Merge pull request #3 from BarclayII/gq-pytorch

graph draft

graph.py

+import networkx as nx
+import torch as T
+import torch.nn as NN
+
+class DiGraph(nx.DiGraph, NN.Module):
+    '''
+    Reserved attributes:
+    * state: node state vectors during message passing iterations
+        edges does not have "state vectors"; the "state" field is reserved for storing messages
+    * tag: node-/edge-specific feature tensors or other data
+    '''
+    def __init__(self, data=None, **attr):
+        NN.Module.__init__(self)
+        nx.DiGraph.__init__(self, data=data, **attr)
+
+        self.message_funcs = []
+        self.update_funcs = []
+
+    def add_node(self, n, state=None, tag=None, attr_dict=None, **attr):
+        nx.DiGraph.add_node(self, n, state=state, tag=None, attr_dict=attr_dict, **attr)
+
+    def add_nodes_from(self, nodes, state=None, tag=None, **attr):
+        nx.DiGraph.add_nodes_from(self, nodes, state=state, tag=tag, **attr)
+
+    def add_edge(self, u, v, tag=None, attr_dict=None, **attr):
+        nx.DiGraph.add_edge(self, u, v, tag=tag, attr_dict=attr_dict, **attr)
+
+    def add_edges_from(self, ebunch, tag=tag, attr_dict=None, **attr):
+        nx.DiGraph.add_edges_from(self, ebunch, tag=tag, attr_dict=attr_dict, **attr)
+
+    def _nodes_or_all(self, nodes='all'):
+        return self.nodes() if nodes == 'all' else nodes
+
+    def _edges_or_all(self, edges='all'):
+        return self.edges() if edges == 'all' else edges
+
+    def _node_tag_name(self, v):
+        return '(%s)' % v
+
+    def _edge_tag_name(self, u, v):
+        return '(%s, %s)' % (min(u, v), max(u, v))
+
+    def zero_node_state(self, state_dims, batch_size=None, nodes='all'):
+        shape = (
+                [batch_size] + list(state_dims)
+                if batch_size is not None
+                else state_dims
+                )
+        nodes = self._nodes_or_all(nodes)
+
+        for v in nodes:
+            self.node[v]['state'] = T.zeros(shape)
+
+    def init_node_tag_with(self, shape, init_func, dtype=T.float32, nodes='all', args=()):
+        nodes = self._nodes_or_all(nodes)
+
+        for v in nodes:
+            self.node[v]['tag'] = init_func(NN.Parameter(T.zeros(shape, dtype=dtype)), *args)
+            self.register_parameter(self._node_tag_name(v), self.node[v]['tag'])
+
+    def init_edge_tag_with(self, shape, init_func, dtype=T.float32, edges='all', args=()):
+        edges = self._edges_or_all(edges)
+
+        for u, v in edges:
+            self[u][v]['tag'] = init_func(NN.Parameter(T.zeros(shape, dtype=dtype)), *args)
+            self.register_parameter(self._edge_tag_name(u, v), self[u][v]['tag'])
+
+    def remove_node_tag(self, nodes='all'):
+        nodes = self._nodes_or_all(nodes)
+
+        for v in nodes:
+            delattr(self, self._node_tag_name(v))
+            del self.node[v]['tag']
+
+    def remove_edge_tag(self, edges='all'):
+        edges = self._edges_or_all(edges)
+
+        for u, v in edges:
+            delattr(self, self._edge_tag_name(u, v))
+            del self[u][v]['tag']
+
+    def edge_tags(self):
+        for u, v in self.edges():
+            yield self[u][v]['tag']
+
+    def node_tags(self):
+        for v in self.nodes():
+            yield self.node[v]['tag']
+
+    def states(self):
+        for v in self.nodes():
+            yield self.node[v]['state']
+
+    def named_edge_tags(self):
+        for u, v in self.edges():
+            yield ((u, v), self[u][v]['tag'])
+
+    def named_node_tags(self):
+        for v in self.nodes():
+            yield (v, self.node[v]['tag'])
+
+    def named_states(self):
+        for v in self.nodes():
+            yield (v, self.node[v]['state'])
+
+    def register_message_func(self, message_func, edges='all', batched=False):
+        '''
+        batched: whether to do a single batched computation instead of iterating
+        message function: accepts source state tensor and edge tag tensor, and
+        returns a message tensor
+        '''
+        self.message_funcs.append((self._edges_or_all(edges), message_func, batched))
+
+    def register_update_func(self, update_func, nodes='all', batched=False):
+        '''
+        batched: whether to do a single batched computation instead of iterating
+        update function: accepts a node attribute dictionary (including state and tag),
+        and a dictionary of edge attribute dictionaries
+        '''
+        self.update_funcs.append((self._nodes_or_all(nodes), update_func, batched))
+
+    def step(self):
+        # update message
+        for ebunch, f, batched in self.message_funcs:
+            if batched:
+                # FIXME: need to optimize since we are repeatedly stacking and
+                # unpacking
+                source = T.stack([self.node[u]['state'] for u, _ in ebunch])
+                edge_tag = T.stack([self[u][v]['tag'] for u, v in ebunch])
+                message = f(source, edge_tag)
+                for i, (u, v) in enumerate(ebunch):
+                    self[u][v]['state'] = message[i]
+            else:
+                for u, v in ebunch:
+                    self[u][v]['state'] = f(
+                            self.node[u]['state'],
+                            self[u][v]['tag']
+                            )
+
+        # update state
+        # TODO: does it make sense to batch update the nodes?
+        for v, f in self.update_funcs:
+            self.node[v]['state'] = f(self.node[v], self[v])

