An example for Timm ViT

examples/hf_transformers/prune_vit.py

@@ -0,0 +1,57 @@
+from transformers import ViTImageProcessor, ViTForImageClassification
+from transformers.models.vit.modeling_vit import ViTSelfAttention
+import torch_pruning as tp
+from PIL import Image
+import requests
+
+url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
+image = Image.open(requests.get(url, stream=True).raw)
+
+processor = ViTImageProcessor.from_pretrained('google/vit-base-patch16-224')
+model = ViTForImageClassification.from_pretrained('google/vit-base-patch16-224')
+example_inputs = processor(images=image, return_tensors="pt")["pixel_values"]
+#outputs = model(example_inputs)
+#logits = outputs.logits
+# model predicts one of the 1000 ImageNet classes
+#predicted_class_idx = logits.argmax(-1).item()
+#print("Predicted class:", model.config.id2label[predicted_class_idx])
+
+print(model)
+imp = tp.importance.MagnitudeImportance(p=2, group_reduction="mean")
+base_macs, base_params = tp.utils.count_ops_and_params(model, example_inputs)
+channel_groups = {}
+
+# All heads should be pruned simultaneously, so we group channels by head.
+for m in model.modules():
+    if isinstance(m, ViTSelfAttention):
+        channel_groups[m.query] = m.num_attention_heads
+        channel_groups[m.key] = m.num_attention_heads
+        channel_groups[m.value] = m.num_attention_heads
+
+pruner = tp.pruner.MagnitudePruner(
+                model, 
+                example_inputs, 
+                global_pruning=False, # If False, a uniform sparsity will be assigned to different layers.
+                importance=imp, # importance criterion for parameter selection
+                iterative_steps=1, # the number of iterations to achieve target sparsity
+                ch_sparsity=0.5,
+                channel_groups=channel_groups,
+                output_transform=lambda out: out.logits.sum(),
+                ignored_layers=[model.classifier],
+            )
+
+for g in pruner.step(interactive=True):
+    #print(g)
+    g.prune()
+
+# Modify the attention head size and all head size aftering pruning
+for m in model.modules():
+    if isinstance(m, ViTSelfAttention):
+        m.attention_head_size = m.query.out_features // m.num_attention_heads
+        m.all_head_size = m.query.out_features
+
+print(model)
+test_output = model(example_inputs)
+pruned_macs, pruned_params = tp.utils.count_ops_and_params(model, example_inputs)
+print("Base MACs: %d G, Pruned MACs: %d G"%(base_macs/1e9, pruned_macs/1e9))
+print("Base Params: %d M, Pruned Params: %d M"%(base_params/1e6, pruned_params/1e6))

examples/hf_transformers/readme.md

@@ -0,0 +1,6 @@
+# Example for HuggingFace ViT
+
+## Pruning
+```bash
+python prune_vit.py
+```

torch_pruning/dependency.py

@@ -634,7 +634,7 @@ def _detect_unwrapped_parameters(self, unwrapped_parameters):
             unwrapped_parameters = []
         unwrapped_detected = list( set(unwrapped_detected) - set([p for (p, _) in unwrapped_parameters]) )
         if len(unwrapped_detected)>0 and self.verbose:
-            warning_str = "Unwrapped parameters detected: {}.\n Torch-Pruning will prune the last non-singleton dimension of a parameter. If you wish to customize this behavior, please provide an unwrapped_parameters argument.".format([_param_to_name[p] for p in unwrapped_detected])
+            warning_str = "Unwrapped parameters detected: {}.\n Torch-Pruning will prune the last non-singleton dimension of these parameters. If you wish to change this behavior, please provide an unwrapped_parameters argument.".format([_param_to_name[p] for p in unwrapped_detected])
             warnings.warn(warning_str)
 
         # set default pruning dim for unwrapped parameters
@@ -782,9 +782,6 @@ def create_node_if_not_exists(grad_fn):
                     module = ops._SplitOp(self._op_id)
                     self._op_id+=1
                 elif "view" in grad_fn.name().lower() or 'reshape' in grad_fn.name().lower():
-                    #if 'reshape' in grad_fn.name().lower():
-                        #print(grad_fn.__dir__())
-                        #print(grad_fn._saved_self_sizes)
                     module = ops._ReshapeOp(self._op_id)
                     self._op_id+=1
                 else: