn2, beta_bn2, W_conv_down, gamma_bn3, beta_bn3) |
| mode = "train" |
| |
| ema_mean_bn1 = matrix(0, rows=C_base, cols=1) |
| ema_var_bn1 = matrix(0, rows=C_base, cols=1) |
| ema_mean_bn2 = matrix(0, rows=C_base, cols=1) |
| ema_var_bn2 = matrix(0, rows=C_base, cols=1) |
| ema_mean_bn3 = matrix(0, rows=C_base, cols=1) |
| ema_var_bn3 = matrix(0, rows=C_base, cols=1) |
| ema_means_vars = list(ema_mean_bn1, ema_var_bn1, ema_mean_bn2, ema_var_bn2, ema_mean_bn3, ema_var_bn3) |
| |
| [out, Hout, Wout, ema_means_vars_up, cached_out, cached_means_vars] = resnet::basic_block_forward(X, weights, C_in, |
| C_base, Hin, Win, strideh, stridew, mode, ema_means_vars) |
| |
| dOut = matrix("-0.0625 -0.0625 |
| -0.0625 -0.05511813 |
| |
| -0.03883262 -0.05729559 |
| 0.04194808 -0.03308601 |
| |
| -0.04001789 -0.0625 |
| -0.0625 -0.03310877 |
| |
| -0.0625 0.02698019 |
| -0.00352125 0.04049902 |
| |
| |
| -0.0625 0.04098879 |
| 0.02273786 0.10461447 |
| |
| -0.0625 -0.0625 |
| 0.00151762 -0.0625 |
| |
| -0.00950899 -0.05501665 |
| -0.0625 -0.0141529 |
| |
| -0.0625 -0.0625 |
| -0.0625 -0.03729285", |
| rows=N, cols=C_base*Hout*Wout) |
| |
| [dX, gradients] = resnet::basic_block_backward(dOut, cached_out, weights, C_in, C_base, |
| strideh, stridew, cached_means_vars) |
| |
| dW_conv1_exp = matrix("-0.02246146 0.10423277 0.1029001 |
| -0.05089705 0.10958684 0.10958683 |
| -0.07130328 0.10958684 0.10958684 |
| |
| -0.10713803 0.08291006 0.08157738 |
| -0.13252197 0.10958683 0.10958685 |
| -0.1529282 0.10958683 0.10958683 |
| |
| |
| -0.02584631 0.06710234 0.06838104 |
| -0.34364033 -0.17614923 -0.17614923 |
| -0.3919655 -0.17614925 -0.17614923 |
| |
| -0.1790472 0.0875615 0.08884021 |
| -0.53694105 -0.17614923 -0.17614923 |
| -0.58526623 -0.17614922 -0.17614925 |
| |
| |
| -0.13346207 -0.03711003 -0.04283796 |
| -0.17548679 -0.3563866 -0.35638663 |
| -0.1589087 -0.3563866 -0.35638663 |
| |
| -0.21392906 -0.12875691 -0.13448483 |
| -0.1091744 -0.35638666 -0.35638663 |
| -0.09259631 -0.3563867 -0.3563866 |
| |
| |
| -0.10748374 -0.15856032 -0.15569028 |
| -0.15130743 -0.26214042 -0.2621404 |
| -0.18142992 -0.2621404 -0.26214042 |
| |
| -0.01991791 -0.1126398 -0.10976979 |
| -0.27179736 -0.26214045 -0.26214042 |
| -0.30191985 -0.26214045 -0.2621404 ", |
| rows=C_base, cols=C_in*3*3) |
| test_util::check_all_close(as.matrix(gradients[1]), dW_conv1_exp, 0.0001) |
| |
| dW_bn1_exp = matrix(" 0.11060466 -0.02107318 -0.0812627 -0.00814073", rows=C_base, cols=1) |
| test_util::check_all_close(as.matrix(gradients[2]), dW_bn1_exp, 0.0001) |
| |
| db_bn1_exp = matrix(" 0.01407476 -0.07577097 0.09522599 0.07362167", rows=C_base, cols=1) |
| test_util::check_all_close(as.matrix(gradients[3]), db_bn1_exp, 0.0001) |
| |
| dW_conv2_exp = matrix(" 0.00000000e+00 6.27728775e-02 4.55777068e-03 |
| 7.02145845e-02 5.98357618e-01 6.22697920e-02 |
| 1.20493323e-02 1.22186027e-01 1.86003119e-01 |
| |
| -2.30462596e-01 -3.63258511e-01 -1.50781767e-02 |
| -1.57529697e-01 -2.15406895e-01 2.45092548e-02 |
| -1.97134074e-02 -2.55168434e-02 -7.79616646e-04 |
| |
| 2.15687096e-01 1.35737220e-02 -2.27542143e-04 |
| 1.13740548e-01 1.09955356e-01 -2.35504449e-05 |