model.py

+import torch as T
+import torch.nn as NN
+import networkx as nx
+from graph import DiGraph
+
+class TreeGlimpsedClassifier(NN.Module):
+    def __init__(self,
+                 n_children=2,
+                 n_depth=3,
+                 h_dims=128,
+                 node_tag_dims=128,
+                 edge_tag_dims=128,
+                 h_dims=128,
+                 n_classes=10,
+                 steps=5,
+                 ):
+        '''
+        Basic idea:
+        * We detect objects through an undirected graphical model.
+        * The graphical model consists of a balanced tree of latent variables h
+        * Each h is then connected to a bbox variable b and a class variable y
+        * b of the root is fixed to cover the entire canvas
+        * All other h, b and y are updated through message passing
+        * The loss function should be either (not completed yet)
+            * multiset loss, or
+            * maximum bipartite matching (like Order Matters paper)
+        '''
+        NN.Module.__init__(self)
+        self.n_children = n_children
+        self.n_depth = n_depth
+        self.h_dims = h_dims
+        self.node_tag_dims = node_tag_dims
+        self.edge_tag_dims = edge_tag_dims
+        self.h_dims = h_dims
+        self.n_classes = n_classes
+
+        # Create graph of latent variables
+        G = nx.balanced_tree(self.n_children, self.n_depth)
+        nx.relabel_nodes(G,
+                         {i: 'h%d' % i for i in range(self.G.nodes())},
+                         False
+                         )
+        h_nodes_list = G.nodes()
+        for h in h_nodes_list:
+            G.node[h]['type'] = 'h'
+        b_nodes_list = ['b%d' % i for i in range(len(h_nodes_list))]
+        y_nodes_list = ['y%d' % i for i in range(len(h_nodes_list))]
+        hy_edge_list = [(h, y) for h, y in zip(h_nodes_list, y_nodes_list)]
+        hb_edge_list = [(h, b) for h, b in zip(h_nodes_list, y_nodes_list)]
+        yh_edge_list = [(y, h) for y, h in zip(y_nodes_list, h_nodes_list)]
+        bh_edge_list = [(b, h) for b, h in zip(b_nodes_list, h_nodes_list)]
+
+        G.add_nodes_from(b_nodes_list, type='b')
+        G.add_nodes_from(y_nodes_list, type='y')
+        G.add_edges_from(hy_edge_list)
+        G.add_edges_from(hb_edge_list)
+
+        self.G = DiGraph(G)
+        hh_edge_list = [(u, v)
+                        for u, v in self.G.edges()
+                        if self.G.node[u]['type'] == self.G.node[v]['type'] == 'h']
+
+        self.G.init_node_tag_with(node_tag_dims, T.nn.init.uniform_, args=(-.01, .01))
+        self.G.init_edge_tag_with(
+                edge_tag_dims,
+                T.nn.init.uniform_,
+                args=(-.01, .01),
+                edges=hy_edge_list + hb_edge_list + yh_edge_list, bh_edge_list
+                )
+
+        # y -> h.  An attention over embeddings dynamically generated through edge tags
+        self.yh_emb = NN.Sequential(
+                NN.Linear(edge_tag_dims, h_dims),
+                NN.ReLU(),
+                NN.Linear(h_dims, n_classes * h_dims),
+                )
+        self.G.register_message_func(self._y_to_h, edges=yh_edge_list, batched=True)
+
+        # b -> h.  Projects b and edge tag to the same dimension, then concatenates and projects to h
+        self.bh_1 = NN.Linear(self.glimpse.att_params, h_dims)
+        self.bh_2 = NN.Linear(edge_tag_dims, h_dims)
+        self.bh_all = NN.Linear(3 * h_dims, h_dims)
+        self.G.register_message_func(self._b_to_h, edges=bh_edge_list, batched=True)
+
+        # h -> h.  Just passes h itself
+        self.G.register_message_func(self._h_to_h, edges=hh_edge_list, batched=True)
+
+        # h -> b.  Concatenates h with edge tag and go through MLP.
+        # Produces Δb
+        self.hb = NN.Linear(hidden_layers + edge_tag_dims, self.glimpse.att_params)
+        self.G.register_message_func(self._h_to_b, edges=hb_edge_list, batched=True)
+
+        # h -> y.  Concatenates h with edge tag and go through MLP.
+        # Produces Δy
