├── .gitignore ├── LICENSE.md ├── README.md ├── data_loader.py ├── deep_gradient_compression.py └── main.py /.gitignore: -------------------------------------------------------------------------------- 1 | #project specific 2 | ./data 3 | # Byte-compiled / optimized / DLL files 4 | __pycache__/ 5 | *.py[cod] 6 | *$py.class 7 | 8 | # C extensions 9 | *.so 10 | 11 | # Distribution / packaging 12 | .Python 13 | build/ 14 | develop-eggs/ 15 | dist/ 16 | downloads/ 17 | eggs/ 18 | .eggs/ 19 | lib/ 20 | lib64/ 21 | parts/ 22 | sdist/ 23 | var/ 24 | wheels/ 25 | *.egg-info/ 26 | .installed.cfg 27 | *.egg 28 | MANIFEST 29 | 30 | # PyInstaller 31 | # Usually these files are written by a python script from a template 32 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 33 | *.manifest 34 | *.spec 35 | 36 | # Installer logs 37 | pip-log.txt 38 | pip-delete-this-directory.txt 39 | 40 | # Unit test / coverage reports 41 | htmlcov/ 42 | .tox/ 43 | .coverage 44 | .coverage.* 45 | .cache 46 | nosetests.xml 47 | coverage.xml 48 | *.cover 49 | .hypothesis/ 50 | .pytest_cache/ 51 | 52 | # Translations 53 | *.mo 54 | *.pot 55 | 56 | # Django stuff: 57 | *.log 58 | local_settings.py 59 | db.sqlite3 60 | 61 | # Flask stuff: 62 | instance/ 63 | .webassets-cache 64 | 65 | # Scrapy stuff: 66 | .scrapy 67 | 68 | # Sphinx documentation 69 | docs/_build/ 70 | 71 | # PyBuilder 72 | target/ 73 | 74 | # Jupyter Notebook 75 | .ipynb_checkpoints 76 | 77 | # pyenv 78 | .python-version 79 | 80 | # celery beat schedule file 81 | celerybeat-schedule 82 | 83 | # SageMath parsed files 84 | *.sage.py 85 | 86 | # Environments 87 | .env 88 | .venv 89 | env/ 90 | venv/ 91 | ENV/ 92 | env.bak/ 93 | venv.bak/ 94 | 95 | # Spyder project settings 96 | .spyderproject 97 | .spyproject 98 | 99 | # Rope project settings 100 | .ropeproject 101 | 102 | # mkdocs documentation 103 | /site 104 | 105 | # mypy 106 | .mypy_cache/ 107 | 108 | #ide related 109 | .idea 110 | -------------------------------------------------------------------------------- /LICENSE.md: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright [yyyy] [name of copyright owner] 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Deep Gradient Compression 2 | Partial implementation of paper "DEEP GRADIENT COMPRESSION: REDUCING THE COMMUNICATION BANDWIDTH FOR DISTRIBUTED TRAINING" 3 | 4 | ## Installation 5 | for installing required packages run 6 | ` pip3 install -r requirements.txt` 7 | 8 | ## Run project 9 | `python main.py` 10 | 11 | ## Implementation 12 | Current implementation consist of only 13 | * large gradients selection and update 14 | * small gradients accumulation 15 | * momentum corelation 16 | * momentum factor masking 17 | 18 | 19 | ## References 20 | [DEEP GRADIENT COMPRESSION:REDUCING THE COMMUNICATION BANDWIDTH FOR DISTRIBUTED TRAINING](https://arxiv.org/pdf/1712.01887.pdf) 21 | [Pytorch tutorial on distributed training](https://pytorch.org/tutorials/intermediate/dist_tuto.html) 22 | -------------------------------------------------------------------------------- /data_loader.py: -------------------------------------------------------------------------------- 1 | from random import Random 2 | import numpy as np 3 | 4 | 5 | class Partition: 6 | def __init__(self, data, index): 7 | self.data = data 8 | self.index = index 9 | 10 | def __len__(self): 11 | return len(self.index) 12 | 13 | def __getitem__(self, item): 14 | data_idx = self.index[item] 15 | return self.data[data_idx] 16 | 17 | 18 | class DataPartitioner: 19 | def __init__(self, data, sizes=[0.7, 0.2, 0.1], seed=1111): 20 | self.data = data 21 | self.partitions = [] 22 | rand = Random() 23 | rand.seed(seed) 24 | data_len = len(data) 25 | indexes = np.arange(data_len) 26 | rand.shuffle(indexes) 