concept-learning/run.py at main · aw4309/concept-learning · GitHub

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
import yaml
import torch
import torch.nn as nn
import numpy as np
import argparse
import os
import shutil
import copy
import tqdm

from mltools.utils import cuda_tools
from mltools import ml_utils

import utils

def get_yaml_path():
    parser=argparse.ArgumentParser()
    parser.add_argument("config_path",type=str,default="")
    args=parser.parse_args()
    yaml_path=args.config_path
    assert os.path.exists(yaml_path)
    assert yaml_path.endswith(".yaml")
    return yaml_path

if __name__ == "__main__":
    device=cuda_tools.get_freer_device(verbose=False)
    yaml_path=get_yaml_path()
    print("Running:",yaml_path)
    config=utils.load_config(yaml_path)
    experiment_directory=config["experiment_directory"]
    if os.path.exists(experiment_directory):
        shutil.rmtree(experiment_directory)
    os.makedirs(experiment_directory)
    shutil.copyfile(yaml_path,os.path.join(experiment_directory,os.path.split(yaml_path)[-1]))

    seed=config["seed"]
    if seed is not None:
        ml_utils.seed_all(seed)

    x_tr,y_tr,l_tr,x_te,y_te,l_te=utils.generate_data(config)
    x_tr_gen,y_tr_gen,l_tr_gen,x_te_gen,y_te_gen,l_te_gen=utils.generate_data(config,forgen=True)
    dataset_tr=torch.utils.data.TensorDataset(torch.tensor(x_tr).float(),torch.tensor(y_tr).float())
    dataset_te=torch.utils.data.TensorDataset(torch.tensor(x_te).float(),torch.tensor(y_te).float())

    utils.draw_setup(config,x_tr,x_te,os.path.join(experiment_directory,"setup.png"),l_tr=l_tr,l_te=l_te)

    n=len(dataset_tr)
    n_tr=int(n*config["train_ratio"])
    n_val=n-n_tr
    dataset_tr_tr,dataset_tr_val=torch.utils.data.random_split(dataset_tr,[n_tr,n_val])
    #print(dataset_tr_tr[0])

    model=utils.get_model(config)
    model=model.to(device)
    #print(model)

    batch_size=config["batch_size"]
    tr_loader=torch.utils.data.DataLoader(dataset_tr_tr,batch_size=batch_size,shuffle=True,num_workers=4)
    val_loader=torch.utils.data.DataLoader(dataset_tr_val,batch_size=batch_size,shuffle=False,num_workers=4)
    te_loader=torch.utils.data.DataLoader(dataset_te,batch_size=batch_size,shuffle=False,num_workers=4)

    ckpt_fol=os.path.join(experiment_directory,f"ckpts")
    os.makedirs(ckpt_fol)
    figure_fol=os.path.join(experiment_directory,f"figures")
    os.makedirs(figure_fol)
    generation_fol=os.path.join(experiment_directory,f"generations")
    os.makedirs(generation_fol)

    unconditioned=config.get("unconditioned",False)

    def batch_to_kwargs(batch):
        x,c=batch
        x=x.to(device)
        c=c.to(device)
        if unconditioned:
            c[:]=0.
        return {"x":x,"c":c}

    save_steps=[]
    val_losses=[]
    te_losses=[]
    ckpt_paths=[]
    def append_save_step(step,model):
        model.eval()
        save_steps.append(step)
    def append_val_te_loss(step,model):
        model.eval()
        with torch.no_grad():
            val_loss=0
            for batch in val_loader:
                loss=model.get_loss(**batch_to_kwargs(batch))
                val_loss+=loss.item()
            val_loss/=len(val_loader)
            te_loss=0
            for batch in te_loader:
                loss=model.get_loss(**batch_to_kwargs(batch))
                te_loss+=loss.item()
            te_loss/=len(te_loader)
        val_losses.append(val_loss)
        te_losses.append(te_loss)
    def save_ckpt(step,model):
        model.eval()
        ckpt_path=os.path.join(ckpt_fol,f"ckpt_step={step}.pth")
        torch.save(model.state_dict(),ckpt_path)
        ckpt_paths.append(ckpt_path)
    def generate(step,model):
        model.eval()
        c_tr=torch.tensor(y_tr_gen).float().to(device)
        c_te=torch.tensor(y_te_gen).float().to(device)
        if unconditioned:
            c_tr[:]=0.
            c_te[:]=0.
        gen_tr=model.generate(c=c_tr).detach().cpu().numpy()
        gen_te=model.generate(c=c_te).detach().cpu().numpy()
        gens={}
        gens["gen_tr"]=gen_tr
        gens["gen_te"]=gen_te
        gens["y_tr_gen"]=y_tr_gen
        gens["y_te_gen"]=y_te_gen
        gens["l_tr_gen"]=l_tr_gen
        gens["l_te_gen"]=l_te_gen
        generation_path=os.path.join(generation_fol,f"gen_{step}.pth")
        torch.save(gens,generation_path)
        figure_path=os.path.join(figure_fol,f"gen_{step}.png")
        utils.draw_setup(config,gen_tr,gen_te,figure_path,l_tr=l_tr_gen,l_te=l_te_gen)
    if config.get("save_ckpts",True):
        callbacks=[append_save_step,append_val_te_loss,save_ckpt,generate]
    else:
        callbacks=[append_save_step,append_val_te_loss,generate]

    train_results=ml_utils.train(model=model,dl_tr=tr_loader,batch_to_kwargs=batch_to_kwargs,
    n_steps=config["num_steps"],callback_steps=config["save_steps"],callbacks=callbacks,device=device)
    time_tr=train_results["time_tr"]
    time_callbacks=train_results["time_callbacks"]
    t_tr_r=time_tr/(time_tr+time_callbacks)*100
    t_cb_r=time_callbacks/(time_tr+time_callbacks)*100
    print(f"Time Split: Train:{t_tr_r:.2f}% Callback:{t_cb_r:.2f}%")

    logs={}
    logs.update(train_results)
    if hasattr(model,"data_noise"):#variational diffusion model
        logs["min_vlb"]=model.get_min_vlb()
    logs["losses"]=np.array(train_results["online_losses"])
    logs["save_steps"]=np.array(save_steps)
    logs["val_losses"]=np.array(val_losses)
    logs["te_losses"]=np.array(te_losses)
    logs["ckpt_paths"]=ckpt_paths
    torch.save(logs,os.path.join(experiment_directory,"logs.pth"))

    utils.plot_losses(logs,os.path.join(experiment_directory,"losses.png"))

    model.eval()