nabla_ml/
nab_model.rs

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
use crate::nab_array::NDArray;
use crate::nab_layers::NabLayer;
use crate::nab_optimizers::NablaOptimizer;
use crate::nab_loss::NabLoss;
use std::collections::HashMap;
use serde::{Serialize, Deserialize};
use std::path::Path;
use serde_json;
use flate2::write::GzEncoder;
use flate2::read::GzDecoder;
use flate2::Compression;
use std::io::{Write, Read};

static mut NEXT_NODE_ID: usize = 0;

/// Represents a node in the computation graph
pub struct Node {
    pub layer: NabLayer,
    pub inputs: Vec<usize>,  // Indices of input nodes
    pub output_shape: Vec<usize>,
}

/// Represents a model using the Functional API
/// 
/// # Examples
/// 
/// ```rust
/// use nabla_ml::nab_model::NabModel;
/// use nabla_ml::nab_layers::NabLayer;
/// 
/// // Create model architecture
/// let input = NabModel::input(vec![784]);
/// let dense1 = NabLayer::dense(784, 512, Some("relu"), Some("dense1"));
/// let x = input.apply(dense1);
/// let output_layer = NabLayer::dense(512, 10, Some("softmax"), Some("output"));
/// let output = x.apply(output_layer);
/// 
/// // Create and compile model
/// let mut model = NabModel::new_functional(vec![input], vec![output]);
/// model.compile(
///     "sgd",
///     0.1,
///     "categorical_crossentropy",
///     vec!["accuracy".to_string()]
/// );
/// ```
#[allow(dead_code)]
#[derive(Clone)]
pub struct NabModel {
    layers: Vec<NabLayer>,
    optimizer_type: String,
    learning_rate: f64,
    loss_type: String,  // e.g. "mse", "categorical_crossentropy"
    metrics: Vec<String>,
}

/// Represents an input node in the computation graph
#[derive(Clone)]
pub struct Input {
    shape: Vec<usize>,
    node_index: Option<usize>,
}

/// Represents an output node in the computation graph
#[derive(Clone)]
#[allow(dead_code)]
pub struct Output {
    layer: NabLayer,
    inputs: Vec<usize>,
    output_shape: Vec<usize>,
    previous_output: Option<Box<Output>>,
}

impl Input {
    /// Applies a layer to this input, preserving node connectivity
    pub fn apply<L: Into<NabLayer>>(&self, layer: L) -> Output {
        let mut layer = layer.into();
        let output_shape = layer.compute_output_shape(&self.shape);
        
        // Get next ID safely
        let layer_id = unsafe {
            NEXT_NODE_ID += 1;
            NEXT_NODE_ID
        };
        
        layer.set_node_index(layer_id);
        
        println!("Connecting layer {} (id: {}) to input (id: {})", 
            layer.get_name(), 
            layer_id,
            self.node_index.unwrap()
        );
        
        Output {
            layer,
            inputs: vec![self.node_index.unwrap()],
            output_shape,
            previous_output: None,
        }
    }

    /// Returns the input shape of this Input node
    /// 
    /// # Returns
    /// 
    /// A reference to the shape vector
    pub fn get_input_shape(&self) -> &Vec<usize> {
        &self.shape
    }
}

impl Output {
    /// Applies a layer to this output, maintaining the graph structure
    pub fn apply<L: Into<NabLayer>>(&self, layer: L) -> Output {
        let mut layer = layer.into();
        let output_shape = layer.compute_output_shape(&self.output_shape);
        
        // Get next ID safely
        let layer_id = unsafe {
            NEXT_NODE_ID += 1;
            NEXT_NODE_ID
        };
        
        layer.set_node_index(layer_id);
        
        println!("Connecting layer {} (id: {}) to {} (id: {})", 
            layer.get_name(), 
            layer_id,
            self.layer.get_name(),
            self.layer.node_index.unwrap()
        );
        
        Output {
            layer,
            inputs: vec![self.layer.node_index.unwrap()],
            output_shape,
            previous_output: Some(Box::new(self.clone())),
        }
    }

