Update hubconf.py with valid NGC checkpoints

[SSD][hubconf] contiguous input tensor

PyTorch/SpeechSynthesis/Tacotron2/tacotron2/model.py

@@ -530,8 +530,8 @@ class Decoder(nn.Module):
 
         self.initialize_decoder_states(memory, mask=None)
 
-        mel_lengths = torch.zeros([memory.size(0)], dtype=torch.int32).cuda()
+        mel_lengths = torch.zeros([memory.size(0)], dtype=torch.int32).to(memory.device)
-        not_finished = torch.ones([memory.size(0)], dtype=torch.int32).cuda()
+        not_finished = torch.ones([memory.size(0)], dtype=torch.int32).to(memory.device)
         mel_outputs, gate_outputs, alignments = [], [], []
         while True:
             decoder_input = self.prenet(decoder_input)

hubconf.py

@@ -3,6 +3,7 @@ import torch
 import os
 import sys
 
+
 # from https://github.com/NVIDIA/DeepLearningExamples/blob/master/PyTorch/SpeechSynthesis/Tacotron2/inference.py
 def checkpoint_from_distributed(state_dict):
     """
@@ -35,6 +36,17 @@ def unwrap_distributed(state_dict):
     return new_state_dict
 
 
+def _download_checkpoint(checkpoint, force_reload):
+    model_dir = os.path.join(torch.hub._get_torch_home(), 'checkpoints')
+    if not os.path.exists(model_dir):
+        os.makedirs(model_dir)
+    ckpt_file = os.path.join(model_dir, os.path.basename(checkpoint))
+    if not os.path.exists(ckpt_file) or force_reload:
+        sys.stderr.write('Downloading checkpoint from {}\n'.format(checkpoint))
+        urllib.request.urlretrieve(checkpoint, ckpt_file)
+    return ckpt_file
+
+
 dependencies = ['torch']
 
 
@@ -66,10 +78,7 @@ def nvidia_ncf(pretrained=True, **kwargs):
             checkpoint = 'https://developer.nvidia.com/joc-ncf-fp16-pyt-20190225'
         else:
             checkpoint = 'https://developer.nvidia.com/joc-ncf-fp32-pyt-20190225'
-        ckpt_file = os.path.basename(checkpoint)
+        ckpt_file = _download_checkpoint(checkpoint, force_reload)
-        if not os.path.exists(ckpt_file) or force_reload:
-            sys.stderr.write('Downloading checkpoint from {}\n'.format(checkpoint))
-            urllib.request.urlretrieve(checkpoint, ckpt_file)
         ckpt = torch.load(ckpt_file)
 
         if checkpoint_from_distributed(ckpt):
@@ -127,13 +136,10 @@ def nvidia_tacotron2(pretrained=True, **kwargs):
 
     if pretrained:
         if fp16:
-            checkpoint = 'https://developer.nvidia.com/joc-tacotron2-fp16-pyt-20190306'
+            checkpoint = 'https://api.ngc.nvidia.com/v2/models/nvidia/tacotron2_pyt_ckpt_amp/versions/19.09.0/files/nvidia_tacotron2pyt_fp16_20190427'
         else:
-            checkpoint = 'https://developer.nvidia.com/joc-tacotron2-fp32-pyt-20190306'
+            checkpoint = 'https://api.ngc.nvidia.com/v2/models/nvidia/tacotron2_pyt_ckpt_fp32/versions/19.09.0/files/nvidia_tacotron2pyt_fp32_20190427'
-        ckpt_file = os.path.basename(checkpoint)
+        ckpt_file = _download_checkpoint(checkpoint, force_reload)
-        if not os.path.exists(ckpt_file) or force_reload:
-            sys.stderr.write('Downloading checkpoint from {}\n'.format(checkpoint))
-            urllib.request.urlretrieve(checkpoint, ckpt_file)
         ckpt = torch.load(ckpt_file)
         state_dict = ckpt['state_dict']
         if checkpoint_from_distributed(state_dict):
@@ -187,13 +193,10 @@ def nvidia_waveglow(pretrained=True, **kwargs):
 
