!2199 ssd_mobilenetV1_gpu_PR

Merge pull request !2199 from 陈宇凡/master

official/cv/ssd/README.md

@@ -254,6 +254,7 @@ Then you can run everything just like on ascend.
       │   ├── local_adapter.py            ## local adapter
       │   ├── moxing_adapter.py           ## moxing adapter
     ├─ config
+        ├─ ssd_mobilenet_v1_300_config_gpu.yaml ## parameter configuration
         ├─ ssd_mobilenet_v1_fpn_config.yaml ## parameter configuration
         ├─ ssd_resnet50_fpn_config.yaml ## parameter configuration
         ├─ ssd_vgg16_config.yaml ## parameter configuration
@@ -692,6 +693,24 @@ mAP: 0.23657619676441116
 | configuration           | ssd_vgg16_config.yaml      | ssd_vgg16_config_gpu.yaml      |
 | Scripts             | <https://gitee.com/mindspore/models/tree/master/official/cv/ssd> |
 
+| Parameters          |                        GPU                         |
+| ------------------- | ------------------------------------------------------ |
+| Model Version       | SSD MobileNetV1          |
+| Resource            | GPU(Tesla V100 PCIE)，CPU 2.1GHz 64 cores，Memory 128G   |
+| Uploaded Date       | 03/03/2022 (month/day/year) |
+| MindSpore Version   | 1.5.0                     |
+| Dataset             | COCO2017                    |
+| Training Parameters | epoch = 500,  batch_size = 32  |
+| Optimizer           | Momentum                        |
+| Loss Function       | Sigmoid Cross Entropy,SmoothL1Loss    |
+| Speed               | 8pcs: 108 ms/step                     |
+| Total time          | 8pcs: 6.87hours                      |
+| outputs             | mAP                                          |
+| Accuracy            | IoU=0.50: 21.5%                       |
+| Model for inference | 88M(.ckpt file)                        |
+| configuration           | ssd_mobilenet_v1_300_config_gpu.yaml    |
+| Scripts             | <https://gitee.com/mindspore/models/tree/master/official/cv/ssd> |
+
 ## [Description of Random Situation](#contents)
 
 In dataset.py, we set the seed inside “create_dataset" function. We also use random seed in train.py.

official/cv/ssd/README_CN.md

@@ -202,6 +202,7 @@ bash run_eval_gpu.sh [DATASET] [CHECKPOINT_PATH] [DEVICE_ID] [CONFIG_PATH]
       │   ├──local_adapter.py             ## 本地设备配置
       │   ├──moxing_adapter.py            ## modelarts设备配置
     ├─ config
+        ├─ ssd_mobilenet_v1_300_config_gpu.yaml ## 参数配置
         ├─ ssd_mobilenet_v1_fpn_config.yaml ## 参数配置
         ├─ ssd_resnet50_fpn_config.yaml ## 参数配置
         ├─ ssd_vgg16_config.yaml ## 参数配置
@@ -579,6 +580,24 @@ mAP: 0.23657619676441116
 | 参数文件           | ssd_vgg16_config.yaml      | ssd_vgg16_config_gpu.yaml      |
 | 脚本链接             | <https://gitee.com/mindspore/models/tree/master/official/cv/ssd> |
 
+| 参数           |                        GPU                         |
+| ------------------- | ------------------------------------------------------ |
+| 网络       | SSD MobileNetV1          |
+| 资源          | GPU(Tesla V100 PCIE)，CPU 2.1GHz 64 cores，Memory 128G   |
+| 上传日期       | 03/03/2022 (month/day/year) |
+| 昇思版本   | 1.5.0                     |
+| 数据集        | COCO2017                    |
+| 训练参数 | epoch = 500,  batch_size = 32  |
+| 优化器          | Momentum                        |
+| 损失函数      | Sigmoid Cross Entropy,SmoothL1Loss    |
+| 性能               | 8pcs: 108 ms/step                     |
+| 训练耗时          | 8pcs: 6.87hours                      |
+| 推理输出               | mAP                                          |
+| 评价指标            | IoU=0.50: 21.5%                       |
+| 推理模型大小 | 88M(.ckpt file)                |
+| 参数文件          | ssd_mobilenet_v1_300_config_gpu.yaml    |
+| 脚本链接             | <https://gitee.com/mindspore/models/tree/master/official/cv/ssd> |
+
 # 随机情况说明
 
