Refactor graph visualization (#165)

* Move graph_utils.py to relaying

* Add tracing wrapper of YOLOModule

* Fixing lint

* Make IR visualizer as a Class

* Reset the unseen_ops

* Refactor the dot initialization

* Adopt the f-strings

* Make get_node_names explicit

* Adding minor utils

* Move predictions to attributes

* Move self.predictions above the make_graph

* Fix status initialization

* Add get_trace_module unit-test

* Update tvm relay and graph visualize notebooks

* Add more assertions in unit-test

* Add docstring for get_trace_module

Author: zhiqwang

notebooks/export-relay-inference-tvm.ipynb

@@ -0,0 +1,11 @@
+from torch.jit._trace import TopLevelTracedModule
+
+from yolort.models import yolov5s
+from yolort.relaying import get_trace_module
+
+
+def test_get_trace_module():
+    model_func = yolov5s(pretrained=True)
+    script_module = get_trace_module(model_func, input_shape=(416, 320))
+    assert isinstance(script_module, TopLevelTracedModule)
+    assert script_module.code is not None

notebooks/model-graph-visualization.ipynb

@@ -0,0 +1,2 @@
+# Copyright (c) 2021, Zhiqiang Wang. All Rights Reserved.
+from .trace_wrapper import get_trace_module