    /// Returns the previous layer that produced this output
    pub fn get_previous_layer(&self) -> Option<&NabLayer> {
        // Return layer that produced this output
        None // TODO: Implement layer tracking
    }
}

#[allow(dead_code)]
impl NabModel {
    /// Creates a new input layer with specified shape
    /// 
    /// # Arguments
    /// * `shape` - Shape of input excluding batch dimension
    /// 
    /// # Examples
    /// ```ignore
    /// let input = NabModel::input(vec![784]); // For MNIST images
    /// ```
    pub fn input(shape: Vec<usize>) -> Input {
        let node_index = unsafe {
            NEXT_NODE_ID += 1;
            NEXT_NODE_ID
        };
        
        Input {
            shape,
            node_index: Some(node_index),
        }
    }

    /// Creates a new model
    pub fn new() -> Self {
        NabModel {
            layers: Vec::new(),
            optimizer_type: String::new(),
            learning_rate: 0.0,
            loss_type: String::new(),
            metrics: Vec::new(),
        }
    }

    /// Adds a layer to the model
    pub fn add(&mut self, layer: NabLayer) -> &mut Self {
        self.layers.push(layer);
        self
    }

    /// Compiles the model with training configuration
    /// 
    /// # Arguments
    /// * `optimizer_type` - Optimization algorithm ("sgd", "adam", etc)
    /// * `learning_rate` - Learning rate for optimization
    /// * `loss_type` - Loss function ("mse", "categorical_crossentropy")
    /// * `metrics` - Metrics to track during training
    pub fn compile(&mut self, optimizer_type: &str, learning_rate: f64, 
                  loss_type: &str, metrics: Vec<String>) {
        self.optimizer_type = optimizer_type.to_string();
        self.learning_rate = learning_rate;
        self.loss_type = loss_type.to_string();
        self.metrics = metrics;
    }

    /// Trains for one epoch
    fn train_epoch(&mut self, x: &NDArray, y: &NDArray, batch_size: usize) -> HashMap<String, f64> {
        let mut metrics = HashMap::new();
        let mut total_loss = 0.0;
        let mut total_correct = 0;
        let num_samples = x.shape()[0];
        let num_batches = (num_samples + batch_size - 1) / batch_size;

        // Process mini-batches
        for batch_idx in 0..num_batches {
            let start_idx = batch_idx * batch_size;
            let end_idx = (start_idx + batch_size).min(num_samples);
            
            // Get batch data
            let x_batch = x.slice(start_idx, end_idx);
            let y_batch = y.slice(start_idx, end_idx);
            
            // Forward and backward pass as one operation
            let (predictions, loss) = self.forward_backward(&x_batch, &y_batch);
            
            // Accumulate metrics
            total_loss += loss * (end_idx - start_idx) as f64;
            total_correct += self.count_correct(&predictions, &y_batch);
        }
        
        // Calculate average metrics
        metrics.insert("loss".to_string(), total_loss / num_samples as f64);
        metrics.insert("accuracy".to_string(), total_correct as f64 / num_samples as f64);
        
        metrics
    }

    fn forward_backward(&mut self, x_batch: &NDArray, y_batch: &NDArray) -> (NDArray, f64) {
        // Forward pass
        let predictions = self.predict(x_batch);
        let loss = self.calculate_loss(&predictions, y_batch);
        let loss_grad = self.calculate_loss_gradient(&predictions, y_batch);
        
        // Backward pass
        let mut gradient = loss_grad;
        let learning_rate = self.learning_rate;  // Cache learning rate
        
        for layer in self.layers.iter_mut().rev() {
            if layer.is_trainable() {
                gradient = layer.backward(&gradient);
                
                // Update weights using cached learning rate
                if let Some(weights) = layer.weights.as_mut() {
                    let weight_grads = layer.weight_gradients.as_ref().unwrap();
                    NablaOptimizer::sgd_update(weights, weight_grads, learning_rate);
                }
                if let Some(biases) = layer.biases.as_mut() {
                    let bias_grads = layer.bias_gradients.as_ref().unwrap();
                    NablaOptimizer::sgd_update(biases, bias_grads, learning_rate);
                }
            }
        }
        
