Merge pull request #4 from BarclayII/gq-pytorch

more changes

graph.py

 import networkx as nx
 import torch as T
 import torch.nn as NN
+from util import *
 
 class DiGraph(nx.DiGraph, NN.Module):
     '''
@@ -25,7 +26,7 @@ class DiGraph(nx.DiGraph, NN.Module):
     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):
+    def add_edges_from(self, ebunch, tag=None, 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'):
@@ -49,20 +50,20 @@ class DiGraph(nx.DiGraph, NN.Module):
         nodes = self._nodes_or_all(nodes)
 
         for v in nodes:
-            self.node[v]['state'] = T.zeros(shape)
+            self.node[v]['state'] = tovar(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.node[v]['tag'] = init_func(NN.Parameter(tovar(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[u][v]['tag'] = init_func(NN.Parameter(tovar(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'):
@@ -115,7 +116,7 @@ class DiGraph(nx.DiGraph, NN.Module):
         '''
         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
+        and a list of tuples (source node, target node, edge attribute dictionary)
         '''
         self.update_funcs.append((self._nodes_or_all(nodes), update_func, batched))
 
@@ -126,7 +127,11 @@ class DiGraph(nx.DiGraph, NN.Module):
                 # 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])
+                edge_tags = [self[u][v]['tag'] for u, v in ebunch]
+                if all(t is None for t in edge_tags):
+                    edge_tag = None
+                else:
+                    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]
@@ -139,5 +144,6 @@ class DiGraph(nx.DiGraph, NN.Module):
 
         # 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])
+        for vbunch, f, batched in self.update_funcs:
+            for v in vbunch:
+                self.node[v]['state'] = f(self.node[v], self.in_edges(v, data=True))

model.py

+
 import torch as T
 import torch.nn as NN
-import networkx as nx
+import torch.nn.init as INIT
+import torch.nn.functional as F
+import numpy as NP
+import numpy.random as RNG
+from util import *
+from glimpse import create_glimpse
+from zoneout import ZoneoutLSTMCell
+from collections import namedtuple
+import os
 from graph import DiGraph
+import networkx as nx
+
+no_msg = os.getenv('NOMSG', False)
+
+def build_cnn(**config):
+    cnn_list = []
+    filters = config['filters']
+    kernel_size = config['kernel_size']
+    in_channels = config.get('in_channels', 3)
+    final_pool_size = config['final_pool_size']
+
+    for i in range(len(filters)):
+        module = NN.Conv2d(
+            in_channels if i == 0 else filters[i-1],
+            filters[i],
+            kernel_size,
+            padding=tuple((_ - 1) // 2 for _ in kernel_size),
+            )
+        INIT.xavier_uniform(module.weight)
+        INIT.constant(module.bias, 0)
+        cnn_list.append(module)
+        if i < len(filters) - 1:
+            cnn_list.append(NN.LeakyReLU())
+    cnn_list.append(NN.AdaptiveMaxPool2d(final_pool_size))
+
+    return NN.Sequential(*cnn_list)
+
 
 class TreeGlimpsedClassifier(NN.Module):
     def __init__(self,
@@ -10,9 +46,13 @@ class TreeGlimpsedClassifier(NN.Module):
                  h_dims=128,
                  node_tag_dims=128,
                  edge_tag_dims=128,
-                 h_dims=128,
                  n_classes=10,
                  steps=5,
+                 filters=[16, 32, 64, 128, 256],
+                 kernel_size=(3, 3),
+                 final_pool_size=(2, 2),
+                 glimpse_type='gaussian',
+                 glimpse_size=(15, 15),
                  ):
         '''
         Basic idea:
@@ -33,20 +73,30 @@ class TreeGlimpsedClassifier(NN.Module):
         self.edge_tag_dims = edge_tag_dims
         self.h_dims = h_dims
         self.n_classes = n_classes
+        self.glimpse = create_glimpse(glimpse_type, glimpse_size)
+        self.steps = steps
+
+        self.cnn = build_cnn(
+                filters=filters,
+                kernel_size=kernel_size,
+                final_pool_size=final_pool_size,
+                )
 
         # 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())},
+                         {i: 'h%d' % i for i in range(len(G.nodes()))},
                          False
                          )
-        h_nodes_list = G.nodes()
+        self.h_nodes_list = 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))]
+        self.b_nodes_list = b_nodes_list
+        self.y_nodes_list = y_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)]
+        hb_edge_list = [(h, b) for h, b in zip(h_nodes_list, b_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)]
 