| 1.39786359e-02 6.76771551e-02 0.00000000e+00 |
| |
| -9.22712237e-02 -2.82834589e-01 -1.38832042e-02 |
| -1.52194882e-02 2.60263085e-02 9.82659869e-03 |
| 0.00000000e+00 1.41863674e-02 1.05086461e-01 |
| |
| |
| 0.00000000e+00 5.18248565e-02 -6.42012283e-02 |
| 5.79686798e-02 4.33321059e-01 -5.52755892e-01 |
| 2.53020395e-02 2.54378200e-01 3.95552069e-01 |
| |
| -1.31259233e-01 8.39732587e-02 4.01174039e-01 |
| -2.59595156e-01 -3.81430775e-01 7.20100850e-02 |
| -9.93236750e-02 -1.69444025e-01 -5.55632524e-02 |
| |
| 2.05957323e-01 -2.13401914e-01 5.07378485e-03 |
| 1.25175878e-01 1.10367373e-01 -1.67843944e-03 |
| -3.87407765e-02 9.45824534e-02 0.00000000e+00 |
| |
| -5.25527745e-02 -5.58004305e-02 3.09570760e-01 |
| -7.66815916e-02 -1.37685180e-01 -1.52672842e-01 |
| 0.00000000e+00 -9.84215960e-02 1.30307078e-01 |
| |
| |
| 0.00000000e+00 1.21219526e-03 5.99100720e-03 |
| 1.35590055e-03 -3.04231290e-02 7.05613419e-02 |
| -5.36800846e-02 -4.34169441e-01 1.18018150e-01 |
| |
| -1.83993857e-02 2.74298918e-02 1.34550676e-01 |
| 5.64235747e-02 3.97945419e-02 4.81129810e-03 |
| 3.42092514e-02 2.82005444e-02 -1.89568941e-02 |
| |
| -2.42714863e-03 5.99674359e-02 1.22890016e-03 |
| -2.24024475e-01 1.11650631e-01 -5.72644640e-04 |
| -1.25298709e-01 2.60425620e-02 0.00000000e+00 |
| |
| -7.36663491e-03 2.65299832e-03 7.49798343e-02 |
| 2.64108200e-02 5.34383431e-02 6.85021579e-02 |
| 0.00000000e+00 -1.81864589e-01 3.47672515e-02 |
| |
| |
| 0.00000000e+00 -7.43892491e-02 -7.71372812e-03 |
| -8.32080767e-02 -7.09796607e-01 -8.08604732e-02 |
| -1.06588183e-02 -9.39640850e-02 -7.12726489e-02 |
| |
| 2.23416343e-01 2.80079544e-01 -8.82177129e-02 |
| 2.54184335e-01 3.47144812e-01 -1.16981082e-01 |
| 7.17329308e-02 6.99866414e-02 -2.60418989e-02 |
| |
| -2.79086232e-01 -5.31585366e-02 -7.40691903e-04 |
| -1.04264751e-01 -1.04167528e-01 -7.86666467e-04 |
| 5.14573716e-02 -5.00418246e-02 0.00000000e+00 |
| |
| 8.94500837e-02 2.54736274e-01 -4.51924093e-02 |
| 5.53805046e-02 8.24335217e-03 -1.32365167e-01 |
| 0.00000000e+00 1.07619762e-01 -8.57934728e-02", |
| rows=C_base, cols=C_base*3*3) |
| test_util::check_all_close(as.matrix(gradients[4]), dW_conv2_exp, 0.0001) |
| |
| dW_bn2_exp = matrix(" 0.1440372 0.14261805 -0.06086085 -0.01561027", rows=C_base, cols=1) |
| test_util::check_all_close(as.matrix(gradients[5]), dW_bn2_exp, 0.0001) |
| |
| db_bn2_exp = matrix(" 0.11322299 -0.08574851 -0.15180519 0.02666511", rows=C_base, cols=1) |
| test_util::check_all_close(as.matrix(gradients[6]), db_bn2_exp, 0.0001) |
| |
| dW_conv3_exp = matrix(" 8.9406967e-08 |
| |
| 1.1920929e-07 |
| |
| |
| 2.5063753e-05 |
| |
| 2.4765730e-05 |
| |
| |
| -1.1175871e-08 |
| |
| -2.9802322e-08 |
| |
| |
| -2.2351742e-08 |
| |
| -2.9802322e-08", rows=C_base, cols=C_in) |
| test_util::check_all_close(as.matrix(gradients[7]), dW_conv3_exp, 0.0001) |
| |
| dW_bn3_exp = matrix(" 0.22805437 -0.10600837 0.00529851 0.10261551", rows=C_base, cols=1) |
| test_util::check_all_close(as.matrix(gradients[8]), dW_bn3_exp, 0.0001) |
| |
| db_bn3_exp = matrix(" 0.11322299 -0.08574851 -0.15180519 0.02666511", rows=C_base, cols=1) |
