Combine interfaces

app.py

@@ -42,6 +42,7 @@ class DropOutput(Callback):
 
 
 # initialize only once
+# Takes about 2 minutes (126 seconds) to generate an image in Huggingface spaces on CPU
 model = MinDalle(
     models_root='./pretrained',
     dtype=float32,
@@ -52,7 +53,9 @@ model = MinDalle(
 
 
 def gen_image(prompt):
-    torch.set_grad_enabled(False)  # Hugging Space faces seems to run out of memory if grads are not disabled
+    # See https://huggingface.co/spaces/pootow/min-dalle/blob/main/app.py
+    # Hugging Space faces seems to run out of memory if grads are not disabled
+    torch.set_grad_enabled(False)
     print(f'RUNNING gen_image with pronpt: {prompt}')
     images = model.generate_images(
         text=prompt,
@@ -70,24 +73,30 @@ def gen_image(prompt):
     return Image.fromarray(images[0])
 
 
-# gpu = False
-# # init only once
-# learner = load_learner('export.pkl',
-#                        cpu=not gpu)  # cpu=False uses GPU; make sure installed torch is GPU e.g. `pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116`
-#
-#
-# def run_model(name):
-#     prompt = f"Name: {name}\r\n"
-#     prompt_ids = tokenizer.encode(prompt)
-#     if gpu:
-#         inp = tensor(prompt_ids)[None].cuda()  # Use .cuda() for torch GPU
-#     else:
-#         inp = tensor(prompt_ids)[None]
-#     preds = learner.model.generate(inp, max_length=1024, num_beams=5, temperature=1.5, do_sample=True)
-#     result = tokenizer.decode(preds[0].cpu().numpy())
-#     result = result.split('###')[0].replace(r'\r\n', '\n')
-#     return result
-
-
-iface = gr.Interface(fn=gen_image, inputs="text", outputs="pil")
+gpu = False
+# init only once
+learner = load_learner('export.pkl',
+                       cpu=not gpu)  # cpu=False uses GPU; make sure installed torch is GPU e.g. `pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116`
+
+
+def run_model(name):
+    prompt = f"Name: {name}\r\n"
+    prompt_ids = tokenizer.encode(prompt)
+    if gpu:
+        inp = tensor(prompt_ids)[None].cuda()  # Use .cuda() for torch GPU
+    else:
+        inp = tensor(prompt_ids)[None]
+    preds = learner.model.generate(inp, max_length=1024, num_beams=5, temperature=1.5, do_sample=True)
+    result = tokenizer.decode(preds[0].cpu().numpy())
+    result = result.split('###')[0].replace(r'\r\n', '\n')
+    return result
+
+
+def run(name):
+    text = run_model(name)
+    pil = gen_image('smiling dog')
+    return text, pil
+
+
+iface = gr.Interface(fn=run, inputs="text", outputs=["text", "pil"])
 iface.launch()