Adds the ability to load PyTorch models (read more)

- Python script de-cudas the models and exports them in the
understandable "zip" format. (The models output by MelGan, Tacotron, etc
do not conform to something Rust's libtorch bindings can understand by
default)

- Adds the Rust code to load but not the cargo deps since that needs a
refactor. The version is `tch = "0.1.3"`
This commit is contained in:
Brandon Thomas committed 2019-12-23 15:01:39 -05:00
1 parent b77de37029
commit 5d0d22c5f8
3 files changed
+122

No files matched your search

+1
View File
@@ -0,0 +1 @@
torch==1.3.1
+32
View File
@@ -0,0 +1,32 @@
extern crate tch;
//const TACOTRON_MODEL_PATH : &'static str = "/home/bt/models/tacotron2-nvidia/tacotron2_statedict.pt";
//const MELGAN_MODEL_PATH : &'static str = "/home/bt/models/melgan-swpark/firstgo_a7c2351_1100.pt";
const WRAPPED_MODEL_PATH : &'static str = "/home/bt/dev/voder/container.pt";
pub fn main() {
println!("Tacotron2 + MelGan");
// Create the model and load the weights from the file.
let mut vs = tch::nn::VarStore::new(tch::Device::Cpu);
println!("Loading Wrapped model: {}", WRAPPED_MODEL_PATH);
vs.load(WRAPPED_MODEL_PATH).unwrap();
println!("loaded");
// TODO: Now how do I evaluate the model?
/*// Apply the forward pass of the model to get the logits.
let output = net
.forward_t(&image.unsqueeze(0), /*train=*/ false)
.softmax(-1, tch::Kind::Float); // Convert to probability.
// Print the top 5 categories for this image.
for (probability, class) in imagenet::top(&output, 5).iter() {
println!("{:50} {:5.2}%", class, 100.0 * probability)
}
Ok(())*/
}
+89
View File
@@ -0,0 +1,89 @@
#!/usr/bin/env python3
# From: https://github.com/pytorch/pytorch/issues/20356#issuecomment-545572400
import torch
from collections import OrderedDict
class Container(torch.nn.Module):
def __init__(self, my_values):
super().__init__()
for key in my_values:
setattr(self, key, my_values[key])
def forward(self, x):
# NB(bt): This wasn't originally provided. I'm assuming this is a torch thing
#y = [self.layer1(x[0]), self.layer2(x[1])]
#return y
return x
# my_values = {
# 'a': torch.ones(2, 2),
# 'b': torch.ones(2, 2) + 10,
# 'c': 'hello',
# 'd': 6
#}
print('Loading melgan model...')
melgan_model_file = '/home/bt/models/melgan-swpark/firstgo_a7c2351_1100.pt'
melgan_model = torch.load(melgan_model_file)
def cuda_to_cpu(model):
"""Recursively make everything non-CUDA"""
if isinstance(model, dict) or isinstance(model, OrderedDict):
for key, value in model.items():
model[key] = cuda_to_cpu(value)
return model
if isinstance(model, list):
for i, value in enumerate(model):
model[i] = cuda_to_cpu(value)
elif isinstance(model, torch.Tensor):
return model.cpu()
else:
#print(type(model))
return model
print('Converitng to CPU model...')
cuda_to_cpu(melgan_model)
"""
for model_key, model_payload in melgan_model.items():
print("===== Model Key: {} =====".format(model_key))
if isinstance(model_payload, dict) or isinstance(model_payload, OrderedDict):
for k, v in model_payload.items():
#print("Key: {}".format(k))
#print("Type v: {}".format(type(v)))
if isinstance(v, torch.Tensor):
v = v.cpu()
#print(v)
for model_key, model_payload in melgan_model.items():
print("===== Model Key: {} =====".format(model_key))
if isinstance(model_payload, dict) or isinstance(model_payload, OrderedDict):
for k, v in model_payload.items():
pass
#print("Key: {}".format(k))
#print("Type v: {}".format(type(v)))
#print(v)
"""
print('Containerizing model...')
# Save arbitrary values supported by TorchScript
# https://pytorch.org/docs/master/jit.html#supported-type
graph = {
# NB: You can't save a toplevel dictionary of dictionaries,
# like so: 'melgan': melgan_model,
'melgan_model_g': melgan_model['model_g'],
'melgan_model_d': melgan_model['model_d'],
'melgan_optim_g': melgan_model['optim_g'],
'melgan_optim_d': melgan_model['optim_d'],
'melgan_step': melgan_model['step'],
'melgan_epoch': melgan_model['epoch'],
'melgan_hp_str': melgan_model['hp_str'],
'melgan_githash': melgan_model['githash'],
}
container = torch.jit.script(Container(graph))
print('Saving model...')
container.save("container.pt")