        (predictions, loss)
    }

    fn count_correct(&self, predictions: &NDArray, targets: &NDArray) -> usize {
        let pred_classes = predictions.argmax(Some(1));
        let true_classes = targets.argmax(Some(1));
        
        pred_classes.iter()
            .zip(true_classes.iter())
            .filter(|(&p, &t)| p == t)
            .count()
    }

    /// Creates a new model from input and output nodes
    pub fn new_functional(inputs: Vec<Input>, outputs: Vec<Output>) -> Self {
        let mut layers = Vec::new();
        let mut visited = std::collections::HashSet::new();
        
        // First add input layers
        for input in inputs {
            let mut layer = NabLayer::input(input.shape.clone(), None);
            layer.set_node_index(input.node_index.unwrap());
            visited.insert(input.node_index.unwrap());
            layers.push(layer);
        }
        
        // Then add remaining layers by traversing backwards from each output
        for output in outputs {
            let mut current = Some(output);
            let mut layer_stack = Vec::new();
            
            // Build stack of layers from output to input
            while let Some(curr) = current {
                if !visited.contains(&curr.layer.node_index.unwrap()) {
                    visited.insert(curr.layer.node_index.unwrap());
                    layer_stack.push(curr.layer);
                }
                current = curr.previous_output.map(|prev| *prev);
            }
            
            // Add layers in reverse order (from input to output)
            layers.extend(layer_stack.into_iter().rev());
        }

        NabModel {
            layers,
            optimizer_type: String::new(),
            learning_rate: 0.0,
            loss_type: String::new(),
            metrics: Vec::new(),
        }
    }

    /// Trains the model on input data
    /// 
    /// # Arguments
    /// * `x_train` - Training features
    /// * `y_train` - Training labels 
    /// * `batch_size` - Mini-batch size
    /// * `epochs` - Number of training epochs
    /// * `validation_data` - Optional validation dataset
    /// 
    /// # Returns
    /// HashMap containing training history metrics
    /// 
    /// # Examples
    /// ```ignore
    /// let history = model.fit(
    ///     &x_train,
    ///     &y_train, 
    ///     64,    // batch_size
    ///     5,     // epochs
    ///     Some((&x_test, &y_test))
    /// );
    /// ```
    pub fn fit(&mut self, x_train: &NDArray, y_train: &NDArray,
               batch_size: usize, epochs: usize,
               validation_data: Option<(&NDArray, &NDArray)>) 
               -> HashMap<String, Vec<f64>> {
        let mut history = HashMap::new();
        let mut train_metrics = Vec::new();
        let mut val_metrics = Vec::new();

        for epoch in 0..epochs {
            // Training phase
            let metrics = self.train_epoch(x_train, y_train, batch_size);
            train_metrics.push(metrics);

            // Validation phase
            if let Some((x_val, y_val)) = validation_data {
                let val_metric = self.evaluate(x_val, y_val, batch_size);
                val_metrics.push(val_metric);
            }

            // Print progress
            self.print_progress(epoch + 1, epochs, &train_metrics[epoch], 
                              val_metrics.last());
        }

        // Store history
        history.insert("loss".to_string(), 
            train_metrics.iter().map(|m| m["loss"]).collect());
        history.insert("accuracy".to_string(), 
            train_metrics.iter().map(|m| m["accuracy"]).collect());

        if !val_metrics.is_empty() {
            history.insert("val_loss".to_string(), 
                val_metrics.iter().map(|m| m["loss"]).collect());
            history.insert("val_accuracy".to_string(), 
                val_metrics.iter().map(|m| m["accuracy"]).collect());
        }

        history
    }

    /// Prints training progress
    fn print_progress(
        &self,
        epoch: usize,
        total_epochs: usize,
        train_metrics: &HashMap<String, f64>,
        val_metrics: Option<&HashMap<String, f64>>,
    ) {
        print!("Epoch {}/{} - ", epoch, total_epochs);
        for (name, value) in train_metrics {
            print!("{}: {:.4} ", name, value);
        }
        if let Some(val_metrics) = val_metrics {
            for (name, value) in val_metrics {
                print!("val_{}: {:.4} ", name, value);
            }
        }
        println!();
    }

