A toy experiment of Global Convolutional Network given VOC2012 Datasets.
git clone https://github.com/JamesHsu333/Global_Convolutional_Network.git
cd
pip install -r requirements.txt- Download from Global_Convolutional_Network VOC2012 Dataset
- Extract directory
JPEGImagesandSegmentationClassunder directorydata/ - Extract
train.txtandval.txtfromVOC2012/ImageSets/Segmentationand put them underdata/
The images of VOC2012 are 500x225 pixels. Due to GPU memory limitations, they are resized to 224x224.
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 64, 112, 112] 9,408
BatchNorm2d-2 [-1, 64, 112, 112] 128
ReLU-3 [-1, 64, 112, 112] 0
Conv2d-4 [-1, 64, 112, 112] 4,096
BatchNorm2d-5 [-1, 64, 112, 112] 128
ReLU-6 [-1, 64, 112, 112] 0
Conv2d-7 [-1, 64, 112, 112] 36,864
BatchNorm2d-8 [-1, 64, 112, 112] 128
ReLU-9 [-1, 64, 112, 112] 0
Conv2d-10 [-1, 256, 112, 112] 16,384
BatchNorm2d-11 [-1, 256, 112, 112] 512
Conv2d-12 [-1, 256, 112, 112] 16,384
BatchNorm2d-13 [-1, 256, 112, 112] 512
ReLU-14 [-1, 256, 112, 112] 0
Bottleneck-15 [-1, 256, 112, 112] 0
Conv2d-16 [-1, 64, 112, 112] 16,384
BatchNorm2d-17 [-1, 64, 112, 112] 128
ReLU-18 [-1, 64, 112, 112] 0
Conv2d-19 [-1, 64, 112, 112] 36,864
BatchNorm2d-20 [-1, 64, 112, 112] 128
ReLU-21 [-1, 64, 112, 112] 0
Conv2d-22 [-1, 256, 112, 112] 16,384
BatchNorm2d-23 [-1, 256, 112, 112] 512
ReLU-24 [-1, 256, 112, 112] 0
Bottleneck-25 [-1, 256, 112, 112] 0
Conv2d-26 [-1, 64, 112, 112] 16,384
BatchNorm2d-27 [-1, 64, 112, 112] 128
ReLU-28 [-1, 64, 112, 112] 0
Conv2d-29 [-1, 64, 112, 112] 36,864
BatchNorm2d-30 [-1, 64, 112, 112] 128
ReLU-31 [-1, 64, 112, 112] 0
Conv2d-32 [-1, 256, 112, 112] 16,384
BatchNorm2d-33 [-1, 256, 112, 112] 512
ReLU-34 [-1, 256, 112, 112] 0
Bottleneck-35 [-1, 256, 112, 112] 0
Conv2d-36 [-1, 128, 112, 112] 32,768
BatchNorm2d-37 [-1, 128, 112, 112] 256
ReLU-38 [-1, 128, 112, 112] 0
Conv2d-39 [-1, 128, 56, 56] 147,456
BatchNorm2d-40 [-1, 128, 56, 56] 256
ReLU-41 [-1, 128, 56, 56] 0
Conv2d-42 [-1, 512, 56, 56] 65,536
BatchNorm2d-43 [-1, 512, 56, 56] 1,024
Conv2d-44 [-1, 512, 56, 56] 131,072
BatchNorm2d-45 [-1, 512, 56, 56] 1,024
ReLU-46 [-1, 512, 56, 56] 0
Bottleneck-47 [-1, 512, 56, 56] 0
Conv2d-48 [-1, 128, 56, 56] 65,536
BatchNorm2d-49 [-1, 128, 56, 56] 256
ReLU-50 [-1, 128, 56, 56] 0
Conv2d-51 [-1, 128, 56, 56] 147,456
BatchNorm2d-52 [-1, 128, 56, 56] 256
ReLU-53 [-1, 128, 56, 56] 0
Conv2d-54 [-1, 512, 56, 56] 65,536
BatchNorm2d-55 [-1, 512, 56, 56] 1,024
ReLU-56 [-1, 512, 56, 56] 0
Bottleneck-57 [-1, 512, 56, 56] 0
