Model Training - Program Graphs for Code Analysis

Description

Learn variable semantics by running Gated Graph Neural Networks over program graphs using task-specific graph modifications and maximum likelihood training.

Activities in this Phase

Order	Name	Actions
Train GGNN
#1	Train with maximum likelihood	Read the 'TGN-1 Train with maximum likelihood' activity from playbook 'Program Graphs for Code Analysis' (workflow 'Train GGNN'). Follow its guidance when assisting me. VARNAMING: average final SLOT node states; use as GRU initial state to predict variable name as subtoken sequence via MLE. VARMISUSE: compute context c(t)=h(v_SLOT) and usage u(t,v)=h(v_t,v); train argmax of c(t)^T u(t,v) via MLE. Report accuracy (exact match) and F1 (subtoken match) for VARNAMING; accuracy and PR-AUC for VARMISUSE.
#1	Run GGNN propagation	Read the 'TGN-1 Run GGNN propagation' activity from playbook 'Program Graphs for Code Analysis' (workflow 'Train GGNN'). Follow its guidance when assisting me. For each step: compute messages m(v)_k = linear(h(v)) per edge type. Aggregate incoming messages via elementwise summation. Update node state via GRU: h'(v) = GRU(aggregated_messages, h(v)). Repeat for exactly 8 steps — fewer is insufficient; more does not help.
#1	Batch graphs as disconnected components	Read the 'TGN-1 Batch graphs as disconnected components' activity from playbook 'Program Graphs for Code Analysis' (workflow 'Train GGNN'). Follow its guidance when assisting me. Represent each graph as a sparse adjacency list to minimize memory. Combine a minibatch into one large graph with disconnected components by making node IDs unique per batch. Prepare minibatches on a separate CPU thread to avoid GPU stalls during training.
#1	Configure task-specific graph	Read the 'TGN-1 Configure task-specific graph' activity from playbook 'Program Graphs for Code Analysis' (workflow 'Train GGNN'). Follow its guidance when assisting me. For VARNAMING: replace all tokens of the target variable with a special SLOT token. For VARMISUSE: insert a v_SLOT node at the target location connected via all non-variable edges; insert candidate nodes v_t,v for each type-compatible variable connected via LastUse, LastWrite, LastLexicalUse edges. Set an extra bit to 1 for candidate nodes.