    /// Evaluates model performance on test data
    /// 
    /// # Arguments
    /// * `x_test` - Test features
    /// * `y_test` - Test labels
    /// * `batch_size` - Batch size for evaluation
    /// 
    /// # Returns
    /// HashMap containing evaluation metrics
    #[allow(unused_variables)]
    pub fn evaluate(&mut self, x_test: &NDArray, y_test: &NDArray,
                   batch_size: usize) -> HashMap<String, f64> {
        let mut metrics = HashMap::new();
        let predictions = self.predict(x_test);
        
        // Calculate loss
        let loss = self.calculate_loss(&predictions, y_test);
        metrics.insert("loss".to_string(), loss);
        
        // Calculate other metrics
        for metric in &self.metrics {
            match metric.as_str() {
                "accuracy" => {
                    let acc = self.calculate_accuracy(&predictions, y_test);
                    metrics.insert("accuracy".to_string(), acc);
                }
                _ => {}
            }
        }
        
        metrics
    }

    /// Calculates accuracy for classification tasks
    fn calculate_accuracy(&self, predictions: &NDArray, targets: &NDArray) -> f64 {
        let pred_classes = predictions.argmax(Some(1));
        let true_classes = targets.argmax(Some(1));
        
        let correct = pred_classes.iter()
            .zip(true_classes.iter())
            .filter(|(&p, &t)| p == t)
            .count();
            
        correct as f64 / predictions.shape()[0] as f64
    }

    /// Makes predictions on input data
    /// 
    /// # Arguments
    /// * `x` - Input features to predict on
    /// 
    /// # Returns
    /// NDArray of model predictions
    pub fn predict(&mut self, x: &NDArray) -> NDArray {
        let mut current = x.clone();
        for layer in &mut self.layers {
            current = layer.forward(&current, false);
        }
        current
    }

    fn calculate_loss(&self, predictions: &NDArray, targets: &NDArray) -> f64 {
        match self.loss_type.as_str() {
            "mse" => NabLoss::mean_squared_error(predictions, targets),
            "categorical_crossentropy" => NabLoss::cross_entropy_loss(predictions, targets),
            _ => NabLoss::mean_squared_error(predictions, targets),
        }
    }

    fn calculate_loss_gradient(&self, predictions: &NDArray, targets: &NDArray) -> NDArray {
        match self.loss_type.as_str() {
            "mse" => predictions.subtract(targets).divide_scalar(predictions.shape()[0] as f64),
            "categorical_crossentropy" => predictions.subtract(targets).divide_scalar(predictions.shape()[0] as f64),
            _ => predictions.subtract(targets).divide_scalar(predictions.shape()[0] as f64),
        }
    }

    // Add debug method
    pub fn print_layers(&self) {
        println!("\nLayer stack:");
        for (i, layer) in self.layers.iter().enumerate() {
            println!("{}: {} -> {:?}", i, layer.get_name(), layer.get_output_shape());
        }
    }

    /// Prints a summary of the model's layers and parameters
    /// 
    /// Displays a formatted table showing:
    /// - Layer name and type
    /// - Output shape
    /// - Number of parameters
    /// 
    /// Also shows total parameters, trainable parameters, and non-trainable parameters
    /// 
    /// # Example
    /// ```ignore
    /// use nabla_ml::nab_model::NabModel;
    /// use nabla_ml::nab_layers::NabLayer;
    /// 
    /// let input = NabModel::input(vec![784]);
    /// let dense = NabLayer::dense(784, 128, Some("relu"), Some("dense1"));
    /// let output = input.apply(dense);
    /// let mut model = NabModel::new_functional(vec![input], vec![output]);
    /// 
    /// model.summary();
    /// // Model: "sequential"
    /// // ─────────────────────────────────────────────────────
    /// // Layer (type)          Output Shape         Param #   
    /// // =================================================
    /// // input                 (None, 784)          0         
    /// // dense1 (Dense)        (None, 128)          100,480   
    /// // =================================================
    /// // Total params: 100,480
    /// // Trainable params: 100,480
    /// // Non-trainable params: 0
    /// ```
    pub fn summary(&self) {
        println!("Model: \"functional\"");
        println!("─────────────────────────────────────────────────────");
        println!("{:<20} {:<18} {:<10}", "Layer (type)", "Output Shape", "Param #");
        println!("=================================================");

        let mut total_params = 0;
        let mut trainable_params = 0;
        let mut non_trainable_params = 0;

