DL-Art-School/codes/trainer/injectors/base_injectors.py
James Betker 596a62fe01 Apply fix to gpt_asr_hf and prep it for inference
Fix is that we were predicting two characters in advance, not next character
2021-11-04 10:09:24 -06:00

574 lines
22 KiB
Python

import random
import torch.nn
from kornia.augmentation import RandomResizedCrop
from torch.cuda.amp import autocast
from trainer.inject import Injector, create_injector
from trainer.losses import extract_params_from_state
from utils.util import opt_get
from utils.weight_scheduler import get_scheduler_for_opt
class PadInjector(Injector):
def __init__(self, opt, env):
super().__init__(opt, env)
self.multiple = opt['multiple']
def forward(self, state):
ldim = state[self.input].shape[-1]
mod = self.multiple-(ldim % self.multiple)
t = state[self.input]
if mod != 0:
t = torch.nn.functional.pad(t, (0, mod))
return {self.output: t}
class SqueezeInjector(Injector):
def __init__(self, opt, env):
super().__init__(opt, env)
self.dim = opt['dim']
def forward(self, state):
return {self.output: state[self.input].squeeze(dim=self.dim)}
# Uses a generator to synthesize an image from [in] and injects the results into [out]
# Note that results are *not* detached.
class GeneratorInjector(Injector):
def __init__(self, opt, env):
super(GeneratorInjector, self).__init__(opt, env)
self.grad = opt['grad'] if 'grad' in opt.keys() else True
self.method = opt_get(opt, ['method'], None) # If specified, this method is called instead of __call__()
self.args = opt_get(opt, ['args'], {})
def forward(self, state):
gen = self.env['generators'][self.opt['generator']]
if self.method is not None and hasattr(gen, 'module'):
gen = gen.module # Dereference DDP wrapper.
method = gen if self.method is None else getattr(gen, self.method)
with autocast(enabled=self.env['opt']['fp16']):
if isinstance(self.input, list):
params = extract_params_from_state(self.input, state)
else:
params = [state[self.input]]
if self.grad:
results = method(*params, **self.args)
else:
with torch.no_grad():
results = method(*params, **self.args)
new_state = {}
if isinstance(self.output, list):
# Only dereference tuples or lists, not tensors. IF YOU REACH THIS ERROR, REMOVE THE BRACES AROUND YOUR OUTPUTS IN THE YAML CONFIG
assert isinstance(results, list) or isinstance(results, tuple)
for i, k in enumerate(self.output):
new_state[k] = results[i]
else:
new_state[self.output] = results
return new_state
# Injects a result from a discriminator network into the state.
class DiscriminatorInjector(Injector):
def __init__(self, opt, env):
super(DiscriminatorInjector, self).__init__(opt, env)
def forward(self, state):
with autocast(enabled=self.env['opt']['fp16']):
d = self.env['discriminators'][self.opt['discriminator']]
if isinstance(self.input, list):
params = [state[i] for i in self.input]
results = d(*params)
else:
results = d(state[self.input])
new_state = {}
if isinstance(self.output, list):
# Only dereference tuples or lists, not tensors.
assert isinstance(results, list) or isinstance(results, tuple)
for i, k in enumerate(self.output):
new_state[k] = results[i]
else:
new_state[self.output] = results
return new_state
# Injects a scalar that is modulated with a specified schedule. Useful for increasing or decreasing the influence
# of something over time.
class ScheduledScalarInjector(Injector):
def __init__(self, opt, env):
super(ScheduledScalarInjector, self).__init__(opt, env)
self.scheduler = get_scheduler_for_opt(opt['scheduler'])
def forward(self, state):
return {self.opt['out']: self.scheduler.get_weight_for_step(self.env['step'])}
# Adds gaussian noise to [in], scales it to [0,[scale]] and injects into [out]
class AddNoiseInjector(Injector):
def __init__(self, opt, env):
super(AddNoiseInjector, self).__init__(opt, env)
self.mode = opt['mode'] if 'mode' in opt.keys() else 'normal'
def forward(self, state):
