lixiangchun
commited on
Commit
·
25f71fa
1
Parent(s):
32d3c5a
first commit
Browse files- README.md +9 -0
- data/log-reads-200-patients-trn200-val200-test622-tiny.txt +1 -0
- data/test.csv.gz +3 -0
- data/trn.csv.gz +3 -0
- data/val.csv.gz +3 -0
- model.py +81 -0
- sequence_embedding/config.json +29 -0
- sequence_embedding/pytorch_model.bin +3 -0
- sequence_embedding/special_tokens_map.json +6 -0
- sequence_embedding/tokenizer.json +153 -0
- sequence_embedding/tokenizer_config.json +8 -0
- train.py +258 -0
- train.sh +10 -0
README.md
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# DECIDIA-code
|
2 |
+
Source code for weakly supervised classification of cancer versus control exclusively from bisulfite sequencing reads.
|
3 |
+
|
4 |
+
|
5 |
+
## Training
|
6 |
+
```bash
|
7 |
+
bash train.sh
|
8 |
+
```
|
9 |
+
|
data/log-reads-200-patients-trn200-val200-test622-tiny.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
epoch train_loss train_acc val_loss val_acc eval_loss eval_acc
|
data/test.csv.gz
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:43aacd0406822240dd50dbc6c24123fe188f5672c1243574eec0643b38c44d6b
|
3 |
+
size 15856551
|
data/trn.csv.gz
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:04d798b3dd0735b3a777abcfacd1fe71600a49a8e11666bf410ee307a6f8c176
|
3 |
+
size 25539107
|
data/val.csv.gz
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:6dc2370881027c81e206cabbf3d361ba2f28ccb94ed7b1e801762bea579eed63
|
3 |
+
size 5028508
|
model.py
ADDED
@@ -0,0 +1,81 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
|
5 |
+
def exists(val):
|
6 |
+
return val is not None
|
7 |
+
|
8 |
+
def initialize_weights(module):
|
9 |
+
for m in module.modules():
|
10 |
+
if isinstance(m, nn.Linear):
|
11 |
+
nn.init.xavier_normal_(m.weight)
|
12 |
+
if m.bias is not None:
|
13 |
+
m.bias.data.zero_()
|
14 |
+
|
15 |
+
"""
|
16 |
+
Attention Network with Sigmoid Gating (3 fc layers)
|
17 |
+
args:
|
18 |
+
L: input feature dimension
|
19 |
+
D: hidden layer dimension
|
20 |
+
dropout: whether to use dropout (p = 0.25)
|
21 |
+
n_classes: number of classes
|
22 |
+
"""
|
23 |
+
class Attn_Net_Gated(nn.Module):
|
24 |
+
|
25 |
+
def __init__(self, L = 1024, D = 256, n_tasks = 1):
|
26 |
+
super(Attn_Net_Gated, self).__init__()
|
27 |
+
self.attention_a = nn.Sequential(nn.Linear(L, D), nn.Tanh(), nn.Dropout(0.25))
|
28 |
+
self.attention_b = nn.Sequential(nn.Linear(L, D), nn.Sigmoid(), nn.Dropout(0.25))
|
29 |
+
self.attention_c = nn.Linear(D, n_tasks)
|
30 |
+
|
31 |
+
def forward(self, x):
|
32 |
+
a = self.attention_a(x)
|
33 |
+
b = self.attention_b(x)
|
34 |
+
A = a.mul(b)
|
35 |
+
A = self.attention_c(A) # N x n_classes
|
36 |
+
return A, x
|
37 |
+
|
38 |
+
|
39 |
+
"""
|
40 |
+
Code borrow from: https://github.com/mahmoodlab/TOAD
|
41 |
+
|
42 |
+
args:
|
43 |
+
gate: whether to use gating in attention network
|
44 |
+
size_args: size config of attention network
|
45 |
+
dropout: whether to use dropout in attention network
