Feat: learn flax

.gitignore

@@ -3,3 +3,5 @@ t5_*/
 exports/
 modified_t5_model/
 traces/
+ruff.toml
+settings.json

learn_flax/flax_basics.py

@@ -0,0 +1,332 @@
+# ---
+# jupyter:
+#   jupytext:
+#     formats: ipynb,py:percent
+#     text_representation:
+#       extension: .py
+#       format_name: percent
+#       format_version: '1.3'
+#       jupytext_version: 1.16.4
+# ---
+
+# %% [markdown]
+# # Flax Basics
+
+# %% [markdown]
+# # Linear regression with Flax
+# In this example we perform linear regression with gradient descent. It is
+# done by first coming up with a model, generating data from the model (because
+# data generated from known parameters will "match" the model predictions
+# perfectly albeit with a little noise).
+
+
+# %%
+# import packages
+import jax
+from typing import Any, Callable, Sequence
+from jax import random, numpy as jnp
+import flax
+from flax import linen as nn
+
+# %%
+# define model
+# notice there is no input size declaration
+model = nn.Dense(features=5)
+
+# model paramaters & initialization
+key1, key2 = random.split(random.key(0))
+# dummy input to trigger shape inference
+x = random.normal(key1, (10,))
+params = model.init(key2, x)  # initialization call
+
+# check output shapes
+print(jax.tree_util.tree_map(lambda x: x.shape, params))
+
+model.apply(params, x)
+
+
+# %%
+# example of gradient descent
+
+# set problem dimensions
+n_samples: int = 20
+x_dim: int = 10
+y_dim: int = 5
+
+# generate random ground truth W and b
+key = random.key(0)
+k1, k2 = random.split(key)
+W = random.normal(k1, (x_dim, y_dim))
+b = random.normal(k2, (y_dim,))
+# store parameters in frozendict pytree
+true_params = flax.core.freeze({"params": {"bias": b, "kernel": W}})
+print(jax.tree_util.tree_map(lambda x: x.shape, true_params))
+
+# Generate samples with additional noise.
+key_sample, key_noise = random.split(k1)
+x_samples = random.normal(key_sample, (n_samples, x_dim))
+# Wx + b + epsilon (noise)
+y_samples = (
+    jnp.dot(x_samples, W) + b + 0.1 *
+    random.normal(key_noise, (n_samples, y_dim))
+)
+print('x shape:', x_samples.shape, '; y shape:', y_samples.shape)
+
+
+# %%
+# get model output
+@jax.jit
+def mse(params, x_batched, y_batched):
+    # define squared loss for a single pair
+    def squared_error(x, y):
+        # use model.apply to get model_f(x)
+        pred = model.apply(params, x)
+        return jnp.inner(y - pred, y - pred) / 2.0
+    # vmap squared_error function across larger array along axis 0
+    return jnp.mean(jax.vmap(squared_error)(x_batched, y_batched), axis=0)
+
+
+# %%
+
+# perform gradient descent
+learning_rate = 0.3  # Gradient step size.
+print('Loss for "true" W,b: ', mse(true_params, x_samples, y_samples))
+# get the gradient function for backprop
+loss_grad_fn = jax.value_and_grad(mse)
+
+
+@jax.jit
+def update_params(params, learning_rate, grads):
+    params = jax.tree_util.tree_map(
+        lambda p, g: p - learning_rate * g, params, grads)
+    return params
+
+
+for i in range(101):
+    # Perform one gradient update.
+    loss_val, grads = loss_grad_fn(params, x_samples, y_samples)
+    params = update_params(params, learning_rate, grads)
+    if i % 10 == 0:
+        print(f'Loss step {i}: ', loss_val)
+
+# %% [markdown]
+# # Optimization with Optax
+
+# %%
+# init optax object
+import optax
+tx = optax.adam(learning_rate=learning_rate)
+# we initialize the optimizer with model params
+opt_state = tx.init(params)
+loss_grad_fn = jax.value_and_grad(mse)
+
+# %%
+# perform gradient descent with optax optimizer instead of update_params function
+for i in range(101):
+    loss_val, grads = loss_grad_fn(params, x_samples, y_samples)
+    updates, opt_state = tx.update(grads, opt_state)
+    params = optax.apply_updates(params, updates)
+    if i % 10 == 0:
+        print('Loss step {}: '.format(i), loss_val)   
+
+# %% [markdown]
+# Serializing the result - aka saving the model
+from flax import serialization
+bytes_output = serialization.to_bytes(params)
+dict_output = serialization.to_state_dict(params)
+print('Dict output')
+print(dict_output)
+print('Bytes output')
+print(bytes_output)
+
+
+# %%
+# load back from dict output
+# we first get the param structure
+model = nn.Dense(features=5)
+key1, key2 = random.split(random.key(0))
+# dummy input to trigger shape inference
+x = random.normal(key1, (10,))
+params = model.init(key2, x)  # initialization call
+
+# then we get the saved weights
+# from state_dict
+serialization.from_state_dict(params, dict_output)
+print(params)
+# from bytes
+serialization.from_bytes(params, bytes_output)
+print(params)
+
+
+# %% [markdown]
+# Defining your own models
+# A nn.Module subclass is composed of:
+# 
+# 1. collection of data fields (argument parameters)
+# 2. setup() method to setup structures to be called in the forward pass
+# 3. __call__() function that implements the forward pass
+#
+# there is a params structure for the model in the form of a pytree of
+# parameters
+
+# %%
+# module basics
+class ExplicitMLP(nn.Module):
+    # model arguments
+    # we have a list of feature sizes for the dense layer
+    features: Sequence[int]
+