@@ -65,21 +115,22 @@ class TreeGlimpsedClassifier(NN.Module):
                 edge_tag_dims,
                 T.nn.init.uniform_,
                 args=(-.01, .01),
-                edges=hy_edge_list + hb_edge_list + yh_edge_list, bh_edge_list
+                edges=hy_edge_list + hb_edge_list + bh_edge_list
+                )
+        self.G.init_edge_tag_with(
+                h_dims * n_classes,
+                T.nn.init.uniform_,
+                args=(-.01, .01),
+                edges=yh_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.bh_all = NN.Linear(2 * h_dims + filters[-1] * NP.prod(final_pool_size), h_dims)
         self.G.register_message_func(self._b_to_h, edges=bh_edge_list, batched=True)
 
         # h -> h.  Just passes h itself
@@ -87,12 +138,12 @@ class TreeGlimpsedClassifier(NN.Module):
 
         # 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.hb = NN.Linear(h_dims + 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.hy = NN.Linear(h_dims + 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
@@ -111,13 +162,7 @@ class TreeGlimpsedClassifier(NN.Module):
         '''
         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 = edge_tag.reshape(n_yh_edges, 1, self.n_classes, self.h_dims)
         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)
@@ -128,10 +173,11 @@ class TreeGlimpsedClassifier(NN.Module):
         edge_tag: (n_bh_edges, edge_tag_dims)
         '''
         n_bh_edges, batch_size, _ = source.shape
+        # FIXME: really using self.x is a bad design here
         _, nchan, nrows, ncols = self.x.size()
-        source = source.reshape(-1, self.glimpse.att_params)
+        _source = source.reshape(-1, self.glimpse.att_params)
 
-        m_b = T.relu(self.bh_1(source))
+        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)
@@ -140,9 +186,12 @@ class TreeGlimpsedClassifier(NN.Module):
         # 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)
+        grows, gcols = g.size()[-2:]
+        g = g.reshape(n_bh_edges * batch_size, nchan, grows, gcols)
         phi = self.cnn(g).reshape(n_bh_edges * batch_size, -1)
 
+        # TODO: add an attribute (g) to h
+
         m = self.bh_all(T.cat([m_b, m_t, phi], 1))
         m = m.reshape(n_bh_edges, batch_size, self.h_dims)
 
@@ -156,40 +205,59 @@ class TreeGlimpsedClassifier(NN.Module):
         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
+        db = self.hb(I)
+        return db.reshape(n_hb_edges, batch_size, -1)
 
     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
+        edge_tag = edge_tag.expand(n_hy_edges, batch_size, self.edge_tag_dims)
+        I = T.cat([source, edge_tag], -1).reshape(n_hy_edges * batch_size, -1)
+        dy = self.hy(I)
+        return dy.reshape(n_hy_edges, batch_size, -1)
 
     def _update_b(self, b, b_n):
-        return b['state'] + list(b_n.values())[0]['state']
+        return b['state'] + b_n[0][2]['state']
 
     def _update_y(self, y, y_n):
-        return y['state'] + list(y_n.values())[0]['state']
+        return y['state'] + y_n[0][2]['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)
+        m = T.stack([e[2]['state'] for e in h_n]).mean(0)
+        return T.relu(h['state'] + m)
 
-    def forward(self, x):
+    def forward(self, x, y=None):
+        self.x = x
         batch_size = x.shape[0]
 
-        self.G.zero_node_state(self.h_dims, batch_size, nodes=h_nodes_list)
+        self.G.zero_node_state((self.h_dims,), batch_size, nodes=self.h_nodes_list)
+        self.G.zero_node_state((self.n_classes,), batch_size, nodes=self.y_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':
+            if self.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
+
+        self.y_pre = T.stack(
+                [self.G.node['y%d' % i]['state'] for i in range(self.n_nodes - 1, self.n_nodes - self.n_leaves - 1, -1)],
+                1
+                )
+        self.v_B = T.stack(
+                [self.G.node['b%d' % i]['state'] for i in range(self.n_nodes)],
+                1,
+                )
+        self.y_logprob = F.log_softmax(self.y_pre)
+        return self.G.node['h0']['state']
+
+    @property
+    def n_nodes(self):
+        return (self.n_children ** self.n_depth - 1) // (self.n_children - 1)
+
+    @property
+    def n_leaves(self):
+        return self.n_children ** (self.n_depth - 1)