File size: 8,570 Bytes
8d18b7c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Additional Utility Metrics"""

import logging
from typing import Any, Dict, List, Optional

import numpy as np
import torch
from sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score

logger = logging.getLogger(__name__)


def compute_all_metrics(

    predictions: np.ndarray,

    references: np.ndarray,

    task: str = "classification",

) -> Dict[str, float]:
    """Compute comprehensive metrics based on task type."""
    if task == "classification":
        return compute_classification_metrics(predictions, references)
    elif task == "regression":
        return compute_regression_metrics(predictions, references)
    elif task == "code_generation":
        return compute_code_metrics(predictions, references)
    elif task == "reasoning":
        return compute_reasoning_metrics(predictions, references)
    else:
        raise ValueError(f"Unknown task: {task}")


def compute_classification_metrics(

    predictions: np.ndarray,

    references: np.ndarray,

    average: str = "macro",

) -> Dict[str, float]:
    """Compute classification metrics."""
    from sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score

    accuracy = accuracy_score(references, predictions)
    f1 = f1_score(references, predictions, average=average, zero_division=0)
    precision = precision_score(references, predictions, average=average, zero_division=0)
    recall = recall_score(references, predictions, average=average, zero_division=0)

    return {
        "accuracy": accuracy,
        "f1": f1,
        "precision": precision,
        "recall": recall,
    }


def compute_regression_metrics(

    predictions: np.ndarray,

    references: np.ndarray,

) -> Dict[str, float]:
    """Compute regression metrics."""
    mae = mean_absolute_error(references, predictions)
    mse = mean_squared_error(references, predictions)
    rmse = np.sqrt(mse)
    r2 = r2_score(references, predictions)

    return {
        "mae": mae,
        "mse": mse,
        "rmse": rmse,
        "r2": r2,
    }


def compute_code_metrics(

    predictions: List[str],

    references: List[str],

) -> Dict[str, float]:
    """Compute code generation metrics."""
    # Exact match
    exact_matches = sum(p.strip() == r.strip() for p, r in zip(predictions, references))
    exact_match_rate = exact_matches / len(predictions) if predictions else 0.0

    # BLEU score (simplified)
    try:
        from nltk.translate.bleu_score import corpus_bleu

        # Tokenize
        pred_tokens = [p.split() for p in predictions]
        ref_tokens = [[r.split()] for r in references]

        bleu = corpus_bleu(ref_tokens, pred_tokens)
    except ImportError:
        bleu = 0.0

    return {
        "exact_match": exact_match_rate,
        "bleu": bleu,
    }


def compute_reasoning_metrics(

    predictions: List[str],

    references: List[str],

    steps_predictions: Optional[List[List[str]]] = None,

    steps_references: Optional[List[List[str]]] = None,

) -> Dict[str, float]:
    """Compute reasoning-specific metrics."""
    # Exact match
    exact_matches = sum(p.strip() == r.strip() for p, r in zip(predictions, references))
    exact_match_rate = exact_matches / len(predictions) if predictions else 0.0

    # Step-level accuracy if available
    step_accuracy = 0.0
    if steps_predictions and steps_references:
        step_scores = []
        for pred_steps, ref_steps in zip(steps_predictions, steps_references):
            # Jaccard similarity
            pred_set = set(pred_steps)
            ref_set = set(ref_steps)
            if ref_set:
                intersection = pred_set & ref_set
                union = pred_set | ref_set
                step_scores.append(len(intersection) / len(union))
        if step_scores:
            step_accuracy = np.mean(step_scores)

    return {
        "exact_match": exact_match_rate,
        "step_accuracy": step_accuracy,
    }


def compute_perplexity_from_loss(loss: float) -> float:
    """Convert loss to perplexity."""
    return float(torch.exp(torch.tensor(loss)).item())


def compute_parameter_count(model: torch.nn.Module) -> Dict[str, int]:
    """Count parameters by type."""
    total_params = sum(p.numel() for p in model.parameters())
    trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)