+    def setup(self):
+        self.layers = [nn.Dense(feat) for feat in self.features]
+    
+    def __call__(self, inputs):
+        x = inputs
+        # go through each layer
+        for i, lyr in enumerate(self.layers):
+            x = lyr(x)
+            # add an activating function at the end of each layer
+            if i != len(self.layers) - 1:
+                x = nn.relu(x)
+        return x
+
+# init the model
+key1, key2 = random.split(random.key(117), 2)
+# any shape is ok
+x = random.uniform(key1, (4,4))
+print(x)
+
+# instantiate the model
+model = ExplicitMLP(features=[3,4,5])
+# init model with an rng key and a test data
+params = model.init(key2, x)
+y = model.apply(params, x)
+
+print(
+    'initialized parameter shapes:\n',
+    jax.tree_util.tree_map(jnp.shape, flax.core.unfreeze(params)),
+)
+print('output:\n', y)
+
+# %%
+# using nn.Module with @nn.compact
+
+class SimpleMLP(nn.Module):
+    features: Sequence[int]
+
+    # nn.compact style is declaring submodules inline
+    @nn.compact
+    def __call__(self, inputs):
+        x = inputs
+        for i, feat in enumerate(self.features):
+            x = nn.Dense(feat, name=f'layers_{i}')(x)
+            if i != len(self.features) - 1:
+                x = nn.relu(x)
+            # providing a name is optional though!
+            # the default autonames would be "Dense_0", "Dense_1", ...
+        return x
+
+key1, key2 = random.split(random.key(117), 2)
+x = random.uniform(key1, (4,4))
+
+model = SimpleMLP(features=[3,4,5])
+params = model.init(key2, x)
+y = model.apply(params, x)
+
+print('initialized parameter shapes:\n', jax.tree_util.tree_map(jnp.shape, flax.core.unfreeze(params)))
+print('output:\n', y)
+
+
+# %%
+# creating your own layers
+# this is a guide to show you how to create your own dense layer as an example
+class SimpleDense(nn.Module):
+    features: int
+    kernel_init: Callable = nn.initializers.lecun_normal()
+    bias_init: Callable = nn.initializers.zeros_init()
+
+    @nn.compact
+    def __call__(self, inputs):
+        # utilize the nn.Module.param function to declare a module
+        # * because params are in __call__, shape inference is possible *
+        # the 'W'
+        kernel = self.param(
+            'kernel',  # name
+            self.kernel_init,  # Initialization function
+            (inputs.shape[-1], self.features),  # shape
+        )
+        # 'Wx'
+        y = jnp.dot(inputs, kernel)
+        # 'Wx + b'
+        bias = self.param('bias', self.bias_init, (self.features,))
+        y = y + bias
+        return y
+
+key1, key2 = random.split(random.key(0), 2)
+x = random.uniform(key1, (4,4))
+
+model = SimpleDense(features=3)
+# init_fn takes (prng key, *init args, **init kwargs)
+params = model.init(key2, x)
+y = model.apply(params, x)
+
+print('initialized parameters:\n', params)
+print('output:\n', y)
+
+
+# %%
+# handling state in your module
+# jax does not work well with side-effects and hidden state
+# jax functions are supposed to be pure functions with determinate input outputs
+# we introduce "mutable" to separate the function and the state
+
+# note: there are 2 collections: 
+# 'params' for trainable, 'batch_stats' for non-trainable
+class BiasAdderWithRunningMean(nn.Module):
+    decay: float = 0.99
+
+    @nn.compact
+    def __call__(self, x):
+        # easy pattern to detect if we're initializing via empty variable tree
+        # check for variable 'mean' in the 'batch_stats' collection
+        is_initialized = self.has_variable('batch_stats', 'mean')
+        # declare variable outside model parameters
+        # declare variable mean in the 'batch_stats' collection
+        # ra_mean now means batch_stats->mean
+        ra_mean = self.variable(
+            'batch_stats', 
+            'mean', 
+            lambda s: jnp.zeros(s), 
+            x.shape[1:]  # dim of data after 0th element e.g. (10,5) -> 5
+        )
+        bias = self.param('bias', lambda rng, shape: jnp.zeros(shape), x.shape[1:])
+        if is_initialized:
+            ra_mean.value = (
+                self.decay * ra_mean.value 
+                + (1.0 - self.decay) * jnp.mean(x, axis=0, keepdims=True)
+            )
+
+        return x - ra_mean.value + bias
+
+
+key1, key2 = random.split(random.key(0), 2)
+x = jnp.ones((10,5))
+model = BiasAdderWithRunningMean()
+variables = model.init(key1, x)
+print('initialized variables:\n', variables)
+# setting mutable means you also get the updated_state of mutable
+# hence there is no true hidden state in the function
+# this serves as a escape hatch for implementing pure functions with state
+# state is therefore kept away from the module
+y, updated_state = model.apply(variables, x, mutable=['batch_stats'])
+print('updated state:\n', updated_state)
+
+
+# %%
+for val in [1.0, 2.0, 3.0]:
+    x = val * jnp.ones((10, 5))
+    y, updated_state = model.apply(variables, x, mutable=['batch_stats'])
+    old_state, params = flax.core.pop(variables, 'params')
+    variables = flax.core.freeze({'params': params, **updated_state})
+    print('updated state:\n', updated_state)  # Shows only the mutable part
+
+# %%

