使用 Grain 高效加载数据：JAX/Flax NNX 练习

# Environment Setup
# This cell configures the environment to simulate multiple CPU devices
# and installs necessary libraries.
# IMPORTANT: RUN THIS CELL FIRST. If you encounter issues with JAX device
# counts later, try 'Runtime -> Restart runtime' in the Colab menu
# and run this cell again before any others.
import os

# Configure JAX to see 8 virtual CPU devices.
# This must be done before JAX is imported for the first time in a session.
os.environ["XLA_FLAGS"] = "--xla_force_host_platform_device_count=8"

# Install libraries
!pip install -q jax-ai-stack==2025.9.3
print("Libraries installed.")

# Now, import chex and attempt to set_n_cpu_devices.
# This must be called after setting XLA_FLAGS and before JAX initializes its backends.
import chex

try:
  chex.set_n_cpu_devices(8)
  print("chex.set_n_cpu_devices(8) called successfully.")
except RuntimeError as e:
  print(f"Note on chex.set_n_cpu_devices: {e}")
  print("This usually means JAX was already initialized. The XLA_FLAGS environment variable should still apply.")
  print("If you see issues with device counts, ensure you 'Restart runtime' and run this cell first.")

# Verify JAX device count
import jax
print(f"JAX version: {jax.__version__}")
print(f"JAX found {jax.device_count()} devices.")
print(f"JAX devices: {jax.devices()}")

if jax.device_count() != 8:
  print("\nWARNING: JAX does not see 8 devices. Parallelism/sharding exercises might not behave as expected.")
  print("Please try 'Runtime -> Restart runtime' and run this setup cell again first.")

# Common imports for the exercises
import jax.numpy as jnp
import numpy as np
import grain.python as grain # Main Grain API
from flax import nnx # Flax NNX API
import time # For simulating work and observing performance
import copy # For checkpointing example
from typing import Dict, Any, List # For type hints
import dataclasses # For ShardOptions if needed manually
import functools # For functools.partial
print("Imports complete. Setup finished.")

# @title Exercise 1: Student Code
# 1. Define MySource
class MySource(grain.RandomAccessDataSource):
  def __init__(self, num_records: int = 1000):
    self._num_records = num_records
  def __len__(self) -> int:
      # TODO: Return the total number of records
      # YOUR CODE HERE
      return 0 # Replace this

  def __getitem__(self, idx: int) -> Dict[str, Any]:
      # TODO: Handle potential index wrap-around for multiple epochs
      # effective_idx = ...
      # YOUR CODE HERE
      effective_idx = idx # Replace this

      # TODO: Simulate loading data: an image and a label
      # image = np.ones(...) * (effective_idx % 255)
      # label = effective_idx % 10
      # YOUR CODE HERE
      image = np.zeros((32,32,3), dtype=np.uint8) # Replace this
      label = 0 # Replace this
      return {'image': image, 'label': label}

# 2. Instantiate MySource
# TODO: Create an instance of MySource
# source = ...
# YOUR CODE HERE
source = None # Replace this

# 3. Create an IndexSampler
# TODO: Create an IndexSampler that shuffles, runs for 1 epoch, and uses seed 42.
# num_records should be len(source).
# index_sampler = grain.IndexSampler(...)
# YOUR CODE HERE
index_sampler = None # Replace this

# 4. Define Operations
# TODO: Define ConvertToFloat transform
class ConvertToFloat(grain.MapTransform):
  def map(self, features: Dict[str, Any]) -> Dict[str, Any]:
    # TODO: Convert 'image' to float32 and normalize to [0, 1].
    # Keep 'label' as is.
    # YOUR CODE HERE
    image = features['image'] # Replace this
    return {'image': image.astype(np.float32) / 255.0, 'label': features['label']}

# TODO: Create a list of transformations: ConvertToFloat instance, then grain.Batch
batch_size = 64, drop_remainder = True
# transformations = [...]
# YOUR CODE HERE
transformations = [] # Replace this

# 5. Instantiate DataLoader
# TODO: Create a DataLoader with worker_count=0 and appropriate read_options
# data_loader_sequential = grain.DataLoader(...)
# YOUR CODE HERE
data_loader_sequential = None # Replace this

# 6. Iterate and print batch info
if data_loader_sequential:
  print("DataLoader configured sequentially.")
  data_iterator_seq = iter(data_loader_sequential)
  try:
    first_batch_seq = next(data_iterator_seq)
    print(f"Sequential - First batch image shape: {first_batch_seq['image'].shape}")
    print(f"Sequential - First batch label shape: {first_batch_seq['label'].shape}")
    # Example: Check a value from the first image of the first batch
    print(f"Sequential - Example image value (first item, [0,0,0]): {first_batch_seq['image'][0, 0, 0, 0]}")
    print(f"Sequential - Example label value (first item): {first_batch_seq['label'][0]}")
  except StopIteration:
    print("Sequential DataLoader is empty or exhausted.")
else:
  print("Sequential DataLoader not configured yet.")

# @title Exercise 1: Solution
# 1. Define MySource
class MySource(grain.RandomAccessDataSource):
  def __init__(self, num_records: int = 1000):
    self._num_records = num_records

  def __len__(self) -> int:
    return self._num_records

  def __getitem__(self, idx: int) -> Dict[str, Any]:
    effective_idx = idx % self._num_records # Handle wrap-around
    image = np.ones((32, 32, 3), dtype=np.uint8) * (effective_idx % 255)
    label = effective_idx % 10
    return {'image': image, 'label': label}

# 2. Instantiate MySource
source = MySource(num_records=1000)
print(f"DataSource created with {len(source)} records.")

# 3. Create an IndexSampler
index_sampler = grain.IndexSampler(
    num_records=len(source),
    shard_options=grain.NoSharding(), # No sharding for this exercise
    shuffle=True,
    num_epochs=1, # Run for 1 epoch
    seed=42
    )
print("IndexSampler created.")

# 4. Define Operations
class ConvertToFloat(grain.MapTransform):
  def map(self, features: Dict[str, Any]) -> Dict[str, Any]:
    # Convert 'image' to float32 and normalize to [0, 1].
    # Keep 'label' as is.
    image = features['image'].astype(np.float32) / 255.0
    return {'image': image, 'label': features['label']}

transformations = [
    ConvertToFloat(),
    grain.Batch(batch_size=64, drop_remainder=True)
    ]
print("Transformations defined.")

# 5. Instantiate DataLoader
data_loader_sequential = grain.DataLoader(
    data_source=source,
    operations=transformations,
    sampler=index_sampler,
    worker_count=0, # Sequential mode
    shard_options=grain.NoSharding(), # Explicitly no sharding for this loader instance
    read_options=grain.ReadOptions(num_threads=0) # Dataset is in-memory
)

# 6. Iterate and print batch info
if data_loader_sequential:
  print("DataLoader configured sequentially.")
  data_iterator_seq = iter(data_loader_sequential)
  try:
    first_batch_seq = next(data_iterator_seq)
    print(f"Sequential - First batch image shape: {first_batch_seq['image'].shape}") # Expected: (64, 32, 32, 3)
    print(f"Sequential - First batch label shape: {first_batch_seq['label'].shape}") # Expected: (64,)
    # Example: Check a value from the first image of the first batch
    print(f"Sequential - Example image value (first item, [0,0,0]): {first_batch_seq['image'][0, 0, 0, 0]}")
    print(f"Sequential - Example label value (first item): {first_batch_seq['label'][0]}")
  except StopIteration:
    print("Sequential DataLoader is empty or exhausted.")
else:
  print("Sequential DataLoader not configured yet.")

# @title Exercise 2: Student Code
# 1. Reuse/Recreate components (DataSource with simulated work, Sampler, Operations)
# TODO: Define MySourceWithWork, adding time.sleep(0.01) in getitem
class MySourceWithWork(grain.RandomAccessDataSource):
  def __init__(self, num_records: int = 1000):
    self._num_records = num_records

  def __len__(self) -> int:
    # YOUR CODE HERE
    return self._num_records

  def __getitem__(self, idx: int) -> Dict[str, Any]:
    effective_idx = idx % self._num_records
    # TODO: Add time.sleep(0.01) to simulate work
    # YOUR CODE HERE

    image = np.ones((32, 32, 3), dtype=np.uint8) * (effective_idx % 255)
    label = effective_idx % 10
    return {'image': image, 'label': label}

# TODO: Instantiate MySourceWithWork
# source_with_work = ...
# YOUR CODE HERE
source_with_work = None # Replace this

# TODO: Create a new IndexSampler (e.g., for indefinite epochs, num_epochs=None)
# Or reuse the one from Ex1 if you reset it or it's for multiple epochs.
# For simplicity, let's create one for indefinite epochs.
# parallel_sampler = grain.IndexSampler(...)
# YOUR CODE HERE
parallel_sampler = None # Replace this

# Transformations can be reused from Exercise 1
# transformations = [ConvertToFloat(), grain.Batch(batch_size=64, drop_remainder=True)]
# (Assuming ConvertToFloat is defined from Ex1 solution)

# 2. Instantiate DataLoader with worker_count > 0
# TODO: Set num_workers (e.g., 4)
# num_workers = ...
# YOUR CODE HERE
num_workers = 0 # Replace this

# TODO: Create data_loader_parallel
# data_loader_parallel = grain.DataLoader(...)
# YOUR CODE HERE
data_loader_parallel = None # Replace this

# 3. Iterate and print batch info
if data_loader_parallel:
  print(f"DataLoader configured with worker_count={num_workers}.")
  data_iterator_parallel = iter(data_loader_parallel)
  try:
    first_batch_parallel = next(data_iterator_parallel)
    print(f"Parallel - First batch image shape: {first_batch_parallel['image'].shape}")
    print(f"Parallel - First batch label shape: {first_batch_parallel['label'].shape}")
  except StopIteration:
    print("Parallel DataLoader is empty or exhausted.")
else:
  print("Parallel DataLoader not configured yet.")

# 4. (Optional) Timing comparison
# Re-create sequential loader with MySourceWithWork for a fair comparison
if source_with_work and transformations and index_sampler: # index_sampler from Ex1
  data_loader_seq_with_work = grain.DataLoader(
      data_source=source_with_work,
      operations=transformations, # Reusing from Ex1
      sampler=index_sampler, # Reusing from Ex1 (ensure it's fresh or allows re-iteration)
      worker_count=0,
      shard_options=grain.NoSharding(),
      read_options=grain.ReadOptions(num_threads=0)
      )
  num_batches_to_test = 5 # Small number for quick test

if data_loader_seq_with_work:
    print(f"\nTiming test for {num_batches_to_test} batches:")
    # Sequential
    iterator_seq = iter(data_loader_seq_with_work)
    start_time = time.time()
    try:
      for i in range(num_batches_to_test):
        batch = next(iterator_seq)
        if i == 0: print(f"  Seq batch 1 label sum: {batch['label'].sum()}") # to ensure work is done
    except StopIteration:
      print("Sequential loader exhausted early.")
    end_time = time.time()
    print(f"Sequential ({num_batches_to_test} batches) took: {end_time - start_time:.4f} seconds")

if data_loader_parallel:
    # Parallel
    # Ensure sampler is fresh for parallel loader if it was used above
    # For this optional part, let's use a fresh sampler for the parallel loader
    # to avoid StopIteration if the previous sampler was single-epoch and exhausted.
    fresh_parallel_sampler = grain.IndexSampler(
        num_records=len(source_with_work),
        shard_options=grain.NoSharding(),
        shuffle=True,
        num_epochs=None, # Indefinite
        seed=43 # Different seed or same, for this test it's about speed
    )
    data_loader_parallel_for_timing = grain.DataLoader(
        data_source=source_with_work,
        operations=transformations, # Reusing from Ex1
        sampler=fresh_parallel_sampler,
        worker_count=num_workers if num_workers > 0 else 2, # Ensure parallelism
        shard_options=grain.NoSharding(),
        read_options=grain.ReadOptions(num_threads=0)
    )
    iterator_parallel = iter(data_loader_parallel_for_timing)
    start_time = time.time()
    try:
      for i in range(num_batches_to_test):
        batch = next(iterator_parallel)
        if i == 0: print(f"  Parallel batch 1 label sum: {batch['label'].sum()}") # to ensure work is done
    except StopIteration:
      print("Parallel loader exhausted early.")
    end_time = time.time()
    print(f"Parallel ({num_batches_to_test} batches, {num_workers if num_workers > 0 else 2} workers) took: {end_time - start_time:.4f} seconds")
else:
  print("Skipping optional timing: source_with_work, transformations, or index_sampler not defined.")

# @title Exercise 2: Solution
# 1. Reuse/Recreate components
# Define MySourceWithWork, adding time.sleep(0.01) in getitem
class MySourceWithWork(grain.RandomAccessDataSource):
  def __init__(self, num_records: int = 1000):
    self._num_records = num_records

  def __len__(self) -> int:
    return self._num_records

  def __getitem__(self, idx: int) -> Dict[str, Any]:
    effective_idx = idx % self._num_records
    time.sleep(0.01) # Simulate 10ms of work per item
    image = np.ones((32, 32, 3), dtype=np.uint8) * (effective_idx % 255)
    label = effective_idx % 10
    return {'image': image, 'label': label}

source_with_work = MySourceWithWork(num_records=1000)
print(f"MySourceWithWork created with {len(source_with_work)} records.")

# Sampler for parallel loading (indefinite epochs for robust testing)
parallel_sampler = grain.IndexSampler(
    num_records=len(source_with_work),
    shard_options=grain.NoSharding(),
    shuffle=True,
    num_epochs=None, # Run indefinitely
    seed=42
    )
print("Parallel IndexSampler created.")

# Transformations can be reused from Exercise 1 solution
# Ensure ConvertToFloat is defined (it was in Ex1 solution cell)
if 'ConvertToFloat' not in globals(): # Basic check
  class ConvertToFloat(grain.MapTransform): # Redefine if not in current scope
    def map(self, features: Dict[str, Any]) -> Dict[str, Any]:
      image = features['image'].astype(np.float32) / 255.0
      return {'image': image, 'label': features['label']}

    print("Redefined ConvertToFloat for safety.")

transformations_ex2 = [
    ConvertToFloat(),
    grain.Batch(batch_size=64, drop_remainder=True)
    ]
print("Transformations for Ex2 ready.")

# 2. Instantiate DataLoader with worker_count > 0
num_workers = 4 # Use 4 workers; JAX is configured for 8 virtual CPUs
# Max useful workers often related to num CPU cores available.
data_loader_parallel = grain.DataLoader(
    data_source=source_with_work,
    operations=transformations_ex2,
    sampler=parallel_sampler,
    worker_count=num_workers,
    shard_options=grain.NoSharding(),
    read_options=grain.ReadOptions(num_threads=0) # Data source simulates work but is "in-memory"
    )

# 3. Iterate and print batch info
if data_loader_parallel:
  print(f"DataLoader configured with worker_count={num_workers}.")
  data_iterator_parallel = iter(data_loader_parallel)
  try:
    first_batch_parallel = next(data_iterator_parallel)
    print(f"Parallel - First batch image shape: {first_batch_parallel['image'].shape}")
    print(f"Parallel - First batch label shape: {first_batch_parallel['label'].shape}")
  except StopIteration:
    print("Parallel DataLoader is empty or exhausted.")
else:
  print("Parallel DataLoader not configured yet.")

# 4. (Optional) Timing comparison
# Create a fresh IndexSampler for the sequential loader for a fair comparison start
# (num_epochs=1 to match typical test for sequential pass)
seq_sampler_for_timing = grain.IndexSampler(
    num_records=len(source_with_work),
    shard_options=grain.NoSharding(),
    shuffle=True,
    num_epochs=1, # Single epoch for this timing test
    seed=42
    )

data_loader_seq_with_work = grain.DataLoader(
    data_source=source_with_work,
    operations=transformations_ex2,
    sampler=seq_sampler_for_timing,
    worker_count=0,
    shard_options=grain.NoSharding(),
    read_options=grain.ReadOptions(num_threads=0)
    )
num_batches_to_test = 5 # Number of batches to fetch for timing
print(f"\nTiming test for {num_batches_to_test} batches (each item has 0.01s simulated work):")

# Sequential
iterator_seq = iter(data_loader_seq_with_work)
start_time_seq = time.time()
try:
  for i in range(num_batches_to_test):
    batch_seq = next(iterator_seq)
    if i == 0 and num_batches_to_test > 0 : print(f" Seq batch 1 label sum: {batch_seq['label'].sum()}") # to ensure work is done
except StopIteration:
  print(f"Sequential loader exhausted before {num_batches_to_test} batches.")
end_time_seq = time.time()
print(f"Sequential ({num_batches_to_test} batches) took: {end_time_seq - start_time_seq:.4f} seconds")

