File size: 3,524 Bytes
fed1832
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python
import os
import torch
import numpy as np
import matplotlib.pyplot as plt

torch.set_printoptions(profile="full")  

FILTER_RATE = 0.95
TOP_RATE = 0.5
ACTIVATION_BAR_RATIO = 0.95

langs = ["en", "eu"]
base_path = "new_activations"

n, over_zero = [], []

model_name = None
checkpoint = None

for lang in langs:
    # File path
    path = os.path.join(base_path, f"activation.{lang}.train.l2-7b-eu.pt")
    data = torch.load(path)
    n.append(data["n"])
    over_zero.append(data["over_zero"])

    # Extract model_name and checkpoint only once
    if model_name is None:
        model_name = os.path.basename(os.path.dirname(path))  # folder name
        filename = os.path.basename(path)
        parts = filename.split('.')
        checkpoint = parts[-1]  # 'qwen-checkpoint-1200'

# Convert to tensors
n = torch.Tensor(n)  # (lang_num)
over_zero = torch.stack(over_zero, dim=-1)  # (layer_num, neuron_num, lang_num)

num_layers, intermediate_size, lang_num = over_zero.size()

# 1. Activation probability
activation_probs = over_zero / n  # broadcast

# 2. Normalized activation probability
normed_activation_probs = activation_probs / activation_probs.sum(dim=-1, keepdim=True)

# 3. LAPE (entropy)
log_prop = torch.where(normed_activation_probs > 0,
                       normed_activation_probs.log(),
                       torch.zeros_like(normed_activation_probs))
entropy = -(normed_activation_probs * log_prop).sum(dim=-1)

# 4. Filter neurons using 95th percentile
flat_probs = activation_probs.flatten()
thresh = flat_probs.kthvalue(int(flat_probs.numel() * FILTER_RATE)).values
valid_mask = (activation_probs > thresh).any(dim=-1)  # [layers, neurons]
entropy[~valid_mask] = float("inf")

# 5. Select top-k neurons with lowest entropy
flat_entropy = entropy.flatten()
topk = int(flat_entropy.numel() * TOP_RATE)
_, idx = flat_entropy.topk(topk, largest=False) 

layer_idx = idx // intermediate_size
neuron_idx = idx % intermediate_size

# 6. Group by languages
selection_props = activation_probs[layer_idx, neuron_idx]  # [topk, lang_num]
bar = flat_probs.kthvalue(int(flat_probs.numel() * ACTIVATION_BAR_RATIO)).values
lang_mask = (selection_props > bar).T  # [lang_num, topk]

final_mask = {}
for i, lang in enumerate(langs):
    neuron_ids = torch.where(lang_mask[i])[0]
    layer_lists = [[] for _ in range(num_layers)]
    for nid in neuron_ids.tolist():
        l = layer_idx[nid].item()
        h = neuron_idx[nid].item()
        layer_lists[l].append(h)
    final_mask[lang] = [torch.tensor(lst, dtype=torch.long) for lst in layer_lists]

# =========================
# Plot number of neurons per layer (bar chart)
# =========================
plt.figure(figsize=(12, 6))
x = np.arange(num_layers)
width = 0.35

bars_list = []
for i, lang_key in enumerate(langs):
    counts = [len(layer) for layer in final_mask[lang_key]]
    bars = plt.bar(x + i * width, counts, width=width, label=lang_key)
    bars_list.append(bars)

    # Thêm số lên mỗi bar
    for bar in bars:
        height = bar.get_height()
        plt.text(bar.get_x() + bar.get_width()/2.0, height, f'{int(height)}',
                 ha='center', va='bottom', fontsize=9)

plt.xlabel("Layer Index")
plt.ylabel("Number of Neurons")
plt.title(f"Number of Language-Specific Neurons per Layer\nModel: {model_name}, Checkpoint: {checkpoint}")
plt.xticks(x + width / 2, x)
plt.legend()
plt.grid(alpha=0.3, axis='y')
plt.tight_layout()

plt.savefig(f"{model_name}_{checkpoint}_neurons_bar.png", dpi=300)
plt.close()