Upload 6 files

data_preparation.py

@@ -0,0 +1,102 @@
+import random
+
+import numpy as np
+import torch
+import torch.utils.data
+
+import nn_layers
+from scipy.io.wavfile import read
+from text import text_to_sequence
+from hyper_parameters import tacotron_params
+
+
+class DataPreparation(torch.utils.data.Dataset):
+
+    def __init__(self, audiopaths_and_text, tacotron_hyperparams):
+        self.audiopaths_and_text = audiopaths_and_text
+        self.audio_text_parameters = tacotron_hyperparams
+        self.stft = nn_layers.TacotronSTFT(tacotron_hyperparams['filter_length'], tacotron_hyperparams['hop_length'],
+                                           tacotron_hyperparams['win_length'], tacotron_hyperparams['n_mel_channels'],
+                                           self.audio_text_parameters['sampling_rate'],
+                                           tacotron_hyperparams['mel_fmin'], tacotron_hyperparams['mel_fmax'])
+        random.seed(1234)
+        random.shuffle(self.audiopaths_and_text)
+
+    def load_audiowav_torch(self, audiopath, samp_rate):
+        sr, data = read(audiopath)
+        assert samp_rate == sr, "Sample rate does not match with the configuration"
+
+        return torch.FloatTensor(data.astype(np.float32))
+
+    def melspec_textSequence_pair(self, audiopath_and_text):
+        wav_path, sentence = audiopath_and_text[0], audiopath_and_text[1]
+        # wav to torch tensor
+        wav_torch = self.load_audiowav_torch(wav_path, self.audio_text_parameters['sampling_rate'])
+        wav_torch_norm = wav_torch / self.audio_text_parameters['max_wav_value']
+        wav_torch_norm = wav_torch_norm.unsqueeze(0)
+        wav_torch_norm = torch.autograd.Variable(wav_torch_norm, requires_grad=False)
+        mel_spec = self.stft.mel_spectrogram(wav_torch_norm)
+        mel_spec = torch.squeeze(mel_spec, 0)
+        # text to torch integer tensor sequence
+        sentence_sequence = torch.IntTensor(text_to_sequence(sentence, self.audio_text_parameters['text_cleaners']))
+
+        return sentence_sequence, mel_spec
+
+    def __getitem__(self, index):
+        return self.melspec_textSequence_pair(self.audiopaths_and_text[index])
+
+    def __len__(self):
+        return len(self.audiopaths_and_text)
+
+
+class DataCollate:
+
+    def __init__(self, number_frames_step):
+        self.number_frames_step = number_frames_step
+
+    def __call__(self, batch):
+        inp_lengths, sorted_decreasing = torch.sort(torch.LongTensor([len(x[0]) for x in batch]),
+                                                   dim=0, descending=True)
+        max_length_in = inp_lengths[0]
+
+        # padding sentences sequences for a fixed-length tensor size
+        sentences_padded = torch.LongTensor(len(batch), max_length_in)
+        sentences_padded.zero_()
+        for i in range(len(sorted_decreasing)):
+            int_seq_sentence = batch[sorted_decreasing[i]][0]
+            # all slots of a line until the end of the sentence. The rest, 0's
+            sentences_padded[i, :int_seq_sentence.size(0)] = int_seq_sentence
+
+        # length of the mel filterbank used
+        num_melfilters = batch[0][1].size(0)
+
+        # longest recorded spectrogram representation + 1 space to mark the end
+        max_length_target = max([x[1].size(1) for x in batch])  # THERE IS A CHANGE FROM THE ORIGINAL CODE!!!