# Parallel
# Use a fresh sampler for the parallel loader for timing to ensure it's not exhausted
# and runs for enough batches.
parallel_sampler_for_timing = grain.IndexSampler(
    num_records=len(source_with_work),
    shard_options=grain.NoSharding(),
    shuffle=True,
    num_epochs=None, # Indefinite, or ensure enough for num_batches_to_test
    seed=43 # Can be same or different seed
    )

data_loader_parallel_for_timing = grain.DataLoader(
    data_source=source_with_work,
    operations=transformations_ex2,
    sampler=parallel_sampler_for_timing,
    worker_count=num_workers,
    shard_options=grain.NoSharding(),
    read_options=grain.ReadOptions(num_threads=0)
    )

iterator_parallel_timed = iter(data_loader_parallel_for_timing)
start_time_parallel = time.time()
try:
  for i in range(num_batches_to_test):
    batch_par = next(iterator_parallel_timed)
    if i == 0 and num_batches_to_test > 0 : print(f" Parallel batch 1 label sum: {batch_par['label'].sum()}") # to ensure work is done
except StopIteration:
  print(f"Parallel loader exhausted before {num_batches_to_test} batches.")
end_time_parallel = time.time()
print(f"Parallel ({num_batches_to_test} batches, {num_workers} workers) took: {end_time_parallel - start_time_parallel:.4f} seconds")

if end_time_parallel - start_time_parallel < end_time_seq - start_time_seq:
  print("Parallel loading was faster, as expected!")
else:
  print("Parallel loading was not significantly faster. This might happen for very small num_batches_to_test due to overhead, or if simulated work is too little.")

# @title Exercise 3: Student Code
# 1. Define OneHotEncodeLabel
class OneHotEncodeLabel(grain.MapTransform):
  def __init__(self, num_classes: int):
    # TODO: Store num_classes
    # YOUR CODE HERE
    self._num_classes = 0 # Replace this

  def map(self, features: Dict[str, Any]) -> Dict[str, Any]:
    label = features['label']
    # TODO: Create one-hot encoded version of the label
    # one_hot_label = np.zeros(...)
    # one_hot_label[label] = 1.0
    # YOUR CODE HERE
    one_hot_label = np.array([label]) # Replace this

    features['label'] = one_hot_label
    return features

# 2. Reuse/Create DataSource and Sampler
# TODO: Instantiate MySource (from Ex1, no sleep)
# source_ex3 = ...
# YOUR CODE HERE
source_ex3 = None # Replace this

# TODO: Instantiate an IndexSampler (e.g., from Ex1, or a new one)
# sampler_ex3 = grain.IndexSampler(...)
# YOUR CODE HERE
sampler_ex3 = None # Replace this

# 3. Create new list of operations
# TODO: Instantiate OneHotEncodeLabel
num_classes_for_ohe = 10
one_hot_encoder = OneHotEncodeLabel(num_classes=num_classes_for_ohe)
# YOUR CODE HERE
one_hot_encoder = None # Replace this

# TODO: Define transformations_ex3 list including one_hot_encoder,
# ConvertToFloat (if not already applied), and grain.Batch
# (Assuming ConvertToFloat is defined from Ex1 solution)
# transformations_ex3 = [...]
# YOUR CODE HERE
transformations_ex3 = [] # Replace this

# 4. Instantiate DataLoader
# TODO: Create data_loader_ex3
# data_loader_ex3 = grain.DataLoader(...)
# YOUR CODE HERE
data_loader_ex3 = None # Replace this

# 5. Iterate and print batch info
if data_loader_ex3:
  print("DataLoader with OneHotEncodeLabel configured.")
  iterator_ex3 = iter(data_loader_ex3)
  try:
    first_batch_ex3 = next(iterator_ex3)
    print(f"Custom MapTransform - Batch image shape: {first_batch_ex3['image'].shape}")
    print(f"Custom MapTransform - Batch label shape: {first_batch_ex3['label'].shape}") # Expected: (batch_size, num_classes)
    if first_batch_ex3['label'].size > 0:
      print(f"Custom MapTransform - Example one-hot label: {first_batch_ex3['label'][0]}")
  except StopIteration:
    print("DataLoader for Ex3 is empty or exhausted.")
else:
  print("DataLoader for Ex3 not configured yet.")

# @title Exercise 3: Solution
# 1. Define OneHotEncodeLabel
class OneHotEncodeLabel(grain.MapTransform):
  def __init__(self, num_classes: int):
    self._num_classes = num_classes

  def map(self, features: Dict[str, Any]) -> Dict[str, Any]:
    label_scalar = features['label']
    one_hot_label = np.zeros(self._num_classes, dtype=np.float32)
    one_hot_label[label_scalar] = 1.0

    # Create a new dictionary to avoid modifying the input dict in place if it's reused
    # by other transforms or parts of the pipeline, though often direct modification is fine.
    # For safety and clarity, let's return a new dict or an updated copy.
    updated_features = features.copy()
    updated_features['label'] = one_hot_label
    return updated_features

# 2. Reuse/Create DataSource and Sampler
# Using MySource from Exercise 1 solution (no artificial sleep)
if 'MySource' not in globals(): # Basic check
  class MySource(grain.RandomAccessDataSource): # Redefine if not in current scope
    def __init__(self, num_records: int = 1000):
      self._num_records = num_records
    def __len__(self) -> int:
      return self._num_records

    def __getitem__(self, idx: int) -> Dict[str, Any]:
      effective_idx = idx % self._num_records
      image = np.ones((32, 32, 3), dtype=np.uint8) * (effective_idx % 255)
      label = effective_idx % 10
      return {'image': image, 'label': label}
  print("Redefined MySource for Ex3.")

source_ex3 = MySource(num_records=1000)
sampler_ex3 = grain.IndexSampler(
  num_records=len(source_ex3),
  shard_options=grain.NoSharding(),
  shuffle=True,
  num_epochs=1,
  seed=42
  )
print("DataSource and Sampler for Ex3 ready.")

# 3. Create new list of operations
num_classes_for_ohe = 10 # Matches idx % 10 in MySource
one_hot_encoder = OneHotEncodeLabel(num_classes=num_classes_for_ohe)

# Ensure ConvertToFloat is defined
if 'ConvertToFloat' not in globals():
  class ConvertToFloat(grain.MapTransform):
    def map(self, features: Dict[str, Any]) -> Dict[str, Any]:
      image = features['image'].astype(np.float32) / 255.0
      return {'image': image, 'label': features['label']} # Pass label through
print("Redefined ConvertToFloat for Ex3.")

transformations_ex3 = [
  ConvertToFloat(), # Apply first to have float images
  one_hot_encoder, # Then one-hot encode labels
  grain.Batch(batch_size=64, drop_remainder=True)
  ]
print("Transformations for Ex3 defined.")

# 4. Instantiate DataLoader
data_loader_ex3 = grain.DataLoader(
  data_source=source_ex3,
  operations=transformations_ex3,
  sampler=sampler_ex3,
  worker_count=0, # Can be > 0 as well, OneHotEncodeLabel is picklable
  shard_options=grain.NoSharding(),
  read_options=grain.ReadOptions(num_threads=0)
  )

# 5. Iterate and print batch info
if data_loader_ex3:
  print("DataLoader with OneHotEncodeLabel configured.")
  iterator_ex3 = iter(data_loader_ex3)
  try:
    first_batch_ex3 = next(iterator_ex3)
    print(f"Custom MapTransform - Batch image shape: {first_batch_ex3['image'].shape}")
    print(f"Custom MapTransform - Batch label shape: {first_batch_ex3['label'].shape}") # Expected: (64, 10)
    if first_batch_ex3['label'].size > 0:
      print(f"Custom MapTransform - Example one-hot label (first item): {first_batch_ex3['label'][0]}")
    original_label_example = np.argmax(first_batch_ex3['label'][0])
    print(f"Custom MapTransform - Decoded original label (first item): {original_label_example}")
  except StopIteration:
    print("DataLoader for Ex3 is empty or exhausted.")
else:
  print("DataLoader for Ex3 not configured yet.")