 dataset.py中设置了“create_dataset”函数内的种子，同时还使用了train.py中的随机种子。

official/cv/ssd/config/ssd_mobilenet_v1_300_config_gpu.yaml

+# Builtin Configurations(DO NOT CHANGE THESE CONFIGURATIONS unless you know exactly what you are doing)
+enable_modelarts: False
+# Url for modelarts
+data_url: ""
+train_url: ""
+checkpoint_url: ""
+# Path for local
+run_distribute: False
+enable_profiling: False
+data_path: "/DATA_1/cyf"
+output_path: "/cache/train"
+load_path: "/cache/checkpoint_path/"
+device_target: "GPU"
+checkpoint_path: "./checkpoint/"
+checkpoint_file_path: "ssd-500_458.ckpt"
+
+# ==============================================================================
+# Training options
+model_name: "ssd_mobilenet_v1"
+img_shape: [300, 300]
+num_ssd_boxes: 1917
+match_threshold: 0.5
+nms_threshold: 0.6
+min_score: 0.1
+max_boxes: 100
+all_reduce_fusion_config: [29, 58, 89]
+use_float16: True
+
+# learing rate settings
+lr_init: 0.001
+lr_end_rate: 0.001
+warmup_epochs: 2
+momentum: 0.9
+weight_decay: 0.00015
+ssd_vgg_bn: False
+pretrain_vgg_bn: False
+
+
+# network
+num_addition_layers: 4
+extras_strides: [1, 1, 2, 2, 2, 2]
+num_default: [3, 6, 6, 6, 6, 6]
+extras_in_channels: [256, 512, 1024, 512, 256, 256]
+extras_out_channels: [512, 1024, 512, 256, 256, 128]
+extras_strides: [1, 1, 2, 2, 2, 2]
+extras_ratio: [0.2, 0.2, 0.2, 0.25, 0.5, 0.25]
+feature_size: [19, 10, 5, 3, 2, 1]
+min_scale: 0.2
+max_scale: 0.95
+aspect_ratios: [[], [2, 3], [2, 3], [2, 3], [2, 3], [2, 3]]
+steps: [16, 32, 64, 100, 150, 300]
+prior_scaling: [0.1, 0.2]
+gamma: 2.0
+alpha: 0.75
+
+dataset: "coco"
+lr: 0.05
+mode_sink: "sink"
+device_id: 0
+device_num: 1
+epoch_size: 500
+batch_size: 32
+loss_scale: 1024
+pre_trained: ""
+pre_trained_epoch_size: 0
+save_checkpoint_epochs: 10
+only_create_dataset: False
+eval_start_epoch: 40
+eval_interval: 1
+run_eval: False
+filter_weight: False
+freeze_layer: None
+save_best_ckpt: True
+
+result_path: ""
+img_path: ""
+drop: False
+
+# `mindrecord_dir` and `coco_root` are better to use absolute path.
+feature_extractor_base_param: ""
+checkpoint_filter_list: ['multi_loc_layers', 'multi_cls_layers']
+mindrecord_dir: "/DATA_1/cyf/MindRecord_COCO"
+coco_root: "/DATA_1/cyf/"
+train_data_type: "train2017"
+val_data_type: "val2017"
+instances_set: "annotations/instances_{}.json"
+classes: ['background', 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus',
+                'train', 'truck', 'boat', 'traffic light', 'fire hydrant',
+                'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog',
+                'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra',
+                'giraffe', 'backpack', 'umbrella', 'handbag', 'tie',
+                'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball',
+                'kite', 'baseball bat', 'baseball glove', 'skateboard',
+                'surfboard', 'tennis racket', 'bottle', 'wine glass', 'cup',
+                'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',
+                'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza',
+                'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed',
+                'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote',
+                'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink',
+                'refrigerator', 'book', 'clock', 'vase', 'scissors',
+                'teddy bear', 'hair drier', 'toothbrush']
+num_classes: 81
+# The annotation.json position of voc validation dataset.
+voc_json: "annotations/voc_instances_val.json"
+# voc original dataset.
+voc_root: "/data/voc_dataset"
+# if coco or voc used, `image_dir` and `anno_path` are useless.
+image_dir: ""
+anno_path: ""
+file_name: "ssd"
+file_format: "AIR"
+
+---
+# Help description for each configuration
+enable_modelarts: "Whether training on modelarts, default: False"
+data_url: "Dataset url for obs"
+train_url: "Training output url for obs"
+checkpoint_url: "The location of checkpoint for obs"
+data_path: "Dataset path for local"
+output_path: "Training output path for local"
+load_path: "The location of checkpoint for obs"
+device_target: "Target device type, available: [Ascend, GPU, CPU]"
+enable_profiling: "Whether enable profiling while training, default: False"
+num_classes: "Class for dataset"
+batch_size: "Batch size for training and evaluation"
+epoch_size: "Total training epochs."
+keep_checkpoint_max: "keep the last keep_checkpoint_max checkpoint"
+checkpoint_path: "The location of the checkpoint file."
+checkpoint_file_path: "The location of the checkpoint file."