+        # add extra space if the number of frames per step is higher than 1
+        if max_length_target % self.number_frames_step != 0:
+            max_length_target += self.number_frames_step - max_length_target % self.number_frames_step
+            assert max_length_target % self.number_frames_step == 0
+
+        # padding mel spectrogram representations. The output is a 3D tensor
+        melspec_padded = torch.FloatTensor(len(batch), num_melfilters, max_length_target)
+        melspec_padded.zero_()
+
+        # GST new prosody matrices definition with zero padding:
+        prosody_padded = torch.FloatTensor(len(batch), num_melfilters, max_length_target)
+        prosody_padded.zero_()
+
+        gate_padded = torch.FloatTensor(len(batch), max_length_target)
+        gate_padded.zero_()
+        output_lengths = torch.LongTensor(len(batch))
+
+        for j in range(len(sorted_decreasing)):
+            melspec = batch[sorted_decreasing[j]][1]
+            melspec_padded[j, :, :melspec.size(1)] = melspec
+
+            # GST filling padded prosody matrix:
+            prosody_padded[j, :, :melspec.size(1)] = melspec
+
+            gate_padded[j, melspec.size(1) - 1:] = 1
+            output_lengths[j] = melspec.size(1)
+
+        return sentences_padded, inp_lengths, melspec_padded, gate_padded, output_lengths, prosody_padded

distributed.py

@@ -0,0 +1,180 @@
+import torch
+import torch.distributed as dist
+from torch.nn.modules import Module
+from torch.autograd import Variable
+
+
+def _flatten_dense_tensors(tensors):
+    """Flatten dense tensors into a contiguous 1D buffer. Assume tensors are of
+    same dense type.
+    Since inputs are dense, the resulting tensor will be a concatenated 1D
+    buffer. Element-wise operation on this buffer will be equivalent to
+    operating individually.
+    Arguments:
+        tensors (Iterable[Tensor]): dense tensors to flatten.
+    Returns:
+        A contiguous 1D buffer containing input tensors.
+    """
+    if len(tensors) == 1:
+        return tensors[0].contiguous().view(-1)
+    flat = torch.cat([t.contiguous().view(-1) for t in tensors], dim=0)
+    return flat
+
+
+def _unflatten_dense_tensors(flat, tensors):
+    """View a flat buffer using the sizes of tensors. Assume that tensors are of
+    same dense type, and that flat is given by _flatten_dense_tensors.
+    Arguments:
+        flat (Tensor): flattened dense tensors to unflatten.
+        tensors (Iterable[Tensor]): dense tensors whose sizes will be used to
+          unflatten flat.
+    Returns:
+        Unflattened dense tensors with sizes same as tensors and values from
+        flat.
+    """
+    outputs = []
+    offset = 0
+    for tensor in tensors:
+        numel = tensor.numel()
+        outputs.append(flat.narrow(0, offset, numel).view_as(tensor))
+        offset += numel
+    return tuple(outputs)
+
+
+'''
+This version of DistributedDataParallel is designed to be used in conjunction with the multiproc.py
+launcher included with this example. It assumes that your run is using multiprocess with 1
+GPU/process, that the model is on the correct device, and that torch.set_device has been
+used to set the device.
+Parameters are broadcasted to the other processes on initialization of DistributedDataParallel,
+and will be allreduced at the finish of the backward pass.
+'''
+
+
+class DistributedDataParallel(Module):
+
+    def __init__(self, module):
+        super(DistributedDataParallel, self).__init__()
+        # fallback for PyTorch 0.3
+        if not hasattr(dist, '_backend'):
+            self.warn_on_half = True
+        else:
+            self.warn_on_half = True if dist._backend == dist.dist_backend.GLOO else False
+
+        self.module = module
+
+        for p in self.module.state_dict().values():
+            if not torch.is_tensor(p):
+                continue
+            dist.broadcast(p, 0)
+
+        def allreduce_params():
+            if(self.needs_reduction):
+                self.needs_reduction = False
+                buckets = {}
+                for param in self.module.parameters():
+                    if param.requires_grad and param.grad is not None:
+                        tp = type(param.data)
+                        if tp not in buckets:
+                            buckets[tp] = []
+                        buckets[tp].append(param)
+                if self.warn_on_half:
+                    if torch.cuda.HalfTensor in buckets:
+                        print("WARNING: gloo dist backend for half parameters may be extremely slow." +
+                              " It is recommended to use the NCCL backend in this case. This currently requires" +
+                              "PyTorch built from top of tree master.")