# @title Exercise 4: Student Code
# 1. Define RandomBrightnessAdjust
class RandomBrightnessAdjust(grain.RandomMapTransform):
  def random_map(self, features: Dict[str, Any], rng: np.random.Generator) -> Dict[str, Any]:
    # TODO: Ensure image is float (e.g. by placing ConvertToFloat before this in ops)
    image = features['image']

    # TODO: Generate a random brightness factor using the provided rng
    # brightness_factor = rng.uniform(...)
    # YOUR CODE HERE
    brightness_factor = 1.0 # Replace this

    # TODO: Apply brightness adjustment and clip
    # adjusted_image = np.clip(...)
    # YOUR CODE HERE
    adjusted_image = image # Replace this

    # Create a new dictionary or update a copy
    updated_features = features.copy()
    updated_features['image'] = adjusted_image
    return updated_features

# 2. Reuse/Create DataSource, Sampler, ConvertToFloat
# TODO: Instantiate MySource (from Ex1)
# source_ex4 = ...
# YOUR CODE HERE
source_ex4 = None # Replace this

# TODO: Instantiate an IndexSampler with a seed (e.g., seed=42, num_epochs=1 or None)
# sampler_ex4_seed42 = grain.IndexSampler(...)
# YOUR CODE HERE
sampler_ex4_seed42 = None # Replace this

# (Assuming ConvertToFloat is defined from Ex1 solution)

# 3. Create list of operations
# TODO: Instantiate RandomBrightnessAdjust
# random_brightness_adjuster = ...
# YOUR CODE HERE
random_brightness_adjuster = None # Replace this

# TODO: Define transformations_ex4 list: ConvertToFloat, random_brightness_adjuster, grain.Batch
# transformations_ex4 = [...]
# YOUR CODE HERE
transformations_ex4 = [] # Replace this

# 4. Instantiate two DataLoaders with the same config
# TODO: Create dl_run1
# dl_run1 = grain.DataLoader(...)
# YOUR CODE HERE
dl_run1 = None # Replace this

# TODO: Create dl_run2 (using the exact same sampler instance or a new one with the same seed)
# dl_run2 = grain.DataLoader(...)
# YOUR CODE HERE
dl_run2 = None # Replace this

# 5. & 6. Iterate and compare
pixel_to_check = (0, 0, 0, 0) # Batch_idx, H, W, C
if dl_run1:
  print("--- Run 1 (seed 42) ---")
  iterator_run1 = iter(dl_run1)
  try:
    batch1_run1 = next(iterator_run1)
    value_run1 = batch1_run1['image'][pixel_to_check]
    print(f"Run 1 - Pixel {pixel_to_check} value: {value_run1}")
  except StopIteration:
    print("dl_run1 exhausted.")
    value_run1 = None
else:
  print("dl_run1 not configured.")
  value_run1 = None

if dl_run2:
  print("\n--- Run 2 (seed 42, same sampler) ---")
  # If sampler_ex4_seed42 was single-epoch and already used by dl_run1,
  # dl_run2 might be empty. For robust test, ensure sampler allows re-iteration
  # or use a new sampler instance with the same seed.
  # If sampler_ex4_seed42 had num_epochs=None, iter(dl_run2) is fine.
  # If num_epochs=1, you might need to re-create sampler_ex4_seed42 for dl_run2
  # or ensure dl_run1 didn't exhaust it (e.g. by not fully iterating it).
  # For this exercise, assume sampler_ex4_seed42 can be re-used or is fresh for dl_run2.
  iterator_run2 = iter(dl_run2)
  try:
    batch1_run2 = next(iterator_run2)
    value_run2 = batch1_run2['image'][pixel_to_check]
    print(f"Run 2 - Pixel {pixel_to_check} value: {value_run2}")

    # 7. Verify
    if value_run1 is not None and value_run2 is not None:
      if np.allclose(value_run1, value_run2):
        print("\nSUCCESS: Pixel values are identical. Randomness is reproducible!")
      else:
        print(f"\nFAILURE: Pixel values differ. value1={value_run1}, value2={value_run2}")
  except StopIteration:
    print("dl_run2 exhausted. This might happen if the sampler was single-epoch and already used.")
    value_run2 = None
else:
  print("dl_run2 not configured.")

# 8. (Optional) Test with a different seed
# TODO: Create sampler_ex4_seed100 (seed=100)
# sampler_ex4_seed100 = grain.IndexSampler(...)
# YOUR CODE HERE
sampler_ex4_seed100 = None # Replace this

# TODO: Create dl_run3 with sampler_ex4_seed100
# dl_run3 = grain.DataLoader(...)
# YOUR CODE HERE
dl_run3 = None # Replace this

if dl_run3:
  print("\n--- Run 3 (seed 100) ---")
  iterator_run3 = iter(dl_run3)
  try:
    batch1_run3 = next(iterator_run3)
    value_run3 = batch1_run3['image'][pixel_to_check]
    print(f"Run 3 - Pixel {pixel_to_check} value: {value_run3}")
    if value_run1 is not None and not np.allclose(value_run1, value_run3):
      print("SUCCESS: Pixel values differ from Run 1 (seed 42), as expected with a new seed.")
    elif value_run1 is not None:
      print("NOTE: Pixel values are the same as Run 1. Check seed or logic.")
  except StopIteration:
    print("dl_run3 exhausted.")
else:
  print("\nOptional part (dl_run3) not configured.")

# @title Exercise 4: Solution
# 1. Define RandomBrightnessAdjust
class RandomBrightnessAdjust(grain.RandomMapTransform):
  def random_map(self, features: Dict[str, Any], rng: np.random.Generator) -> Dict[str, Any]:
    image = features['image'] # Assumes image is already float, e.g. from ConvertToFloat
    # Generate a random brightness factor using the provided rng
    brightness_factor = rng.uniform(0.7, 1.3)

    # Apply brightness adjustment and clip
    adjusted_image = image * brightness_factor
    adjusted_image = np.clip(adjusted_image, 0.0, 1.0)

    updated_features = features.copy()
    updated_features['image'] = adjusted_image
    return updated_features

# 2. Reuse/Create DataSource, Sampler, ConvertToFloat
if 'MySource' not in globals(): # Basic check for MySource
  class MySource(grain.RandomAccessDataSource):
    def __init__(self, num_records: int = 1000):
      self._num_records = num_records
    def __len__(self) -> int:
      return self._num_records
    def __getitem__(self, idx: int) -> Dict[str, Any]:
      effective_idx = idx % self._num_records
      image = np.ones((32, 32, 3), dtype=np.uint8) * (effective_idx % 255)
      label = effective_idx % 10
      return {'image': image, 'label': label}
  print("Redefined MySource for Ex4.")
source_ex4 = MySource(num_records=1000)

# Sampler with a fixed seed. num_epochs=None allows re-iteration for multiple DataLoaders.
# If num_epochs=1, the sampler instance can only be fully iterated once.
# For this test, using num_epochs=None or re-creating the sampler for each DataLoader is safest.
# Let's use num_epochs=None to allow the same sampler instance to be used.
sampler_ex4_seed42 = grain.IndexSampler(
  num_records=len(source_ex4),
  shard_options=grain.NoSharding(),
  shuffle=True,
  num_epochs=None, # Allow indefinite iteration
  seed=42
  )
print("DataSource and Sampler (seed 42) for Ex4 ready.")

if 'ConvertToFloat' not in globals(): # Basic check for ConvertToFloat
  class ConvertToFloat(grain.MapTransform):
    def map(self, features: Dict[str, Any]) -> Dict[str, Any]:
      image = features['image'].astype(np.float32) / 255.0
      return {'image': image, 'label': features['label']}
  print("Redefined ConvertToFloat for Ex4.")

# 3. Create list of operations
random_brightness_adjuster = RandomBrightnessAdjust()
transformations_ex4 = [
  ConvertToFloat(),
  random_brightness_adjuster,
  grain.Batch(batch_size=64, drop_remainder=True)
]
print("Transformations for Ex4 defined.")

# 4. Instantiate two DataLoaders with the same config
# Using worker_count > 0 to also test picklability of RandomBrightnessAdjust
num_workers_ex4 = 2
dl_run1 = grain.DataLoader(
  data_source=source_ex4,
  operations=transformations_ex4,
  sampler=sampler_ex4_seed42, # Same sampler instance
  worker_count=num_workers_ex4,
  shard_options=grain.NoSharding(),
  read_options=grain.ReadOptions(num_threads=0)
  )

dl_run2 = grain.DataLoader(
  data_source=source_ex4,
  operations=transformations_ex4,
  sampler=sampler_ex4_seed42, # Same sampler instance
  worker_count=num_workers_ex4,
  shard_options=grain.NoSharding(),
  read_options=grain.ReadOptions(num_threads=0)
  )
print(f"DataLoaders for Run1 and Run2 created with worker_count={num_workers_ex4}.")

