wip

experiments/mnist/flax/train.py

@@ -98,6 +98,13 @@ def train_epoch(state, train_ds, batch_size, rng):
     for perm in perms:
         batch_images = train_ds["image"][perm, ...]
         batch_labels = train_ds["label"][perm, ...]
+
+        # jaxpr = jax.make_jaxpr(apply_and_update_model)(state, batch_images, batch_labels).jaxpr
+        # print(jaxpr)
+
+        # print(set([v.primitive for v in jaxpr.eqns]))
+        # assert False
+
         state, loss, accuracy = apply_and_update_model(state, batch_images, batch_labels)
         epoch_loss.append(loss)
         epoch_accuracy.append(accuracy)
@@ -145,9 +152,9 @@ def train_and_evaluate(config: ml_collections.ConfigDict, workdir: str) -> train
     init_rng = jax.random.PRNGKey(1)
 
     state = create_train_state(init_rng, config)
-    print(state)
-    print(jax.tree_util.tree_flatten(state))
-    state = jsa.as_scaled_array(state)
+    # print(state)
+    # print(jax.tree_util.tree_flatten(state))
+    # state = jsa.as_scaled_array(state)
 
     logging.info("Start Flax MNIST training...")