Conv2d-58 [-1, 128, 56, 56] 65,536
BatchNorm2d-59 [-1, 128, 56, 56] 256
ReLU-60 [-1, 128, 56, 56] 0
Conv2d-61 [-1, 128, 56, 56] 147,456
BatchNorm2d-62 [-1, 128, 56, 56] 256
ReLU-63 [-1, 128, 56, 56] 0
Conv2d-64 [-1, 512, 56, 56] 65,536
BatchNorm2d-65 [-1, 512, 56, 56] 1,024
ReLU-66 [-1, 512, 56, 56] 0
Bottleneck-67 [-1, 512, 56, 56] 0
Conv2d-68 [-1, 128, 56, 56] 65,536
BatchNorm2d-69 [-1, 128, 56, 56] 256
ReLU-70 [-1, 128, 56, 56] 0
Conv2d-71 [-1, 128, 56, 56] 147,456
BatchNorm2d-72 [-1, 128, 56, 56] 256
ReLU-73 [-1, 128, 56, 56] 0
Conv2d-74 [-1, 512, 56, 56] 65,536
BatchNorm2d-75 [-1, 512, 56, 56] 1,024
ReLU-76 [-1, 512, 56, 56] 0
Bottleneck-77 [-1, 512, 56, 56] 0
Conv2d-78 [-1, 256, 56, 56] 131,072
BatchNorm2d-79 [-1, 256, 56, 56] 512
ReLU-80 [-1, 256, 56, 56] 0
Conv2d-81 [-1, 256, 28, 28] 589,824
BatchNorm2d-82 [-1, 256, 28, 28] 512
ReLU-83 [-1, 256, 28, 28] 0
Conv2d-84 [-1, 1024, 28, 28] 262,144
BatchNorm2d-85 [-1, 1024, 28, 28] 2,048
Conv2d-86 [-1, 1024, 28, 28] 524,288
BatchNorm2d-87 [-1, 1024, 28, 28] 2,048
ReLU-88 [-1, 1024, 28, 28] 0
Bottleneck-89 [-1, 1024, 28, 28] 0
Conv2d-90 [-1, 256, 28, 28] 262,144
BatchNorm2d-91 [-1, 256, 28, 28] 512
ReLU-92 [-1, 256, 28, 28] 0
Conv2d-93 [-1, 256, 28, 28] 589,824
BatchNorm2d-94 [-1, 256, 28, 28] 512
ReLU-95 [-1, 256, 28, 28] 0
Conv2d-96 [-1, 1024, 28, 28] 262,144
BatchNorm2d-97 [-1, 1024, 28, 28] 2,048
ReLU-98 [-1, 1024, 28, 28] 0
Bottleneck-99 [-1, 1024, 28, 28] 0
Conv2d-100 [-1, 256, 28, 28] 262,144
BatchNorm2d-101 [-1, 256, 28, 28] 512
ReLU-102 [-1, 256, 28, 28] 0
Conv2d-103 [-1, 256, 28, 28] 589,824
BatchNorm2d-104 [-1, 256, 28, 28] 512
ReLU-105 [-1, 256, 28, 28] 0
Conv2d-106 [-1, 1024, 28, 28] 262,144
BatchNorm2d-107 [-1, 1024, 28, 28] 2,048
ReLU-108 [-1, 1024, 28, 28] 0
Bottleneck-109 [-1, 1024, 28, 28] 0
Conv2d-110 [-1, 256, 28, 28] 262,144
BatchNorm2d-111 [-1, 256, 28, 28] 512
ReLU-112 [-1, 256, 28, 28] 0
Conv2d-113 [-1, 256, 28, 28] 589,824
BatchNorm2d-114 [-1, 256, 28, 28] 512
ReLU-115 [-1, 256, 28, 28] 0
Conv2d-116 [-1, 1024, 28, 28] 262,144
BatchNorm2d-117 [-1, 1024, 28, 28] 2,048
ReLU-118 [-1, 1024, 28, 28] 0
Bottleneck-119 [-1, 1024, 28, 28] 0
Conv2d-120 [-1, 256, 28, 28] 262,144
BatchNorm2d-121 [-1, 256, 28, 28] 512
ReLU-122 [-1, 256, 28, 28] 0
Conv2d-123 [-1, 256, 28, 28] 589,824
BatchNorm2d-124 [-1, 256, 28, 28] 512
ReLU-125 [-1, 256, 28, 28] 0
Conv2d-126 [-1, 1024, 28, 28] 262,144
BatchNorm2d-127 [-1, 1024, 28, 28] 2,048
ReLU-128 [-1, 1024, 28, 28] 0
Bottleneck-129 [-1, 1024, 28, 28] 0
Conv2d-130 [-1, 256, 28, 28] 262,144
BatchNorm2d-131 [-1, 256, 28, 28] 512
ReLU-132 [-1, 256, 28, 28] 0
Conv2d-133 [-1, 256, 28, 28] 589,824
BatchNorm2d-134 [-1, 256, 28, 28] 512
ReLU-135 [-1, 256, 28, 28] 0
Conv2d-136 [-1, 1024, 28, 28] 262,144
BatchNorm2d-137 [-1, 1024, 28, 28] 2,048
ReLU-138 [-1, 1024, 28, 28] 0
Bottleneck-139 [-1, 1024, 28, 28] 0
Conv2d-140 [-1, 512, 28, 28] 524,288
BatchNorm2d-141 [-1, 512, 28, 28] 1,024
ReLU-142 [-1, 512, 28, 28] 0
Conv2d-143 [-1, 512, 14, 14] 2,359,296
BatchNorm2d-144 [-1, 512, 14, 14] 1,024
ReLU-145 [-1, 512, 14, 14] 0
Conv2d-146 [-1, 2048, 14, 14] 1,048,576
BatchNorm2d-147 [-1, 2048, 14, 14] 4,096
Conv2d-148 [-1, 2048, 14, 14] 2,097,152
BatchNorm2d-149 [-1, 2048, 14, 14] 4,096
ReLU-150 [-1, 2048, 14, 14] 0
Bottleneck-151 [-1, 2048, 14, 14] 0
Conv2d-152 [-1, 512, 14, 14] 1,048,576
BatchNorm2d-153 [-1, 512, 14, 14] 1,024
ReLU-154 [-1, 512, 14, 14] 0
Conv2d-155 [-1, 512, 14, 14] 2,359,296
BatchNorm2d-156 [-1, 512, 14, 14] 1,024
ReLU-157 [-1, 512, 14, 14] 0
Conv2d-158 [-1, 2048, 14, 14] 1,048,576
BatchNorm2d-159 [-1, 2048, 14, 14] 4,096
ReLU-160 [-1, 2048, 14, 14] 0
Bottleneck-161 [-1, 2048, 14, 14] 0
Conv2d-162 [-1, 512, 14, 14] 1,048,576
BatchNorm2d-163 [-1, 512, 14, 14] 1,024
ReLU-164 [-1, 512, 14, 14] 0
Conv2d-165 [-1, 512, 14, 14] 2,359,296
BatchNorm2d-166 [-1, 512, 14, 14] 1,024
ReLU-167 [-1, 512, 14, 14] 0
Conv2d-168 [-1, 2048, 14, 14] 1,048,576
BatchNorm2d-169 [-1, 2048, 14, 14] 4,096
ReLU-170 [-1, 2048, 14, 14] 0
Bottleneck-171 [-1, 2048, 14, 14] 0
Conv2d-172 [-1, 21, 112, 112] 37,653
Conv2d-173 [-1, 21, 112, 112] 3,108
Conv2d-174 [-1, 21, 118, 106] 37,653
Conv2d-175 [-1, 21, 112, 112] 3,108
GCN-176 [-1, 21, 112, 112] 0
Conv2d-177 [-1, 21, 112, 112] 3,990
ReLU-178 [-1, 21, 112, 112] 0
Conv2d-179 [-1, 21, 112, 112] 3,990
BR-180 [-1, 21, 112, 112] 0
Conv2d-181 [-1, 21, 56, 56] 75,285
Conv2d-182 [-1, 21, 56, 56] 3,108
Conv2d-183 [-1, 21, 62, 50] 75,285
Conv2d-184 [-1, 21, 56, 56] 3,108
GCN-185 [-1, 21, 56, 56] 0
Conv2d-186 [-1, 21, 56, 56] 3,990
ReLU-187 [-1, 21, 56, 56] 0
Conv2d-188 [-1, 21, 56, 56] 3,990
BR-189 [-1, 21, 56, 56] 0
Conv2d-190 [-1, 21, 28, 28] 150,549
Conv2d-191 [-1, 21, 28, 28] 3,108
Conv2d-192 [-1, 21, 34, 22] 150,549
Conv2d-193 [-1, 21, 28, 28] 3,108
GCN-194 [-1, 21, 28, 28] 0
Conv2d-195 [-1, 21, 28, 28] 3,990
ReLU-196 [-1, 21, 28, 28] 0
Conv2d-197 [-1, 21, 28, 28] 3,990
BR-198 [-1, 21, 28, 28] 0
Conv2d-199 [-1, 21, 14, 14] 301,077
Conv2d-200 [-1, 21, 14, 14] 3,108
Conv2d-201 [-1, 21, 20, 8] 301,077
Conv2d-202 [-1, 21, 14, 14] 3,108
GCN-203 [-1, 21, 14, 14] 0