# 5. & 6. Iterate and compare
pixel_to_check = (0, 0, 0, 0) # Batch_idx=0, H=0, W=0, C=0
print("\n--- Run 1 (seed 42) ---")
iterator_run1 = iter(dl_run1)

try:
  batch1_run1 = next(iterator_run1)
  value_run1 = batch1_run1['image'][pixel_to_check]
  print(f"Run 1 - Pixel {pixel_to_check} value: {value_run1}")
except StopIteration:
  print("dl_run1 exhausted.")
  value_run1 = None

print("\n--- Run 2 (seed 42, same sampler instance) ---")
iterator_run2 = iter(dl_run2) # Gets a new iterator from the DataLoader

try:
  batch1_run2 = next(iterator_run2)
  value_run2 = batch1_run2['image'][pixel_to_check]
  print(f"Run 2 - Pixel {pixel_to_check} value: {value_run2}")
  # 7. Verify
  if value_run1 is not None and value_run2 is not None:
    if np.allclose(value_run1, value_run2):
      print("\nSUCCESS: Pixel values are identical. Randomness is reproducible with the same sampler instance!")
    else:
      print(f"\nFAILURE: Pixel values differ. value1={value_run1}, value2={value_run2}. This shouldn't happen if sampler is re-used correctly.")
except StopIteration:
  print("dl_run2 exhausted.")

# 8. (Optional) Test with a different seed
sampler_ex4_seed100 = grain.IndexSampler(
  num_records=len(source_ex4),
  shard_options=grain.NoSharding(),
  shuffle=True,
  num_epochs=None,
  seed=100 # Different seed
  )

dl_run3 = grain.DataLoader(
  data_source=source_ex4,
  operations=transformations_ex4,
  sampler=sampler_ex4_seed100, # Sampler with different seed
  worker_count=num_workers_ex4,
  shard_options=grain.NoSharding(),
  read_options=grain.ReadOptions(num_threads=0)
  )
print("\nDataLoader for Run3 (seed 100) created.")
print("\n--- Run 3 (seed 100) ---")
iterator_run3 = iter(dl_run3)

try:
  batch1_run3 = next(iterator_run3)
  value_run3 = batch1_run3['image'][pixel_to_check]
  print(f"Run 3 - Pixel {pixel_to_check} value: {value_run3}")
  if value_run1 is not None and not np.allclose(value_run1, value_run3):
    print("SUCCESS: Pixel values differ from Run 1 (seed 42), as expected with a new sampler seed.")
  elif value_run1 is not None:
    print("NOTE: Pixel values are the same as Run 1. This is unexpected if seeds are different.")
except StopIteration:
  print("dl_run3 exhausted.")

# @title Exercise 5: Student Code
# 1. Reuse DataSource and basic transformations
# TODO: Instantiate MySource (from Ex1)
# source_ex5 = ...
# YOUR CODE HERE
source_ex5 = None # Replace this

# TODO: Define basic_transformations_ex5 (e.g., ConvertToFloat, Batch)
# (Assuming ConvertToFloat is defined)
# basic_transformations_ex5 = [...]
# YOUR CODE HERE
basic_transformations_ex5 = [] # Replace this

# 2. Define shard_count
shard_count = 2
common_seed = 42
num_epochs_for_sharding_test = 1 # To make collection of all labels feasible

# 3. Simulate Process 0
# TODO: Create shard_options_p0
# shard_options_p0 = grain.ShardOptions(...)
# YOUR CODE HERE
shard_options_p0 = None # Replace this

# TODO: Create sampler_p0. Pass shard_options_p0 to the IndexSampler.
# sampler_p0 = grain.IndexSampler(...)
# YOUR CODE HERE
sampler_p0 = None # Replace this

# TODO: Create dl_p0. Pass shard_options_p0 to the DataLoader as well.
# dl_p0 = grain.DataLoader(...)
# YOUR CODE HERE
dl_p0 = None # Replace this

labels_p0 = set()
if dl_p0:
  print("--- Simulating Process 0 ---")
  # YOUR CODE HERE: Iterate through dl_p0 and collect all unique original labels.
  # Remember that labels might be batched. You need to iterate through items in a batch.
  # For simplicity, if your MySource generates labels like idx % 10,
  # you can try to collect the indices that were sampled.
  # Or, more directly, collect the 'label' field from each item.
  # To get original indices, you might need a transform that passes index through.
  # Let's collect the 'label' values directly.
  pass # Replace with iteration logic

# 4. Simulate Process 1
# TODO: Create shard_options_p1
# shard_options_p1 = grain.ShardOptions(...)
# YOUR CODE HERE
shard_options_p1 = None # Replace this

# TODO: Create sampler_p1
# sampler_p1 = grain.IndexSampler(...)
# YOUR CODE HERE
sampler_p1 = None # Replace this

# TODO: Create dl_p1
# dl_p1 = grain.DataLoader(...)
# YOUR CODE HERE
dl_p1 = None # Replace this

labels_p1 = set()
if dl_p1:
  print("\n--- Simulating Process 1 ---")
  # YOUR CODE HERE: Iterate through dl_p1 and collect all unique labels.
  pass # Replace with iteration logic

# 5. Verify
print(f"\n--- Verification (Total records in source: {len(source_ex5) if source_ex5 else 'N/A'}) ---")
print(f"Unique labels collected by Process 0 (count {len(labels_p0)}): sorted {sorted(list(labels_p0))[:20]}...")
print(f"Unique labels collected by Process 1 (count {len(labels_p1)}): sorted {sorted(list(labels_p1))[:20]}...")
if labels_p0 and labels_p1:
  intersection = labels_p0.intersection(labels_p1)
  if not intersection:
    print("\nSUCCESS: No overlap in labels between Process 0 and Process 1. Sharding works as expected!")
  else:
    print(f"\nNOTE: Some overlap in labels found (count {len(intersection)}): {intersection}.")
    print("This can happen if labels are not unique per index, or if sharding logic has issues.")
    print("With MySource's label = idx % 10, an overlap in labels is expected even if indices are disjoint.")
    print("A better test would be to collect original indices if possible.")

# For a more direct test of sharding of indices:
# We can define a DataSource that returns the index itself.
class IndexSource(grain.RandomAccessDataSource):
  def __init__(self, num_records: int):
    self._num_records = num_records
  def __len__(self) -> int:
    return self._num_records
  def __getitem__(self, idx: int) -> int:
    return idx % self._num_records # Return the index
index_source = IndexSource(num_records=100) # Smaller source for easier inspection

idx_sampler_p0 = grain.IndexSampler(len(index_source), shard_options_p0, shuffle=False, num_epochs=1, seed=common_seed)
idx_sampler_p1 = grain.IndexSampler(len(index_source), shard_options_p1, shuffle=False, num_epochs=1, seed=common_seed)

# DataLoader for indices (no batching, just to see raw sampled indices)
# Note: DataLoader expects dicts. Let's make IndexSource return {'index': idx}
class IndexDictSource(grain.RandomAccessDataSource):
  def __init__(self, num_records: int): self._num_records = num_records
  def __len__(self) -> int:
    return self._num_records
  def __getitem__(self, idx: int) -> Dict[str,int]:
    return {'index': idx % self._num_records}
index_dict_source = IndexDictSource(num_records=100)

# Samplers for IndexDictSource
idx_dict_sampler_p0 = grain.IndexSampler(len(index_dict_source), shard_options_p0, shuffle=False, num_epochs=1, seed=common_seed)
idx_dict_sampler_p1 = grain.IndexSampler(len(index_dict_source), shard_options_p1, shuffle=False, num_epochs=1, seed=common_seed)

# DataLoaders for IndexDictSource
# Pass shard_options to DataLoader as well.
if shard_options_p0 and shard_options_p1:
  dl_indices_p0 = grain.DataLoader(index_dict_source, [], idx_dict_sampler_p0, worker_count=0, shard_options=shard_options_p0)
  dl_indices_p1 = grain.DataLoader(index_dict_source, [], idx_dict_sampler_p1, worker_count=0, shard_options=shard_options_p1)
  indices_from_p0 = {item['index'] for item in dl_indices_p0} if dl_indices_p0 else set()
  indices_from_p1 = {item['index'] for item in dl_indices_p1} if dl_indices_p1 else set()

print(f"\n--- Verification of INDICES (Source size: {len(index_dict_source)}) ---")
print(f"Indices from P0 (count {len(indices_from_p0)}, shuffle=False): {sorted(list(indices_from_p0))}")
print(f"Indices from P1 (count {len(indices_from_p1)}, shuffle=False): {sorted(list(indices_from_p1))}")

if indices_from_p0 and indices_from_p1:
  idx_intersection = indices_from_p0.intersection(indices_from_p1)
  if not idx_intersection:
    print("SUCCESS: No overlap in INDICES. Sharding of data sources works correctly!")
  else:
    print(f"FAILURE: Overlap in INDICES found: {idx_intersection}")
else:
  print("Skipping index verification part as shard_options are not defined.")
  print("\nReminder: In a real distributed setup, you'd use grain.sharding.ShardByJaxProcess() "
  "and JAX would manage jax.process_index() automatically for each process.")