|
46 |
+
n_classes: number of classes
|
47 |
+
"""
|
48 |
+
|
49 |
+
class DeepAttnMIL(nn.Module):
|
50 |
+
|
51 |
+
def __init__(self, input_dim = 1024, size_arg = "big", n_classes = 2):
|
52 |
+
super(DeepAttnMIL, self).__init__()
|
53 |
+
self.size_dict = {"small": [input_dim, 512, 256], "big": [input_dim, 512, 384]}
|
54 |
+
size = self.size_dict[size_arg]
|
55 |
+
|
56 |
+
self.attention_net = nn.Sequential(
|
57 |
+
nn.Linear(size[0], size[1]),
|
58 |
+
nn.ReLU(),
|
59 |
+
nn.Dropout(0.25),
|
60 |
+
Attn_Net_Gated(L = size[1], D = size[2], n_tasks = 1))
|
61 |
+
|
62 |
+
self.classifier = nn.Linear(size[1], n_classes)
|
63 |
+
|
64 |
+
initialize_weights(self)
|
65 |
+
|
66 |
+
def forward(self, h, return_features=False, attention_only=False):
|
67 |
+
A, h = self.attention_net(h)
|
68 |
+
A = torch.transpose(A, 1, 0)
|
69 |
+
if attention_only:
|
70 |
+
return A[0]
|
71 |
+
|
72 |
+
A = F.softmax(A, dim=1)
|
73 |
+
M = torch.mm(A, h)
|
74 |
+
|
75 |
+
if return_features:
|
76 |
+
return M
|
77 |
+
|
78 |
+
logits = self.classifier(M)
|
79 |
+
|
80 |
+
return logits
|
81 |
+
|
sequence_embedding/config.json
ADDED
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_name_or_path": "facebook/opt",
|
3 |
+
"_remove_final_layer_norm": false,
|
4 |
+
"activation_dropout": 0.0,
|
5 |
+
"activation_function": "relu",
|
6 |
+
"architectures": [
|
7 |
+
"OPTForCausalLM"
|
8 |
+
],
|
9 |
+
"attention_dropout": 0.0,
|
10 |
+
"bos_token_id": 2,
|
11 |
+
"do_layer_norm_before": true,
|
12 |
+
"dropout": 0.1,
|
13 |
+
"eos_token_id": 2,
|
14 |
+
"ffn_dim": 1536,
|
15 |
+
"hidden_size": 384,
|
16 |
+
"init_std": 0.02,
|
17 |
+
"layerdrop": 0.0,
|
18 |
+
"max_position_embeddings": 512,
|
19 |
+
"model_type": "opt",
|
20 |
+
"num_attention_heads": 12,
|
21 |
+
"num_hidden_layers": 1,
|
22 |
+
"pad_token_id": 1,
|
23 |
+
"prefix": "</s>",
|
24 |
+
"torch_dtype": "float32",
|
25 |
+
"transformers_version": "4.21.1",
|
26 |
+
"use_cache": true,
|
27 |
+
"vocab_size": 30,
|
28 |
+
"word_embed_proj_dim": 384
|
29 |
+
}
|
sequence_embedding/pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:fd73e58e7cfeb3faf6b3b3062ad16295066c115a9f1a57ed08d3aa5ce495fe6d
|
3 |
+
size 7942569
|
sequence_embedding/special_tokens_map.json
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"bos_token": "</s>",
|
3 |
+
"eos_token": "</s>",
|
4 |
+
"pad_token": "<pad>",
|
5 |
+
"unk_token": "</s>"
|
6 |
+
}
|
sequence_embedding/tokenizer.json
ADDED
@@ -0,0 +1,153 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"version": "1.0",
|
3 |
+
"truncation": null,
|
4 |
+
"padding": null,
|
5 |
+
"added_tokens": [
|
6 |
+
{
|
7 |
+
"id": 0,
|
8 |
+
"content": "<s>",
|
9 |
+
"single_word": false,
|
10 |
+
"lstrip": false,
|
11 |
+
"rstrip": false,
|
12 |
+
"normalized": false,
|
13 |
+
"special": true
|
14 |
+
},
|
15 |
+
{
|
16 |
+
"id": 1,
|
17 |
+
"content": "<pad>",
|
18 |
+
"single_word": false,
|
19 |
+
"lstrip": false,
|
20 |
+
