Skip to content

GitLab

Commit 1d5a34cf authored by wanglch's avatar wanglch
Browse files

Initial commit

parents
Pipeline #1446 canceled with stages
Hide whitespace changes
Inline Side-by-side
Showing with 723 additions and 0 deletions
clip_benchmark/data/mscoco_captions/coco-cn_test.json 0 → 100644
View file @ 1d5a34cf
This source diff could not be displayed because it is too large. You can view the blob instead.
clip_benchmark/probe_benchmark/PROBES.md 0 → 100644
View file @ 1d5a34cf
Steps to run.
1. Navigate to `CLIP_benchmark`.
2. Run `export PYTHONPATH=$PWD`.
3. (Optional) To re-run the experiments, run `python probe_benchmark/scaling_experiments.py`. You'll have to change line
51 to point to your data.
4. (Optional) To generate the results, run `python probe_benchmark/build_df_scaling_experiments.py`.
5. (Optional) VTAB requires post-processing to average. Run `python probe_benchmark/process_vtab.py`.
6. Generate plots with `python probe_benchmark/scaling_plot.py`.
7. Generate table with `python probe_benchark/generate_table.py`.
clip_benchmark/probe_benchmark/build_df_scaling_experiments.py 0 → 100644
View file @ 1d5a34cf
import json
import os
import pandas as pd
if __name__ == '__main__':
compute_df = pd.read_csv('probe_benchmark/clip_table_2.csv')
# mdf = pd.read_csv("https://gist.githubusercontent.com/mehdidc/58dee67cecd5431a80ee3a2346c9c165/raw/45288ebccaacc34a97f580f8bf16fb3274927f2c/gistfile1.txt")
mdf = pd.read_csv('probe_benchmark/openclip_results.csv')
info = []
# import pdb; pdb.set_trace()
models = ['ViT-B-32-quickgelu,laion400m_e32',
'ViT-B-32,openai',
'ViT-B-32,laion2b_s34b_b79k',
'ViT-B-16,laion400m_e32',
'ViT-B-16-plus-240,laion400m_e32',
'ViT-B-16,openai',
# 'ViT-L-14-336,openai',
'ViT-L-14,openai',
'ViT-B-32,laion2b_e16',
'ViT-L-14,laion400m_e32',
'ViT-L-14,laion2b_s32b_b82k',
'ViT-H-14,laion2b_s32b_b79k',
'ViT-g-14,laion2b_s12b_b42k',
]
alt_models = ['B/32 400M',
'B/32 CLIP WIT',
'B/32 2B',
'B/16 400M',
'B/16+ 400M',
'B/16 CLIP WIT',
# 'ViT-L-14-336,openai',
'L/14 CLIP WIT',
'B/32 2B',
'L/14 400M',
'L/14 2B',
'H/14 2B',
'g/14 2B',
]
datasets = ['imagenet1k-unverified', 'cifar100']
datasets = datasets + [
'vtab/caltech101',
'vtab/cifar10',
'vtab/cifar100',
'vtab/clevr_count_all',
'vtab/clevr_closest_object_distance',
'vtab/diabetic_retinopathy',
'vtab/dmlab',
'vtab/dsprites_label_orientation',
'vtab/dsprites_label_x_position',
'vtab/dtd',
'vtab/eurosat',
'vtab/kitti_closest_vehicle_distance',
'vtab/flowers',
'vtab/pets',
'vtab/pcam',
'vtab/resisc45',
'vtab/smallnorb_label_azimuth',
'vtab/smallnorb_label_elevation',