        // Print each layer's info
        for layer in &self.layers {
            let (params, trainable) = self.count_params(layer);
            total_params += params;
            if trainable {
                trainable_params += params;
            } else {
                non_trainable_params += params;
            }

            let shape_str = format!("(None, {})", 
                layer.get_output_shape()
                    .iter()
                    .map(|x| x.to_string())
                    .collect::<Vec<_>>()
                    .join(", ")
            );

            let layer_type = if layer.get_name().contains("input") {
                layer.get_name().to_string()
            } else {
                format!("{} ({})", 
                    layer.get_name(),
                    layer.get_type()
                )
            };

            println!("{:<20} {:<18} {:<10}", 
                layer_type,
                shape_str,
                self.format_number(params)
            );
        }

        println!("=================================================");
        println!("Total params: {}", self.format_number(total_params));
        println!("Trainable params: {}", self.format_number(trainable_params));
        println!("Non-trainable params: {}", self.format_number(non_trainable_params));
    }

    /// Counts parameters for a given layer
    fn count_params(&self, layer: &NabLayer) -> (usize, bool) {
        let mut params = 0;
        
        // Count weights
        if let Some(weights) = &layer.weights {
            params += weights.data().len();
        }
        
        // Count biases
        if let Some(biases) = &layer.biases {
            params += biases.data().len();
        }

        (params, layer.is_trainable())
    }

    /// Formats large numbers with commas
    fn format_number(&self, n: usize) -> String {
        n.to_string()
            .chars()
            .rev()
            .collect::<Vec<_>>()
            .chunks(3)
            .map(|chunk| chunk.iter().collect::<String>())
            .collect::<Vec<_>>()
            .join(",")
            .chars()
            .rev()
            .collect()
    }

    /// Saves the model to a compressed .ez file
    /// 
    /// # Arguments
    /// * `path` - Path to save the model (e.g. "model.ez")
    pub fn save_compressed<P: AsRef<Path>>(&self, path: P) -> std::io::Result<()> {
        // Create encoder with best compression
        let file = std::fs::File::create(path)?;
        let mut encoder = GzEncoder::new(file, Compression::best());
        
        // Prepare and serialize model data
        let model_data = ModelData {
            config: ModelConfig {
                optimizer_type: self.optimizer_type.clone(),
                learning_rate: self.learning_rate,
                loss_type: self.loss_type.clone(),
                metrics: self.metrics.clone(),
            },
            layers: self.layers.iter().map(|layer| LayerState {
                layer_type: layer.get_type().to_string(),
                name: layer.get_name().to_string(),
                input_shape: layer.input_shape.clone(),
                output_shape: layer.output_shape.clone(),
                weights: layer.weights.as_ref().map(|w| w.data().to_vec()),
                biases: layer.biases.as_ref().map(|b| b.data().to_vec()),
                activation: layer.activation.clone(),
            }).collect(),
        };

        // Write serialized data
        let serialized = serde_json::to_string(&model_data)?;
        encoder.write_all(serialized.as_bytes())?;
        encoder.finish()?;

        Ok(())
    }

    /// Loads a model from a compressed .ez file
    /// 
    /// # Arguments
    /// * `path` - Path to the model file (e.g. "model.ez")
    pub fn load_compressed<P: AsRef<Path>>(path: P) -> std::io::Result<Self> {
        // Create decoder
        let file = std::fs::File::open(path)?;
        let mut decoder = GzDecoder::new(file);
        
        // Read and decompress
        let mut contents = String::new();
        decoder.read_to_string(&mut contents)?;
        
