env trained after refactor. script run.py made for easy operations. p…

…ython run.py <env> t for train, e for eval, e -w for extract nn weights and load into .c file, then compile locally and run.

config/ocean/blastar.ini

@@ -8,26 +8,26 @@ rnn_name = Recurrent
 num_envs = 4096
 
 [train]
-anneal_lr = False # 4
-batch_size = 65536 # 253
-bptt_horizon = 4
-checkpoint_interval = 200
+anneal_lr = False
+batch_size = 131072 # 65536
+bptt_horizon = 8
+checkpoint_interval = 600
 clip_coef = 0.2
 clip_vloss = True
-compile = False
+compile = True
 compile_mode = reduce-overhead
 cpu_offload = False
 data_dir = experiments
 device = cuda
 ent_coef = 0.002511261007200052
 env_batch_size = 1
 gae_lambda = 0.9212875655241286
-gamma = 0.9310283509696092
-learning_rate = 0.0004700459984905535
+gamma = 0.9410283509696092
+learning_rate = 0.0010700459984905535
 max_grad_norm = 0.9702296257019044
-minibatch_size = 4096 # 23
-norm_adv = True #5 35
-num_envs = 1 # roihj
+minibatch_size = 4096
+norm_adv = True
+num_envs = 1
 num_workers = 1
 seed = 1
 torch_deterministic = True

pufferlib/ocean/blastar/blastar.c

@@ -1,62 +1,71 @@
-// blastar.c
-#include "blastar_env.h"
-#include "blastar_renderer.h"
-#include "puffernet.h"
+#include "blastar.h"
+
 #include <assert.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <time.h>
 
+#include "puffernet.h"
+
+const char* WEIGHTS_PATH = "/home/daa/pufferlib_testbench/PufferLib/pufferlib/resources/blastar/blastar_weights.bin";
 #define OBSERVATIONS_SIZE 31
 #define ACTIONS_SIZE 6
+#define NUM_WEIGHTS 137095
 
 void get_input(BlastarEnv* env) {
-    if ((IsKeyDown(KEY_DOWN) && IsKeyDown(KEY_RIGHT))) {
-        env->actions[0] = 4; // Move down-right
-    } else if ((IsKeyDown(KEY_DOWN) && IsKeyDown(KEY_LEFT))) {
-        env->actions[0] = 5; // Move down-left
-    } else if (IsKeyDown(KEY_SPACE) && (IsKeyDown(KEY_RIGHT))) {
-        env->actions[0] = 6; // Fire and move right
-    } else if (IsKeyDown(KEY_SPACE) && (IsKeyDown(KEY_LEFT))) {
-        env->actions[0] = 7; // Fire and move left
-    } else if (IsKeyDown(KEY_SPACE)) {
-        env->actions[0] = 3; // Fire
-    } else if (IsKeyDown(KEY_DOWN)) {
-        env->actions[0] = 2; // Move down
-    } else if (IsKeyDown(KEY_LEFT)) {
-        env->actions[0] = 1; // Move left
-    } else if (IsKeyDown(KEY_RIGHT)) {
-        env->actions[0] = 0; // Move right
-    } else {
-        env->actions[0] = 0; // No action
+    // Left
+    if (IsKeyDown(KEY_LEFT) || IsKeyDown(KEY_A)) {
+        env->actions[0] = 1;
+    }
+    // Right
+    else if (IsKeyDown(KEY_RIGHT) || IsKeyDown(KEY_D)) {
+        env->actions[0] = 2;
+    }
+    // Up
+    else if (IsKeyDown(KEY_UP) || IsKeyDown(KEY_W)) {
+        env->actions[0] = 3;
+    }
+    // Down
+    else if (IsKeyDown(KEY_DOWN) || IsKeyDown(KEY_S)) {
+        env->actions[0] = 4;
+    }
+    // Fire
+    else if (IsKeyDown(KEY_SPACE)) {
+        env->actions[0] = 5;
+    }
+    // No action
+    else {
+        env->actions[0] = 0;
     }
 }
 
 int demo() {
-    // Load weights for the AI model
-    Weights* weights = load_weights("./pufferlib/resources/blastar/blastar_weights.bin", 137095);
-    LinearLSTM* net = make_linearlstm(weights, 1, OBSERVATIONS_SIZE, ACTIONS_SIZE);
+    Weights* weights = load_weights(WEIGHTS_PATH, NUM_WEIGHTS);
+    LinearLSTM* net =
+        make_linearlstm(weights, 1, OBSERVATIONS_SIZE, ACTIONS_SIZE);
 
-    BlastarEnv env;
-    init_blastar(&env);
-    allocate_env(&env);
+    BlastarEnv env = {
+        .player.x = SCREEN_WIDTH / 2,
+        .player.y = SCREEN_HEIGHT - player_height,
+    };
+    allocate(&env);
 
     Client* client = make_client(&env);
 
     unsigned int seed = 12345;
     srand(seed);
-    reset_blastar(&env);
+    reset(&env);
 
     int running = 1;
     while (running) {
         if (IsKeyDown(KEY_LEFT_SHIFT)) {
-            get_input(&env); // Human input
+            get_input(&env);  // Human input
         } else {
-            forward_linearlstm(net, env.observations, env.actions); // AI input
+            forward_linearlstm(net, env.observations, env.actions);  // AI input
         }
 
         c_step(&env);
-        c_render(client, &env);
+        render(client, &env);
 
         if (WindowShouldClose() || env.game_over) {
             running = 0;
@@ -66,33 +75,31 @@ int demo() {
     free_linearlstm(net);
     free(weights);
     close_client(client);
-    free_allocated_env(&env);
-    close_blastar(&env);
+    free_allocated(&env);
     return 0;
 }
 
 void perftest(float test_time) {
     BlastarEnv env;
-    init_blastar(&env);
-    allocate_env(&env);
+    init(&env);
+    allocate(&env);
 
     unsigned int seed = 12345;
     srand(seed);
-    reset_blastar(&env);
+    reset(&env);
 
     int start = time(NULL);
     int steps = 0;
     while (time(NULL) - start < test_time) {
-        env.actions[0] = rand() % ACTIONS_SIZE; // Random actions
+        env.actions[0] = rand() % ACTIONS_SIZE;  // Random actions
         c_step(&env);
         steps++;
     }
 
     int end = time(NULL);
     printf("Steps per second: %f\n", steps / (float)(end - start));
 
-    free_allocated_env(&env);
-    close_blastar(&env);
+    free_allocated(&env);
 }
 
 int main() {