Upload RxRx3-core_inference.ipynb
Browse files- RxRx3-core_inference.ipynb +195 -0
RxRx3-core_inference.ipynb
ADDED
@@ -0,0 +1,195 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "code",
|
5 |
+
"execution_count": null,
|
6 |
+
"id": "5edcb7d2-53dc-4170-9f2f-619c0da0ae4c",
|
7 |
+
"metadata": {},
|
8 |
+
"outputs": [],
|
9 |
+
"source": [
|
10 |
+
"import torch\n",
|
11 |
+
"import numpy as np\n",
|
12 |
+
"from torch.utils.data import DataLoader\n",
|
13 |
+
"import pandas as pd"
|
14 |
+
]
|
15 |
+
},
|
16 |
+
{
|
17 |
+
"cell_type": "markdown",
|
18 |
+
"id": "f839c8fb-b018-4ab6-86a9-7d5bf7883b45",
|
19 |
+
"metadata": {},
|
20 |
+
"source": [
|
21 |
+
"# Load OpenPhenom"
|
22 |
+
]
|
23 |
+
},
|
24 |
+
{
|
25 |
+
"cell_type": "code",
|
26 |
+
"execution_count": null,
|
27 |
+
"id": "84b9324d-fde9-4c43-bc5a-eb66cdb4f891",
|
28 |
+
"metadata": {},
|
29 |
+
"outputs": [],
|
30 |
+
"source": [
|
31 |
+
"# Load model directly\n",
|
32 |
+
"from huggingface_mae import MAEModel\n",
|
33 |
+
"open_phenom = MAEModel.from_pretrained(\".\")"
|
34 |
+
]
|
35 |
+
},
|
36 |
+
{
|
37 |
+
"cell_type": "code",
|
38 |
+
"execution_count": null,
|
39 |
+
"id": "57d918c5-78de-4b36-9f46-4652c5da93f2",
|
40 |
+
"metadata": {},
|
41 |
+
"outputs": [],
|
42 |
+
"source": [
|
43 |
+
"open_phenom.eval()\n",
|
44 |
+
"cuda_available = torch.cuda.is_available()\n",
|
45 |
+
"if cuda_available:\n",
|
46 |
+
" open_phenom.cuda()"
|
47 |
+
]
|
48 |
+
},
|
49 |
+
{
|
50 |
+
"cell_type": "markdown",
|
51 |
+
"id": "7c89d82d-5365-4492-b496-adb3bbd71b32",
|
52 |
+
"metadata": {},
|
53 |
+
"source": [
|
54 |
+
"# Load Rxrx3-core"
|
55 |
+
]
|
56 |
+
},
|
57 |
+
{
|
58 |
+
"cell_type": "code",
|
59 |
+
"execution_count": null,
|
60 |
+
"id": "deeff3a8-db67-4905-a7e9-c43aad614a84",
|
61 |
+
"metadata": {},
|
62 |
+
"outputs": [],
|
63 |
+
"source": [
|
64 |
+
"from datasets import load_dataset\n",
|
65 |
+
"rxrx3_core = load_dataset(\"recursionpharma/rxrx3-core\")['train']"
|
66 |
+
]
|
67 |
+
},
|
68 |
+
{
|
69 |
+
"cell_type": "markdown",
|
70 |
+
"id": "8f2226ce-9415-4dd8-932e-54e4e1bd8c1a",
|
71 |
+
"metadata": {},
|
72 |
+
"source": [
|
73 |
+
"# Infernce loop"
|
74 |
+
]
|
75 |
+
},
|
76 |
+
{
|
77 |
+
"cell_type": "code",
|
78 |
+
"execution_count": null,
|
79 |
+
"id": "aa1218ab-f9cd-413b-9228-c1146df978be",
|
80 |
+
"metadata": {},
|
81 |
+
"outputs": [],
|
82 |
+
"source": [
|
83 |
+
"def convert_path_to_well_id(path_str):\n",
|
84 |
+
" \n",
|
85 |
+
" return path_str.split('_')[0].replace('/','_').replace('Plate','')\n",
|
86 |
+
" \n",
|
87 |
+
"def collate_rxrx3_core(batch):\n",
|
88 |
+
" \n",
|
89 |
+
" images = np.stack([np.array(i['jp2']) for i in batch]).reshape(-1,6,512,512)\n",
|
90 |
+
" images = np.vstack([patch_image(i) for i in images]) # convert to 4 256x256 patches\n",
|
91 |
+
" images = torch.from_numpy(images)\n",
|
92 |
+
" well_ids = [convert_path_to_well_id(i['__key__']) for i in batch[::6]]\n",
|
93 |
+
" return images, well_ids\n",
|
94 |
+
"\n",
|
95 |
+
"def iter_border_patches(width, height, patch_size):\n",
|
96 |
+
" \n",
|
97 |
+
" x_start, x_end, y_start, y_end = (0, width, 0, height)\n",
|
98 |
+
"\n",
|
99 |
+
" for x in range(x_start, x_end - patch_size + 1, patch_size):\n",
|
100 |
+
" for y in range(y_start, y_end - patch_size + 1, patch_size):\n",
|
101 |
+
" yield x, y\n",
|
102 |
+
"\n",
|
103 |
+
"def patch_image(image_array, patch_size=256):\n",
|
104 |
+
" \n",
|
105 |
+