# @title Exercise 5: Solution
# Redefine MySource for Ex5 to include 'original_index'.
class MySource(grain.RandomAccessDataSource):
  def __init__(self, num_records: int = 1000):
    self._num_records = num_records
  def __len__(self) -> int:
    return self._num_records
  def __getitem__(self, idx: int) -> Dict[str, Any]:
    effective_idx = idx % self._num_records
    image = np.ones((32, 32, 3), dtype=np.uint8) * (effective_idx % 255)
    label = effective_idx % 10 # Label is idx % 10
    # For better sharding verification, let's also pass the original index
    return {'image': image, 'label_test': label, 'original_index': effective_idx}

print("Redefined MySource for Ex5 to include 'original_index'.")
source_ex5 = MySource(num_records=1000)

# Redefine ConvertToFloat for Ex5 to include 'original_index'.
class ConvertToFloat(grain.MapTransform):
  def map(self, features: Dict[str, Any]) -> Dict[str, Any]:
    # This transform should pass through all keys it doesn't modify
    updated_features = features.copy()
    updated_features['image'] = features['image'].astype(np.float32) / 255.0
    return updated_features
print("Redefined ConvertToFloat for Ex5.")

# We will collect 'original_index' after batching, so batch must preserve it.
# grain.Batch by default collates features with the same name.
basic_transformations_ex5 = [
  ConvertToFloat(),
  grain.Batch(batch_size=64, drop_remainder=True) # drop_remainder for batching
  ]
print("DataSource and Transformations for Ex5 ready.")

# 2. Define shard_count
shard_count = 2
common_seed = 42
num_epochs_for_sharding_test = 1

# 3. Simulate Process 0
shard_options_p0 = grain.ShardOptions(shard_index=0, shard_count=shard_count, drop_remainder=True) # drop_remainder for sharding

# Sampler for Process 0. It's important that the sampler itself is sharded.
# The DataLoader's shard_options will also apply this sharding if the sampler isn't already sharded,
# or verify consistency if it is.
sampler_p0 = grain.IndexSampler(
  num_records=len(source_ex5),
  shard_options=shard_options_p0, # Shard the sampler
  shuffle=True, # Shuffle for more realistic scenario
  num_epochs=num_epochs_for_sharding_test,
  seed=common_seed
  )

dl_p0 = grain.DataLoader(
  data_source=source_ex5,
  operations=basic_transformations_ex5,
  sampler=sampler_p0,
  worker_count=0, # Keep it simple for verification
  shard_options=shard_options_p0, # Also inform DataLoader about the sharding context
  read_options=grain.ReadOptions(num_threads=0)
  )

indices_p0 = set()
if dl_p0:
  print("--- Simulating Process 0 ---")
  for batch in dl_p0:
    indices_p0.update(batch['original_index'].tolist()) # Collect original indices
  print(f"Process 0 collected {len(indices_p0)} unique indices.")

# 4. Simulate Process 1
shard_options_p1 = grain.ShardOptions(shard_index=1, shard_count=shard_count, drop_remainder=True)
sampler_p1 = grain.IndexSampler(
  num_records=len(source_ex5),
  shard_options=shard_options_p1, # Shard the sampler
  shuffle=True, # Use same shuffle setting and seed for apples-to-apples comparison of sharding logic
  num_epochs=num_epochs_for_sharding_test,
  seed=common_seed # Same seed ensures shuffle order is same before sharding
  )

dl_p1 = grain.DataLoader(
  data_source=source_ex5,
  operations=basic_transformations_ex5,
  sampler=sampler_p1,
  worker_count=0,
  shard_options=shard_options_p1, # Inform DataLoader
  read_options=grain.ReadOptions(num_threads=0)
  )

indices_p1 = set()
if dl_p1:
  print("\n--- Simulating Process 1 ---")
  for batch in dl_p1:
    indices_p1.update(batch['original_index'].tolist()) # Collect original indices
  print(f"Process 1 collected {len(indices_p1)} unique indices.")

# 5. Verify
print(f"\n--- Verification of original_indices (Total records in source: {len(source_ex5)}) ---")
# Showing a few from each for brevity
print(f"Unique original_indices from P0 (first 20 sorted): {sorted(list(indices_p0))[:20]}...")
print(f"Unique original_indices from P1 (first 20 sorted): {sorted(list(indices_p1))[:20]}...")
expected_per_shard = len(source_ex5) // shard_count # Due to drop_remainder=True in ShardOptions
print(f"Expected records per shard (approx, due to drop_remainder in sharding): {expected_per_shard}")
print(f"Actual for P0: {len(indices_p0)}, P1: {len(indices_p1)}")

if indices_p0 and indices_p1:
  intersection = indices_p0.intersection(indices_p1)
  if not intersection:
    print("\nSUCCESS: No overlap in original_indices between Process 0 and Process 1. Sharding works!")
  else:
    print(f"\nFAILURE: Overlap in original_indices found (count {len(intersection)}): {sorted(list(intersection))[:20]}...")
    print("This should not happen if sharding is correct and seeds/shuffle are consistent.")
else:
  print("Could not perform intersection test as one or both sets of indices are empty.")
  total_unique_indices_seen = len(indices_p0.union(indices_p1))
  print(f"Total unique indices seen across both simulated processes: {total_unique_indices_seen}")

# With drop_remainder=True in sharding, total might be less than len(source_ex5)
# if len(source_ex5) is not divisible by shard_count.
# Example: 1000 records, 2 shards. Each gets 500. Total 1000.
# Example: 1001 records, 2 shards. drop_remainder=True means each gets 500. Total 1000. 1 record dropped.
print("\nReminder: In a real distributed JAX application:")
print("1. Each JAX process would run this script (or similar).")
print("2. shard_options = grain.sharding.ShardByJaxProcess(drop_remainder=True) would be used.")
print("3. jax.process_index() and jax.process_count() would provide the correct shard info automatically.")
print("4. The IndexSampler and DataLoader would be configured with these auto-detected shard_options.")

# @title Exercise 6: Student Code
# 1. Define SimpleNNXModel
class SimpleNNXModel(nnx.Module):
  def __init__(self, din: int, dout: int, *, rngs: nnx.Rngs):
    # TODO: Initialize an nnx.Linear layer
    # self.linear = nnx.Linear(...)
    # YOUR CODE HERE
    self.linear = None # Replace this
  def __call__(self, x: jax.Array):
    # TODO: Flatten the input image (if B, H, W, C) and pass through linear layer
    # x_flat = x.reshape((x.shape[0], -1))
    # return self.linear(x_flat)
    # YOUR CODE HERE
    return x # Replace this

# 2. Define train_step
def train_step(model: SimpleNNXModel, batch: Dict[str, jax.Array]):
  # TODO: Perform forward pass: model(batch['image'])
  # logits = ...
  # YOUR CODE HERE
  logits = model(batch['image']) # Assuming model handles it
  # TODO: Calculate a dummy loss (e.g., mean of logits)
  # loss = ...
  # YOUR CODE HERE
  loss = jnp.array(0.0) # Replace this

  # In a real scenario, you'd also compute gradients and update model parameters here.
  # For this exercise, we just return the original model and the loss.
  return model, loss

# 3. Set up DataLoader
# TODO: Instantiate MySource (from Ex1, or the one with 'original_index' if you prefer)
# source_ex6 = ...
# YOUR CODE HERE
source_ex6 = None # Replace this

# TODO: Instantiate an IndexSampler for a few epochs (e.g., 2 epochs)
# sampler_ex6 = grain.IndexSampler(...)
# YOUR CODE HERE
sampler_ex6 = None # Replace this

# TODO: Define transformations_ex6 (e.g., ConvertToFloat, grain.Batch)
# (Assuming ConvertToFloat is defined)
# transformations_ex6 = [...]
# YOUR CODE HERE
transformations_ex6 = [] # Replace this