# Scale can be a fixed float, or a state key (e.g. from ScheduledScalarInjector).
if isinstance(self.opt['scale'], str):
scale = state[self.opt['scale']]
else:
scale = self.opt['scale']
if scale is None:
scale = 1
ref = state[self.opt['in']]
if self.mode == 'normal':
noise = torch.randn_like(ref) * scale
elif self.mode == 'uniform':
noise = torch.FloatTensor(ref.shape).uniform_(0.0, scale).to(ref.device)
return {self.opt['out']: state[self.opt['in']] + noise}
# Averages the channel dimension (1) of [in] and saves to [out]. Dimensions are
# kept the same, the average is simply repeated.
class GreyInjector(Injector):
def __init__(self, opt, env):
super(GreyInjector, self).__init__(opt, env)
def forward(self, state):
mean = torch.mean(state[self.opt['in']], dim=1, keepdim=True)
mean = mean.repeat(1, 3, 1, 1)
return {self.opt['out']: mean}
class InterpolateInjector(Injector):
def __init__(self, opt, env):
super(InterpolateInjector, self).__init__(opt, env)
if 'scale_factor' in opt.keys():
self.scale_factor = opt['scale_factor']
self.size = None
else:
self.scale_factor = None
self.size = (opt['size'], opt['size'])
def forward(self, state):
scaled = torch.nn.functional.interpolate(state[self.opt['in']], scale_factor=self.opt['scale_factor'],
size=self.opt['size'], mode=self.opt['mode'])
return {self.opt['out']: scaled}
# Extracts four patches from the input image, each a square of 'patch_size'. The input images are taken from each
# of the four corners of the image. The intent of this loss is that each patch shares some part of the input, which
# can then be used in the translation invariance loss.
#
# This injector is unique in that it does not only produce the specified output label into state. Instead it produces five
# outputs for the specified label, one for each corner of the input as well as the specified output, which is the top left
# corner. See the code below to find out how this works.
#
# Another note: this injector operates differently in eval mode (e.g. when env['training']=False) - in this case, it
# simply sets all the output state variables to the input. This is so that you can feed the output of this injector
# directly into your generator in training without affecting test performance.
class ImagePatchInjector(Injector):
def __init__(self, opt, env):
super(ImagePatchInjector, self).__init__(opt, env)
self.patch_size = opt['patch_size']
self.resize = opt[
'resize'] if 'resize' in opt.keys() else None # If specified, the output is resized to a square with this size after patch extraction.
def forward(self, state):
im = state[self.opt['in']]
if self.env['training']:
res = {self.opt['out']: im[:, :3, :self.patch_size, :self.patch_size],
'%s_top_left' % (self.opt['out'],): im[:, :, :self.patch_size, :self.patch_size],
'%s_top_right' % (self.opt['out'],): im[:, :, :self.patch_size, -self.patch_size:],
'%s_bottom_left' % (self.opt['out'],): im[:, :, -self.patch_size:, :self.patch_size],
'%s_bottom_right' % (self.opt['out'],): im[:, :, -self.patch_size:, -self.patch_size:]}
else:
res = {self.opt['out']: im,
'%s_top_left' % (self.opt['out'],): im,
'%s_top_right' % (self.opt['out'],): im,
'%s_bottom_left' % (self.opt['out'],): im,
'%s_bottom_right' % (self.opt['out'],): im}
if self.resize is not None:
res2 = {}
for k, v in res.items():
res2[k] = torch.nn.functional.interpolate(v, size=(self.resize, self.resize), mode="nearest")
res = res2
return res
# Concatenates a list of tensors on the specified dimension.
class ConcatenateInjector(Injector):
def __init__(self, opt, env):
super(ConcatenateInjector, self).__init__(opt, env)
self.dim = opt['dim']
def forward(self, state):
input = [state[i] for i in self.input]