test/test_relaying.py

@@ -0,0 +1,222 @@
+# Copyright (c) 2021, Zhiqiang Wang
+# Copyright (c) 2020, Thomas Viehmann
+"""
+Visualizing JIT Modules
+
+Modified from https://github.com/t-vi/pytorch-tvmisc/tree/master/hacks
+with license under the CC-BY-SA 4.0.
+
+Please link to Thomas's blog post or the original github source (linked from the
+blog post) with the attribution notice.
+"""
+from collections import OrderedDict
+from graphviz import Digraph
+
+
+class TorchScriptVisualizer:
+    def __init__(self, module):
+
+        self.module = module
+
+        self.unseen_ops = {
+            'prim::ListConstruct', 'prim::ListUnpack',
+            'prim::TupleConstruct', 'prim::TupleUnpack',
+            'aten::Int',
+            'aten::unbind', 'aten::detach',
+            'aten::contiguous', 'aten::to',
+            'aten::unsqueeze', 'aten::squeeze',
+            'aten::index', 'aten::slice', 'aten::select',
+            'aten::constant_pad_nd',
+            'aten::size', 'aten::split_with_sizes',
+            'aten::expand_as', 'aten::expand',
+            'aten::_shape_as_tensor',
+        }
+        # probably also partially absorbing ops. :/
+        self.absorbing_ops = ('aten::size', 'aten::_shape_as_tensor')
+
+    def render(
+        self,
+        classes_to_visit={'YOLO', 'YOLOHead'},
+        format='svg',
+        labelloc='t',
+        attr_size='8,7',
+    ):
+        self.clean_status()
+
+        model_input = next(self.module.graph.inputs())
+        model_type = self.get_node_names(model_input)[-1]
+        dot = Digraph(
+            format=format,
+            graph_attr={'label': model_type, 'labelloc': labelloc},
+        )
+        self.make_graph(self.module, dot=dot, classes_to_visit=classes_to_visit)
+
+        dot.attr(size=attr_size)
+        return dot
+
+    def clean_status(self):
+        self._seen_edges = set()
+        self._seen_input_names = set()
+        self._predictions = OrderedDict()
+
+    @staticmethod
+    def get_node_names(node):
+        return node.type().str().split('.')
+
+    @staticmethod
+    def get_function_name(node):
+        return node.type().__repr__().split('.')[-1]
+
+    def make_graph(self, module, dot=None, parent_dot=None, prefix="", input_preds=None,
+                   classes_to_visit=None, classes_found=None):
+        graph = module.graph
+
+        self_input = next(graph.inputs())
+        self._predictions[self_input] = (set(), set())  # Stand for `input` and `op` respectively
+
+        for nr, i in enumerate(list(graph.inputs())[1:]):
+            name = f'{prefix}input_{i.debugName()}'
+            self._predictions[i] = {name}, set()
+            dot.node(name, shape='ellipse')
+            if input_preds is not None:
+                pred, op = input_preds[nr]
+                self.make_edges(pred, f'input_{name}', name, op, parent_dot)
+
+        for node in graph.nodes():
+            node_inputs = list(node.inputs())
+            only_first_ops = {'aten::expand_as'}
+            rel_inp_end = 1 if node.kind() in only_first_ops else None
+
+            relevant_inputs = [i for i in node_inputs[:rel_inp_end] if is_relevant_type(i.type())]
+            relevant_outputs = [o for o in node.outputs() if is_relevant_type(o.type())]
+
+            if node.kind() == 'prim::CallMethod':
+                node_names = self.get_node_names(node_inputs[0])
+                fq_submodule_name = '.'.join([nc for nc in node_names if not nc.startswith('__')])
+                submodule_type = node_names[-1]
+                submodule_name = find_name(node_inputs[0], self_input)
+                name = f'{prefix}.{node.output().debugName()}'
+                label = f'{prefix}{submodule_name} ({submodule_type})'
+
+                if classes_found is not None:
+                    classes_found.add(fq_submodule_name)
+
+                if ((classes_to_visit is None and (not fq_submodule_name.startswith('torch.nn')
+                    or fq_submodule_name.startswith('torch.nn.modules.container')))
+                    or (classes_to_visit is not None and (submodule_type in classes_to_visit
+                        or fq_submodule_name in classes_to_visit))):
+
+                    # go into subgraph
+                    sub_prefix = f'{prefix}{submodule_name}.'
+
+                    for i, o in enumerate(node.outputs()):
+                        self._predictions[o] = {f'{sub_prefix}output_{i}'}, set()
+
+                    with dot.subgraph(name=f'cluster_{name}') as sub_dot:
+                        sub_dot.attr(label=label)
+                        sub_module = module
+                        for k in submodule_name.split('.'):
+                            sub_module = getattr(sub_module, k)
+
+                        self.make_graph(
+                            sub_module,
+                            dot=sub_dot,
+                            parent_dot=dot,
+                            prefix=sub_prefix,
+                            input_preds=[self._predictions[i] for i in node_inputs[1:]],
+                            classes_to_visit=classes_to_visit,
+                            classes_found=classes_found,
+                        )
+
+                else:
+                    dot.node(name, label=label, shape='box')
+                    for i in relevant_inputs:
+                        pred, op = self._predictions[i]
+                        self.make_edges(pred, prefix + i.debugName(), name, op, dot)
+                    for o in node.outputs():
+                        self._predictions[o] = {name}, set()
+
+            elif node.kind() == 'prim::CallFunction':
+                name = f'{prefix}.{node.output().debugName()}'
+                fun_name = self.get_function_name(node_inputs[0])
+                dot.node(name, label=fun_name, shape='box')
+                for i in relevant_inputs:
+                    pred, op = self._predictions[i]
+                    self.make_edges(pred, prefix + i.debugName(), name, op, dot)
+                for o in node.outputs():
+                    self._predictions[o] = {name}, set()
+
+            else:
+                label = node.kind().split('::')[-1].rstrip('_')
+                pred, op = set(), set()
+                for i in relevant_inputs:
+                    apred, aop = self._predictions[i]
+                    pred |= apred
+                    op |= aop
+
+                if node.kind() in self.absorbing_ops:
+                    pred, op = set(), set()
+                elif (len(relevant_inputs) > 0
+                      and len(relevant_outputs) > 0
+                      and node.kind() not in self.unseen_ops):
+                    op.add(label)
+                for o in node.outputs():
+                    self._predictions[o] = pred, op
+
+        for i, o in enumerate(graph.outputs()):
+            name = f'{prefix}output_{i}'
+            dot.node(name, shape='ellipse')
+            pred, op = self._predictions[o]
+            self.make_edges(pred, f'input_{name}', name, op, dot)
+
+    def add_edge(self, dot, n1, n2):
+        if (n1, n2) not in self._seen_edges:
+            self._seen_edges.add((n1, n2))
+            dot.edge(n1, n2)
+
+    def make_edges(self, preds, input_name, name, op, edge_dot):
+        if len(op) > 0:
+            if input_name not in self._seen_input_names:
+                self._seen_input_names.add(input_name)
+                label_lines = [[]]
+                line_len = 0
+                for w in op:
+                    if line_len >= 20:
+                        label_lines.append([])
+                        line_len = 0
+                    label_lines[-1].append(w)
+                    line_len += len(w) + 1
+
+                edge_dot.node(
+                    input_name,
+                    label='\n'.join([' '.join(w) for w in label_lines]),
+                    shape='box',
+                    style='rounded',
+                )
+                for p in preds:
+                    self.add_edge(edge_dot, p, input_name)
+            self.add_edge(edge_dot, input_name, name)
+        else:
+            for p in preds:
+                self.add_edge(edge_dot, p, name)
+
+
+def find_name(layer_input, self_input, suffix=None):
+    if layer_input == self_input:
+        return suffix
+    cur = layer_input.node().s('name')
+    if suffix is not None:
+        cur = f'{cur}.{suffix}'
+    of = next(layer_input.node().inputs())
+    return find_name(of, self_input, suffix=cur)
+
+
+def is_relevant_type(t):
+    kind = t.kind()
+    if kind == 'TensorType':
+        return True
+    if kind in ('ListType', 'OptionalType'):
+        return is_relevant_type(t.getElementType())
+    if kind == 'TupleType':
+        return any([is_relevant_type(tt) for tt in t.elements()])
+    return False