# TODO: Instantiate data_loader_ex6
# data_loader_ex6 = grain.DataLoader(...)
# YOUR CODE HERE
data_loader_ex6 = None # Replace this

# 4. Write the Training Loop
if data_loader_ex6: # Proceed only if DataLoader is configured
  # TODO: Initialize SimpleNNXModel
  # image_height, image_width, image_channels = 32, 32, 3
  # num_classes_ex6 = 10
  # model_key = jax.random.key(0)
  # model_ex6 = SimpleNNXModel(...)
  # YOUR CODE HERE
  model_ex6 = None # Replace this

if model_ex6:
  # TODO: Get an iterator from data_loader_ex6
  # grain_iterator_ex6 = ...
  # YOUR CODE HERE
  grain_iterator_ex6 = iter([]) # Replace this

  num_steps = 100
  print(f"\nStarting conceptual training loop for {num_steps} steps...")
  for step in range(num_steps):
    try:
      # TODO: Get next_batch from iterator
      # next_batch = ...
      # YOUR CODE HERE
      next_batch = None # Replace this
      if next_batch is None:
        raise StopIteration # Simulate exhaustion if not implemented
    except StopIteration:
      print(f"DataLoader exhausted at step {step}. Ending loop.")
      break

    # TODO: Call train_step
    # loss, model_ex6 = train_step(model_ex6, next_batch) # model_ex6 isn't actually updated here
    # YOUR CODE HERE
    loss = jnp.array(0.0) # Replace this

    if step % 20 == 0 or step == num_steps - 1:
      print(f"Step {step}: Dummy Loss = {loss.item():.4f}")
      print("Conceptual training loop finished.")
    else:
      print("Model for Ex6 not initialized.")
else:
  print("DataLoader for Ex6 not configured.")

# @title Exercise 6: Solution
# 1. Define SimpleNNXModel
class SimpleNNXModel(nnx.Module):
  def __init__(self, din: int, dout: int, *, rngs: nnx.Rngs):
    self.linear = nnx.Linear(din, dout, rngs=rngs)
  def __call__(self, x: jax.Array) -> jax.Array:
    # Assuming x is (B, H, W, C)
    batch_size = x.shape[0]
    x_flat = x.reshape((batch_size, -1)) # Flatten H, W, C dimensions
    return self.linear(x_flat)

# 2. Define conceptual train_step
def train_step(model: SimpleNNXModel, batch: Dict[str, jax.Array]):
  # Perform forward pass
  logits = model(batch['image']) # model.call is invoked
  # Calculate a dummy loss
  loss = jnp.mean(logits**2) # Example: mean of squared logits

  # In a real training step:
  # # 1. Define a loss function.
  # def loss_fn(model):
  #   logits = model(batch['image'])
  #   # loss_value = ... (e.g., optax.softmax_cross_entropy_with_integer_labels)
  #   return jnp.mean(logits**2) # Using dummy loss from exercise
  #
  # # 2. Calculate gradients.
  # grads = nnx.grad(loss_fn, wrt=nnx.Param)(model)
  #
  # # 3. Update the model's parameters in-place using the optimizer.
  # #    Note: The optimizer is defined outside the train step.
  # #    e.g., optimizer = nnx.Optimizer(model, optax.adam(1e-3), wrt=nnx.Param)
  # optimizer.update(model, grads)
  #
  # # 4. For this exercise, we just return the original model and the loss.
  return model, loss

# 3. Set up DataLoader
# Redefine MySource for Ex6.
class MySource(grain.RandomAccessDataSource):
  def __init__(self, num_records: int = 1000):
    self._num_records = num_records
  def __len__(self) -> int:
    return self._num_records
  def __getitem__(self, idx: int) -> Dict[str, Any]:
    effective_idx = idx % self._num_records
    image = np.ones((32, 32, 3), dtype=np.uint8) * (effective_idx % 255)
    label = effective_idx % 10
    return {'image': image, 'label': label}
print("Redefined MySource for Ex6.")

source_ex6 = MySource(num_records=1000)
sampler_ex6 = grain.IndexSampler(
  num_records=len(source_ex6),
  shard_options=grain.NoSharding(),
  shuffle=True,
  num_epochs=2, # Run for 2 epochs
  seed=42
  )

# Redefine ConvertToFloat for Ex6.
class ConvertToFloat(grain.MapTransform):
  def map(self, features: Dict[str, Any]) -> Dict[str, Any]:
    updated_features = features.copy()
    updated_features['image'] = features['image'].astype(np.float32) / 255.0
    return updated_features
print("Redefined ConvertToFloat for Ex6.")

transformations_ex6 = [
  ConvertToFloat(),
  grain.Batch(batch_size=64, drop_remainder=True)
  ]

data_loader_ex6 = grain.DataLoader(
  data_source=source_ex6,
  operations=transformations_ex6,
  sampler=sampler_ex6,
  worker_count=2, # Use a couple of workers
  shard_options=grain.NoSharding(),
  read_options=grain.ReadOptions(num_threads=0)
  )
print("DataLoader for Ex6 configured.")

# 4. Write the Training Loop
# Define image dimensions and number of classes
image_height, image_width, image_channels = 32, 32, 3
input_dim = image_height * image_width * image_channels
num_classes_ex6 = 10

# Initialize SimpleNNXModel
# NNX modules are typically initialized outside JIT, then their state can be passed.
# For this conceptual example, the model instance itself is passed.
model_key = jax.random.key(0)
model_ex6 = SimpleNNXModel(din=input_dim, dout=num_classes_ex6, rngs=nnx.Rngs(params=model_key))
print(f"SimpleNNXModel initialized. Input dim: {input_dim}, Output dim: {num_classes_ex6}")

# Get an iterator from data_loader_ex6
grain_iterator_ex6 = iter(data_loader_ex6)
num_steps = 100 # Total steps for the conceptual loop

print(f"\nStarting conceptual training loop for {num_steps} steps...")
for step_idx in range(num_steps):
  try:
    # Get next_batch from iterator
    next_batch = next(grain_iterator_ex6)
  except StopIteration:
    print(f"DataLoader exhausted at step {step_idx}. Ending loop.")
    # Example: Re-initialize iterator if sampler allows multiple epochs
    # if sampler_ex6.num_epochs is None or sampler_ex6.num_epochs > 1 (and we tracked current epoch):
      # print("Re-initializing iterator for new epoch...")
      # grain_iterator_ex6 = iter(data_loader_ex6)
      # next_batch = next(grain_iterator_ex6)
    # else:
    break # Exit loop if truly exhausted

  # Call train_step
  # JAX arrays in batch are automatically handled by jax.jit
  _, loss = train_step(model_ex6, next_batch) # model_ex6 state isn't actually updated here

  if step_idx % 20 == 0 or step_idx == num_steps - 1:
    print(f"Step {step_idx}: Dummy Loss = {loss.item():.4f}") # .item() to get Python scalar from JAX array
print("Conceptual training loop finished.")

# @title Exercise 7: Student Code
# 1. Set up DataLoader
# TODO: Instantiate MySource (e.g., from Ex1)
# source_ex7 = ...
# YOUR CODE HERE
source_ex7 = None # Replace this

# TODO: Instantiate an IndexSampler (num_epochs=None, seed=42)
# sampler_ex7 = grain.IndexSampler(...)
# YOUR CODE HERE
sampler_ex7 = None # Replace this

# TODO: Define transformations_ex7 (e.g., ConvertToFloat, Batch)
# (Assuming ConvertToFloat is defined)
# transformations_ex7 = [...]
# YOUR CODE HERE
transformations_ex7 = [] # Replace this

# TODO: Instantiate data_loader_ex7
# data_loader_ex7 = grain.DataLoader(...)
# YOUR CODE HERE
data_loader_ex7 = None # Replace this

if data_loader_ex7:
  # 2. Get iterator1
  # TODO: iterator1 = iter(...)
  # YOUR CODE HERE
  iterator1 = iter([]) # Replace this

# 3. Iterate a few times
num_initial_iterations = 3
print(f"--- Initial Iteration (iterator1) for {num_initial_iterations} batches ---")
last_batch_iterator1 = None
for i in range(num_initial_iterations):
  try:
    # TODO: last_batch_iterator1 = next(...)
    # YOUR CODE HERE
    last_batch_iterator1 = {} # Replace this
    print(f"iterator1, batch {i+1} - first label: {last_batch_iterator1.get('label', [None])[0]}")
  except StopIteration:
    print("iterator1 exhausted prematurely.")
    break

# 4. Save State
# TODO: saved_iterator_state = iterator1.get_state()
# YOUR CODE HERE
saved_iterator_state = None # Replace this
print(f"\nIterator state saved. Type: {type(saved_iterator_state)}")

# For verification: get the *next* batch from iterator1 *after* saving state
expected_next_batch_from_iterator1 = None
if saved_iterator_state is not None: # Ensure state was actually saved
  try:
    # TODO: expected_next_batch_from_iterator1 = next(...)
    # YOUR CODE HERE
    expected_next_batch_from_iterator1 = {} # Replace this
    print(f"Expected next batch (from iterator1 after get_state) - first label: {expected_next_batch_from_iterator1.get('label', [None])[0]}")
  except StopIteration:
    print("iterator1 exhausted when trying to get expected_next_batch.")