return {self.opt['out']: torch.cat(input, dim=self.dim)}
# Removes margins from an image.
class MarginRemoval(Injector):
def __init__(self, opt, env):
super(MarginRemoval, self).__init__(opt, env)
self.margin = opt['margin']
self.random_shift_max = opt['random_shift_max'] if 'random_shift_max' in opt.keys() else 0
def forward(self, state):
input = state[self.input]
if self.random_shift_max > 0:
output = []
# This is a really shitty way of doing this. If it works at all, I should reconsider using Resample2D, for example.
for b in range(input.shape[0]):
shiftleft = random.randint(-self.random_shift_max, self.random_shift_max)
shifttop = random.randint(-self.random_shift_max, self.random_shift_max)
output.append(input[b, :, self.margin + shiftleft:-(self.margin - shiftleft),
self.margin + shifttop:-(self.margin - shifttop)])
output = torch.stack(output, dim=0)
else:
output = input[:, :, self.margin:-self.margin,
self.margin:-self.margin]
return {self.opt['out']: output}
# Produces an injection which is composed of applying a single injector multiple times across a single dimension.
class ForEachInjector(Injector):
def __init__(self, opt, env):
super(ForEachInjector, self).__init__(opt, env)
o = opt.copy()
o['type'] = opt['subtype']
o['in'] = '_in'
o['out'] = '_out'
self.injector = create_injector(o, self.env)
self.aslist = opt['aslist'] if 'aslist' in opt.keys() else False
def forward(self, state):
injs = []
st = state.copy()
inputs = state[self.opt['in']]
for i in range(inputs.shape[1]):
st['_in'] = inputs[:, i]
injs.append(self.injector(st)['_out'])
if self.aslist:
return {self.output: injs}
else:
return {self.output: torch.stack(injs, dim=1)}
class ConstantInjector(Injector):
def __init__(self, opt, env):
super(ConstantInjector, self).__init__(opt, env)
self.constant_type = opt['constant_type']
self.like = opt['like'] # This injector uses this tensor to determine what batch size and device to use.
def forward(self, state):
like = state[self.like]
if self.constant_type == 'zeroes':
out = torch.zeros_like(like)
else:
raise NotImplementedError
return {self.opt['out']: out}
class IndicesExtractor(Injector):
def __init__(self, opt, env):
super(IndicesExtractor, self).__init__(opt, env)
self.dim = opt['dim']
assert self.dim == 1 # Honestly not sure how to support an abstract dim here, so just add yours when needed.
def forward(self, state):
results = {}
for i, o in enumerate(self.output):
if self.dim == 1:
results[o] = state[self.input][:, i]
return results
class RandomShiftInjector(Injector):
def __init__(self, opt, env):
super(RandomShiftInjector, self).__init__(opt, env)
def forward(self, state):
img = state[self.input]
return {self.output: img}
class BatchRotateInjector(Injector):
def __init__(self, opt, env):
super(BatchRotateInjector, self).__init__(opt, env)
def forward(self, state):
img = state[self.input]
return {self.output: torch.roll(img, 1, 0)}
# Injector used to work with image deltas used in diff-SR
class SrDiffsInjector(Injector):
def __init__(self, opt, env):
super(SrDiffsInjector, self).__init__(opt, env)
self.mode = opt['mode']
assert self.mode in ['recombine', 'produce_diff']
self.lq = opt['lq']
self.hq = opt['hq']
if self.mode == 'produce_diff':
self.diff_key = opt['diff']
self.include_combined = opt['include_combined']
def forward(self, state):
resampled_lq = state[self.lq]
hq = state[self.hq]
if self.mode == 'produce_diff':
diff = hq - resampled_lq
if self.include_combined:
res = torch.cat([resampled_lq, diff, hq], dim=1)
else:
res = torch.cat([resampled_lq, diff], dim=1)
return {self.output: res,
self.diff_key: diff}
elif self.mode == 'recombine':
combined = resampled_lq + hq
return {self.output: combined}
class MultiFrameCombiner(Injector):
def __init__(self, opt, env):
super().__init__(opt, env)
self.mode = opt['mode']
