add flops (#3629)

* add flops

configs/runtime.yml

@@ -2,3 +2,4 @@ use_gpu: true
 log_iter: 20
 save_dir: output
 snapshot_epoch: 1
+print_flops: false

docs/tutorials/FAQ.md

@@ -94,3 +94,7 @@ TestDataset:
   !ImageFolder
     anno_path: annotations/instances_val2017.json
 ```
+
+**Q:** 如何打印网络FLOPs？
+
+**A:** 在`configs/runtime.yml`中设置`print_flops: true`，同时需要安装PaddleSlim(比如：pip install paddleslim)，即可打印模型的FLOPs。

ppdet/engine/trainer.py

@@ -325,6 +325,9 @@ class Trainer(object):
             self.cfg.log_iter, fmt='{avg:.4f}')
         self.status['training_staus'] = stats.TrainingStats(self.cfg.log_iter)
 
+        if self.cfg.get('print_flops', False):
+            self._flops(self.loader)
+
         for epoch_id in range(self.start_epoch, self.cfg.epoch):
             self.status['mode'] = 'train'
             self.status['epoch_id'] = epoch_id
@@ -405,6 +408,8 @@ class Trainer(object):
         self._compose_callback.on_epoch_begin(self.status)
         self.status['mode'] = 'eval'
         self.model.eval()
+        if self.cfg.get('print_flops', False):
+            self._flops(loader)
         for step_id, data in enumerate(loader):
             self.status['step_id'] = step_id
             self._compose_callback.on_step_begin(self.status)
@@ -450,6 +455,8 @@ class Trainer(object):
         # Run Infer 
         self.status['mode'] = 'test'
         self.model.eval()
+        if self.cfg.get('print_flops', False):
+            self._flops(loader)
         for step_id, data in enumerate(loader):
             self.status['step_id'] = step_id
             # forward
@@ -587,3 +594,28 @@ class Trainer(object):
                 pass
         paddle.disable_static()
         return pruned_input_spec
+
+    def _flops(self, loader):
+        self.model.eval()
+        try:
+            import paddleslim
+        except Exception as e:
+            logger.warning(
+                'Unable to calculate flops, please install paddleslim, for example: `pip install paddleslim`'
+            )
+            return
+
+        from paddleslim.analysis import dygraph_flops as flops
+        input_data = None
+        for data in loader:
+            input_data = data
+            break
+
+        input_spec = [{
+            "image": input_data['image'][0].unsqueeze(0),
+            "im_shape": input_data['im_shape'][0].unsqueeze(0),
+            "scale_factor": input_data['scale_factor'][0].unsqueeze(0)
+        }]
+        flops = flops(self.model, input_spec) / (1000**3)
+        logger.info(" Model FLOPs : {:.6f}G. (image shape is {})".format(
+            flops, input_data['image'][0].unsqueeze(0).shape))

ppdet/slim/prune.py

@@ -65,7 +65,7 @@ class Pruner(object):
         if self.print_params:
             print_prune_params(model)
 
-        ori_flops = flops(model, input_spec) / 1000
+        ori_flops = flops(model, input_spec) / (1000**3)
         logger.info("FLOPs before pruning: {}GFLOPs".format(ori_flops))
         if self.criterion == 'fpgm':
             pruner = paddleslim.dygraph.FPGMFilterPruner(model, input_spec)
@@ -78,7 +78,7 @@ class Pruner(object):
         for i, param in enumerate(self.pruned_params):
             ratios[param] = pruned_ratios[i]
         pruner.prune_vars(ratios, [0])
-        pruned_flops = flops(model, input_spec) / 1000
+        pruned_flops = flops(model, input_spec) / (1000**3)
         logger.info("FLOPs after pruning: {}GFLOPs; pruned ratio: {}".format(
             pruned_flops, (ori_flops - pruned_flops) / ori_flops))
 

requirements.txt

@@ -14,4 +14,4 @@ lap
 sklearn
 motmetrics
 openpyxl
-decord
\ No newline at end of file
+decord