'vtab/svhn',
]
ks = [10, 25, -1]
lrs = [0.1, 0.01, 0.001]
epoch_vals = [10, 20, 40]
batch_sizes = [32 * 8]
def get_us_dataset(pretrained):
if '2b' in pretrained:
return 'LAION-2B'
elif 'laion' in pretrained:
return 'LAION-400M'
else:
return 'CLIP-WIT'
for dataset in datasets:
dataset_root = '/datasets01/imagenet_full_size/061417' if dataset.startswith(
'imagenet') else '/private/home/mitchellw/git/forks/CLIP_benchmark'
for ii, model_info in enumerate(models):
model_info_split = model_info.split(',')
model, pretrained = model_info_split[0], model_info_split[1]
for epochs in epoch_vals:
for k in ks:
if k == 25 and 'vtab' in dataset:
continue
for lr in lrs:
for bs in batch_sizes:
pth = '/private/home/mitchellw/git/forks/CLIP_benchmark/probe_benchmark/data/' + f'{model}-{pretrained}-{dataset}-{epochs}-{k}-{lr}-{bs}.json'.replace(
'/', '_')
print(pth)
assert os.path.exists(pth)
row = {
'k': k,
'lr': lr,
'bs': bs,
'epochs': epochs,
'model': model.replace('-quickgelu', ''),
'pretrained': pretrained,
'pretrained_short': 'laion2b' if 'laion2b' in pretrained else pretrained,
'pretrained_clean': 'LAION' if 'laion' in pretrained else 'CLIP-WiT',
'dataset': dataset,
'macts': compute_df[compute_df.model == model.replace('-quickgelu', '')][
'image_macts'].values[0],
# 'gmacs_total': mdf[mdf.model_fullname_pretty == alt_models[ii]]['gmacs_total'].values[0],
# 'samples_seen': mdf[mdf.model_fullname_pretty == alt_models[ii]]['samples_seen'].values[0],
'gmacs_total':
mdf[mdf.model_fullname == models[ii].replace(',', ' ')]['gmacs_total'].values[0],
'samples_seen':
mdf[mdf.model_fullname == models[ii].replace(',', ' ')]['samples_seen'].values[0],
'samples_seen_pretty': mdf[mdf.model_fullname == models[ii].replace(',', ' ')][
'samples_seen_pretty'].values[0],
'model_short': models[ii].replace(',', ' '),
'upstream_dataset': get_us_dataset(pretrained)
}
with open(pth, 'r') as f:
row.update(json.load(f)['metrics'])
info.append(row)
with open('probe_benchmark/scaling_experiment_data2.json', 'w') as f:
json.dump(info, f)
clip_benchmark/probe_benchmark/clip_table_2.csv 0 → 100644
View file @ 1d5a34cf
model,image_size,image_width,text_width,embed_dim,gmacs,macts,mparams,image_gmacs,image_macts,image_mparams,text_gmacs,text_macts,text_mparams
ViT-B-32,224,768,512,512,7.4,10.31,151.28,4.41,5.01,87.85,2.98,5.3,63.43
ViT-B-32-plus-256,256,896,640,640,12.43,14.38,210.3,7.79,7.76,119.13,4.64,6.63,91.16
RN50,224,2048,512,1024,9.16,18.29,102.01,6.17,12.98,38.32,2.98,5.3,63.69
ViT-M-16,224,512,512,512,10.99,21.23,102.02,8.0,15.93,38.59,2.98,5.3,63.43
