[WIP]add_benchmark

configs/caddn/caddn_deeplabv3p_resnet101_os8_kitti.yml

@@ -2,6 +2,14 @@ batch_size: 4
 iters: 74240  # 928*80
 sync_bn: true
 
+amp_cfg:
+  use_amp: False
+  # only enable backbone and fpn
+  enable: False
+  level: O1
+  scaler:
+    init_loss_scaling: 32.0
+
 train_dataset:
   type: KittiDepthDataset
   dataset_root: data/kitti

paddle3d/models/detection/caddn/caddn.py

@@ -25,6 +25,7 @@
 from paddle3d.models.layers import ConvBNReLU
 from paddle3d.utils import checkpoint
 from paddle3d.utils.logger import logger
+from paddle3d.utils import dtype2float32
 
 from .bev import BEV
 from .f2v import FrustumToVoxel
@@ -71,9 +72,16 @@ def train_forward(self, data):
             data["batch_size"] = b
 
         # ffe
-        image_features = self.backbone_3d(images)
+        if hasattr(self, 'amp_cfg_'):
+            with paddle.amp.auto_cast(**self.amp_cfg_):
+                image_features = self.backbone_3d(images)
+                depth_logits = self.class_head(image_features, data["image_shape"])
+            image_features = dtype2float32(image_features)
+            depth_logits = dtype2float32(depth_logits)
+        else:
+            image_features = self.backbone_3d(images)
+            depth_logits = self.class_head(image_features, data["image_shape"])
 
-        depth_logits = self.class_head(image_features, data["image_shape"])
         data = self.ffe(image_features[0], depth_logits, data)
 
         # frustum_to_voxel
@@ -91,9 +99,14 @@ def train_forward(self, data):
         data["spatial_features"] = bev_features
 
         # backbone_2d
-        data = self.backbone_2d(data)
-        predictions = self.dense_head(data)
-
+        if hasattr(self, 'amp_cfg_'):
+            with paddle.amp.auto_cast(**self.amp_cfg_):
+                data = self.backbone_2d(data)
+                predictions = self.dense_head(data)
+            predictions = dtype2float32(predictions)
+        else:
+            data = self.backbone_2d(data)
+            predictions = self.dense_head(data)
         loss = self.get_loss(predictions)
         return loss
 

paddle3d/models/transformers/attentions/spatial_cross_attention.py

@@ -418,6 +418,9 @@ def forward(self,
 
         # sampling_locations.stop_gradient = True
         # attention_weights.stop_gradient = True
+        value = value.cast(paddle.float32)
+        sampling_locations = sampling_locations.cast(paddle.float32)
+        attention_weights = attention_weights.cast(paddle.float32)
         output = ms_deform_attn.ms_deform_attn(
             value, sampling_locations, attention_weights, spatial_shapes,
             level_start_index, self.im2col_step)
@@ -627,6 +630,7 @@ def forward(self,
 
         value = value.cast(paddle.float32)
         sampling_locations = sampling_locations.cast(paddle.float32)
+        attention_weights = attention_weights.cast(paddle.float32)
         output = ms_deform_attn.ms_deform_attn(
             value, sampling_locations, attention_weights, spatial_shapes,
             level_start_index, self.im2col_step)

paddle3d/models/transformers/attentions/temporal_self_attention.py

@@ -255,6 +255,7 @@ def forward(self,
         # attention_weights.stop_gradient = True
         value = value.cast(paddle.float32)
         sampling_locations = sampling_locations.cast(paddle.float32)
+        attention_weights = attention_weights.cast(paddle.float32)
         output = ms_deform_attn.ms_deform_attn(
             value, sampling_locations, attention_weights, spatial_shapes,
             level_start_index, self.im2col_step)

paddle3d/models/transformers/utils.py

@@ -232,6 +232,7 @@ def nan_to_num(x, nan=0.0, posinf=None, neginf=None, name=None):
     posinf_value = paddle.full_like(x, float("+inf"))
     neginf_value = paddle.full_like(x, float("-inf"))
     nan = paddle.full_like(x, nan)
+    x = x.astype("float32")
     assert x.dtype in [paddle.float32, paddle.float64]
     is_float32 = x.dtype == paddle.float32
     if posinf is None:

