-
Notifications
You must be signed in to change notification settings - Fork 0
/
visualize.py
135 lines (116 loc) · 5.66 KB
/
visualize.py
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
from net.network import Network
from data.dataloader import DataLoader, MyDataset
from data import dataloader
from torch import argmax
import pandas as pd
from parameters import *
import numpy as np
import torch
import matplotlib.pyplot as plt
from IPython.display import display, IFrame
from matplotlib_venn import venn3
from collections import Counter
import json
import os.path as path
import os
def create_venn(train_loader: DataLoader):
sets = {}
data_size = len(train_loader.dataset.data)
for net_params in [("rnn", 2, 256), ("gru", 2, 32), ("lstm", 1, 256)]:
net = Network.load_network(*net_params)
net.eval()
val = set()
offset = 0
for x, y in train_loader:
correct = net.forward(x).argmax(-1) == y.flatten()
val |= {n + offset for n, b in enumerate(correct) if b}
offset += len(y.flatten())
sets[net.network_name()] = val
venn3(sets.values(), sets.keys(), subset_label_formatter=lambda num: f"{(num / data_size):1.0%}")
plt.show()
def create_model_performance_table(test_loader: DataLoader):
model_names = ["lstm", "rnn", "gru"]
num_layerss = [1, 2, 3]
hidden_sizes = [32, 64, 128, 256]
corrects, losses = [], []
loss_func = torch.nn.functional.cross_entropy
for model_name in model_names:
for num_layer in num_layerss:
for hidden_size in hidden_sizes:
net = Network.load_network(model_name, num_layer, hidden_size)
net.eval() # set the network in evaluation mode.
correct, loss = [], []
for x, y in test_loader:
out = net.forward(x)
predictions = out.argmax(-1)
correct += (predictions == y.flatten()).tolist()
loss += [(loss_func(out, y.flatten())).tolist()]
corrects.append(np.mean(correct))
losses.append(np.mean(loss))
for title, arr in (("accuracy", corrects), ("loss", losses)):
arr = np.array(arr).reshape((len(hidden_sizes), len(model_names) * len(num_layerss)))
columns = pd.MultiIndex.from_product([model_names, num_layerss])
df = pd.DataFrame(arr, columns=columns, index=hidden_sizes)
df.columns.name = title
display(df.round(3))
def plot_gate(gates):
num_gates, num_layers, _, _ = gates.shape
color_base = ["red", "green", "blue"]
gate_names = {2: ["Update", "Reset"], 3: ["Input", "Forget", "Output"]}
gate_names = gate_names[num_gates]
fig = plt.figure()
fig.suptitle('Saturation Plot', fontsize=16)
for gate_id, gate in enumerate(gates):
ax = fig.add_subplot(1, num_gates, gate_id + 1, aspect='equal', title=f'{gate_names[gate_id]} Gate',
xlabel='fraction right saturated', ylabel='fraction left saturated')
for layer, (right, left) in enumerate(gate):
clr = Counter(zip(left, right))
l, r = np.array(list(clr.keys())).T
rad = 100 * np.array(list(clr.values())) / max(clr.values())
c = np.full(len(l), color_base[layer])
plt.scatter(l, r, rad, c, alpha=1.0 / num_layers)
ax.plot(np.linspace(0, 1), np.linspace(1, 0))
plt.draw()
plt.show()
def get_saturation(gate, left_thresh, right_thresh):
return (gate < left_thresh).mean(axis=-1), (gate > right_thresh).mean(axis=-1)
def visualize_gate(*gates):
gates_lr_layer_text = np.array([get_saturation(gate, 0.1, 0.9) for gate in gates])
gate_layer_lr_text = gates_lr_layer_text.transpose((0, 2, 1, 3))
plot_gate(gate_layer_lr_text)
def create_gate_plots(model_name, num_layers, hidden_size, test_loader):
network = Network.load_network(model_name, num_layers, hidden_size)
network.eval()
if model_name == "lstm":
input_gates, forget_gates, cell_gates, output_gates, cell_states = network.extract_from_loader(test_loader)
visualize_gate(input_gates, forget_gates, output_gates)
elif model_name == "gru":
reset_gate, update_gate = network.extract_from_loader(test_loader)
visualize_gate(update_gate, reset_gate)
def visualize_cell(cell, x, hidden_size):
char_cell = {'cell_size': hidden_size, 'seq': ''.join(x)}
char_cell.update({f"cell_layer_{layer + 1}": cell[layer].tolist() for layer in range(len(cell))})
with open(path.join("cell_visualization", 'char_cell.json'), 'w') as json_file:
json.dump(char_cell, json_file)
def create_cell_visualization(model_name, num_layers, hidden_size, test_loader: DataLoader, vocabulary):
network = Network.load_network(model_name, num_layers, hidden_size)
network.eval()
begin = 1
last = 844
if model_name == "lstm":
input_gates, forget_gates, cell_gates, output_gates, cell_states = network.extract_from_loader(test_loader)
visualize_cell(cell_states[:, begin:last],
dataloader.long_tensor_as_data(test_loader.dataset.data[begin:last], vocabulary), hidden_size)
elif model_name == "gru":
reset_gates, update_gates = network.extract_from_loader(test_loader)
visualize_cell(reset_gates[:, begin:last],
dataloader.long_tensor_as_data(test_loader.dataset.data[begin:last], vocabulary), hidden_size)
if __name__ == '__main__':
os.environ['KMP_DUPLICATE_LIB_OK'] = 'True' # prevents errors
file_path = "data/warandpeace.txt"
(train, test, val), vocabulary = dataloader.load_data(file_path, SPLITS, BATCH_SIZE, SEQ_LEN, DEVICE)
# Uncomment to reproduce one of the following.
# create_model_performance_table(test)
# create_venn(test)
# create_gate_plots("gru", 3, 128, test)
# create_cell_visualization("lstm", 1, 256, test, vocabulary)