/
make_cluster_plots.py
212 lines (186 loc) · 8.13 KB
/
make_cluster_plots.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
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
import os
import yaml
from yaml import Loader
import glob
import copy
import pickle
from probe import get_results
import matplotlib.pyplot as plt
from sklearn.manifold import TSNE
from sklearn.decomposition import PCA
import numpy as np
from make_video import prepare_video
import matplotlib.cm as cm
def plot_clusters(
results,
folder,
indices=range(2),
hidden_key="hidden_outputs",
reduction="tsne",
layer_wise=False,
factors=None,
fill_types=None,
colors=None,
):
assert len(results) == 500, f"len(results)={len(results)}"
for i in indices:
example = copy.deepcopy(results[i])
# offset
# example[hidden_key] = [h[300:] for h in example[hidden_key]]
# example["states"] = example["states"][300:]
# example["input"] = example["input"][300:]
# t_hidden = len(example[hidden_key][0])
t_states = len(example["states"])
# take hidden states upto t_states
example[hidden_key] = [h[:t_states].reshape(t_states, -1) for h in example[hidden_key]]
n_layers = len(example[hidden_key])
alphabet = sorted(set(example["input"]).difference({"|"}))
states = sorted(set(example["states"]).difference({"-1"}))
reducer = TSNE if reduction == "tsne" else PCA
# breakpoint()
if layer_wise:
Xs = [
reducer(n_components=2).fit_transform(hidden_outputs)
for hidden_outputs in example[hidden_key]
]
X = np.concatenate(Xs, axis=0)
X = X.reshape(n_layers, t_states, 2)
else:
hidden_outputs = np.concatenate(example[hidden_key], axis=0)
X = reducer(n_components=2).fit_transform(hidden_outputs)
X = X.reshape(n_layers, t_states, 2)
# X = X[:, :t_states, :]
for t in range(20, t_states, 10):
a, _ = factors[n_layers]
axes, fig = plt.subplots(a, a, figsize=(16, 16))
axes.suptitle(
f"T={t}, #states: "
+ str(len(example["dfa"].dfa._transition_function))
+ ", len(vocab): "
+ str(len(alphabet))
)
for layer in range(n_layers - 1, -1, -1):
X_layer = X[layer, :t]
labels = np.array(example["states"][:t])
chars = np.array(list(example["input"][:t]))
# vocab = example["vocab"][1:]
ax = fig[layer // a, layer % a]
for label in states:
if label == -1 or label == "-1":
continue
for ci, v in enumerate(alphabet):
indices = np.where((labels == label) & (chars == v))
if len(indices) > 0:
ax.scatter(
X_layer[indices, 0],
X_layer[indices, 1],
marker=fill_types[states.index(label)],
c=colors[ci],
)
# show legend
# set title
ax.set_title(f"layer={layer}")
plt.savefig(os.path.join(folder, f"e_{i}_{reduction}_t_{t}.jpg"))
plt.close()
video_out_folder = os.path.join(folder, "video")
video_glob = folder + "/" + f"e_{i}_{reduction}_t_*.jpg"
prepare_video(video_glob, video_out_folder, f"e_{i}_{reduction}")
print(f"done at {video_out_folder}")
return video_out_folder
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--exp", type=str, default="s4d")
parser.add_argument("--hidden_key", type=str, default="hidden_outputs")
args = parser.parse_args()
exp_folders_40000 = {
"s4d": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_40000/s4d/generations/144_test.txt",
"rwkv": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_40000/rwkv/generations/110_test.txt",
"retnet": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_40000/retention/generations/36_test.txt",
"lstm": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_40000/lstm/generations/174_test.txt",
"linear_transformer": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_40000/linear_transformer/generations/117_test.txt",
"hyena": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_40000/hyena/generations/28_test.txt",
"h3": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_40000/h3/generations/57_test.txt",
"transformer/1": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_40000/transformer_1/generations/133_test.txt",
"transformer/2": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_40000/transformer_2/generations/177_test.txt",
"transformer/12": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_40000/transformer/generations/184_test.txt",
"transformer/4": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_40000/transformer_4/generations/194_test.txt",
"transformer/8": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_40000/transformer_8_w_hiddens/generations/174_test.txt",
}
exp_folders_2500 = {
"s4d": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_2500/s4d/generations/54_test.txt",
"rwkv": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_2500/rwkv/generations/20_test.txt",
"retnet": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_2500/retention/generations/106_test.txt",
"lstm": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_2500/lstm/generations/142_test.txt",
"linear_transformer": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_2500/linear_transformer/generations/40_test.txt",
"hyena": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_2500/hyena/generations/59_test.txt",
"h3": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_2500/h3/generations/47_test.txt",
"transformer/8": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_2500/transformer/generations/176_test.txt",
"transformer/2": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_2500/transformer_2/generations/197_test.txt",
"transformer/4": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_2500/transformer_4/generations/155_test.txt",
"transformer/1": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_2500/transformer_1/generations/13_val.txt"
# "transformer/2": "/raid/lingo/akyurek/git/iclmodels/experiments/
# "transformer/12": "/raid/lingo/akyurek/git/iclmodels/experiments/
# "transformer/4": "/raid/lingo/akyurek/git/iclmodels/experiments/
}
exp_folders_5000 = {
"transformer/8": "/raid/lingo/akyurek/git/iclmodels/experiments/hiddens_5000/transformer/generations/194_test.txt"
}
fill_types = (
"o",
"8",
"*",
"v",
"X",
"^",
"<",
">",
"s",
"p",
"h",
"H",
"D",
"d",
"P",
)
colors = cm.Set1.colors + (
(0.0, 0.0, 0.0),
cm.Set2.colors[0],
cm.tab20b.colors[-1],
cm.tab20b.colors[-2],
cm.tab20b.colors[-3],
)
factors = {
4: (2, 2),
9: (3, 3),
8: (3, 3),
13: (4, 4),
5: (3, 3),
3: (2, 2),
1: (1, 1),
}
exp_folder = exp_folders_2500[args.exp]
print(exp_folder)
print(args)
results = get_results(
exp_folder, subset="test", key=args.hidden_key, in_states=True
)
# video_folders = []
# for exp_name, folder in exp_folders.items():
exp_folder_main = os.path.dirname(exp_folder)
plot_folder = os.path.join(exp_folder_main, "plots", args.hidden_key)
print(plot_folder)
# make
os.makedirs(plot_folder, exist_ok=True)
output_video_folder = plot_clusters(
results,
plot_folder,
hidden_key=args.hidden_key,
indices=[0,],
reduction="pca",
layer_wise=True,
factors=factors,
fill_types=fill_types,
colors=colors,
)
print(output_video_folder)