| |
| |
| |
| |
| |
| |
| |
| |
| |
|
|
| import os |
| import torch |
| import torch.nn as nn |
| import torch.nn.functional as F |
|
|
| from lib.models.tools.module_helper import ModuleHelper |
| from functools import partial |
|
|
|
|
| class ASPPModuleV2(nn.Module): |
| """ |
| Reference: |
| Chen, Liang-Chieh, et al. |
| *"Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs."* |
| """ |
| def __init__(self, features, inner_features=512, out_features=512, dilations=(12, 24, 36), bn_type=None, dropout=0.1): |
| super(ASPPModuleV2, self).__init__() |
| self.conv_1x1 = nn.Sequential( |
| nn.Conv2d(features, inner_features, kernel_size=1, padding=0, dilation=1, bias=False), |
| ModuleHelper.BNReLU(inner_features, bn_type=bn_type)) |
| self.conv_3x3_1 = nn.Sequential( |
| nn.Conv2d(features, inner_features, kernel_size=3, padding=dilations[0], dilation=dilations[0], bias=False), |
| ModuleHelper.BNReLU(inner_features, bn_type=bn_type)) |
| self.conv_3x3_2 = nn.Sequential( |
| nn.Conv2d(features, inner_features, kernel_size=3, padding=dilations[1], dilation=dilations[1], bias=False), |
| ModuleHelper.BNReLU(inner_features, bn_type=bn_type)) |
| self.conv_3x3_3 = nn.Sequential( |
| nn.Conv2d(features, inner_features, kernel_size=3, padding=dilations[2], dilation=dilations[2], bias=False), |
| ModuleHelper.BNReLU(inner_features, bn_type=bn_type)) |
| self.fuse = nn.Sequential( |
| nn.Conv2d(inner_features * 4, out_features, kernel_size=1, padding=0, dilation=1, bias=False), |
| ModuleHelper.BNReLU(out_features, bn_type=bn_type), |
| nn.Dropout2d(dropout), |
| ) |
|
|
| def forward(self, x): |
| _, _, h, w = x.size() |
| feat1 = self.conv_1x1(x) |
| feat2 = self.conv_3x3_1(x) |
| feat3 = self.conv_3x3_2(x) |
| feat4 = self.conv_3x3_3(x) |
| out = torch.cat((feat1, feat2, feat3, feat4), 1) |
| out = self.fuse(out) |
| return out |
|
|
| class ASPPModule(nn.Module): |
| """ |
| Reference: |
| Chen, Liang-Chieh, et al. *"Rethinking Atrous Convolution for Semantic Image Segmentation."* |
| """ |
| def __init__(self, features, inner_features=256, out_features=256, dilations=(12, 24, 36), bn_type=None, dropout=0.1): |
| super(ASPPModule, self).__init__() |
|
|
| self.conv_gp = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)), |
| nn.Conv2d(features, inner_features, kernel_size=1, padding=0, dilation=1, |
| bias=False), |
| ModuleHelper.BNReLU(inner_features, bn_type=bn_type)) |
|
|
| self.conv_1x1 = nn.Sequential( |
| nn.Conv2d(features, inner_features, kernel_size=1, padding=0, dilation=1, bias=False), |
| ModuleHelper.BNReLU(inner_features, bn_type=bn_type)) |
| self.conv_3x3_1 = nn.Sequential( |
| nn.Conv2d(features, inner_features, kernel_size=3, padding=dilations[0], dilation=dilations[0], bias=False), |
| ModuleHelper.BNReLU(inner_features, bn_type=bn_type)) |
| self.conv_3x3_2 = nn.Sequential( |
| nn.Conv2d(features, inner_features, kernel_size=3, padding=dilations[1], dilation=dilations[1], bias=False), |
| ModuleHelper.BNReLU(inner_features, bn_type=bn_type)) |
| self.conv_3x3_3 = nn.Sequential( |
| nn.Conv2d(features, inner_features, kernel_size=3, padding=dilations[2], dilation=dilations[2], bias=False), |
| ModuleHelper.BNReLU(inner_features, bn_type=bn_type)) |
|
|
| self.fuse = nn.Sequential( |
| nn.Conv2d(inner_features * 5, out_features, kernel_size=1, padding=0, dilation=1, bias=False), |
| ModuleHelper.BNReLU(out_features, bn_type=bn_type), |
| nn.Dropout2d(dropout), |
| ) |
|
|
| def forward(self, x): |
| _, _, h, w = x.size() |
| feat_gp = F.interpolate(self.conv_gp(x), size=(h, w), mode='bilinear', align_corners=True) |
| feat1 = self.conv_1x1(x) |
| feat2 = self.conv_3x3_1(x) |
| feat3 = self.conv_3x3_2(x) |
| feat4 = self.conv_3x3_3(x) |
| out = torch.cat((feat_gp, feat1, feat2, feat3, feat4), 1) |
| out = self.fuse(out) |
| return out |
|
|
|
|
| if __name__ == "__main__": |
| os.environ["CUDA_VISIBLE_DEVICES"] = '0' |
| custom_bn_type = os.environ.get('bn_type', 'inplace_abn') |
|
|
| if int(os.environ.get('eval_os_8', 1)): |
| print("Complexity Evaluation Results for ASPP with input shape [2048 X 128 X 128]") |
| feats = torch.randn((1, 2048, 128, 128)).cuda() |
| aspp_infer = ASPPModule(2048, 256, 256, bn_type=custom_bn_type) |
| else: |
| print("Complexity Evaluation Results for ASPP with input shape [720 X 256 X 512]") |
| feats = torch.randn((1, 720, 256, 512)).cuda() |
| aspp_infer = ASPPModule(720, 256, 256, bn_type=custom_bn_type) |
|
|
| aspp_infer.eval() |
| aspp_infer.cuda() |
|
|
| def count_parameters(model): |
| return sum(p.numel() for p in model.parameters() if p.requires_grad) |
|
|
| avg_time = 0 |
| avg_mem = 0 |
| import time |
| with torch.no_grad(): |
| for i in range(100): |
| start_time = time.time() |
| outputs = aspp_infer(feats) |
| torch.cuda.synchronize() |
| avg_time += (time.time() - start_time) |
| avg_mem += (torch.cuda.max_memory_allocated()-feats.element_size() * feats.nelement()) |
|
|
| print("Average Parameters : {}".format(count_parameters(aspp_infer))) |
| print("Average Running Time: {}".format(avg_time/100)) |
| print("Average GPU Memory: {:.2f} MB".format(avg_mem / 100 / 2**20)) |
| print("\n\n") |