yolort/relaying/__init__.py

@@ -0,0 +1,63 @@
+# Copyright (c) 2021, Zhiqiang Wang. All Rights Reserved.
+from typing import Dict, Tuple, Callable
+
+import torch
+from torch import nn, Tensor
+
+
+def dict_to_tuple(out_dict: Dict[str, Tensor]) -> Tuple:
+    """
+    Convert the model output dictionary to tuple format.
+    """
+    if "masks" in out_dict.keys():
+        return out_dict["boxes"], out_dict["scores"], out_dict["labels"], out_dict["masks"]
+    return out_dict["boxes"], out_dict["scores"], out_dict["labels"]
+
+
+class TraceWrapper(nn.Module):
+    """
+    This is a wrapper for `torch.jit.trace`, as there are some scenarios
+    where `torch.jit.script` support is limited.
+    """
+    def __init__(self, model):
+        super().__init__()
+        self.model = model
+
+    def forward(self, x):
+        out = self.model(x)
+        return dict_to_tuple(out[0])
+
+
+@torch.no_grad()
+def get_trace_module(
+    model_func: Callable[..., nn.Module],
+    input_shape: Tuple[int, int] = (416, 416),
+):
+    """
+    Get the tarcing of a given model function.
+
+    Example:
+
+        >>> from yolort.models import yolov5s
+        >>> from yolort.relaying.trace_wrapper import get_trace_module
+        >>>
+        >>> model = yolov5s(pretrained=True)
+        >>> tracing_module = get_trace_module(model)
+        >>> print(tracing_module.code)
+        def forward(self,
+            x: Tensor) -> Tuple[Tensor, Tensor, Tensor]:
+          _0, _1, _2, = (self.model).forward(x, )
+          return (_0, _1, _2)
+
+    Args:
+        model_func (Callable): The model function to be traced.
+        input_shape (Tuple[int, int]): Shape size of the input image.
+    """
+    model = TraceWrapper(model_func)
+    model.eval()
+
+    dummy_input = torch.rand(1, 3, *input_shape)
+    trace_module = torch.jit.trace(model, dummy_input)
+    trace_module.eval()
+
+    return trace_module