Conv2d-204 [-1, 21, 14, 14] 3,990
ReLU-205 [-1, 21, 14, 14] 0
Conv2d-206 [-1, 21, 14, 14] 3,990
BR-207 [-1, 21, 14, 14] 0
Conv2d-208 [-1, 21, 28, 28] 3,990
ReLU-209 [-1, 21, 28, 28] 0
Conv2d-210 [-1, 21, 28, 28] 3,990
BR-211 [-1, 21, 28, 28] 0
Conv2d-212 [-1, 21, 56, 56] 3,990
ReLU-213 [-1, 21, 56, 56] 0
Conv2d-214 [-1, 21, 56, 56] 3,990
BR-215 [-1, 21, 56, 56] 0
Conv2d-216 [-1, 21, 112, 112] 3,990
ReLU-217 [-1, 21, 112, 112] 0
Conv2d-218 [-1, 21, 112, 112] 3,990
BR-219 [-1, 21, 112, 112] 0
Conv2d-220 [-1, 21, 112, 112] 3,990
ReLU-221 [-1, 21, 112, 112] 0
Conv2d-222 [-1, 21, 112, 112] 3,990
BR-223 [-1, 21, 112, 112] 0
Conv2d-224 [-1, 21, 224, 224] 3,990
ReLU-225 [-1, 21, 224, 224] 0
Conv2d-226 [-1, 21, 224, 224] 3,990
BR-227 [-1, 21, 224, 224] 0
================================================================
Total params: 24,733,844
Trainable params: 24,733,844
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 1159.64
Params size (MB): 94.35
Estimated Total Size (MB): 1254.56
----------------------------------------------------------------
- Created a
base_modeldirectory under theexperimentsdirectory. It contains a fileparams.jsonwhich sets the hyperparameters for the experiment. It looks like
{
"learning_rate": 0.001,
"batch_size": 5,
"num_epochs": 35,
"dropout_rate": 0.0,
"num_channels": 32,
"save_summary_steps": 100,
"num_workers": 4
}- Train your experiment. Run
python train.py- Created a new directory
learning_ratein experiments. Run
python search_hyperparams.py --parent_dir experiments/learning_rateIt will train and evaluate a model with different values of learning rate defined in search_hyperparams.py and create a new directory for each experiment under experiments/learning_rate/.
4. Display the results of the hyperparameters search in a nice format
python synthesize_results.py --parent_dir experiments/learning_rate- Evaluation on the test set Once you've run many experiments and selected your best model and hyperparameters based on the performance on the validation set, you can finally evaluate the performance of your model on the test set. Run
python evaluate.py --data_dir data/64x64_SIGNS --model_dir experiments/base_model- For more Project Structure details, please refer to Deep Learning Project Structure
- Code implementation refers from SConsul/Global_Convolutional_Network
[1] Chao Peng, Xiangyu Zhang, Gang Yu, Guiming Luo, and Jian Sun. Large kernel matters - improve semantic segmentation by global convolutional network. CoRR, abs/1703.02719, 2017.