        // Deserialize
        let model_data: ModelData = serde_json::from_str(&contents)?;
        
        // Reconstruct model
        let mut layers = Vec::new();
        for state in model_data.layers {
            let mut layer = match state.layer_type.as_str() {
                "Input" => NabLayer::input(state.input_shape.clone(), Some(&state.name)),
                "Dense" => NabLayer::dense(
                    state.input_shape[0],
                    state.output_shape[0],
                    state.activation.as_deref(),
                    Some(&state.name)
                ),
                _ => return Err(std::io::Error::new(
                    std::io::ErrorKind::InvalidData,
                    format!("Unknown layer type: {}", state.layer_type)
                )),
            };

            // Restore weights and biases
            if let Some(weights) = state.weights {
                let weight_shape = match state.layer_type.as_str() {
                    "Dense" => vec![state.input_shape[0], state.output_shape[0]],
                    _ => state.input_shape.clone()
                };
                layer.weights = Some(NDArray::new(weights, weight_shape));
            }
            if let Some(biases) = state.biases {
                layer.biases = Some(NDArray::new(biases, vec![state.output_shape[0]]));
            }

            layers.push(layer);
        }

        Ok(NabModel {
            layers,
            optimizer_type: model_data.config.optimizer_type,
            learning_rate: model_data.config.learning_rate,
            loss_type: model_data.config.loss_type,
            metrics: model_data.config.metrics,
        })
    }
}

/// Serializable model configuration
#[derive(Serialize, Deserialize)]
struct ModelConfig {
    optimizer_type: String,
    learning_rate: f64,
    loss_type: String,
    metrics: Vec<String>,
}

/// Serializable layer state
#[derive(Serialize, Deserialize)]
struct LayerState {
    layer_type: String,
    name: String,
    input_shape: Vec<usize>,
    output_shape: Vec<usize>,
    weights: Option<Vec<f64>>,
    biases: Option<Vec<f64>>,
    activation: Option<String>,
}

#[derive(Serialize, Deserialize)]
struct ModelData {
    config: ModelConfig,
    layers: Vec<LayerState>,
}

/// Resets the global node ID counter
/// Used for testing to ensure consistent behavior
pub fn reset_node_id() {
    unsafe {
        NEXT_NODE_ID = 0;
    }
}

#[cfg(test)]
#[allow(unused_imports)]
#[allow(unused_variables)]
mod tests {
    use super::*;
    use crate::nab_activations::NablaActivation;
    use crate::nab_optimizers::NablaOptimizer;
    use crate::nab_loss::NabLoss;
    use crate::nab_mnist::NabMnist;
    use crate::nab_utils::NabUtils;

    #[test]
    fn test_linear_regression() {
        // Reset node ID counter before test
        reset_node_id();
        
        // Create synthetic data for linear regression
        // y = 2x + 1 with some noise
        let x_data = NDArray::from_matrix(vec![
            vec![1.0], vec![2.0], vec![3.0], vec![4.0], vec![5.0]
        ]);
        let y_data = NDArray::from_matrix(vec![
            vec![3.1], vec![5.0], vec![6.9], vec![9.2], vec![11.0]
        ]);

        // Create model architecture
        let input = NabModel::input(vec![1]);
        let output_layer = NabLayer::dense(1, 1, None, Some("linear_output"));
        let output = input.apply(output_layer);

        // Create and compile model
        let mut model = NabModel::new_functional(vec![input], vec![output]);
        model.compile(
            "sgd",
            0.01,
            "mse",
            vec!["mse".to_string()]
        );

        // Train model for multiple epochs
        for _ in 0..100 {  // Increase training iterations
            model.train_epoch(&x_data, &y_data, x_data.shape()[0]); // Use full batch
        }
        
        // Make predictions
        let predictions = model.predict(&x_data);
        
        // Verify predictions follow roughly linear pattern
        let pred_vec = predictions.data();
        for i in 1..pred_vec.len() {
            assert!(pred_vec[i] > pred_vec[i-1], 
                "Predictions should increase monotonically. Found {} <= {} at index {}", 
                pred_vec[i], pred_vec[i-1], i
            );
        }
    }