"rstrip": false,
|
21 |
+
"normalized": false,
|
22 |
+
"special": true
|
23 |
+
},
|
24 |
+
{
|
25 |
+
"id": 2,
|
26 |
+
"content": "</s>",
|
27 |
+
"single_word": false,
|
28 |
+
"lstrip": false,
|
29 |
+
"rstrip": false,
|
30 |
+
"normalized": false,
|
31 |
+
"special": true
|
32 |
+
},
|
33 |
+
{
|
34 |
+
"id": 3,
|
35 |
+
"content": "<unk>",
|
36 |
+
"single_word": false,
|
37 |
+
"lstrip": false,
|
38 |
+
"rstrip": false,
|
39 |
+
"normalized": false,
|
40 |
+
"special": true
|
41 |
+
}
|
42 |
+
],
|
43 |
+
"normalizer": null,
|
44 |
+
"pre_tokenizer": {
|
45 |
+
"type": "Whitespace"
|
46 |
+
},
|
47 |
+
"post_processor": {
|
48 |
+
"type": "TemplateProcessing",
|
49 |
+
"single": [
|
50 |
+
{
|
51 |
+
"SpecialToken": {
|
52 |
+
"id": "</s>",
|
53 |
+
"type_id": 0
|
54 |
+
}
|
55 |
+
},
|
56 |
+
{
|
57 |
+
"Sequence": {
|
58 |
+
"id": "A",
|
59 |
+
"type_id": 0
|
60 |
+
}
|
61 |
+
}
|
62 |
+
],
|
63 |
+
"pair": [
|
64 |
+
{
|
65 |
+
"Sequence": {
|
66 |
+
"id": "A",
|
67 |
+
"type_id": 0
|
68 |
+
}
|
69 |
+
},
|
70 |
+
{
|
71 |
+
"Sequence": {
|
72 |
+
"id": "B",
|
73 |
+
"type_id": 1
|
74 |
+
}
|
75 |
+
}
|
76 |
+
],
|
77 |
+
"special_tokens": {
|
78 |
+
"</s>": {
|
79 |
+
"id": "</s>",
|
80 |
+
"ids": [
|
81 |
+
2
|
82 |
+
],
|
83 |
+
"tokens": [
|
84 |
+
"</s>"
|
85 |
+
]
|
86 |
+
},
|
87 |
+
"<pad>": {
|
88 |
+
"id": "<pad>",
|
89 |
+
"ids": [
|
90 |
+
1
|
91 |
+
],
|
92 |
+
"tokens": [
|
93 |
+
"<pad>"
|
94 |
+
]
|
95 |
+
},
|
96 |
+
"<s>": {
|
97 |
+
"id": "<s>",
|
98 |
+
"ids": [
|
99 |
+
0
|
100 |
+
],
|
101 |
+
"tokens": [
|
102 |
+
"<s>"
|
103 |
+
]
|
104 |
+
},
|
105 |
+
"<unk>": {
|
106 |
+
"id": "<unk>",
|
107 |
+
"ids": [
|
108 |
+
3
|
109 |
+
],
|
110 |
+
"tokens": [
|
111 |
+
"<unk>"
|
112 |
+
]
|
113 |
+
}
|
114 |
+
}
|
115 |
+
},
|
116 |
+
"decoder": null,
|
117 |
+
"model": {
|
118 |
+
"type": "WordLevel",
|
119 |
+
"vocab": {
|
120 |
+
"<s>": 0,
|
121 |
+
"<pad>": 1,
|
122 |
+
"</s>": 2,
|
123 |
+
"<unk>": 3,
|
124 |
+
"A": 4,
|
125 |
+
"B": 5,
|
126 |
+
"C": 6,
|
127 |
+
"D": 7,
|
128 |
+
"E": 8,
|
129 |
+
"F": 9,
|
130 |
+
"G": 10,
|
131 |
+
"H": 11,
|
132 |
+
"I": 12,
|
133 |
+
"J": 13,
|
134 |
+
"K": 14,
|
135 |
+
"L": 15,
|
136 |
+
"M": 16,
|
137 |
+
"N": 17,
|
138 |
+
"O": 18,
|
139 |
+
"P": 19,
|
140 |
+
"Q": 20,
|
141 |
+
"R": 21,
|
142 |
+
"S": 22,
|
143 |
+
"T": 23,
|
144 |
+
"U": 24,
|
145 |
+
"V": 25,
|
146 |
+
"W": 26,
|
147 |
+
"X": 27,
|
148 |
+
"Y": 28,
|
149 |
+
"Z": 29
|
150 |
+
},
|
151 |
+
"unk_token": "<unk>"
|
152 |
+
}
|
153 |
+
}
|
sequence_embedding/tokenizer_config.json
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"bos_token": "</s>",
|
3 |
+
"eos_token": "</s>",
|
4 |
+
"model_max_length": 2048,
|
5 |
+
"pad_token": "<pad>",
|
6 |
+
"tokenizer_class": "PreTrainedTokenizerFast",
|
7 |
+
"unk_token": "</s>"
|
8 |
+
}
|
train.py
ADDED
@@ -0,0 +1,258 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/opt/software/install/miniconda38/bin/python
|
2 |
+
import argparse
|