self.dim = opt['dim'] if 'dim' in opt.keys() else None
self.flow = opt['flow']
self.in_lq_key = opt['in']
self.in_hq_key = opt['in_hq']
self.out_lq_key = opt['out']
self.out_hq_key = opt['out_hq']
from models.flownet2.networks import Resample2d
self.resampler = Resample2d()
def combine(self, state):
flow = self.env['generators'][self.flow]
lq = state[self.in_lq_key]
hq = state[self.in_hq_key]
b, f, c, h, w = lq.shape
center = f // 2
center_img = lq[:, center, :, :, :]
imgs = [center_img]
with torch.no_grad():
for i in range(f):
if i == center:
continue
nimg = lq[:, i, :, :, :]
flowfield = flow(torch.stack([center_img, nimg], dim=2).float())
nimg = self.resampler(nimg, flowfield)
imgs.append(nimg)
hq_out = hq[:, center, :, :, :]
return {self.out_lq_key: torch.cat(imgs, dim=1),
self.out_hq_key: hq_out,
self.out_lq_key + "_flow_sample": torch.cat(imgs, dim=0)}
def synthesize(self, state):
lq = state[self.in_lq_key]
return {
self.out_lq_key: lq.repeat(1, self.dim, 1, 1)
}
def forward(self, state):
if self.mode == "synthesize":
return self.synthesize(state)
elif self.mode == "combine":
return self.combine(state)
else:
raise NotImplementedError
# Combines data from multiple different sources and mixes them along the batch dimension. Labels are then emitted
# according to how the mixing was performed.
class MixAndLabelInjector(Injector):
def __init__(self, opt, env):
super().__init__(opt, env)
self.out_labels = opt['out_labels']
def forward(self, state):
input_tensors = [state[i] for i in self.input]
num_inputs = len(input_tensors)
bs = input_tensors[0].shape[0]
labels = torch.randint(0, num_inputs, (bs,), device=input_tensors[0].device)
# Still don't know of a good way to do this in torch.. TODO make it better..
res = []
for b in range(bs):
res.append(input_tensors[labels[b]][b, :, :, :])
output = torch.stack(res, dim=0)
return {self.out_labels: labels, self.output: output}
# Randomly performs a uniform resize & crop from a base image.
# Never resizes below input resolution or messes with the aspect ratio.
class RandomCropInjector(Injector):
def __init__(self, opt, env):
super().__init__(opt, env)
dim_in = opt['dim_in']
dim_out = opt['dim_out']
scale = dim_out / dim_in
self.operator = RandomResizedCrop(size=(dim_out, dim_out), scale=(scale, 1),
ratio=(.99,1), # An aspect ratio range is required, but .99,1 is effectively "none".
resample='NEAREST')
def forward(self, state):
return {self.output: self.operator(state[self.input])}
class Stylegan2NoiseInjector(Injector):
def __init__(self, opt, env):
super().__init__(opt, env)
self.mix_prob = opt_get(opt, ['mix_probability'], .9)
self.latent_dim = opt_get(opt, ['latent_dim'], 512)
def make_noise(self, batch, latent_dim, n_noise, device):
return torch.randn(n_noise, batch, latent_dim, device=device).unbind(0)
def forward(self, state):
i = state[self.input]
if self.mix_prob > 0 and random.random() < self.mix_prob:
return {self.output: self.make_noise(i.shape[0], self.latent_dim, 2, i.device)}
else:
return {self.output: self.make_noise(i.shape[0], self.latent_dim, 1, i.device)}
class NoiseInjector(Injector):
def __init__(self, opt, env):
super().__init__(opt, env)
self.shape = tuple(opt['shape'])
def forward(self, state):
shape = (state[self.input].shape[0],) + self.shape
return {self.output: torch.randn(shape, device=state[self.input].device)}
# Incorporates the specified dimension into the batch dimension.
class DecomposeDimensionInjector(Injector):
def __init__(self, opt, env):
super().__init__(opt, env)
self.dim = opt['dim']
self.cutoff_dim = opt_get(opt, ['cutoff_dim'], -1)
assert self.dim != 0 # Cannot decompose the batch dimension
def forward(self, state):
inp = state[self.input]
dims = list(range(len(inp.shape))) # Looks like [0,1,2,3]
shape = list(inp.shape)
del dims[self.dim]
del shape[self.dim]