    /// Tests full training pipeline on MNIST dataset
    /// 
    /// This test:
    /// 1. Loads and preprocesses MNIST data
    /// 2. Creates a neural network with:
    ///    - Input layer (784 units)
    ///    - Dense layer (512 units, ReLU)
    ///    - Dense layer (256 units, ReLU) 
    ///    - Output layer (10 units, softmax)
    /// 3. Compiles with SGD optimizer and cross-entropy loss
    /// 4. Trains for 5 epochs
    /// 5. Verifies accuracy exceeds 85%
    #[test]
    fn test_mnist_full_pipeline() {
    //     // Step 1: Load MNIST data
    //     println!("Loading MNIST data...");
    //     let ((x_train, y_train), (x_test, y_test)) = NabUtils::load_and_split_dataset("datasets/mnist_test", 80.0).unwrap();

    //     // Step 2: Normalize input data (scale pixels to 0-1)
    //     println!("Normalizing data...");
    //     let x_train = x_train.divide_scalar(255.0);
    //     let x_test = x_test.divide_scalar(255.0);

    //     // Step 2.5: Reshape input data
    //     let x_train = x_train.reshape(&[x_train.shape()[0], 784])
    //         .expect("Failed to reshape training data");
    //     let x_test = x_test.reshape(&[x_test.shape()[0], 784])
    //         .expect("Failed to reshape test data");

    //     // Step 2.6: One-hot encode target data
    //     println!("One-hot encoding targets...");
    //     let y_train = NDArray::one_hot_encode(&y_train);
    //     let y_test = NDArray::one_hot_encode(&y_test);
            

    //     println!("Data shapes:");
    //     println!("x_train: {:?}", x_train.shape());
    //     println!("y_train: {:?}", y_train.shape());
    //     println!("x_test: {:?}", x_test.shape());
    //     println!("y_test: {:?}", y_test.shape());

    //     // Step 3: Create model architecture
    //     println!("Creating model...");
    //     let input = NabModel::input(vec![784]);  // 28x28 = 784 pixels

    //     // Dense layer with 512 units and ReLU activation
    //     let dense1 = NabLayer::dense(784, 512, Some("relu"), Some("dense1"));
    //     let x = input.apply(dense1);

    //     // Dense layer with 256 units and ReLU activation
    //     let dense2 = NabLayer::dense(512, 256, Some("relu"), Some("dense2"));
    //     let x = x.apply(dense2);

    //     // Output layer with 10 units (one per digit) and softmax activation
    //     let output_layer = NabLayer::dense(256, 10, Some("softmax"), Some("output"));
    //     let output = x.apply(output_layer);

    //     // Step 4: Create and compile model
    //     println!("Compiling model...");
    //     let mut model = NabModel::new_functional(vec![input], vec![output]);
    //     model.compile(
    //         "sgd",                      
    //         0.1,                        // Increase learning rate from 0.01 to 0.1
    //         "categorical_crossentropy", 
    //         vec!["accuracy".to_string()]
    //     );

    //     // Step 5: Train model
    //     println!("Training model...");
    //     let history = model.fit(
    //         &x_train,
    //         &y_train,
    //         64,             // Increase batch size from 32 to 64
    //         5,             // Increase epochs from 2 to 10
    //         Some((&x_test, &y_test))
    //     );

    //     // Step 6: Evaluate final model
    //     println!("Evaluating model...");
    //     let eval_metrics = model.evaluate(&x_test, &y_test, 32);
        
    //     // Print final results
    //     println!("Final test accuracy: {:.2}%", eval_metrics["accuracy"] * 100.0);
        
    //     // Verify model achieved reasonable accuracy (>85%)
    //     assert!(eval_metrics["accuracy"] > 0.85, 
    //         "Model accuracy ({:.2}%) below expected threshold", 
    //         eval_metrics["accuracy"] * 100.0
    //     );

    //     // Verify training history contains expected metrics
    //     assert!(history.contains_key("loss"));
    //     assert!(history.contains_key("accuracy"));
    //     assert!(history.contains_key("val_loss"));
    //     assert!(history.contains_key("val_accuracy"));
    }