3 |
+
parser = argparse.ArgumentParser(description='DECIDIA training program')
|
4 |
+
parser.add_argument('--input_dir', type=str, help='input directory')
|
5 |
+
parser.add_argument('--sequence_embedding', type=str, help='sequence embedding directory')
|
6 |
+
parser.add_argument('--num_hidden_layers', type=int, default=1, help='num_hidden_layers [1]')
|
7 |
+
parser.add_argument('--train_file', type=str, help='training file')
|
8 |
+
parser.add_argument('--val_file', type=str, help='validation file')
|
9 |
+
parser.add_argument('--device', type=str, help='device', default='cuda:1')
|
10 |
+
parser.add_argument('--num_classes', type=int, help='num_classes [32]', default=32)
|
11 |
+
parser.add_argument('--diseases', type=str, default=None, help='diseases included, e.g "LUAD,LUSC"')
|
12 |
+
parser.add_argument('--weight_decay', type=float, help='weight_decay [1e-5]', default=1e-5)
|
13 |
+
parser.add_argument('--modeling_context', action='store_true', help='whether use OPT to model context dependency')
|
14 |
+
parser.add_argument("--lr_scheduler_type", type=str,
|
15 |
+
choices=["linear", "cosine", "cosine_with_restarts", "polynomial", "constant", "constant_with_warmup"],
|
16 |
+
default="constant", help="The scheduler type to use.")
|
17 |
+
parser.add_argument('--pretrained_weight', type=str, help='pretrained weight')
|
18 |
+
parser.add_argument('--pretrained_cls_token', type=str, help='pretrained cls token')
|
19 |
+
parser.add_argument('--epochs', type=int, default=100, help='epochs (default: 100)')
|
20 |
+
parser.add_argument('--num_sequences', type=int, default=None, help='num of sequences to sample from training set')
|
21 |
+
parser.add_argument('--num_train_patients', type=int, default=None, help='num of patients data to sample from training set')
|
22 |
+
|
23 |
+
args = parser.parse_args()
|
24 |
+
|
25 |
+
import os
|
26 |
+
os.environ['TOKENIZERS_PARALLELISM'] = 'false'
|
27 |
+
import sys
|
28 |
+
import glob
|
29 |
+
import torch
|
30 |
+
import torch.nn as nn
|
31 |
+
from tqdm import tqdm
|
32 |
+
from torch.optim import AdamW, Adam, SGD, Adagrad
|
33 |
+
from sklearn.utils import resample
|
34 |
+
from transformers import get_scheduler
|
35 |
+
import numpy as np
|
36 |
+
import pandas as pd
|
37 |
+
import random
|
38 |
+
import time
|
39 |
+
from transformers import (
|
40 |
+
PreTrainedTokenizerFast,
|
41 |
+
OPTForCausalLM
|
42 |
+
)
|
43 |
+
from model import DeepAttnMIL
|
44 |
+
|
45 |
+
torch.set_num_threads(2)
|
46 |
+
device = args.device
|
47 |
+
random.seed(123)
|
48 |
+
|
49 |
+
tokenizer = PreTrainedTokenizerFast.from_pretrained(args.sequence_embedding)
|
50 |
+
net = OPTForCausalLM.from_pretrained(args.sequence_embedding)
|
51 |
+
net = net.to(device)
|
52 |
+
net.eval()
|
53 |
+
|
54 |
+
feature_dim = net.config.hidden_size
|
55 |
+
|
56 |
+
trn_df = pd.read_csv(f'{args.input_dir}/trn.csv.gz')
|
57 |
+
reads_per_patient = trn_df.patient.value_counts().unique()
|
58 |
+