+                        self.warn_on_half = False
+
+                for tp in buckets:
+                    bucket = buckets[tp]
+                    grads = [param.grad.data for param in bucket]
+                    coalesced = _flatten_dense_tensors(grads)
+                    dist.all_reduce(coalesced)
+                    coalesced /= dist.get_world_size()
+                    for buf, synced in zip(grads, _unflatten_dense_tensors(coalesced, grads)):
+                        buf.copy_(synced)
+
+        for param in list(self.module.parameters()):
+            def allreduce_hook(*unused):
+                param._execution_engine.queue_callback(allreduce_params)
+            if param.requires_grad:
+                param.register_hook(allreduce_hook)
+
+    def forward(self, *inputs, **kwargs):
+        self.needs_reduction = True
+        return self.module(*inputs, **kwargs)
+
+    '''
+    def _sync_buffers(self):
+        buffers = list(self.module._all_buffers())
+        if len(buffers) > 0:
+            # cross-node buffer sync
+            flat_buffers = _flatten_dense_tensors(buffers)
+            dist.broadcast(flat_buffers, 0)
+            for buf, synced in zip(buffers, _unflatten_dense_tensors(flat_buffers, buffers)):
+                buf.copy_(synced)
+     def train(self, mode=True):
+        # Clear NCCL communicator and CUDA event cache of the default group ID,
+        # These cache will be recreated at the later call. This is currently a
+        # work-around for a potential NCCL deadlock.
+        if dist._backend == dist.dist_backend.NCCL:
+            dist._clear_group_cache()
+        super(DistributedDataParallel, self).train(mode)
+        self.module.train(mode)
+    '''
+
+
+'''
+Modifies existing model to do gradient allreduce, but doesn't change class
+so you don't need "module"
+'''
+
+
+def apply_gradient_allreduce(module):
+        if not hasattr(dist, '_backend'):
+            module.warn_on_half = True
+        else:
+            module.warn_on_half = True if dist._backend == dist.dist_backend.GLOO else False
+
+        for p in module.state_dict().values():
+            if not torch.is_tensor(p):
+                continue
+            dist.broadcast(p, 0)
+
+        def allreduce_params():
+            if module.needs_reduction:
+                module.needs_reduction = False
+                buckets = {}
+                for param in module.parameters():
+                    if param.requires_grad and param.grad is not None:
+                        tp = type(param.data)
+                        if tp not in buckets:
+                            buckets[tp] = []
+                        buckets[tp].append(param)
+                if module.warn_on_half:
+                    if torch.cuda.HalfTensor in buckets:
+                        print("WARNING: gloo dist backend for half parameters may be extremely slow." +
+                              " It is recommended to use the NCCL backend in this case. This currently requires" +
+                              "PyTorch built from top of tree master.")
+                        module.warn_on_half = False
+
+                for tp in buckets:
+                    bucket = buckets[tp]
+                    grads = [param.grad.data for param in bucket]
+                    coalesced = _flatten_dense_tensors(grads)
+                    dist.all_reduce(coalesced)
+                    coalesced /= dist.get_world_size()
+                    for buf, synced in zip(grads, _unflatten_dense_tensors(coalesced, grads)):
+                        buf.copy_(synced)
+
+        for param in list(module.parameters()):
+            def allreduce_hook(*unused):
+                Variable._execution_engine.queue_callback(allreduce_params)
+            if param.requires_grad:
+                param.register_hook(allreduce_hook)
+
+        def set_needs_reduction(self, input, output):
+            self.needs_reduction = True
+
+        module.register_forward_hook(set_needs_reduction)
+        return module

fp16_optimizer.py

@@ -0,0 +1,385 @@
+import torch
+from torch import nn
+from torch.autograd import Variable
+from torch.nn.parameter import Parameter
+from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors
+
+from loss_scaler import DynamicLossScaler, LossScaler
+
+FLOAT_TYPES = (torch.FloatTensor, torch.cuda.FloatTensor)
+HALF_TYPES = (torch.HalfTensor, torch.cuda.HalfTensor)
+
+
+def conversion_helper(val, conversion):
+    """Apply conversion to val. Recursively apply conversion if `val` is a nested tuple/list structure."""
