-
Notifications
You must be signed in to change notification settings - Fork 1
/
example_pytorch_feature_extraction.py
320 lines (260 loc) · 14.1 KB
/
example_pytorch_feature_extraction.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
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
# This code shows the process of extracting features from a PyTorch model
# developed by Meta researchers
# For details about the model or paper, see https://github.com/facebookresearch/ConvNeXt
# https://arxiv.org/abs/2201.03545
# Ref: Islam et al, Revealing Hidden Patterns in Deep Neural Network Feature Space Continuum
# via Manifold Learning, in press, Nature Communications, 2023
import argparse
import datetime
import numpy as np
import time
import torch
import torch.nn as nn
import torch.backends.cudnn as cudnn
import json
import torch.distributed as dist
import os
os.environ['MASTER_ADDR'] = 'localhost'
os.environ['MASTER_PORT'] = '12345'
from pathlib import Path
from timm.data.mixup import Mixup
from timm.models import create_model
from timm.loss import LabelSmoothingCrossEntropy, SoftTargetCrossEntropy
from timm.utils import ModelEma
from optim_factory import create_optimizer, LayerDecayValueAssigner
from datasets import build_dataset
from engine import train_one_epoch, evaluate
from utils import NativeScalerWithGradNormCount as NativeScaler
import utils
import models.convnext
import models.convnext_isotropic
import torch
from torchvision import datasets, transforms
from torch.utils.data import DataLoader, random_split
import numpy as np
def str2bool(v):
"""
Converts string to bool type; enables command line
arguments in the format of '--arg1 true --arg2 false'
"""
if isinstance(v, bool):
return v
if v.lower() in ('yes', 'true', 't', 'y', '1'):
return True
elif v.lower() in ('no', 'false', 'f', 'n', '0'):
return False
else:
raise argparse.ArgumentTypeError('Boolean value expected.')
def get_args_parser():
parser = argparse.ArgumentParser('ConvNeXt training and evaluation script for image classification', add_help=False)
parser.add_argument('--batch_size', default=64, type=int,
help='Per GPU batch size')
parser.add_argument('--epochs', default=300, type=int)
parser.add_argument('--update_freq', default=1, type=int,
help='gradient accumulation steps')
# Model parameters
parser.add_argument('--model', default='convnext_base', type=str, metavar='MODEL',
help='Name of model to train')
parser.add_argument('--drop_path', type=float, default=0, metavar='PCT',
help='Drop path rate (default: 0.0)')
parser.add_argument('--input_size', default=224, type=int,
help='image input size')
parser.add_argument('--layer_scale_init_value', default=1e-6, type=float,
help="Layer scale initial values")
# EMA related parameters
parser.add_argument('--model_ema', type=str2bool, default=False)
parser.add_argument('--model_ema_decay', type=float, default=0.9999, help='')
parser.add_argument('--model_ema_force_cpu', type=str2bool, default=False, help='')
parser.add_argument('--model_ema_eval', type=str2bool, default=False, help='Using ema to eval during training.')
# Optimization parameters
parser.add_argument('--opt', default='adamw', type=str, metavar='OPTIMIZER',
help='Optimizer (default: "adamw"')
parser.add_argument('--opt_eps', default=1e-8, type=float, metavar='EPSILON',
help='Optimizer Epsilon (default: 1e-8)')
parser.add_argument('--opt_betas', default=None, type=float, nargs='+', metavar='BETA',
help='Optimizer Betas (default: None, use opt default)')
parser.add_argument('--clip_grad', type=float, default=None, metavar='NORM',
help='Clip gradient norm (default: None, no clipping)')
parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
help='SGD momentum (default: 0.9)')
parser.add_argument('--weight_decay', type=float, default=0.05,
help='weight decay (default: 0.05)')
parser.add_argument('--weight_decay_end', type=float, default=None, help="""Final value of the
weight decay. We use a cosine schedule for WD and using a larger decay by
the end of training improves performance for ViTs.""")