27 | 28 | for frac in sizes: 29 | part_len = int(frac * data_len) 30 | self.partitions.append(indexes[:part_len]) 31 | indexes = indexes[part_len:] 32 | 33 | def use(self, partition): 34 | return Partition(self.data, self.partitions[partition]) 35 | 36 | 37 | 38 | -------------------------------------------------------------------------------- /deep_gradient_compression.py: -------------------------------------------------------------------------------- 1 | import torch 2 | torch.manual_seed(0) 3 | from torch import nn 4 | import torch.distributed as dist 5 | import numpy as np 6 | 7 | def weights_init(m): 8 | if isinstance(m, nn.Conv2d): 9 | torch.nn.init.xavier_uniform_(m.weight.data) 10 | 11 | 12 | class DGC(nn.Module): 13 | def __init__(self, model, rank, size,device_id, momentum, full_update_layers, persentages, itreations): 14 | """Class for performing sparse distributed gradient updates before backpropogation 15 | :parameter 16 | model : torch.Sequentioal, main model to be trained in data-parallel distributed manner 17 | rank : int, rank of the process on which class object will be allocated 18 | size: int, overall number of processes 19 | momentum : int, value of the momentum correlation 20 | full_update_layers : list of ints, layer indexes which will be updated without sparsification 21 | persentages : list of floats, persentages of sparsification 22 | iterations : list of ints, iterations at which persentages of sparsification will be changed 23 | """ 24 | super(DGC, self).__init__() 25 | self.layers = {} 26 | self.shapes = [] 27 | 28 | self.rank = rank 29 | self.size = size 30 | self.device_id = device_id 31 | self.main_model = model 32 | self.momentum = momentum 33 | self.compressed_size = None 34 | self.full_update_layers = full_update_layers 35 | 36 | self.percentages = persentages 37 | self.iterations = itreations 38 | self.current_persentage = None 39 | 40 | for i, (name, layer) in enumerate(model.named_parameters()): 41 | if layer.requires_grad and len(layer.size()) == 4 and i not in self.full_update_layers: 42 | self.layers[name] = torch.zeros(layer.size()) 43 | self.shapes.append(layer.size()) 44 | 45 | 46 | 47 | def avarage_gradient_dense(self): 48 | """ Gradient averaging of layers without sparsification """ 49 | 50 | for i, (name, p) in enumerate(self.main_model.named_parameters()): 51 | tensor = p.grad.data.cpu() 52 | 53 | if i in self.full_update_layers or len(tensor.shape) != 4: 54 | dist.all_reduce(tensor, op=dist.ReduceOp.SUM) 55 | tensor /= float(self.size) 56 | 57 | 58 | def select_top_values_and_indices(self, tensor, name=None, momentum=None): 59 | """Selecting top k gradient per layer 60 | :parameter 61 | tensor : 4D tensor, tensor of gradients at the certain layer 62 | name : string, name of the layer 63 | momentum : float, value of momentum correlation 64 | :return 65 | top_indices : tensor, indices with top values at current layer 66 | top_values : tensor, values at top_indices 67 | """ 68 | current_layer = tensor + self.layers[name] 69 | 70 | current_layer = current_layer.flatten() 71 | kbig = int(len(current_layer) * self.current_persentage / 100) 72 | if kbig == 0: 73 | kbig = 10 74 | _, top_indices_unsorted = torch.topk(torch.abs(current_layer), kbig) 75 | top_values_unsorted = torch.take(current_layer, top_indices_unsorted) 76 | 77 | indices_sorted = torch.argsort(top_values_unsorted) 78 | top_values = top_values_unsorted[indices_sorted] 79 | top_indices = top_indices_unsorted[indices_sorted] 80 | 81 | small_values_tensor = tensor.clone() 82 | small_values_tensor = small_values_tensor.put_(top_indices, torch.zeros(len(top_indices))) 83 | 84 | self.layers[name] = self.layers[name].put_(top_indices, torch.zeros(len(top_indices))) 85 | self.layers[name] += small_values_tensor * momentum 86 | return top_indices, top_values 87 | 88 | def accumulate_gradients(self): 89 | """Accumulation gradients per iterations 90 | :return 91 | top_gradients : list [tensor], selected gradient values from all layers 92 | top_indices : list [tensor], indices of top_gradient values at the original gradient 93 | amounts_per_layer : list [int], numbers of elements selected from each layer""" 94 | gradient_tensors = [] 95 | gradient_indices = [] 96 | gradient_amounts = [] 97 | layer_idx = 0 98 | 99 | for i, (name, p) in enumerate(self.main_model.named_parameters()): 100 | if i in self.full_update_layers: 101 | continue 102 | if len(p.grad.data.cpu().shape) == 4: 103 | top_indices, top_values = self.select_top_values_and_indices(p.grad.data.cpu(), 104 | name, self.momentum) 105 | 106 | gradient_tensors.extend(top_values) 107 | gradient_indices.extend(top_indices) 108 | gradient_amounts.extend(np.ones(len(top_values)) * layer_idx) 109 | layer_idx += 1 110 | top_gradients = torch.FloatTensor(gradient_tensors)[None, None, None, ...] 111 | top_indices = torch.LongTensor(gradient_indices) 112 | amounts_per_layer = torch.LongTensor(gradient_amounts) 113 | 114 | return top_gradients, top_indices, amounts_per_layer 115 | 116 | def update_gradients(self, value): 117 | """update the final constructed sparse gradient before optimization 118 | :parameter 119 | value : list [tensor], constructed sparse gradients""" 120 | depth_idx = 0 121 | for i, (name, param) in enumerate(self.main_model.named_parameters()): 122 | if i in self.full_update_layers: 123 | continue 124 | if param.requires_grad: 125 | if len(param.grad.data.shape) == 4: 126 | param.grad.data = value[depth_idx] 127 | depth_idx += 1 128 | 129 | def avarage_gradients_sparse(self, value): 130 | """Avaraging sparse gradients obtained from the all nodes (synchronized) 131 | :parameter 132 | value : list [tensor], constracted sparse gradients of separate nodes 133 | :return 134 | avg_grads : tensor, avaraged over all nodes gradient tensor 135 | """ 136 | idx = 0 137 | avg_grads = [] 138 | for i, (name, param) in enumerate(self.main_model.named_parameters()): 139 | if i in self.full_update_layers: 140 | continue 141 | if len(param.grad.data.cpu().shape) == 4: 142 | layer = value[idx] 143 | idx += 1 144 | if self.rank == 0: 145 | g_list = [] 146 | for i in range(self.size): 147 | g_list.append(torch.zeros(layer.shape).to('cpu')) 148 | dist.gather(tensor=layer.to('cpu'), dst=0, gather_list=g_list) 149 | div = torch.zeros_like(g_list[0]) 150 | for i in range(len(g_list)): 151 | div += (g_list[i] != 0).float() 152 | 153 | div = torch.clamp(div, 1., len(g_list)) 154 | updated_grad = torch.sum(torch.stack(g_list), dim=0) / div 155 | 156 | avg_grads.append(updated_grad.cuda(self.device_id)) 157 | else: 158 | dist.gather(tensor=layer.to('cpu'), dst=0, gather_list=[]) 159 | 160 | return avg_grads 161 | 162 | def construct_grads(self, grads, indices,amounts): 163 | """Constructions from the separate indices sparse gradient tensor, 164 | with the same shape as original gradient, with top values at choosen indices, 165 | and zeros elsewhere 166 | :parameter 167 | grads : list [tensor], selected top values from each layer 168 | indices : list [tensor], indices of selected values 169 | amounts : list [int], numbers of elements selected from each layer 170 | :return 171 | new_grads : list [tensor], constructed sparse gradients 172 | """ 173 | grads = grads[0, 0, 0] 174 | new_grads = [] 175 | indices = indices.cuda(self.device_id) 176 | indices_amounts = amounts.cuda(self.device_id) 177 | conv_layer_idx = 0 178 | for i, (name, p) in enumerate(self.main_model.named_parameters()): 179 | if i in self.full_update_layers: 180 | continue 181 | tensor = p.grad.data.cpu() 182 | if len(tensor.shape) == 4: 183 | idc = indices[indices_amounts == conv_layer_idx] 184 | layer_grad = grads[indices_amounts == conv_layer_idx] 185 | 186 | updated_grad = torch.zeros_like(p.grad.data) 187 | updated_grad = updated_grad.put_(idc, layer_grad) 188 | new_grads.append(updated_grad) 189 | conv_layer_idx += 1 190 | return new_grads 191 | 192 | def transfer_gradients(self, grad_update_conv): 193 | """transfering avaraged sparse gradient to the all nodes for the optimization of the model at each node 194 | :parameter 195 | grad_update_conv : tensor, final avaraged sparse gradient tensor 196 | :return 197 | upd_grads : final gradient accessable at each node 198 | """ 199 | upd_grads = [] 200 | 201 | for idx in range(len(self.shapes)): 202 | updated = torch.zeros(self.shapes[idx]) 203 | 204 | if self.rank == 0: 205 | reciever_list = [] 206 | for i in range(self.size): 207 | reciever_list.append(grad_update_conv[idx].to('cpu')) 208 | 209 | dist.scatter(tensor=updated, src=0, scatter_list=reciever_list) 210 | else: 211 | dist.scatter(tensor=updated, src=0, scatter_list=[]) 212 | upd_grads.append(updated.cuda(self.device_id)) 213 | 214 | return upd_grads 215 | 216 | 217 | def gradient_update(self, it): 218 | """main function for the gradient sparsification 219 | :parameter 220 | it : iteration of training 221 | """ 222 | if it in self.iterations: 223 | self.current_persentage = self.percentages[self.iterations.index(it)] 224 | 225 | gradient_vector, gradient_indices, gradient_amounts = self.accumulate_gradients() 226 | 227 | updated_grads = self.construct_grads(gradient_vector.cuda(self.device_id), indices=gradient_indices, amounts=gradient_amounts) 228 | grad_update_conv = self.avarage_gradients_sparse(value=updated_grads) 229 | updated = self.transfer_gradients(grad_update_conv) 230 | self.update_gradients(updated) 231 | self.avarage_gradient_dense() 232 | 233 | 234 | 235 | 236 | 237 | 238 | 239 | 240 | 241 | 242 | 243 | 244 | 245 | 246 | 247 | 248 | -------------------------------------------------------------------------------- /main.py: -------------------------------------------------------------------------------- 1 | from data_loader import * 2 | import os 3 | import torch 4 | import torch.distributed as dist 5 | import torch.nn as nn 6 | import torch.optim as optim 7 | from torchvision import datasets, transforms 8 | import torch.nn.functional as F 9 | import torch.utils.data as data 10 | from math import ceil 11 | from torch.autograd import Variable 12 | from torch.multiprocessing import Process 13 | import argparse 14 | from deep_gradient_compression import DGC 15 | torch.manual_seed(0) 16 | torch.backends.cudnn.deterministic = True 17 | torch.backends.cudnn.benchmark = False 18 | 19 | parser = argparse.ArgumentParser() 20 | parser.add_argument('--data_dir', default='./data', help='path where to download, or from where to read data') 21 | parser.add_argument('--momentum', default=0.6, type=int, help='momentum correlation for accumulated gradients') 22 | parser.add_argument('--lr', default=1e-2, type=int, help='learning rate') 23 | parser.add_argument('--epoch', default=10, type=int, help='number of epochs to train') 24 | parser.add_argument('--batch_size', default=128, type=int, help='batch size which will be divided to number of model instances') 25 | parser.add_argument('--world_size', default=2, type=int, help='number of model instances to be run parallel') 26 | parser.add_argument('--persentages', default=[25, 6.25, 1.5625, 0.4, 0.1], type=list, help='exponential decreasing persentages of the gradients for top k selection') 27 | parser.add_argument('--iters', default=[0, 50, 100, 200, 300], type=list, help='iterations at which persentage will be decreased (for args.persentages)') 28 | 29 | args = parser.parse_args() 30 | 31 | 32 | class Net_CIFAR(nn.Module): 33 | """Dummy Convolutional network for CIFAR10 classification""" 34 | def __init__(self): 35 | super(Net_CIFAR, self).