| test_util::check_all_close(as.matrix(gradients[9]), db_bn3_exp, 0.0001) |
| } |
| |
| |
| /* |
| * **** Basic Block Shape Handling Testing **** |
| */ |
| |
| /* |
| * Test case 1: |
| * Basic block forward computation shouldn't raise errors |
| * when given valid inputs with valid shapes without having |
| * to downsample the input and the output should have the |
| * expected shape. |
| */ |
| basic_test_basic_block_forward(C_in=4, C_base=4, strideh=1, stridew=1, Hin=4, Win=4, N=3, downsample=FALSE) |
| |
| /* |
| * Test case 2: |
| * Basic block forward computation shouldn't raise errors |
| * when given valid inputs with having to downsample inputs |
| * because of non-matching channels and the output should |
| * have the expected (downsampled) shapes. |
| */ |
| basic_test_basic_block_forward(C_in=2, C_base=4, strideh=1, stridew=1, Hin=4, Win=4, N=3, downsample=TRUE) |
| |
| /* |
| * Test case 3: |
| * Basic block forward computation shouldn't raise errors |
| * when given valid inputs with having to downsample inputs |
| * because of stride bigger than 1 and the output should |
| * have the expected (downsampled) shapes. |
| */ |
| basic_test_basic_block_forward(C_in=4, C_base=4, strideh=2, stridew=2, Hin=4, Win=4, N=3, downsample=TRUE) |
| |
| /* |
| * **** Basic Block Value Testing **** |
| * In these test cases, we compare the forward pass |
| * computation of a basic residual block against the |
| * PyTorch implementation. We calculate the PyTorch |
| * values with the NN module |
| * torchvision.models.resnet.BasicBlock and then |
| * hard-code the randomly initialized weights and |
| * biases and the expected output computed by PyTorch |
| * into this file. |
| */ |
| |
| /* |
| * Test case 1: |
| * A simple forward pass of basic block with same |
| * input and output channels and the same input and |
| * output dimensions, i.e. stride 1, in train mode. |
| */ |
| values_test_basic_block_forward_1() |
| |
| /* |
| * Test case 2: |
| * A simple forward pass of basic block with |
| * downsampling through a stride of 2 but same number |
| * of channels in train mode. |
| */ |
| values_test_basic_block_forward_2() |
| |
| /* |
| * Test case 2: |
| * A simple forward pass of basic block with |
| * downsampling through non-matching channel sizes |
| * and different number of channels in train mode. |
| */ |
| values_test_basic_block_forward_3() |
| |
| /* |
| * *** Residual Layer Value Testing *** |
| * A residual layer is a sequence of residual blocks |
| * which all have the same number of base channels. In |
| * residual networks, there are 4 different residual layer. |
| * With this test, we test the correct computation of the |
| * shape and values of the output by comparing it to PyTorches |
| * residual layers. We modified the PyTorch implementation to |
| * extract the residual layer. |
| */ |
| |
| /* |
| * Test case 1: |
| * A residual layer forward pass with downsampling through |
| * a stride of 2 and non-matching input and base channels |
| * of 2 and 4. |
| */ |
| values_test_residual_layer_forward() |
| |
| /* |
| * *** Residual Layer Backward Pass Testing *** |
| */ |
| |
| values_test_basic_block_backward() |
| |
