Merge branch 'develop' into pact_clip

demo/quant/pact_quant_aware/README.md

@@ -159,7 +159,7 @@ compiled_train_prog = compiled_train_prog.with_data_parallel(
 
 普通量化：
 ```
-python train.py --model MobileNetV3_large_x1_0 --pretrained_model ./pretrain/MobileNetV3_large_x1_0_ssld_pretrained --checkpoint_dir ./output/MobileNetV3_large_x1_0 --num_epochs 30 --lr 0.0001 --use_pact False
+python train.py --model MobileNetV3_large_x1_0 --pretrained_model ./pretrain/MobileNetV3_large_x1_0_ssld_pretrained --num_epochs 30 --lr 0.0001 --use_pact False
 
 ```
 
@@ -179,7 +179,7 @@ python train.py --model MobileNetV3_large_x1_0 --pretrained_model ./pretrain/Mob
 
 使用PACT量化训练
 ```
-python train.py --model MobileNetV3_large_x1_0 --pretrained_model ./pretrain/MobileNetV3_large_x1_0_ssld_pretrained --checkpoint_dir ./output/MobileNetV3_large_x1_0 --num_epochs 30 --lr 0.0001 --use_pact True --batch_size 128 --lr_strategy=piecewise_decay --step_epochs 20 --l2_decay 1e-5
+python train.py --model MobileNetV3_large_x1_0 --pretrained_model ./pretrain/MobileNetV3_large_x1_0_ssld_pretrained --num_epochs 30 --lr 0.0001 --use_pact True --batch_size 128 --lr_strategy=piecewise_decay --step_epochs 20 --l2_decay 1e-5
 ```
 
 输出结果为

demo/quant/pact_quant_aware/train.py

@@ -55,8 +55,12 @@ add_arg('data',             str, "imagenet",
         "Which data to use. 'mnist' or 'imagenet'")
 add_arg('log_period',       int, 10,
         "Log period in batches.")
-add_arg('checkpoint_dir',         str, "output",
-        "checkpoint save dir")
+add_arg('checkpoint_dir',         str, None,
+        "checkpoint dir")
+add_arg('checkpoint_epoch',         int, None,
+        "checkpoint epoch")
+add_arg('output_dir',         str, "output/MobileNetV3_large_x1_0",
+        "model save dir")
 add_arg('use_pact',          bool, True,
         "Whether to use PACT or not.")
 
@@ -288,6 +292,7 @@ def compress(args):
                 compiled_train_prog,
                 feed=data,
                 fetch_list=[avg_cost.name, acc_top1.name, acc_top5.name])
+
             end_time = time.time()
             loss_n = np.mean(loss_n)
             acc_top1_n = np.mean(acc_top1_n)
@@ -322,24 +327,37 @@ def compress(args):
     # train loop
     best_acc1 = 0.0
     best_epoch = 0
-    for i in range(args.num_epochs):
+
+    start_epoch = 0
+    if args.checkpoint_dir is not None:
+        ckpt_path = args.checkpoint_dir
+        assert args.checkpoint_epoch is not None, "checkpoint_epoch must be set"
+        start_epoch = args.checkpoint_epoch
+        fluid.io.load_persistables(
+            exe, dirname=args.checkpoint_dir, main_program=val_program)
+        start_step = start_epoch * int(
+            math.ceil(float(args.total_images) / args.batch_size))
+        v = fluid.global_scope().find_var('@LR_DECAY_COUNTER@').get_tensor()
+        v.set(np.array([start_step]).astype(np.float32), place)
+
+    for i in range(start_epoch, args.num_epochs):
         train(i, compiled_train_prog)
         acc1 = test(i, val_program)
         fluid.io.save_persistables(
             exe,
-            dirname=os.path.join(args.checkpoint_dir, str(i)),
+            dirname=os.path.join(args.output_dir, str(i)),
             main_program=val_program)
         if acc1 > best_acc1:
             best_acc1 = acc1
             best_epoch = i
             fluid.io.save_persistables(
                 exe,
-                dirname=os.path.join(args.checkpoint_dir, 'best_model'),
+                dirname=os.path.join(args.output_dir, 'best_model'),
                 main_program=val_program)
-    if os.path.exists(os.path.join(args.checkpoint_dir, 'best_model')):
+    if os.path.exists(os.path.join(args.output_dir, 'best_model')):
         fluid.io.load_persistables(
             exe,
-            dirname=os.path.join(args.checkpoint_dir, 'best_model'),
+            dirname=os.path.join(args.output_dir, 'best_model'),
             main_program=val_program)
 
