relax learning rate constraint, as @rom1504 wants to try a higher one

dalle2_pytorch/dalle2_pytorch.py

@@ -335,6 +335,10 @@ def approx_standard_normal_cdf(x):
 def discretized_gaussian_log_likelihood(x, *, means, log_scales, thres = 0.999):
     assert x.shape == means.shape == log_scales.shape
 
+    # attempting to correct nan gradients when learned variance is turned on
+    # in the setting of deepspeed fp16
+    eps = 1e-12 if x.dtype == torch.float32 else 1e-5
+
     centered_x = x - means
     inv_stdv = torch.exp(-log_scales)
     plus_in = inv_stdv * (centered_x + 1. / 255.)
@@ -349,7 +353,7 @@ def discretized_gaussian_log_likelihood(x, *, means, log_scales, thres = 0.999):
         log_cdf_plus,
         torch.where(x > thres,
             log_one_minus_cdf_min,
-            log(cdf_delta)))
+            log(cdf_delta, eps = eps)))
 
     return log_probs
 
@@ -704,7 +708,7 @@ class Attention(nn.Module):
         sim = sim - sim.amax(dim = -1, keepdim = True).detach()
         sim = sim * self.pb_relax_alpha
 
-        attn = sim.softmax(dim = -1, dtype = torch.float32)
+        attn = sim.softmax(dim = -1)
         attn = self.dropout(attn)
 
         # aggregate values
@@ -1127,11 +1131,12 @@ class SinusoidalPosEmb(nn.Module):
         self.dim = dim
 
     def forward(self, x):
+        dtype, device = x.dtype, x.device
         half_dim = self.dim // 2
         emb = math.log(10000) / (half_dim - 1)
-        emb = torch.exp(torch.arange(half_dim, device = x.device) * -emb)
+        emb = torch.exp(torch.arange(half_dim, device = device, dtype = dtype) * -emb)
         emb = rearrange(x, 'i -> i 1') * rearrange(emb, 'j -> 1 j')
-        return torch.cat((emb.sin(), emb.cos()), dim = -1)
+        return torch.cat((emb.sin(), emb.cos()), dim = -1).type(dtype)
 
 class Block(nn.Module):
     def __init__(
@@ -1272,7 +1277,7 @@ class CrossAttention(nn.Module):
         sim = sim - sim.amax(dim = -1, keepdim = True).detach()
         sim = sim * self.pb_relax_alpha
 
-        attn = sim.softmax(dim = -1, dtype = torch.float32)
+        attn = sim.softmax(dim = -1)
 
         out = einsum('b h i j, b h j d -> b h i d', attn, v)
         out = rearrange(out, 'b h n d -> b n (h d)')
@@ -1626,6 +1631,7 @@ class Unet(nn.Module):
 
         # time conditioning
 
+        time = time.type_as(x)
         time_hiddens = self.to_time_hiddens(time)
 
         time_tokens = self.to_time_tokens(time_hiddens)

dalle2_pytorch/trainer.py

@@ -173,14 +173,26 @@ class DiffusionPriorTrainer(nn.Module):
         super().__init__()
         assert isinstance(diffusion_prior, DiffusionPrior)
         assert not exists(accelerator) or isinstance(accelerator, Accelerator)
-        assert exists(accelerator) or exists(device), "You must supply some method of obtaining a device."
         ema_kwargs, kwargs = groupby_prefix_and_trim('ema_', kwargs)
 
+        # verbosity
+
+        self.verbose = verbose
+
         # assign some helpful member vars
+
         self.accelerator = accelerator
-        self.device = accelerator.device if exists(accelerator) else device
         self.text_conditioned = diffusion_prior.condition_on_text_encodings
 
+        # setting the device
+
+        if not exists(accelerator) and not exists(device):
+            diffusion_prior_device = next(diffusion_prior.parameters()).device
+            self.print(f'accelerator not given, and device not specified: defaulting to device of diffusion prior parameters - {diffusion_prior_device}')
+            self.device = diffusion_prior_device
+        else:
+            self.device = accelerator.device if exists(accelerator) else device
+
         # save model
 
         self.diffusion_prior = diffusion_prior
@@ -214,13 +226,9 @@ class DiffusionPriorTrainer(nn.Module):
 
         self.max_grad_norm = max_grad_norm
 
-        # verbosity
-
-        self.verbose = verbose
-
         # track steps internally
 
-        self.register_buffer('step', torch.tensor([0]))
+        self.register_buffer('step', torch.tensor([0], device = self.device))
 
     # accelerator wrappers
 
@@ -465,7 +473,7 @@ class DecoderTrainer(nn.Module):
 
         lr, wd, eps, warmup_steps = map(partial(cast_tuple, length = self.num_unets), (lr, wd, eps, warmup_steps))
 
-        assert all([unet_lr < 1e-3 for unet_lr in lr]), 'your learning rate is too high, recommend sticking with 1e-4, at most 5e-4'
+        assert all([unet_lr <= 1e-2 for unet_lr in lr]), 'your learning rate is too high, recommend sticking with 1e-4, at most 5e-4'
 
         optimizers = []
         schedulers = []
@@ -550,7 +558,7 @@ class DecoderTrainer(nn.Module):
         if only_model:
             return loaded_obj
 
-        for ind, last_step in zip(range(0, self.num_unets), self.steps.cpu().unbind()):
+        for ind, last_step in zip(range(0, self.num_unets), self.steps.tolist()):
 
             optimizer_key = f'optim{ind}'
             optimizer = getattr(self, optimizer_key)

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '0.16.5'
+__version__ = '0.16.13'