assert len(reads_per_patient) == 1
|
59 |
+
reads_per_patient = reads_per_patient[0]
|
60 |
+
if args.num_sequences < reads_per_patient:
|
61 |
+
trn_df = pd.concat([df.sample(args.num_sequences, random_state=123) for patient, df in trn_df.groupby('patient')])
|
62 |
+
|
63 |
+
num_train_samples = len(trn_df.patient.unique())
|
64 |
+
if args.num_train_patients is None:
|
65 |
+
args.num_train_patients = num_train_samples
|
66 |
+
if args.num_train_patients < num_train_samples:
|
67 |
+
trn_df = trn_df[trn_df.patient.isin(random.sample(trn_df.patient.unique().tolist(), args.num_train_patients))]
|
68 |
+
|
69 |
+
trn_x = torch.zeros(args.num_train_patients, args.num_sequences, feature_dim)
|
70 |
+
trn_y = torch.as_tensor([-1] * args.num_train_patients)
|
71 |
+
|
72 |
+
test_df = pd.read_csv(f'{args.input_dir}/test.csv.gz')
|
73 |
+
num_test_samples = len(test_df.patient.unique())
|
74 |
+
test_x = torch.zeros(num_test_samples, reads_per_patient, feature_dim)
|
75 |
+
test_y = torch.as_tensor([-1] * num_test_samples)
|
76 |
+
test_patients = []
|
77 |
+
|
78 |
+
val_df = pd.read_csv(f'{args.input_dir}/val.csv.gz')
|
79 |
+
num_val_samples = len(val_df.patient.unique())
|
80 |
+
val_x = torch.zeros(num_val_samples, reads_per_patient, feature_dim)
|
81 |
+
val_y = torch.as_tensor([-1] * num_val_samples)
|
82 |
+
val_patients = []
|
83 |
+
|
84 |
+
|
85 |
+
pad_token_id = net.config.pad_token_id
|
86 |
+
|
87 |
+
|
88 |
+
for i, (patient, e) in tqdm(enumerate(trn_df.groupby('patient')), total=args.num_train_patients):
|
89 |
+
a = [' '.join(list(s)) for s in e.seq]
|
90 |
+
inputs = tokenizer(a, max_length=100, padding='max_length', truncation=True, return_tensors='pt', return_token_type_ids=False)
|
91 |
+
for k, v in inputs.items():inputs[k] = v.to(device)
|
92 |
+
with torch.inference_mode():
|
93 |
+
out = net.model(**inputs)
|
94 |
+
features = out.last_hidden_state.mean(1).cpu()
|
95 |
+
trn_x[i] = features
|
96 |
+
trn_y[i] = e.label.tolist()[0]
|
97 |
+
|
98 |
+
|
99 |
+
for i, (patient, e) in tqdm(enumerate(test_df.groupby('patient')), total=num_test_samples):
|
100 |
+
a = [' '.join(list(s)) for s in e.seq]
|
101 |
+
inputs = tokenizer(a, max_length=100, padding='max_length', truncation=True, return_tensors='pt', return_token_type_ids=False)
|
102 |
+
for k, v in inputs.items():inputs[k] = v.to(device)
|
103 |
+
with torch.inference_mode():
|
104 |
+
out = net.model(**inputs)
|
105 |
+
features = out.last_hidden_state.mean(1).cpu()
|
106 |
+
test_x[i] = features
|
107 |
+
test_y[i] = e.label.tolist()[0]
|
108 |
+
test_patients.append(patient)
|
109 |
+
|
110 |
+
for i, (patient, e) in tqdm(enumerate(val_df.groupby('patient')), total=num_val_samples):
|
111 |
+
a = [' '.join(list(s)) for s in e.seq]
|
112 |
+
inputs = tokenizer(a, max_length=100, padding='max_length', truncation=True, return_tensors='pt', return_token_type_ids=False)
|
113 |
+
for k, v in inputs.items():inputs[k] = v.to(device)
|
114 |
+
with torch.inference_mode():