     # 3. Freeze the graph after training by adjusting the quantize

docs/zh_cn/quick_start/quant_aware_tutorial.md

@@ -126,14 +126,14 @@ test(val_quant_program)
 
 ```python
 float_prog, int8_prog = slim.quant.convert(val_quant_program, exe.place, save_int8=True)
-target_vars = [float_prog.global_block().var(name) for name in outputs]
+target_vars = [float_prog.global_block().var(outputs[-1])]
 fluid.io.save_inference_model(dirname='./inference_model/float',
-        feeded_var_names=[var.name for var in inputs],
+        feeded_var_names=[inputs[0].name],
         target_vars=target_vars,
         executor=exe,
         main_program=float_prog)
 fluid.io.save_inference_model(dirname='./inference_model/int8',
-        feeded_var_names=[var.name for var in inputs],
+        feeded_var_names=[inputs[0].name],
         target_vars=target_vars,
         executor=exe,
         main_program=int8_prog)

docs/zh_cn/tutorials/image_classification_mkldnn_quant_tutorial.md

 # Intel CPU上部署量化模型教程
 
-在Intel Casecade Lake机器上（如：Intel(R) Xeon(R) Gold 6271），经过量化和DNNL加速，INT8模型在单线程上性能为FP32模型的3~3.7倍；在Intel SkyLake机器上（如：Intel(R) Xeon(R) Gold 6148），单线程性能为FP32模型的1.5倍，而精度仅有极小下降。图像分类量化的样例教程请参考[图像分类INT8模型在CPU优化部署和预测](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/demo/mkldnn_quant/README.md)。自然语言处理模型的量化请参考[ERNIE INT8 模型精度与性能复现](https://github.com/PaddlePaddle/benchmark/tree/master/Inference/c%2B%2B/ernie/mkldnn)
+在Intel Casecade Lake机器上（如：Intel(R) Xeon(R) Gold 6271），经过量化和DNNL加速，INT8模型在单线程上性能为FP32模型的3~3.7倍；在Intel SkyLake机器上（如：Intel(R) Xeon(R) Gold 6148），单线程性能为FP32模型的1.5倍，而精度仅有极小下降。图像分类量化的样例教程请参考[图像分类INT8模型在CPU优化部署和预测](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/demo/mkldnn_quant/)。自然语言处理模型的量化请参考[ERNIE INT8 模型精度与性能复现](https://github.com/PaddlePaddle/benchmark/tree/master/Inference/c%2B%2B/ernie/mkldnn)
 
 ## 图像分类INT8模型在 Xeon(R) 6271 上的精度和性能
 

paddleslim/models/util.py

@@ -29,4 +29,4 @@ def image_classification(model, image_shape, class_num, use_gpu=False):
         exe = fluid.Executor(place)
         exe.run(fluid.default_startup_program())
     return exe, train_program, val_program, (image, label), (
-        acc_top1.name, acc_top5.name, avg_cost.name)
+        acc_top1.name, acc_top5.name, avg_cost.name, out.name)

paddleslim/prune/group_param.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 
 from ..core import GraphWrapper
-from .prune_walker import conv2d as conv2d_walker
+from .prune_walker import PRUNE_WORKER
 
 __all__ = ["collect_convs"]
 
@@ -55,8 +55,9 @@ def collect_convs(params, graph, visited={}):
         pruned_params = []
         param = graph.var(param)
         conv_op = param.outputs()[0]
-        walker = conv2d_walker(
-            conv_op, pruned_params=pruned_params, visited=visited)
+
+        cls = PRUNE_WORKER.get(conv_op.type())
+        walker = cls(conv_op, pruned_params=pruned_params, visited=visited)
         walker.prune(param, pruned_axis=0, pruned_idx=[0])
         groups.append(pruned_params)
     visited = set()

paddleslim/prune/prune_walker.py

@@ -84,9 +84,7 @@ class PruneWorker(object):
             cls = PRUNE_WORKER.get("default_walker")
         _logger.debug("\nfrom: {}\nto: {}\npruned_axis: {}; var: {}".format(
             self.op, op, pruned_axis, var.name()))