t5_jax.py

@@ -96,7 +96,7 @@ split_datasets = load_from_disk(file_path)
 training_size = len(split_datasets['train'])
 # Store some constant
 seed = 117
-num_epochs = 80
+num_epochs = 5
 batch_size = 384  # 384 is the best
 num_train_epochs = num_epochs
 per_device_train_batch_size = batch_size
@@ -314,16 +314,16 @@ def data_loader(rng: jax.random.PRNGKey, dataset: Dataset, batch_size: int, shuf
 
     for idx in batch_idx:
         batch = dataset[idx]
-        batch = {k: np.array(v) for k, v in batch.items()}
+        batch = {k: jnp.array(v) for k, v in batch.items()}
 
         yield batch
 
 
 # %% [markdown]
 # # Model
-
-
-
+#
+#
+#
 
 # %%
 
@@ -442,16 +442,17 @@ def train_step(state, batch, label_smoothing_factor=0.0):
     num_labels = jax.lax.psum(num_labels, "batch")
 
     # true loss = total loss / total samples
-    loss = jax.lax.psum(loss, "batch")
-    loss = jax.tree_util.tree_map(lambda x: x / num_labels, loss)
+    # loss = jax.lax.psum(loss, "batch")
+    # loss = jax.tree_util.tree_map(lambda x: x / num_labels, loss)
 
     # true grad = total grad / total samples
     grad = jax.lax.psum(grad, "batch")
     grad = jax.tree_util.tree_map(lambda x: x / num_labels, grad)
     new_state = state.apply_gradients(grads=grad, dropout_rng=new_dropout_rng)
 
-    metrics = {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}
-    return new_state, metrics
+    # metrics = {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}
+    # return new_state, metrics
+    return new_state
 
 # Define generation function
 max_length = (
@@ -505,19 +506,20 @@ for epoch in epochs:
     for _ in tqdm(range(steps_per_epoch), desc="Training...", position=1, leave=False):
         batch = next(train_loader)
         batch = shard(batch)
-        state, train_metric = p_train_step(state, batch)
-        train_metrics.append(train_metric)
+        state = p_train_step(state, batch)
+        # train_metrics.append(train_metric)
 
     train_time = time.time() - train_start
 
-    train_metric = unreplicate(train_metric)
-    train_metric['loss'].block_until_ready()
+    # train_metric = unreplicate(train_metric)
+    # train_metric['loss'].block_until_ready()
 
 
 
     epochs.write(
-        f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {train_metric['loss']}, "
-        f"Learning Rate:{train_metric['learning_rate']}, "
+        # f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {train_metric['loss']}, "
+        f"Epoch... ({epoch + 1}/{num_epochs} |  "
+        # f"Learning Rate:{train_metric['learning_rate']}, "
         f"Last train time: {train_time})"
     )
 # jax.profiler.stop_trace()

t5_jax_prediction.py

@@ -30,7 +30,7 @@ from typing import Callable, Optional
 import math
 
 # jax.config.update("jax_default_matmul_precision", "tensorfloat32")
-jax.config.update("jax_default_matmul_precision", "high")
+jax.config.update("jax_default_matmul_precision", "bfloat16")
 
 jax.config.update("jax_enable_x64", False)