parser.add_argument('--lr', type=float, default=4e-3, metavar='LR',
help='learning rate (default: 4e-3), with total batch size 4096')
parser.add_argument('--layer_decay', type=float, default=1.0)
parser.add_argument('--min_lr', type=float, default=1e-6, metavar='LR',
help='lower lr bound for cyclic schedulers that hit 0 (1e-6)')
parser.add_argument('--warmup_epochs', type=int, default=20, metavar='N',
help='epochs to warmup LR, if scheduler supports')
parser.add_argument('--warmup_steps', type=int, default=-1, metavar='N',
help='num of steps to warmup LR, will overload warmup_epochs if set > 0')
# Augmentation parameters
parser.add_argument('--color_jitter', type=float, default=0.4, metavar='PCT',
help='Color jitter factor (default: 0.4)')
parser.add_argument('--aa', type=str, default='rand-m9-mstd0.5-inc1', metavar='NAME',
help='Use AutoAugment policy. "v0" or "original". " + "(default: rand-m9-mstd0.5-inc1)'),
parser.add_argument('--smoothing', type=float, default=0.1,
help='Label smoothing (default: 0.1)')
parser.add_argument('--train_interpolation', type=str, default='bicubic',
help='Training interpolation (random, bilinear, bicubic default: "bicubic")')
# Evaluation parameters
parser.add_argument('--crop_pct', type=float, default=None)
# * Random Erase params
parser.add_argument('--reprob', type=float, default=0.25, metavar='PCT',
help='Random erase prob (default: 0.25)')
parser.add_argument('--remode', type=str, default='pixel',
help='Random erase mode (default: "pixel")')
parser.add_argument('--recount', type=int, default=1,
help='Random erase count (default: 1)')
parser.add_argument('--resplit', type=str2bool, default=False,
help='Do not random erase first (clean) augmentation split')
# * Mixup params
parser.add_argument('--mixup', type=float, default=0.8,
help='mixup alpha, mixup enabled if > 0.')
parser.add_argument('--cutmix', type=float, default=1.0,
help='cutmix alpha, cutmix enabled if > 0.')
parser.add_argument('--cutmix_minmax', type=float, nargs='+', default=None,
help='cutmix min/max ratio, overrides alpha and enables cutmix if set (default: None)')
parser.add_argument('--mixup_prob', type=float, default=1.0,
help='Probability of performing mixup or cutmix when either/both is enabled')
parser.add_argument('--mixup_switch_prob', type=float, default=0.5,
help='Probability of switching to cutmix when both mixup and cutmix enabled')
parser.add_argument('--mixup_mode', type=str, default='batch',
help='How to apply mixup/cutmix params. Per "batch", "pair", or "elem"')
# * Finetuning params
parser.add_argument('--finetune', default='',
help='finetune from checkpoint')
parser.add_argument('--head_init_scale', default=1.0, type=float,
help='classifier head initial scale, typically adjusted in fine-tuning')
parser.add_argument('--model_key', default='model|module', type=str,
help='which key to load from saved state dict, usually model or model_ema')
parser.add_argument('--model_prefix', default='', type=str)
# Dataset parameters
parser.add_argument('--data_path', default='E:\imagenet_structured', type=str,
help='dataset path')
parser.add_argument('--eval_data_path', default=None, type=str,
help='dataset path for evaluation')
parser.add_argument('--nb_classes', default=1000, type=int,
help='number of the classification types')
parser.add_argument('--imagenet_default_mean_and_std', type=str2bool, default=True)
parser.add_argument('--data_set', default='IMNET', choices=['CIFAR', 'IMNET', 'image_folder'],
type=str, help='ImageNet dataset path')
parser.add_argument('--output_dir', default='',
help='path where to save, empty for no saving')
parser.add_argument('--log_dir', default=None,
help='path where to tensorboard log')
parser.add_argument('--device', default='cuda',
help='device to use for training / testing')
parser.add_argument('--seed', default=0, type=int)
parser.add_argument('--resume', default='',
help='resume from checkpoint')
parser.add_argument('--auto_resume', type=str2bool, default=True)
parser.add_argument('--save_ckpt', type=str2bool, default=True)
parser.add_argument('--save_ckpt_freq', default=1, type=int)
parser.add_argument('--save_ckpt_num', default=3, type=int)
parser.add_argument('--start_epoch', default=0, type=int, metavar='N',
help='start epoch')
parser.add_argument('--eval', type=str2bool, default=False,
help='Perform evaluation only')
parser.add_argument('--dist_eval', type=str2bool, default=True,
help='Enabling distributed evaluation')
parser.add_argument('--disable_eval', type=str2bool, default=False,
help='Disabling evaluation during training')
parser.add_argument('--num_workers', default=10, type=int)
parser.add_argument('--pin_mem', type=str2bool, default=True,
help='Pin CPU memory in DataLoader for more efficient (sometimes) transfer to GPU.')