-        walker = cls(op,
-                     pruned_params=self.pruned_params,
-                     visited=self.visited)
+        walker = cls(op, pruned_params=self.pruned_params, visited=self.visited)
         walker.prune(var, pruned_axis, pruned_idx)
 
 
@@ -175,29 +173,8 @@ class conv2d_transpose(PruneWorker):
                 self._prune_op(op, filter_var, 0, pruned_idx)
 
         elif var in self.op.inputs("Filter"):
-            assert pruned_axis in [0, 1]
-
-            self.pruned_params.append((var, pruned_axis, pruned_idx))
-
-            for op in var.outputs():
-                self._prune_op(op, var, pruned_axis, pruned_idx)
-
-            if pruned_axis == 1:
-                if len(self.op.inputs("Bias")) > 0:
-                    self.pruned_params.append(
-                        (self.op.inputs("Bias"), channel_axis, pruned_idx))
-                output_var = self.op.outputs("Output")[0]
-                self._visit(output_var, channel_axis)
-                next_ops = output_var.outputs()
-                for op in next_ops:
-                    self._prune_op(op, output_var, channel_axis, pruned_idx)
-
-            elif pruned_axis == 0:
-                input_var = self.op.inputs("Input")[0]
-                self._visit(input_var, channel_axis)
-                pre_ops = input_var.inputs()
-                for op in pre_ops:
-                    self._prune_op(op, input_var, channel_axis, pruned_idx)
+            _logger.warn("Skip pruning output channels of conv2d_transpose!")
+            return
         elif var in self.op.outputs("Output"):
             assert pruned_axis == channel_axis, "pruned_axis: {}; var: {}".format(
                 pruned_axis, var.name())

tests/test_prune.py

@@ -41,6 +41,9 @@ class TestPrune(unittest.TestCase):
             conv5 = conv_bn_layer(sum2, 8, 3, "conv5")
             conv6 = conv_bn_layer(conv5, 8, 3, "conv6")
 
+            conv7 = fluid.layers.conv2d_transpose(
+                input=conv6, num_filters=16, filter_size=2, stride=2)
+
         shapes = {}
         for param in main_program.global_block().all_parameters():
             shapes[param.name] = param.shape
@@ -53,8 +56,8 @@ class TestPrune(unittest.TestCase):
         main_program, _, _ = pruner.prune(
             main_program,
             scope,
-            params=["conv4_weights"],
-            ratios=[0.5],
+            params=["conv4_weights", "conv2d_transpose_0.w_0"],
+            ratios=[0.5, 0.6],
             place=place,
             lazy=False,
             only_graph=False,
@@ -67,11 +70,12 @@ class TestPrune(unittest.TestCase):
             "conv3_weights": (8, 4, 3, 3),
             "conv4_weights": (4, 8, 3, 3),
             "conv5_weights": (8, 4, 3, 3),
-            "conv6_weights": (8, 8, 3, 3)
+            "conv6_weights": (8, 8, 3, 3),
+            "conv2d_transpose_0.w_0": (8, 16, 2, 2),
         }
 
         for param in main_program.global_block().all_parameters():
-            if "weights" in param.name:
+            if param.name in shapes:
                 print("param: {}; param shape: {}".format(param.name,
                                                           param.shape))
                 self.assertTrue(param.shape == shapes[param.name])