official/cv/ssd/eval.py

@@ -18,7 +18,7 @@
 import os
 import mindspore as ms
 from mindspore import Tensor
-from src.ssd import SSD300, SsdInferWithDecoder, ssd_mobilenet_v2, ssd_mobilenet_v1_fpn, ssd_resnet50_fpn, ssd_vgg16
+from src.ssd import SSD300, SsdInferWithDecoder, ssd_mobilenet_v2, ssd_mobilenet_v1_fpn, ssd_mobilenet_v1, ssd_resnet50_fpn, ssd_vgg16
 from src.dataset import create_ssd_dataset, create_mindrecord
 from src.eval_utils import apply_eval
 from src.box_utils import default_boxes
@@ -38,6 +38,8 @@ def ssd_eval(dataset_path, ckpt_path, anno_json):
         net = ssd_mobilenet_v1_fpn(config=config)
     elif config.model_name == "ssd_resnet50_fpn":
         net = ssd_resnet50_fpn(config=config)
+    elif config.model_name == "ssd_mobilenet_v1":
+        net = ssd_mobilenet_v1(config=config)
     else:
         raise ValueError(f'config.model: {config.model_name} is not supported')
     net = SsdInferWithDecoder(net, Tensor(default_boxes), config)

official/cv/ssd/export.py

@@ -18,7 +18,7 @@ import numpy as np
 
 import mindspore as ms
 from mindspore import Tensor
-from src.ssd import SSD300, SsdInferWithDecoder, ssd_mobilenet_v2, ssd_mobilenet_v1_fpn, ssd_resnet50_fpn, ssd_vgg16
+from src.ssd import SSD300, SsdInferWithDecoder, ssd_mobilenet_v2, ssd_mobilenet_v1_fpn, ssd_mobilenet_v1, ssd_resnet50_fpn, ssd_vgg16
 from src.model_utils.config import config
 from src.model_utils.moxing_adapter import moxing_wrapper
 from src.box_utils import default_boxes
@@ -49,6 +49,8 @@ def run_export():
         net = ssd_mobilenet_v1_fpn(config=config)
     elif config.model_name == "ssd_resnet50_fpn":
         net = ssd_resnet50_fpn(config=config)
+    elif config.model_name == "ssd_mobilenet_v1":
+        net = ssd_mobilenet_v1(config=config)
     else:
         raise ValueError(f'config.model: {config.model_name} is not supported')
 

official/cv/ssd/src/mobilenet_v1.py

@@ -87,5 +87,39 @@ class MobileNetV1(nn.Cell):
         output = self.fc(output)
         return output
 
+class FeatureSelector(nn.Cell):
+    """
+    Select specific layers from an entire feature list
+    """
+    def __init__(self, feature_idxes):
+        super(FeatureSelector, self).__init__()
+        self.feature_idxes = feature_idxes
+
+    def construct(self, feature_list):
+        selected = ()
+        for i in self.feature_idxes:
+            selected = selected + (feature_list[i],)
+        return selected
+
+class MobileNetV1Feature(nn.Cell):
+    """
+    MobileNetV1 with FPN as SSD backbone.
+    """
+    def __init__(self, config):
+        super(MobileNetV1Feature, self).__init__()
+        self.mobilenet_v1 = MobileNetV1(features_only=True)
+
+        self.selector = FeatureSelector([14, 26])
+
+        self.layer_indexs = [14, 26]
+
+    def construct(self, x):
+        features = self.mobilenet_v1(x)
+        features = self.selector(features)
+        return features
+
 def mobilenet_v1(class_num=1001):
     return MobileNetV1(class_num)
+
+def mobilenet_v1_Feature(config):
+    return MobileNetV1Feature(config)

official/cv/ssd/src/ssd.py

@@ -24,6 +24,7 @@ from mindspore.communication.management import get_group_size
 import mindspore.ops as ops
 from .fpn import mobilenet_v1_fpn, resnet50_fpn
 from .vgg16 import vgg16
+from .mobilenet_v1 import mobilenet_v1_Feature
 
 
 def _make_divisible(v, divisor, min_value=None):
@@ -367,6 +368,56 @@ class SsdMobilenetV1Fpn(nn.Cell):
         pred_label = ops.cast(pred_label, ms.float32)
         return pred_loc, pred_label
 