RN101,224,2048,512,512,12.84,23.38,119.69,9.86,18.08,56.26,2.98,5.3,63.43
ViT-M-16-256,256,512,512,512,13.62,27.56,102.05,10.63,22.26,38.62,2.98,5.3,63.43
ViT-B-16,224,768,512,512,20.57,29.2,149.62,17.58,23.9,86.19,2.98,5.3,63.43
ViT-B-16-plus,224,896,640,640,28.41,34.5,208.35,23.77,27.88,117.19,4.64,6.63,91.16
ViT-B-16-plus-240,240,896,640,640,32.05,39.71,208.38,27.41,33.08,117.21,4.64,6.63,91.16
RN50x4,288,2560,640,640,26.09,41.9,178.3,21.45,35.27,87.14,4.64,6.63,91.16
ViT-L-16,224,1024,768,768,68.26,71.47,427.74,61.6,63.52,304.09,6.66,7.95,123.65
ViT-L-14,224,1024,768,768,87.73,96.74,427.62,81.08,88.79,303.97,6.66,7.95,123.65
RN50x16,384,3072,768,768,81.86,111.49,290.98,75.2,103.54,167.33,6.66,7.95,123.65
ViT-H-16,224,1280,1024,1024,150.96,122.01,986.26,127.4,100.81,632.23,23.57,21.2,354.03
ViT-H-14,224,1280,1024,1024,190.97,160.61,986.11,167.4,139.41,632.08,23.57,21.2,354.03
ViT-L-14-280,280,1024,768,768,136.0,168.66,427.76,129.34,160.71,304.11,6.66,7.95,123.65
RN50x64,448,4096,1024,1024,193.4,199.15,500.28,181.61,188.55,297.4,11.78,10.6,202.88
ViT-g-14,224,1408,1024,1024,290.74,213.84,1366.68,267.18,192.64,1012.65,23.57,21.2,354.03
ViT-H-14-280,280,1280,1024,1024,289.49,268.29,986.29,265.93,247.09,632.26,23.57,21.2,354.03
ViT-L-14-336,336,1024,768,768,197.76,278.19,427.94,191.1,270.24,304.29,6.66,7.95,123.65
ViT-g-14-280,280,1408,1024,1024,446.95,358.73,1366.88,423.38,337.53,1012.85,23.57,21.2,354.03
ViT-H-14-336,336,1280,1024,1024,414.53,428.74,986.52,390.97,407.54,632.49,23.57,21.2,354.03
ViT-g-14-336,336,1408,1024,1024,644.21,571.87,1367.13,620.65,550.67,1013.1,23.57,21.2,354.03
clip_benchmark/probe_benchmark/generate_table.py 0 → 100644
View file @ 1d5a34cf
import pandas as pd
# make a new version of vtab
if __name__ == '__main__':
df_full = pd.read_json('probe_benchmark/scaling_experiment_data2.json')
df = df_full[df_full.fewshot_k == -1]
df25 = df_full[df_full.fewshot_k == 25]
df10 = df_full[df_full.fewshot_k == 10]
datasets = [
'vtab/caltech101',
'vtab/cifar10',
'vtab/cifar100',
'vtab/clevr_count_all',
'vtab/clevr_closest_object_distance',
'vtab/diabetic_retinopathy',
'vtab/dmlab',
'vtab/dsprites_label_orientation',
'vtab/dsprites_label_x_position',
'vtab/dtd',
'vtab/eurosat',
'vtab/kitti_closest_vehicle_distance',
'vtab/flowers',
'vtab/pets',
'vtab/pcam',
'vtab/resisc45',
'vtab/smallnorb_label_azimuth',
'vtab/smallnorb_label_elevation',
'vtab_svhn',
]
datasets2 = [
'imagenet1k-unverified', 'cifar100'
]
all_info = []
cols = []
first = True
for n, g in df_full.groupby(['model', 'pretrained', 'samples_seen_pretty']):
count = 0
total = 0.