# Compute the reverse permutation and shape arguments needed to undo this operation.
rev_shape = [inp.shape[self.dim]] + shape.copy()
rev_permute = list(range(len(inp.shape)))[1:] # Looks like [1,2,3]
rev_permute = rev_permute[:self.dim] + [0] + (rev_permute[self.dim:] if self.dim < len(rev_permute) else [])
out = inp.permute([self.dim] + dims).reshape((-1,) + tuple(shape[1:]))
if self.cutoff_dim > -1:
out = out[:self.cutoff_dim]
return {self.output: out,
f'{self.output}_reverse_shape': rev_shape,
f'{self.output}_reverse_permute': rev_permute}
# Undoes a decompose.
class RecomposeDimensionInjector(Injector):
def __init__(self, opt, env):
super().__init__(opt, env)
self.rev_shape_key = opt['reverse_shape']
self.rev_permute_key = opt['reverse_permute']
def forward(self, state):
inp = state[self.input]
rev_shape = state[self.rev_shape_key]
rev_permute = state[self.rev_permute_key]
out = inp.reshape(rev_shape)
out = out.permute(rev_permute).contiguous()
return {self.output: out}
# Performs normalization across fixed constants.
class NormalizeInjector(Injector):
def __init__(self, opt, env):
super().__init__(opt, env)
self.shift = opt['shift']
self.scale = opt['scale']
def forward(self, state):
inp = state[self.input]
out = (inp - self.shift) / self.scale
return {self.output: out}
# Performs frequency-bin normalization for spectrograms.
class FrequencyBinNormalizeInjector(Injector):
def __init__(self, opt, env):
super().__init__(opt, env)
self.shift, self.scale = torch.load(opt['stats_file'])
self.shift = self.shift.view(1,-1,1)
self.scale = self.scale.view(1,-1,1)
def forward(self, state):
inp = state[self.input]
self.shift = self.shift.to(inp.device)
self.scale = self.scale.to(inp.device)
out = (inp - self.shift) / self.scale
return {self.output: out}
# Performs normalization across fixed constants.
class DenormalizeInjector(Injector):
def __init__(self, opt, env):
super().__init__(opt, env)
self.shift = opt['shift']
self.scale = opt['scale']
def forward(self, state):
inp = state[self.input]
out = inp * self.scale + self.shift
return {self.output: out}
# Performs normalization across fixed constants.
class MelSpectrogramInjector(Injector):
def __init__(self, opt, env):
super().__init__(opt, env)
from models.tacotron2.layers import TacotronSTFT
# These are the default tacotron values for the MEL spectrogram.
filter_length = opt_get(opt, ['filter_length'], 1024)
hop_length = opt_get(opt, ['hop_length'], 256)
win_length = opt_get(opt, ['win_length'], 1024)
n_mel_channels = opt_get(opt, ['n_mel_channels'], 80)
mel_fmin = opt_get(opt, ['mel_fmin'], 0)
mel_fmax = opt_get(opt, ['mel_fmax'], 8000)
sampling_rate = opt_get(opt, ['sampling_rate'], 22050)
self.stft = TacotronSTFT(filter_length, hop_length, win_length, n_mel_channels, sampling_rate, mel_fmin, mel_fmax)
def forward(self, state):
inp = state[self.input]
if len(inp.shape) == 3: # Automatically squeeze out the channels dimension if it is present (assuming mono-audio)
inp = inp.squeeze(1)
assert len(inp.shape) == 2
self.stft = self.stft.to(inp.device)
return {self.output: self.stft.mel_spectrogram(inp)}
class RandomAudioCropInjector(Injector):
def __init__(self, opt, env):
super().__init__(opt, env)
self.crop_sz = opt['crop_size']
def forward(self, state):
inp = state[self.input]
len = inp.shape[-1]
margin = len - self.crop_sz
start = random.randint(0, margin)
return {self.output: inp[:, :, start:start+self.crop_sz]}
if __name__ == '__main__':
inj = MelSpectrogramInjector({'in': 'x', 'out': 'y'}, None)
print(inj({'x':torch.rand(10,1,40800)})['y'].shape)