# 5. Simulate Resumption
# TODO: Get iterator2 from the same data_loader_ex7
# iterator2 = iter(...)
# YOUR CODE HERE
iterator2 = iter([]) # Replace this

if saved_iterator_state is not None:
  # TODO: iterator2.set_state(...)
  # YOUR CODE HERE
  print("\n--- Resumed Iteration (iterator2) ---")
  print("Iterator state restored to iterator2.")
else:
  print("\nSkipping resumption, saved_iterator_state is None.")

# 6. Iterate once from iterator2
resumed_batch = None
if saved_iterator_state is not None: # Only if state was set
  try:
    # TODO: resumed_batch = next(...)
    # YOUR CODE HERE
    resumed_batch = {} # Replace this
    print(f"Resumed batch (from iterator2 after set_state) - first label: {resumed_batch.get('label', [None])[0]}")
  except StopIteration:
    print("iterator2 exhausted immediately after set_state.")

# 7. Verify
if expected_next_batch_from_iterator1 is not None and resumed_batch is not None:
  # Compare 'image' data of the first element in the batch
  # TODO: Perform comparison (e.g., np.allclose on image data)
  # are_identical = np.allclose(...)
  # YOUR CODE HERE
  are_identical = False # Replace this

  if are_identical:
    print("\nSUCCESS: Resumed batch is identical to the expected next batch. Checkpointing works!")
  else:
    print("\nFAILURE: Resumed batch differs from the expected next batch.")
    # print(f"Expected image data (sample): {expected_next_batch_from_iterator1['image'][0,0,0,:3]}")
    # print(f"Resumed image data (sample): {resumed_batch['image'][0,0,0,:3]}")
elif saved_iterator_state is not None:
  print("\nVerification inconclusive: could not obtain both expected and resumed batches.")
else:
  print("DataLoader for Ex7 not configured.")

# @title Exercise 7: Solution
# 1. Set up DataLoader
# Redefine MySource for Ex7.
class MySource(grain.RandomAccessDataSource):
  def __init__(self, num_records: int = 1000):
    self._num_records = num_records
  def __len__(self) -> int:
    return self._num_records
  def __getitem__(self, idx: int) -> Dict[str, Any]:
    effective_idx = idx % self._num_records
    image = np.ones((32, 32, 3), dtype=np.uint8) * (effective_idx % 255)
    label = effective_idx % 10
    # Add original_index for easier verification if needed, though label might suffice
    return {'image': image, 'label': label, 'original_index': effective_idx}
print("Redefined MySource for Ex7.")

source_ex7 = MySource(num_records=1000)
sampler_ex7 = grain.IndexSampler(
  num_records=len(source_ex7),
  shard_options=grain.NoSharding(),
  shuffle=True, # Shuffling makes the test more robust
  num_epochs=None, # Indefinite iteration
  seed=42
  )

# Redefine ConvertToFloat for Ex7.
class ConvertToFloat(grain.MapTransform):
  def map(self, features: Dict[str, Any]) -> Dict[str, Any]:
    updated_features = features.copy()
    updated_features['image'] = features['image'].astype(np.float32) / 255.0
    return updated_features
print("Redefined ConvertToFloat for Ex7.")

transformations_ex7 = [
  ConvertToFloat(),
  grain.Batch(batch_size=64, drop_remainder=True)
  ]

data_loader_ex7 = grain.DataLoader(
  data_source=source_ex7,
  operations=transformations_ex7,
  sampler=sampler_ex7,
  worker_count=0, # Simpler for state verification, but works with >0 too
  shard_options=grain.NoSharding(),
  read_options=grain.ReadOptions(num_threads=0)
  )
print("DataLoader for Ex7 configured.")

# 2. Get iterator1
iterator1 = iter(data_loader_ex7)

# 3. Iterate a few times
num_initial_iterations = 3
print(f"--- Initial Iteration (iterator1) for {num_initial_iterations} batches ---")
last_batch_iterator1 = None
for i in range(num_initial_iterations):
  try:
    last_batch_iterator1 = next(iterator1)
    # Using 'original_index' for more robust check than just 'label'
    print(f"iterator1, batch {i+1} - first original_index: {last_batch_iterator1['original_index'][0]}")
  except StopIteration:
    print("iterator1 exhausted prematurely.")
    break

# 4. Save State
# Make a deep copy if you plan to continue using iterator1 and don't want
# its state object to be modified if Python passes by reference internally (usually not an issue for simple state).
# For PyGrainDatasetIterator, get_state() returns a new state object.
saved_iterator_state = iterator1.get_state()
print(f"\nIterator state saved. Type: {type(saved_iterator_state)}")

# For verification: get the next batch from iterator1 after saving state
expected_next_batch_from_iterator1 = None
if saved_iterator_state is not None:
  try:
    expected_next_batch_from_iterator1 = next(iterator1)
    print(f"Expected next batch (from iterator1 after get_state) - first original_index: {expected_next_batch_from_iterator1['original_index'][0]}")
  except StopIteration:
    print("iterator1 exhausted when trying to get expected_next_batch.")

# 5. Simulate Resumption
# Get a new iterator from the same DataLoader instance.
iterator2 = iter(data_loader_ex7)
if saved_iterator_state is not None:
  iterator2.set_state(saved_iterator_state)
  print("\n--- Resumed Iteration (iterator2) ---")
  print("Iterator state restored to iterator2.")
else:
  print("\nSkipping resumption, saved_iterator_state is None.")

# 6. Iterate once from iterator2
resumed_batch = None
if saved_iterator_state is not None:
  try:
    resumed_batch = next(iterator2)
    print(f"Resumed batch (from iterator2 after set_state) - first original_index: {resumed_batch['original_index'][0]}")
  except StopIteration:
    print("iterator2 exhausted immediately after set_state. This means the saved state was at the very end.")

# 7. Verify
if expected_next_batch_from_iterator1 is not None and resumed_batch is not None:
  # Compare 'image' data and 'original_index' of the first element in the batch for robustness
  # (Labels might repeat, indices are better for this check if available)
  expected_img_sample = expected_next_batch_from_iterator1['image'][0]
  resumed_img_sample = resumed_batch['image'][0]
  expected_idx_sample = expected_next_batch_from_iterator1['original_index'][0]
  resumed_idx_sample = resumed_batch['original_index'][0]
  are_indices_identical = (expected_idx_sample == resumed_idx_sample)
  are_images_identical = np.allclose(expected_img_sample, resumed_img_sample)

  are_identical = are_indices_identical and are_images_identical

  if are_identical:
      print("\nSUCCESS: Resumed batch is identical to the expected next batch. Checkpointing works!")
  else:
      print("\nFAILURE: Resumed batch differs from the expected next batch.")
      if not are_indices_identical:
          print(f"  - Mismatch in first original_index: Expected {expected_idx_sample}, Got {resumed_idx_sample}")
      if not are_images_identical:
          print(f"  - Mismatch in image data for first element.")
          # print(f"    Expected image data (sample [0,0,0]): {expected_img_sample[0,0,0]}")
          # print(f"    Resumed image data (sample [0,0,0]): {resumed_img_sample[0,0,0]}")
elif saved_iterator_state is not None: # If state was saved but verification couldn't complete
  if expected_next_batch_from_iterator1 is None and resumed_batch is None:
    print("\nVERIFICATION NOTE: Both iterators seem to be at the end of the dataset after the initial iterations. This is valid if the dataset was short.")
  else:
    print("\nVerification inconclusive: could not obtain both expected and resumed batches for comparison.")
    print(f" expected_next_batch_from_iterator1 is None: {expected_next_batch_from_iterator1 is None}")
    print(f" resumed_batch is None: {resumed_batch is None}")
else: # If DataLoader itself wasn't configured
  print("DataLoader for Ex7 not configured.")