__init__() 36 | self.conv1 = nn.Conv2d(3, 10, kernel_size=5, bias=False) 37 | self.conv2 = nn.Conv2d(10, 20, kernel_size=5, bias=False) 38 | self.conv2_drop = nn.Dropout2d() 39 | self.fc1 = nn.Linear(500, 50, bias=False) 40 | self.fc2 = nn.Linear(50, 10, bias=False) 41 | 42 | def forward(self, x): 43 | x = F.relu(F.max_pool2d(self.conv1(x), 2)) 44 | x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2)) 45 | x = x.view(-1, 500) 46 | x = F.relu(self.fc1(x)) 47 | x = F.dropout(x, training=self.training) 48 | x = self.fc2(x) 49 | return F.log_softmax(x) 50 | 51 | 52 | def partition_dataset(): 53 | """Load (or download) dataset and divide the data into partitions to feed into different branches 54 | :return 55 | train_set : function, partition of the dataset depending on rank of the process 56 | b_size : int, batch size at each node 57 | """ 58 | dataset = datasets.CIFAR10(root=args.data_dir, train=True, 59 | download=True, transform=transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])) 60 | size = dist.get_world_size() 61 | # split batch size in two equal parts 62 | b_size = int(args.batch_size / float(size)) 63 | 64 | # partition dataset to the number of parallel instances 65 | partition_size = [1. / size for _ in range(size)] 66 | partition = DataPartitioner(dataset, partition_size) 67 | partition = partition.use(dist.get_rank()) 68 | train_set = data.DataLoader(partition, batch_size=b_size, shuffle=True) 69 | 70 | return train_set, b_size 71 | 72 | 73 | def run(rank, world_size): 74 | """main training function""" 75 | device_id = rank 76 | train_set, b_size = partition_dataset() 77 | model = Net_CIFAR().cuda(device_id) 78 | optimizer = optim.SGD(model.parameters(), lr=args.lr) 79 | num_batches = ceil(len(train_set.dataset) / float(b_size)) 80 | 81 | dgc_trainer = DGC(model=model, rank=rank, size=world_size, device_id=device_id, 82 | momentum=args.momentum, full_update_layers=[4],persentages=args.persentages, itreations=args.iters) 83 | 84 | 85 | for epoch in range(args.epoch): 86 | epoch_loss = 0 87 | for batch_idx, (data, target) in enumerate(train_set): 88 | it = epoch * len(train_set) + batch_idx 89 | 90 | data, target = Variable(data.cuda(device_id)), Variable(target.cuda(device_id)) 91 | optimizer.zero_grad() 92 | output = model(data) 93 | loss = F.nll_loss(output, target) 94 | epoch_loss += loss 95 | loss.backward() 96 | dgc_trainer.gradient_update(it) 97 | 98 | optimizer.step() 99 | 100 | print('Rank ',dist.get_rank(), ', epoch ', epoch, ': ',epoch_loss / num_batches) 101 | 102 | 103 | def init_processing(rank, size, fn, backend='gloo'): 104 | """initiale each process by indicate where the master node is located(by ip and port) and run main function 105 | :parameter 106 | rank : int , rank of current process 107 | size : int, overall number of processes 108 | fn : function, function to run at each node 109 | backend : string, name of the backend for distributed operations 110 | """ 111 | os.environ['MASTER_ADDR'] = '127.0.0.1' 112 | os.environ['MASTER_PORT'] = '29500' 113 | dist.init_process_group(backend=backend, rank=rank, world_size=size) 114 | fn(rank, size) 115 | 116 | 117 | if __name__ == '__main__': 118 | processes = [] 119 | for rank in range(args.world_size): 120 | p = Process(target=init_processing, args=(rank, args.world_size, run)) 121 | p.start() 122 | processes.append(p) 123 | 124 | for p in processes: 125 | p.join() 126 | --------------------------------------------------------------------------------