File size: 3,627 Bytes
67ba414 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 | # Gradient Analysis Plotting
This folder contains the plotting utilities for the gradient-analysis workflow.
There are two plotting entry points:
1. [plot_gradient_analysis.py](./plot_gradient_analysis.py)
- pulls one W&B run directly
- exports local `json` / `csv`
- writes per-step PNG plots
2. [plot_icml_steps.py](./plot_icml_steps.py)
- builds a fixed 3-step comparison figure from already-exported `metrics.json` files
- intended for paper-style summary figures
For the training-side workflow and arguments, see:
- [docs/guide_gradient_analysis.md](../docs/guide_gradient_analysis.md)
Current default training behavior from [config/base.yaml](../config/base.yaml):
- `trainer.gradient_analysis_mode=True`
- `trainer.gradient_analysis_every=50`
- `trainer.gradient_analysis_env_groups=null`
- `trainer.gradient_analysis_group_size=null`
- `trainer.exit_after_gradient_analysis=False`
## Typical Workflow
### 1. Run one analysis job
Example helper runner:
```bash
bash scripts/runs/run_sokoban_ppo_filter_grad_analysis.sh \
--gpus 0,1,2,3,4,5,6,7
```
That job:
- trains for `101` steps
- validates before training and every `10` steps
- runs gradient analysis at steps `1`, `51`, and `101`
- uses a training batch of `8x16`
- uses a separate gradient-analysis batch of `128x16`
### 2. List available analysis steps in W&B
```bash
python gradient_analysis/plot_gradient_analysis.py \
--wandb-path deimos-xing/ragen_gradient_analysis/<run_id> \
--list-steps
```
### 3. Plot all analysis steps from that run
```bash
python gradient_analysis/plot_gradient_analysis.py \
--wandb-path deimos-xing/ragen_gradient_analysis/<run_id>
```
Default output directory:
```text
gradient_analysis_outputs/<run_name>_<run_id>/
```
### 4. Plot only one step
```bash
python gradient_analysis/plot_gradient_analysis.py \
--wandb-path deimos-xing/ragen_gradient_analysis/<run_id> \
--step 1
```
### 5. Choose your own output directory
```bash
python gradient_analysis/plot_gradient_analysis.py \
--wandb-path deimos-xing/ragen_gradient_analysis/<run_id> \
--step 1 \
--output-dir gradient_analysis_outputs/my_custom_dir
```
## Files Produced By `gradient_analysis/plot_gradient_analysis.py`
For each selected step, the script writes:
- `gradient_analysis_summary_step_<N>.png`
- `gradient_analysis_plots_step_<N>.png`
- `gradient_analysis_loss_plots_step_<N>.png`
- `gradient_analysis_reward_std_step_<N>.png`
- `gradient_analysis_normed_grads_step_<N>.png`
- `gradient_analysis_metrics_step_<N>.json`
- `gradient_analysis_bucket_rv_table_step_<N>.csv`
The `metrics.json` export is the bridge to the paper-style plotting script.
## Building A 3-Step Comparison Figure
If you have three exported step directories and want the fixed grid figure:
```bash
python gradient_analysis/plot_icml_steps.py \
--mode ppo \
--step0-dir /path/to/step0 \
--step20-dir /path/to/step20 \
--step40-dir /path/to/step40 \
--out gradient_analysis_outputs/ppo_step0_20_40.png
```
Each step directory must contain:
```text
metrics.json
```
If your exported file is named `gradient_analysis_metrics_step_<N>.json`, copy or rename it to `metrics.json` inside each step directory before calling `plot_icml_steps.py`.
## What To Inspect First
For a new run, start with:
1. `gradient_analysis_summary_step_<N>.png`
2. `gradient_analysis_plots_step_<N>.png`
3. `gradient_analysis_metrics_step_<N>.json`
Those three are usually enough to tell:
- how many buckets were populated
- whether task gradients dominate regularizer gradients
- whether gradient magnitude is monotonic or non-monotonic in reward variance
|