|
115 |
+
out = net.model(**inputs)
|
116 |
+
features = out.last_hidden_state.mean(1).cpu()
|
117 |
+
val_x[i] = features
|
118 |
+
val_y[i] = e.label.tolist()[0]
|
119 |
+
val_patients.append(patient)
|
120 |
+
|
121 |
+
|
122 |
+
|
123 |
+
fout = open(f'{args.input_dir}/log-reads-{args.num_sequences}-patients-trn{args.num_train_patients}-val{num_val_samples}-test{num_test_samples}-tiny.txt', 'w')
|
124 |
+
print("epoch\ttrain_loss\ttrain_acc\tval_loss\tval_acc\teval_loss\teval_acc", file=fout)
|
125 |
+
|
126 |
+
model = DeepAttnMIL(input_dim=feature_dim, n_classes=args.num_classes, size_arg='big')
|
127 |
+
|
128 |
+
|
129 |
+
if args.pretrained_weight:
|
130 |
+
state_dict = torch.load(args.pretrained_weight, map_location='cpu')
|
131 |
+
if state_dict['classifier.weight'].size(0) != args.num_classes:
|
132 |
+
del state_dict['classifier.weight']
|
133 |
+
del state_dict['classifier.bias']
|
134 |
+
|
135 |
+
msg = model.load_state_dict(state_dict, strict=False)
|
136 |
+
print(msg)#, file=fout)
|
137 |
+
|
138 |
+
model = model.to(device)
|
139 |
+
|
140 |
+
print(model)#, file=fout)
|
141 |
+
|
142 |
+
|
143 |
+
criterion = nn.CrossEntropyLoss()
|
144 |
+
|
145 |
+
no_decay = ["bias", "LayerNorm.weight"]
|
146 |
+
optimizer_grouped_parameters = [
|
147 |
+
{
|
148 |
+
"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
|
149 |
+
"weight_decay": 1e-5,
|
150 |
+
},
|
151 |
+
{
|
152 |
+
"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)],
|
153 |
+
"weight_decay": 0.0,
|
154 |
+
},
|
155 |
+
]
|
156 |
+
opt = AdamW(optimizer_grouped_parameters, lr=2e-5)
|
157 |
+
|
158 |
+
|
159 |
+
num_update_steps_per_epoch = len(trn_df)
|
160 |
+
max_train_steps = args.epochs * num_update_steps_per_epoch
|
161 |
+
lr_scheduler = get_scheduler(name=args.lr_scheduler_type, optimizer=opt, num_warmup_steps=num_update_steps_per_epoch*1, num_training_steps=max_train_steps)
|
162 |
+
|
163 |
+
|
164 |
+
best_eval_acc = 0.0
|
165 |
+
best_eval_loss = 100000.0
|
166 |
+
best_val_loss = 100000.0
|
167 |
+
for epoch in range(args.epochs):
|
168 |
+
model.train()
|
169 |
+
total_loss, total_batch, total_num, correct_k = 0, 0, 0, 0
|
170 |
+
idxs = random.sample(range(len(trn_y)), len(trn_y))
|
171 |
+
for idx in idxs:
|
172 |
+
x = trn_x[idx]
|
173 |
+
y = trn_y[idx].unsqueeze(0)
|
174 |
+
x = x.to(device)
|
175 |
+
y = y.to(device)
|
176 |
+
|
177 |
+
logit = model(x)
|
178 |
+
loss = criterion(logit, y)
|
179 |
+
|
180 |
+
opt.zero_grad()
|
181 |
+
loss.backward()
|
182 |
+
opt.step()
|
183 |
+
lr_scheduler.step()
|
184 |
+
|
185 |
+
total_loss += loss.item()
|
186 |
+
total_batch += 1
|
187 |
+
total_num += len(y)
|
188 |
+
correct_k += logit.argmax(1).eq(y).sum()
|
189 |
+
|
190 |
+
train_acc = correct_k / total_num
|
191 |
+
train_loss = total_loss / total_batch
|
192 |
+
|
193 |
+
#######Evalutate on test set ################
|
194 |
+
model.eval()
|
195 |
+
total_loss, total_batch, total_num, correct_k = 0, 0, 0, 0
|
196 |
+