    #[test]
    fn test_model_summary() {
        // Reset node ID counter before test
        reset_node_id();
        
        // Create a simple model
        let input = NabModel::input(vec![784]);
        let dense1 = NabLayer::dense(784, 32, Some("relu"), Some("dense1"));
        let x = input.apply(dense1);

        let dense2 = NabLayer::dense(32, 32, Some("relu"), Some("dense2"));
        let x = x.apply(dense2);

        let output_layer = NabLayer::dense(32, 10, Some("softmax"), Some("output"));
        let output = x.apply(output_layer);
        
        let model = NabModel::new_functional(vec![input], vec![output]);

        // Capture stdout to verify summary output
        let output = std::io::stdout();
        let handle = output.lock();
        
        model.summary();

        // Verify parameter counts
        let total_params: usize = model.layers.iter()
            .map(|l| model.count_params(l).0)
            .sum();
        
        assert_eq!(total_params, 784*32 + 32 + 32*32 + 32 + 32*10 + 10); // weights + biases
    }

    #[test]
    fn test_model_save_load() {
        // Reset node ID counter before test
        reset_node_id();
        
        // Create a simple model
        let input = NabModel::input(vec![784]);
        let dense1 = NabLayer::dense(784, 32, Some("relu"), Some("dense1"));
        let x = input.apply(dense1);

        let dense2 = NabLayer::dense(32, 32, Some("relu"), Some("dense2"));
        let x = x.apply(dense2);

        let output_layer = NabLayer::dense(32, 10, Some("softmax"), Some("output"));
        let output = x.apply(output_layer);

        let mut model = NabModel::new_functional(vec![input], vec![output]);
        model.compile("sgd", 0.1, "categorical_crossentropy", vec!["accuracy".to_string()]);


        // Save the model
        model.save_compressed("test_model.ez").expect("Failed to save model");

        // Load the model
        let loaded_model = NabModel::load_compressed("test_model.ez").expect("Failed to load model");

        // Verify model configuration
        assert_eq!(loaded_model.optimizer_type, model.optimizer_type);
        assert_eq!(loaded_model.learning_rate, model.learning_rate);
        assert_eq!(loaded_model.loss_type, model.loss_type);
        assert_eq!(loaded_model.metrics, model.metrics);

        // Verify layers
        assert_eq!(loaded_model.layers.len(), model.layers.len());
        for (loaded, original) in loaded_model.layers.iter().zip(model.layers.iter()) {
            assert_eq!(loaded.get_type(), original.get_type());
            assert_eq!(loaded.get_output_shape(), original.get_output_shape());
            
            if let (Some(w1), Some(w2)) = (&loaded.weights, &original.weights) {
                assert_eq!(w1.shape(), w2.shape(), "Weight shapes don't match");
                assert!(w1.data().iter().zip(w2.data().iter())
                    .all(|(a, b)| (a - b).abs() < 1e-6), 
                    "Weight values don't match");
            }

            if let (Some(b1), Some(b2)) = (&loaded.biases, &original.biases) {
                assert_eq!(b1.shape(), b2.shape(), "Bias shapes don't match");
                assert!(b1.data().iter().zip(b2.data().iter())
                    .all(|(a, b)| (a - b).abs() < 1e-6),
                    "Bias values don't match");
            }
        }

        // Clean up test file
        std::fs::remove_file("test_model.ez").expect("Failed to clean up test file");
    }

    #[test]
    fn test_input_shape() {
        // Reset node ID counter before test
        reset_node_id();
        
        let shape = vec![784, 32];
        let input = NabModel::input(shape.clone());
        
        // Test that get_input_shape returns the correct shape
        assert_eq!(input.get_input_shape(), &shape);
        
        // Test that the shape is preserved when applying a layer
        let dense = NabLayer::dense(784, 128, Some("relu"), Some("dense1"));
        let output = input.apply(dense);
        assert_eq!(input.get_input_shape(), &shape, "Input shape should remain unchanged after applying layer");
    }
}