" _, width, height = image_array.shape\n",
|
106 |
+
" output_patches = []\n",
|
107 |
+
" patch_count = 0\n",
|
108 |
+
" for x, y in iter_border_patches(width, height, patch_size):\n",
|
109 |
+
" patch = image_array[:, y : y + patch_size, x : x + patch_size].copy()\n",
|
110 |
+
" output_patches.append(patch)\n",
|
111 |
+
" \n",
|
112 |
+
" output_patches = np.stack(output_patches)\n",
|
113 |
+
" \n",
|
114 |
+
" return output_patches"
|
115 |
+
]
|
116 |
+
},
|
117 |
+
{
|
118 |
+
"cell_type": "code",
|
119 |
+
"execution_count": null,
|
120 |
+
"id": "de308003-bcfc-4b59-9715-dd884b9b2536",
|
121 |
+
"metadata": {},
|
122 |
+
"outputs": [],
|
123 |
+
"source": [
|
124 |
+
"# Convert to PyTorch DataLoader\n",
|
125 |
+
"batch_size = 128\n",
|
126 |
+
"num_workers = 4\n",
|
127 |
+
"rxrx3_core_dataloader = DataLoader(rxrx3_core, batch_size=batch_size*6, shuffle=False, \n",
|
128 |
+
" collate_fn=collate_rxrx3_core, num_workers=num_workers)"
|
129 |
+
]
|
130 |
+
},
|
131 |
+
{
|
132 |
+
"cell_type": "code",
|
133 |
+
"execution_count": null,
|
134 |
+
"id": "9e3ea6c2-d1aa-4e20-a175-d72ea636153e",
|
135 |
+
"metadata": {},
|
136 |
+
"outputs": [],
|
137 |
+
"source": [
|
138 |
+
"# Inference loop\n",
|
139 |
+
"num_features = 384\n",
|
140 |
+
"n_crops = 4\n",
|
141 |
+
"well_ids = []\n",
|
142 |
+
"emb_ind = 0\n",
|
143 |
+
"embeddings = np.zeros(\n",
|
144 |
+
" ((len(rxrx3_core_dataloader.dataset)//6), num_features), dtype=np.float32\n",
|
145 |
+
")\n",
|
146 |
+
"forward_pass_counter = 0\n",
|
147 |
+
"\n",
|
148 |
+
"for imgs, batch_well_ids in rxrx3_core_dataloader:\n",
|
149 |
+
"\n",
|
150 |
+
" if cuda_available:\n",
|
151 |
+
" with torch.amp.autocast(\"cuda\"), torch.no_grad():\n",
|
152 |
+
" latent = open_phenom.predict(imgs.cuda())\n",
|
153 |
+
" else:\n",
|
154 |
+
" latent = open_phenom.predict(imgs)\n",
|
155 |
+
" \n",
|
156 |
+
" latent = latent.view(-1, n_crops, num_features).mean(dim=1) # average over 4 256x256 crops per image\n",
|
157 |
+
" embeddings[emb_ind : (emb_ind + len(latent))] = latent.detach().cpu().numpy()\n",
|
158 |
+
" well_ids.extend(batch_well_ids)\n",
|
159 |
+
"\n",
|
160 |
+
" emb_ind += len(latent)\n",
|
161 |
+
" forward_pass_counter += 1\n",
|
162 |
+
" if forward_pass_counter % 5 == 0:\n",
|
163 |
+
" print(f\"forward pass {forward_pass_counter} of {len(rxrx3_core_dataloader)} done, wells inferenced {emb_ind}\")\n",
|
164 |
+
"\n",
|
165 |
+
"embedding_df = embeddings[:emb_ind]\n",
|
166 |
+
"embedding_df = pd.DataFrame(embedding_df)\n",
|
167 |
+
"embedding_df.columns = [f\"feature_{i}\" for i in range(num_features)]\n",
|
168 |
+
"embedding_df['well_id'] = well_ids\n",
|
169 |
+
"embedding_df = embedding_df[['well_id']+[f\"feature_{i}\" for i in range(num_features)]]\n",
|
170 |
+
"embedding_df.to_parquet('OpenPhenom_rxrx3-core_embeddings.parquet')"
|
171 |
+
]
|
172 |
+
}
|
173 |
+
],
|
174 |
+
"metadata": {
|
175 |
+
"kernelspec": {
|
176 |
+
"display_name": "photo2",
|
177 |
+
"language": "python",
|
178 |
+
"name": "photo2"
|
179 |
+
},
|
180 |
+
"language_info": {
|
181 |
+
"codemirror_mode": {
|
182 |
+
"name": "ipython",
|
183 |
+
"version": 3
|
184 |
+
},
|
185 |
+
"file_extension": ".py",
|
186 |
+
"mimetype": "text/x-python",
|
187 |
+
"name": "python",
|
188 |
+
"nbconvert_exporter": "python",
|
189 |
+
"pygments_lexer": "ipython3",
|
190 |
+
"version": "3.10.14"
|
191 |
+
}
|
192 |
+
},
|
193 |
+
"nbformat": 4,
|
194 |
+
"nbformat_minor": 5
|
195 |
+
}
|