+    if not isinstance(val, (tuple, list)):
+        return conversion(val)
+    rtn =  [conversion_helper(v, conversion) for v in val]
+    if isinstance(val, tuple):
+        rtn = tuple(rtn)
+    return rtn
+
+
+def fp32_to_fp16(val):
+    """Convert fp32 `val` to fp16"""
+    def half_conversion(val):
+        val_typecheck = val
+        if isinstance(val_typecheck, (Parameter, Variable)):
+            val_typecheck = val.data
+        if isinstance(val_typecheck, FLOAT_TYPES):
+            val = val.half()
+        return val
+    return conversion_helper(val, half_conversion)
+
+
+def fp16_to_fp32(val):
+    """Convert fp16 `val` to fp32"""
+    def float_conversion(val):
+        val_typecheck = val
+        if isinstance(val_typecheck, (Parameter, Variable)):
+            val_typecheck = val.data
+        if isinstance(val_typecheck, HALF_TYPES):
+            val = val.float()
+        return val
+    return conversion_helper(val, float_conversion)
+
+
+class FP16_Module(nn.Module):
+    def __init__(self, module):
+        super(FP16_Module, self).__init__()
+        self.add_module('module', module.half())
+
+    def forward(self, *inputs, **kwargs):
+        return fp16_to_fp32(self.module(*(fp32_to_fp16(inputs)), **kwargs))
+
+
+class FP16_Optimizer(object):
+    """
+    FP16_Optimizer is designed to wrap an existing PyTorch optimizer,
+    and enable an fp16 model to be trained using a master copy of fp32 weights.
+
+    Args:
+        optimizer (torch.optim.optimizer):  Existing optimizer containing initialized fp16 parameters.  Internally, FP16_Optimizer replaces the passed optimizer's fp16 parameters with new fp32 parameters copied from the original ones.  FP16_Optimizer also stores references to the original fp16 parameters, and updates these fp16 parameters from the master fp32 copy after each step.
+        static_loss_scale (float, optional, default=1.0):  Loss scale used internally to scale fp16 gradients computed by the model.  Scaled gradients will be copied to fp32, then downscaled before being applied to the fp32 master params, so static_loss_scale should not affect learning rate.
+        dynamic_loss_scale (bool, optional, default=False):  Use dynamic loss scaling.  If True, this will override any static_loss_scale option.
+
+    """
+
+    def __init__(self, optimizer, static_loss_scale=1.0, dynamic_loss_scale=False):
+        if not torch.cuda.is_available:
+            raise SystemError('Cannot use fp16 without CUDA')
+
+        self.fp16_param_groups = []
+        self.fp32_param_groups = []
+        self.fp32_flattened_groups = []
+        for i, param_group in enumerate(optimizer.param_groups):
+            print("FP16_Optimizer processing param group {}:".format(i))
+            fp16_params_this_group = []
+            fp32_params_this_group = []
+            for param in param_group['params']:
+                if param.requires_grad:
+                    if param.type() == 'torch.cuda.HalfTensor':
+                        print("FP16_Optimizer received torch.cuda.HalfTensor with {}"
+                              .format(param.size()))
+                        fp16_params_this_group.append(param)
+                    elif param.type() == 'torch.cuda.FloatTensor':
+                        print("FP16_Optimizer received torch.cuda.FloatTensor with {}"
+                              .format(param.size()))
+                        fp32_params_this_group.append(param)
+                    else:
+                        raise TypeError("Wrapped parameters must be either "
+                                        "torch.cuda.FloatTensor or torch.cuda.HalfTensor. "
+                                        "Received {}".format(param.type()))
+
+            fp32_flattened_this_group = None
+            if len(fp16_params_this_group) > 0:
+                fp32_flattened_this_group = _flatten_dense_tensors(
+                    [param.detach().data.clone().float() for param in fp16_params_this_group])
+
+                fp32_flattened_this_group = Variable(fp32_flattened_this_group, requires_grad = True)
+
+                fp32_flattened_this_group.grad = fp32_flattened_this_group.new(
+                    *fp32_flattened_this_group.size())
+
+            # python's lovely list concatenation via +
+            if fp32_flattened_this_group is not None:
+                param_group['params'] = [fp32_flattened_this_group] + fp32_params_this_group
+            else:
+                param_group['params'] = fp32_params_this_group
+
+            self.fp16_param_groups.append(fp16_params_this_group)
+            self.fp32_param_groups.append(fp32_params_this_group)
+            self.fp32_flattened_groups.append(fp32_flattened_this_group)
+
+        # print("self.fp32_flattened_groups = ", self.fp32_flattened_groups)
+        # print("self.fp16_param_groups = ", self.fp16_param_groups)
+
+        self.optimizer = optimizer.__class__(optimizer.param_groups)
+
+        # self.optimizer.load_state_dict(optimizer.state_dict())
+
+        self.param_groups = self.optimizer.param_groups
+
+        if dynamic_loss_scale:
+            self.dynamic_loss_scale = True
+            self.loss_scaler = DynamicLossScaler()
+        else:
+            self.dynamic_loss_scale = False
+            self.loss_scaler = LossScaler(static_loss_scale)
+
+        self.overflow = False
+        self.first_closure_call_this_step = True
+
+    def zero_grad(self):
+        """
+        Zero fp32 and fp16 parameter grads.
+        """
+        self.optimizer.zero_grad()
+        for fp16_group in self.fp16_param_groups:
+            for param in fp16_group:
+                if param.grad is not None:
+                    param.grad.detach_() # This does appear in torch.optim.optimizer.zero_grad(),
+                                         # but I'm not sure why it's needed.
+                    param.grad.zero_()
+
+    def _check_overflow(self):
+        params = []
+        for group in self.fp16_param_groups:
+            for param in group:
+                params.append(param)
+        for group in self.fp32_param_groups:
+            for param in group:
+                params.append(param)
+        self.overflow = self.loss_scaler.has_overflow(params)
+
+    def _update_scale(self, has_overflow=False):
+        self.loss_scaler.update_scale(has_overflow)
+
+    def _copy_grads_fp16_to_fp32(self):
+        for fp32_group, fp16_group in zip(self.fp32_flattened_groups, self.fp16_param_groups):
+            if len(fp16_group) > 0:
+                # This might incur one more deep copy than is necessary.
+                fp32_group.grad.data.copy_(
+                    _flatten_dense_tensors([fp16_param.grad.data for fp16_param in fp16_group]))
+
+    def _downscale_fp32(self):
+        if self.loss_scale != 1.0:
+            for param_group in self.optimizer.param_groups:
+                for param in param_group['params']:
+                    param.grad.data.mul_(1./self.loss_scale)
+
+    def clip_fp32_grads(self, clip=-1):
+        if not self.overflow:
+            fp32_params = []
+            for param_group in self.optimizer.param_groups:
+                for param in param_group['params']:
+                    fp32_params.append(param)
+            if clip > 0:
+                return torch.nn.utils.clip_grad_norm_(fp32_params, clip)
+
+    def _copy_params_fp32_to_fp16(self):
+        for fp16_group, fp32_group in zip(self.fp16_param_groups, self.fp32_flattened_groups):
+            if len(fp16_group) > 0:
+                for fp16_param, fp32_data in zip(fp16_group, _unflatten_dense_tensors(fp32_group.data, fp16_group)):
+                    fp16_param.data.copy_(fp32_data)
+
+    def state_dict(self):
+        """
+        Returns a dict containing the current state of this FP16_Optimizer instance.
+        This dict contains attributes of FP16_Optimizer, as well as the state_dict
+        of the contained Pytorch optimizer.
+
+        Untested.
+        """
+        state_dict = {}
+        state_dict['loss_scaler'] = self.loss_scaler
+        state_dict['dynamic_loss_scale'] = self.dynamic_loss_scale
+        state_dict['overflow'] = self.overflow
+        state_dict['first_closure_call_this_step'] = self.first_closure_call_this_step
+        state_dict['optimizer_state_dict'] = self.optimizer.state_dict()
+        return state_dict
+
+    def load_state_dict(self, state_dict):
+        """
+        Loads a state_dict created by an earlier call to state_dict.
+
+        Untested.
+        """
+        self.loss_scaler = state_dict['loss_scaler']
+        self.dynamic_loss_scale = state_dict['dynamic_loss_scale']
+        self.overflow = state_dict['overflow']
+        self.first_closure_call_this_step = state_dict['first_closure_call_this_step']
+        self.optimizer.load_state_dict(state_dict['optimizer_state_dict'])
+
+    def step(self, closure=None): # could add clip option.
+        """
+        If no closure is supplied, step should be called after fp16_optimizer_obj.backward(loss).
+        step updates the fp32 master copy of parameters using the optimizer supplied to
+        FP16_Optimizer's constructor, then copies the updated fp32 params into the fp16 params
+        originally referenced by Fp16_Optimizer's constructor, so the user may immediately run
+        another forward pass using their model.
+
+        If a closure is supplied, step may be called without a prior call to self.backward(loss).
+        However, the user should take care that any loss.backward() call within the closure
+        has been replaced by fp16_optimizer_obj.backward(loss).
+
+        Args:
+           closure (optional):  Closure that will be supplied to the underlying optimizer originally passed to FP16_Optimizer's constructor.  closure should call zero_grad on the FP16_Optimizer object, compute the loss, call .backward(loss), and return the loss.
+
+        Closure example::
+
+            # optimizer is assumed to be an FP16_Optimizer object, previously constructed from an
+            # existing pytorch optimizer.
+            for input, target in dataset:
+                def closure():
+                    optimizer.zero_grad()
+                    output = model(input)
+                    loss = loss_fn(output, target)
+                    optimizer.backward(loss)
+                    return loss
+                optimizer.step(closure)
+
+        .. note::
+            The only changes that need to be made compared to
+            `ordinary optimizer closures`_ are that "optimizer" itself should be an instance of
+            FP16_Optimizer, and that the call to loss.backward should be replaced by
+            optimizer.backward(loss).
+
+        .. warning::
+            Currently, calling step with a closure is not compatible with dynamic loss scaling.
+
+        .. _`ordinary optimizer closures`:
+            http://pytorch.org/docs/master/optim.html#optimizer-step-closure
+        """
+        if closure is not None and isinstance(self.loss_scaler, DynamicLossScaler):
+            raise TypeError("Using step with a closure is currently not "
+                            "compatible with dynamic loss scaling.")
+
+        scale = self.loss_scaler.loss_scale
+        self._update_scale(self.overflow)
+
+        if self.overflow:
+            print("OVERFLOW! Skipping step. Attempted loss scale: {}".format(scale))
+            return
+
+        if closure is not None:
+            self._step_with_closure(closure)
+        else:
+            self.optimizer.step()
+
+        self._copy_params_fp32_to_fp16()
+
+        return
+
+    def _step_with_closure(self, closure):
+        def wrapped_closure():
+            if self.first_closure_call_this_step:
+                """
+                We expect that the fp16 params are initially fresh on entering self.step(),
+                so _copy_params_fp32_to_fp16() is unnecessary the first time wrapped_closure()
+                is called within self.optimizer.step().
+                """
+                self.first_closure_call_this_step = False
+            else:
+                """
+                If self.optimizer.step() internally calls wrapped_closure more than once,
+                it may update the fp32 params after each call.  However, self.optimizer
+                doesn't know about the fp16 params at all.  If the fp32 params get updated,
+                we can't rely on self.optimizer to refresh the fp16 params.  We need
+                to handle that manually:
+                """
+                self._copy_params_fp32_to_fp16()
+
+            """
+            Our API expects the user to give us ownership of the backward() call by
+            replacing all calls to loss.backward() with optimizer.backward(loss).
+            This requirement holds whether or not the call to backward() is made within
+            a closure.
+            If the user is properly calling optimizer.backward(loss) within "closure,"
+            calling closure() here will give the fp32 master params fresh gradients
+            for the optimizer to play with,
+            so all wrapped_closure needs to do is call closure() and return the loss.
+            """
+            temp_loss = closure()
+            return temp_loss
+
+        self.optimizer.step(wrapped_closure)
+
+        self.first_closure_call_this_step = True
+
+    def backward(self, loss, update_fp32_grads=True):
+        """
+        fp16_optimizer_obj.backward performs the following conceptual operations:
+
+        fp32_loss = loss.float() (see first Note below)
+
+        scaled_loss = fp32_loss*loss_scale
+
+        scaled_loss.backward(), which accumulates scaled gradients into the .grad attributes of the
+        fp16 model's leaves.
+
+        fp16 grads are then copied to the stored fp32 params' .grad attributes (see second Note).
+
+        Finally, fp32 grads are divided by loss_scale.
+
+        In this way, after fp16_optimizer_obj.backward, the fp32 parameters have fresh gradients,
+        and fp16_optimizer_obj.step may be called.
+
+        .. note::
+            Converting the loss to fp32 before applying the loss scale provides some
+            additional safety against overflow if the user has supplied an fp16 value.
+            However, for maximum overflow safety, the user should
+            compute the loss criterion (MSE, cross entropy, etc) in fp32 before supplying it to
+            fp16_optimizer_obj.backward.
+
+        .. note::
+            The gradients found in an fp16 model's leaves after a call to
+            fp16_optimizer_obj.backward should not be regarded as valid in general,
+            because it's possible
+            they have been scaled (and in the case of dynamic loss scaling,
+            the scale factor may silently change over time).
+            If the user wants to inspect gradients after a call to fp16_optimizer_obj.backward,
+            he/she should query the .grad attribute of FP16_Optimizer's stored fp32 parameters.
+
+        Args:
+            loss:  The loss output by the user's model.  loss may be either float or half (but see first Note above).
+            update_fp32_grads (bool, optional, default=True):  Option to copy fp16 grads to fp32 grads on this call.  By setting this to False, the user can delay this copy, which is useful to eliminate redundant fp16->fp32 grad copies if fp16_optimizer_obj.backward is being called on multiple losses in one iteration.  If set to False, the user becomes responsible for calling fp16_optimizer_obj.update_fp32_grads before calling fp16_optimizer_obj.step.
+
+        Example::
+
+            # Ordinary operation:
+            optimizer.backward(loss)
+
+            # Naive operation with multiple losses (technically valid, but less efficient):
+            # fp32 grads will be correct after the second call,  but
+            # the first call incurs an unnecessary fp16->fp32 grad copy.
+            optimizer.backward(loss1)
+            optimizer.backward(loss2)
+
+            # More efficient way to handle multiple losses:
+            # The fp16->fp32 grad copy is delayed until fp16 grads from all
+            # losses have been accumulated.
+            optimizer.backward(loss1, update_fp32_grads=False)
+            optimizer.backward(loss2, update_fp32_grads=False)
+            optimizer.update_fp32_grads()
+        """
+        self.loss_scaler.backward(loss.float())
+        if update_fp32_grads:
+            self.update_fp32_grads()
+
+    def update_fp32_grads(self):
+        """
+        Copy the .grad attribute from stored references to fp16 parameters to
+        the .grad attribute of the master fp32 parameters that are directly
+        updated by the optimizer.  :attr:`update_fp32_grads` only needs to be called if
+        fp16_optimizer_obj.backward was called with update_fp32_grads=False.
+        """
+        if self.dynamic_loss_scale:
+            self._check_overflow()
+            if self.overflow: return
+        self._copy_grads_fp16_to_fp32()
+        self._downscale_fp32()
+
+    @property
+    def loss_scale(self):
+        return self.loss_scaler.loss_scale

loss_function.py

@@ -0,0 +1,25 @@
+from torch import nn
+
+
+class Tacotron2Loss(nn.Module):
+    def __init__(self):
+        super(Tacotron2Loss, self).__init__()
+
+    def forward(self, model_output, targets):
+        mel_target, gate_target = targets[0], targets[1]
+        mel_target.requires_grad = False
+        gate_target.requires_grad = False
+        # Ensures dimension 1 will be size 1, the rest can be adapted. It is a column of length 189 with all zeroes
+        # till the end of the current sequence, which is filled with 1's
+        gate_target = gate_target.view(-1, 1)
+
+        mel_out, mel_out_postnet, gate_out, _, _ = model_output
+        gate_out = gate_out.view(-1, 1)
+        # Mean Square Error (L2) loss function for decoder generation + post net generation
+        mel_loss = nn.MSELoss()(mel_out, mel_target) + \
+            nn.MSELoss()(mel_out_postnet, mel_target)
+        # Binary Cross Entropy with a Sigmoid layer combined. It is more efficient than using a plain Sigmoid
+        # followed by a BCELoss as, by combining the operations into one layer, we take advantage of the log-sum-exp
+        # trick for numerical stability
+        gate_loss = nn.BCEWithLogitsLoss()(gate_out, gate_target)
+        return mel_loss + gate_loss

loss_scaler.py

@@ -0,0 +1,79 @@
+
+
+class LossScaler:
+
+    def __init__(self, scale=1):
+        self.cur_scale = scale
+
+    # `params` is a list / generator of torch.Variable
+    def has_overflow(self, params):
+        return False
+
+    # `x` is a torch.Tensor
+    def _has_inf_or_nan(x):
+        return False
+
+    # `overflow` is boolean indicating whether we overflowed in gradient
+    def update_scale(self, overflow):
+        pass
+
+    @property
+    def loss_scale(self):
+        return self.cur_scale
+
+    def scale_gradient(self, module, grad_in, grad_out):
+        return tuple(self.loss_scale * g for g in grad_in)
+
+    def backward(self, loss):
+        scaled_loss = loss*self.loss_scale
+        scaled_loss.backward()
+
+class DynamicLossScaler:
+
+    def __init__(self,
+                 init_scale=2**32,
+                 scale_factor=2.,
+                 scale_window=1000):
+        self.cur_scale = init_scale
+        self.cur_iter = 0
+        self.last_overflow_iter = -1
+        self.scale_factor = scale_factor
+        self.scale_window = scale_window
+
+    # `params` is a list / generator of torch.Variable
+    def has_overflow(self, params):
+        for p in params:
+            if p.grad is not None and DynamicLossScaler._has_inf_or_nan(p.grad.data):
+                return True
+
+        return False
+
+    # `x` is a torch.Tensor
+    def _has_inf_or_nan(x):
+        cpu_sum = float(x.float().sum())
+        if cpu_sum == float('inf') or cpu_sum == -float('inf') or cpu_sum != cpu_sum:
+            return True
+        return False
+
+    # `overflow` is boolean indicating whether we overflowed in gradient
+    def update_scale(self, overflow):
+        if overflow:
+            #self.cur_scale /= self.scale_factor
+            self.cur_scale = max(self.cur_scale/self.scale_factor, 1)
+            self.last_overflow_iter = self.cur_iter
+        else:
+            if (self.cur_iter - self.last_overflow_iter) % self.scale_window == 0:
+                self.cur_scale *= self.scale_factor
+#        self.cur_scale = 1
+        self.cur_iter += 1
+
+    @property
+    def loss_scale(self):
+        return self.cur_scale
+
+    def scale_gradient(self, module, grad_in, grad_out):
+        return tuple(self.loss_scale * g for g in grad_in)
+
+    def backward(self, loss):
+        scaled_loss = loss*self.loss_scale
+        scaled_loss.backward()

multiproc.py

@@ -0,0 +1,23 @@
+import time
+import torch
+import sys
+import subprocess
+
+argslist = list(sys.argv)[1:]
+num_gpus = torch.cuda.device_count()
+argslist.append('--n_gpus={}'.format(num_gpus))
+workers = []
+job_id = time.strftime("%Y_%m_%d-%H%M%S")
+argslist.append("--group_name=group_{}".format(job_id))
+
+for i in range(num_gpus):
+    argslist.append('--rank={}'.format(i))
+    stdout = None if i == 0 else open("logs/{}_GPU_{}.log".format(job_id, i),
+                                      "w")
+    print(argslist)
+    p = subprocess.Popen([str(sys.executable)]+argslist, stdout=stdout)
+    workers.append(p)
+    argslist = argslist[:-1]
+
+for p in workers:
+    p.wait()