-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_wab.py
183 lines (165 loc) · 7.54 KB
/
train_wab.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
import torch
import torch.nn as nn
import torch.optim as optim
import numpy as np
from tqdm import tqdm
from NewDataPrep import *
from metric import *
from model2 import MultiVAE,loss_function
from torch.utils.data import DataLoader
import wandb
import random
import yaml
from collections import defaultdict
import traceback
import os
# os.environ["CUDA_LAUNCH_BLOCKING"] = "1"
# os.environ["TORCH_USE_CUDA_DSA"] = "1"
config1 = yaml.load(open("config.yaml","r"),Loader=yaml.FullLoader)
print(config1)
sweep_id = wandb.sweep(config1, project="recsys")
device = torch.device("cpu")
if torch.cuda.is_available():
device = torch.device("cuda:0")
print(device)
###############################################################################
# Weights and Biases
###############################################################################
def train_wrapper(config=None):
run = wandb.init(config=config)
try:
train()
except Exception as e:
print(traceback.format_exc())
def train(config=None):
config = wandb.config
lr = config.lr
batch_size = config.batch_size
epochs = config.epochs
metric_names = ["ndcg@100","bi_ndcg@100" ,"hit_rate@20","bi_hit_rate@20"]
###############################################################################
# Load data
###############################################################################
data = pd.read_feather("ratings_ipc_processed.ipc")
side_info = False
# side_info = pd.read_feather("user_stats.ipc")
users = data['username'].unique()
np.random.shuffle(users)
splits = np.split(users,[int(.8*len(users)),int(.9*len(users))])
train_users,val_users,test_users = data[data['username'].isin(splits[0])],data[data['username'].isin(splits[1])],data[data['username'].isin(splits[2])]
# train_side,val_side,test_side = side_info[side_info['username'].isin(splits[0])],side_info[side_info['username'].isin(splits[1])],side_info[side_info['username'].isin(splits[2])]
print("split data")
#n_items safer to obtain from item_map.csv
train_loader = TrainDataloader(train_users,batch_size=batch_size,n_items=10662)
# test_loader =TestDataloader(test_users,batch_size=batch_size,mode="80/20",n_items=10661)
valid_loader = TestDataloader(val_users,batch_size=batch_size//2,mode="80/20",n_items=10662) #,side_info=val_side.copy(deep=True)
n_items = train_loader.n_items
# idxlist = list(range(N))
print("Data loaded")
###############################################################################
# Build the model
###############################################################################
anneal_cap = config.anneal_cap
q_dims = [n_items+ (side_info.shape[1]-1) if side_info else 0,config.layer1, config.layer2]
p_dims = [config.layer2, config.layer1, n_items]
total_anneal_steps = 10000
del data,users,train_users,val_users,test_users
#del train_side,val_side,test_side
model = MultiVAE(p_dims,dropout=config.dropout).to(device)
print(f"Model: {p_dims}, dropout: {config.dropout}")
optimizer = optim.Adam(model.parameters(), lr=lr)
criterion = loss_function
###############################################################################
# Training code
###############################################################################
# Turn on training mode
model.train()
scaler = torch.cuda.amp.GradScaler()
total_anneal_steps = (len(train_loader) * epochs) // (1.5*anneal_cap)
update_count = 0
best_n100 = -np.inf
train_loss_hist = torch.zeros(epochs)
# np.random.shuffle(idxlist)
positives,negative = torch.zeros((batch_size,n_items),device=device),torch.zeros((batch_size,n_items),device=device)
for epoch in range(1, epochs + 1):
train_loss = 0.0
for data,side in iter(tqdm(train_loader,position=0, leave=True)):
positives.zero_()
negative.zero_()
positives[data==1] = 1
negative[data==-1] = 1
if side is not None:
data = torch.cat([data,side],1)
# print(data.shape,data[0].shape)
if total_anneal_steps > 0:
anneal = min(anneal_cap,
1. * update_count / total_anneal_steps)
else:
anneal = anneal_cap
optimizer.zero_grad()
with torch.cuda.amp.autocast(dtype=torch.float16):
recon_batch, mu, logvar = model(data)
loss = criterion(recon_batch, positives, mu, logvar, anneal)
scaler.scale(loss).backward()
train_loss += loss.item()
scaler.step(optimizer)
scaler.update()
update_count += 1
wandb.log({"train total_loss":train_loss})
tqdm.write(f"Epoch {epoch}: train loss {train_loss}, anneal: {anneal}")
metrics = evaluate(model, valid_loader,anneal,metric_names,prefix="validation ")
# Save the model if the n100 is the best we've seen so far.
if metrics["validation bi_ndcg@100"] > best_n100:
with open("models/best_multvae.pt"+str(model.p_dims), 'wb') as f:
torch.save(model.state_dict(), f)
tqdm.write("Saving model (new best validation n100)")
#wand file upload
best_n100 = metrics["validation bi_ndcg@100"]
#Check for increases in ndcg for early stopping
# train_loss_hist[epoch - 1] = -val_ndcg
# diffs = torch.diff(train_loss_hist)
# if (diffs > 0).sum() > 4:
# tqdm.write("Early stopping")
# return
# evaluate(model, test_loader,anneal,metric_names,prefix="test ")
art = wandb.Artifact(f"multvae-{wandb.run.id}", type="model")
art.add_file("models/best_multvae.pt"+str(model.p_dims))
wandb.log_artifact(art)
wandb.finish()
def evaluate(model, valid_loader,anneal,metric_names,criterion=loss_function,prefix="test "):
# Turn on evaluation mode
model.eval()
metric_logs = defaultdict(list)
total_loss = 0.0
positives,negative = torch.zeros((valid_loader.batch_size,valid_loader.n_items),device=device),torch.zeros((valid_loader.batch_size,valid_loader.n_items),device=device)
positives_heldout,negatives_heldout = torch.zeros((valid_loader.batch_size,valid_loader.n_items),device=device),torch.zeros((valid_loader.batch_size,valid_loader.n_items),device=device)
with torch.no_grad():
for data,heldout_data,side in iter(tqdm(valid_loader,position=0, leave=True,mininterval=2.0,miniters=1)):
positives.zero_()
negative.zero_()
positives_heldout.zero_()
negatives_heldout.zero_()
positives[data==1] = 1
negative[data==-1] = 1
positives_heldout[heldout_data==1] = 1
negatives_heldout[heldout_data==-1] = 1
if side is not None:
data = torch.cat([data,side],1)
recon_batch, mu, logvar = model(data)
loss = criterion(recon_batch, positives, mu, logvar, anneal)
total_loss += loss.item()
# Exclude examples from training set because they will not be recommended
recon_batch[positives == 1] = -torch.inf
metrics = Metric(recon_batch,positives_heldout,negatives_heldout,metric_names,device=device)
res = metrics.calculate()
for k,v in res.items():
metric_logs[k].append(v)
metric_logs_out = dict()
for k,v in metric_logs.items():
metric_logs_out[prefix+k] = np.mean(v)
metric_logs_out[prefix+'total_loss'] = total_loss
wandb.log(metric_logs_out)
tqdm.write(Metric.prettify(metric_logs_out))
model.train()
return metric_logs_out
wandb.agent(sweep_id, train_wrapper, count=16)