eval_probs = []
|
197 |
+
for x, y, pid in zip(test_x, test_y, test_patients):
|
198 |
+
y = y.unsqueeze(0).to(device)
|
199 |
+
x = x.to(device)
|
200 |
+
|
201 |
+
with torch.inference_mode():
|
202 |
+
logit = model(x)
|
203 |
+
loss = criterion(logit, y)
|
204 |
+
|
205 |
+
eval_probs.append(logit.flatten().softmax(0).tolist())
|
206 |
+
|
207 |
+
total_loss += loss.item()
|
208 |
+
total_batch += 1
|
209 |
+
total_num += len(y)
|
210 |
+
correct_k += logit.argmax(1).eq(y).sum()
|
211 |
+
|
212 |
+
eval_acc = correct_k / total_num
|
213 |
+
eval_loss = total_loss / total_batch
|
214 |
+
|
215 |
+
#######Evalutate on val set ################
|
216 |
+
model.eval()
|
217 |
+
total_loss, total_batch, total_num, correct_k = 0, 0, 0, 0
|
218 |
+
val_probs = []
|
219 |
+
for x, y, pid in zip(val_x, val_y, val_patients):
|
220 |
+
y = y.unsqueeze(0).to(device)
|
221 |
+
x = x.to(device)
|
222 |
+
|
223 |
+
with torch.inference_mode():
|
224 |
+
logit = model(x)
|
225 |
+
loss = criterion(logit, y)
|
226 |
+
|
227 |
+
val_probs.append(logit.flatten().softmax(0).tolist())
|
228 |
+
|
229 |
+
total_loss += loss.item()
|
230 |
+
total_batch += 1
|
231 |
+
total_num += len(y)
|
232 |
+
correct_k += logit.argmax(1).eq(y).sum()
|
233 |
+
|
234 |
+
val_acc = correct_k / total_num
|
235 |
+
val_loss = total_loss / total_batch
|
236 |
+
|
237 |
+
|
238 |
+
print(f"{epoch+1}\t{train_loss}\t{train_acc}\t{val_loss}\t{val_acc}\t{eval_loss}\t{eval_acc}", file=fout)
|
239 |
+
fout.flush()
|
240 |
+
|
241 |
+
if val_loss < best_val_loss:
|
242 |
+
torch.save(model.state_dict(), f'{args.input_dir}/model-reads-{args.num_sequences}-patients-trn{args.num_train_patients}-val{num_val_samples}-test{num_test_samples}-tiny.pt')
|
243 |
+
best_val_loss = val_loss
|
244 |
+
|
245 |
+
eval_probs = pd.DataFrame(eval_probs, columns=['p_normal', 'p_cancer'])
|
246 |
+
info = pd.DataFrame({'patient':test_patients, 'label':test_y.tolist()})
|
247 |
+
info = pd.concat([info, eval_probs], axis=1)
|
248 |
+
info.to_csv(f'{args.input_dir}/test_prediction-reads-{args.num_sequences}-patients-trn{args.num_train_patients}-val{num_val_samples}-test{num_test_samples}-tiny.csv', index=False)
|
249 |
+
|
250 |
+
val_probs = pd.DataFrame(val_probs, columns=['p_normal', 'p_cancer'])
|
251 |
+
info = pd.DataFrame({'patient':val_patients, 'label':val_y.tolist()})
|
252 |
+
info = pd.concat([info, val_probs], axis=1)
|
253 |
+
info.to_csv(f'{args.input_dir}/val_prediction-reads-{args.num_sequences}-patients-trn{args.num_train_patients}-val{num_val_samples}-test{num_test_samples}-tiny.csv', index=False)
|
254 |
+
|
255 |
+
fout.close()
|
256 |
+
|
257 |
+
|
258 |
+
|
train.sh
ADDED
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/bin/bash
|
2 |
+
|
3 |
+
python train.py \
|
4 |
+
--sequence_embedding sequence_embedding \
|
5 |
+
--input_dir data \
|
6 |
+
--num_classes 2 \
|
7 |
+
--device 'cuda:0' \
|
8 |
+
--num_train_patients 200 \
|
9 |
+
--num_sequences 200
|
10 |
+
|