for d in datasets:
g_filter = g[(g.dataset == d) & (g.fewshot_k == -1)]
count += 1
total += g_filter.lp_acc1.max()
avg = total / count
info = {'VTAB acc': avg}
if first:
cols.append('VTAB acc')
for d in datasets2:
for k in [10, 25, -1]:
g_filter = g[(g.dataset == d) & (g.fewshot_k == k)]
info[f'{d}: {k} shot'] = g_filter.lp_acc1.max()
if first:
cols.append(f'{d}: {k} shot')
for k in ['model', 'pretrained', 'upstream_dataset', 'gmacs_total', 'samples_seen_pretty']:
info[k] = g[k].values[0]
all_info.append(info)
first = False
df = pd.DataFrame(all_info)
formatters = {}
print(df.keys())
columns = ['model', 'samples_seen_pretty', 'upstream_dataset']
df = df.sort_values(by=['model', 'samples_seen_pretty', 'upstream_dataset'])
for ds in cols:
columns.append(ds)
formatters[ds] = lambda x: f'{100 * x:.2f}'
latex = df.to_latex(columns=columns, formatters=formatters)
print(latex)
# with open('probe_benchmark/scaling_experiment_data_combined.json', 'w') as f:
# json.dump(all_info, f)
clip_benchmark/probe_benchmark/laion5b_fewshot_experiments.py 0 → 100644
View file @ 1d5a34cf
import os
from clip_benchmark.cli import get_parser_args, run
# /private/home/mitchellw/miniconda3/envs/cb/bin/python probe_benchmark/laion5b_fewshot_experiments.py
if __name__ == '__main__':
models = ['ViT-B-32-quickgelu,laion400m_e32',
'ViT-B-32,openai',
'ViT-B-32,laion2b_s34b_b79k',
'ViT-B-16,laion400m_e32',
# 'ViT-B-16-plus-240,laion400m_e32',
'ViT-B-16,openai',
# 'ViT-L-14-336,openai',
'ViT-L-14,openai',
# 'ViT-B-32,laion2b_e16',
'ViT-L-14,laion400m_e32',
'ViT-L-14,laion2b_s32b_b82k',
'ViT-H-14,laion2b_s32b_b79k',
]
datasets = ['imagenet1k-unverified']
ks = [1, 2, 4, 8, 16, 32, 64, 128]
lrs = [0.1, 0.01, 0.001, 0.0001]
epoch_vals = [10, 20, 40, 80]
batch_sizes = [32 * 8]
for epochs in epoch_vals:
for dataset in datasets:
dataset_root = '/datasets01/imagenet_full_size/061417' if dataset.startswith(
'imagenet') else '/private/home/mitchellw/git/forks/CLIP_benchmark'
for model_info in models:
model_info_split = model_info.split(',')
model, pretrained = model_info_split[0], model_info_split[1]
for k in ks:
for lr in lrs:
for bs in batch_sizes:
args = get_parser_args()
args.dataset_root = dataset_root
args.dataset = dataset
args.task = 'linear_probe'
args.pretrained = pretrained
args.model = model
args.output = '/private/home/mitchellw/git/forks/CLIP_benchmark/probe_benchmark/data/' + f'{model}-{pretrained}-{dataset}-{epochs}-{k}-{lr}-{bs}.json'.replace(
'/', '_')
if os.path.exists(args.output):
print('skipping - exists.')
args.fewshot_k = k
args.fewshot_epochs = epochs
args.fewshot_lr = lr
args.batch_size = bs
args.skip_load = True # NOTE
run(args)
clip_benchmark/probe_benchmark/openclip_results.csv 0 → 100644
View file @ 1d5a34cf
This source diff could not be displayed because it is too large. You can view the blob instead.
clip_benchmark/probe_benchmark/process_vtab.py 0 → 100644
View file @ 1d5a34cf
import json
import pandas as pd
# make a new version of vtab
if __name__ == '__main__':
df = pd.read_json('probe_benchmark/scaling_experiment_data2.json')
df = df[df.fewshot_k == -1]
datasets = [
'vtab/caltech101',
'vtab/cifar10',
'vtab/cifar100',
'vtab/clevr_count_all',
'vtab/clevr_closest_object_distance',
'vtab/diabetic_retinopathy',
'vtab/dmlab',
'vtab/dsprites_label_orientation',
'vtab/dsprites_label_x_position',
'vtab/dtd',
'vtab/eurosat',
'vtab/kitti_closest_vehicle_distance',
'vtab/flowers',
'vtab/pets',
'vtab/pcam',
'vtab/resisc45',
'vtab/smallnorb_label_azimuth',
'vtab/smallnorb_label_elevation',
'vtab/svhn',
]
all_info = []
for n, g in df.groupby(['model', 'pretrained', 'samples_seen_pretty']):
count = 0
total = 0.
