diff options
Diffstat (limited to 'train_colab.ipynb')
| -rw-r--r-- | train_colab.ipynb | 52 |
1 files changed, 4 insertions, 48 deletions
diff --git a/train_colab.ipynb b/train_colab.ipynb index 24ae0dc..afae94d 100644 --- a/train_colab.ipynb +++ b/train_colab.ipynb @@ -15,11 +15,7 @@ { "cell_type": "markdown", "metadata": {}, - "source": [ - "# Blazing Eights - Colab GPU Training\n", - "\n", - "Clone repo → Train PPO agent on GPU → Push trained model back to GitHub" - ] + "source": "# Blazing Eights - Colab GPU Training\n\nClone repo → Train PPO agent (CPU collect, GPU update) → Push trained model back to GitHub\n\n**Game**: UNO variant with custom special cards (8=Wild, K=All draw, J=Skip, Swap=Swap hands)." }, { "cell_type": "markdown", @@ -58,30 +54,14 @@ { "cell_type": "code", "metadata": {}, - "source": [ - "# 2-player training: GPU makes the PPO update faster\n", - "!python train.py --num_players 2 --episodes 200000 --save_path blazing_ppo_2p" - ], + "source": "# 2-player training with greedy warmup + CSV logging\n# Game simulation on CPU, PPO updates on GPU automatically\n!python train.py --num_players 2 --episodes 200000 --save_path blazing_ppo_2p\n\n# Show training log\nimport pandas as pd\ndf = pd.read_csv(\"blazing_ppo_2p_log.csv\")\nprint(df.to_string(index=False))", "execution_count": null, "outputs": [] }, { "cell_type": "code", "metadata": {}, - "source": [ - "# (Optional) 3-player training\n", - "# !python train.py --num_players 3 --episodes 300000 --save_path blazing_ppo_3p" - ], - "execution_count": null, - "outputs": [] - }, - { - "cell_type": "code", - "metadata": {}, - "source": [ - "# (Optional) 4-player training\n", - "# !python train.py --num_players 4 --episodes 400000 --lr 1e-4 --ent_coef 0.02 --save_path blazing_ppo_4p" - ], + "source": "# (Optional) 3-player training\n# !python train.py --num_players 3 --episodes 300000 --save_path blazing_ppo_3p\n\n# (Optional) Skip greedy warmup\n# !python train.py --num_players 2 --episodes 200000 --greedy_warmup 0 --save_path blazing_ppo_2p_no_warmup", "execution_count": null, "outputs": [] }, @@ -163,31 +143,7 @@ { "cell_type": "code", "metadata": {}, - "source": [ - "import sys\n", - "sys.path.insert(0, \".\")\n", - "from train import PolicyValueNet, evaluate_vs_random\n", - "\n", - "device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n", - "model = PolicyValueNet().to(device)\n", - "\n", - "# Load the trained model\n", - "import glob\n", - "final_models = glob.glob(\"*_final.pt\") + glob.glob(\"models/*_final.pt\")\n", - "if final_models:\n", - " ckpt = torch.load(final_models[0], map_location=device, weights_only=True)\n", - " model.load_state_dict(ckpt[\"model\"])\n", - " model.eval()\n", - " print(f\"Loaded: {final_models[0]}\")\n", - " print(f\"Trained for {ckpt.get('episode', '?')} episodes\")\n", - " print()\n", - "\n", - " for n in [2, 3, 4]:\n", - " wr = evaluate_vs_random(model, num_players=n, num_games=2000, device=device)\n", - " print(f\" {n} players: win rate = {wr:.1%} (random baseline: {1/n:.1%})\")\n", - "else:\n", - " print(\"No model found. Train first!\")" - ], + "source": "import sys\nsys.path.insert(0, \".\")\nfrom train import PolicyValueNet, evaluate_vs_random\n\ndevice = \"cpu\" # eval on CPU (single-sample inference)\nmodel = PolicyValueNet().to(device)\n\nimport glob\nfinal_models = glob.glob(\"*_final.pt\") + glob.glob(\"models/*_final.pt\")\nif final_models:\n ckpt = torch.load(final_models[0], map_location=device, weights_only=True)\n model.load_state_dict(ckpt[\"model\"])\n model.eval()\n print(f\"Loaded: {final_models[0]}\")\n print(f\"Trained for {ckpt.get('episode', '?')} episodes\")\n print()\n\n for n in [2, 3, 4]:\n wr = evaluate_vs_random(model, num_players=n, num_games=2000, device=device)\n print(f\" {n} players: win rate = {wr:.1%} (random baseline: {1/n:.1%})\")\nelse:\n print(\"No model found. Train first!\")", "execution_count": null, "outputs": [] } |