add a new image_size parameter in train_dalle and generate

VAE models can be use with patches of any size.
For example a model trained on 16x16 patches can still be used on 32x32 patches
that increase the seq length from 256 to 1024 in dalle

dalle_pytorch/vae.py

@@ -96,14 +96,14 @@ def download(url, filename = None, root = CACHE_PATH):
 # pretrained Discrete VAE from OpenAI
 
 class OpenAIDiscreteVAE(nn.Module):
-    def __init__(self):
+    def __init__(self, image_size=256):
         super().__init__()
 
         self.enc = load_model(download(OPENAI_VAE_ENCODER_PATH))
         self.dec = load_model(download(OPENAI_VAE_DECODER_PATH))
 
         self.num_layers = 3
-        self.image_size = 256
+        self.image_size = image_size
         self.num_tokens = 8192
 
     @torch.no_grad()
@@ -130,7 +130,7 @@ def forward(self, img):
 # https://arxiv.org/abs/2012.09841
 
 class VQGanVAE(nn.Module):
-    def __init__(self, vqgan_model_path, vqgan_config_path):
+    def __init__(self, image_size=256, vqgan_model_path=None, vqgan_config_path=None):
         super().__init__()
 
         if vqgan_model_path is None:
@@ -155,7 +155,7 @@ def __init__(self, vqgan_model_path, vqgan_config_path):
         self.model = model
 
         self.num_layers = int(log(config.model.params.ddconfig.attn_resolutions[0])/log(2))
-        self.image_size = 256
+        self.image_size = image_size
         self.num_tokens = config.model.params.n_embed
 
         self._register_external_parameters()

generate.py

@@ -43,6 +43,9 @@
 parser.add_argument('--top_k', type = float, default = 0.9, required = False,
                     help='top k filter threshold')
 
+parser.add_argument('--image_size', type = int, default = 256, required = False,
+                    help='image size')
+
 parser.add_argument('--outputs_dir', type = str, default = './outputs', required = False,
                     help='output directory')
 
@@ -81,12 +84,14 @@ def exists(val):
 
 dalle_params.pop('vae', None) # cleanup later
 
+IMAGE_SIZE = args.image_size
+
 if vae_params is not None:
-    vae = DiscreteVAE(**vae_params)
+    vae = DiscreteVAE(IMAGE_SIZE, **vae_params[1:])
 elif not args.taming:
-    vae = OpenAIDiscreteVAE()
+    vae = OpenAIDiscreteVAE(IMAGE_SIZE)
 else:
-    vae = VQGanVAE(args.vqgan_model_path, args.vqgan_config_path)
+    vae = VQGanVAE(IMAGE_SIZE, args.vqgan_model_path, args.vqgan_config_path)
 
 
 dalle = DALLE(vae = vae, **dalle_params).cuda()
@@ -95,8 +100,6 @@ def exists(val):
 
 # generate images
 
-image_size = vae.image_size
-
 texts = args.text.split('|')
 
 for text in tqdm(texts):

train_dalle.py

@@ -108,6 +108,8 @@
 
 model_group.add_argument('--loss_img_weight', default = 7, type = int, help = 'Image loss weight')
 
+model_group.add_argument('--image_size', default = 256, type = int, help = 'Image size')
+
 model_group.add_argument('--attn_types', default = 'full', type = str, help = 'comma separated list of attention types. attention type can be: full or sparse or axial_row or axial_col or conv_like.')
 
 args = parser.parse_args()
@@ -155,6 +157,7 @@ def cp_path_to_dir(cp_path, tag):
 SAVE_EVERY_N_STEPS = args.save_every_n_steps
 KEEP_N_CHECKPOINTS = args.keep_n_checkpoints
 
+IMAGE_SIZE = args.image_size
 MODEL_DIM = args.dim
 TEXT_SEQ_LEN = args.text_seq_len
 DEPTH = args.depth
@@ -202,17 +205,16 @@ def cp_path_to_dir(cp_path, tag):
     scheduler_state = loaded_obj.get('scheduler_state')
 
     if vae_params is not None:
-        vae = DiscreteVAE(**vae_params)
+        vae = DiscreteVAE(IMAGE_SIZE, **vae_params[1:])
     else:
         if args.taming:
-            vae = VQGanVAE(VQGAN_MODEL_PATH, VQGAN_CONFIG_PATH)
+            vae = VQGanVAE(IMAGE_SIZE, VQGAN_MODEL_PATH, VQGAN_CONFIG_PATH)
         else:
-            vae = OpenAIDiscreteVAE()
+            vae = OpenAIDiscreteVAE(IMAGE_SIZE)
 
     dalle_params = dict(
         **dalle_params
     )
-    IMAGE_SIZE = vae.image_size
     resume_epoch = loaded_obj.get('epoch', 0)
 else:
     if exists(VAE_PATH):
@@ -228,19 +230,17 @@ def cp_path_to_dir(cp_path, tag):
 
         vae_params, weights = loaded_obj['hparams'], loaded_obj['weights']
 
-        vae = DiscreteVAE(**vae_params)
+        vae = DiscreteVAE(IMAGE_SIZE, **vae_params[1:])
         vae.load_state_dict(weights)
     else:
         if distr_backend.is_root_worker():
             print('using pretrained VAE for encoding images to tokens')
         vae_params = None
 
         if args.taming:
-            vae = VQGanVAE(VQGAN_MODEL_PATH, VQGAN_CONFIG_PATH)
+            vae = VQGanVAE(IMAGE_SIZE, VQGAN_MODEL_PATH, VQGAN_CONFIG_PATH)
         else:
-            vae = OpenAIDiscreteVAE()
-
-    IMAGE_SIZE = vae.image_size
+            vae = OpenAIDiscreteVAE(IMAGE_SIZE)
 
     dalle_params = dict(
         num_text_tokens=tokenizer.vocab_size,