# distributed training parameters
parser.add_argument('--world_size', default=1, type=int,
help='number of distributed processes')
parser.add_argument('--local_rank', default=-1, type=int)
parser.add_argument('--dist_on_itp', type=str2bool, default=False)
parser.add_argument('--dist_url', default='env://',
help='url used to set up distributed training')
parser.add_argument('--backend', default='gloo',
help='backend added by tauhid')
parser.add_argument('--use_amp', type=str2bool, default=False,
help="Use PyTorch's AMP (Automatic Mixed Precision) or not")
# Weights and Biases arguments
parser.add_argument('--enable_wandb', type=str2bool, default=False,
help="enable logging to Weights and Biases")
parser.add_argument('--project', default='convnext', type=str,
help="The name of the W&B project where you're sending the new run.")
parser.add_argument('--wandb_ckpt', type=str2bool, default=False,
help="Save model checkpoints as W&B Artifacts.")
return parser
parser = argparse.ArgumentParser('ConvNeXt training and evaluation script', parents=[get_args_parser()])
args = parser.parse_args()
# Load the trained model
state_dict = torch.load('save/convnext_base_22k_1k_224.pth', map_location='cuda')
# Adjust the keys if the model was trained using nn.DataParallel or nn.DistributedDataParallel
new_state_dict = {}
for k, v in state_dict.items():
name = k[7:] if k.startswith('module.') else k # remove `module.` if present
new_state_dict[name] = v
model = create_model(
args.model,
pretrained=False,
num_classes=args.nb_classes,
drop_path_rate=args.drop_path,
layer_scale_init_value=args.layer_scale_init_value,
head_init_scale=args.head_init_scale,
)
model.load_state_dict(new_state_dict['model'])
device=torch.device('cuda')
model.to(device)
# Define your transformations
transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
# Add other necessary transformations
])
# Load the dataset
full_dataset = datasets.ImageFolder('E:/imagenet_structured/val_sub', transform=transform)
# Create a DataLoader for the subset
subset_loader = DataLoader(full_dataset, batch_size=32, shuffle=False)
# Now you can use `subset_loader` to process the images
# Assuming `model` is your pre-loaded model
model.eval() # Set the model to evaluation mode
def get_activation(name):
def hook(model, input, output):
print(f"Hook activated for {name}")
activation[name] = output.detach()
print(f"Output shape: {output.detach().shape}")
return hook
activation = {}
layer = model.stages[2][0].norm
layer_string='stages.2.0.norm'
hook = layer.register_forward_hook(get_activation(layer_string))
features = []
labelsV=[]
with torch.no_grad():
for inputs, labels in subset_loader:
inputs = inputs.cuda()
outputs = model(inputs) # Extract features
featuresX = activation[layer_string]
features.extend(featuresX.cpu().numpy())
labelsV.extend(labels)
feat_array=np.array(features);
mdic = {"feat_array": feat_array, 'labels':labelsV}
np.save('feature_matrix_s1norm1_val.npy', feat_array)
hook.remove()
activation = {}
layer = model.stages[2][2].norm
layer_string='stages.2.2.norm'
hook = layer.register_forward_hook(get_activation(layer_string))
features = []
labelsV=[]
with torch.no_grad():
for inputs, labels in subset_loader:
inputs = inputs.cuda()
outputs = model(inputs) # Extract features
featuresX = activation[layer_string]
features.extend(featuresX.cpu().numpy())
labelsV.extend(labels)
feat_array=np.array(features);
mdic = {"feat_array": feat_array, 'labels':labelsV}
np.save('feature_matrix_s1norm2_val.npy', feat_array)
hook.remove()