test_tipc/configs/bevformer/train_infer_python.txt

@@ -0,0 +1,58 @@
+===========================train_params===========================
+model_name:bevformer
+python:python3.7
+gpu_list:0
+##
+auto_cast:null
+--iters:lite_train_lite_infer=100
+--save_dir:./output/
+--batch_size:lite_train_lite_infer=4
+pretrained_model:null
+train_model_name:iter_100/model.pdparams
+train_infer_img_dir:./datasets/nuscenes
+null:null
+##
+trainer:norm_train
+norm_train:tools/train.py --config configs/bevformer/bevformer_tiny_r50_fpn_fp16_nuscenes.yml --num_workers 20 --log_interval 1 --seed 123 --save_interval 100
+pact_train:null
+fpgm_train:null
+distill_train:null
+null:null
+null:null
+##
+===========================eval_params===========================
+eval:null
+null:null
+##
+===========================infer_params===========================
+--output_dir:./output/
+load:null
+norm_export:null
+quant_export:null
+fpgm_export:null
+distill_export:null
+export1:null
+export2:null
+inference_dir:null
+infer_model:null
+infer_export:null
+infer_quant:False
+inference:null
+null:null
+null:null
+null:null
+null:null
+null:null
+--model_path:
+null:null
+null:null
+--benchmark:True
+null:null
+===========================to_static_train_benchmark_params===========================
+to_static_train:--to_static True
+===========================train_benchmark_params==========================
+batch_size:2
+fp_items:fp32|fp16
+iters:100
+--profiler_options:batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile
+flags:FLAGS_cudnn_exhaustive_search=0

test_tipc/configs/caddn/train_infer_python.txt

@@ -0,0 +1,58 @@
+===========================train_params===========================
+model_name:caddn
+python:python3.7
+gpu_list:0
+##
+auto_cast:null
+--iters:lite_train_lite_infer=100
+--save_dir:./output/
+--batch_size:lite_train_lite_infer=4
+pretrained_model:null
+train_model_name:iter_100/model.pdparams
+train_infer_img_dir:./datasets/nuscenes
+null:null
+##
+trainer:norm_train
+norm_train:tools/train.py --config configs/caddn/caddn_deeplabv3p_resnet101_os8_kitti.yml --num_workers 4 --log_interval 1 --seed 123 --save_interval 100
+pact_train:null
+fpgm_train:null
+distill_train:null
+null:null
+null:null
+##
+===========================eval_params===========================
+eval:null
+null:null
+##
+===========================infer_params===========================
+--output_dir:./output/
+load:null
+norm_export:null
+quant_export:null
+fpgm_export:null
+distill_export:null
+export1:null
+export2:null
+inference_dir:null
+infer_model:null
+infer_export:null
+infer_quant:False
+inference:null
+null:null
+null:null
+null:null
+null:null
+null:null
+--model_path:
+null:null
+null:null
+--benchmark:True
+null:null
+===========================to_static_train_benchmark_params===========================
+to_static_train:--to_static True
+===========================train_benchmark_params==========================
+batch_size:4
+fp_items:fp32|fp16
+iters:100
+--profiler_options:batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile
+flags:FLAGS_cudnn_exhaustive_search=0

test_tipc/prepare.sh

@@ -30,6 +30,7 @@ trainer_list=$(func_parser_value "${lines[14]}")
 
 if [ ${MODE} = "benchmark_train" ];then
     pip install -r requirements.txt
+    pip install -e .
     MODE="lite_train_lite_infer"
 fi
 
@@ -39,13 +40,36 @@ if [ ${MODE} = "lite_train_lite_infer" ];then
         mkdir -p ./test_tipc/data/mini_modelnet40
         cd ./test_tipc/data/mini_modelnet40 && tar xf ../mini_modelnet40.tar.gz && cd ../../
     elif [ ${model_name} = "petrv2" ]; then
-        rm -rf ./data/nuscenes
+
         wget -nc  -P ./ https://paddle3d.bj.bcebos.com/pretrained/fcos3d_vovnet_imgbackbone-remapped.pdparams --no-check-certificate
-        cd ./data/ && unzip nuscenes.zip && cd ../
+        rm -rf ./data
+        # download data
+        # cd Paddle3D
+        mkdir data && cd data
+        python ${BENCHMARK_ROOT}/paddlecloud/file_upload_download.py \
+            --remote-path /user/userdata/benchmark/data/Paddle3D/petrv2/ \
+            --local-path ./ \
+            --mode download
+        unzip nuscenes.zip && cd ../
     elif [ ${model_name} = "centerpoint" ]; then
         rm -rf ./datasets/KITTI
         wget -nc -P ./datasets/ https://paddle3d.bj.bcebos.com/TIPC/dataset/kitti_mini_centerpoint.tar.gz --no-check-certificate
-        cd ./datasets/ && tar -xzf kitti_mini_centerpoint.tar.gz && cd ../ ;;
+        cd ./datasets/ && tar -xzf kitti_mini_centerpoint.tar.gz && cd ../
+    elif [ ${model_name} = "bevformer" ]; then
+        rm -rf datasets && mkdir datasets && cd datasets
+        python ${BENCHMARK_ROOT}/paddlecloud/file_upload_download.py \
+            --remote-path /user/userdata/benchmark/data/Paddle3D/bevformer/ \
+            --local-path ./ \
+            --mode download
+        tar xf nuscenes.tar && cd ../
+    elif [ ${model_name} = "caddn" ]; then
+        rm -rf data && mkdir data && cd data
+        python ${BENCHMARK_ROOT}/paddlecloud/file_upload_download.py \
+            --remote-path /user/userdata/benchmark/data/Paddle3D/caddn/ \
+            --local-path ./ \
+            --mode download
+        tar xf KITTI.tar &&  mv KITTI kitti
+        cd ../;
     else
         echo "Not added into TIPC yet."
     fi

test_tipc/test_train_inference_python.sh

@@ -256,11 +256,11 @@ else
                 fi
                 set_save_model=$(func_set_params "${save_model_key}" "${save_log}")
                 if [ ${#gpu} -le 2 ];then  # train with cpu or single gpu
-                    cmd="${python} ${run_train} ${set_use_gpu}  ${set_save_model} ${set_train_params1} ${set_epoch} ${set_pretrain} ${set_batchsize} ${set_amp_config} ${set_amp_level}"
+                    cmd="${python} ${run_train} ${set_use_gpu}  ${set_save_model} ${set_epoch} ${set_pretrain} ${set_batchsize} ${set_amp_config} ${set_amp_level} ${set_train_params1}"
                 elif [ ${#ips} -le 26 ];then  # train with multi-gpu
-                    cmd="${python} -m paddle.distributed.launch --gpus=${gpu} ${run_train} ${set_use_gpu} ${set_save_model} ${set_train_params1} ${set_epoch} ${set_pretrain} ${set_batchsize} ${set_amp_config} ${set_amp_level}"
+                    cmd="${python} -m paddle.distributed.launch --gpus=${gpu} ${run_train} ${set_use_gpu} ${set_train_params1} ${set_epoch} ${set_pretrain} ${set_batchsize} ${set_amp_config} ${set_amp_level} ${set_train_params1}"
                 else     # train with multi-machine
-                    cmd="${python} -m paddle.distributed.launch --ips=${ips} --gpus=${gpu} ${run_train} ${set_use_gpu} ${set_save_model} ${set_train_params1} ${set_pretrain} ${set_epoch} ${set_batchsize} ${set_amp_config} ${set_amp_level}"
+                    cmd="${python} -m paddle.distributed.launch --ips=${ips} --gpus=${gpu} ${run_train} ${set_use_gpu} ${set_save_model} ${set_pretrain} ${set_epoch} ${set_batchsize} ${set_amp_config} ${set_amp_level} ${set_train_params1}"
                 fi
                 # run train
                 export FLAGS_cudnn_deterministic=True