+
+class SsdMobilenetV1Feature(nn.Cell):
+    """
+    SSD Network using mobilenetV1 with fpn to extract features
+
+    Args:
+        config (dict): The default config of SSD.
+        is_training (bool): Used for training, default is True.
+
+    Returns:
+        Tensor, localization predictions.
+        Tensor, class conf scores.
+
+    Examples:backbone
+         SsdMobilenetV1Feature(config, True).
+    """
+    def __init__(self, config, is_training=True):
+        super(SsdMobilenetV1Feature, self).__init__()
+        self.multi_box = MultiBox(config)
+        self.activation = ops.Sigmoid()
+        self.feature_extractor = mobilenet_v1_Feature(config)
+        in_channels = config.extras_in_channels
+        out_channels = config.extras_out_channels
+        strides = config.extras_strides
+        residual_list = []
+        for i in range(2, len(in_channels)):
+            residual = ConvBNReLU(in_channels[i], out_channels[i], stride=strides[i],
+                                  )
+            residual_list.append(residual)
+        self.multi_residual = nn.layer.CellList(residual_list)
+        self.multi_box = MultiBox(config)
+        self.is_training = is_training
+        if not is_training:
+            self.activation = ops.Sigmoid()
+
+    def construct(self, x):
+        feature, output = self.feature_extractor(x)
+        multi_feature = (feature, output)
+        feature = output
+        for residual in self.multi_residual:
+            feature = residual(feature)
+            multi_feature += (feature,)
+        pred_loc, pred_label = self.multi_box(multi_feature)
+        if not self.training:
+            pred_label = self.activation(pred_label)
+        pred_loc = ops.cast(pred_loc, ms.float32)
+        pred_label = ops.cast(pred_label, ms.float32)
+        return pred_loc, pred_label
+
+
 class SsdResNet50Fpn(nn.Cell):
     """
     SSD Network using ResNet50 with fpn to extract features
@@ -633,6 +684,9 @@ class SsdInferWithDecoder(nn.Cell):
 def ssd_mobilenet_v1_fpn(**kwargs):
     return SsdMobilenetV1Fpn(**kwargs)
 
+def ssd_mobilenet_v1(**kwargs):
+    return SsdMobilenetV1Feature(**kwargs)
+
 def ssd_resnet50_fpn(**kwargs):
     return SsdResNet50Fpn(**kwargs)
 

official/cv/ssd/train.py

@@ -25,7 +25,7 @@ from mindspore.train import Model
 from mindspore.context import ParallelMode
 from mindspore.common import set_seed, dtype
 from src.ssd import SSD300, SsdInferWithDecoder, SSDWithLossCell, TrainingWrapper, ssd_mobilenet_v2,\
-    ssd_mobilenet_v1_fpn, ssd_resnet50_fpn, ssd_vgg16
+    ssd_mobilenet_v1_fpn, ssd_mobilenet_v1, ssd_resnet50_fpn, ssd_vgg16
 from src.dataset import create_ssd_dataset, create_mindrecord
 from src.lr_schedule import get_lr
 from src.init_params import init_net_param, filter_checkpoint_parameter_by_list
@@ -54,6 +54,15 @@ def ssd_model_build():
                 param_dict["network.feature_extractor.mobilenet_v1." + x] = param_dict[x]
                 del param_dict[x]
             ms.load_param_into_net(ssd.feature_extractor.mobilenet_v1.network, param_dict)
+    elif config.model_name == "ssd_mobilenet_v1":
+        ssd = ssd_mobilenet_v1(config=config)
+        init_net_param(ssd)
+        if config.feature_extractor_base_param != "":
+            param_dict = ms.load_checkpoint(config.feature_extractor_base_param)
+            for x in list(param_dict.keys()):
+                param_dict["network.feature_extractor.mobilenet_v1." + x] = param_dict[x]
+                del param_dict[x]
+            ms.load_param_into_net(ssd.feature_extractor.mobilenet_v1.network, param_dict)
     elif config.model_name == "ssd_resnet50_fpn":
         ssd = ssd_resnet50_fpn(config=config)
         init_net_param(ssd)
@@ -83,6 +92,10 @@ def set_graph_kernel_context(device_target, model):
         # Enable graph kernel for default model ssd300 on GPU back-end.
         ms.set_context(enable_graph_kernel=True,
                        graph_kernel_flags="--enable_parallel_fusion --enable_expand_ops=Conv2D")
+    if device_target == "GPU" and model == "ssd_mobilenet_v1":
+        # Enable graph kernel for default model ssd300 on GPU back-end.
+        ms.context.set_context(enable_graph_kernel=True,
+                               graph_kernel_flags="--enable_parallel_fusion --enable_expand_ops=Conv2D")
 
 @moxing_wrapper()
 def train_net():