for d in datasets:
g_filter = g[g.dataset == d]
count += 1
total += g_filter.lp_acc1.max()
avg = total / count
info = {'dataset': 'vtab', 'lp_acc1': avg, 'fewshot_k': -1}
for k in ['model', 'pretrained', 'upstream_dataset', 'gmacs_total', 'samples_seen_pretty']:
info[k] = g[k].values[0]
all_info.append(info)
with open('probe_benchmark/scaling_experiment_data_vtab.json', 'w') as f:
json.dump(all_info, f)
clip_benchmark/probe_benchmark/scaling_experiment_data2.json 0 → 100644
View file @ 1d5a34cf
This source diff could not be displayed because it is too large. You can view the blob instead.
clip_benchmark/probe_benchmark/scaling_experiment_data_vtab.json 0 → 100644
View file @ 1d5a34cf
[
{
"dataset": "vtab",
"lp_acc1": 0.7272385796110142,
"fewshot_k": -1,
"model": "ViT-B-16",
"pretrained": "laion400m_e32",
"upstream_dataset": "LAION-400M",
"gmacs_total": 268122270972.16,
"samples_seen_pretty": "13B"
},
{
"dataset": "vtab",
"lp_acc1": 0.7125347395825511,
"fewshot_k": -1,
"model": "ViT-B-16",
"pretrained": "openai",
"upstream_dataset": "CLIP-WIT",
"gmacs_total": 263296000000.0,
"samples_seen_pretty": "13B"
},
{
"dataset": "vtab",
"lp_acc1": 0.7332202011443508,
"fewshot_k": -1,
"model": "ViT-B-16-plus-240",
"pretrained": "laion400m_e32",
"upstream_dataset": "LAION-400M",
"gmacs_total": 370313744206.08,
"samples_seen_pretty": "13B"
},
{
"dataset": "vtab",
"lp_acc1": 0.7143166719197058,
"fewshot_k": -1,
"model": "ViT-B-32",
"pretrained": "laion2b_e16",
"upstream_dataset": "LAION-2B",
"gmacs_total": 256967931347.2,
"samples_seen_pretty": "34B"
},
{
"dataset": "vtab",
"lp_acc1": 0.7152995214130362,
"fewshot_k": -1,
"model": "ViT-B-32",
"pretrained": "laion2b_s34b_b79k",
"upstream_dataset": "LAION-2B",
"gmacs_total": 291096483388.0,
"samples_seen_pretty": "34B"
},
{
"dataset": "vtab",
"lp_acc1": 0.7183753019516755,
"fewshot_k": -1,
"model": "ViT-B-32",
"pretrained": "laion400m_e32",
"upstream_dataset": "LAION-400M",
"gmacs_total": 96456237491.2,
"samples_seen_pretty": "13B"
},
{
"dataset": "vtab",
"lp_acc1": 0.6971394911855741,
"fewshot_k": -1,
"model": "ViT-B-32",
"pretrained": "openai",
"upstream_dataset": "CLIP-WIT",
"gmacs_total": 94720000000.0,
"samples_seen_pretty": "13B"
},
{
"dataset": "vtab",
"lp_acc1": 0.7596462313700938,
"fewshot_k": -1,
"model": "ViT-H-14",
"pretrained": "laion2b_s32b_b79k",
"upstream_dataset": "LAION-2B",
"gmacs_total": 6631508868008.96,
"samples_seen_pretty": "34B"
},
{
"dataset": "vtab",
"lp_acc1": 0.744758325311516,
"fewshot_k": -1,
"model": "ViT-L-14",
"pretrained": "laion2b_s32b_b82k",
"upstream_dataset": "LAION-2B",
"gmacs_total": 2807360000000.0,
"samples_seen_pretty": "34B"
},
{
"dataset": "vtab",
"lp_acc1": 0.7397637678783028,
"fewshot_k": -1,
"model": "ViT-L-14",
"pretrained": "laion400m_e32",
"upstream_dataset": "LAION-400M",
