[PaddleSlim]Fix name conficts in distillation. (#2013)

* Fix name conficts in distillation.

* Download pretrain model in run_quant.sh

PaddleSlim/compress.py

@@ -69,14 +69,9 @@ def compress(args):
         momentum=0.9,
         learning_rate=fluid.layers.piecewise_decay(
             boundaries=[5000 * 30, 5000 * 60, 5000 * 90],
-            values=[0.01, 0.001, 0.0001, 0.00001]),
+            values=[0.1, 0.01, 0.001, 0.0001]),
         regularization=fluid.regularizer.L2Decay(4e-5))
 
-    #    opt = fluid.optimizer.Momentum(
-    #        momentum=0.9,
-    #        learning_rate=0.01,
-    #        regularization=fluid.regularizer.L2Decay(4e-5))
-
     place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
     exe = fluid.Executor(place)
     exe.run(fluid.default_startup_program())
@@ -107,7 +102,10 @@ def compress(args):
         with fluid.program_guard(teacher_program, startup_program):
             img = teacher_program.global_block()._clone_variable(
                 image, force_persistable=False)
-            predict = teacher_model.net(img, class_dim=args.class_dim)
+            predict = teacher_model.net(img,
+                                        class_dim=args.class_dim,
+                                        conv1_name='res_conv1',
+                                        fc_name='res_fc')
         exe.run(startup_program)
         assert args.teacher_pretrained_model and os.path.exists(
             args.teacher_pretrained_model

PaddleSlim/models/resnet.py

@@ -26,7 +26,7 @@ class ResNet():
         self.params = train_parameters
         self.layers = layers
 
-    def net(self, input, class_dim=1000):
+    def net(self, input, class_dim=1000, conv1_name='conv1', fc_name=None):
         layers = self.layers
         supported_layers = [50, 101, 152]
         assert layers in supported_layers, \
@@ -48,7 +48,7 @@ class ResNet():
             filter_size=7,
             stride=2,
             act='relu',
-            name="conv1")
+            name=conv1_name)
         conv = fluid.layers.pool2d(
             input=conv,
             pool_size=3,
@@ -77,6 +77,7 @@ class ResNet():
         out = fluid.layers.fc(input=pool,
                               size=class_dim,
                               act='softmax',
+                              name=fc_name,
                               param_attr=fluid.param_attr.ParamAttr(
                                   initializer=fluid.initializer.Uniform(-stdv,
                                                                         stdv)))

PaddleSlim/quant_low_level_api/run_quant.sh

 #!/usr/bin/env bash
 
+# download pretrain model
+root_url="http://paddle-imagenet-models-name.bj.bcebos.com"
+MobileNetV1="MobileNetV1_pretrained.zip"
+ResNet50="ResNet50_pretrained.zip"
+pretrain_dir='../pretrain'
+
+if [ ! -d ${pretrain_dir} ]; then
+  mkdir ${pretrain_dir}
+fi
+
+cd ${pretrain_dir}
+
+if [ ! -f ${MobileNetV1} ]; then
+    wget ${root_url}/${MobileNetV1}
+    unzip ${MobileNetV1}
+fi
+
+if [ ! -f ${ResNet50} ]; then
+    wget ${root_url}/${ResNet50}
+    unzip ${ResNet50}
+fi
+
+cd -
+
+
 export CUDA_VISIBLE_DEVICES=0,1,2,3
 
 #MobileNet v1:
 python quant.py \
        --model=MobileNet \
-       --pretrained_fp32_model=../data/pretrain/MobileNetV1_pretrained \
+       --pretrained_fp32_model=${pretrain_dir}/MobileNetV1_pretrained \
        --use_gpu=True \
        --data_dir=../data/ILSVRC2012 \
        --batch_size=256 \
@@ -23,7 +48,7 @@ python quant.py \
 #ResNet50:
 #python quant.py \
 #       --model=ResNet50 \
-#       --pretrained_fp32_model=../data/pretrain/ResNet50_pretrained \
+#       --pretrained_fp32_model=${pretrain_dir}/ResNet50_pretrained \
 #       --use_gpu=True \
 #       --data_dir=../data/ILSVRC2012 \
 #       --batch_size=128 \

PaddleSlim/run.sh

@@ -26,13 +26,22 @@ cd -
 
 # for distillation
 #-----------------
-export CUDA_VISIBLE_DEVICES=0
+export CUDA_VISIBLE_DEVICES=0,1,2,3
+
+
+# Fixing name conflicts in distillation
+mv ResNet50_pretrained/conv1_weights ResNet50_pretrained/res_conv1_weights
+mv ResNet50_pretrained/fc_0.w_0 ResNet50_pretrained/res_fc.w_0
+mv ResNet50_pretrained/fc_0.b_0 ResNet50_pretrained/res_fc.b_0
 python compress.py \
 --model "MobileNet" \
 --teacher_model "ResNet50" \
 --teacher_pretrained_model ./pretrain/ResNet50_pretrained \
 --compress_config ./configs/mobilenetv1_resnet50_distillation.yaml
 
+mv ResNet50_pretrained/res_conv1_weights ResNet50_pretrained/conv1_weights
+mv ResNet50_pretrained/res_fc.w_0 ResNet50_pretrained/fc_0.w_0
+mv ResNet50_pretrained/res_fc.b_0 ResNet50_pretrained/fc_0.b_0
 
 # for sensitivity filter pruning
 #-------------------------------
@@ -62,12 +71,21 @@ python compress.py \
 # for distillation with quantization
 #-----------------------------------
 #export CUDA_VISIBLE_DEVICES=0
+#
+## Fixing name conflicts in distillation
+#mv ResNet50_pretrained/conv1_weights ResNet50_pretrained/res_conv1_weights
+#mv ResNet50_pretrained/fc_0.w_0 ResNet50_pretrained/res_fc.w_0
+#mv ResNet50_pretrained/fc_0.b_0 ResNet50_pretrained/res_fc.b_0
+#
 #python compress.py \
 #--model "MobileNet" \
 #--teacher_model "ResNet50" \
 #--teacher_pretrained_model ./data/pretrain/ResNet50_pretrained \
 #--compress_config ./configs/quantization_dist.yaml
-
+#
+#mv ResNet50_pretrained/res_conv1_weights ResNet50_pretrained/conv1_weights
+#mv ResNet50_pretrained/res_fc.w_0 ResNet50_pretrained/fc_0.w_0
+#mv ResNet50_pretrained/res_fc.b_0 ResNet50_pretrained/fc_0.b_0
 
 # for uniform filter pruning with quantization
 #---------------------------------------------