+        self.hy = NN.Linear(hidden_layers + edge_tag_dims, self.n_classes)
+        self.G.register_message_func(self._h_to_y, edges=hy_edge_list, batched=True)
+
+        # b update: just adds the original b by Δb
+        self.G.register_update_func(self._update_b, nodes=b_nodes_list, batched=False)
+
+        # y update: also adds y by Δy
+        self.G.register_update_func(self._update_y, nodes=y_nodes_list, batched=False)
+
+        # h update: simply adds h by the average messages and then passes it through ReLU
+        self.G.register_update_func(self._update_h, nodes=h_nodes_list, batched=False)
+
+    def _y_to_h(self, source, edge_tag):
+        '''
+        source: (n_yh_edges, batch_size, 10) logits
+        edge_tag: (n_yh_edges, edge_tag_dims)
+        '''
+        n_yh_edges, batch_size, _ = source.shape
+
+        if not self._yh_emb_cached:
+            self._yh_emb_cached = True
+            self._yh_emb_w = self.yh_emb(edge_tag)
+            self._yh_emb_w = self._yh_emb_w.reshape(
+                    n_yh_edges, self.n_classes, self.h_dims)
+
+        w = self._yh_emb_w[:, None]
+        w = w.expand(n_yh_edges, batch_size, self.n_classes, self.h_dims)
+        source = source[:, :, None, :]
+        return (F.softmax(source) @ w).reshape(n_yh_edges, batch_size, self.h_dims)
+
+    def _b_to_h(self, source, edge_tag):
+        '''
+        source: (n_bh_edges, batch_size, 6) bboxes
+        edge_tag: (n_bh_edges, edge_tag_dims)
+        '''
+        n_bh_edges, batch_size, _ = source.shape
+        _, nchan, nrows, ncols = self.x.size()
+        source = source.reshape(-1, self.glimpse.att_params)
+
+        m_b = T.relu(self.bh_1(source))
+        m_t = T.relu(self.bh_2(edge_tag))
+        m_t = m_t[:, None, :].expand(n_bh_edges, batch_size, self.h_dims)
+        m_t = m_t.reshape(-1, self.h_dims)
+
+        # glimpse takes batch dimension first, glimpse dimension second.
+        # here, the dimension of @source is n_bh_edges (# of glimpses), then
+        # batch size, so we transpose them
+        g = self.glimpse(self.x, source.transpose(0, 1)).transpose(0, 1)
+        g = g.reshape(n_bh_edges * batch_size, nchan, nrows, ncols)
+        phi = self.cnn(g).reshape(n_bh_edges * batch_size, -1)
+
+        m = self.bh_all(T.cat([m_b, m_t, phi], 1))
+        m = m.reshape(n_bh_edges, batch_size, self.h_dims)
+
+        return m
+
+    def _h_to_h(self, source, edge_tag):
+        return source
+
+    def _h_to_b(self, source, edge_tag):
+        n_hb_edges, batch_size, _ = source.shape
+        edge_tag = edge_tag[:, None]
+        edge_tag = edge_tag.expand(n_hb_edges, batch_size, self.edge_tag_dims)
+        I = T.cat([source, edge_tag], -1).reshape(n_hb_edges * batch_size, -1)
+        b = self.hb(I)
+        return db
+
+    def _h_to_y(self, source, edge_tag):
+        n_hy_edges, batch_size, _ = source.shape
+        edge_tag = edge_tag[:, None]
+        edge_tag = edge_tag.expand(n_hb_edges, batch_size, self.edge_tag_dims)
+        I = T.cat([source, edge_tag], -1).reshape(n_hb_edges * batch_size, -1)
+        y = self.hy(I)
+        return dy
+
+    def _update_b(self, b, b_n):
+        return b['state'] + list(b_n.values())[0]['state']
+
+    def _update_y(self, y, y_n):
+        return y['state'] + list(y_n.values())[0]['state']
+
+    def _update_h(self, h, h_n):
+        m = T.stack([e['state'] for e in h_n]).mean(0)
+        return T.relu(h + m)
+
+    def forward(self, x):
+        batch_size = x.shape[0]
+
+        self.G.zero_node_state(self.h_dims, batch_size, nodes=h_nodes_list)
+        full = self.glimpse.full().unsqueeze(0).expand(batch_size, self.glimpse.att_params)
+        for v in self.G.nodes():
+            if G.node[v]['type'] == 'b':
+                # Initialize bbox variables to cover the entire canvas
+                self.G.node[v]['state'] = full
+
+        self._yh_emb_cached = False
+
+        for t in range(self.steps):
+            self.G.step()
+            # We don't change b of the root
+            self.G.node['b0']['state'] = full