"gmacs_total": 1143527799338.24,
"samples_seen_pretty": "13B"
},
{
"dataset": "vtab",
"lp_acc1": 0.7376775015037333,
"fewshot_k": -1,
"model": "ViT-L-14",
"pretrained": "openai",
"upstream_dataset": "CLIP-WIT",
"gmacs_total": 1122944000000.0,
"samples_seen_pretty": "13B"
},
{
"dataset": "vtab",
"lp_acc1": 0.7517780869059744,
"fewshot_k": -1,
"model": "ViT-g-14",
"pretrained": "laion2b_s12b_b42k",
"upstream_dataset": "LAION-2B",
"gmacs_total": 3549396664594.8,
"samples_seen_pretty": "13B"
}
]
clip_benchmark/probe_benchmark/scaling_experiments.py 0 → 100644
View file @ 1d5a34cf
import os
from clip_benchmark.cli import get_parser_args, run
if __name__ == '__main__':
models = ['ViT-B-32-quickgelu,laion400m_e32',
'ViT-B-32,openai',
'ViT-B-32,laion2b_s34b_b79k',
'ViT-B-16,laion400m_e32',
'ViT-B-16-plus-240,laion400m_e32',
'ViT-B-16,openai',
'ViT-L-14-336,openai',
'ViT-L-14,openai',
'ViT-B-32,laion2b_e16',
'ViT-L-14,laion400m_e32',
'ViT-L-14,laion2b_s32b_b82k',
'ViT-H-14,laion2b_s32b_b79k',
'ViT-g-14,laion2b_s12b_b42k',
]
datasets = ['imagenet1k-unverified', 'cifar100']
datasets = datasets + [
'vtab/caltech101',
'vtab/cifar10',
'vtab/cifar100',
'vtab/clevr_count_all',
'vtab/clevr_closest_object_distance',
'vtab/diabetic_retinopathy',
'vtab/dmlab',
'vtab/dsprites_label_orientation',
'vtab/dsprites_label_x_position',
'vtab/dtd',
'vtab/eurosat',
'vtab/kitti_closest_vehicle_distance',
'vtab/flowers',
'vtab/pets',
'vtab/pcam',
'vtab/resisc45',
'vtab/smallnorb_label_azimuth',
'vtab/smallnorb_label_elevation',
'vtab/svhn',
]
ks = [10, 25, -1]
lrs = [0.1, 0.01, 0.001]
epoch_vals = [10, 20, 40]
batch_sizes = [32 * 8]
if not os.path.exists('probe_benchmark/data'):
os.mkdir('probe_benchmark/data')
for dataset in datasets:
dataset_root = 'datasets/' + dataset.split('/')[-1] # TODO: change!
print(dataset_root)
for model_info in models:
model_info_split = model_info.split(',')
model, pretrained = model_info_split[0], model_info_split[1]
for epochs in epoch_vals:
# For VTAB, do not run >= 25 shot.
for k in ks:
if k >= 25 and dataset.startswith('vtab'):
continue
for lr in lrs:
for bs in batch_sizes:
args = get_parser_args()
args.dataset_root = dataset_root
args.dataset = dataset
args.task = 'linear_probe'
args.pretrained = pretrained
args.model = model
args.output = f'probe_benchmark/data/' + f'{model}-{pretrained}-{dataset}-{epochs}-{k}-{lr}-{bs}.json'.replace(
'/', '_')
if os.path.exists(args.output):
print('skipping - exists.')
continue
args.fewshot_k = k
args.fewshot_epochs = epochs
args.fewshot_lr = lr
args.batch_size = bs
run(args)
print(dataset, model, pretrained, epochs, k, lr, bs)
clip_benchmark/probe_benchmark/scaling_plot.ipynb 0 → 100644
View file @ 1d5a34cf
This source diff could not be displayed because it is too large. You can view the blob instead.