     if pretrained:
         if fp16:
-            checkpoint = 'https://developer.nvidia.com/joc-waveglow-fp16-pyt-20190306'
+            checkpoint = 'https://api.ngc.nvidia.com/v2/models/nvidia/waveglow_ckpt_amp/versions/19.09.0/files/nvidia_waveglowpyt_fp16_20190427'
         else:
-            checkpoint = 'https://developer.nvidia.com/joc-waveglow-fp32-pyt-20190306'
+            checkpoint = 'https://api.ngc.nvidia.com/v2/models/nvidia/waveglow_ckpt_fp32/versions/19.09.0/files/nvidia_waveglowpyt_fp32_20190427'
-        ckpt_file = os.path.basename(checkpoint)
+        ckpt_file = _download_checkpoint(checkpoint, force_reload)
-        if not os.path.exists(ckpt_file) or force_reload:
-            sys.stderr.write('Downloading checkpoint from {}\n'.format(checkpoint))
-            urllib.request.urlretrieve(checkpoint, ckpt_file)
         ckpt = torch.load(ckpt_file)
         state_dict = ckpt['state_dict']
         if checkpoint_from_distributed(state_dict):
@@ -225,13 +228,15 @@ def nvidia_waveglow(pretrained=True, **kwargs):
 def nvidia_ssd_processing_utils():
     import numpy as np
     import skimage
+    from skimage import io, transform
+
     from PyTorch.Detection.SSD.src.utils import dboxes300_coco, Encoder
 
     class Processing:
         @staticmethod
         def load_image(image_path):
             """Code from Loading_Pretrained_Models.ipynb - a Caffe2 tutorial"""
-            img = skimage.img_as_float(skimage.io.imread(image_path))
+            img = skimage.img_as_float(io.imread(image_path))
             if len(img.shape) == 2:
                 img = np.array([img, img, img]).swapaxes(0, 2)
             return img
@@ -243,13 +248,13 @@ def nvidia_ssd_processing_utils():
             if (aspect > 1):
                 # landscape orientation - wide image
                 res = int(aspect * input_height)
-                imgScaled = skimage.transform.resize(img, (input_width, res))
+                imgScaled = transform.resize(img, (input_width, res))
             if (aspect < 1):
                 # portrait orientation - tall image
                 res = int(input_width / aspect)
-                imgScaled = skimage.transform.resize(img, (res, input_height))
+                imgScaled = transform.resize(img, (res, input_height))
             if (aspect == 1):
-                imgScaled = skimage.transform.resize(img, (input_width, input_height))
+                imgScaled = transform.resize(img, (input_width, input_height))
             return imgScaled
 
         @staticmethod
@@ -270,6 +275,7 @@ def nvidia_ssd_processing_utils():
             NHWC = np.array(inputs)
             NCHW = np.swapaxes(np.swapaxes(NHWC, 1, 3), 2, 3)
             tensor = torch.from_numpy(NCHW)
+            tensor = tensor.contiguous()
             tensor = tensor.cuda()
             tensor = tensor.float()
             if fp16:
@@ -326,6 +332,7 @@ def nvidia_ssd_processing_utils():
 
     return Processing()
 
+
 def nvidia_ssd(pretrained=True, **kwargs):
     """Constructs an SSD300 model.
     For detailed information on model input and output, training recipies, inference and performance
@@ -356,14 +363,9 @@ def nvidia_ssd(pretrained=True, **kwargs):
         m = batchnorm_to_float(m)
 
     if pretrained:
-        if fp16:
+        checkpoint = 'https://api.ngc.nvidia.com/v2/models/nvidia/ssd_pyt_ckpt_amp/versions/19.09.0/files/nvidia_ssdpyt_fp16_190826.pt'
-            checkpoint = 'https://developer.nvidia.com/joc-ssd-fp16-pyt-20190225'
+        # ckpt = torch.hub.load_state_dict_from_url(checkpoint, progress=True, check_hash=False)
-        else:
+        ckpt_file = _download_checkpoint(checkpoint, force_reload)
-            checkpoint = 'https://developer.nvidia.com/joc-ssd-fp32-pyt-20190225'
-        ckpt_file = os.path.basename(checkpoint)
-        if not os.path.exists(ckpt_file) or force_reload:
-            sys.stderr.write('Downloading checkpoint from {}\n'.format(checkpoint))
-            urllib.request.urlretrieve(checkpoint, ckpt_file)
         ckpt = torch.load(ckpt_file)
         ckpt = ckpt['model']
         if checkpoint_from_distributed(ckpt):