    # Count by module type
    module_counts = {}
    for name, module in model.named_modules():
        module_type = type(module).__name__
        if module_type not in module_counts:
            module_counts[module_type] = 0
        module_counts[module_type] += sum(p.numel() for p in module.parameters())

    return {
        "total": total_params,
        "trainable": trainable_params,
        "frozen": total_params - trainable_params,
        "by_module": module_counts,
    }


def compute_flops(

    model: torch.nn.Module,

    input_shape: tuple,

    forward_pass: bool = True,

) -> Dict[str, float]:
    """Estimate FLOPs for a forward pass."""
    # Simplified FLOPs estimation
    # For transformer: 6 * batch_size * seq_len * d_model^2 per layer (approx)
    total_params = sum(p.numel() for p in model.parameters())

    # Rough estimate: 2 * params per token
    batch_size, seq_len = input_shape[0], input_shape[1]
    flops_per_token = 2 * total_params
    total_flops = flops_per_token * seq_len * batch_size

    return {
        "total_flops": total_flops,
        "flops_per_token": flops_per_token,
        "gflops": total_flops / 1e9,
    }


def compute_memory_usage(

    model: torch.nn.Module,

    batch_size: int,

    seq_len: int,

    dtype: str = "bfloat16",

) -> Dict[str, float]:
    """Estimate memory usage."""
    # Parameter memory
    param_bytes = {
        "float32": 4,
        "float16": 2,
        "bfloat16": 2,
        "int8": 1,
        "int4": 0.5,
    }[dtype]

    param_memory = sum(p.numel() for p in model.parameters()) * param_bytes / 1e9  # GB

    # Activation memory (rough estimate: batch_size * seq_len * d_model * 2 * num_layers)
    # Assuming 2x for activations
    d_model = getattr(model.config, "d_model", 2048)
    num_layers = getattr(model.config, "num_hidden_layers", 24)
    activation_memory = batch_size * seq_len * d_model * 2 * num_layers * param_bytes / 1e9

    # Gradient memory (same as parameters if not using gradient checkpointing)
    gradient_memory = param_memory

    total_memory = param_memory + activation_memory + gradient_memory

    return {
        "parameters_gb": param_memory,
        "activations_gb": activation_memory,
        "gradients_gb": gradient_memory,
        "total_gb": total_memory,
    }


def track_gradient_norms(

    model: torch.nn.Module,

    norm_type: float = 2.0,

) -> Dict[str, float]:
    """Compute gradient norms for debugging."""
    total_norm = 0.0
    param_norms = {}

    for name, param in model.named_parameters():
        if param.grad is not None:
            param_norm = param.grad.data.norm(norm_type).item()
            param_norms[name] = param_norm
            total_norm += param_norm ** norm_type

    total_norm = total_norm ** (1.0 / norm_type)

    return {
        "total_grad_norm": total_norm,
        "param_grad_norms": param_norms,
    }


def compute_parameter_distribution(model: torch.nn.Module) -> Dict[str, Any]:
    """Analyze parameter distribution (mean, std, min, max)."""
    stats = {
        "mean": [],
        "std": [],
        "min": [],
        "max": [],
        "num_zeros": [],
    }

    for name, param in model.named_parameters():
        if param.requires_grad:
            data = param.data.cpu().numpy().flatten()
            stats["mean"].append(float(np.mean(data)))
            stats["std"].append(float(np.std(data)))
            stats["min"].append(float(np.min(data)))
            stats["max"].append(float(np.max(data)))
            stats["num_zeros"].append(int(np.sum(data == 0)))

    # Aggregate
    return {
        "overall_mean": float(np.mean(stats["mean"])),
        "overall_std": float(np.mean(stats["std"])),
        "overall_min": float(np.min(stats["min"])),
        "overall_max": float(np.max(stats["max"])),
        "total_zeros": sum(stats["num_zeros"]),
        "zero_percentage": sum(stats["num_zeros"]) / sum(p.numel() for p in model.parameters() if p.requires_grad),
    }