clip_benchmark/requirements.txt 0 → 100644
View file @ 1d5a34cf
open_clip_torch>=0.2.1
opencv-python
peft>=0.6.2
protobuf==3.20.3
pycocoevalcap
pyyaml
scikit-learn>=1.0,<2
scikit-learn
scipy
task_adaptation
tensorflow==2.11.0
termcolor
tqdm>=2
transformers>=4.32.0
webdataset>=0.2.31
yacs
clip_benchmark/setup.cfg 0 → 100644
View file @ 1d5a34cf
[bumpversion]
current_version = 0.1.0
commit = True
tag = True
[bumpversion:file:setup.py]
search = version='{current_version}'
replace = version='{new_version}'
[bumpversion:file:clip_benchmark/__init__.py]
search = __version__ = '{current_version}'
replace = __version__ = '{new_version}'
[bdist_wheel]
universal = 1
[flake8]
exclude = docs
clip_benchmark/setup.py 0 → 100644
View file @ 1d5a34cf
#!/usr/bin/env python
"""The setup script."""
from setuptools import find_packages, setup
with open('README.md') as readme_file:
readme = readme_file.read()
with open('HISTORY.rst') as history_file:
history = history_file.read()
def load_requirements(f):
return [l.strip() for l in open(f).readlines()]
requirements = load_requirements('requirements.txt')
test_requirements = requirements + ['pytest', 'pytest-runner']
setup(
author='Mehdi Cherti',
author_email='mehdicherti@gmail.com',
python_requires='>=3.6',
classifiers=[
'Development Status :: 2 - Pre-Alpha',
'Intended Audience :: Developers',
'License :: OSI Approved :: MIT License',
'Natural Language :: English',
'Programming Language :: Python :: 3',
'Programming Language :: Python :: 3.6',
'Programming Language :: Python :: 3.7',
'Programming Language :: Python :: 3.8',
],
description='CLIP-like models benchmarks on various datasets',
entry_points={
'console_scripts': [
'clip_benchmark=clip_benchmark.cli:main',
'clip_benchmark_export_wds=clip_benchmark.webdataset_builder:main',
],
},
install_requires=requirements,
license='MIT license',
long_description=readme + '\n\n' + history,
long_description_content_type='text/markdown',
include_package_data=True,
keywords='clip_benchmark',
name='clip_benchmark',
packages=find_packages(include=['clip_benchmark', 'clip_benchmark.*']),
test_suite='tests',
tests_require=test_requirements,
url='https://github.com/mehdidc/clip_benchmark',
version='1.4.0',
zip_safe=False,
extra_require={
'vtab': ['task_adaptation==0.1', 'timm>=0.5.4'],
'tfds': ['tfds-nightly', 'timm>=0.5.4'],
'coco': ['pycocotools>=2.0.4'],
}
)
clip_benchmark/test_internvl_c_classification.sh 0 → 100644
View file @ 1d5a34cf
set -x
PARTITION=${PARTITION:-'INTERN4'}
alias s1a="srun -p ${PARTITION} -N 1 --gres=gpu:1 --cpus-per-task 10 --quotatype=auto"
export PYTHONPATH="${PYTHONPATH}:$(pwd)"
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "birdsnap" --dataset_root ./data/birdsnap/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "cifar10" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "cifar100" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "food101" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "sun397" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "cars" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "fgvc_aircraft" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "dtd" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "pets" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "caltech101" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "mnist" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "stl10" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "eurosat" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "gtsrb" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "country211" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "pcam" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "renderedsst2" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "fer2013" --dataset_root ./data/fer2013 --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "voc2007" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "vtab/flowers" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
s1a --async python3 clip_benchmark/cli.py eval --model_type internvl --language "en" --task "zeroshot_classification" \
--dataset "vtab/resisc45" --dataset_root ./data/ --model internvl_c_classification \
--pretrained ./pretrained/internvl_c_13b_224px.pth --output result.json
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment